content
stringlengths 1
15.9M
|
|---|
\section{Introduction}
Selecting $k$ diverse items out of a larger collection of $n$ items
is a classical problem in computer science which naturally emerges in many
tasks such as summarization (select $k$ phrases) and recommendation (select
$k$ articles/ads to show to the user). An increasingly popular approach
to model and quantify diversity in this
subset selection problem is that of determinantal point processes (DPPs).
Given a set $[n] \stackrel{\text{\tiny{def}}}{=} \{1, \ldots, n\}$ of $n$ items and a target
size $k$, one can define a DPP of size $k$ (known as a $k$-DPP)
through an $n\times n$ posivite semi-definite (PSD) similarity matrix
$\mathbf{L}$ (also known as the kernel matrix).
The matrix $\mathbf{L}$ encodes the similarities between items, and the user must choose
it so that $[\mathbf{L}]_{ij}$ is larger the more the $i$-th and $j$-th items are similar.
Given $k$ and $\mathbf{L}$, we define $S \sim k$-${\mathrm{DPP}}(\mathbf{L})$
as a distribution over all ${n \choose k}$ index subsets
$S\subseteq [n]$ of size $k$, such that $\Pr(S) \propto \det(\mathbf{L}_S)$
is proportional to the determinant of the sub-matrix $\mathbf{L}_S$ induced by
the subset. DPPs have found numerous applications in machine
learning, not only for summarization \cite{dpp-summarization,dpp-video,dpp-salient-threads,pmlr-v80-celis18a}
and recommendation \cite{dpp-shopping,NIPS2018_7805}, but also in experimental
design \cite{bayesian-experimental-design,symmetric-polynomials},
stochastic optimization \cite{dpp-minibatch,randomized-newton},
Gaussian Process optimization \cite{NIPS2016_6452},
low-rank approximation
\cite{pca-volume-sampling,more-efficient-volume-sampling,nystrom-multiple-descent}, and
more (recent surveys include
\cite{dpp-ml,dpp-stats,dpp-randnla}). Note that early work on DPPs
focused on a \emph{random-size} variant, which we
denote $S\sim {\mathrm{DPP}}(\mathbf{L})$, where the subset size is allowed to take any value
between $0$ and $n$, and the role of parameter $k$ is replaced by the expected size $\mathbb E[|S|]=
d_{\textnormal{eff}}(\mathbf{L}) \stackrel{\text{\tiny{def}}}{=} \mathrm{tr}\left(\mathbf{L}(\mathbf{L} + \mathbf I)^{-1}\right)$.
The quantity $d_{\textnormal{eff}}(\mathbf{L})$ is known in randomized linear algebra
\cite{ridge-leverage-scores,dpp-randnla} and learning theory
\cite{caponnetto2007optimal} as the effective dimension. While
random-size DPPs exhibit deep connections to many scientific domains
\cite{dpp-independence}, the \emph{fixed-size} $k$-DPPs are typically
more practical from a machine learning stand-point \cite{k-dpp}.
\textbf{Sampling from a $k$-DPP.} The first $k$-DPP samplers
scaled poorly, as they all relied on an eigendecomposition \cite{k-dpp} of
$\mathbf{L}$ taking $\mathcal{O}(n^3)$ time.
Replacing the eigendecomposition with a Cholesky factorization can increase numerical stability \cite{dpp-noeig},
and empirical performance \cite{poulson2019high-performance} thanks to
dynamically-scheduled, shared-memory parallelizations, but still
ultimately require $\mathcal{O}(n^3)$ time. A number of methods have
been proposed which use approximate eigendecomposition
\cite{dpp-coreset,dpp-nystrom} to reduce the computational cost,
however these approaches provide limited guarantees on the accuracy of sampling.
To improve scalability, several approaches
based on Monte-Carlo sampling were introduced, using rejection or Gibbs sampling.
The fastest MCMC sampler for $k$-DPPs, to the best of our knowledge, is
by \cite{rayleigh-mcmc} and has $n\cdot{\mathrm{poly}}(k)$
complexity, i.e., asymptotically much faster than the cost of eigendecomposition.
However these MCMC methods do \emph{not} sample exactly from the $k$-DPP distribution,
and can only guarantee that the final sample will be close in distribution
to a $k$-DPP. Moreover these guarantees only hold \emph{after mixing}, which is difficult
to verify and requires at least $\mathcal{O}(nk^2)$ time, making MCMC methods not applicable
when $n$ is large.
\begin{wrapfigure}{R}{.44\textwidth}
\centering
\vspace{-2\baselineskip}
\begin{tabular}{r|l}
&Complexity
\\
\hline
\cite{k-dpp,dpp-noeig,poulson2019high-performance,more-efficient-volume-sampling}
&$n^3$\\[1mm]
\textsc{DPP-VFX}\xspace \cite{derezinski2019exact}& $n\cdot k^{10} + k^{15}$\\[1mm]
$\alpha$-DPP (this paper)& $(\beta n\cdot k^6+k^9)\sqrt k$
\end{tabular}
\captionof{table}{\small Runtime comparison of exact $k$-DPP
sampling algorithms. Here, $\beta\leq 1$ is the fraction of
items observed by $\alpha$-DPP (see Theorem~\ref{thm:informal-main}).}
\vspace{-3mm}
\label{tab:comp}
\end{wrapfigure}
A recent line of works \cite{derezinski2019exact,dpp-intermediate},
using the ideas from \cite{leveraged-volume-sampling,correcting-bias},
developed sampling algorithms
specially designed for a \emph{random-size} DPP (as opposed to a $k$-DPP), which avoid
expensive decomposition of the kernel while sampling exactly from $S\sim{\mathrm{DPP}}(\mathbf{L})$.
In particular, they showed that it is sufficient to first choose an intermediate subset
$\sigma \subseteq [n]$ sampled i.i.d.\,from the \emph{marginal}
distribution of the DPP, i.e.,\xspace $\mathbf P(i \in \sigma) \approx \mathbf P(i \in S)$, and
then sample from a DPP restricted to the items indexed by
$\sigma$. Since the size of $\sigma$ is typically much less
than $n$, this leads to a more efficient algorithm.
Note that rescaling ${\mathrm{DPP}}(\mathbf{L})$ into ${\mathrm{DPP}}(\alpha\mathbf{L})$ using some constant~$\alpha$
only changes the expected size of $S$ from $d_{\textnormal{eff}}(\mathbf{L})$ to $d_{\textnormal{eff}}(\alpha\mathbf{L})$.
By accurately choosing an appropriate $\alpha_{\star}$, one can boost the probability that the random
size of $S$ is exactly $k$, and convert a DPP sampler into a $k$-DPP sampler
by repeatedly sampling $S \sim {\mathrm{DPP}}(\alpha_{\star}\mathbf{L})$ until $S$ has
size~$k$. Based on this reduction, \citet{derezinski2019exact}
gave the first algorithm (\textsc{DPP-VFX}\xspace) which is capable of exact sampling
from a $k$-DPP in time $n\cdot{\mathrm{poly}}(k)$. However,
when sampling from $k$-DPPs, the approach of
\cite{derezinski2019exact} has \emph{two major limitations}:
\begin{enumerate}[leftmargin=*]
\item \textsc{DPP-VFX}\xspace has an $\Omega(n)$ runtime bottleneck, since it requires
computing all $n$ marginals, one for each item, in
order to define the i.i.d.\/ distribution of $\sigma$, which may
be infeasible for very large $n$.
\item
The reduction used by \cite{derezinski2019exact} to convert a DPP
sampler into a $k$-DPP sampler
increases the time complexity by a factor
of at least $k^4$, resulting in a $\acO(n\cdot k^{10}+k^{15})$ runtime.
\end{enumerate}
In this paper, we address both of these limitations by introducing a
new algorithm called $\alpha$-DPP, which 1) does not need to compute all
of the marginals, and 2) uses a new efficient reduction to convert
from a random-size DPP to a fixed-size $k$-DPP (see Table
\ref{tab:comp} for comparison).
\textbf{Main contribution: uniform intermediate sampling for $k$-DPPs.}
To resolve the $\Omega(n)$ runtime bottleneck, we use an additional
intermediate sample $\rho$ based on \emph{uniform} sub-sampling.
Since uniform sampling can be implemented without looking at the actual items
in the collection, this means that we do not even have to look at any
item outside of $\rho$.
The only necessary assumption required by our approach is that the maximum
entry (i.e.,\xspace similarity) of $\mathbf{L}$ is bounded by a constant $\kappa^2$. However,
to simplify exposition we also assume w.l.o.g.~that $d_{\textnormal{eff}}(\mathbf{L}) \geq k$ (see \Cref{s:binary-search}).
In particular, we 1) sample $\rho$ uniformly out of $[n]$,
then 2) only approximate the marginal probabilities of items in $\rho$
to compute $\sigma$, and finally 3) downsample $\sigma$ into
a DPP sample $S$. To guarantee that $S$ is distributed exactly according to the DPP
it is crucial that $\rho$ is diverse enough. We show
that sampling a $k^2/d_{\textnormal{eff}}(\mathbf{L})$ fraction of $[n]$ into $\rho$ (i.e.,\xspace $|\rho| \approx k^2/d_{\textnormal{eff}}(\mathbf{L}) \cdot n$)
is enough. Since all the expensive computation is performed only
on $\rho$, this gives us a $d_{\textnormal{eff}}(\mathbf{L})/k^2$ speedup
over existing methods.
\vspace{2mm}
\begin{theorem}\label{thm:informal-main}
Given any $\mathbf{L} \succeq \mathbf{0}$ with $\max_{ij} \mathbf{L}_{ij} \leq
\kappa^2$ and $1 \leq k \leq d_{\textnormal{eff}}(\mathbf{L})$, there exists an algorithm that returns
$S \sim k$-${\mathrm{DPP}}(\mathbf{L})$, and with probability $1 - \delta$ runs in time
$$\acO\big((\beta n\cdot k^6+k^9)\sqrt k\log(1/\delta)\big),$$
where $\beta \leq\min\big\{k^2\kappa^2 /d_{\textnormal{eff}}(\mathbf{L}),1\big\}$ is the fraction
of items observed by the algorithm.
\end{theorem}
In the derivation of \Cref{thm:informal-main} we make several novel contributions.
First, we provide a DPP sampler that
given $\mathbf{L}$ and a rescaling $\alpha \leq 1$ leverages a mixture of uniform and rejection sampling
to sample from ${\mathrm{DPP}}(\alpha\mathbf{L})$ observing only an $\alpha\kappa^2k$ fraction of
the items. We then show that the optimal rescaling $\alpha_{\star}$ required
by the reduction from $k$-DPP to DPP can be bounded with $\alpha_{\star} \leq \mathcal{O}(k/d_{\textnormal{eff}}(\mathbf{L}))$,
and thus our rescaling-aware sampler can sample from $k$-DPPs looking
only at a $k^2/d_{\textnormal{eff}}(\mathbf{L})$ fraction of the items. Finally, we provide
an efficient search algorithm to find a close approximation $\hat{\alpha}$
of $\alpha_{\star}$.
\textbf{Model misspecification and computational free lunch.}
Our result can be also interpreted from a perspective of model misspecification.
Note that every time the users define a $k$-DPP they
also implicitly define a random size ${\mathrm{DPP}}(\mathbf{L})$.
Moreover, the natural expected sample size (i.e.,\xspace implicit number of unique items in $[n]$)
of ${\mathrm{DPP}}(\mathbf{L})$ is $d_{\textnormal{eff}}(\mathbf{L})$,
which does not depend on the desired size~$k$.
Therefore, if $\mathbf{L}$ is not chosen appropriately
$d_{\textnormal{eff}}(\mathbf{L})$ might be much larger than $k$, and the $k$-DPP is selecting
$k$ unique items out of a much larger implicit pool of $d_{\textnormal{eff}}(\mathbf{L}) \gg k$ unique items.
In this case, it is possible to consider only a small $k^2/d_{\textnormal{eff}}(\mathbf{L})$ fraction of the items
selected uniformly at random and still have enough unique items to sample a diverse $k$-subset.
Our result shows for the first time that it is possible to take advantage of this modeling disagreement between $k$ and $d_{\textnormal{eff}}(\mathbf{L})$ to gain computational
savings while still sampling \emph{exactly} from the DPP, i.e.,\xspace a computational free lunch.
\textbf{Binary search reduction from k-DPP to DPP.} Both our approach and the one of \citet{derezinski2019exact} rely on first implementing
an efficient random-size DPP sampler, followed by the usage of a black-box construction
based on rejection sampling to transform the DPP sampler into a $k$-DPP sampler.
However the reduction of \citet{derezinski2019exact} requires access to a
high-precision estimate of $d_{\textnormal{eff}}(\alpha\mathbf{L})$ in order to appropriately
tune $\alpha$. This makes optimizing $\alpha$ the bottleneck in the reduction from $k$-DPP to DPP,
and therefore there is a large computational gap between the two problems.
We close this gap thanks to a novel approach to find a suitable rescaling $\alpha$
based not on optimization but rather on binary search. Crucially, to find a suitable $\alpha$
this approach does not require an estimate of $d_{\textnormal{eff}}(\alpha\mathbf{L})$, but only $\mathcal{O}(\sqrt{k}\log(n))$ black-box calls
to a DPP sampler. Therefore, it can transform any random size DPP sampler
into a $k$-DPP sampler with only a $\sqrt{k}$ overhead,
and could be applied to any future improved sampler beyond this paper.
\section{Sampling from a rescaled DPP with intermediate uniform subsampling}
\label{s:alpha-sampler}
In this section we focus on a specific class of DPPs, $S \sim {\mathrm{DPP}}(\alpha\mathbf{L})$,
specified using a rescaling $\alpha \leq 1$
and a similarity matrix $\mathbf{L}$, which we refer to as rescaled DPPs. The main result
of the section is showing that a sufficiently large subset selected uniformly
at random can be used as an intermediate sample to sample from a
rescaled DPP without looking at all of the items.
The main reason to focus on rescaled DPPs is because they naturally appear when
reducing $k$-DPP sampling to DPP sampling, where rescaling is used to align
the random size of the DPP and $k$. This is going to be the focus of the next section.
However the approach proposed in this section
is not limited to rescaled DPPs, but under the right assumptions can be extended
to accelerate sampling from generic DPPs. We will discuss these extensions
at the end of the section.\vspace{-1mm}
\paragraph{Notation}
We use $[n]$ to denote the set $\{1,\dots,n\}$. For a matrix
$\mathbf B\in\mathbb R^{n\times m}$ and index sets $C$, $D$, we use $\mathbf B_{C,D}$
to denote the submatrix of $\mathbf B$ consisting of the intersection of rows
indexed by $C$ with columns indexed by $D$. If $C=D$, we use a
shorthand $\mathbf B_{C}$ and if $D=[m]$, we may write $\mathbf B_{C,[m]}$.
Finally, we also allow $C,D$ to be multisets or
sequences, in which case each row/column is duplicated in the matrix
according to its multiplicity (and in the case of sequences, we
order the rows/columns as they appear in the sequence).
Note that with this notation if $\mathbf{L}=\mathbf B\B^\top$ then
$\mathbf{L}_{C,D}=\mathbf B_{C,[n]}\mathbf B_{D,[n]}^\top$.
\vspace{-2mm}
\subsection{Background: distortion-free intermediate sampling.}\label{subs:background}
Rather than sampling directly from the target DPP, intermediate
sampling \cite{dpp-intermediate,leveraged-volume-sampling}
first selects an intermediate subset $\sigma$ from $[n]$, and then refines it
by extracting $S$ from $\sigma$. Crucially, if $\sigma$ is selected according
to a so-called Regularized DPP (${\mathrm{R\textnormal{-}DPP}}$), this is equivalent to sampling $S$ from a DPP.
\begin{definition}\label{d:r-dpp}
For any psd matrix $\mathbf{L} \in \mathbb R^{n\times n}$, distribution
$p\stackrel{\text{\tiny{def}}}{=} \{p_i\}_{i=1}^n$ and $r>0$, define $\widetilde{\mathbf{L}} \in \mathbb R^{n\times n}$ with~$\widetilde{\mathbf{L}}_{i,j} \stackrel{\text{\tiny{def}}}{=} \frac{\mathbf{L}_{i,j}}{r\sqrt{p_ip_j}}$. We define an ${\mathrm{R\textnormal{-}DPP}}_p^r(\mathbf{L})$
as distribution over events $A\subseteq\bigcup_{k=0}^\infty[n]^k$ such that\vspace{-2mm}
\begin{align*}
\Pr(A) \stackrel{\text{\tiny{def}}}{=}\mathbb E_{\sigma}\big[\mathbf 1_{[\sigma\in A]}\det(\mathbf I +
\widetilde{\mathbf{L}}_\sigma)\big]/{\det(\mathbf I + \mathbf{L})}\mathbin{\raisebox{0.5ex}{,}}\quad\text{for}
\quad\sigma=(\sigma_1,\dots,\sigma_t)\simiid p,\quad
t\sim\mathrm{Poisson}(r).
\end{align*}
\end{definition}
\begin{proposition}[{\citealp[Theorem 8]{dpp-intermediate}}]\label{t:composition}
For any $\mathbf{L}$, $p$, $r,$ and $\widetilde{\mathbf{L}}$ defined as in Definition
\ref{d:r-dpp},\vspace{-1mm}
\begin{align*}
\text{if}\quad \sigma\sim{\mathrm{R\textnormal{-}DPP}}_p^r(\mathbf{L})\quad\text{and}\quad
S\sim{\mathrm{DPP}}(\widetilde{\mathbf{L}}_\sigma)\quad\text{then}\quad
\{\sigma_i:i\!\in\! S\}\sim{\mathrm{DPP}}(\mathbf{L}).
\end{align*}
\end{proposition}
A computationally inefficient but conceptually simple approach to rejection sample $\sigma$ is the following:\\*
1) compute all marginals $\mathbf P(i \in S) = \ell_i(\mathbf{L}) \stackrel{\text{\tiny{def}}}{=} [\mathbf{L}(\mathbf I+\mathbf{L})^{-1}]_{i}$ and sum to $\sum_{i=1}^n\ell_i(\mathbf{L}) = d_{\textnormal{eff}}(\mathbf{L})$ \cite{ridge-leverage-scores};\\*
2) sample $t \sim \text{Poisson}(c)$ and $\sigma \sim \text{Multinomial}\left(t, \frac{\ell_1(\mathbf{L})}{d_{\textnormal{eff}}(\mathbf{L})}, \ldots, \frac{\ell_n(\mathbf{L})}{d_{\textnormal{eff}}(\mathbf{L})}\right)$ for an appropriate constant $c$;\\*
3) accept $\sigma$ w.p.~$\tfrac{\det(\mathbf I + \wt{\mathbf{L}}_{\sigma})}{C\det(\mathbf I + \mathbf{L})}$, where $C$ is an appropriate constant used to make the rejection step valid.
All existing intermediate sampling algorithms \cite{dpp-intermediate,derezinski2019exact,leveraged-volume-sampling,correcting-bias} rely on this approach,
refining it to make use of efficient approximations of the marginals
$\ell_i(\mathbf{L})$ and adapting the constants $c$ and $C$ to the data. However they all share a common bottleneck:
to sample $\sigma$ i.i.d.~they need to approximate all marginals $\ell_i(\mathbf{L})$
and the normalization constant $d_{\textnormal{eff}}(\mathbf{L})$, and therefore the final
runtime scales as $n\cdot{\mathrm{poly}}(k)$.
While this is much smaller than the $\mathcal{O}(n^3)$ required by an exact sampler,
it still becomes quickly unfeasible when $n$ is very large. In what follows
we will introduce another approach to sample from an ${\mathrm{R\textnormal{-}DPP}}$ that does not
require to approximate the marginals of all items, but only the items selected
in a preliminary \emph{uniform} intermediate sample.
\subsection{Faster DPP sampling with uniform intermediate sampling}
\begin{wrapfigure}[21]{R}{.45\textwidth}
\vspace{-2\baselineskip}
\begin{minipage}{1.02\linewidth}
\begin{algorithm}[H]
\caption{\textsc{$\alpha$-DPP}\xspace sampler}
\label{alg:alpha-dpp-sampler}
\begin{algorithmic}[1]
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\REQUIRE $\alpha$, $\mathbf{L}$, $\mathcal{D}$, $\mathbf W$, $r \geq 1$
\STATE Set $\widehat{\mathbf{L}} = \mathbf W^{1/2}\mathbf{L}_{\mathcal{D},\mathcal{D}}\mathbf W^{1/2} \in \mathbb R^{m \times m}$
\REPEAT
\STATE Sample $u \sim \text{Poisson}(re^{1/r}\alpha n\kappa^2)$
\STATE Sample $\rho = \text{Uniform}(u, [n])$
\FOR{$j = \{1, \dots, u\}$}
\STATE Compute $l_{\rho_j}$ using Eq.~(\ref{eq:rls-approx})
\STATE
Sample $z_j \sim \text{Bernoulli}(l_{\rho_j}/(\alpha\kappa^2))$\label{line:reject-binomial}\hspace*{-1cm}
\ENDFOR
\STATE Set $\sigma = \{\rho_j: z_j = 1\}$, $t = |\sigma|$
\STATE Set $[\widetilde{\mathbf{L}}_{\sigma}]_{ij} = \tfrac{1}{r\sqrt{l_{\sigma_i} l_{\sigma_j}}}[\mathbf{L}]_{\sigma_i\sigma_j}$
\STATE $\textit{Acc}\sim\!
\text{Bernoulli}\Big(\tfrac{{\mathbb E}^{\tilde{s}}\det(\mathbf I+ \alpha\widetilde{\mathbf{L}}_\sigma)}{{\mathbb E}^{t/r}\det(\mathbf I+\alpha\widehat{\mathbf{L}})}
\Big)$ \label{line:acc}
\UNTIL{$\textit{Acc}=\text{true}$\label{line:rep2}}
\STATE Sample $\widetilde{S}\sim {\mathrm{DPP}}\big(\alpha\widetilde{\mathbf{L}}_\sigma\big)$\label{line:sub}
\RETURN $S = \{\sigma_i: i\!\in\! \widetilde{S}\}$
\end{algorithmic}
\end{algorithm}
\end{minipage}
\end{wrapfigure}
We now introduce our novel $\alpha$-rescaled DPP sampler, called
\textsc{$\alpha$-DPP}\xspace (see \Cref{alg:alpha-dpp-sampler}).
It requires as input a rescaling $\alpha$, a similarity matrix $\mathbf{L}$ and a parameter
$r$ that will be used to tune the Poisson sampling step of \Cref{t:composition}
approach.
It also requires as input a \emph{dictionary} $\mathcal{D}$ containing $m$ elements,
and set of weights stored in a diagonal matrix $\mathbf W \in \mathbb R^{m \times m}$.
A dictionary is a subset of items $\mathcal{D} \subseteq [n]$
such that reweighting the items in $\mathcal{D}$ by $\mathbf W$ provides a good approximation
of $\mathbf{L}$, so that the approximate marginals
\begin{align}\label{eq:rls-approx}
l_i \stackrel{\text{\tiny{def}}}{=} \alpha[\mathbf{L} - \alpha\mathbf{L}_{[n],\mathcal{D}}^\top(\alpha\mathbf{L}_{\mathcal{D}} + \mathbf W^{-1})^{-1}\mathbf{L}_{[n],\mathcal{D}}]_{i}
\vspace{-.25\baselineskip}
\end{align}
computed using $\mathcal{D}$ and $\mathbf W$ are close to the true marginals $\ell_i$ (see \Cref{a:misc-proofs}).
Compared to the meta-approach of \Cref{t:composition}, the main technical difference
is that rather than sampling directly $t$ from an appropriate Poisson,
and then $\sigma$ from a Multinomial, we introduce an intermediate uniform sampling
step. In particular, we first sample a Poisson $u$, and then uniformly sample
a subset $\rho$ containing $u$ items. We then compute an approximation
$l_i$ of the marginal $\ell_i$ only for the items in $\rho$,
and downsample $\rho$ into $\sigma$ using rejection sampling (\Cref{line:reject-binomial}).
Finally, we accept or reject $\sigma$ (\Cref{line:acc})
and then downsample $\sigma$ into $S$ using a standard DPP sampler on
the smaller $\wt{\mathbf{L}}_{\sigma}$.
\Cref{alg:alpha-dpp-sampler} is not simply a different
implementation of the approach of \Cref{t:composition}, since even if Multinomial sampling is implemented with lazy evaluations of $l_i$,
we would still need to compute the \emph{normalization} constant of the Multinomial,
which strictly requires computing all $l_i$. Similarly, the rejection test of \Cref{line:acc} is also designed to accept as many candidates as possible without requiring the computation of the normalization constant as in \cite{derezinski2019exact}. Rather
our approach is a novel method to sample from an ${\mathrm{R\textnormal{-}DPP}}$
using Poisson rejection sampling. In particular, we prove not only that
it always returns an $S$ sampled according to the exact DPP distribution,
but also that if the dictionary satisfies certain conditions,
the main of which is $(\varepsilon,\alpha)$-accuracy (see \Cref{a:misc-proofs} and \cite{calandriello_disqueak_2017}),
then the algorithm will generate $S$ quickly.
\begin{theorem}\label{thm:alpha-sampling}
Given any $\mathbf{L} \succeq \mathbf{0}$, dictionary $\mathcal{D}$, $\mathbf W\succ \mathbf{0}$, $r \geq 1$ and $\alpha > 0$, \textsc{$\alpha$-DPP}\xspace returns
$S \sim {\mathrm{DPP}}(\alpha\mathbf{L})$. Moreover, if $r \geq d_{\textnormal{eff}}(\alpha\mathbf{L}) \geq 1/2$, $\mathcal{D}$ and $\mathbf W$ are $(1/d_{\textnormal{eff}}(\alpha\mathbf{L}), \alpha)$-accurate,
$\mathcal{D}$ satisfies $|\mathcal{D}| \leq 10d_{\textnormal{eff}}(\alpha\mathbf{L})$, and $d_{\textnormal{eff}}(\alpha\widehat{\mathbf{L}}) \leq 10d_{\textnormal{eff}}(\alpha\mathbf{L})$, w.p.~$1 - \delta$ \textsc{$\alpha$-DPP}\xspace runs in time\vspace{-1mm}
\begin{align*}
\mathcal{O}\Big(\left[ \min\{\alpha\kappa^2d_{\textnormal{eff}}(\alpha\mathbf{L}), 1\}\cdot n \cdot
d_{\textnormal{eff}}(\alpha\mathbf{L})^6\log^2(n/\delta) +
d_{\textnormal{eff}}(\alpha\mathbf{L})^9\log^3(n/\delta)\right ] \cdot \log(1/\delta)\Big).
\end{align*}
\end{theorem}
\vspace{-1mm}
The main implication of our result is that the intermediate distribution based
on marginals can be replaced more and more accurately with a uniform distribution
as $\alpha$ becomes smaller. This results in having to compute marginals
only for a $\min\{\alpha\kappa^2d_{\textnormal{eff}}(\alpha\mathbf{L}), 1\}$ fraction of the $n$ items.
This speedup can be significant when the rescaling $\alpha$ is very small,
as is the case when we want to sample a small number of items out of a large
collection.
Compared to other exact DPP samplers, such as \textsc{DPP-VFX}\xspace,
our \textsc{$\alpha$-DPP}\xspace is strictly faster by roughly a $1/(\alpha\kappa^2d_{\textnormal{eff}}(\alpha\mathbf{L}))$ factor
when implemented with an appropriate caching strategy for the
estimates $l_i$ (see \Cref{a:misc-proofs}).
Further, unlike MCMC samplers, \textsc{$\alpha$-DPP}\xspace is an \emph{exact} sampler. Moreover, there is no
known MCMC approach that can achieve a runtime sub-linear in $n$ when $\alpha$
is small as \textsc{$\alpha$-DPP}\xspace.
An $(\varepsilon, \alpha)$-accurate dictionary
that also satisfies the other conditions can be generated using
a slight modification of the \textsc{BLESS}\xspace algorithm \cite{NIPS2018_7810}, that we call \textsc{BLESS-I}\xspace algorithm, presented
in \Cref{a:poisson-bless}.
However, note that since the marginals $\ell_i$ are equivalent to
the ridge leverage scores \cite{ridge-leverage-scores} of item $i$,
we can replace \textsc{BLESS-I}\xspace with any present or future algorithm for leverage score sampling that can
be modified to be rescaling-aware \cite{calandriello_disqueak_2017,NIPS2018_7810}.
Moreover, note that \textsc{BLESS-I}\xspace also returns an estimate
of $d_{\textnormal{eff}}(\alpha\mathbf{L})$ that is sufficiently accurate to tune $r$ and $\varepsilon$.
At the same time, our analysis could be excessively conservative, and instead of trying
to set $r$ and $\varepsilon$ using $d_{\textnormal{eff}}(\alpha\mathbf{L})$ as suggested by \Cref{thm:alpha-sampling},
a more practical strategy is to start with a constant
$r$ and increase it slowly if the sampler is rejecting with a too low probability,
using a doubling schedule to preserve overall time complexity.
\textbf{Proof sketch.} The proof is divided in two parts, proving that \textsc{$\alpha$-DPP}\xspace is an exact sampler
(\Cref{lem:a-dpp-exact}) and that under the right conditions it is efficient (\Cref{lem:a-dpp-efficient}).
For the first part we once again rely on the approach of \Cref{t:composition},
but with the added difficulty of not being allowed to compute all the marginals.
To avoid this bottleneck, we show that:\vspace{-1mm}
\begin{enumerate}
\item[A)] sampling $t \sim \text{Poisson}(r)$ and $\sigma \sim
\text{Multinomial}(t, \{\ell_i/d_{\textnormal{eff}}(\alpha\mathbf{L})\}_{i=1}^n)$;\quad and
\item[B)] sampling $n$ independent $s_i \sim \text{Poisson}(r'\ell_i(\alpha\mathbf{L}))$, and adding $s_i$ copies of item $i$ to $\sigma$,
\end{enumerate}\vspace{-1mm}
are equivalent for an appropriate choice of $r$ and $r'$, i.e.,\xspace
we prove that the $\sigma$ generated by both approach A and B follow the same distribution.
However, unlike approach A, approach B does not require computing
a normalization constant, i.e.,\xspace it samples from \emph{unnormalized} probabilities.
Moreover, if we know an upper bound on the marginals we can further reduce the number of marginals that need
to be computed. In our case we use
the bound $\ell_i \leq \alpha\kappa^2$, and show that
\vspace{-1mm}
\begin{enumerate}
\item[C)] sampling $n$ Poisson independently $u_i \sim
\text{Poisson}(r'\alpha\kappa^2)$, only if $u_i > 0$ computing
$\ell_i$ and sampling $s_i \sim \text{Binomial}(u_i,
\ell_i/(\alpha\kappa^2))$, and adding $s_i$ copies of item $i$ to
$\sigma$
\end{enumerate}\vspace{-1mm}
once again generates $\sigma$ strictly equivalent to the ones of approach B and A.
The added advantage of approach C over the others is that only the marginals of items with $u_i > 0$
are actually computed, and there is no need to compute a normalization constant.
Starting from this new approach, to obtain our \textsc{$\alpha$-DPP}\xspace sampler (\Cref{alg:alpha-dpp-sampler})
we simply replace the $n$ Poisson $u_i \sim \text{Poisson}(r'\alpha\kappa^2)$ with a single
$u \sim \text{Poisson}(r'\alpha\kappa^2 n)$ followed by uniform sampling,
and replace the exact $\ell_i$ with approximate $l_i$.
For the second part we derive a lower bound on the acceptance probability
similar to the one from \citet{derezinski2019exact}. However, while they
use an $n \times d_{\textnormal{eff}}(\alpha\mathbf{L})$ Nystr\"om\xspace approximation of the matrix $\mathbf{L}$,
to avoid direct dependencies on $n$ we are forced to use a less stable
approximation $\widehat{\mathbf{L}}$. As a result, controlling $d_{\textnormal{eff}}(\alpha\widehat{\mathbf{L}})$
requires a more careful analysis.
\textbf{Beyond uniform subsampling.} One of the implications of our analysis
is that more adaptive upper bounds on the marginals $\ell_i$ could further
speedup our \textsc{$\alpha$-DPP}\xspace sampling approach.
In particular, we chose uniform sampling, i.e.,\xspace a uniform upper bound,
for its conceptual simplicity and because knowing an upper bound $\kappa^2$
on the entries of the similarity matrix usually does not require looking at
the items, e.g.,\xspace $\kappa^2$ is always equal to 1 for Gaussian similarity, Cosine similarity or other
self normalized similarities. However for other similarities, such as linear
similarity, this bound could be very loose. A simple replacement is using
the actual diagonal of $\mathbf{L}$, which requires to look at all items and $\mathcal{O}(n)$
time to compute but is usually very scalable. Ideally, one could imagine
designing a sequence of upper bounds starting from cheaper to more computationally
expensive, where more advanced techniques such as random projection are used
near the end to further filter candidate items.
\section{Efficient reduction from k-DPP to rescaled DPP via binary search}
\label{s:binary-search}
Given our fast DPP sampler, we can see a $k$-DPP as a sampling process where we first sample
$S \sim {\mathrm{DPP}}(\alpha\mathbf{L})$, check if the sample size $|S|$ is equal to $k$, and then
accept or reject the sample accordingly. Rescaling
$\mathbf{L}$ by a constant factor $\alpha$ only changes the expected size
$d_{\textnormal{eff}}(\alpha\mathbf{L})$ (and not the $k$-DPP), with
$\alpha > 1$ increasing the expected size and $\alpha < 1$ decreasing it.
Thus, it is natural to imagine that there exists some $\alpha_\star$ for which
the acceptance probability is high. Indeed this was recently proven to be possible.
\citet{derezinski2019exact} show that if the \emph{mode}
$m_{\alpha_{\star}}$ of $S_{\alpha_\star}\sim{\mathrm{DPP}}(\alpha_\star \mathbf{L})$ is equal to
$k$, then we will accept with probability at least $\Omega(1/\sqrt{k})$.
They also provide an
algorithm to find such an $\alpha_{\star}$. However,
this algorithm has a prohibitively high computational cost,
$\acO(nk^{10} + k^{15})$, because ensuring that the mode of
${\mathrm{DPP}}(\alpha_\star\mathbf{L})$ is \emph{exactly} $k$ requires an extremely
accurate approximation of $\mathbf{L}$. Instead, our approach is to run a binary search
to find a good rescaling $\alpha$, which will terminate once the
acceptance probability is high enough, regardless of whether $k$ is
exactly the mode. Crucially, this binary search only
requires a black box ${\mathrm{DPP}}(\alpha\mathbf{L})$ sampler (such as
our \textsc{$\alpha$-DPP}\xspace), and it only queries the sampler $\acO(\sqrt k)$ many
times. To prove that the binary search finds a good $\alpha$ in a
small number of steps, we establish a new property (Lemma~\ref{l:pb-new}) of the Poisson
Binomial distribution (the distribution of the subset sizes of
${\mathrm{DPP}}(\alpha\mathbf{L})$), which should be of independent interest.
\subsection{Binary search}
Our main result in this subsection is \Cref{alg:binary-search}, which requires only oracle
access to the samples from a random-size DPP, and finds a rescaling
$\hat\alpha$ which enables efficient rejection sampling from a
$k$-DPP. Note that the provided oracle sampler does not have to be
our \textsc{$\alpha$-DPP}\xspace sampler, so the algorithm could be paired with other samplers.
\begin{lemma}[{restate=[name=restated]binarysearch}]\label{l:binary-search}
Suppose that we are given an integer $k$, a range
$I=[\alpha_{\min},\alpha_{\max}]$ where
$\alpha_{\max}=\gamma\alpha_{\min}$, and access
to an oracle which, for any
$\alpha\in I$, returns $S\sim
{\mathrm{DPP}}(\alpha \mathbf{L})$. If there exists $\alpha_\star\in I$ such that $k$
is the mode of $|S|$ for $S\sim{\mathrm{DPP}}(\alpha_\star \mathbf{L})$, then using
$O\big(\sqrt k\log^2(k\log(\gamma)/\delta)\big)$
calls to the oracle we can find $\hat\alpha\in I$ such that with
probability $1-\delta$ we have
\begin{align*}
\Pr(|S| = k) = \Omega\big(\tfrac1{\sqrt k}\big),\qquad\text{for}\quad
S\sim{\mathrm{DPP}}(\hat\alpha \mathbf{L}).
\end{align*}
\end{lemma}\vspace{-2mm}
The distribution of subset size $|S|$ for $S\sim{\mathrm{DPP}}(\mathbf{L})$
can be defined via the eigenvalues $\lambda_1\geq\lambda_2\geq ...$ of
$\mathbf{L}$ (see \cite{dpp-independence}): if we let $b_i\sim
\mathrm{Bernoulli}(\frac{\lambda_i}{\lambda_i+1})$ for $i\geq 1$, then
$\sum_i b_i$ is distributed identically to $|S|$. This distribution is known as the \emph{Poisson
Binomial}, and it has been extensively studied in the probability
literature \cite{pb-review}. The recent result of
\cite{derezinski2019exact} on the probability of the mode of a Poisson
Binomial shows that it is possible to find $\hat\alpha$ satisfying the
condition of Lemma \ref{l:binary-search}.
\begin{lemma}[{restate=[name=restated]pbmode}]\label{l:pb-mode}
There is an absolute constant $0<c<1$ such that for any Poisson
Binomial distribution $p:\mathbb Z_{\geq 0}\rightarrow\mathbb R_{\geq 0}$,
with mode $k^*$ we have $p(k^*)\geq \frac{c}{\sqrt{k^*\!+1}}\cdot$
\end{lemma}\vspace{-2mm}
This result, however, does not provide an efficient way of
finding an $\hat\alpha$ such that the mode of the subset size distribution of
${\mathrm{DPP}}(\hat\alpha\mathbf{L})$ is $k$. We circumvent this problem by performing
a binary search (\Cref{alg:binary-search}) that looks for such an
$\hat\alpha$, but stops early when it reaches a
sufficiently good candidate, avoiding excess computations. To make this rigorous, we establish the
following new property of the Poisson Binomial distribution, which
should be of independent interest.
\begin{lemma}[{restate=[name=restated]pbnew}]\label{l:pb-new}
Let $p:\mathbb Z_{\geq 0}\rightarrow\mathbb R_{\geq 0}$ be a Poisson Binomial distribution,
and let $k\geq 1$ satisfy
$p(k)<\frac{c}{12\sqrt{3(k+1)}}\mathbin{\raisebox{0.5ex}{,}}$ where $c$ comes from Lemma
\ref{l:pb-mode}. Then, $P_{<k}=\sum_{i<k}p(k)$ and
$P_{>k}=\sum_{i>k}p(k)$ satisfy:
\begin{enumerate}[leftmargin=1cm]
\vspace{-3mm}
\item if the mode of $p$ is less than $k$, then
$P_{>k}\leq \frac12 - \frac c{12}$; \vspace{-2mm}
\item if the mode of $p$ is greater than $k$, then $P_{<k}\leq \frac12
- \frac{c}{12}\cdot$
\end{enumerate}
\end{lemma}\vspace{-1mm}
Informally, the above result states the following:
For any $k$, either its probability under the given Poisson Binomial
is at least $\Omega(\frac1{\sqrt k})$, or this $k$ splits the probability mass into two
uneven parts, with the larger one containing the mode. Thus, as long
as our candidate $\alpha$ does not yield high acceptance probability
for $k$, it is easy to make the branching decision in the binary search by estimating the
quantities $P_{>k}$ and $P_{<k}$ simply by repeated sampling from
${\mathrm{DPP}}(\alpha\mathbf{L})$. Note that if the condition on $p(k)$ is not
satisfied, then performing the branching decision could be very
expensive, but our algorithm avoids this possibility. The proof of
\Cref{l:binary-search} (\Cref{a:bin-search}) follows from Lemmas
\ref{l:pb-mode} and \ref{l:pb-new}.
\setlength{\textfloatsep}{5pt}
\begin{algorithm}[t]
\caption{Binary search for initializing the $k$-${\mathrm{DPP}}(\mathbf{L})$ sampler}
\label{alg:binary-search}
{
\setstretch{1.4}
\begin{algorithmic}[1]
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\REQUIRE $0<\alpha_{\min}<\alpha_{\max}$, sampling oracle for
${\mathrm{DPP}}(\alpha \mathbf{L})$, integer $k$ and constants $C>0, \delta\in(0,1)$
\ENSURE $\hat\alpha$ such that ${\mathrm{DPP}}(\hat\alpha \mathbf{L})$ can be used
to efficiently sample $k$-${\mathrm{DPP}}(\mathbf{L})$
\FOR{$s = \{1, \ldots, \lceil\log(\gamma)\rceil\}$}
\STATE \textbf{if} {$\alpha_{\max}/\alpha_{\min} < (1+\frac1{(k+3)^2})$\label{line:if1}} \textbf{then return} $\hat\alpha = \alpha_{\min}$
\STATE $\bar\alpha \leftarrow\sqrt{\alpha_{\min}\alpha_{\max}}$
\STATE Sample $S_1,...,S_t\simiid {\mathrm{DPP}}(\bar\alpha \mathbf{L})$\quad
where\quad $t = C\sqrt k\log(s/\delta)$
\STATE $\hat P_k \leftarrow \frac1t \sum_{i=1}^t\mathbf 1_{[|S_i|=k]}$
\STATE \textbf{if} {$\hat P_k\geq \frac12\cdot \frac{c}{12\sqrt{3(k+1)}}$\label{line:if2}} \textbf{then return}
$\hat\alpha = \bar\alpha$
\STATE $(\hat P_{<k}, \hat P_{>k}) \leftarrow \big(\frac1t
\sum_{i=1}^t\mathbf 1_{[|S_i|<k]},\ \frac1t
\sum_{i=1}^t\mathbf 1_{[|S_i|>k]}\big)$
\vspace{.15cm}
\STATE \textbf{if} $\hat P_{<k}> \hat P_{>k}$ \textbf{then} $(\alpha_{\min},\alpha_{\max}) = (\bar\alpha,\alpha_{\max})$ \textbf{else} $(\alpha_{\min},\alpha_{\max}) = (\alpha_{\min},\bar\alpha)$
\label{line:branch}
\ENDFOR
\end{algorithmic}
}
\end{algorithm}
\subsection{Constructing the initial interval}
To initiate our binary search, we must first find a range of values
$[\alpha_{\min},\alpha_{\max}]$, which contains the desired $\alpha_\star$, and also
construct a sampling oracle for ${\mathrm{DPP}}(\alpha\mathbf{L})$.
The binary search procedure is deliberately presented in a way that is agnostic to
how these two steps are accomplished, because a number of existing DPP
samplers could be adapted to take advantage of
\Cref{alg:binary-search}, including
\cite{dpp-noeig,poulson2019high-performance,dpp-intermediate,derezinski2019exact}.
Our implementation of these two steps is different than these previous
approaches in that it takes advantage of the structure of the kernel
so that it only has to look at a potentially small fraction of the
data points. We achieve this with a modified version
of the BLESS algorithm \cite{NIPS2018_7810}.
\begin{lemma}[{restate=[name=restated]algovalidinterval}]\label{lem:algo-valid-interval}
W.p.~$1-\delta$ \textsc{BLESS-I}\xspace runs in time
$\acO\left(\min\{\alpha_{\max}\kappa^2, 1\}nk^6 + k^9\right)$
and satisfies:
\vspace{-.5\baselineskip}
{
\setlength{\itemsep}{-2pt}
\begin{enumerate}[leftmargin=1cm]
\item The interval $[\alpha_{\min},\alpha_{\max}]$ is bounded by $\frac14(k-1)/\mathrm{tr}(\mathbf{L}) \leq \alpha_{\min} \leq \alpha_{\max} \leq 8(k + 2)/d_{\textnormal{eff}}(\mathbf{L})$
\item There is $\alpha_\star\in[\alpha_{\min},\alpha_{\max}]$
for which $k$ is the mode of $|S|$ where
$S\sim{\mathrm{DPP}}(\alpha_\star\mathbf{L})$;
\item The dictionary $\mathcal{D}^{\alpha_{\max}}$ satisfies the
conditions from \Cref{thm:alpha-sampling} for any $\alpha\in[\alpha_{\min},\alpha_{\max}]$.
\end{enumerate}
}
\end{lemma}
The first two parts of the lemma ensure that the interval
$I=[\alpha_{\min},\alpha_{\max}]$ is a valid input for
the binary search in \Cref{alg:binary-search} and that its size
$\gamma = \alpha_{\max}/\alpha_{\min} \leq 4\mathrm{tr}(\mathbf{L})/d_{\textnormal{eff}}(\mathbf{L})$
is bounded in the log-scale. The last part implies that
$\alpha$-DPP can be used by that algorithm as the oracle sampler.
At a high level, \Cref{alg:adjusted-bless} proceeds by starting with a
small $\alpha^0$ that is guaranteed to be a valid lower bound for the interval,
and for which a dictionary $\mathcal{D}^0$ can be constructed simply via
uniform sampling. Then we repeatedly double the $\alpha$ and refine
the dictionary, until we reach $\alpha^i$ such that we can ensure that
with high probability $d_{\textnormal{eff}}(\alpha^i\mathbf{L}) \geq k + 1$ which makes it
a valid upper bound for the interval (then, this $\alpha^i$
becomes $\alpha_{\max}$).
\subsection{Overall time complexity of $k$-DPP sampling}
Putting together all the results from the previous sections, we can
finally bound the computational complexity of our $k$-DPP sampler,
which first uses \textsc{BLESS-I}\xspace (\Cref{alg:adjusted-bless}) to construct a dictionary and search interval,
and then applies the binary search of (\Cref{alg:binary-search}) using
our $\alpha$-DPP sampler (\Cref{alg:alpha-dpp-sampler}) as the sampling oracle.
Once again note that in the following computational analysis we will use conservative
values for many parameters, notably $r$ from \textsc{$\alpha$-DPP}\xspace and $q$ from \textsc{BLESS-I}\xspace,
as they are suggested from the theory. However in practice it is always better to
start from a more optimistic value, and keep doubling them only if the sampler
repeatedly fails to accept. Importantly, samples generated this way will still
be exactly distributed according to the DPP, as all the approximations used
in our approach only influence the runtime of our algorithm, and not the correctness of
its acceptance, which always holds.
By \Cref{lem:algo-valid-interval}, the preprocessing step of running
\textsc{BLESS-I}\xspace takes $\acO\left(\min\{\alpha_{\max}\kappa^2,1\}nk^6 + k^9\right)$
and generates a dictionary $\mathcal{D}$ with size $\acO(k^3)$.
Since $d_{\textnormal{eff}}(\alpha\mathbf{L}) \leq d_{\textnormal{eff}}(\alpha_{\max}\mathbf{L}) \leq \mathcal{O}(k)$ for all
$\alpha$ in the search interval, each call to the \textsc{$\alpha$-DPP}\xspace sampler also requires
at most $\acO\left(\min\{\alpha_{\max}\kappa^2k, 1\}nk^6 + k^9\right)$.
Finally, the binary search invokes $\alpha$-DPP at most $\acO(\sqrt{k})$ times
so the overall runtime is
$\acO\big((\min\{\alpha_{\max}\kappa^2k, 1\}nk^{6} +
k^{9})\cdot\sqrt{k}\big)$.
We now provide a bound on $\alpha_{\max}$.
\begin{lemma}[{restate=[name=restated]lambdalowerbound}]\label{lem:lambda-lower-bound}
For any matrix $\mathbf{L}$ and $0 < \alpha \leq 1$, we have
$d_{\textnormal{eff}}(\alpha\mathbf{L})/d_{\textnormal{eff}}(\mathbf{L}) \geq \alpha \geq d_{\textnormal{eff}}(\alpha\mathbf{L})/\mathrm{tr}(\mathbf{L})$.
\end{lemma}
Applied to $\alpha_{\max}$, we obtain $\alpha_{\max} \leq d_{\textnormal{eff}}(\alpha_{\max}\mathbf{L})/d_{\textnormal{eff}}(\mathbf{L}) \leq
\mathcal{O}(k/d_{\textnormal{eff}}(\mathbf{L}))$, giving us the final runtime of
$\acO\big((\min\{k^2\kappa^2/d_{\textnormal{eff}}(\mathbf{L}), 1\}nk^{6} + k^{9})\cdot\sqrt{k}\big)$
reported in \Cref{thm:informal-main}.
\noindent\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=.95\textwidth]{fig/fig_runtime_repro_noresample_first_k_dpp.pdf}
\vspace{-3mm}
\captionof{figure}{Small scale experiment}\label{fig:small-scale}
\vspace{-2mm}
\end{minipage}\hfill
\begin{minipage}{0.49\textwidth}
\centering
\includegraphics[width=.95\textwidth]{fig/fig_runtime_repro_k_succ_rbf_time.pdf}
\vspace{-3mm}
\captionof{figure}{Large scale experiment.}\label{fig:large-scale}
\vspace{-2mm}
\end{minipage}
\section{Experiments}
\vspace{-.75\baselineskip}
In this section, we evaluate our \textsc{$\alpha$-DPP}\xspace sampler on a
benchmark\footnote{\url{https://github.com/LCSL/dpp-vfx}}
introduced by \cite{derezinski2019exact} (see
\Cref{a:experiments}). The benchmark uses
subsets of the infinite MNIST dataset \cite{loosli-canu-bottou-2006}
with $d = 784$ and $n$ varying up to $10^6$.
All experiments are executed on a 28 core Xeon E5-2680 v4. Each experiment is repeated multiple times, and we report mean
values and a 95\% confidence interval.
\textbf{Baselines:} we compare \textsc{$\alpha$-DPP}\xspace with \textsc{DPP-VFX}\xspace \cite{derezinski2019exact}, an MCMC sampler \cite{rayleigh-mcmc}
and a sampler based on eigendecompositions \cite{dpp-independence, gillenwater2014approximate}.
All algorithms are implemented in \texttt{python} as part
\ifisarxiv
\else
\footnote{We provide our modified version of the library which includes \textsc{$\alpha$-DPP}\xspace in the supplementary material.}
\fi
of \texttt{DPPy} \cite{gautier2018dppy}.
Due to their similar input, we use the same oversampling parameters (see \Cref{a:experiments}) for \textsc{$\alpha$-DPP}\xspace and \textsc{DPP-VFX}\xspace.
We run the MCMC sampler for $\mathcal{O}(nk)$ iterations to guarantee mixing \cite{rayleigh-mcmc}. For more details
on hyperparameter tuning we refer to \Cref{a:experiments}.
\textbf{Results} \ We begin by reporting results on a smaller subset of data (\Cref{fig:small-scale})
where even the non-efficient samplers can be run. We use an \texttt{rbf}
similarity with $\sigma = \sqrt{3d}$, and set $k=10$ to match the number of
digit classes in MNIST. Note that for $n = 70000$ \textsc{BLESS-I}\xspace estimates
$d_{\textnormal{eff}}(\mathbf{L}) \approx 300$, validating our assumption of $d_{\textnormal{eff}}(\mathbf{L}) \gg k$.
Thanks to this mismatch, we can see how \textsc{$\alpha$-DPP}\xspace maintains a \emph{constant}
runtime as $n$ grows, and increasingly matches or outpaces competing baselines as $n$ grows.
In particular, it becomes faster than the eigendecomposition based sampler (which cannot scale beyond $n = 24000$)
or the MCMC sampler. However, the gap is still sufficiently small that
\textsc{DPP-VFX}\xspace, the previously fastest $k$-DPP sampler
available, remains competitive.
For larger datasets we consider only the scalable samplers, \textsc{$\alpha$-DPP}\xspace and
\textsc{DPP-VFX}\xspace. We consider again an \texttt{rbf} similarity, but this time we choose $n$ up
to $10^6$ and $\sigma = \sqrt{10}$. This further increases the gap between $k$ and $d_{\textnormal{eff}}(\mathbf{L})$,
with \textsc{BLESS-I}\xspace estimating $d_{\textnormal{eff}}(\mathbf{L}) \approx 1000$.
We report results in \Cref{fig:large-scale}, with runtime shown in
log-scale. In this regime, the gap between \textsc{DPP-VFX}\xspace and
\textsc{$\alpha$-DPP}\xspace widens, as \textsc{DPP-VFX}\xspace
cannot use rescaling to reduce the final dictionary size from $d_{\textnormal{eff}}(\mathbf{L})$ to
$d_{\textnormal{eff}}(\hat\alpha\mathbf{L}) \approx k$, and has to compute $n$ marginal probabilities
since it does not leverage uniform intermediate subsampling.
In particular, thanks to the uniform sampling step, we see that \textsc{$\alpha$-DPP}\xspace's
runtime does not grow as $n$ grows, since all the expensive computations
are performed in the small intermediate subset which is hardly
sensitive to $n$. We note that, due to using a smaller dictionary,
\textsc{$\alpha$-DPP}\xspace requires about 2-5x more trials in the rejection sampling
step, which leads to larger variance in the runtime.
\begin{wrapfigure}[12]{R}{0.43\textwidth}
\begin{minipage}{\linewidth}
\vspace{-5mm}
\centering
\includegraphics[width=.95\textwidth]{fig/fig_runtime_repro_k_succ_rbf_fraction.pdf}
\vspace{-2mm}
\captionsetup{font=small}
\captionof{figure}{Fraction of items observed by \textsc{$\alpha$-DPP}\xspace.}\label{fig:rejections}
\vspace{-4mm}
\end{minipage}
\end{wrapfigure}
In \Cref{fig:rejections}, we report the fraction of data that
is observed by \textsc{$\alpha$-DPP}\xspace in the large scale experiment. This
quantity, denoted as $\beta$ in \Cref{thm:informal-main}, is
responsible for much of the computational gains of the algorithm over
\textsc{DPP-VFX}\xspace, reported in \Cref{fig:large-scale}. Note that the remaining
$1-\beta$ portion of the data does not ever need to be loaded into
the program's memory, which leads to a significant reduction in
memory accesses. We observe that as the
data size increases, the fraction of items observed by \textsc{$\alpha$-DPP}\xspace goes down to as
little as $1$\% for $n=10^6$, which is why the
runtime of \textsc{$\alpha$-DPP}\xspace stays roughly flat, whereas the runtime of
$\textsc{DPP-VFX}\xspace$ grows.
\newpage
\section*{Broader impact}
DPPs were discovered in the 70s by Odile Macchi to model repulsion of
particle distributions in fermions, so improvements in samplers may help
in modelling physical simulations. In bringing faster DPP samplers to
machine learning we aim to enable a
better handling of diversity through this rigorous theoretical framework.
\begin{ack}
MD thanks the NSF for funding via the NSF TRIPODS program.
\end{ack}
\bibliographystyle{plainnat}
|
\section{Introduction}
Born from World War II and the haunting and despicable practices of Nazi era experimentation \cite{Nazianat71:online} the \textit{1947 Nuremberg code}~\cite{weindling2001origins_informed_consent} and the subsequent \textit{1964 Helsinki declaration}~\cite{Timetodi89:online_Nature}, helped establish the doctrine of \textbf{Informed Consent} which builds on the fundamental notions of human dignity and agency to control dissemination of information about oneself. This has shepherded data collection endeavors in the medical and psychological sciences concerning human subjects, including photographic data \cite{naidoo2009informed, blain2002informed}, for the past several decades. A less stringent version of informed consent, \textit{broad consent}, proposed in 45 CFR 46.116(d) of the \textit{Revised Common Rule}~\cite{eCFR_Cod99_online_common_rule}, has been recently introduced that still affords the basic safeguards towards protecting one's identity in large scale databases. However, in the age of \textit{Big Data}, the fundamentals of informed consent, privacy, or agency of the individual have gradually been eroded. Institutions, academia, and industry alike, amass millions of images of people without consent and often for unstated purposes under the guise of anonymization. These claims are misleading given there is weak anonymity and privacy in aggregate data in general \cite{narayanan2008robust_de-anonymization} and more crucially, images of faces are not the type of data that can be aggregated. As can be seen in Table~\ref{tab:datasets}, several tens of millions of images of people are found in peer-reviewed literature. These images are obtained without consent or awareness of the individuals or IRB approval for collection. In \textit{Section 5-B} of~\cite{torralba200880_tiny}, for instance, the authors nonchalantly state ``\textit{As many images on the web contain pictures of people, a large fraction (23\%) of the 79 million images in our dataset have people in them}''. With this background, we now focus on one of the most celebrated and canonical large scale image datasets: the \textit{ImageNet} dataset. From the questionable ways images were sourced, to troublesome labeling of people in images, to the downstream effects of training AI models using such images, ImageNet and large scale vision datasets (LSVD) in general \textit{constitute a Pyrrhic win} for computer vision. We argue, this win has come at the expense of harm to \textit{minoritized groups} and further aided the gradual erosion of privacy, consent, and agency of both the individual and the collective.
\subsection{ImageNet: A brief overview}
\begin{table*}[ht]
\centering
\begin{threeparttable}
\caption{Large scale image datasets containing people's images}
\begin{tabular}{|l|l|l|l|}
\hline
Dataset &
\begin{tabular}[c]{@{}l@{}}Number of images\\ (in millions)\end{tabular} &
\begin{tabular}[c]{@{}l@{}}Number of \\ categories\\ (in thousands)\end{tabular} &
\begin{tabular}[c]{@{}l@{}}Number of \\ consensual\\ images\end{tabular} \\ \hline
JFT-300M (\cite{hinton2015distilling_jft}) & 300+ & 18 & 0 \\ \hline
Open Images (\cite{kuznetsova2018open}) & 9 & 20 & 0 \\ \hline
Tiny-Images (\cite{torralba200880_tiny}) & 79 & 76 & 0 \\ \hline
Tencent-ML (\cite{wu2019tencent}) & 18 & 11 & 0 \\ \hline
ImageNet-(21k,11k\tnote{1},1k) (\cite{russakovsky2015imagenet}) & (14, 12, 1) & (22, 11, 1) & 0 \\ \hline
Places (\cite{zhou2017places}) & 11 & 0.4 & 0 \\ \hline
\end{tabular}%
\begin{tablenotes}
\item [1] See \url{https://modelzoo.co/model/resnet-mxnet}
\end{tablenotes}
\label{tab:datasets}
\end{threeparttable}
\end{table*}
The emergence of the ImageNet dataset \cite{deng2009imagenet} is widely considered a pivotal moment\footnote{``The data that transformed AI research—and possibly the world'': \url{https://bit.ly/2VRxx3L}} in the Deep Learning revolution that transformed Computer Vision (CV), and Artificial Intelligence (AI) in general.
Prior to ImageNet, computer vision and image processing researchers trained image classification models on small datasets such as CalTech101 (9k images), PASCAL-VOC (30k images), LabelMe (37k images), and the SUN (131k images) dataset (see slide-37 in \cite{imagenet30:online}). ImageNet, with over 14 million images spread across 21,841 synsets, replete with 1,034,908 bounding box annotations, brought in an aspect of scale that was previously missing. A subset of 1.2 million images across 1000 classes was carved out from this dataset to form the ImageNet-1k dataset (popularly called \texttt{ILSVRC-2012}) which formed the basis for the \textit{Task-1: classification} challenge in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC). This soon became widely touted as the \textit{Computer Vision Olympics}\footnote{ https://engineering.missouri.edu/2014/01/team-takes-top-rankings-in-computer-vision-olympics/}. The vastness of this dataset allowed a Convolutional Neural Network (CNN) with 60 million parameters \cite{krizhevsky2012imagenet} trained by the \textit{SuperVision} team from University of Toronto to usher in the rebirth of the CNN-era (see \cite{alom2018history_alexnet}), which is now widely dubbed the \textit{AlexNet moment} in AI.
Although ImageNet was created over a decade ago, it remains one of the most influential and powerful image databases available today. Its power and magnitude is matched by its unprecedented societal impact. Although an \textit{a posteriori} audit might seem redundant a decade after its creation, ImageNet's continued significance and the culture it has fostered for other large scale datasets warrants an ongoing critical dialogue.
The rest of this paper is structured as follows. In section 2, we cover related work that has explored the ethical dimensions that arise with LSVD. In section 3, we describe the landscape of both the immediate and long term threats individuals and society as a whole encounter due to ill-considered LSVD curation. In Section 4, we propose a set of solutions which might assuage some of the concerns raised in section 3. In Section 5, we present a template quantitative auditing procedure using the ILSVRC2012 dataset as an example and describe the data assets we have curated for the computer vision community to build on. We conclude with broad reflections on LSVDs, society, ethics, and justice.
\section{Background and related work}
The very declaration of a taxonomy brings some things into existence while rendering others invisible \cite{bowker2000sorting}. A gender classification system that conforms to essentialist binaries, for example, operationalizes gender in a cis-centric way resulting in exclusion of non-binary and transgender people \cite{keyes2018misgendering}. Categories simplify and freeze nuanced and complex narratives, obscuring political and moral reasoning behind a category. Over time, messy and contingent histories hidden behind a category are forgotten and trivialized \cite{star2007enacting}. With the adoption of taxonomy sources, image datasets inherit seemingly invisible yet profoundly consequential shortcomings. The dataset creation process, its implication for ML systems, and subsequently, the societal impact of these systems has attracted a substantial body of critique. We categorize such body of work into two groups that complement one another. While the first group can be seen as concerned with the broad downstream effects, the other concentrates mainly on the dataset creation process itself.
\subsection{Broad critiques}
The absence of critical engagement with canonical datasets disproportionately negatively impacts women, racial and ethnic minorities, and vulnerable individuals and communities at the margins of society~\cite{birhane2019algorithmic}. For example, image search results both exaggerate stereotypes and systematically under-represent women in search results for occupations~\cite{kay2015unequal}; object detection systems designed to detect pedestrians display higher error rates for recognition of demographic groups with dark skin tones~\cite{wilson2019predictive}; and gender classification systems show disparities in image classification accuracy where lighter-skin males are classified with the highest accuracy while darker-skin females suffer the most misclassification~\cite{buolamwini2018gender}. Gender classification systems that lean on binary and cis-genderist constructs operationalize gender in a trans-exclusive way resulting in tangible harm to trans people~\citep{keyes2018misgendering, scheuerman2019computers}. With a persistent trend where minoritized and vulnerable individuals and communities often disproportionately suffer the negative outcomes of ML systems, D'Ignazio and Klein \cite{d2020data} have called for a shift in rethinking ethics not just as a fairness metric to mitigate the narrow concept of bias but as a practice that results in justice for the most negatively impacted. Similarly, Kasy and Abebe \cite{kasy2020fairness} contend that perspectives that acknowledge existing inequality and aim to redistribute power are pertinent as opposed to fairness-based perspectives. Such understanding of \textit{ethics as justice} then requires a focus beyond `bias' and fairnesss' in LSVDs and requires questioning of how images are sourced, labelled, and what it means for models to be trained on them.
One of the most thorough investigations in this regard can be found in \cite{Excavati19:online}. In this recent work, Crawford and Paglen present an in-depth critical examination of ImageNet including the dark and troubling results of classifying people as if they are objects. Offensive and derogatory labels that perpetuate historical and current prejudices are assigned to people’s actual images. The authors emphasise that not only are images that were scraped across the web appropriated as data for computer vision tasks, but also the very act of assigning labels to people based on physical features raises fundamental concerns around reviving long-discredited pseudo-scientific ideologies of physiognomy~\cite{y2017physiognomy}.
\subsection{Critiques of the curation phase}
Within the dataset creation process, \textit{taxonomy sources} pass on their limitations and underlying assumptions that are problematic. The adoption of underlying structures presents a challenge where --- without critical examination of the architecture --- ethically dubious taxonomies are inherited. This has been one of the main challenges for ImageNet given that the dataset is built on the backbone of WordNet's structure. Acknowledging some of the problems, the authors from the ImageNet team did recently attempt to address \cite{yang2020towards_facct} the stagnant concept vocabulary of WordNet. They admitted that only 158 out of the 2,832 existing synsets should remain in the person sub-tree\footnote{In order to prune all the nodes. They also took into account the \textit{imageability} of the synsets and the skewed representation in the images pertaining to the \textit{Image retrieval} phase}. Nonetheless, some serious problems remain untouched. This motivates us to address in greater depth the overbearing presence of the \textit{WordNet effect} on image datasets.
\subsection{The WordNet Effect}
ImageNet is not the only large scale vision dataset that has inherited the shortcomings of the WordNet taxonomy. The \textit{80 million Tiny Images} dataset \cite{torralba200880_tiny} which grandfathered the CIFAR-10/100 datasets and the Tencent ML-images dataset \cite{wu2019tencent} also used the same path. Unlike ImageNet, these datasets have never been audited\footnote{In response to the mainstream media covering a pre-print of this work, we were informed that the curators of the dataset have withdrawn the dataset with a note accessible here: \url{https://groups.csail.mit.edu/vision/TinyImages/}} or scrutinized and some of the sordid results from inclusion of \textit{ethnophaulisms} in Tiny-Images dataset's label space are displayed in Figure \ref{fig:tiny_images}. The figure demonstrates both the number of images in a subset of the \textit{offensive} classes (sub-figure(a)) and the exemplar images (sub-figure(b)) that show the images in the noun-class labelled \texttt{n****r\footnote{Due to its offensiveness, we have censored this word (and other words throughout the paper), however, it remains uncensored on the website at the time of writing.}}, a fact that serves as a stark reminder that a great deal of work remains to be done by the ML community at large.
\begin{figure}[ht!]
\centering
\subfloat[Class-wise counts of the offensive classes]{
\includegraphics[width=0.45\textwidth]{plots/tiny_images/count_tiny_images.pdf}
}
\subfloat[Samples from the class labelled \texttt{n****r}]{
\includegraphics[width=0.375\textwidth]{plots/tiny_images/tiny_images_nig.pdf}
}
\caption{Results from the \textit{80 Million Tiny Images} dataset exemplifying the toxicities of it's label space}
\label{fig:tiny_images}
\end{figure}
Similarly, we found at least 315 classes\footnote{See \url{https://bit.ly/30DybmF}} of the potentially 1593 classes deemed to be \textit{non-imageable} by the ImageNet curators in \cite{yang2020towards_facct} still retained in the Tencent-ML-Images dataset that includes image classes such as \texttt{[transvestite, bad person, fornicatress, orphan, mamma's boy, and enchantress]}.
Finally, the \textit{labeling and validation} of the curation process also present ethical challenges. Recent work such as~\cite{gray2019ghost} has explored the intentionally hidden labour, which they have termed as \textit{Ghost Work}, behind such tasks. Image labeling and validation requires the use of crowd-sourced platforms such as MTurk, often contributing to the exploitation of underpaid and undervalued \textit{gig workers}. Within the topic of image labeling but with a different dimension and focus, recent work such as \cite{imagenet_mit} and \cite{2006.07159_arewedone} has focused on the shortcomings of human-annotation procedures used during the ImageNet dataset curation. These shortcomings, the authors point out, include single label per-image procedure that causes problems given that real-world images often contain multiple objects, and inaccuracies due to ``overly restrictive label proposals''.
\section{The threat landscape}
\label{sec: threats}
In this section, we survey the landscape of harm and threats, both immediate and long term, that emerge with dataset curation practices in the absence of careful ethical considerations and anticipation for negative societal consequences. Our goal here is to bring awareness to the ML and AI community regarding the severity of the threats and to motivate a sense of urgency to act on them. We hope this will result in practices such as the mandatory constitution of Institutional Review Boards (IRB) for large scale dataset curation processes.
\subsection{The rise of reverse image search engines, loss of privacy, and the blackmailing threat}
Large image datasets, when built without careful consideration of societal implications, pose a threat to the welfare and well-being of individuals. Most often, vulnerable people and marginalised populations pay a disproportionately high price. Reverse image search engines that allow face search such as \cite{Facesear28:online} have gotten remarkably and worryingly efficient in the past year. For a small fee, anyone can use their portal or their API\footnote{Please refer to the supplementary material in Appendix \ref{appendix:privacy_loss} for the screenshots} to run an automated process to uncover the ``real-world'' identities of the \textit{humans of ImageNet} dataset. For example, in societies where sex work is socially condemned or legally criminalized, re-identification of a sex worker through image search, for example, bears a real danger for the individual victim. %
Harmful discourse such as \textit{revenge porn}, are part of a broader continuum of image-based sexual abuse \cite{mcglynn2017beyond}. To further emphasize this specific point, many of the images in classes such as \texttt{maillot}, \texttt{brassiere}, and \texttt{bikini} contain images of beach voyeurism and other non-consensual cases of digital image gathering (covered in detail in Section \ref{sec:quant}). We were able to (unfortunately) easily map the victims, most of whom are women, in the pictures to ``real-world'' identities of people belonging to a myriad of backgrounds including teachers, medical professionals, and academic professors using reverse image search engines such as \cite{Facesear8:online}. Paying heed to the possibility of the \textit{Streisand effect}\footnote{The Streisand effect \textit{``is a social phenomenon that occurs when an attempt to hide, remove, or censor information has the unintended consequence of further publicizing that information, often via the Internet''} \cite{Streisan38:online}
}, we took the decision not to divulge any further quantitative or qualitative details on the extent or the location of such images in the dataset besides alerting the curators of the dataset(s) and making a passionate plea to the community not to underestimate the severity of this particular threat vector.
\subsection{The emergence of even larger and more opaque datasets} The attempt to build computer vision has been gradual and can be traced as far back as 1966 to Papert's \textit{The Summer Vision Project}~\citep{papert1966summer}, if not earlier. However, ImageNet, with its vast amounts of data, has not only erected a canonical landmark in the history of AI, it has also paved the way for even bigger, more powerful, and suspiciously opaque datasets. The lack of scrutiny of the ImageNet dataset by the wider computer vision community has only served to embolden institutions, both academic and commercial, to build far bigger datasets without scrutiny (see Table \ref{tab:datasets}). Various highly cited and celebrated papers in recent years~\cite{hinton2015distilling_jft,chollet2017xception_jft,brock2018large_jft,sun2017revisiting_jft}, for example, have used the \textit{unspoken unicorn} amongst large scale vision datasets, that is, the JFT-300M dataset [?]\footnote{We have decided to purposefully leave the '?' in place and plan to revisit it only after the dataset's creator(s) publish the details of it's curation}. This dataset is inscrutable and operates in the dark, to the extent that there has not even been official communication as to what \textit{JFT-300M} stands for. All that the ML community knows is it purportedly boasts more than 300M images spread across 18k categories. The open source variant(s) of this, the \textit{Open Images V4-5-6}~\cite{kuznetsova2018open} contains a subset of 30.1M images covering 20k categories (and also has an extension dataset with 478k crowd-sourced images across more than 6000 categories).
While parsing through some of the images, we found \textbf{verifiably}\footnote{See \url{https://bit.ly/2y1sC7i}. We performed verification with the uploader of the image via the Flickr link shared.} non-consensual images of children that were siphoned off of \textit{flickr} hinting towards the prevalence of similar issues for JFT-300M from which this was sourced. Besides the other large datasets in Table \ref{tab:datasets}, we have cases such as the \textit{CelebA-HQ} dataset, which is actually a \textit{heavily processed} dataset whose grey-box curation process only appears in Appendix-C of \cite{karras2017progressive} where no clarification is provided on this \textit{"frequency based visual quality metric"} used to sort the images based on \textit{quality}. Benchmarking any downstream algorithm of such an opaque, biased and (semi-)synthetic dataset will only result in controversial scenarios such as \cite{pulse}, where the authors had to hurriedly incorporate addendums admitting biased results.
Hence, it is important to reemphasize that the existence and use of such datasets bear direct and indirect impact on people, given that decision making on social outcomes increasingly leans on ubiquitously integrated AI systems trained and validated on such datasets. Yet, despite such profound consequences, critical questions such as where the data comes from or whether the images were obtained consensually are hardly considered part of the LSVD curation process.
The more nuanced and perhaps indirect impact of ImageNet is the \textit{culture} that it has cultivated within the broader AI community; a culture where the appropriation of images of real people as raw material free for the taking has come be to perceived as \textit{the norm}. Such norm and lack of scrutiny has played a role towards the creation of monstrous and secretive datasets without much resistance, prompting further questions such as `what other secretive datasets currently exist hidden and guarded under the guise of proprietary assets?'.
Current work that has sprung out of secretive datasets, such as Clearview AI~\cite{clearview_nyt}
\footnote{Clearview AI is a US based privately owned technology company that provides facial recognition services to various customers including North American law enforcement agencies. With more than 3 billion photos scraped from the web, the company operated in the dark until its services to law enforcement was reported in late 2019},
points to a deeply worrying and insidious threat not only to vulnerable groups but also to the very meaning of privacy as we know it \cite{ACLUsues41:online_khari}.
\subsection{The Creative Commons fallacy} In May 2007 the iconic case of \textit{Chang versus Virgin mobile: The school girl, the billboard, and virgin} \cite{ChangvVirgin59:online} unraveled in front of the world, leading to widespread debate on the uneasy relationship between personal privacy, consent, and image copyright, initiating a substantial corpus of academic debate (see \cite{corbett2009creative, corbett2011creative, hietanen2011creative, carroll2011school_creative}). A Creative Commons license addresses only copyright issues -- not privacy rights or consent to use images for training. Yet, many of the efforts beyond ImageNet, including the Open Images dataset \cite{kuznetsova2018open}, have been built on top of the \textit{Creative commons} loophole that large scale dataset curation agencies interpret as a \textit{free for all, consent-included} green flag. This, we argue, is fundamentally fallacious as is evinced in the views presented in \cite{UseandFa55:CCBY} by the Creative commons organization that reads:
\textit{``CC licenses were designed to address a specific constraint, which they do very well: unlocking restrictive copyright. But copyright is not a good tool to protect individual privacy, to address research ethics in AI development, or to regulate the use of surveillance tools employed online.''}. Datasets culpable of this \textit{CC-BY heist} such as \textit{MegaFace} and \textit{IBM's Diversity in Faces} have now been deleted in response to the investigations (see \cite{2001.03071_ibm_ms} for a survey) lending further support to the Creative Commons fallacy.
\subsection{Blood diamond effect in models trained on this dataset} Akin to the \textit{ivory carving-illegal poaching}, and \textit{diamond jewelry art-blood diamond} nexuses, we posit that there is a similar moral conundrum at play here that effects all downstream applications entailing models trained using a \textit{tainted} dataset. Often, these transgressions may be rather subtle. In this regard, we pick an examplar field of application that on the surface appears to be a low risk application area: \textit{Neural generative art}. Neural generative art created using tools such as BigGAN \cite{brock2018large_jft} and Art-breeder \cite{GANbreeder} that in turn use pre-trained deep-learning models trained on ethically dubious datasets, bear the downstream burden\footnote{Please refer to the appendix ( Section \ref{appendix:blood_diamond}) where we demonstrate one such real-world experiment entailing unethically generated neural art replete with responses obtained from human critiques as to what they felt about the imagery being displayed.} of the problematic residues from non-consensual image siphoning, thus running afoul of the Wittgensteinian edict of \textit{ethics and aesthetics being one and the same.} \cite{eaglestone2004one}. We also note that there is a privacy-leakage facet to this \textit{downstream burden}. In the context of face recognition, works such as \cite{song2017machine_toomuch} have demonstrated that CNNs with high predictive power unwittingly accommodate accurate extraction of subsets of the facial images that they were trained on, thus abetting dataset leakage\footnote{We'd like to especially highlight the \textit{megapixel.cc} project \cite{MegaPixels.cc} for the ground-breaking work on datasets to train such facial recognition systems }.
\subsection{Perpetuation of unjust and harmful stereotypes} Finally, zooming out and taking a broad perspective allows us to see that the very practice of embarking on a classification, taxonomization, and labeling task endows the classifier with the power to decide what is a legitimate, normal, or correct way of being, acting, and behaving in the social world \cite{bowker2000sorting}. For any given society, what comes to be perceived as \textit{normal} or \textit{acceptable} is often dictated by dominant ideologies. Systems of classification, which operate within a power asymmetrical social hierarchy, necessarily embed and amplify historical and cultural prejudices, injustices, and biases \cite{star2007enacting}. In western societies, ``desirable'', ``positive'', and ``normal'' characteristics and ways of being are constructed and maintained in ways that align with the dominant narrative, giving advantage to those that fit the status quo. Groups and individuals on the margins, on the other hand, are often perceived as the ``outlier'' and the ``deviant''. Image classification and labelling practices, without the necessary precautions and awareness of these problematic histories, pick up these stereotypes and prejudices and perpetuate them \cite{o2016weapons, noble2018algorithms, eubanks2018automating}. AI systems trained on such data amplify and normalize these stereotypes, inflicting unprecedented harm on those that are already on the margins of society. While the ImageNet team did initiate strong efforts towards course-correction \cite{yang2020towards_facct}, the Tiny Images dataset still contains harmful slurs and offensive labels. And worse, we remain in the dark regarding the secretive and opaque LSVDs.
\input{audit_card}
\section {Candidate solutions: The path ahead}
Decades of work within the fields of Science and Technology Studies (STS) and the Social Sciences show that there is no single straightforward solution to most of the wider social and ethical challenges that we have discussed \cite{suchman2007human,benjamin2019race,d2020data}. These challenges are deeply rooted in social and cultural structures and form part of the fundamental social fabric. Feeding AI systems on the world's beauty, ugliness, and cruelty, but expecting it to reflect only the beauty is a fantasy \cite{benjamin2019race}. These challenges and tensions will exist as long as humanity continues to operate. Given the breadth of the challenges that we have faced, any attempt for a quick fix risks concealing the problem and providing a false sense of solution. The idea of a complete removal of biases, for example, might in reality be simply hiding them out of sight \cite{gonen2019lipstick}. Furthermore, many of the challenges (bias, discrimination, injustice) vary with context, history, and place, and are concepts that continually shift and change constituting a moving target \cite{birhane2019algorithmic}. The pursuit of panacea in this context, therefore, is not only unattainable but also misguided. Having said that, there are remedies that can be applied to overcome the specific harms that we have discussed in this paper, which eventually potentially play constituent roles in improving the wider and bigger social and structural issues in the long run.
\subsection{Remove, replace, and open strategy}
In \cite{yang2020towards_facct}, the authors concluded that within the \textit{person sub-tree} of the ImageNet dataset, 1593 of the 2832 people categories were \textit{potentially offensive} labels and planned to \textit{"remove all of these from ImageNet."}. We strongly advocate a similar path for the offensive noun classes in the Tiny Images dataset that we have identified in section 2.1, as well as images that fall into the categories of verifiably\footnote{We use the term \textit{verifiably} to denote only those NSFW images that were hand-annotated by the volunteers indicating that they also contained the textual context that was of pornographic phraseology. We have an example grid of these images in the Appendix.} pornographic, shot in a non-consensual setting (up-skirt), beach voyeuristic, and exposed genitalia in the ImageNet-ILSVRC-2012 dataset. In cases where the image category is retained but the images are not, the option of \textit{replacement with consensually shot financially compensated images} arises. It is possible that some of the people in these images might come forward to consent and contribute their images in exchange for fair financial compensation, credit, or out of sheer altruism \cite{brown2019people}. We re-emphasize that our consternation focuses on the non-consensual aspect of the images and not on the category-class and the ensuing content of the images in it. This solution, however, brings forth further questions: does this make image datasets accessible only to those who can afford it? Will we end up with pool of images with a predominantly financially disadvantaged participants?
Science is self-correcting so long as it is accessible and open to critical engagement. We have tried to engage critically and map actionable ways forward given what we know of these LSVDs. The secretive and opaque LSVDs, however, thread a dangerous territory, given that they directly or indirectly impact society but remain hidden and inaccessible. Although the net benefit of the open science movement remains controversial, we strongly contend that making LSVDs open and accessible allows audits of these datasets, which is a first step towards a responsible scientific endeavour.
\subsection{Automated downstream removal from reverse search engines that allow for image deletion requests}
We found that some of the reverse image search engines do allow for \textit{users to remove particular image from our [sic] index} via their \textit{"Report abuse"} portals\footnote{See \url{https://pimeyes.com/en/faq/remove-from-database}}. This allows for dataset auditors to enlist images found in their dataset(s) containing identifiable individuals and direct them towards a guided image removal process from the reverse image search engine(s), in order to mitigate some aspects of immediate harm.
\subsection{Differentially private obfuscation of the faces}
This path entails harnessing techniques such as DP-Blur \cite{fan2018image} with quantifiable privacy guarantees to obfuscate the identity of the humans in the image. The \textit{Inclusive images challenge} \cite{shankar2017no_diversity}, for example, already incorporated blurring during dataset curation\footnote{\url{https://www.kaggle.com/c/inclusive-images-challenge} } and addressed the downstream effects surrounding change in predictive power of the models trained on the blurred versions of the dataset curated. We believe that replication of this template that also clearly included avenues for recourse in case of an erroneously non-blurred image being sighted by a researcher will be a step in the right direction for the community at large.
\subsection{Synthetic-to-real and Dataset distillation}
The basic idea here is to utilize (or augment) synthetic images in lieu of real images during model training.
Approaches include using hand-drawn sketch images (\textit{ImageNet-Sketch}~\citep{wang2019learning}), using GAN generated images \cite{GAN_aug} and techniques such as \textit{Dataset distillation}~\cite{wang2018dataset}, where a dataset or a subset of a dataset is distilled down to a few representative \textit{\textbf{synthetic}} samples. This is a nascent field with some promising results emerging in unsupervised domain adaptation across visual domains \cite{peng2018visda} and universal digit classification \cite{prabhu2019fonts}.
\subsection{Ethics-reinforced filtering during the curation}
The specific ethical transgressions that emerged during our longitudinal analysis of ImageNet could have been prevented if there were explicit instructions provided to the \textit{MTurkers} during the dataset curation phase to enable filtering of these images at the source (See Fig.9 in~\cite{recht2019imagenet_v2} for example). We hope ethics checks become an integral part of the User-Interface deployed during the humans-in-the-loop validation phase for future dataset curation endeavors.
\subsection{Dataset audit cards}
As emphasized in Section 4, context is crucial in determining whether a certain dataset ethical or problematic as it provides a vital background information and datasheets are an effective way of providing context. Much along the lines of \textit{model cards} \cite{model_cards} and \textit{datasheet for datasets} \cite{gebru2018datasheets}, we propose dissemination of \textit{dataset audit cards}. This allows large scale image dataset curators to publish the goals, curation procedures, known shortcomings and caveats alongside their dataset dissemination. In Figure \ref{fig:audit_card}, we have curated an example dataset audit card for the ImageNet dataset using the quantitative analyses carried out in Section~\ref{sec:quant}
\section{Quantitative dataset auditing: ImageNet as a template}
\label{sec:quant}
\begin{table*}[ht!]
\centering
\caption{Meta datasets curated during the audit processes}
\begin{tabularx}{\textwidth}{|l|l|X|}
\hline
file\_name &
shape &
file\_contents \\ \hline
df\_insightface\_stats.csv &
(1000, 30) &
24 classwise statistical parameters obtained by running the \texttt{InsightFace} model (\cite{insight-face}) on the ImageNet dataset \\ \hline
df\_audit\_age\_gender\_dex.csv &
(1000, 12) &
11 classwise (ordered by the wordnet-id) statistical parameters obtained from the json files (of the DEX paper) \cite{rothe2018deep_dex} \\ \hline
df\_nsfw.csv &
(1000, 5) &
The mean and std of the NSFW scores of the train and val images arranged per-class. (Unnamed: 0: WordNetID of the class) \\ \hline
df\_acc\_classwise\_resnet50.csv &
(1000, 7) &
Classwise accuracy metrics (\& the image level preds) obtained by running the ResNet50 model on ImageNet train and Val sets \\ \hline
df\_acc\_classwise\_NasNet\_mobile.csv &
(1000, 7) &
Classwise accuracy metrics (\& the image level preds) obtained by running the NasNet model on ImageNet train and Val sets \\ \hline
df\_imagenet\_names\_umap.csv &
(1000, 5) &
Dataframe with 2D UMAP embeddings of the Glove vectors of the classes of the ImageNet dataset \\ \hline
df\_census\_imagenet\_61.csv &
(1000, 61) &
The MAIN census dataframe covering class-wise metrics across 61 parameters, all of which are explained in df\_census\_columns\_interpretation.csv \\ \hline
df\_census\_columns\_interpretation.csv &
(61, 2) &
The interpretations of the 61 metrics of the census dataframe above \\ \hline
df\_hand\_survey.csv &
(61, 3) &
Dataframe contaimning the details of the 61 images unearthed via hand survey (Do not pay heed to 61. it is a mere coincidence) \\ \hline
df\_classes\_tiny\_images\_3.csv &
(75846, 3) &
Dataframe containing the class\_ind, class\_name (wordnet noun) and n\_images \\ \hline
df\_dog\_analysis.csv &
(7, 4) &
Dataframe containing breed, gender\_ratio and survey result from the paper Breed differences in canine aggression' \\ \hline
\end{tabularx}
\label{tab:csv_census}
\end{table*}
We performed a cross-categorical quantitative analysis of ImageNet to assess the extent of the ethical transgressions and the feasibility of model-annotation based approaches. This resulted in an \textit{ImageNet census}, entailing both image-level as well as class-level analysis across the $57$ different metrics (see supplementary section) covering Count, Age and Gender (CAG), NSFW-scoring, semanticity of class labels and accuracy of classification using pre-trained models. We have distilled the important revelations of this census as a \textit{dataset audit card} presented in Figure \ref{fig:audit_card}. This audit also entailed a human-in-the-loop based hybrid-approach that the pre-trained-model annotations (along the lines of \cite{dulhanty2019auditing,yang2020towards_facct}) to segment the large dataset into smaller sub-sets and hand-label the smaller subsets to generate two lists covering 62 misogynistic images and 30 image-classes with co-occuring children.
We used the \texttt{DEX} \cite{rothe2018deep_dex} and the \texttt{InsightFace} \cite{insight-face} pre-trained models\footnote{While harnessing these pre-trained gender classification models, we would like to \textbf{strongly emphasize} that the specific models and the \textit{problems} that they were intended to solve, when taken in isolation, stand on ethically dubious grounds themselves. In this regard, we strongly concur with previous work such as \cite{ainow_1} that \textit{gender classification} based on appearance of a person in a digital image is both \textbf{scientifically flawed} and is a technology that bears a high risk of systemic abuse.} to generate the cardinality, gender skewness, and age-distribution results captured in Figure~\ref{fig:cag_card}. This resulted in discovery of \textbf{83,436} images with persons, encompassing \textbf{101,070 to 132,201} individuals, thus constituting $8-10\%$ of the dataset. Further, we munged together gender, age, class semanticity\footnote{ Obtained using \textit{GloVe embeddings} \cite{pennington2014glove} on the labels} and NSFW content flagging information from the pre-trained \texttt{NSFW-MobileNet-v2} model \cite{GantMann83:NSFW} to help perform a guided search of misogynistic consent-violating transgressions. This resulted in discovery of five dozen plus images\footnote{Listed in \texttt{df\_hand\_survey.csv}} across four categories: \textit{beach-voyeur-photography, exposed-private-parts, verifiably pornographic and upskirt} in the following classes: \textit{445-Bikini, 638 -maillot, 639-tank suit, 655-miniskirt} and \textit{459-brassiere} (see Figure \ref{fig:agn_card}).
Lastly, we harnessed literature from areas spanning from dog-ownership bias (\cite{hirschman1994_dog_gender_1},\cite{ramirez2006_dog_gender_2}) to engendering of musical instruments (\cite{wrape2016gender_music_gender}, \cite{bullerjahn2016masculine_music}) to generate analysis of subtle forms of \textit{human co-occurrence}-based gender bias in Figure \ref{fig:co_occur_card}.
\\Captured in Table \ref{tab:csv_census} are the details of the \textit{csv} formatted data assets curated for the community to build on. The CAG statistics are covered in { df\_insightface\_stats.csv} and {df\_audit\_age\_gender\_dex.csv}. Similarly, we have also curated NSFW scoring ({df\_nsfw.csv}), Accuracy ({df\_acc\_classwise\_resnet50/\_NasNet\_mobile.csv}) and Semanticity ({df\_imagenet\_names\_umap.csv}) datasets as well. {df\_census\_imagenet\_61.csv} contains the 61 cumulative paramaters for each of the 1000 classes (with their column interpretations in {df\_census\_columns\_interpretation.csv}). We have duly open-sourced these meta-datasets and 14 tutorial-styled Jupyter notebooks (spanning both ImageNet and Tiny-Images datasets) for community access\footnote{ Link: \url{https://rb.gy/zccdps}}.
\begin{table*}[]
\centering
\begin{tabular}{|l|l|}
\hline
Metric & Models used \\ \hline
Count, Age and Gender & \texttt{DEX} \cite{rothe2018deep_dex}, \texttt{InsightFace} \cite{insight-face}, \texttt{RetinaFace} \cite{deng2019retinaface}, \texttt{ArcFace} \cite{deng2018arcface} \\ \hline
NSFW-scoring & \texttt{NSFW-MobileNet-V2-224} \cite{GantMann83:NSFW} \\ \hline
Semanticity & \texttt{Glove} \cite{pennington2014glove}, \texttt{UMAP} \cite{mcinnes2018umap} \\ \hline
Classification Accuracy & \texttt{Resent-50} \cite{he2016deep:ResNet}, \texttt{NasNet-mobile} \cite{zoph2018:NasNet} \\ \hline
\end{tabular}
\caption{Metrics considered and pre-trained models used}
\label{tab:models}
\end{table*}
\begin{table*}
\centering
\captionsetup{justification=centering}
\begin{tabular}{rrrrrrrrr}
\toprule
$N_{train-O}^{(dex)}$ & $n_{train-O}^{(if)}$ & $n_{val-O}^{(if)}$ & $N_{train-O}^{(if)}$ & $N_{val-O}^{(if)}$ & $N_{train-W}^{(if)}$ & $N_{train-M}^{(if)}$ & $N_{val-W}^{(if)}$ & $N_{val-M}^{(if)}$\\
\midrule
132,201 & 80,340 & 3,096 & 97,678 & 3,392 & 26,195 & 71,439 & 645 & 2,307 \\
\bottomrule
\end{tabular}
\caption{ \textit{Humans of the imagenet dataset}: How many? \\ Key: $\left\{ {n/N} \right\}_{\{ train/val\} - \{ O/W/M\} }^{(\left\{ {dex/if} \right\})}$.(O:Overall,W:Women,M: Men) }
\label{tab:summary}
\end{table*}
\begin{table*}[]
\begin{tabular*}{\textwidth}{c @{\extracolsep{\fill}} |c|c|c|c|c|c|}
\toprule
class\_number &label & mean\_gender\_audit & mean\_age\_audit & mean\_nsfw\_train \\
\midrule
445 & bikini, two-piece & 0.18 & 24.89 & 0.859 \\
638 & maillot & 0.18 & 25.91 & 0.802 \\
639 & maillot, tank suit & 0.18 & 26.67 & 0.769 \\
655 & miniskirt, mini & 0.19 & 29.95 & 0.62 \\
459 & brassiere, bra, bandeau & 0.16 & 25.03 & 0.61 \\
\bottomrule
\end{tabular*}
\caption{Table of the 5 classes for further investigation that emerged from the NSFW analysis}
\label{tab:agn}
\end{table*}
\section{Conclusion and discussion}
We have sought to draw the attention of the machine learning community towards the societal and ethical implications of large scale datasets, such as the problem of non-consensual images and the oft-hidden problems of categorizing people. ImageNet has been championed as one of the most incredible breakthroughs in computer vision, and AI in general. We indeed celebrate ImageNet's achievement and recognize the creators' efforts to grapple with some ethical questions. Nonetheless, ImageNet as well as other large image datasets remain troublesome.
In hindsight, perhaps the ideal time to have raised ethical concerns regarding LSVD curation would have been in 1966 at the birth of \textit{The Summer Vision Project} \cite{papert1966summer}. The right time after that was when the creators of ImageNet embarked on the project to ``map out the entire world of objects''. Nonetheless, these are crucial conversations that the computer vision community needs to engage with \textbf{now} given the rapid democratization of imaging scraping tools (\cite{bingscra68:online_pip1,ImageScr23:online_pip2,imagebot60:online_pip3}) and \textit{dataset-zoos} (\cite{Google_Dataset_search:online,DS_TensorFl10:online,DS_torchvis46:online}). The continued silence will only serve to cause more harm than good in the future.
In this regard, we have outlined a few solutions, including \textit{audit cards}, that can be considered to ameliorate some of the concerns raised. We have also curated meta-datasets and open-sourced the code to carry out quantitative auditing using the ILSVRC2012 dataset as a template. However, we posit that the deeper problems are rooted in the wider structural traditions, incentives, and discourse of a field that treats ethical issues as an afterthought. A field where \textit{in the wild} is often a euphemism for \textit{without consent}. We are up against a system that has veritably mastered \textit{ethics shopping}, \textit{ethics bluewashing}, \textit{ethics lobbying}, \textit{ethics dumping}, and \textit{ethics shirking} \cite{floridi2019translating}.
Within such an ingrained tradition, even the most thoughtful scholar can find it challenging to pursue work outside the frame of the ``tradition''. Subsequently, radical ethics that challenge deeply ingrained traditions need to be incentivised and rewarded in order to bring about a shift in culture that centres justice and the welfare of disproportionately impacted communities. We urge the machine learning community to pay close attention to the direct and indirect impact of our work on society, especially on vulnerable groups. Awareness of historical antecedents, contextual, and political dimensions of current work is imperative is this regard. We hope this work contributes in raising awareness regarding the need to cultivate a justice centred practice and motivates the constitution of IRBs for large scale dataset curation processes.
\section{Acknowledgements}
This work was supported, in part, by Science Foundation Ireland grant 13/RC/2094 and co-funded under the European Regional Development Fund through the Southern \& Eastern Regional Operational Programme to Lero - the Irish Software Research Centre (www.lero.ie).
The authors would like to thank Alex Hanna, Andrea E. Martin, Anthony Ventresque, Elayne Ruane, John Whaley, Mariya Vasileva, Nicolas Le Roux, Olivia Guest, Os Keyes, Reubs J. Walsh, Sang Han, and Thomas Laurent for their useful feedback on an earlier version of this manuscript.
\begin{appendices}
\section{Risk of privacy loss via reverse search engines}
\label{appendix:privacy_loss}
As covered in the main paper, reverse image search engines that facilitate face search such as \cite{Facesear28:online} have gotten remarkably and worryingly efficient in the past year. For a small fee, anyone can use their portal or their API to run an automated process and uncover the ``real-world'' identities of the \textit{humans of ImageNet} dataset. While all genders in the imagenet dataset are under this risk, there is asymmetric risk here as the high NSFW classes such as \texttt{bra, bikini and maillot} are often the ones with higher female-to-men ratio (See Figure \ref{fig:nsfw_scatter}). Figure \ref{fig:reverse_image} showcases a snapshot image of one such reverse image search portal to demonstrate how easy it is for anyone to access their GUI and uncover ``real world'' identities of people which can lead to catastrophic downstream risks such as blackmailing and other forms on online abuse.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.5\textwidth]{plots/example_blurred.png}
\caption{Snapshot of a popular reverse image search website}
\label{fig:reverse_image}
\end{figure}
\section{Quantitative auditing}
\label{appendix:Quantitative_auditing}
In this section, we cover the details of performing the quantitative analysis on the ImageNet dataset including the following metrics: Person CAG (Count -Age - Gender) , NSFW scoring of the images, Semanticity and classification accuracy. The pre-trained models used in this endeavor are covered in Table \ref{tab:models}.
All of these analyses and the generated meta-datasets have been open sourced at \url{https://rb.gy/zccdps}.
Figure \ref{fig:assets_map} covers the details of all the jupyter notebooks authored to generate the datasets covered in Table \ref{tab:csv_census}.
\begin{figure*}[ht!]
\centering
\includegraphics[width=0.95\textwidth]{Supplement/assets_map.pdf}
\caption{Visualization of all the notebooks and dataset assets curated during the quantitative analysis}
\label{fig:assets_map}
\end{figure*}
\subsection{Count, Age and Gender}
\begin{figure}[!h]
\centering
\includegraphics[width=0.5\textwidth]{Supplement/dex_errors.png}
\caption{An example image with the output bounding boxes and the confidence scores of the humans detected in the image by the \texttt{DEX} model(\cite{rothe2018deep_dex})}
\label{fig:dex_errors}
\end{figure}
In order to perform a human-centric census covering metrics such as count, age, and gender, we used the \texttt{InsightFace} toolkit for face analysis \cite{insight-face}, that provided implementations of: \texttt{ArcFace} for deep face recognition \cite{deng2019retinaface} and \texttt{Retina-Face} for face localisation (bounding-box generation) \cite{deng2019retinaface}. We then combined the results of these models with the results obtained from \cite{dulhanty2019auditing} that used the \texttt{DEX} \cite{rothe2018deep_dex} model. The results are as shown in Table \ref{tab:summary} that captures the summary statistics for the \texttt{ILSVRC2012} dataset. In this table, the lower case $n$ denotes the number of \textit{images with persons} identified in them whereas $N$ indicates the \textit{number of persons}\footnote{The difference is simply on account of more than one person being identified by the model in a given image}. The superscript indicates the algorithm used (\texttt{DEX} or \texttt{InsightFace} (if) ) whereas the subscript has two fields: The train or validation subset indicator and the census gender-category. For example, $n_{val-O}^{(if)}=3096$ implies that there were 3096 images in the ImageNet validation set (out of $50000$) where the \texttt{InsightFace} models were able to detect a person's face.
As shown, the \texttt{InsightFace} model identified 101,070 persons across 83,436 images (including the train and validation subsets) which puts the prevalence rate of persons whose presence in the dataset exists sans explicit consent to be around $7.6\%$ which is less aggressive compared to the $10.3\%$ predicted by the DEX model (that focussed on the training subset), which has a higher identification false positive rate. An example of this can be seen in Fig \ref{fig:dex_errors} which showcases an example image with the bounding boxes of the detected persons in the image.
Much akin to \cite{dulhanty2019auditing}, we found a strong bias towards (relatively older) male presence (73,746 with a mean age of 33.24 compared to 26,840 with a mean age of 25.58). At this juncture, we would like to reemphasize that these \textit{high accuracy} pre-trained models can indeed be highly error prone conditioned on the ethnicity of the person, as analyzed in \cite{dulhanty2019auditing,buolamwini2018gender} and we would like to invite the community to re-audit these images with better and more ethically responsible tools (See Fig \ref{fig:age_errors} for example of errors we could spot during the inference stage).
Figure \ref{fig:dex_if_a}, presents the class-wise estimates of the number of persons in the dataset using the DEX and the InsightFace models. In Figure \ref{fig:dex_if_b}, we capture the variation in the estimates of count, gender and age of the \texttt{DEX} and the \texttt{InsightFace} models.
Before delving in to the discussions of the results obtained, we define the parameters that were measured.
To begin, we denote $\phi_i$ to be the binary \textit{face-present} indicator variable(
$
{\phi _i} =
\begin{cases}
\text{1} &\text{if face present }\\
\text{0} &\text{otherwise} \\
\end{cases}
$) with regards to the image indexed $i$,
$(A)$ (in the superscripts) to be the algorithm used ($A \in \{\texttt{DEX,INSIGHTFACE}\}$), and $N_c$ to be the number of images in the class $c$. Now, we define the class-level mean person count ($\eta _c^{(A)}$), mean-gender-skewness score ($\xi _c^{(A)}$) and mean-age ($\alpha_c^{(A)}$) to be,
\begin{align*}
\eta _c^{(A)} & = \frac{1}{{{N_c}}}\sum\limits_{i = 1}^{{N_c}} {{\phi _i}}\\
\alpha _c^{(A)} & = \frac{1}{{{N_c}}}\sum\limits_{i = 1}^{{N_c}} {{\phi _i}a_i^{(A)}}\\
\xi _c^{(A)} & = \frac{1}{{{N_c}}}{\sum\limits_{i = 1}^{{N_c}} {{\phi _i}\left( {\frac{{g_i^{(A)} - \mu _c^{(A)}}}{{\sigma _c^{(A)}}}} \right)} ^3}
\end{align*}
Here, $a_i^{(A)}$ is the age-estimate of the person generated by algorithm $(A)$ in the $i^{th}$ image and ${\mu _c^{(A)}}$ and ${\sigma_c^{(A)}}$ represent the mean and standard-deviation of the gender-estimate of the images belonging to class $c$ and estimated by algorithm $(A)$ respectively.\\
\begin{figure}[!ht]
\centering
\includegraphics[width=0.5\textwidth]{Supplement/age_error_blurred.png}
\caption{An example image with the output bounding boxes and the estimated ages/ (binarized) genders of the persons detected in the image by the \texttt{InsightFace} model. (Here 0: female and 1: Male)}
\label{fig:age_errors}
\end{figure}
\begin{figure*}[ht!]
\centering
\subfloat[Class-wise estimates of number of humans in the images]{
\includegraphics[width=0.95\textwidth]{plots/census/1a_dex_if_Nfaces.pdf}
\label{fig:dex_if_a}
}
\hspace{0mm}
\subfloat[Scatter-plots with correlations covering the cardinality, age and gender estimates]{
\includegraphics[width=0.95\textwidth]{plots/census/1b_dex_if_comparisons.pdf}
\label{fig:dex_if_b}
}
\caption{Juxtaposing the results from the \texttt{DEX} and the \texttt{InsightFace} models}
\label{fig:dex_if}
\end{figure*}
With regards to the first scatter-plot in Figure \ref{fig:dex_if}(b), we observe that the estimated class-wise counts of persons ($\eta _c^{(A)}$) detected by the \texttt{DEX} and \texttt{InsightFace} models in the images were in strong agreement ($Pearson-r=0.973(0.0)$) which helps to further establish the global person prevalence rate in the images to be in the order of $7.6-10.3\%$. These scatter-plots constitute Figure \ref{fig:cag_card} of the dataset audit card (Figure \ref{fig:audit_card}).
\\Now, we would like to draw the attention of the reader towards the weaker correlation ($Pearson-r=0.723(0.0)$) when it came to gender-skewness ($\xi _c^{(A)}$) and the mean age-estimates ($\alpha _c^{(A)};\ Pearson-r=0.567(0.0)$) scatter-plots in Figure \ref{fig:dex_if}(b). Given that the algorithms used are state-of-the-art with regards to the datasets they have been trained on (see \cite{rothe2018deep_dex} and \cite{insight-face}), the high disagreement on a ``neutral'' dataset like ImageNet exposes the frailties of these algorithmic pipelines upon experiencing population shifts in the test dataset. This, we believe, lends further credence to the studies that have demonstrated poor reliability of these so-termed accurate models upon change of the underlying demographics (see \cite{dulhanty2019auditing} and \cite{buolamwini2018gender}) and further supports the need to move away from gender classification on account of not just the inherent moral and ethical repugnance of the task itself but also on its lack of merit for \textit{scientific validity} \cite{ainow_1}.
\subsection{NSFW scoring aided misogynistic imagery hand-labeling}
Previous journalistic efforts (see \cite{Insideth82:online_register}) had revealed the presence of strongly misogynistic content in the ImageNet dataset, specifically in the categories of \texttt{beach-voyeur-photography, upskirt images, verifiably pornographic and exposed private-parts}. These specific four categories have been well researched in digital criminology and intersectional feminism (see \cite{henry2017not,mcglynn2017beyond,powell2018image,powell2010configuring}) and have formed the backbone of several legislations worldwide (see \cite{mcglynn2017more_law1},\cite{gillespie2019tackling_law2}). In order to help generate a hand labelled dataset of these images amongst more than 1.3 million images, we used a hybrid human-in-the-loop approach where we first formed a smaller subset of images from image classes filtered using a model-annotated NSFW-average score as a proxy. For this, we used the \texttt{NSFW-Mobilenet-v2} model \cite{GantMann83:NSFW} which is an image-classification model with the output classes being \texttt{[drawings, hentai, neutral, porn, sexy]}. We defined the NSFW score of an image by summing up the softmax values of the \texttt{[hentai, porn, sexy]} subset of classes and estimated the mean-NSFW score of all of the images of a class to obtain the results portrayed in Figure \ref{fig:agn}. On the left hand side of Figure \ref{fig:agn}, we see the scatter-plot of the mean-NSFW scores plotted against the mean-gender scores (obtained from the DEX model estimates) for the 1000 imagenet classes. We then found five natural clusters upon using the \textit{Affinity Propagation} algorithm \cite{frey2007clustering_ap}. Given the \texttt{0:FEMALE|1:MALE} gender assignments in the model we used (see \cite{dulhanty2019auditing}), classes with lower mean-gender scores allude towards a \textit{women-majority class}). The specific details of the highlighted cluster in the scatter-plot in Figure \ref{fig:agn} are displayed in Table \ref{tab:agn}. Further introducing the age dimension (by way of utilising the mean-age metric for each class), we see in the right hand side of Figure \ref{fig:agn}, that the classes with the highest NSFW scores were those where the dominating demographic was that of young women. With this shortlisting methodology, we were left with approximately 7000 images which were then hand labelled by a team of five volunteers (three male, two female, all aged between 23-45) to curate a list of $61$ images where there was complete agreement over the 4 class assignment. We have open-sourced the hand-curated list (see Table \ref{tab:hand_survey}), and the summary results are as showcased in Figure \ref{fig:hand_labelled}. In sub-figure Figure \ref{fig:hand_labelled_grid}, we see the cross-tabulated class-wise counts of the four categories of images\footnote{ This constitutes Figure \ref{fig:agn_card}( in the data audit card)} across the imagenet classes and in Figure \ref{fig:hand_labelled_histo}, we present the histogram-plots of these 61 hand-labelled images across the imagenet classes. As seen, the \texttt{bikini, two-piece} class with a mean NSFW score of $0.859$ was the \textit{main} image class with 24 confirmed \texttt{beach-voyeur} pictures.
Here, we would like to strongly reemphasise that we are disseminating this list as a community resource so as to facilitate further scholarly engagement and also, if need be, to allow scholars in countries where incriminating laws (see \cite{durham2015opposing}) may exist, to deal with in the appropriate topical way deemed fit. We certainly admit to the primacy of \textbf{context} in which the \textit{objectionable} content appears. For example, the image \texttt{n03617480\_6206.jpeg} in the class \texttt{n03617480 - kimono} that contained genital exposure, turned out to be a photographic bondage art piece shot by Nobuyoshi Araki\cite{ozaki2008shashinjinsei} that straddles the fine line between \textit{scopophilic eroticism} and pornography. But, as explored in \cite{durham2015opposing}, the mere possession of a digital copy of this picture would be punishable by law in other nations and we believe that these factors have to be considered contextually while disseminating a large scale image dataset and should be detailed as caveats in the dissemination document.
\subsubsection{NSFW and semanticity of classes}
\label{appendix:nsfw_semanticity}
We also analyzed the relationship between the semanticity of classes and NSFW scores. Firstly, we obtained a representative word for each of the 1000 class labels in ILSVRC2012 and used \cite{pennington2014glove} to generate dense word-vector \textit{Glove} embeddings in 300-D. Further, in order to generate the 2D/3D scatter-plots in Figure \ref{fig:nsfw_scatter}, we used the UMAP \cite{mcinnes2018umap} algorithm to perform dimensionality reduction. \texttt{df\_imagenet\_names\_umap.csv } contains the 2D UMAP embeddings of the resultant Glove vectors of the classes that are then visualized in Figure \ref{fig:nsfw_scatter} (a). In Figure \ref{fig:nsfw_scatter} (b), we see the 3D surface plot of the 2D UMAP semantic dimensions versus the NSFW scores. As seen, it is peaky in specific points of the semantic space of the label categories mapping to classes such as \texttt{brassier, bikini and maillot}.
\begin{figure*}[ht!]
\centering
\includegraphics[width=0.95\textwidth]{plots/census/3b_umap_nsfw_train_scatter.pdf}
\caption{Figure showcasing the relationship between the semanticity of classes and the class-wise mean NSFW scores}
\label{fig:nsfw_scatter}
\end{figure*}
\subsection{Dogs to musical instruments: Co-occurrence based gender biases}
\label{appendix:dogs_musical}
Social, historical, and cultural biases prevalent in the society feed into datasets and the statistical models trained on them. In the context of Natural Language Processing (NLP), the framework of lexical co-occurrence has been harnessed to tease out these biases, especially in the context of gender biases. In \cite{tan2019assessing}, the authors analyzed occupation words stereotypically perceived as male (that they termed as \textit{M-biased} words) as well as occupation words stereotypically perceived as female (\textit{F-biased} words) in large text corpora and the ensuing downstream effects when used to generate contextual word representations in SoTA models such as such as BERT and GPT-2. Further, in \cite{Rekabsaz_Henderson_West_Hanbury_2020}, direct normalized co-occurrence associations between the word and the representative concept words were proposed as a novel corpus bias measurement method, and its efficacy was demonstrated with regards to the actual gender bias statistics of the U.S. job market and its estimates measured via the text corpora. In the context of the ImageNet dataset, we investigated if such co-occurrence biases do exist in the context of human co-occurrence in the images. Previously, in \cite{stock2018convnets_basketball}, the authors had explored the biased representation learning of an ImageNet trained model by considering the class \texttt{basketball} where images containing black persons were deemed \textit{prototypical}. Here, we tried to investigate if the gender of the person co-occurring in the background alongside the non-person class was skewed along the lines that it is purported to be in related academic work. We performed these investigations in the context of person-occurrence with regards to dog-breeds as well as musical instruments. Presented in Figure \ref{fig:co_occur_gender_bias} (a) are the conditional violin plots relating the dog-breed group of the image class of a subset of the ImageNet dataset in comparison with the with the mean gender score obtained from the \texttt{DEX} model analyses. We obtained these measurements in two phases. In the first phase, we grouped the $120$ ImageNet classes of dog-breeds in to the following 7 groups: \texttt{[Toy, Hound ,Sporting, Terrier, Non-Sporting, Working, Herding]} following the formal American Kennel Club\footnote{AKC claims that registered breeds are \textit{assigned to one of seven groups representing characteristics and functions the breeds were originally bred for.}} (AKC) groupings (see \cite{ListofBr98:AKC}). The remaining breeds not in the AKC list were placed into the \texttt{Unknown} group. Once grouped, we computed the gender-conditioned population spreads of person-concurrence using the mean-gender value of the constituent image classes obtained estimated from \cite{dulhanty2019auditing}. Prior literature (see \cite{hirschman1994_dog_gender_1,ramirez2006_dog_gender_2}) has explored the nexus between the perceived \textit{manliness} of dog groups and the ownership gender. These stereotypical associations were indeed reflected in the person co-occurrence gender distributions in Figure \ref{fig:co_occur_gender_bias_dog}, where we see that the so perceived \textit{masculine} dog groups belonging to the set \texttt{[Non-Sporting, Working, Herding]} had a stronger male-gender co-occurrence bias. \\
In a similar vein, in Figure \ref{fig:co_occur_gender_bias_instrument} we present the gender-skewness ($\xi _c^{(DEX)} = \frac{1}{{{N_c}}}{\sum\limits_{i = 1}^{{N_c}} {I[{\phi _i}]\left( {\frac{{g_i^{(DEX)} - {\mu _c}}}{{{\sigma _c}}}} \right)} ^3}$) variation amongst the co-occurring persons across the 17 imagenet musical instrument classes. Works such as \cite{delzell1992gender_music}, \cite{zervoudakes1994gender_music} and \cite{bullerjahn2016masculine_music} have explored in depth, the gender biases there exist in musical instrument selection. As stated in \cite{wrape2016gender_music_gender}, instruments such as the \texttt{cello}, \texttt{oboe}, \texttt{flute} and \texttt{violin} have been stereotypically tagged to be \textit{feminine} whereas instruments such as the \texttt{drum}, \texttt{banjo}, \texttt{trombone}, \texttt{trumpet} and the \texttt{saxophone} were the so-termed \textit{masculine} instruments in the western context. While these stereotypes represent current and historical norms, the \textit{west-centric-bias} \footnote{See \url{https://www.kaggle.com/c/inclusive-images-challenge}} of the search engine used to curate the dataset has resulted in the mirroring of these topical real-world association biases. As seen in Figure \ref{fig:co_occur_gender_bias_instrument}, \texttt{harp}, \texttt{cello}, \texttt{oboe}, \texttt{flute} and \texttt{violin} indeed had the strongest \textit{pro-women} bias where as \texttt{drum}, \texttt{banjo}, \texttt{trombone}, \texttt{trumpet} and \texttt{saxophone} were the classes with the strongest \textit{male leaning} skewness scores.
\subsection{Classes containing pictures of infants}
\label{appendix:infants}
We found this category to be particularly pertinent both under the wake of strong legislations protecting privacy of children's digital images as well as the extent of it. We found pictures of infants and children across the following 30 image classes (and possibly more): ['\texttt{bassinet}',
'\texttt{cradle}',
'\texttt{crib}',
'\texttt{bib}',
'\texttt{diaper}',
'\texttt{bubble}',
'\texttt{sunscreen}',
'\texttt{plastic bag}',
'\texttt{hamper}',
'\texttt{seat belt}',
'\texttt{bath towel}',
'\texttt{mask}',
'\texttt{bow-tie}',
'\texttt{tub}',
'\texttt{bucket}',
'\texttt{umbrella}',
'\texttt{punching bag}',
'\texttt{maillot - tank suit}',
'\texttt{swing}',
'\texttt{pajama}',
'\texttt{horizontal bar}',
'\texttt{computer keyboard}',
'\texttt{shoe-shop}',
'\texttt{soccer ball}',
'\texttt{croquet ball}',
'\texttt{sunglasses}',
'\texttt{ladles}',
'\texttt{tricycle - trike - velocipede}',
'\texttt{screwdriver}',
'\texttt{carousel}'].
What was particularly unsettling was the prevalence of entire classes such as '\texttt{bassinet}', '\texttt{cradle}', '\texttt{crib}' and '\texttt{bib}' that had a very high density of images of infants. We believe this might have legal ramifications as well. For example, Article 8 of the European Union General Data Protection Regulation (GDPR), specifically deals with the \textit{conditions applicable to child’s consent in relation to information society services} \cite{EURLex3255:GDPR}. %
The associated \textit{Recital 38} states verbatim that \textit{Children merit specific protection with regard to their personal data, as they may be less aware of the risks, consequences and safeguards concerned and their rights in relation to the processing of personal data. Such specific protection should, in particular, apply to the use of personal data of children for the purposes of marketing or creating personality or user profiles and the collection of personal data with regard to children when using services offered directly to a child}. Further, \textbf{Article 14} of GDPR explicitly states: \textit{Information to be provided where personal data have not been obtained from the data subject}. We advocate allying with the legal community in this regard to address the concerns raised above.
\subsection{ Blood diamond effect in models trained on this dataset}
\label{appendix:blood_diamond}
Akin to the \textit{ivory carving-illegal poaching} and \textit{diamond jewelry art-blood diamond} nexuses, we posit there is a similar moral conundrum at play here and would like to instigate a conversation amongst the neural artists in the community.
The emergence of tools such as BigGAN \cite{brock2018large_jft} and GAN-breeder \cite{GANbreeder} has ushered in an exciting new flavor of generative digital art \cite{boden2009generative}, generated using deep neural networks (see \cite{hertzmann2019aesthetics} for a survey). A cursory search on twitter\footnote{\url{https://twitter.com/hashtag/biggan?lang=en }} reveals hundreds of interesting art-works created using BigGANs. There are many detailed blog-posts\footnote{\url{https://rb.gy/pr9pwb}} on generating neural art by beginning with seed images and performing nifty experiments in the latent space of BigGANs. At the point of writing the final version of this paper, (6/26/2020, 10:34 PM PST), users on the \textit{ArtBreeder} app\footnote{\url{https://ganbreeder.app}} had \textit{generated} 64683549 images.
Further, \textit{Christie's}, the British auction house behemoth, recently hailed the selling of the neural network generated \textit{Portrait of Edmond Belamy} for an incredible $\$432,500$ as \textit{signalling the arrival of AI art on the world auction stage}\cite{christies}.
Given the rapid growth of this field, we believe this is the right time to have a critical conversation about a particularly dark ethical consequence of using such frameworks that entail models trained on the \textit{ImageNet} dataset which has many images that are pornographic, non-consensual, voyeuristic and also entail underage nudity. We argue the use of ill-considered seed images to train the models trickles down to the final art-form in a way similar to the blood-diamond syndrome in jewelry art \cite{fishman2004diamond}.
\textbf{An example:} Consider the neural art image in Figure \ref{fig:example} we generated using the \textit{GanBreeder} app. On first appearance, it is not very evident as to what the constituent \textit{seed} classes are that went into the creation of this neural artwork image. When we solicited volunteers online to critique the artwork (see the collection of responses in Table \ref{tab:responses}), none had an inkling regarding a rather sinister trickle down effect at play here. As it turns out, we craftily generated this image using hand-picked specific instances of \textit{children} images emanating from what we have shown are two problematic \textit{seed} image classes: \texttt{Bikini} and \texttt{Brassiere}. More specifically, for this particular image, we set the \textit{Gene weights} to be: [\textit{Bikini}: 42.35, \textit{Brassiere}: 31.66, \textit{Comic Book} - 84.84 ]. We would like to strongly emphasize at this juncture that the problem does not emanate from a visual patriarchal mindset \cite{baran2017visual}, whereby we associate female undergarment imagery to be somehow unethical, but the root cause lies in the fact that many of the images curated into the dataset (at least with regards to the 2 above mentioned classes) were voyeuristic, pornographic, non-consensual and also entailed underage nudity.
\begin{figure*}[ht!]
\centering
\includegraphics[width=0.975\textwidth]{plots/census/3a_AGN_3d_scatter_flat.pdf}
\caption{Class-wise cross-categorical scatter-plots across the age, gender and NSFW score estimates}
\label{fig:agn}
\end{figure*}
\begin{figure*}[ht!]
\centering
\subfloat[Cross-tabulated grid-plot of the co-occurrence of the imagenet classes with the hand-labelled categories]{
\includegraphics[width=0.4\textwidth]{plots/census/3c_crosstab_hand_survey.pdf}
\label{fig:hand_labelled_grid}
}
\subfloat[Histogram-plots of the hand-labelled images]{
\includegraphics[width=0.6\textwidth]{plots/census/3c_count_hand_survey.pdf}
\label{fig:hand_labelled_histo}
}
\caption{Plots showcasing the statistics of the hand-survey across the \texttt{beach-voyeur, exposed-private-parts, upskirt, verifiably-pornographic} image categories}
\label{fig:hand_labelled}
\end{figure*}
\begin{figure*}[ht!]
\centering
\captionsetup{justification=centering}
\subfloat[Categorized violin plot demonstrating the class-wise mean gender scores across the dog-breed image groups]{
\includegraphics[width=0.5\textwidth]{plots/census/4a_dog_breeds.pdf}
\label{fig:co_occur_gender_bias_dog}
}
\subfloat[Gender skewness scores across the different musical instrument image classes ]{
\includegraphics[width=0.5\textwidth]{plots/census/4b_music_skewness.pdf}
\label{fig:co_occur_gender_bias_instrument}
}
\caption{Plots showcasing the \textit{human co-occurrence} based gender-bias analysis}
\label{fig:co_occur_gender_bias}
\end{figure*}
\begin{table*}[ht]
\centering
\begin{adjustbox}{width=0.85\textwidth}
\small
\begin{tabular}{llrll}
\toprule
wordnet\_id & label & mean\_nsfw\_train & category & file\_names \\
\midrule
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_11383.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_12451.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_13794.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_14133.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_15158.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_15170.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_15864.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_17.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_17291.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_17410.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_18107.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_18124.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_18260.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_20096.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_22044.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_283.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_3414.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_3536.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_4.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_5713.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_9181.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & beach\_voyeur & n02837789\_9859.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & exposed\_private\_parts & n02837789\_17069.JPEG \\
n02837789 & bikini, two-piece & 0.859369 & exposed\_private\_parts & n02837789\_19619.JPEG \\
n02892767 & brassiere, bra, bandeau & 0.610233 & exposed\_private\_parts & n02892767\_19629.JPEG \\
n02892767 & brassiere, bra, bandeau & 0.610233 & exposed\_private\_parts & n02892767\_3235.JPEG \\
n02892767 & brassiere, bra, bandeau & 0.610233 & upskirt & n02892767\_17717.JPEG \\
n02892767 & brassiere, bra, bandeau & 0.610233 & verifiably\_pornographic & n02892767\_5914.JPEG \\
n03527444 & holster & 0.058000 & exposed\_private\_parts & n03527444\_12661.JPEG \\
n03617480 & kimono & 0.091925 & exposed\_private\_parts & n03617480\_6206.JPEG \\
n03710637 & maillot & 0.801976 & beach\_voyeur & ILSVRC2012\_val\_00021081.JPEG \\
n03710637 & maillot & 0.801976 & beach\_voyeur & n03710637\_15836.JPEG \\
n03710637 & maillot & 0.801976 & beach\_voyeur & n03710637\_272.JPEG \\
n03710637 & maillot & 0.801976 & beach\_voyeur & n03710637\_3832.JPEG \\
n03710637 & maillot & 0.801976 & beach\_voyeur & n03710637\_5095.JPEG \\
n03710637 & maillot & 0.801976 & beach\_voyeur & n03710637\_5373.JPEG \\
n03710637 & maillot & 0.801976 & beach\_voyeur & n03710637\_5386.JPEG \\
n03710637 & maillot & 0.801976 & beach\_voyeur & n03710637\_66.JPEG \\
n03710637 & maillot & 0.801976 & beach\_voyeur & n03710637\_7074.JPEG \\
n03710637 & maillot & 0.801976 & exposed\_private\_parts & n03710637\_6756.JPEG \\
n03710721 & maillot, tank suit & 0.768278 & beach\_voyeur & n03710721\_1812.JPEG \\
n03710721 & maillot, tank suit & 0.768278 & beach\_voyeur & n03710721\_3040.JPEG \\
n03710721 & maillot, tank suit & 0.768278 & beach\_voyeur & n03710721\_3488.JPEG \\
n03710721 & maillot, tank suit & 0.768278 & beach\_voyeur & n03710721\_7542.JPEG \\
n03710721 & maillot, tank suit & 0.768278 & beach\_voyeur & n03710721\_8122.JPEG \\
n03770439 & miniskirt, mini & 0.619425 & upskirt & n03770439\_10283.JPEG \\
n03770439 & miniskirt, mini & 0.619425 & upskirt & n03770439\_18237.JPEG \\
n03770439 & miniskirt, mini & 0.619425 & upskirt & n03770439\_2462.JPEG \\
n03770439 & miniskirt, mini & 0.619425 & upskirt & n03770439\_2920.JPEG \\
n03770439 & miniskirt, mini & 0.619425 & upskirt & n03770439\_3615.JPEG \\
n03770439 & miniskirt, mini & 0.619425 & upskirt & n03770439\_4096.JPEG \\
n03770439 & miniskirt, mini & 0.619425 & upskirt & n03770439\_4203.JPEG \\
n03770439 & miniskirt, mini & 0.619425 & upskirt & n03770439\_6214.JPEG \\
n03770439 & miniskirt, mini & 0.619425 & upskirt & n03770439\_8550.JPEG \\
n03770439 & miniskirt, mini & 0.619425 & upskirt & n03770439\_9676.JPEG \\
n03770439 & miniskirt, mini & 0.619425 & verifiably\_pornographic & n03770439\_12003.JPEG \\
n03770439 & miniskirt, mini & 0.619425 & verifiably\_pornographic & n03770439\_1347.JPEG \\
n04209133 & shower cap & 0.130216 & exposed\_private\_parts & n04209133\_10606.JPEG \\
n04209133 & shower cap & 0.130216 & exposed\_private\_parts & n04209133\_206.JPEG \\
n04209133 & shower cap & 0.130216 & exposed\_private\_parts & n04209133\_716.JPEG \\
\bottomrule
\end{tabular}
\end{adjustbox}
\caption{ Table containing the results of hand surveyed images}
\label{tab:hand_survey}
\end{table*}
\begin{figure}[ht!]
\centering
\includegraphics[width=3in]{plots/gan_breeder.jpeg}
\caption{An example neural art image generated by the authors using the \textit{ArtBreeder} app [Gene weights: Bikini: 42.35, Brassiere: 31.66, Comic Book - 84.84 ]}
\label{fig:example}
\end{figure}
\begin{table*}[]
\begin{tabular}{|l|l|}
\hline
Reviewer-ID & Review \\ \hline
A- Grad student, CMU SCS & \begin{tabular}[c]{@{}l@{}}This one reminds me of a mix between graffiti and paper mache using \\ newspaper with color images or magazines . My attention is immediately \\ drawn to near the top of the image which, at first glance, appears to be a \\ red halo of sorts, but upon further consideration, looks to be long black \\ branching horns on a glowing red background. \\ My attention then went to the center top portion, where the "horns" were \\ coming from, which appeared to be the head or skull of a moose or \\ something similar. The body of the creature appears to be of human-like\\ form in a crucifix position, of sorts. The image appears more and more \\ chaotic the further down one looks.\end{tabular} \\ \hline
B- Grad student, Stanford CS & \begin{tabular}[c]{@{}l@{}}Antisymmetric: left side is very artistic, rich in flavor and shades; \\ right is more monotonic but has more texture. \\ Reminds me of the two different sides of the brain through the anti-symmetry\end{tabular} \\ \hline
C- Data Scientist, Facebook Inc & Futurism \\ \hline
D- CS undergrad, U-Michigan & \begin{tabular}[c]{@{}l@{}}It’s visually confusing in the sense that I couldn’t tell if I was \\ looking at a 3D object with a colorful background or a painting. \\ It’s not just abstract, but also mysteriously detailed\\ in areas to the point that I doubt that a human created these\end{tabular} \\ \hline
E - Senior software engineer, Mt View & \begin{tabular}[c]{@{}l@{}}The symmetry implies a sort of intentionally. \\ I get a sense of Picasso mixed with Frieda Callo[sic] here.\end{tabular} \\ \hline
F- Data Scientist, SF & \begin{tabular}[c]{@{}l@{}}Reminds me of a bee and very colorful flowers, but with some \\ nightmarish masks hidden in some places. Very tropical\end{tabular} \\ \hline
\end{tabular}
\caption{Responses received for the neural art image in Fig \ref{fig:example}}
\label{tab:responses}
\end{table*}
\subsection{Error analysis}
\begin{figure*}[ht!]
\centering
\includegraphics[width=1\textwidth,]{Supplement/acc_ttests.pdf}
\caption{On accuracy variations and \textit{human delta}}
\label{fig:acc_human_delta}
\end{figure*}
Given how besotted the computer vision community is with regards to classification accuracy metrics, we decided to indulge in \textit{devil's advocacy} by delving into the nature of variation of class-wise top-5 accuracies in those classes where humans co-occur asymmetrically between the training and validation sets. For this, we performed inference using the ResNet50 \cite{he2016deep:ResNet} and NasNet \cite{zoph2018:NasNet} models and sorted all the 1000 classes as per the $N_{train}^{persons}/N_{val}^{persons}$ ratios (termed \textit{human-delta} in the figure) and compared their accuracies with regards to the general population (amongst the 1000 classes). As gathered from Figure \ref{fig:acc_human_delta}, we saw a statistically significant drop in top-5 accuracies ($T-test \in (-3.87,-3.06)$) for the \textit{top-25 human-delta classes}, thereby motivating that even for the purveyors of \textit{scientism} fuelled pragmatism, there is motivation here to pay heed to the problem of humans in images. We would like to reemphasize that we are most certainly not advocating this to be the \textit{prima causa} for instigating a cultural change in the computer vision community, but are sharing these resources and nuances for further investigation.
\section{Broader impact statement and a wish list}
\label{appendix:impact_statement}
We embarked on this project with an aspiration to illustrate how problematic large scale image dataset curations are both in academia and industry and the need for a fundamental change. Through the course of this work, we solicited and incorporated feedback from scholars in the field who have pointed us towards three valid critiques that we would like to address first. To begin with, we solemnly acknowledge the moral paradox in our use of pre-trained gender classification models for auditing the dataset and duly address this in the previous section. Secondly, as covered in Section \ref{sec: threats} on the \textit{threat landscape}, we also considered the risks of the possible \textit{Streissand effect} with regards to deanonymization of the persons in the dataset that ultimately lead us to not dive further into the quantitative or qualitative aspects of our findings in this regard, besides conveying a specific example via email to the curator of the dataset from which the deanonymization arose. Thirdly, we would like to acknowledge the continued efforts of ImageNet curators to improve the dataset. Although there remains much work to be done, in the large scheme of things and compared to secretive and opaque datasets, the ImageNet dataset allows examinations. Having said that, curating large datasets comes with responsibility (especially given such dataset directly or indirectly impact individual lives and the social world) and all curators need to be held accountable for what they create. With these caveats firmly in tow, we now proceed to conclude with the following Wish List of the impact we hope this work may bring about.
\subsection{Proactive approach over reactive course corrections}
We aspire to see the institutions and individulas curating these large scale datasets to be proactive in establishing the primacy of ethics in the dataset curation process and not just reacting to exposes and pursing posthoc course corrections as an afterthought. We would be well served to remind ourselves that it took the community 11 years to go from the first peer-reviewed dissemination \cite{deng2009imagenet} of the imagenet dataset to achieving the first meaningful course correction in \cite{yang2020towards_facct} whereas the \textit{number of floating-point operations required to train a classifier to AlexNet-level performance on ImageNet had decreased by a factor of 44x between 2012 and 2019} \cite{openai_moores}. This, we believe, demonstrates where the priorities lie and this is precisely where we seek to see the most impact.
\subsection{Bluewashing of AI ethics and revisiting the enterprise of Big data}
At the outset, we question if \textit{Big Data} can ever operate in a manner that caters the needs and welfares of marginalized communities - those disproportionately impacted by algorithmic injustice. Automated large scale data harvesting forays, by their very volition, tend to be \textit{BIG}, in the sense that they are inherently prone to \textbf{B}ias, are \textbf{I}mperceptive to the lessons of human condition and recorded history of vulnerable people and \textbf{G}uileful to exploit the loopholes of legal frameworks that allow siphoning off of lived experiences of disfranchised individuals who have little to no agency and recourse to contend Big Data practices. Both collective silence and empty lip service \footnote{\url{https://www.media.mit.edu/articles/beware-corporate-machinewashing-of-ai/}}, i.e. caricatured appropriations of ethical transgressions entailing \textit{ethics shopping}, \textit{ethics bluewashing}, \textit{ethics lobbying}, \textit{ethics dumping} and \textit{ethics shirking} \cite{floridi2019translating} cause harm and damage. Given that these datasets emerged from institutions such as Google, Stanford, NYU and MIT, all with a substantial number of staff researching AI ethics and policy, we cannot help but feel that this hints towards not just compartmentalization and fetishization of ethics as a \textit{hot topic} but also shrewd usage of the ethicists as agents of \textit{activism outsourcing}.
\subsection{Arresting the creative commons loot}
As covered in the main paper, we could like to see this trend of using the creative commons loophole as an excuse for circumventing the difficult terrain of informed consent. We should, as a field, aspire to treat consent in the same rigorous way as researchers and practitioners in fields such as anthropological studies or medical studies.
In this work, we have sought to draw the attention of the Machine Learning community towards the societal and ethical implications of large scale datasets, such as the problem of non-consensual images and the oft-hidden problems of categorizing people. We were inspired by the adage of \textit{Secrecy begets tyranny}\footnote{From Robert A. Heinlein's 1961 science fiction novel titled \textit{Stranger in a Strange Land \cite{heinlein2014stranger}}} and wanted to issue this as a call to the Machine Learning community to pay close attention to the direct and indirect impact of our work on society, especially on vulnerable groups. We hope this work contributes to raising awareness and adds to a continued discussion of ethics in Machine Learning, along with many other scholars that have been elucidating algorithmic bias, injustice, and harm.
\end{appendices}
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
One of the key ideas of many disciplines including materials science
in particular is that structure and properties of materials tend to be
related. \cite{cheng2011atomic,delince2007structure}
For instance, much of metallurgy is about tuning materials
microstructure to design materials with desired mechanical properties.
\cite{delince2007structure} Typically the relation between, say, the precipitate content of a
precipitation-hardened alloy \cite{ardell1985precipitation}
and its yield strength is understood
as an average property, and indeed in the case of macroscopic
bulk samples the sample-to-sample variations in the yield stress of
identically prepared samples tend to be small. However, this changes when
dealing with samples with sizes down to the micrometer range and below:
Recent micropillar compression experiments have revealed the fluctuating,
irregular nature of small-scale crystal plasticity, \cite{uchic2004sample,
dimiduk2006scale} originating from
the collective, critical-like dynamics of interacting dislocations mediating
the deformation process. The fact that plasticity of
small crystals proceeds via an irregular sequence of strain bursts with
a broad size distribution implies also that sample-to-sample variations
of the plastic response or the stress--strain curves might be considerable, \cite{dimiduk2005size}
even if the samples have been prepared and the deformation experiments
performed using the same protocol.
This raises the important general question of what can be said
about the relation between initial structure and mechanical properties for individual samples of small crystals,
when random variations in the initial
(micro)structure become important. One way of framing the issue is to
consider what we refer to as {\it deformation predictability}: Given a small
crystalline sample with a specific arrangement of pre-existing dislocations
and possibly of some other defects interfering with dislocation motion,
how well can one use such information to predict the plastic response of
that sample? A key challenge here is the high dimensionality of any reasonable description of the disordered initial state, combined with the possibility of non-linearities in the mapping from the initial state properties to the ensuing response. This likely implies that simple empirical laws relating sample properties to its plastic response may be difficult to formulate. Challenges of this type are a key factor behind the recent emergence of machine learning (ML) algorithms as an important part of the toolbox of scientists in a wide range of fields including also physics and materials science. \cite{zdeborova2017machine} ML has been demonstrated to be an efficient approach to address a wide spectrum of problems including materials property prediction, discovery of novel materials, etc. \cite{papanikolaou2019spatial,
raissi2018hidden,baldi2014searching,butler2018machine,raccuglia2016machine,pilania2013accelerating,liu2020material} These developments have lead to the emergence of a novel research field referred to as {\it materials informatics}, \cite{ramprasad2017machine} where informatics methods -- including ML -- are used to determine material properties that are hard to measure or compute using traditional methods.
In general, many supervised ML algorithms including neural networks are capable of learning non-linear mappings from a high-dimensional feature vector to a desired output, and hence these methods should be useful tools when addressing the question of deformation predictability. An important opening in this direction was recently achieved by Salmenjoki {\it et al.},
\cite{salmenjoki2018machine} who applied ML to study
deformation predictability in the case of simple two-dimensional discrete
dislocation dynamics (DDD) simulations. The key idea of Ref.
\cite{salmenjoki2018machine} was that exploiting ML algorithms provides a
useful set of methods to quantify predictability of complex systems such as plastically deforming crystals. This was achieved by training ML models including neural networks and support vector machines to infer the mapping from the initial dislocation microstructure to the response of the system to applied stresses, characterized by the stress--strain curve. Following Ref. \cite{salmenjoki2018machine}, ML has been further applied to predicting stress--strain curves of plastically deforming crystals, \cite{yang2020learning} and learning the interaction kernel of dislocations. \cite{salmenjoki2018mimicking} Recently ML has also been applied to the closely related problems of, e.g., predicting the local yielding dynamics of dry foams, \cite{viitanen2020machine} the creep failure time of disordered materials, \cite{biswas2020prediction} as well as the occurrence times of "laboratory earthquakes". \cite{rouet2017machine}
By its nature, the problem of predicting the plastic deformation process of
crystalline samples depends on details such as whether the crystal only contains pre-existing glissile dislocations
(this was the case in Ref. \cite{salmenjoki2018machine}), or if other
defect populations interacting with the dislocations are present. The latter
could include solute atoms, precipitates, voids, etc., or even grain
boundaries in the case of polycrystals. If present, a description of
these additional defects needs to be included in the
initial state of the system used as input for
the predictive ML models. The presence of such static defects within
the crystal may also change the response
of the crystal to applied stresses and thus the process to be predicted in a fundamental manner. It has been shown that the deformation dynamics of ``pure'' dislocation systems are governed by dislocation jamming, \cite{miguel2002dislocation,laurson2010dynamical} resulting in glassy dislocation dynamics characterized by ``extended criticality'', with
the cutoff scale of the size distribution of dislocation avalanches diverging with the system size at any stress level. \cite{ispanovity2014avalanches,lehtinen2016glassy,ovaska2017excitation} This can be contrasted with systems where significant
pinning of dislocations due to other defects within the crystal may instead
induce a depinning phase transition of the dislocation assembly, resulting
in critical-like dislocation dynamics only in the immediate proximity of the critical point (stress) of the
depinning transition. \cite{ovaska2015quenched,salmenjoki2020precipitate}
In this paper, we study deformation predictability in a one-dimensional
periodic model of a dislocation pileup, interacting with a quenched (frozen) random
pinning landscape. The model is perhaps the simplest possible system including
both interacting mobile dislocations and a quenched random pinning field interfering with
dislocation motion, and therefore it serves
as a useful playground to explore the ideas of ML-based deformation
predictability discussed above. It is known to exhibit a depinning phase transition at a critical flow stress value $\sigma_\textrm{flow}$, separating pinned and moving phases of the dislocation system. \cite{moretti2004depinning} We generate a large database of
stress--strain curves, each corresponding to a unique randomly generated
initial microstructure, by simulating the model with a quasistatically
increasing applied stress for different system sizes (or different numbers of dislocations, $N$). First, the statistical properties
of the stress--strain curves and the strain bursts are analyzed, followed by training and testing of various predictive models ranging from linear regression to convolutional neural networks (CNNs) to establish mappings from the initial random microstructure (defined by both the static pinning landscape and the initial configuration of the dislocations) to the response of the system to applied stresses (i.e., the stress--strain curve).
Our results show that the different predictive models employed
are capable of learning the relation between the input and the stress--strain curves quite well, as measured by the correlation coefficient between the predicted and simulated stress value at a given strain. In particular, the sample flow stress (i.e., the sample-dependent finite size critical point of the depinning transition) can be predicted surprisingly well, with the correlation coefficient reaching values as high as 0.89 for a
regularized CNN. We also explore the predictability of individual strain bursts taking place during the deformation process. Critical avalanches are expected to be inherently unpredictable, and indeed we find approximately zero predictability for most of the strain bursts belonging to
the power law scaling regime of their size distribution. Interestingly, we also find that strain bursts belonging to the cutoffs of the distributions as well as those taking place very close to the flow stress exhibit non-vanishing predictability.
\section{The model: edge dislocation pileup with quenched disorder}
The dislocation pileup model we study here is similar to the one in Refs. \cite{moretti2004depinning,
leoni2009slip}. It describes a system of $N$ straight, parallel edge
dislocations subject to an external shear stress $\sigma$ gliding along the
direction set by their Burgers vector (here, the $x$ direction) within
a given glide plane. By neglecting roughness of the dislocation lines,
the system may be described by a set of point dislocations with coordinates
$x_i$ representing cross sections of the dislocation lines moving in one
dimension. The dislocations interact repulsively with each other via
long-range stress fields with the stress field magnitude inversely
proportional to their mutual distance. \cite{moretti2004depinning}
In addition, the dislocations
are taken to interact via short-range forces with a set of $N_p$ randomly
positioned Gaussian pinning centres, playing here the role of quenched
disorder. In real crystals these pinning centres could be solute atoms, or immobile forest dislocations threading the plane of the pileup. The overdamped equations of motion of the $N$ point dislocations thus read
\begin{equation}
\chi \frac{\mathrm{d}x_i}{\mathrm{d}t} = \mu b^2 \sum_{j \neq i}
\frac{1}{x_i-x_j} + b\sigma + F(x_i),
\end{equation}
where $\chi$ is the effective viscosity, $x_i$ the position of the
$i$th dislocation, $\mu$ the shear modulus, $b$ the Burgers vector
magnitude, and $F$ is the pinning force landscape. We choose a pinning force consisting of Gaussian potential wells centered at pinning sites $\set{x_p}$, given by
\begin{equation}
\label{eq:pf}
F(x) = -\od{E}{x},\quad
E(x) = -\sum_{p=1}^{N_p} E_p\mathrm{e}^{-\frac{1}{2}\left(\frac{x-x_p}{s_p}\right)^2},
\end{equation}
where $x_p$ is the position of the $p$th pinning point, $E_p$ a pinning
energy scale, and $s_p$ the standard deviation of the Gaussian. For simplicity, we set
$\chi=\mu=b=1$, and $E_p = s_p^2 = 0.5$. Periodic
boundary conditions within a system of size $L$ are employed by performing
an infinite sum over the stress fields of the periodic images of each
dislocation, \cite{moretti2004depinning}
\begin{equation}
\sum_{k=-\infty}^{\infty} \frac{1}{x+kL} = \frac{\pi}{L\tan (\pi x/L)}.
\end{equation}
The average spacing between dislocations is set to $L/N=16$ and that between
pinning sites to $L/N_p=2$. The sum over pinning sites is truncated so that
only pinning sites that are closer than a cut-off distance of 8 are
included in the sum.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\columnwidth]{Fig1_v2.pdf}
\caption{(a) An example of a relaxed dislocation configuration in the
simulation model with $N=4$ (and hence of linear size $L=16N=64$). The
randomly positioned pinning sites (there are $N_p = L/2 = 32$ of them)
are shown as circles, and the resulting pinning energy landscape $E(x)$
is shown with the solid line. (b) Example stress--strain curves
$\sigma (\epsilon)$ for three different system sizes ($N=4$, 16 and 64),
illustrating the characteristic irregular staircase shape of the curves, as
well as the size effect where smaller systems exhibit fewer but larger strain
bursts.}
\label{fig:1}
\end{center}
\end{figure}
The simulations are carried out by first choosing the positions of the $N_p$ pinning sites at random from a uniform distribution
along the one-dimensional system of length $L$. $N$ dislocations
are then placed in the system with the constant spacing of $L/N=16$; this equally
spaced dislocation configuration would be the minimum energy state of the
pileup in the absence of pinning sites. The dislocations are initially allowed to
relax to a metastable configuration with $\sigma = 0$; see Fig.~\ref{fig:1}(a) for an example of a relaxed
configuration with $N=4$, showing also the pinning energy landscape. Then
$\sigma$ is increased quasistatically from zero with the rate
\begin{equation}
\frac{\mathrm{d}\sigma}{\mathrm{d}t} = \frac{m_{\sigma}}
{\mathrm{e}^{\frac{r[\bar{v}(t)-v_{\mathrm{th}}]}{v_{\mathrm{th}}}}+1},
\end{equation}
where $m_{\sigma} = 10^{-4}$ is the maximum stress rate,
$v_{\mathrm{th}}=2\cdot 10^{-4}$ is a threshold velocity value,
$\bar{v}(t)=(1/N)\sum_{i=1}^N v_i$ (with $v_i = \mathrm{d}x_i/\mathrm{d}t$)
is the spatially averaged dislocation velocity, and $r=100$ is a shape
parameter. In practice this amounts to a continuous, smooth approximation of
the step function, with $\mathrm{d}\sigma/\mathrm{d}t \approx m_{\sigma}$ for
$\bar{v}<v_{\mathrm{th}}$ and $\mathrm{d}\sigma/\mathrm{d}t \approx 0$ for
$\bar{v}>v_{\mathrm{th}}$, such that $\sigma$ is increased only in between
strain bursts and kept constant during them. Strain
$\epsilon$ is accumulated by dislocation motion,
\begin{equation}
\epsilon(t) = \frac{1}{N}\sum_{i=1}^{N}[x_i(t)-x_i(0)].
\end{equation}
The simulations are run until the dislocations start to flow indefinitely,
i.e., the strain burst size diverges, which happens at a sample-dependent
flow stress $\sigma_\textrm{flow}$. The stress--strain curves $\sigma(\epsilon)$ are recorded by storing
$\sigma(t)$ and $\epsilon(t)$ every $20$ dimensionless time unit during the simulations. Example stress--strain curves are shown in Fig.~\ref{fig:1}(b) for
three different system sizes. Notice the characteristic
irregular staircase-like structure of the curves, as well as the larger
size of strain bursts for smaller systems.
The resulting stress--strain curves, each corresponding
to a different realization of the random pinning landscape and relaxed
initial dislocation positions, are stored in a database consisting of $10\,000$ stress--strain curves for each system size, to be used as training and testing data for the predictive models. For the largest systems ($N = 64\ldots 512$), we also generate a separate dataset of $1\,000$ stress--strain curves stored at a finer time resolution of $\Delta t = 2$, in order to better detect individual strain bursts. This latter data set is used to analyze the statistical properties of the strain bursts, as well as to train and test ML models for predicting strain bursts.
\section{Predictive models: from linear regression to convolutional neural networks}
In what follows we will introduce the predictive models used in this study.
These encompass both linear and non-linear models using hand-picked features as input, as well as a CNN
that requires less feature engineering as it uses a complete description of the system's initial state as input.
Using these different models allows us to assess the dependence of our results on any particular model.
Before training, each input feature (each input channel in the CNN case) is standardized by subtracting the mean and dividing by the standard deviation, and the dataset is divided randomly into a training set containing 80\% of the data points, and a testing set containing the remaining 20\%.
Predictability is measured by calculating the correlation coefficient between predicted and target outputs of the testing set and averaging the result over 5 training instances. Predictive models are implemented using the Python libraries \textit{scikit-learn} \cite{scikit-learn} for LASSO and \textit{Keras} \cite{keras} for simple neural networks and CNNs.
The prediction models work by minimizing the \emph{loss}, which mainly consists of the mean squared error. In addition, most models are $L^1$-regularized, which adds a penalty term to the loss to prevent overlearning. $L^1$ regularization often improves the result and encourages the model to focus on the most useful/promising input features while discarding the unimportant, and as a byproduct it allows us to determine which input features are used by the model.
\subsection{Linear regression: LASSO}
A linear regression model performs an affine transformation, where input features are first linearly combined by multiplication with a (weight) matrix and then biased by adding a translation vector. If the model is optimized using $L^1$-regularization for the weight matrix, it is commonly called LASSO (linear absolute shrinkage and selection operator \cite{tibshirani1996regression}). In this study, this model is applied by using \verb|sklearn.linear_model.Lasso|, which has a built-in optimizer and only requires the user to choose the regularization parameter $\alpha$. We found that values of $\alpha$ in the range $10^{-3} \lesssim \alpha \lesssim 10^{-2}$ gave good results, and chose $\alpha = 10^{-2}$.
\subsection{Simple neural networks}
When two linear regression models are stacked in such a way that the intermediate result is \textit{activated} by applying a non-linear activation function, such as the \textit{rectified linear unit (ReLU)}, the combined model becomes a simple neural network which can, to some extent, learn general non-linear mappings. An intermediate result of such model is called a \textit{hidden layer}, and elements of the intermediate result are referred to as \textit{hidden units}. $L^1$-regularization is applied to each layer, with a parameter $\alpha_n$ describing the strength of regularization at each layer.
Tested variants for stress--strain curve prediction differ by the number of hidden units, hidden layers and regularization. Most have one hidden layer with 64 hidden units and use
$\alpha_1 = 10^{-2}$ for the first weight matrix and $\alpha_n = 10^{-5}$ for the rest. All models use ReLU-activations for the hidden layers. One variant has less regularization, meaning that $\alpha_1 = 10^{-3}$ instead of $10^{-2}$. Theoretically, one hidden layer is enough for reproducing any continuous mapping given enough hidden units, but a model with 3 hidden layers is employed here as well for comparison to other studies. The models are trained using the \textit{Adam} optimizer \cite{kingma2014adam} with learning rate $10^{-3}$ for 100 epochs.
The predictability of strain bursts is studied using a simple neural network having one hidden layer with 64 hidden units. Model output is in this case very big, consisting of a vectorized (one-dimensional) version of an avalanche map that is originally two-dimensional (see the Results section below for details). We found that all parameters collapsed to 0 if $\alpha$ is too high, so a value of $10^{-5}$ is used for all weight matrices. A lower $\alpha$ parameter makes the model susceptible to overlearning. Therefore, models are trained only for 10 epochs but with a higher learning rate of $10^{-2}$.
\subsubsection{Selecting features for prediction}
The LASSO and simple neural network models require hand-picked features for predicting stress--strain curves and strain bursts. It turns out that
good, information-rich features for flow stress prediction are given by
quantiles of distributions derived from the sample-specific pinning landscapes. Fig.~\ref{fig:4}(a) shows the correlation of the flow stress with
quantiles of the pinning energy $E$, the pinning force $F$, as well as $DF = \od{F}{x}$ and $D^2F = \od[2]{F}{x}$.
In a non-interacting system, the flow stress would be controlled by the most negative pinning force $F$, which provides the strongest obstacle for dislocation motion. In an interacting system, it is possible for the dislocations to push on each other and overcome the strongest pinning force at a smaller applied stress. As a result, the correlation between the most negative value of $F$ [given by the $0\%$ quantile of $F$ in Fig.~\ref{fig:4}(a)] and the flow stress is quite small.
Interestingly, the 5\% quantile of $F$ exhibits a much larger negative correlation of approximately $-0.8$ with $\sigma_{\textrm{flow}}$, independently of system size [Fig.~\ref{fig:4}(b)]. Other quantiles of the pinning force as well as of the other quantities listed
in Fig.~\ref{fig:4}(a) also appear to contain useful information for flow stress
prediction. We therefore give the entire quantile curve (sampled on 64 uniformly chosen points) of each field as input features.
In addition to these input features describing the pinning landscape, we also give input features that describe the initial relaxed configuration of dislocations $\{x_j\}_{j=1}^N$. In particular, we use the values $E(x_j), F(x_j), DF(x_j)$ and $D^2F(x_j)$, as well as the distances between successive dislocations, $Dx_j = x_{j+1} - x_j$ (with $x_{N+1} = x_1 + L$). These fields are given as sorted lists, so they can be interpreted as quantiles.
Finally, we choose a set of features given by local extrema of $E$ and $F$, as well as the successive differences in energy extrema $D(E \text{ extrema})$, also given as 64 uniformly spaced quantiles.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\columnwidth]{Fig4_v3.pdf}
\caption{(a) Correlation between flow stress and quantiles of pinning potential energy $E$, pinning force $F$, derivative of the pinning force $D(F)$, and the second derivative of the pinning force $D^2(F)$ for a system of $N=64$ dislocations. Dotted lines show the absolute value of the correlation for easier comparison of magnitudes. (b) demonstrates that the correlations of the pinning force quantiles are not dependent on the system size $N$ (legend). These quantiles are used as hand-picked features for linear regression and simple neural networks. Notice that the 5\% quantile of the pinning force exhibits a high (negative) correlation of $\approx 0.8$ with the flow stress.}
\label{fig:4}
\end{center}
\end{figure}
\subsection{Convolutional neural networks}
These models take discretized fields as input and have the ability to find out if the target output depends on the spatial variations within the input fields. The CNN structure used in this study is adapted for periodic input signals by applying a convolution-activation-pooling scheme that continues to downsample the input array down to a single spatial element. Initially, an input field has 1\,024 = $2^{10}$ spatial elements, so there are 10 convolutional layers (each layer halves the number of spatial elements). This makes the result effectively independent of the particular choice of origin of the input fields.
All convolutional layers employed here use a window size of 3, periodic boundary conditions, ReLU-activation and non-overlapping max-pooling of size 2 (taking the maximum value within non-overlapping bins), and have 16 hidden units. The leftover hidden units after the last pooling operation are directly connected to the output layer. When predicting a single output feature from 2 input fields (channels), the CNN has 7\,185 learnable parameters.
The CNN is trained using the same \textit{Adam} optimizer as for other neural network models, with a learning rate of $10^{-3}$, learning rate decay of $10^{-5}$ and 50 training epochs. We found that even weak $L^1$-regularization caused the parameters to collapse to 0 in most cases, so we do not regularize these models. However, in the case of predicting flow stress only, a kernel regularization value of $5\cdot 10^{-4}$ for the convolutional layers is found to work.
The input to the CNNs consists of the pinning landscape fields $E, F, DF$ and $D^2F$ (although we found that giving one pinning field, such as $E$ or $F$, suffices because the CNN can compute the higher derivatives by differentiation), as well as a field $J$ defined as a sum of Gaussians centered on the relaxed dislocation positions $x_i$,
\begin{equation}
J(x) = \sum_{i=1}^N \frac{1}{\sqrt{2\pi}}\mathrm{e}^{-\frac{(x-x_i)^2}{2}}.
\end{equation}
This field $J$ serves to describe the dislocation positions in terms of an input field with similar structure as the pinning landscape, and which is independent of the spatial resolution [see Fig.~\ref{fig:7}(a) for examples of these input fields]. Notice that by contrast to the case of hand-picked features, these fields are given in order rather than sorted.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\columnwidth]{Fig3_v2.pdf}
\caption{(a), (b) and (c) show the bivariate histograms of the stress--strain curves for different system sizes [$N=16$ in (a), $N=64$ in (b), and $N=256$ in (c)]. (d) displays the probability distributions of the sample flow stresses $\sigma_{\mathrm{flow}}$, showing how the distributions
get increasingly narrow upon increasing the system size.}
\label{fig:2}
\end{center}
\end{figure}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\columnwidth]{Fig2_v5.pdf}
\caption{Left panels: Bivariate avalanche histograms (avalanche counts) as a function of the stress $\sigma$ and avalanche size $\Delta \epsilon$. Right
panels: The stress-resolved probability distributions of the strain bursts sizes $P(\Delta \epsilon,\sigma)$, considering three stress bins: 0.59-0.60 (blue), 0.63-0.64 (red), and 0.67-0.68 (yellow). The average flow stress is around 0.69. (a) $N=64$, (b) $N=128$, (c) $N=256$ and (d) $N=512$. The right panel of (d) shows that the highest stress bin data in the largest system size exhibits a power-law part consistent with an exponent $\tau \approx 1.0$ [see Eq. (\ref{eq:ps})].}
\label{fig:3}
\end{center}
\end{figure}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\columnwidth]{Fig5_v3.pdf}
\caption{Predictability of stress--strain curves for the system size $N=64$ using
hand-picked features,
measured by the correlation between predicted and actual stresses in the test set. The shown result is a sample mean over 5 trained models, each using
their own random division into training and testing sets. Each line
corresponds to a given predictive model as indicated in the caption. The plateau at large strains corresponds to the predictability of the flow stress.}
\label{fig:5}
\end{center}
\end{figure}
\iffalse
\begin{table}[t]
\centering
\begin{tabular}{ |p{6cm}||p{3cm}|p{1.5cm}|p{1.5cm}|p{1.5cm}|p{1.5cm}| }
\hline
Model & Operator & $L^1(M_1)$ & $L^1(M_2)$ & $L^1(M_3)$ & $L^1(M_4)$\\
\hline
\hline
Linear regression (LASSO) & $L=BM$ & $1\cdot 10^{-2}$ & - & - & -\\
Neural network (16 hidden units) & $LN=BM_2AM_1$ & $1\cdot 10^{-2}$ & $1\cdot 10^{-5} $& - & -\\
Neural network (64 hidden units) & $LN$ & $1\cdot 10^{-2}$ & $1\cdot 10^{-5}$ & - & -\\
Neural network (64 hidden units) v2 & $LN$ & $1\cdot 10^{-3}$ & $1\cdot 10^{-5}$ & - & -\\
Neural network ($3\times 64$ hidden units) & $LN_3N_2N_1$ & $1\cdot 10^{-2}$ & $1\cdot 10^{-5}$ & $1\cdot 10^{-5}$ & 1$\cdot 10^{-5}$\\
\hline
\end{tabular}
\caption{Parameters of the LASSO model and different simple neural networks used. The number of hidden units is the number of features (here elements, generally channels) in the intermediate result (change in dimensionality is a consequence of matrix operators $M_i$). The fifth model has 3 intermediate results (hidden layers) having 64 hidden units each. The second column shows the equivalent model operator using matrix formalism (see text). Columns 3-6 show the $L_1$ penalty parameter applied to each weight matrix of the models.}
\label{tab:1}
\end{table}
\fi
\section{Results}
In this section, we first consider the statistical properties of our dataset of
stress--strain curves. Then we report our results for training the predictive models, and employ them to explore the predictability of the deformation
dynamics of the pileup model. Finally, we also discuss our results on
the problem of predicting individual strain bursts.
\subsection{Statistical properties of the stress--strain curves}
We start by quantifying the system size dependent sample-to-sample variations
in the plastic response of the system. Figs.~\ref{fig:2}(a-c) show bivariate
histograms of the stress--strain curves for three system sizes (with $N=16$, 64
and 256, respectively), illustrating that the sample-to-sample variability
in the plastic response is more pronounced in smaller systems. The same
applies to the probability densities of the sample flow stresses
$\sigma_\textrm{flow}$, shown in Fig.~\ref{fig:2}(d): The distributions are
quite broad for small systems and get narrower upon increasing the system
size, presumably approaching a delta peak in the thermodynamic limit. The
average magnitude of the $\sigma_\textrm{flow}$-values is in agreement with
previous results. \cite{moretti2004depinning}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\columnwidth]{Fig5_5_v2.pdf}
\caption{Weight analysis of the trained LASSO model for the $N=64$ system. Color indicates the total $L^1$ penalty ($L^1$ norm) from each feature category’s weights as a function of strain, averaged over 5 training instances. Notice that properties of the initial dislocation positions are most useful at small strains, while more generic features of the force landscape become more useful at larger strains.
}
\label{fig:6}
\end{center}
\end{figure}
As the stress--strain curves consist of a sequence of discrete strain bursts
or dislocation avalanches $\Delta \epsilon$, we consider also their
stress-dependent size distributions $P (\Delta \epsilon,\sigma)$. Quite
generally for systems exhibiting a
depinning phase transition, one expects them to obey the scaling form
\begin{equation}
\label{eq:ps}
P(\Delta \epsilon,\sigma) = (\Delta \epsilon)^{-\tau} f\left(\frac{\Delta \epsilon}
{\Delta \epsilon^*(\sigma)}\right),
\end{equation}
where $\tau$ is the critical exponent of the avalanche size distribution,
$f(x)$ is a scaling function that obeys $f(x)=\mathrm{const.}$ for $x \ll 1$
and $f(x) \rightarrow 0$ for $x \gg 1$, and $\Delta \epsilon^*(\sigma)$ is a
stress-dependent cutoff avalanche size. Left panels of Figs.~\ref{fig:3}(a-d)
show the bivariate histograms of the avalanche count as a function of $\sigma$
and avalanche size $\Delta \epsilon$ for different system sizes (with
$N=64$, 128, 256 and 512, respectively). These illustrate that indeed the
maximum strain burst size increases upon approaching the flow stress from
below, and the increase close to the flow stress is more clear-cut for
larger systems. Notice also that these histograms show an excess of strain bursts around $\sigma \approx 0.4$. This feature is likely related to the characteristic pinning force magnitude due to an isolated pinning point given by Eq. (\ref{eq:pf}). Right panels of Figs.~\ref{fig:3}(a-d) show the corresponding
stress-resolved strain burst size distributions $P(\Delta \epsilon,\sigma)$,
considering three different stress bins below the flow stress for each system
size. For the largest system sizes considered, the data appears to be
consistent with a power-law part characterized by an exponent $\tau \approx 1.0$ [see right panels of Figs.~\ref{fig:3}(c) and (d)]. This value is somewhat lower than the expectation of $\tau \approx 1.25$ based on the scaling of the elastic energy of the pileup in Fourier space, \cite{moretti2004depinning} suggesting that the pileup model would be in the same universality class with depinning of contact lines \cite{joanny1984model} and planar cracks. \cite{bonamy2008crackling,laurson2010avalanches} As we focus here on studying rather small systems, this difference in likely due to finite size effects.
Overall, the statistical properties of the plastic response of the pileup
model shows two key features relevant for deformation predictability: (i)
There are sample-to-sample variations in the response (that we will try to predict using machine learning), and (ii) the distribution of strain bursts $\Delta \epsilon$
displays characteristics of a critical system, suggesting that individual
strain bursts may be inherently unpredictable. In what follows we will
use our trained predictive models to explore how well (i) the sample
stress--strain curves and (ii) individual strain bursts can be predicted.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\columnwidth]{Fig6_v5.pdf}
\caption{(a) Example input fields for a CNN with $N=16$. Dot markers show the spatial resolution. (b) Stress--strain curve predictability at $N=16$ for different models: LASSO (blue line); a "generalist" CNN predicting the entire stress--strain curve at once (red line); individual "specialist" CNNs trained specifically to predict the stress at a single strain (yellow circles); and an $L^1$-regularized CNN trained to only predict the flow stress (purple circle).}
\label{fig:7}
\end{center}
\end{figure}
\subsection{Predictability of the sample stress--strain curves}
\subsubsection{LASSO and simple neural networks}
Fig.~\ref{fig:5} shows the mean correlation coefficient between the predicted and actual stress values as a function of strain for $N=64$,
using linear regression (LASSO) and different variants of simple neural networks.
First, we note that the flow stress prediction is only slightly better than the magnitude of the correlation with the 5\% quantile of the pinning force landscape found above ($\sim 0.8$, see Fig. \ref{fig:4}), with the correlation coefficient assuming values in the range of 0.825-0.84 depending on the model used (the large-strain plateau in Fig.~\ref{fig:5}). Second, the correlation clearly depends on strain, exhibiting a rather pronounced dip reaching values as low as 0.65 or so for intermediate strains, along with two additional, lesser dips for smaller strains, before reaching a value close to 1 as the strain goes to zero. The latter is true especially for LASSO, while the neural network models seem to struggle a bit in capturing the deformation dynamics for very small strains. The reduced predictability for intermediate strains could be related to the importance of inherently unpredictable dislocation avalanches, in analogy with results of Ref. \cite{salmenjoki2018machine}.
It is interesting to analyze which features the predictive models end up using for prediction, and how that might depend on strain. Since the models utilize heavy $L^1$-regularization, analysis of the learned weights becomes straightforward. Features that contribute to predictability are allowed to have high regularization penalty in exchange for decreasing loss, while unimportant features will not decrease loss sufficiently to justify a high regularization penalty. Fig.~\ref{fig:6} shows how much penalty is generated by the weights corresponding to each feature category for the LASSO model. At the start, the pinning force derivative at the relaxed dislocation positions, $DF(x_j)$, is the most active feature, but also the pinning energy at the initial dislocation positions, $E(x_j)$, is used. This is reasonable as such quantities could be used to estimate the initial linear response of the dislocations. Interestingly, and perhaps somewhat counter-intuitively, the pinning energy at the initial dislocation positions becomes even more important for intermediate strains, in addition to a larger diversity of features used. Flow stress prediction uses mainly the quantiles of the pinning force $F$, as well as its local extrema points. This dependence of the important features on strain indicates that predicting the response of the system to infinitesimal stresses is of a different nature compared to predicting the flow stress.
\subsubsection{Predictability of stress--strain curves: convolutional neural networks}
We then proceed to employ CNNs as predictive models. Fig.~\ref{fig:7}(b) shows the mean correlation as a function of strain as obtained using three different types of CNNs, along with the LASSO result for reference. Here, the system size $N=16$ is considered.
First, we consider a CNN which predicts the entire stress--strain curve at once [a "generalist" CNN, shown as a red line in Fig.~\ref{fig:7}(b)]. This model actually underperforms the simple linear regression model for majority of the strain values. This could be due to the fundamentally different nature of the problem for different strain values, as illustrated by Fig.~\ref{fig:6}, showing that different features are important for different strains. Hence, the "generalist" CNN might get exhausted and not be able to optimally learn the entire stress--strain curve at once.
Thus, we consider also a "specialist" CNN which is trained separately for each strain value to learn the corresponding value of stress. The result is shown as yellow circles in Fig.~\ref{fig:7}(b). Indeed, as compared to the generalist model, the specialist model is able to learn the deformation dynamics of the pileup model much better, especially for intermediate strains, where the generalist CNN exhibits a deep dip in the correlation coefficient, presumably due to the importance of hard to predict dislocation avalanches for those strains. It is interesting that the specialist model is able to reach a correlation coefficient exceeding 0.8 also in this regime.
Notably, neither of the models discussed above is able to predict the flow stress much better than by just considering the correlation with the 5\% quantile of the pinning force $F$ (see Fig.~\ref{fig:4}). We therefore try to further improve the specialist CNN by switching on $L^1$-regularization and using only the pinning force $F$ as the input field (this was shown above to be the most informative quantity for flow stress prediction). The individual purple point at a large strain value in Fig. \ref{fig:7}(b) is the resulting flow stress prediction, with a correlation coefficient of 0.89, i.e., significantly better correlation than using any of the other models considered.
\subsection{Predictability of strain bursts}
Finally, we explore the possibility of predicting individual
strain bursts along the stress--strain curves. First, we need to formulate a problem that is suitable for analysis using the predictive models at hand. To this end, we generalize the bivariate avalanche histograms (left panels of Fig.~\ref{fig:3}) to the case of individual samples. As such, these would be rather sparse sets of points in the space spanned by avalanche size and stress. We therefore smooth out these maps by convolving with a Gaussian with a standard deviation of 0.03 units of stress in the vertical direction, and of 0.15 units of $\log_{10}(\text{avalanche size})$ in the horizontal direction; both correspond to 3 times the bin size used in Figs.~\ref{fig:3} and \ref{fig:8}. The studied region is limited to bins where at least one avalanche appears within the dataset, so that the Gaussian smoothing does not cause distribution boundaries to extend to impossible areas. Lastly, the maps are standardized by removing the mean and dividing by the standard deviation at each bin. We then employ a simple neural network with 64 hidden units to find a mapping from the hand-picked input features to the smoothed sample-specific bivariate avalanche histograms.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\columnwidth]{Fig9_v5.pdf}
\caption{Three examples (rows) of strain burst maps of individual systems for $N$=512. Left column (a) shows 3 target maps, and on the right (b) are the corresponding predictions of a simple neural network.}
\label{fig:8}
\end{center}
\end{figure}
Three different examples of smoothed avalanche maps are shown in the left column of Fig.~\ref{fig:8}, with corresponding predicted avalanche maps in the right column. While the predictions are clearly not perfect, the algorithm is able to capture some key features of these avalanche maps.
Fig.~\ref{fig:9} shows correlation maps between the predicted and actual avalanche densities at 4 different system sizes, averaged over 5 training instances. These show a number of interesting features: (i) Large parts of the sample-specific avalanche histograms appear to be completely unpredictable, with the correlation coefficient assuming a value very close to zero (shown in black in Fig.~\ref{fig:9}). This is especially true for most of the regions corresponding to the power-law scaling regime of the avalanche size distribution, in agreement with the idea that critical avalanches should be intrinsically hard to predict. \cite{pun2020prediction} (ii) Avalanches with the largest size for each stress value (i.e., those belonging to the cutoff of the stress-resolved avalanche size distribution) appear to exhibit some degree of predictability, although the related correlation coefficient values are somewhat lower than those found for predictability of the stress--strain curves. This is agreement with the idea that avalanches in the cutoff are not critical and hence somewhat predictable. (iii) Surprisingly, we also find non-vanishing predictability of avalanches of any size occurring in the immediate proximity of the flow stress (bright horizontal segments in the upper parts of the panels of Fig.~\ref{fig:9}). This is surprising because those events are expected to be critical and hence should be unpredictable.
In order to shed some light on the origin of the key features of avalanche predictability discussed above, we finish by considering the correlation between the avalanche count and flow stress. The logic here is that since we have demonstrated above that the sample-specific flow stress can be predicted quite well, if the avalanche maps are correlated with the flow stress, these should be predictable as well. Fig.~\ref{fig:10} reveals two stand-out features: (i) The avalanches taking place in the immediate proximity of the flow stress exhibit clear positive correlation with the flow stress, and (ii) the largest avalanches for stresses smaller than the flow stress (i.e., avalanches belonging to the stress-dependent cut-off of the avalanche size distribution) are negatively correlated with the flow stress. This means that both a {\it higher} than average number of avalanches very close to the flow stress and a {\it lower} than average number of large avalanches at smaller stresses imply a higher than average flow stress value. The good level of predictability of the flow stress demonstrated above thus translates into reasonable predictability of these features of the sample-specific bivariate avalanche histograms.
Notice that the widths of the predictable bands in the avalanche maps (Fig.~\ref{fig:9}) appear to become thinner with increasing system size, so we cannot exclude the possibility that this avalanche predictability would be a finite size effect.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\columnwidth]{Fig7_v2.pdf}
\caption{Correlation maps between the predicted and actual
sample-specific avalanche maps for different system sizes, (a) $N=64$, (b) $N=128$, (c) $N=256$ and (d) $N=512$. Notice the positive correlations for the largest avalanches for each stress value (avalanche cut-off) and for most of the avalanches very close to the flow stress, as well as the approximately zero correlation for avalanches belonging to the power law scaling regime (see Fig.~\ref{fig:3}).}
\label{fig:9}
\end{center}
\end{figure}
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\columnwidth]{Fig10_v5_alt.pdf}
\caption{Correlation maps between the avalanche count and flow stress for the avalanche maps considered in Fig.~\ref{fig:8}. (a) $N=64$, (b) $N=128$, (c) $N=256$ and (d) $N=512$. Notice the positive correlation between the flow stress and avalanches close to the flow stress, as well as the negative correlation between the flow stress and the largest avalanches occurring for smaller stresses.}
\label{fig:10}
\end{center}
\end{figure}
\section{Discussion and conclusions}
We have established that predictive models ranging from linear regression to CNNs can be used to predict the stress--strain curves of the pileup model with a good accuracy. While the different models employed give rise to somewhat different results, a practical conclusion is that the model giving the highest correlation coefficient for a given strain provides a lower limit for deformation predictability of the pileup system. Notably the "specialist" CNN discussed above results in correlation coefficients exceeding 0.8 over the entire range of strains considered, and adding regularization further improves the correlation coefficient of the flow stress prediction to 0.89. These should be interpreted as lower limits of deformation predictability as we obviously cannot exclude the possibility that a hypothetical predictive model not considered here might outperform our models.
While the predictability of the stress--strain curves turns out to be quite good, it is not perfect. This is to be expected because the stress--strain curves consist of critical-like strain bursts which are expected to be inherently hard to predict. Indeed, our attempts to predict the sample-level bivariate avalanche count resulted in essentially zero predictability for most of the "critical" dislocation avalanches belonging to the power-law distributed part of the avalanche size distribution. On the other hand, these sample-level avalanche distributions also showed an interesting regularity: A larger than average flow stress value indicates lower than average number of the largest avalanches for stresses below the flow stress, and larger than average number of avalanches in the immediate vicinity of the flow stress. It would be interesting to extend such analysis from our stress-controlled simulations to strain-controlled ones.
The pileup model studied here has the advantage that the dislocation system will sample the entire one-dimensional random energy landscape, and hence extracting relevant descriptors of the sample-specific quenched disorder is straightforward.
This is in contrast to other systems exhibiting a depinning phase transition such as magnetic domain walls in disordered ferromagnetic thin films, \cite{zapperi1998dynamics} planar crack fronts in disordered solids, \cite{bonamy2008crackling,laurson2010avalanches,laurson2013evolution} or individual dislocation lines interacting with point defects. \cite{zapperi2001depinning} In such systems, driven by an external
force in a direction perpendicular to the average elastic line direction, the disorder sampled by a given realization is not known {\it a priori}. Nevertheless, it would be interesting to extend the present study to consider the predictability of depinning dynamics in such systems. Finally, we point out two crucial outstanding questions related to deformation predictability: The problem remains to be addressed in realistic three-dimensional DDD simulations, with \cite{salmenjoki2020precipitate,lehtinen2018effects} or without \cite{lehtinen2016glassy} additional defects interfering with dislocation motion, as well as in experiments where descriptors of the initial sample microstructure could be extracted, e.g., by X-ray measurement techniques. \cite{ludwig2001three,levine2006x,schafler2005second}\\
\begin{acknowledgments}
The authors wish to thank Henri Salmenjoki for technical help on
machine learning. We acknowledge the financial support of the Academy
of Finland via the Academy Project COPLAST (Project no. 322405).
\end{acknowledgments}
\section*{Data availability}
The data that support the findings of this study are available from the corresponding author upon reasonable request.
\section*{References}
|
\section{Introduction}\label{ref1}
Given a finite set $\Sigma$ of square matrices of the same dimension and with real entries, the \newword{joint spectral radius} of $\Sigma$ is
\[
\tilde\rho(\Sigma)=\lim_{n\to\infty}\max\set{\norm{C}:C\in\Sigma^n},
\]
where $\Sigma^n$ is the set of all products of $n$ matrices from $\Sigma$, repetitions allowed, and $\norm{\,\operatorname{--}\,}$ is the operator norm induced from some vector norm, whose choice is irrelevant.
In short, $\tilde\rho(\Sigma)$ measures the maximal exponential growth rate of vectors under the action of $\Sigma$. Its range of applicability is large and still growing; we refer to~\cite{jungers09}, \cite{guglielmizennaro14} and references therein for a broad panorama and proofs of basic statements.
Despite its simple definition, the computation of the joint spectral radius is a notoriously difficult problem (indeed it is NP-hard~\cite{tsitsiklisblondel97}), even in the restricted form of just determining whether it is nonzero. By the Berger-Wang theorem we have the equivalent characterization
\begin{equation}\label{eq2}
\tilde\rho(\Sigma)=\sup_n\max\set{\rho(C)^{1/n}:C\in\Sigma^n},
\end{equation}
where $\rho(C)$ is the spectral radius of $C$, and in~\cite[p.~19]{lagariaswang95} Lagarias and Wang put forward the \newword{finiteness conjecture}, namely the possibility that the supremum in~\eqref{eq2} is always a maximum. Although in its full generality the conjecture was refuted in~\cite{bouschmairesse02}, counterexamples are difficult to construct, and are widely believed to be rare. The complexity of the matter already appears in the simplest setting, namely sets~$\Sigma$ containing just two $2\times2$ matrices. Indeed, such sets appear in the literature both as finiteness counterexamples~\cite{blondeltheysvladimirov03}, \cite{heremorrissidorovtheys11},
\cite{jenkinsonpollicott18}, \cite{oregon-reyes18},
as well as families of finiteness examples~\cite{jungersblondel08}, \cite{cicone_et_al10}, \cite{kozyakin16}.
In this paper we deal with sets $\Sigma=\set{A,B}$ of matrices in $\SL_2\mathbb{R}$, the group of $2\times2$ matrices with real entries and determinant one. Such matrices act on the hyperbolic plane $\mc{H}=\set{z\in\mathbb{C}:\im z>0}$ via M\"obius isometries $\ppmatrix{a}{b}{c}{d}*z=(az+b)/(cz+d)$, and whenever the group $\Gamma$ generated by $\Sigma$ is fuchsian (i.e., acts on $\mc{H}$ in a properly discontinuous way) the quotient $X=\Gamma\backslash\mc{H}$ is a complete hyperbolic surface. In this case, asking about the joint spectral radius of $\Sigma$ amounts to asking about the maximal \newword{mean free path} along closed geodesics on $X$, namely about the maximal mean time interval between successive crossings of fixed cuts of~$X$ (corresponding to the generators $A,B$ of $\Gamma$) that can be realized among closed geodesics; see~\cite[\S10]{panti20b}. This geometric point of view appears also in~\cite[\S6]{breuillardsert}, where it is discussed the case of two hyperbolic translations well oriented and with disjoint axes (corresponding to $X$ being a pair of pants, provided the two axes are sufficiently far apart). It also appears in~\cite{gekhtmantaylortiozzo19}, although the authors are concerned there with the limiting distribution of mean free paths (which turns out to be gaussian), rather than with their maximal value.
We summarize our results as follows, referring to the following sections for detailed statements.
Fix $\Sigma=\set{A,B}\subset\SL_2\mathbb{R}$ with $\mathrm{tr}(A),\mathrm{tr}(B)\ge2$. We say that $C\in\Sigma^n$ is an \newword{optimal product} if $\tilde\rho(\Sigma)=\rho(C)^{1/n}$ and for no $1\le k<n$ and $D\in\Sigma^k$ we have $\tilde\rho(\Sigma)=\rho(D)^{1/k}$. The existence of optimal products amounts to the validity of the finiteness conjecture; their uniqueness ---which may or may not hold--- is intended up to conjugation.
We assume that $A,B$ are coherently oriented; this is a geometric condition (see Definition~\ref{ref2}) that turns out to be equivalent to the fact that $A,B$ are simultaneously conjugate to a pair of nonnegative matrices.
Discarding the trivial case in which $A$ and $B$ commute and hence are simultaneously diagonalizable (or triangularizable, if parabolic),
we will prove the following results.
\begin{itemize}
\item[(I)] If $A$ and $B$ are hyperbolic with intersecting axes, then the one with larger trace is the unique optimal product. If they have the same trace, they are both optimal.
\item[(II)] If $A$ and $B$ are hyperbolic with asymptotically parallel axes, then the one with larger trace is the unique optimal product (this also holds if one of the two is parabolic with fixed point equal to one of the two fixed points of the other). If they have the same trace, then:
\begin{itemize}
\item[(II.1)] If the attracting fixed point of one of the two is repelling for the other, then $A$ and $B$ are both optimal, and no other product is optimal;
\item[(II.2)] Otherwise,
every product which is not a power is optimal.
\end{itemize}
\item[(III)] If $A$ and $B$ have the same trace and the pair $A,B^{-1}$ is \emph{not} coherently oriented, then $AB$ is the unique optimal product.
\item[(IV)] The above statements leave uncovered the cases in which $A,B$ are both hyperbolic with different trace and ultraparallel axes, or one of the two is parabolic with fixed point distinct from the two fixed points of the other. Assume we are in one of these cases with $\mathrm{tr}(A)<\mathrm{tr}(B)$, and further assume that $A$ and $B$ have integer entries.
\begin{itemize}
\item[(IV.1)] If $\mathrm{tr}(AB)<\mathrm{tr}(B^2)$, then $B$ is the unique optimal product;
\item[(IV.2)] If $\mathrm{tr}(AB)=\mathrm{tr}(B^2)$, then $AB^2$ is the unique optimal product;
\item[(IV.3)] If $\mathrm{tr}(AB) > \mathrm{tr}(B^2)$, then $\mathrm{tr}((AB)^3)$ and $\mathrm{tr}((AB^2)^2)$ differ at least by~$2$; if the former is larger, then $AB$ is the unique optimal product, otherwise so is $AB^2$;
\item[(IV.4)] The statements (IV.1), (IV.2), (IV.3) are false if the assumption about integer entries is dropped.
\end{itemize}
\end{itemize}
Putting together the above statements, we obtain the result stated in the abstract.
This paper is organized as follows: in~\S\ref{ref14} we give the definition of coherent orientation for pairs of nonelliptic matrices in $\SL_2\mathbb{R}$ and prove the equivalence alluded to above. We then establish
in Theorem~\ref{ref3} inequalities relating the translation length of a matrix product with the sum of the translation lengths of the factors. In~\S\ref{ref4} we recast the finiteness property in terms of the existence of maximal elements for a certain preorder defined in the free semigroup on two generators; this interpretation allows us to replace optimal matrix products with better behaved optimal words. We prove statements (I), (II), (III) above in Theorems~\ref{ref7}, \ref{ref9}, and \ref{ref10}. In~\S\ref{ref11} we provide counterexamples and settle~(IV.4). In~\S\ref{ref12} we restrict attention to integer matrices and move from geometric arguments to combinatorial ones, establishing~(IV.1) in Theorem~\ref{ref15}.
The statements (IV.2) and (IV.3) are more involved, requiring a section each, and are established in Theorems~\ref{ref16} and~\ref{ref18}.
\section{Coherently oriented nonelliptic matrices}\label{ref14}
The M\"obius action of $\SL_2\mathbb{R}$ cited in \S\ref{ref1} extends naturally to the euclidean boundary of the hyperbolic plane, namely the real projective line $\PP^1\mathbb{R}=\partial\mc{H}$. A nonidentity matrix $A\in\SL_2\mathbb{R}$ is then \newword{elliptic}, \newword{parabolic}, or \newword{hyperbolic} according to the number of fixed points ---either zero, one, or two--- it has in $\partial\mc{H}$; equivalently, according to the absolute value of its trace being less than, equal to, or greater than~$2$.
Note that replacing $A$ with $-A$ does not change the action and is irrelevant with respect to anything related to spectral radii.
If $A$ is hyperbolic, one of its fixed points is attracting and we will be denoted by $\alpha^+$, the repelling one being denoted~$\alpha^-$; similar conventions hold for other letters $B,C,\ldots$. If $A$ is parabolic, we agree that $\alpha^+=\alpha^-$ is the only fixed point of $A$.
Let $A$ be a nonidentity matrix in $\SL_2\mathbb{R}$ of trace $\ge2$, and let $d$ denote hyperbolic distance (see~\cite{fenchel89} or~\cite{katok92} for basics of hyperbolic geometry). The \newword{translation length} of $A$ is
\[
\ell(A)=\inf\set{d(z,A*z):z\in\mc{H}}.
\]
It has value $0$ if and only if the infimum is not realized by any $z$, if and only if $A$ is parabolic. If $A$ is hyperbolic, then the set of points $z$ realizing the infimum are precisely those points that lie on the \newword{translation axis} of $A$, namely the unique geodesic of ideal endpoints $\alpha^+$ and
$\alpha^-$. For $A$ as above, spectral radius, trace, and translation length have neat relationships, namely
\begin{equation}\label{eq1}
\begin{aligned}
\rho &=\mathrm{tr}/2+\sqrt{(\mathrm{tr}/2)^2-1} = \exp(\ell/2),\\
\mathrm{tr} &=\rho+\rho^{-1} = 2\cosh(\ell/2),\\
\ell &= 2\arccosh(\mathrm{tr}/2) = 2\log\rho.
\end{aligned}
\end{equation}
Since the functions involved in~\eqref{eq1} are order-preserving bijections between the intervals $[1,\infty)$ (for spectral radius), $[2,\infty)$ (for trace), and $[0,\infty)$ (for translation length), comparing nonelliptic matrices with respect to one of these characteristics is the same as comparing them with respect to any other. Moreover, for every $A,B\in\SL_2\mathbb{R}$ with trace $\ge2$, we have $\rho(A)<\rho(B)$ if and only if $\rho(A^n)<\rho(B^n)$ for some (equivalently, for all) $n\ge1$, and the same statement holds for trace and for translation length.
We look at the ideal boundary $\partial\mc{H}$ as a topological circle, cyclically ordered by the ternary betweenness relation $\alpha\prec \beta\prec \gamma$, which reads ``$\alpha,\beta,\gamma$ are pairwise distinct, and traveling from $\alpha$ to~$\gamma$ counterclockwise we meet $\beta$''. Every pair of distinct points $\alpha,\beta$ determines two closed intervals, namely $[\alpha,\beta]=\set{\alpha,\beta}\cup\set{x:\alpha\prec x\prec \beta}$ and $[\beta,\alpha]=\set{\beta,\alpha}\cup\set{x:\beta\prec x\prec \alpha}$.
\begin{definition}
Let\label{ref2} $A,B$ be noncommuting matrices in $\SL_2\mathbb{R}$, both with trace greater than or equal to~$2$. If $\alpha^+=\beta^+$, let $I^+=\set{\alpha^+}$. If $\alpha^+\ne\beta^+$, let $I^+$ be the one, of the two intervals $[\alpha^+,\beta^+]$ and $[\beta^+,\alpha^+]$, which is mapped into itself by both $A$ and $B$, if such an interval exists (if it does then it is unique, since $AB\ne BA$ implies $\set{\alpha^+,\alpha^-}\ne\set{\beta^+,\beta^-}$). If such an interval does not exists, leave $I^+$ undefined. Replace in the above lines $A$, $B$ with $A^{-1}$, $B^{-1}$, and $\alpha^+,\beta^+$ with $\alpha^-,\beta^-$, obtaining the definition of~$I^-$. If both of $I^+$ and $I^-$ are defined, then we say that the pair $A,B$ is
\newword{coherently oriented}.
If $A,B$ are coherently oriented, but $A,B^{-1}$ are not, then we say that $A,B$ are \newword{well oriented}.
\end{definition}
It is clear that $A,B$ are coherently oriented if and only if so are $A^{-1},B^{-1}$. Coherently oriented hyperbolic pairs with ultraparallel axes are necessarily well oriented; see Example~\ref{ref20} and Figure~\ref{fig3} for taxonomy.
\begin{lemma}
Let\label{ref19} $A,B$ be noncommuting matrices in $\SL_2\mathbb{R}$, both with trace $\ge2$. Then they are coherently oriented if and only if there exists $C\in\SL_2\mathbb{R}$ such that $CAC^{-1}$ and $CBC^{-1}$ have nonnegative entries.
\end{lemma}
\begin{proof}
Assume $A,B$ are coherently oriented with $I^+,I^-$ as in Definition~\ref{ref2}. Since $\set{\alpha^+,\alpha^-}\ne\set{\beta^+,\beta^-}$, at least one of $I^+,I^-$ is not a singleton, say $I^+$. Let $K$ be the closure of the complement of $I^+$. Then $\alpha^-,\beta^-\in K$; indeed if, say, $\alpha^-$ were in the interior of $I^+$ we would have $A*\beta^+\notin I^+$, which is impossible.
This fact implies that $A^{-1}[K]\cup B^{-1}[K]\subseteq K$.
Let now $C$ be any matrix in $\SL_2\mathbb{R}$ such that $C[I^+]=[0,\infty]$. Setting $D=CAC^{-1}$ and $E=CBC^{-1}$ we obtain
$D[0,\infty]\cup E[0,\infty]\subseteq [0,\infty]$
and
$D^{-1}[\infty,0]\cup E^{-1}[\infty,0]\subseteq [\infty,0]$.
Write $D=\ppmatrix{a}{b}{c}{d}$; we want to prove that $a,b,c,d\ge0$. Since $D[0,\infty]\subseteq[0,\infty]$, we have that $a$ and $c$ have the same sign, and so do $b$ and $d$. The involution $S=\ppmatrix{}{-1}{1}{}$ exchanges $[0,\infty]$ with $[\infty,0]$. As a consequence,
$SD^{-1} S^{-1}=\ppmatrix{a}{c}{b}{d}$ maps $[0,\infty]$ into itself, which implies that $a$ and $b$ have the same sign, and so do $c$ and $d$.
We conclude that all of $a,b,c,d$ have the same sign, which must be positive, since $\mathrm{tr}(D)\ge2$; the same argument works for $E$.
Conversely, let $A,B$ have nonnegative entries; then $\alpha^+,\beta^+\in[0,\infty]$ and $\alpha^-,\beta^-\in[\infty,0]$. Taking $I^+$ to be the interval of endpoints $\alpha^+,\beta^+$ which is contained in $[0,\infty]$, and analogously for $I^-$, we see that $A,B$ satisfy the conditions of coherent orientation, which are plainly preserved under conjugation.
\end{proof}
\begin{example}
Consider\label{ref20} the following matrices, where zero entries are replaced by spaces:
\begin{gather*}
C=\begin{pmatrix}
9 & 8\\
1 & 1
\end{pmatrix},\quad
D=\begin{pmatrix}
5 & -1\\
1 &
\end{pmatrix},\quad
E=\frac{1}{10}\begin{pmatrix}
17-2\sqrt{6} & -12+2\sqrt{6} \\
-3-2\sqrt{6} & 8+2\sqrt{6}
\end{pmatrix},\\
G=\begin{pmatrix}
5 & -4\\
4 & -3
\end{pmatrix},\quad
L=\begin{pmatrix}
1 & \\
1 & 1
\end{pmatrix}.
\end{gather*}
We draw in Figure~\ref{fig3} the oriented translation axes of the hyperbolic $C,D,E$, as well as oriented horocycles corresponding to the parabolic $G,L$; note that although we work in the upper-plane model~$\mc{H}$, we draw pictures in the Poincar\'e disk model.
\begin{figure}
\begin{tikzpicture}[scale=2.4]
\coordinate (cp) at (0.2219422, 0.97505980);
\coordinate (cm) at (-0.994356, -0.10609428);
\coordinate (ccontrol) at (0.05555949, 0.56253986);
\coordinate (decontrol) at (0.2087727, -0.01665920);
\coordinate (dp) at (0.3999999, 0.91651513);
\coordinate (dm) at (0.4, -0.9165151);
\coordinate (ep) at (-0.99435607, -0.1060942);
\coordinate (em) at (1,0);
\draw (0,0) circle [radius=1cm];
\draw[middlearrow={latex}] (cm) to[arc through ccw=(ccontrol)] (cp);
\draw[middlearrow={latex}] (dm) to[arc through cw=(decontrol)] (dp);
\draw[middlearrow={latex}] (em) to[arc through ccw=(decontrol)] (ep);
\draw[middlearrow={latex}] (em) arc (0:-360:0.16cm);
\draw[middlearrow={latex}] (0,-1) arc (270:-90:0.16cm);
\node[left] at (cm) {$\gamma^-=\epsilon^+$};
\node at ($(cp) + (0.03,0.14)$) {$\gamma^+$};
\node at ($(dm) + (0.12,-0.12)$) {$\delta^-$};
\node at ($(dp) + (0.12,0.12)$) {$\delta^+$};
\node[right] at (em) {$\epsilon^-=1$};
\node[below] at (0,-1) {$0$};
\end{tikzpicture}
\caption{Examples of coherently oriented pairs}
\label{fig3}
\end{figure}
Direct checking shows that coherently oriented pairs can be classified in six subcases as follows, three of them being well oriented.
\begin{itemize}
\item The parabolic-parabolic case, which is necessarily well oriented. The pair $G^{-1},L$ above is an example (with $I^+=[0,1]$ and $I^-=[1,0]$); note that $G,L$ are not coherently oriented. This case is covered by
Theorem~\ref{ref10}.
\item The parabolic-hyperbolic case, which splits in two.
A pair may be well oriented (e.g., $C,G$ with $I^+=[1,\gamma^+]$, $I^-=[\gamma^-,1]$), or coherently oriented but not well oriented (e.g., $E,G$).
The first subcase is covered by Theorems~\ref{ref15}, \ref{ref16}, \ref{ref18}, and the second by Theorem~\ref{ref9}.
\item The hyperbolic-hyperbolic case. This splits in three, the subcases of intersecting (such as $D^{\pm1},E^{\pm1}$) or asymptotically parallel axes (such as $C^{\pm1},E^{\pm1}$) being not well oriented. The remaining case, exemplified by $C,D$, is of course well oriented. The first subcase is covered by Theorem~\ref{ref7}, the second by Theorem~\ref{ref9}, and the third by Theorems~\ref{ref15}, \ref{ref16}, \ref{ref18}.
\end{itemize}
We remark that if two matrices in $\SL_2\mathbb{Z}$ are coherently oriented, it is not necessarily true that they can be conjugated to a pair with nonnegative \emph{integer} entries; the pair $C,D$ above is one such example.
\end{example}
\begin{lemma}
Let\label{ref8} $A,B$ have trace $\ge2$, and assume they are coherently oriented; let $C$ be a product of $A$ and $B$.
\begin{enumerate}
\item We have $\mathrm{tr}(C)\ge2$, $\gamma^+\in I^+$, $\gamma^-\in I^-$; in particular, if $I^+\cap I^-=\emptyset$ then $C$ is hyperbolic.
\item If $\alpha^+\ne\beta^+$ and $C$ is not a power of $B$, then $\gamma^+\ne\beta^+$; an analogous statement holds for repelling fixed points.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) Surely $\mathrm{tr}(C)>0$ by Lemma~\ref{ref19}. Since $A[I^+]\cup B[I^+]\subseteq I^+$ we have $C[I^+]\subseteq I^+$ and a descending chain $I^+\supseteq C[I^+]\supseteq C^2[I^+]\supseteq\cdots$ that shrinks to $\gamma^+\in I^+$; thus $\mathrm{tr}(C)\ge2$. Inverting both $A$ and $B$ we get $\gamma^-\in I^-$.
(2) We have $C=DAB^k$, for some $k\ge0$ and some product $D$ of $A$ and $B$. Then $\beta^+$ is an endpoint of the interval $B^k[I^+]$, and does not belong to $AB^k[I^+]$. Any further application of $A$ and $B$
to $AB^k[I^+]$ leaves $\beta^+$ outside,
and therefore $\beta^+\notin C[I^+]$, which implies $\gamma^+\ne\beta^+$.
\end{proof}
\begin{theorem}
Let\label{ref3} $A,B$ be hyperbolic and coherently oriented with $I^+\cap I^-=\emptyset$. Then $\ell(AB)$ is less than, equal to, or greater than $\ell(A)+\ell(B)$ if and only if the axes of $A$ and $B$ are intersecting, asymptotically parallel, or ultraparallel, respectively.
\end{theorem}
\begin{proof}
Assume that the axes are asymptotically parallel. Then, possibly replacing $A,B$ with $A^{-1},B^{-1}$, we may conjugate and assume $\alpha^+=\beta^+=\infty$ (since $I^+$ and $I^-$ do not intersect, $\alpha^+=\beta^-$ is excluded). We then have
\[
A=\begin{pmatrix}
r & t\\
& r^{-1}
\end{pmatrix},\quad
B=\begin{pmatrix}
s & u\\
& s^{-1}
\end{pmatrix},
\]
with $r,s>1$. Since $\ell(A)=2\log r$, and analogously for $B$ and $AB$, we obtain $\ell(AB)=\ell(A)+\ell(B)$.
\begin{figure}
\begin{tikzpicture}[scale=2.4]
\coordinate (cp) at (0.94115,-0.33796);
\coordinate (cm) at (-0.94115,0.33796);
\coordinate (dp) at (0.91036,0.41380);
\coordinate (dm) at (-0.91036,-0.41380);
\coordinate (x) at (0,0);
\coordinate (em) at (-0.93956,-0.34236);
\coordinate (ep) at (0.97341,-0.22904);
\coordinate (p1) at (-0.33796,-0.94115);
\coordinate (p3) at (0.33796,0.94115);
\coordinate (p2) at (0.41380, -0.91036);
\coordinate (p4) at (-0.41380, 0.91036);
\coordinate (q2) at (0.91697,0.39892);
\coordinate (q1) at (0.45372,-0.89113);
\coordinate (q3) at (-0.97537,0.22055);
\coordinate (q4) at (-0.47517,-0.87988);
\coordinate (y) at (-0.45205,-0.20547);
\coordinate (z) at (0.40665,-0.14602);
\draw (0,0) circle [radius=1cm];
\draw[line width=0.5mm] (cm) to (cp);
\draw[line width=0.5mm] (dm) to (dp);
\draw[line width=0.5mm] (em) to[arc through cw=(y)] (ep);
\draw (q3) to[arc through cw=(y)] (q4);
\draw (q1) to[arc through cw=(z)] (q2);
\draw (p1) to (p3);
\draw (p2) to (p4);
\node at ($(x) + (0,0.17)$) {$x$};
\node at ($(y) + (-0.03,-0.13)$) {$y$};
\node at ($(z) + (0.03,-0.11)$) {$z$};
\node at ($(cp) + (0.1,-0.06)$) {$\alpha^+$};
\node at ($(dp) + (0.12,0.1)$) {$\beta^+$};
\node at ($(cm) + (-0.1,0.1)$) {$\alpha^-$};
\node at ($(dm) + (-0.1,-0.12)$) {$\beta^-$};
\node at ($(p4) + (0.2,-0.1)$) {$\sigma_1$};
\node at ($(p3) + (0.06,-0.2)$) {$\sigma_4$};
\end{tikzpicture}
\hspace{0.6cm}
\begin{tikzpicture}[scale=2.4]
\coordinate (cp) at (0.86602,-0.5);
\coordinate (cm) at (-0.86602,-0.5);
\coordinate (dp) at (0.93629,0.35121);
\coordinate (dm) at (-0.52662,0.85);
\coordinate (em) at (-0.70912,0.70508);
\coordinate (ep) at (0.94089,-0.33869);
\coordinate (x) at (0.2114481, 0.310447);
\coordinate (q2) at (0.9572039, 0.2894142);
\coordinate (q1) at (0.466348,-0.884601);
\coordinate (q3) at (0.10452, 0.994522);
\coordinate (q4) at (-0.952337,- 0.3050460);
\coordinate (y) at (0.154240, 0.033880);
\coordinate (z) at (0.11888,- 0.272033);
\coordinate (w) at (0.39264, - 0.31304);
\coordinate (t) at (-0.073642, 0.4218477);
\coordinate (u) at (0.14726, - 0.98909);
\coordinate (v) at (0.42332, 0.90598);
\draw (0,0) circle [radius=1cm];
\draw[line width=0.5mm] (cm) to[arc through cw=(z)] (cp);
\draw[line width=0.5mm] (dm) to[arc through ccw=(x)] (dp);
\draw[line width=0.5mm] (em) to[arc through ccw=(y)] (ep);
\draw (q1) to[arc through cw=(w)] (q2);
\draw (q4) to[arc through ccw=(t)] (q3);
\draw (u) to[arc through cw=(y)] (v);
\node at ($(cp) + (0.1,-0.1)$) {$\alpha^+$};
\node at ($(dp) + (0.12,0.1)$) {$\beta^+$};
\node at ($(cm) + (-0.1,-0.1)$) {$\alpha^-$};
\node at ($(dm) + (-0.02,0.1)$) {$\beta^-$};
\node at ($(x) + (0.1,0.1)$) {$x$};
\node at ($(w) + (0.08,-0.1)$) {$w$};
\node at ($(y) + (0.04,-0.42)$) {$y$};
\node at ($(x) + (-0.32,0.24)$) {$z$};
\node at ($(y) + (-0.7,0)$) {$\sigma_2$};
\node at ($(y) + (-0.14,-0.66)$) {$\sigma_1$};
\node at ($(w) + (0.32,0.32)$) {$\sigma_3$};
\end{tikzpicture}
\caption{Coherently oriented geodesics, intersecting case left, nonintersecting right}
\label{fig1}
\end{figure}
Assume the axes intersect; denote by $x$ the intersection point, by $y$ the point at distance $\ell(B)/2$ from $x$ moving towards $\beta^-$, and by $z$ the point at distance $\ell(A)/2$ from $x$ moving towards $\alpha^+$. We sketch the situation in Figure~\ref{fig1} left. Let $\sigma_1$ and $\sigma_2$ be the geodesics perpendicular to the axis $\sigma_3$ of $B$ and passing through $x$ and $y$, respectively. Also, let $\sigma_4$ and $\sigma_5$ be the geodesics perpendicular to the axis $\sigma_6$ of $A$ and passing through $x$ and $z$, respectively. For each $i=1,\ldots,6$, the reflection $S_i$ through $\sigma_i$ is an isometric involution of $\mc{H}$, and we have $A=S_5S_4$ and $B=S_1S_2$. The composition $S_4S_1$ is a rotation about $x$, and equals the composition $S_6S_3$, because the pair $(\sigma_4,\sigma_1)$ is mapped to $(\sigma_6,\sigma_3)$ by a rotation of $\pi/2$ about $x$. Summing up, we obtain
\[
AB=(S_5S_4)(S_1S_2)=S_5(S_4S_1)S_2=S_5(S_6S_3)S_2=(S_5S_6)(S_3S_2),
\]
which is the composition of a rotation of $\pi$ about $y$, followed by a rotation of $\pi$ about $z$. These two rotations leave the geodesic through $y$ and $z$ invariant, and thus this geodesic is the axis of $AB$; moreover,
\[
\ell(AB)=d(y,AB*y)=d(y,S_5S_6*y)=2d(y,z).
\]
By the triangle inequality we conclude
\[
\frac{\ell(AB)}{2}=d(y,z)<d(y,x)+d(x,z)=\frac{\ell(A)}{2}+
\frac{\ell(B)}{2},
\]
as desired.
Assume now that the axes of $A$ and $B$ are ultraparallel; see Figure~\ref{fig1} right. Then they determine a unique common perpendicular, denoted by $\sigma_1$, which intersects the axis of $B$ in $x$ and the axis of $A$ in $y$. Let $z$ be the point at distance $\ell(B)/2$ from $x$ moving towards $\beta^-$, and $w$ the point at distance $\ell(A)/2$ from $y$ towards $\alpha^+$. Draw the perpendicular $\sigma_2$ at $z$ to the axis of $B$, and the perpendicular $\sigma_3$ at $w$ to the axis of $A$. Defining as above $S_i$ to be the reflection of mirror $\sigma_i$, we have $A=S_3S_1$, $B=S_1S_2$, and $AB=S_3S_2$, because $S_1$ cancels. Ultraparallel geodesics have a well-defined hyperbolic distance, still denoted by $d$, and we have $d(\sigma_1,\sigma_2)=\ell(B)/2$, $d(\sigma_1,\sigma_3)=\ell(A)/2$, and $d(\sigma_2,\sigma_3)=\ell(AB)/2$.
Denote by $\overline{\sigma}_i$ the euclidean circle (possibly a straight line) in $\mathbb{C}$ of which $\sigma_i$ is an arc. Then the $\overline{\sigma}_i$s are pairwise nonintersecting (because so are the $\sigma_i$s, and an intersection outside $\mc{H}$ would produce an intersection inside, by M\"obius inversion through $\partial\mc{H}$), and $\overline{\sigma}_1$ separates $\overline{\sigma}_2$ from $\overline{\sigma}_3$, meaning that any circle intersecting $\overline{\sigma}_2$ and $\overline{\sigma}_3$ intersects $\overline{\sigma}_1$ too.
The circles $\partial\mc{H}$ and $\overline{\mathrm{axis}(AB)}$ are distinct and perpendicular to both $\overline{\sigma}_2$ and $\overline{\sigma}_3$; thus the set of circles perpendicular to both
$\partial\mc{H}$ and $\overline{\mathrm{axis}(AB)}$ constitute the \newword{coaxial pencil} determined by the pair
$\overline{\sigma}_2,\overline{\sigma}_3$~\cite[\S4]{coxeter71}. The key observation here is that $\overline{\sigma}_1$ does not belong to this pencil, since it is not perpendicular to $\mathrm{axis}(AB)$ (because common perpendiculars to pairs of ultraparallel geodesics are unique, and $A$, $B$, $AB$ have distinct axes by Lemma~\ref{ref8}(3)).
Now, while points in the hyperbolic plane obey the triangle inequality, ultraparallel geodesics obey the \newword{non-triangle inequality}~\cite[\S6]{coxeter71}, according to which the distance between ultraparallel geodesics is strictly greater than the sum of distances between the two given geodesics and a third one, separating the two but not coaxial to them. In our case we get
\[
\frac{\ell(AB)}{2}=d(\sigma_2,\sigma_3)>d(\sigma_2,\sigma_1)+d(\sigma_1,\sigma_3)=\frac{\ell(B)}{2}+
\frac{\ell(A)}{2},
\]
again as desired.
\end{proof}
\section{Words}\label{ref4}
As anticipated in~\S\ref{ref1}, some caution is required in defining the length of matrix products. The problem is, of course, that a given pair $A,B\in\SL_2\mathbb{R}$ (even with nonnegative entries) may fail to generate not only a free group ---a tolerable fault--- but even a free semigroup. For example, the matrices
\[
A=\begin{pmatrix}
1 & \sqrt{6} \\
& 1
\end{pmatrix},\quad
B=\begin{pmatrix}
1 & \\
\sqrt{6} & 1
\end{pmatrix},
\]
satisfy the nontrivial identity
\[
A^2B^3A^2=BA^6B=
\begin{pmatrix}
2 & \sqrt{6} \\
\sqrt{6}/2 & 2
\end{pmatrix};
\]
see~\cite{brennercharnow78} for other examples.
We deal with the issue by working with free semigroups of words, as follows.
Let $\set{a,b}$ be a two-letter alphabet, $F_2^+$ the free semigroup of words $w$ of length $\abs{w}\ge1$, and $F_2$ the enveloping free group. Once a pair $A,B\in\SL_2\mathbb{R}$ has been fixed, we consider the group homomorphism $\phi:F_2\to\SL_2\mathbb{R}$ that maps $a$ to $A$ and $b$ to $B$, and the induced character $[\,\operatorname{--}\,]:F_2\to\mathbb{R}$ defined by $[w]=\mathrm{tr}(\phi(w))$.
\begin{lemma}
The\label{ref5} following statements are true.
\begin{enumerate}
\item Let $\phi'$, $[\,\operatorname{--}\,]'$ be induced by another matrix choice $A',B'\in\SL_2\mathbb{R}$. If $[a]=[a]'$, $[b]=[b]'$, $[ab]=[ab]'$, then $[\,\operatorname{--}\,]=[\,\operatorname{--}\,]'$.
\item $[w][u]=[wu]+[wu^{-1}]$.
\item Given $w$, let $u$ be either $w^{-1}$, or a rotation of $w$, or the reversal of $w$ (that is, $w$ written backwards). Then $[u]=[w]$.
\item $\ttr{wuv}=\ttr{wv}\ttr{u}-\ttr{wu^{-1} v}$, and thus
$[wu^2v]=[wuv][u]-[wv]$.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) follows from the fact~\cite[Theorem~3.1]{horowitz72} that, for a fixed $w$, there exists a polynomial $f_w\in\mathbb{Z}[x,y,z]$ such that, for varying $\phi$, we have $\mathrm{tr}(\phi(w))=f_w\bigl(\mathrm{tr}(\phi(a)),\mathrm{tr}(\phi(b)),\mathrm{tr}(\phi(ab))\bigr)$.
(2) and the identity $[w]=[w^{-1}]$ are well known, and the invariance of $[\,\operatorname{--}\,]$ under word rotation follows from the invariance of the trace under conjugation. Define $\phi'(a)=A^{-1}$, $\phi'(b)=B^{-1}$; then $[\,\operatorname{--}\,]'=[\,\operatorname{--}\,]$ by (1) and invariance under rotation and group inversion. Letting $u$ be the reversal of $w$, we obtain $[u]=[w^{-1}]'=[w^{-1}]=[w]$, which proves (3). Finally, (4) follows from (2) and rotation invariance.
\end{proof}
A key feature of our formalism is that, not only word length in $F_2^+$ is better behaved than matrix product length, but the implicit comparison of spectral radii in~\eqref{eq2} becomes an explicit preorder on words, as follows.
\begin{definition}
Let $A,B\in\SL_2\mathbb{R}$ have trace greater than or equal to $2$, and assume that they are coherently oriented. Define $\phi$, $\ttr{\,\operatorname{--}\,}$ as above; by Lemma~\ref{ref8}, $\ttr{\,\operatorname{--}\,}$ takes values in $\mathbb{R}_{\ge0}$. We define a binary relation $\preceq_{A,B}$ on $F_2^+$ by
\[
w\preceq_{A,B} u\quad\text{if and only if}\quad\bigl[w^{\abs{u}}\bigr]
\le \bigl[u^{\abs{w}}\bigr].
\]
By saying that a word is \newword{maximal} we mean maximal with respect to $\preceq$ (for simplicity's sake we are dropping in the notation the dependence from $A$ and $B$).
A \newword{complete set of optimal words} is a possibly infinite subset $\set{v_1,v_2,\ldots}$ of $F_2^+$ such that:
\begin{itemize}
\item every $v_i$ is maximal, and is a \newword{Lyndon word}, i.e., is strictly smaller in the lexicographic order than any of its proper rotations (in particular, it is not a power);
\item $v_i\ne v_j$ for $i\ne j$;
\item every maximal $w$ is a power of a rotation of some (necessarily unique) $v_i$.
\end{itemize}
A word that belongs to a complete set of optimal words is an~\newword{optimal} word.
\end{definition}
\begin{lemma}
\begin{enumerate}
\item We\label{ref6} have $w\preceq u$ if and only if $\bigl[w^{m/\abs{w}}\bigr]
\le \bigl[u^{m/\abs{u}}\bigr]$, where $m$ is any common multiple of $\abs{w}$ and $\abs{u}$.
\item The relation $\preceq$ on $F_2^+$ is a preorder, and every two words are comparable.
\item If a complete set of optimal words exists, it is unique.
\item The finiteness conjecture holds for $\Sigma=\set{A,B}$ precisely when $F_2^+$ contains maximal ---equivalently, optimal--- words.
\end{enumerate}
\end{lemma}
\begin{proof}
Let $W=\phi(w)$, and analogously for $u$ and $v$
(later on we will apply this uppercase/lowercase convention without further notice).
By the remarks following Equations~\eqref{eq1}, we have $w\preceq u$ if and only if $\rho\bigl(W^{\abs{u}}\bigr)\le \rho\bigl(U^{\abs{w}}\bigr)$
if and only if
$\rho\bigl(W^{m/\abs{w}}\bigr)\le \rho\bigl(U^{m/\abs{u}}\bigr)$
if and only if $\bigl[w^{m/\abs{w}}\bigr]\le
\bigl[u^{m/\abs{u}}\bigr]$.
It is clear that $\preceq$ is reflexive and every two words are comparable. Assuming $w\preceq u\preceq v$, and letting $m$ be a common multiple of $\abs{w},\abs{u},\abs{v}$, we obtain
$\bigl[w^{m/\abs{w}}\bigr]\le\bigl[u^{m/\abs{u}}\bigr]\le
\bigl[v^{m/\abs{v}}\bigr]$, and thus $w\preceq v$.
Let $S$ and $S'$ be two complete sets of optimal words, and let $v\in S$. Since $v$ is maximal, it is a power of a rotation of some $v'\in S'$; by the elementary properties of Lyndon words, $v=v'$.
The remaining assertions follow straight from the definitions; note that every word is both greater and less than any of its powers. In particular, if the maximal word $w$ is a power of $u$, then $u$ is maximal as well.
\end{proof}
We can now make precise and prove~(I), (II) and (III) in~\S\ref{ref1}. We stipulate for the rest of this paper, and without further repetitions, that $A,B$ are noncommuting matrices in $\SL_2\mathbb{R}$, of trace greater than or equal to $2$, and coherently oriented.
The following theorem settles~(I).
\begin{theorem}
Let\label{ref7} $A,B$ be both hyperbolic, and assume that the translation axes intersect. If $[a]\le [b]$ then $b$ is an optimal word, and so is $a$ provided $[a]=[b]$. There are no other optimal words.
\end{theorem}
\begin{proof}
We show by induction that for every word $w$ of length $n\ge1$ we have $w\preceq b$. For $n=1$ or $w\in\set{a^n,b^n}$ this is true. Let $n>1$ and $w=au$ without loss of generality, with $u$ not a power of $a$. By Lemma~\ref{ref8}, $\upsilon^+\in I^+\setminus\set{\alpha^+}$ and $\upsilon^-\in I^-\setminus\set{\alpha^-}$. Therefore the axes of $A$ and of $U$ intersect, and by Theorem~\ref{ref3} and inductive hypothesis we have $\ell(W) < \ell(A)+\ell(U)\le n\ell(B)$. Since $\ell(A^n)=n\ell(B)$ if and only if $[a]=[b]$, this also shows uniqueness.
\end{proof}
\begin{theorem}
Let\label{ref9} $B$ be hyperbolic.
\begin{itemize}
\item[(1)] If $A$ is hyperbolic as well and the translation axes are asymptotically parallel, then:
\begin{itemize}
\item[(1.1)] If $[a]\ne[b]$, then the only optimal word is the one among $a$ and $b$ that corresponds to the matrix with larger trace.
\item[(1.2)] If $[a]=[b]$ and $I^+\cap I^-=\emptyset$, then every word which is not a power is optimal. Conversely, if $I^+$ and $I^-$ intersect (necessarily in a singleton), then both $a$ and $b$ are optimal, and there are no other optimal words.
\end{itemize}
\item[(2)] If $A$ is parabolic with $\alpha^+=\alpha^-\in\set{\beta^+,\beta^-}$, then $b$ is the only optimal word.
\end{itemize}
\end{theorem}
\begin{proof}
(1) By conjugating, possibly exchanging $A$ with $B$ and inverting both of them, we may assume $\alpha^+=\beta^+=\infty$
when $I^+\cap I^-=\emptyset$ and $\alpha^+=\beta^-=\infty$ when $I^+\cap I^-\neq \emptyset$.
Let $r=\rho(A)$ and $s=\rho(B)$.
In the first case, after a further conjugation by a parabolic matrix fixing $\infty$, and by a diagonal matrix, we obtain
\[
A=\begin{pmatrix}
r & 1 \\
& r^{-1}
\end{pmatrix},\quad
B=\begin{pmatrix}
s & \\
& s^{-1}
\end{pmatrix}.
\]
In the second case we similarly obtain
\[
A=\begin{pmatrix}
r & 1 \\
& r^{-1}
\end{pmatrix},\quad
B=\begin{pmatrix}
s^{-1} & \\
& s
\end{pmatrix}.
\]
It remains to check our claims (1.1) and (1.2) on these two pairs, which is easily done by direct inspection.
(2) is obvious: up to a conjugation we have
\[
A=\begin{pmatrix}
1 & r \\
& 1
\end{pmatrix},\quad
B=\begin{pmatrix}
s & \\
& s^{-1}
\end{pmatrix},
\]
for some $r\in\mathbb{R}\setminus\set{0}$ and $s\in\mathbb{R}_{>0}\setminus\set{1}$.
\end{proof}
\begin{theorem}
Let\label{ref10} $\mathrm{tr}(A)=\mathrm{tr}(B)$ and assume that the pair $A,B$ is well oriented. Then the only optimal word is $ab$.
\end{theorem}
\begin{proof}
By possibly exchanging $A$ with $B$, and after an appropriate conjugation, we reduce our matrices to the standard form
\[
A=\begin{pmatrix}
1 & \\
r & 1
\end{pmatrix},\quad
B=\begin{pmatrix}
1 & r \\
& 1
\end{pmatrix},
\]
for some $r>0$ in the parabolic case, or
\[
A=D^{-1} H D,\quad
B=D H D^{-1},
\]
in the hyperbolic one; here we set
\[
H=
\begin{pmatrix}
\cosh(\ell/2) & \sinh(\ell/2) \\
\sinh(\ell/2) & \cosh(\ell/2)
\end{pmatrix},\quad
D=\begin{pmatrix}
\exp(d/4) & \\
& \exp(-d/4)
\end{pmatrix},
\]
with $\ell=\ell(A)=\ell(B)>0$ and $d=d(\mathrm{axis}(A),\mathrm{axis}(B))>0$.
We will establish the result by showing that, for every word $u$ which is not a power of $ab$ or or $ba$, we have $u\prec ab$.
Fix such a $u$ of length $n$, and let $n_a,n_b$ be the number of occurrences of $a$ ---respectively $b$--- in it. If one of $n_a,n_b$ is zero, our claim is true: this is clear in the parabolic case (because $AB$ is hyperbolic), and follows from Theorem~\ref{ref3} in the hyperbolic one.
\paragraph{\emph{Claim}} Let $\mc{W}(n_a,n_b)$ be the set of words containing $n_a$ occurrences of $a$ and $n_b$ of $b$, and
assume without loss of generality $n_a\ge n_b$.
Let $w\in\mc{W}(n_a,n_b)$ with $[w]$ maximal among words in $\mc{W}(n_a,n_b)$. Then every occurrence of $b$ in $w$ is isolated, that is, is preceded and followed, in the cyclic order, by occurrences of $a$.
\paragraph{\emph{Proof of Claim}} In the hyperbolic case this is the content of~\cite[Lemma~5-3]{jorgensensmith90}. The same statement holds in the parabolic case as well. Indeed, the proof of~\cite[Lemma~5-3]{jorgensensmith90} works by repeatedly applying the identity in Lemma~\ref{ref5}(2), while making use of the following facts (references being relative to the quoted paper).
\begin{enumerate}
\item Equation~(5.1), namely
\begin{equation}\label{eq4}
[a^pb^qa^tb^s]-[a^{p+t}b^{q+s}]=pqts\bigl([aba^{-1} b^{-1}]-2\bigr),
\end{equation}
(in the parabolic case, the Chebychev polynomials $\alpha_k,\beta_k$ of~\cite[\S2]{jorgensensmith90} are both equal to $k$).
By explicit computation, in our case we have
\begin{align*}
[a^pb^qa^tb^s] &= 2+(p+t)(q+s)r^2+pqtsr^4,\\
[a^{p+t}b^{q+s}] &= 2+(p+t)(q+s)r^2, \\
[aba^{-1} b^{-1}] &= r^4+2,
\end{align*}
and~\eqref{eq4} remains true.
\item Lemma~5-2, which carries through.
\item Lemma~4-3, which is only used through the inequality $[a^pb^q]>[a^{p-1}b^{q-1}]$; by direct computation one easily checks that this inequality still holds.
\end{enumerate}
Having proved our claim, we may safely assume that all appearances of $b$ in $u$ are isolated. Since by assumption $u$ is not a power of $ab$ or of $ba$, not all occurrences of $a$ are isolated; therefore, up to a rotation, we have
\[
u=a^{k_1}ba^{k_2}\cdots ba^{k_t},
\]
for some $t\ge2$ and $k_1,\ldots,k_t\ge1$.
We shall show that $\upsilon^+$ is in the interior of $I=[0,1]$ and $\upsilon^-$ in the interior of $[-1,0]$.
Let $k\ge1$; it is clear that $A^k[I]\subset I$ both in the parabolic and in the hyperbolic case. We also have $A^kB[I]\subset I$; indeed, it suffice to consider $k=1$. In the parabolic case one easily computes
\[
AB[I]=\biggl[\frac{r}{r^2+1},\frac{r+1}{r^2+r+1}\biggr]\subset I.
\]
The hyperbolic case reduces to the computation of $AB*1$, since $0<AB*0<AB*1$ anyway. Let $AB\cppvector{1}{1}=\cppvector{s}{t}\in\mathbb{R}_{>0}^2$. Then a little help from SageMath establishes that
\begin{align*}
t-s &= \sinh(-d)+\frac{1}{2}\sinh(d+l)+\frac{1}{2}\sinh(d-l) \\
&= \sinh(-d)+\sinh(d)\cosh(l) \\
&= \sinh(d)\bigl(\cosh(l)-1\bigr)>0,
\end{align*}
and thus $AB*1<1$. As in the proof of Lemma~\ref{ref8} we obtain
$A^{k_1}BA^{k_2}\cdots BA^{k_t}[I]\subset I$, and we conclude that $\upsilon^+$ is in the interior of $I$.
The same argument, applied to the reversal $v$ of $u$, shows that the attracting fixed point of $\phi(v)$ is in the interior of $I$ as well. Letting $J=\ppmatrix{-1}{}{}{1}$, we see that $J$ conjugates $A$ with $A^{-1}$ and $B$ with $B^{-1}$, so that $\phi(v)=J\phi(u^{-1})J$. This implies that $\upsilon^-$ is the $J$-image of this attracting fixed point, and thus is in the interior of $[-1,0]$.
We now exchange $a$ with $b$ in $u$, obtaining $u'$. This corresponds to conjugating $A$ and $B$ by $F=\ppmatrix{}{1}{1}{}$; in particular, the attracting fixed point of $U'=\phi(u')$ is in the interior of $[1,\infty]$, and the repelling one in $[\infty,-1]$. As a consequence, the translation axes of $U$ and of $U'$ are ultraparallel.
By Lemma~\ref{ref5}(1) we have $\ell(U)=\ell(U')$, and by Theorem~\ref{ref3} $\ell(UU')>2\ell(U)=\ell(U^2)$. We have thus found a word, namely $uu'$, that strictly dominates $u$ in the $\preceq$ preorder. Since $uu'$ contains the same number $n=n_a+n_b$ of occurrences of $a$ and of $b$, we apply again our Claim above and infer $uu'\preceq ab$. This yields $u\prec ab$, as required.
\end{proof}
\section{Counterexamples}\label{ref11}
We have thus proved (I), (II), (III) in~\S\ref{ref1}, covering all cases in which $A,B$ are coherently oriented but not well oriented.
From now on we restrict attention to well oriented
matrices with integer entries, and prove~(IV); our tools, and the overall tone of our paper, will perceptibly move from geometry to combinatorics. This is unavoidable, since finiteness counterexample do indeed exist for well oriented pairs in $\SL_2\mathbb{R}$, wildly popping out as the pair varies smoothly in certain $1$-parameters families of perfectly tame well oriented translations; this is the case, e.g. of the Morris example in~\cite[\S2.2]{oregon-reyes18}. In order to make this tone shift more palatable to the reader, we begin by providing counterexamples, that is, by discussing~(IV.4).
Let us first note that any triple $(r,s,t)\in\mathbb{R}_{\ge2}^3$ determines uniquely up to conjugation a pair $A,B\in\SL_2\mathbb{R}$
such that $\mathrm{tr}(A)=r$, $\mathrm{tr}(B)=s$, $\mathrm{tr}(AB)=t$; indeed such triples give coordinates for the Teichm\"uller space of hyperbolic pair of pants.
Let us fix $\mathrm{tr}(A)=101/50$, and vary $\mathrm{tr}(B)=x$ in the interval $[101/50,113/50]$. Adjusting the distance between the axes we may impose that the difference
\[
\Delta=\mathrm{tr}(AB)-\mathrm{tr}(B^2)
\]
be constant, in particular equal to $0$ or to any small positive or negative number; once $\Delta$ is fixed we can compare words in $F_2^+$ in the $\preceq$ order.
We fix $\Delta=0$, so that $\ttr{ab}=\ttr{b^2}=x^2-2$, and compare $ab^2$ with $ab^3$. We must compute $\bigl[(ab^2)^4\bigr]$ and $\bigl[(ab^3)^3\bigr]$.
We have $[ab^2]=[ab][b]-[a]=(x^2-2)x-101/50$ and
$\bigl[(ab^2)^4\bigr]=T_4\bigl([ab^2]\bigr)$, where $T_k(y)$ is the
degree~$k$ polynomial defined recursively by $T_0(y)=2$, $T_1(y)=y$, $T_k(y)=yT_{k-1}(y)-T_{k-2}(y)$. Thus
$\bigl[(ab^2)^4\bigr]$ is a polynomial in $x$ of degree $12$, and so is $\bigl[(ab^3)^3\bigr]$. Explicit computation gives
\begin{multline*}
\bigl[(ab^2)^4\bigr]-\bigl[(ab^3)^3\bigr]=
x^{10} - 101/50 x^9 - 9 x^8 + 303/25 x^7\\
+ 98103/2500 x^6 - 909/50 x^5 - 46103/500 x^4 - 2080903/125000 x^3\\
+ 105559/1250 x^2 + 1618727/31250 x + 2050401/6250000,
\end{multline*}
whose graph appears in Figure~\ref{fig2}.
\begin{figure}
\includegraphics[width=7.5cm]{figure2}
\caption{Graph of $\bigl[(ab^2)^4\bigr]-\bigl[(ab^3)^3\bigr]$
as a function of $\ttr{b}$.}
\label{fig2}
\end{figure}
Therefore, for $x$ ranging in an appropriate interval, we have $ab^2\prec ab^3$ and the word $ab^2$ is not maximal, contrary to~(IV.2).
Fix now $\Delta=-1/50$, so that $\mathrm{tr}(AB)$ is slightly less than $\mathrm{tr}(B^2)$. Then $[ab^2]=(x^2-2-1/50)x-101/50$, while $[b^3]=T_3(x)=x^3-3x$. Thus, for large $x$, the word $ab^2$ dominates $b$, contrary to~(IV.1).
Finally, let $\Delta=1/50$; by analogous computations we obtain
\begin{align*}
\bigl[(ab)^3\bigr]&=x^6 - 297/50 x^4 + 21903/2500 x^2 - 227799/125000,\\
\bigl[(ab^2)^2\bigr]&=x^6 - 99/25 x^4 - 101/25 x^3 + 9801/2500 x^2 \\
&\quad + 9999/1250 x + 5201/2500,
\end{align*}
which have the same value at $x_0\sim 2.0255364739899213\ldots$.
We compare the two words $ab$ and $ab^2$ with their concatenation $abab^2$ by computing the
differences $\bigl[(ab)^5\bigr]-\bigl[(abab^2)^2\bigr]$ and
$\bigl[(ab^2)^5\bigr]-\bigl[(abab^2)^3\bigr]$, which are polynomials in $x$ of degree $8$ and $13$, respectively. These polynomials are negative at $x_0$, so both $ab$ and $ab^2$ are strictly dominated by $abab^2$, and~(IV.3) fails.
\section{Integer matrices and Case~(IV.1)}\label{ref12}
Since in Theorems~\ref{ref7}, \ref{ref9}, \ref{ref10} we covered the case in which $A$ and $B$ have equal trace, we assume from now on, without loss of generality, that $A,B\in\SL_2\mathbb{Z}$ are well oriented and satisfy $2\le\mathrm{tr}(A)<\mathrm{tr}(B)$.
For completeness's sake we provide one specimen for each of the cases (IV.1)--(IV.3) of the Introduction; let~$L$ be as in Example~\ref{ref20} and $N=\ppmatrix{1}{1}{}{1}$. Then we have the following examples:
\begin{align*}
&A=L,B=LN & & &\mathrm{tr}(AB)=4<7=\mathrm{tr}(B^2),\\
&A=LNL,B=NLN^3 & & &\mathrm{tr}(AB)=34=\mathrm{tr}(B^2),\\
&A=L^3N,B=N^2LN^2 & & &\mathrm{tr}(AB)=40>34=\mathrm{tr}(B^2)\\
& & & &\mathrm{tr}((AB)^3)=63880>55223=\mathrm{tr}((AB^2)^2),\\
&A=L^{11},B=LNL & & &\mathrm{tr}(AB)=15>14=\mathrm{tr}(B^2)\\
& & & &\mathrm{tr}((AB)^3)=3330<3362=\mathrm{tr}((AB^2)^2).
\end{align*}
\begin{definition}
A \newword{subword} of the word $w\in F_2^+$ is a possibly empty word obtained from $w$ by deleting one or more not necessarily contiguous letters.
\end{definition}
Since we are now working with matrices having integer entries, the range of $\ttr{\,\operatorname{--}\,}$ is $\mathbb{Z}_{\ge2}$. The remark~(1) in the
following lemma is thus trivial, but key in our proof.
\begin{lemma}
\begin{enumerate}\label{ref13}
\item $\ttr{w}<\ttr{u}$ if and only if $\ttr{w}\le\ttr{u}-1$.
\item If $u$ is a subword of $w$ then $\ttr{u}<\ttr{w}$, exception being made for the case in which $A$ is parabolic and $w$ a power of $a$.
\end{enumerate}
\end{lemma}
\begin{proof}
In order to prove (2) we assume that we are not in the exceptional case, in which $\ttr{u}=\ttr{w}=2$. It suffices to consider the removal of a single letter $c$, which by rotation invariance we may assume being the first one; let then $w=cu$.
If both of $C$ and $U$ are hyperbolic then Theorem~\ref{ref3} applies, yielding
$\ell(W)\ge\ell(C)+\ell(U)$. Thus $\ell(W)>\ell(U)$ and $\ttr{w}>\ttr{u}$ by the remarks following Equation~\eqref{eq1}.
Suppose $C=A$ is parabolic; then, since we are not in the exceptional case, $U$ contains $B$ as a factor and is hyperbolic. Moreover, by Lemma~\ref{ref8}(2), neither of $\upsilon^+,\upsilon^-$ equals the fixed point of $C$.
By Lemma~\ref{ref19} we may assume $C,U\in\SL_2\mathbb{R}_{\ge0}$, and a further conjugation ---if needed-- by the matrix $F$ in the proof of Theorem~\ref{ref10} reduces us to the case
\[
C=\begin{pmatrix}
1 & \\
r & 1
\end{pmatrix},\quad
U=\begin{pmatrix}
a & b \\
c & d
\end{pmatrix},
\]
with $r,a,d>0$ and $b,c\ge0$. Now, $c$ can be $0$ ---it is so precisely when one of $\upsilon^+,\upsilon^-$ equals $\infty$--- but $b$ cannot, because otherwise one of $\upsilon^+,\upsilon^-$ would equal the fixed point $0$ of $C$. We thus obtain $\ttr{w}=a+rb+d>a+d=[u]$. An analogous proof applies if $U$ is parabolic.
\end{proof}
\begin{lemma}
For\label{circle} every integer $s\geq 2$ and every function $f: \mathbb Z/ 2s\mathbb Z \to \mathbb Z$,
there exists $x$ satisfying both the following inequalities
\begin{equation*}
f(x)\geq f(x+2), \quad f(x+1)\le f(x+3).
\end{equation*}
\end{lemma}
\begin{proof}
Assume the negation of our statement: it says that $f(x)\geq f(x+2)$ implies $f(x+1)>f(x+3)$ for every $x$.
This readily leads to a contradiction. Indeed $f(x)<f(x+2)<\cdots$
cannot always increase; hence there must exist $y$ such that $f(y)\geq f(y+2)$. Repeatedly applying to $f(y)\geq f(y+2)$
the negation of out statement
leads to a strictly decreasing sequence $f(y+1) > f(y+3) > f(y+5) >\cdots$, and hence to a contradiction.
\end{proof}
\begin{lemma}
If\label{1-4} $\ttr{ab}<\ttr{b^2}$ then $\ttr {ab^k}< \ttr{b^{k+1}}$, for every $k\geq 1$.
\end{lemma}
\begin{proof}
We work by induction. The case $k=1$ is by hypothesis, and for $k=2$ we have
\[
\ttr{ab^2} = \ttr{ab}\ttr{b}-\ttr{a}
\le \bigl(\ttr{b^2}-1\bigl)\ttr{b} -\ttr{a}< \ttr{b^3} - 1.
\]
Let $k> 2$; repeatedly applying Lemma~\ref{ref5}(2) to the left side we obtain
\begin{align*}
\ttr{ab^k} &= \ttr{ab}\ttr{b^{k-1}} - \ttr{ab^{2-k}}\\
& = \ttr{ab}\ttr{b^{k-1}} - \ttr{a}\ttr{b^{k-2}} + \ttr{ab^{k-2}}\\
& < \bigl(\ttr{b^2}-1\bigr)\ttr{b^{k-1}} - \ttr{b^{k-2}} + \ttr{b^{k-1}}\\
& = \ttr{b^{k+1}} + \ttr{b^{k-3}} -\ttr{b^{k-1}} - \ttr{b^{k-2}} + \ttr{b^{k-1}}\\
& = \ttr{b^{k+1}} + \ttr{b^{k-3}} - \ttr{b^{k-2}}\\
& \le \ttr{b^{k+1}} -1.
\end{align*}
Here the third line follows by induction hypothesis, and the last one from $\ttr{b^{k-3}}<\ttr{b^{k-2}}$, which is valid for $k>2$.
\end{proof}
\begin{lemma}
Among\label{4-5-6} all words of fixed length, the trace-maximizing ones do not contain the factor $a^2$.
\end{lemma}
\begin{proof}
Since $\ttr{a}<\ttr{b}$, no such word $w$ can be a power of $a$.
Assume that $w$ contains $a^2$. Then, up to a rotation, $w=bua^2$, and it is enough to prove $\ttr {bua^2}< \ttr{buab}$.
The right side equals $\ttr{bua} \ttr{b}-\ttr{ua}$, and the other side $\ttr{bua} \ttr{a}-\ttr{bu}$. The difference is
then greater than
$\ttr{bua}-\ttr{ua}+\ttr{bu}$, which is strictly positive by Lemma~\ref{ref13}(2).
\end{proof}
\begin{lemma} \label{41}
If $\ttr{ab}<\ttr{b^2}$ then, for every $s,k_1,\cdots,k_s\geq 1$,
we have
\begin{equation*}
\ttr {ab^{k_1}\cdots ab^{k_s}} < \ttr {b^{k_1+\cdots +k_s+s}}.
\end{equation*}
\end{lemma}
\begin{proof}
We work by induction on $s$, the case $s=1$ having been proved in Lemma \ref{1-4}. Let $s\geq 2$. We can
suppose $k_{s-1}\geq k_{s}$, which ensures that
\begin{align*}
\ttr {ab^{k_1}\cdots ab^{k_{s-1}}(ab^{k_s})^{-1}}=\ttr{ b^{k_1}\cdots ab^{k_{s-1}-k_s}}
\end{align*}
is positive. We thus obtain
\begin{align*}
\ttr {ab^{k_1}\cdots ab^{k_s}} &<\ttr {ab^{k_1}\cdots ab^{k_{s-1}}} \ttr {ab^{k_s}}\\
& \le (\ttr{b^{k_1+\cdots +k_{s-1} + s-1}}-1)(\ttr{b^{k_s+1}}-1)\\
& = \ttr{b^{k_1+\cdots +k_{s-1} + k_s + s}} + \ttr{b^{k_1+\cdots +k_{s-1} - k_s + s-2}}\\
& \qquad - \ttr{b^{k_1+\cdots +k_{s-1} + s-1}} -(\ttr{b^{k_s+1}}-1)\\
&< \ttr{b^{k_1+\cdots +k_{s-1} + k_s + s}}-(\ttr{b^{k_s+1}}-1)\\
&< \ttr{b^{k_1+\cdots +k_{s-1} + k_s + s}}.
\end{align*}
\end{proof}
We can now prove~\S\ref{ref1}(IV.1).
\begin{theorem}
Let\label{ref15} $A,B\in\SL_2\mathbb{Z}_{\ge0}$, and assume $\mathrm{tr}(A)<\mathrm{tr}(B)$ and $\mathrm{tr}(AB)<\mathrm{tr}(B^2)$. The $b$ is the only optimal word.
\end{theorem}
\begin{proof}
Let $w$ be a word of length $n$, trace-maximizing among all words of the same length; we must prove that $w$ does not contain any $a$.
By Lemma~\ref{4-5-6}, $w$ does not contain $a^2$ as a factor. After a rotation we may apply Lemma~\ref{41}, and conclude that $w=b^n$.
\end{proof}
\section{Case (IV.2)}
The remaining cases~(IV.2) and (IV.3) are more involved and require a section each. The standing assumptions in this section for $A,B\in\SL_2\mathbb{Z}_{\ge0}$ are $\mathrm{tr}(A)<\mathrm{tr}(B)$ and $\mathrm{tr}(AB)=\mathrm{tr}(B^2)$. They yield
\begin{equation*}
\ttr{ab} - \ttr{a}\ttr{b} = \ttr{b^2} -\ttr{a}\ttr{b} \geq \ttr{b^2} - (\ttr{b}-1)\ttr{b}
= \ttr{b} - 2 \geq \ttr{a}-1 > 0,
\end{equation*}
or, equivalently, $\ttr{ab^{-1}}<0$. This will be useful several times.
\begin{lemma}
Fix\label{5A} a word $w$ and assume $s\geq 0$. Then we have:
\begin{enumerate}
\item $\ttr{wab(ab^2)^sab^3} < \ttr{wab^2(ab^2)^s ab^2}$;
\item $\ttr{wab^3(ab^2)^sab} < \ttr{wab^2(ab^2)^s ab^2}$, if $w$ is empty or begins with $a$.
\end{enumerate}
\end{lemma}
\begin{proof}
We prove (1). We have
\begin{align*}
\ttr{wab(ab^2)^sab^3}&= \ttr{wab(ab^2)^sab}\ttr{b^2} -
\ttr{wab(ab^2)^sab^{-1}}\\
&=\ttr{wab(ab^2)^sab}\ttr{b}^2-2\ttr{wab(ab^2)^sab}\\
&\quad -\ttr{wab(ab^2)^sa}\ttr{b} + \ttr{wab(ab^2)^sab}\\
&=\ttr{wab(ab^2)^sab}\ttr{b}^2 - \ttr{wab(ab^2)^sab} - \ttr{wab(ab^2)^sa}\ttr{b},
\end{align*}
and, by Lemma~\ref{ref5}(4),
\begin{align*}
\ttr{wab^2(ab^2)^sab^2} &= \ttr{wab(ab^2)^sab^2}\ttr{b} - \ttr{wa(ab^2)^sab^2}\\
& = \ttr{wab(ab^2)^sab}\ttr{b}^2 - \ttr{wab(ab^2)^sa}\ttr{b} - \ttr{wa(ab^2)^s ab^2}.
\end{align*}
Subtracting the first end result from the second we get
\begin{align*}
\ttr{wab(ab^2)^sab} - \ttr{wa(ab^2)^sab^2}
& = \ttr{w(ab^2)^sab}\ttr{ab} -\ttr{w(ab)^{-1}(ab^2)^s ab}\\
&\quad - \ttr{wa(ab^2)^sab}\ttr{b} +\ttr{wa(ab^2)^sa} \\
& = \ttr{w(ab^2)^sab}\bigl(\ttr{ab}-\ttr{a}\ttr{b}\bigr) -\ttr{wb(ab^2)^{s-1}ab} \\
&\qquad + \ttr{wb^2(ab^2)^{s-1}ab}\ttr{b} +\ttr{wa(ab^2)^sa}.
\end{align*}
If $s\geq 1$ this is strictly positive by the observation preceding the lemma and Lemma~\ref{ref13}(2).
This also holds when $s=0$, since the sum of the two middle terms
becomes $-\ttr{w}+\ttr{wb}\ttr{b}>0$.
The proof of (2) is similar, except that in the second expansion we work on the last $ab^2$. We have
\begin{align*}
\ttr{wab^3(ab^2)^sab}
& = \ttr{wab(ab^2)^sab}\ttr{b^2} - \ttr{wab^{-1}(ab^2)^sab}\\
& = \ttr{wab(ab^2)^sab}\ttr{b}^2 - 2\ttr{wab(ab^2)^sab}\\
& \quad -\ttr{wa(ab^2)^sab}\ttr{b} + \ttr{wab(ab^2)^sab}\\
& = \ttr{wab(ab^2)^sab}\ttr{b}^2 - \ttr{wab(ab^2)^sab} - \ttr{wa(ab^2)^sab}\ttr{b},
\end{align*}
and
\begin{align*}
\ttr{wab^2(ab^2)^sab^2} &= \ttr{wab^2(ab^2)^sab}\ttr{b} - \ttr{wab^2(ab^2)^sa}\\
& = \ttr{wab(ab^2)^sab}\ttr{b}^2 - \ttr{wa(ab^2)^sab}\ttr{b} - \ttr{wab^2(ab^2)^s a}.
\end{align*}
As above, subtracting the two end results we get
\begin{align*}
\ttr{wab&(ab^2)^sab} - \ttr{wab^2(ab^2)^s a} \\
& = \ttr{wab(ab^2)^sab} - \ttr{wab(ab^2)^s a}\ttr{b} + \ttr{wa(ab^2)^s a}\\
& = \ttr{wab(ab^2)^sa}\ttr{b} - \ttr{wab(ab^2)^sab^{-1}} - \ttr{wab(ab^2)^s a}\ttr{b} + \ttr{wa(ab^2)^s a}\\
& = \ttr{wa(ab^2)^s a} - \ttr{wab(ab^2)^sab^{-1}} \\
& = \ttr{wa(ab^2)^s a} - \ttr{wab(ab^2)^s}\ttr{ab^{-1}} + \ttr{wab(ab^2)^sb a^{-1}}.
\end{align*}
Since $\ttr{ab^{-1}}<0$ and $w$ is empty or beginning with $a$, this is positive.
\end{proof}
\begin{lemma}
Under\label{5B} the same hypotheses of Lemma~\ref{5A} we have:
\begin{enumerate}
\item $\ttr{wab(ab^2)^sab^4} < \ttr{wab^2(ab^2)^s ab^3}$,
\item $\ttr{wab^4(ab^2)^sab} < \ttr{wab^3(ab^2)^s ab^2}$, if $w$ is empty or begins with $a$.
\end{enumerate}
\end{lemma}
\begin{proof}
(1) follows by Lemma~\ref{5A}(1), applied to the word $bw$.
(2) Write $\widetilde w$ for the reversal of $w$. Then by Lemma~\ref{ref5}(3) we obtain
\begin{align*}
\ttr{wab^4(ab^2)^sab} &= \ttr{ba(b^2a)^sb^4a\widetilde{w}}\\
&= \ttr{a\widetilde{w}b(ab^2)^sab^4}\\
&< \ttr{a\widetilde{w}b^2(ab^2)^sab^3}
\tag{\text{by (1)}}\\
&= \ttr{wab^3a(b^2a)^sb^2}\\
&= \ttr{wab^3(ab^2)^sab^2}.
\end{align*}
Note that the use of~(1) in the third line is legitimate, since
$\widetilde{w}$ ends with $a$, or is empty.
\end{proof}
\begin{lemma}
Let\label{5C} $w$ be a word that is empty or ends with $b$, and let $k,h\geq 0$.
Then we have
\[
\ttr{ab^2wab(ab^2)^kab(ab^2)^hab} < \ttr{ab^2w(ab^2)^{k+h+2}}.
\]
\end{lemma}
\begin{proof}
On the left hand side we have
\begin{equation}\label{eq5}
\begin{split}
\ttr{ab^2w&ab(ab^2)^kab(ab^2)^hab} \\
&= \ttr{bwab(ab^2)^kab(ab^2)^hab}\ttr{ab} - \ttr{bwab(ab^2)^kab(ab^2)^h}\\
&= \ttr{bwab(ab^2)^k(ab^2)^hab}\ttr{ab}^2 - \ttr{bwab(ab^2)^kb(ab^2)^{h-1}ab}\ttr{ab} \\
& \qquad - \ttr{bwab(ab^2)^kab(ab^2)^h}\\
&= \ttr{bw(ab^2)^k(ab^2)^hab}\ttr{ab}^3 - \ttr{bwb(ab^2)^{k+h-1}ab}\ttr{ab}^2\\
& \qquad - \ttr{bwab(ab^2)^kb(ab^2)^{h-1}ab}\ttr{ab} - \ttr{bwab(ab^2)^kab(ab^2)^h}\\
&= \ttr{w(ab^2)^{k+h+1}}\ttr{ab}^3 - \ttr{wb(ab^2)^{k+h}}\ttr{ab}^2\\
& \qquad - \ttr{wab(ab^2)^kb(ab^2)^{h}}\ttr{ab} - \ttr{bwab(ab^2)^kab(ab^2)^h}.
\end{split}
\end{equation}
On the other side we have
\begin{align*}
\ttr{ab^2w(ab^2)^{k+h+2}}
&= \ttr{ab^2w(ab^2)^{k+h}}\ttr{(ab^2)^2}- \ttr{ab^2w(ab^2)^{k+h-2}}\\
&= \ttr{w(ab^2)^{k+h+1}}\ttr{(ab^2)^2}- \ttr{w(ab^2)^{k+h-1}}.
\end{align*}
It is enough to show that $\ttr{w(ab^2)^{k+h-1}}< \ttr{wb(ab^2)^{k+h}}\ttr{ab}^2$ and that $\ttr{ab}^3<\ttr{(ab^2)^2}$.
If $k+h\geq 1$ the first inequality is clear. If $k=h=0$, it amounts to
\[
\ttr{wb^{-1}(ab)^{-1}}<\ttr{wb}\ttr{ab}^2,
\]
or equivalently
\[
\ttr{wb^{-1}}\ttr{ab}<\ttr{wb}\ttr{ab}^2+\ttr{wb^{-1}ab},
\]
which holds, since $w$ is empty or ends with $b$.
We now show $\ttr{ab}^3<\ttr{(ab^2)^2}$. We have
\begin{equation*}
\ttr{ab}^3 = \ttr{b^2}^3
= \ttr{b^2}\bigl(\ttr{b^4}+2\bigr)
= \ttr{b^6}+3\ttr{b^2}.
\end{equation*}
Let $\ttr{b} - \ttr{a}=\Delta \ge1$; then
\begin{multline*}
\ttr{(ab^2)^2} = \ttr{ab^2}^2-2
= \bigl(\ttr{ab}\ttr{b}-\ttr{a}\bigr)^2-2
= \bigl(\ttr{b^2}\ttr{b}-\ttr{a}\bigr)^2-2\\
= \bigl(\ttr{b^3}+\Delta\bigr)^2-2
= \ttr{b^3}^2-2 +2\ttr{b^3}\Delta + \Delta^2
= \ttr{b^6}+2\ttr{b^3}\Delta + \Delta^2.
\end{multline*}
We thus have to show $2\ttr{b^3}\Delta + \Delta^2> 3\ttr{b^2}$,
and it is enough
to prove the case $\Delta =1$,
namely $2\ttr{b^3}-3\ttr{b^2}+1>0$.
We compute
\begin{align*}
2\ttr{b^3}-3\ttr{b^2}+1 &= 2T_3\bigl(\ttr{b}\bigr)-
3T_2\bigl(\ttr{b}\bigr)+1 \\
& = 2\ttr{b}^3 -3\ttr{b}^2 -6\ttr{b} + 7.
\end{align*}
Since the polynomial $2x^3-3x^2-6x+7$ has three real roots, all of them strictly less than $3$, and the trace of $B$ is at least $3$, the desired inequality follows.
\end{proof}
\begin{lemma}
Let\label{5D} $w$ be a word that is empty or begins with $a$, and let $s\geq 0$.
Then we have:
\begin{enumerate}
\item $\ttr{wab^4(ab^2)^sab^3}<\ttr{w(ab^2)^{s+3}}$;
\item $\ttr{wab^3(ab^2)^sab^4}<\ttr{w(ab^2)^{s+3}}$.
\end{enumerate}
\end{lemma}
\begin{proof}
We prove (1). On the left side we have
\begin{align*}
\ttr{wab^4&(ab^2)^sab^3}\\
& = \ttr{wab^2(ab^2)^sab^3}\ttr{b^2} - \ttr{wa(ab^2)^sab^3}\\
& = \ttr{wab^2(ab^2)^sab^2}\ttr{b}\ttr{b^2} - \ttr{wab^2(ab^2)^sab}\ttr{b^2} - \ttr{wa(ab^2)^sab^3}\\
& = \ttr{w(ab^2)^{s+2}}\ttr{b}\ttr{ab} - \ttr{w(ab^2)^{s+1}ab}\ttr{b^2}- \ttr{wa(ab^2)^{s+1}b}\\
& = \ttr{w(ab^2)^{s+2}}\ttr{ab^2} + \ttr{w(ab^2)^{s+2}}\ttr{a} \\
&\qquad - \ttr{w(ab^2)^{s+1}ab}\ttr{b^2}- \ttr{wa(ab^2)^{s+1}b},
\end{align*}
while on the other side we have
\[
\ttr{w(ab^2)^{s+3}} = \ttr{w(ab^2)^{s+2}}\ttr{ab^2} - \ttr{w(ab^2)^{s+1}}.
\]
We obtain $\ttr{w(ab^2)^{s+1}} < \ttr{wa(ab^2)^{s+1}b}$
from Lemma~\ref{ref13}(2). We complete the proof by computing
\begin{align*}
\ttr{w&(ab^2)^{s+1}ab}\ttr{b^2} - \ttr{w(ab^2)^{s+2}}\ttr{a} \\
& = \ttr{w(ab^2)^{s+1}ab}\ttr{ab} - \ttr{w(ab^2)^{s+1}abab}- \ttr{w(ab^2)^{s+1}aba^{-1}b}\\
& = \ttr{w(ab^2)^{s+1}abab} + \ttr{w(ab^2)^{s+1}}- \ttr{w(ab^2)^{s+1}abab}- \ttr{w(ab^2)^{s+1}aba^{-1}b}\\
& = \ttr{w(ab^2)^{s+1}} - \ttr{w(ab^2)^{s+1}aba^{-1}b}\\
& = \ttr{w(ab^2)^{s+1}} - \ttr{w(ab^2)^{s+1}ab}\ttr{a^{-1}b} + \ttr{w(ab^2)^{s+1}a^2},
\end{align*}
whose end result is positive since $\ttr{a^{-1}b}<0$.
(2) can be obtained from (1) as in the proof of Lemma~\ref{5B}.
\end{proof}
\begin{lemma}
Let\label{5E} $w$ be a word that is empty or ends with $b$, and let $k,h\geq 0$.
Then we have
\[
\ttr{wab^4(ab^2)^kab^4(ab^2)^hab^4} < \ttr{w(ab^2)^{k+h+5}}.
\]
\end{lemma}
\begin{proof}
On the left side we have
\begin{align*}
&\ttr{wab^4(ab^2)^kab^4(ab^2)^hab^4}\\
& = \ttr{wab^4(ab^2)^kab^4(ab^2)^hab^2}\ttr{b^2} - \ttr{wab^4(ab^2)^kab^4(ab^2)^ha}\\
& = \ttr{wab^4(ab^2)^kab^2(ab^2)^hab^2}\ttr{b^2}^2
- \ttr{wab^4(ab^2)^ka(ab^2)^hab^2}\ttr{b^2}\\
& \qquad -\ttr{wab^4(ab^2)^kab^4(ab^2)^ha}\\
& = \ttr{wab^2(ab^2)^kab^2(ab^2)^hab^2}\ttr{b^2}^3
- \ttr{wa(ab^2)^kab^2(ab^2)^hab^2}\ttr{b^2}^2\\
& \qquad - \ttr{wab^4(ab^2)^ka(ab^2)^hab^2}\ttr{b^2} -\ttr{wab^4(ab^2)^kab^4(ab^2)^ha}\\
& = \ttr{w(ab^2)^{k+h+3}}\ttr{b^2}^3
- \ttr{wa(ab^2)^{k+h+2}}\ttr{b^2}^2 \\
& \qquad - \ttr{wab^4(ab^2)^ka(ab^2)^{h+1}}\ttr{b^2} -\ttr{wab^4(ab^2)^kab^4(ab^2)^ha}.
\end{align*}
On the other side we have
\[
\ttr{w(ab^2)^{k+h+5}} = \ttr{w(ab^2)^{k+h+3}}\ttr{(ab^2)^2}
- \ttr{w(ab^2)^{k+h+1}}.
\]
The second end result is strictly greater than the first, because
$\ttr{w(ab^2)^{k+h+1}}<\ttr{wa(ab^2)^{k+h+2}}\ttr{b^2}^2$ by Lemma~\ref{ref13}(2), and $\ttr{b^2}^3=\ttr{ab}^3<\ttr{(ab^2)^2}$ by the proof of Lemma~\ref{5C}.
\end{proof}
\begin{lemma}
Let\label{5F} $w$ be a word that begins with $ab$ or is empty, and let $k,h\geq 0$. Then
\[
\ttr{wab^3(ab^2)^kab^3(ab^2)^hab^3}<\ttr{w(ab^2)^{k+h+4}}.
\]
\end{lemma}
\begin{proof}
This time we work on the term on the right hand side; the step in the middle of the following identity chain results from $\ttr{(ab)b^2}+\ttr{(ab)b^{-2}}=\ttr{ab}\ttr{b^2}=\ttr{ab}^2$.
\begin{align*}
\ttr{w(a&b^2)^{k+h+4}} \\
& = \ttr{wab(ba)bb(ab^2)^{k+h+2}} \\
& = \ttr{wab^3(ab^2)^{k+h+2}}\ttr{ba} - \ttr{waba^{-1}b(ab^2)^{k+h+2}}\\
& = \ttr{wab^3(ab^2)^k ab^2 (ab)b (ab^2)^h}\ttr{ab}- \ttr{waba^{-1}b(ab^2)^{k+h+2}}\\
& = \ttr{wab^3(ab^2)^k ab^3 (ab^2)^h}\ttr{ab}^2 - \ttr{wab^3(ab^2)^k aba^{-1}b (ab^2)^h}\ttr{ab}\\
& \qquad - \ttr{waba^{-1}b(ab^2)^{k+h+2}}\\
& = \ttr{wab^3(ab^2)^k ab^3 (ab^2)^h}\bigl(\ttr{ab^3}+\ttr{ab^{-1}}\bigr) \\
& \qquad - \ttr{wab^3(ab^2)^k aba^{-1}b (ab^2)^h}\ttr{ab} - \ttr{waba^{-1}b(ab^2)^{k+h+2}}\\
& = \ttr{wab^3(ab^2)^k ab^3 (ab^2)^hab^3} + \ttr{wab^3(ab^2)^k ab^3 (ab^2)^{h-1}ab^{-1}a^{-1}}\\
& \qquad +\ttr{wab^3(ab^2)^k ab^3 (ab^2)^h}\ttr{ab^{-1}} \\
& \qquad - \ttr{wab^3(ab^2)^k aba^{-1}b (ab^2)^h}\ttr{ab} - \ttr{waba^{-1}b(ab^2)^{k+h+2}}.
\end{align*}
Now, the first summand of the end result is the left side of the desired inequality, and the second is positive due to our hypotheses on $w$. We develop the fourth summand, in order to make the third appear:
\begin{align*}
\ttr{wab^3&(ab^2)^k aba^{-1}b (ab^2)^h}\ttr{ab}\\
& = \ttr{wab^3(ab^2)^k ab (ab^2)^h}\ttr{ab}\ttr{a^{-1}b}
-\ttr{wab^3(ab^2)^k a^2 (ab^2)^h}\ttr{ab}\\
& = \ttr{wab^3(ab^2)^k ab (ab^2)^h}\ttr{b^2}\ttr{a^{-1}b}
-\ttr{wab^3(ab^2)^k a^2 (ab^2)^h}\ttr{ab}\\
& = \ttr{wab^3(ab^2)^k ab^3 (ab^2)^h}\ttr{a^{-1}b}
+\ttr{wab^3(ab^2)^k ab^{-1} (ab^2)^h}\ttr{a^{-1}b}\\
& \qquad -\ttr{wab^3(ab^2)^k a^2 (ab^2)^h}\ttr{ab} \\
& = \ttr{wab^3(ab^2)^k ab^3 (ab^2)^h}\ttr{ab^{-1}} \\
& \qquad +\ttr{wab^3(ab^2)^k a (ab^2)^h}\ttr{a^{-1}b}\ttr{b} - \ttr{wab^3(ab^2)^k ab (ab^2)^h}\ttr{a^{-1}b}\\
& \qquad - \ttr{wab^3(ab^2)^k a^2 (ab^2)^h}\ttr{ab}.
\end{align*}
We are then left with proving that the sum
\begin{multline*}
-\ttr{wab^3(ab^2)^k a (ab^2)^h}\ttr{a^{-1}b}\ttr{b} + \ttr{wab^3(ab^2)^k ab (ab^2)^h}\ttr{a^{-1}b}\\
+\ttr{wab^3(ab^2)^k a^2 (ab^2)^h}\ttr{ab}
-\ttr{wab(ab^2)^{k+h+2}}\ttr{a^{-1}b}+ \ttr{wa^2(ab^2)^{k+h+2}}
\end{multline*}
is positive. The first, third, and last summand surely are, and so is the sum or the second with the fourth, because
\[
\ttr{wab^3(ab^2)^k ab (ab^2)^h} < \ttr{w(ab^2)^{k+h+2}}
<\ttr{wab(ab^2)^{k+h+2}},
\]
by Lemma~\ref{5A}(2) and Lemma~\ref{ref13}(2).
\end{proof}
We finally arrive at~\S\ref{ref1}(IV.2).
\begin{theorem}
Let\label{ref16} $A,B\in\SL_2\mathbb{Z}_{\ge0}$, and assume $\mathrm{tr}(A)<\mathrm{tr}(B)$ and $\mathrm{tr}(AB)=\mathrm{tr}(B^2)$. The $ab^2$ is the only optimal word.
\end{theorem}
\begin{proof}
Let $u$ be a word which is trace-maximizing among words of the same length; by the remarks following Equation~\eqref{eq1} we may assume, possibly replacing $u$ with its cube, that this length is a multiple of $3$.
We have to prove that $u$ is a power of a rotation of $ab^2$.
By Lemma \ref{4-5-6}, $u$ does not contain $a^2$ as a factor, up to rotations. We have
\begin{align*}
\ttr{wab^5}= & \ttr{wab^2}\ttr{b^3}-\ttr{wab^{-1}}\\
&= \ttr{wab^2}\bigl(\ttr{bb}\ttr{b}-\ttr{b}\bigr)-\ttr{wab^{-1}}\\
&= \ttr{wab^2}\bigl(\ttr{ab}\ttr{b}-\ttr{b}\bigr)-\ttr{wab^{-1}}\\
&\le \ttr{wab^2}\bigl(\ttr{ab}\ttr{b}-\ttr{a}-1\bigr)-\ttr{wab^{-1}}\\
&= \ttr{wab^2}\bigl(\ttr{ab^2}-1\bigr)-\ttr{wab^{-1}}\\
&= \ttr{w(ab^2)^2}+\ttr{w} -\ttr{wab^2}-\ttr{wab^{-1}}\\
&= \ttr{w(ab^2)^2}+\ttr{w}-\ttr{wab^2} -\ttr{wa}\ttr{b}+\ttr{wab}\\
&< \ttr{w(ab^2)^2},
\end{align*}
and therefore $b^5$ is also excluded. Moreover
\[
\ttr{b^3} = \ttr{bb}\ttr{b}-\ttr{b}
= \ttr{ab}\ttr{b}-\ttr{b}
= \ttr{ab^2} +\ttr{a}-\ttr{b}
< \ttr{ab^2}
\]
shows that $u$ is not a power of $b$. Summing up, $u$ uniquely factorizes as a product of $ab$, $ab^2$, $ab^3$, and $ab^4$;
we refer to this as the \newword{syllabic decomposition} of $u$.
Suppose that $ab$ occurs as a syllable (occurrences are always intended up to rotations). Since the length is a multiple of $3$, at least one of the following cases must hold:
\begin{itemize}
\item one of $ab^3$ and $ab^4$ occurs as well,
\item $ab$ occurs at least thrice.
\end{itemize}
These occurrences will be separated by zero or more occurrences of $ab^2$. In any case, Lemmas~\ref{5A}, \ref{5B}, and~\ref{5C} apply, and $u$ is not trace-maximizing. Thus the syllable $ab$ does not occur in $u$.
Suppose $ab^4$ occurs. Then
\begin{itemize}
\item either $ab^3$ occurs as well,
\item or $ab^4$ occurs at least thrice.
\end{itemize}
Lemmas \ref{5D} and \ref{5E} treat these cases, and thus exclude $ab^4$.
We have established that the only syllables occurring in $u$ are $ab^2$ and $ab^3$. If the latter occurs,
it must do so at least thrice, and Lemma \ref{5F} applies.
We are then left with occurrences of $ab^2$ only, and the proof is complete.
\end{proof}
\section{Case (IV.3)}\label{ref17}
The standing assumptions in this final section are that $A,B\in\SL_2\mathbb{Z}$ satisfy $2\le\mathrm{tr}(A)<\mathrm{tr}(B)$ and $\mathrm{tr}(AB)>\mathrm{tr}(B^2)$.
First, a dichotomy.
\begin{lemma}\label{dichotomy}
Under the above assumptions the numbers $\ttr{(ab)^3}$ and $\ttr{(ab^2)^2}$ differ by at least $2$.
\end{lemma}
\begin{proof}
We work modulo $\ttr{ab}$. We have
\[
\ttr{(ab)^3} \equiv \ttr{(ab)^2}\ttr{ab}-\ttr{ab}
\equiv 0,
\]
and also
\[
\ttr{(ab(ba)b^2} \equiv \ttr{ab^3}\ttr{ba} - \ttr{aba^{-1}b}
\equiv -\ttr{ab}\ttr{a^{-1}b} +\ttr{a^2}
\equiv \ttr{a^2}.
\]
Taking the difference we get
\[
\ttr{(ab^2)^2}-\ttr{(ab)^3} \equiv \ttr{a^2}.
\]
Now, the coset $\ttr{a^2}+\mathbb Z\ttr{ab}$ does not intersect the interval $\set{-1,0,1}$. Indeed, its points closest to zero are $\ttr{a^2}\ge2$ and $\ttr{a^2}-\ttr{ab}\le \ttr{a^2}-\ttr{b^2}-1\le -2$.
\end{proof}
The case $\ttr{(ab)^3}>\ttr{(ab^2)^2}$ of the dichotomy follows familiar patterns.
\begin{lemma}
Assume\label{61} $\ttr{(ab^2)^2}\le \ttr{(ab)^3}-2$.
Fix a word $w$, and let $k\geq 1$ be such that the length of
$wab^2(ab)^kab^2$ is a multiple of $6$. If $w$ is empty, or is a product of $ab$ and $ab^2$, then we have
\[
\ttr{wab^2(ab)^kab^2}<\ttr{w(ab)^{k+3}}.
\]
\end{lemma}
\begin{proof}
On the left side we have
\begin{align*}
&\ttr{wab^2(ab)^kab^2}\\
&= \ttr{wab^2(ab)^k}\ttr{ab^2} - \ttr{wab^2(ab)^{k-1}ab^{-1}a^{-1}}\\
&=\ttr{w(ab)^k}\ttr{ab^2}^2 - \ttr{wb^{-1}(ab)^{k-1}}\ttr{ab^2} - \ttr{wab^2(ab)^{k-1}ab^{-1}a^{-1}}\\
&=\ttr{w(ab)^k}\bigl(\ttr{(ab^2)^2} + 2\bigr) - \ttr{wb^{-1}(ab)^{k-1}}\ttr{ab^2} - \ttr{wab^2(ab)^{k-1}ab^{-1}a^{-1}}\\
&\le \ttr{w(ab)^k}\ttr{(ab)^3} - \ttr{wb^{-1}(ab)^{k-1}}\ttr{ab^2} - \ttr{wab^2(ab)^{k-1}ab^{-1}a^{-1}}\\
&= \ttr{w(ab)^{k+3}} + \ttr{w(ab)^{k-3}} - \ttr{wb^{-1}(ab)^{k-1}}\ttr{ab^2} - \ttr{wab^2(ab)^{k-1}ab^{-1}a^{-1}}.
\end{align*}
We then have to prove that the sum of all but the first summands
of the above end result is negative. We will prove this fact by showing the following two inequalities:
\begin{align*}
\ttr{w(ab)^{k-3}} &\le \ttr{wb^{-1}(ab)^{k-1}}\ttr{ab^2},\\
\ttr{wab^2(ab)^{k-1}ab^{-1}a^{-1}}&>0.
\end{align*}
Suppose $k\geq 3$.
If $w$ is non empty, it begins with $ab$ and ends with $b$.
Thus all the inverted letters simplify and the inequalities follow from Lemma~\ref{ref13}(2).
This also holds when $w$ is the empty word, with different simplifications.
Suppose $k<3$. Then $w$ cannot be empty, for reasons of length, and the second inequality causes no problems.
If $k=2$ the first inequality becomes $\ttr{wb^{-1}a^{-1}}\le \ttr{wb^{-1}ab}\ttr{ab^2}$, clearly true.
If $k=1$ it becomes
$\ttr{w(ab)^{-2}}\le \ttr{wb^{-1}}\ttr{ab^2}$, which is also true since $\ttr{w(ab)^{-2}}=\ttr{wb^{-1}a^{-1}}\ttr{ab}-\ttr{w}$.
\end{proof}
The other case $\ttr{(ab)^3}<\ttr{(ab^2)^2}$ will be treated via the following lemma.
\begin{lemma}
Assume\label{62} $\ttr{(ab)^3}\le \ttr{(ab^2)^2}-2$.
Fix a word $w$, and let $k,h\geq 0$ be such that the length of
$ab^2wab(ab^2)^kab(ab^2)^hab$ is a multiple of $6$. If $w$ is empty, or $w$ is a product of $ab$ and $ab^2$, then we have
\[
\ttr{ab^2wab(ab^2)^kab(ab^2)^hab} <
\ttr{ab^2w(ab^2)^{k+h+2}}.
\]
\end{lemma}
Unfortunately, although the end inequality is the same, the proof of Lemma~\ref{62} is harder than that of its twin Lemma~\ref{5C}.
This is due to the fact that the final step in the proof of Lemma~\ref{62}, namely the inequality
$\ttr{ab}^3\le\ttr{(ab^2)^2}$, may fail; for example, it fails for the matrices $A=L^{11}$, $B=LNL$ cited in~\S\ref{ref12}. We have thus to make do with the weaker $\ttr{ab}^3\le\ttr{(ab^2)^2}+\ttr{ab}$, that can still be quite narrow: for the above matrices we have
\[
\ttr{ab}^3=3375\le
3377 = 3362 + 15 =
\ttr{(ab^2)^2}+\ttr{ab}.
\]
The following lemma establishes that weaker inequality, as well as the fact that the minimal difference $\mathrm{tr}(AB)=15>14=\mathrm{tr}(B^2)$ for the above matrix pair is no coincidence.
\begin{lemma}
Assume\label{aiuto} $\ttr{(ab^2)^2}\geq \ttr{(ab)^3}+2$. Then the following formulas hold:
\begin{align}
\ttr{ab}&=\ttr{b^2}+1, \label{aiuto1}\\
2\ttr{a}&\le \ttr{b}, \label{aiuto2}\\
\ttr{ab}^3&\le \ttr{(ab^2)^2} +\ttr{ab}.\label{aiuto3}
\end{align}
\end{lemma}
\begin{proof}
Let $\alpha = \ttr{a}$, $\beta = \ttr{b}$, $x = \ttr{ab}$; then
we have $\ttr{(ab)^3}=T_3(x)=x^3-3x$ and $\ttr{(ab^2)^2}=T_2(\beta x-\alpha)=(\beta x-\alpha)^2-2$.
Assuming the negation of~\eqref{aiuto1}, we have
$x\geq\ttr{b^2}+2=\beta^2$ and thus
\begin{align*}
\ttr{(ab)^3}-\ttr{(ab^2)^2}
&= x^3 -\beta^2x^2 + (4\beta-3)x -2 \\
&\ge (4\beta-3)x -2.
\end{align*}
Since $\beta$ and $x$ are both at least $3$, the last term is positive, which is a contradiction;
this establishes~\eqref{aiuto1}.
Applying~\eqref{aiuto1} and its equivalent form $\ttr{b}^2=\ttr{ab}+1$ several times, we compute
\begin{align*}
\ttr{(ab^2)^2} &= \ttr{ab^2ab}\ttr{b} - \ttr{a^2b^2}\\
& = \ttr{abab}\ttr{b}^2 - \ttr{a^2b}\ttr{b} -\ttr{a^2b}\ttr{b} +\ttr{a^2}\\
& = \ttr{(ab)^2}\bigl(\ttr{ab}+1\bigr) -2\bigl(\ttr{a}\ttr{ab}\ttr{b}-\ttr{b}^2\bigr)+\ttr{a}^2-2\\
& = \ttr{(ab)^3} + \ttr{ab} + \ttr{(ab)^2} -2\ttr{a}\ttr{ab}\ttr{b} +2\ttr{b}^2+\ttr{a}^2-2\\
& = \ttr{(ab)^3} + \beta^2-1 + (\beta^2-1)^2-2 -2\alpha\beta(\beta^2-1) +2\beta^2 + \alpha^2-2\\
& = \ttr{(ab)^3} + \beta^4 -2\alpha\beta^3 +\beta^2 +2\alpha\beta +\alpha^2-4;
\end{align*}
therefore our hypothesis yield
\[
2\alpha\beta^3-\beta^4 \le
\beta^2 +2\alpha\beta +\alpha^2 -6.
\]
We change variables by setting $\alpha=\lambda\beta$, and obtain
\begin{align*}
(2\lambda -1)\beta^4 &\le (1+2\lambda +\lambda^2)\beta^2-6,\\
(2\lambda -1)\beta &< \frac{(1+\lambda)^2}{\beta},\\
2\lambda\beta &< \beta + \frac{(1+\lambda)^2}{\beta},\\
2\alpha &\le \beta + \biggl \lfloor \frac{(1+\lambda)^2}{\beta} \biggr \rfloor,
\end{align*}
the last step justified by the fact that $2\lambda\beta=2\alpha$ is an integer.
If $\beta=3$ then $\alpha=2$, $\lambda=2/3$, and the last inequality means $4\le 3$. Therefore $\beta \geq 4$ and the floor part is zero since $1+\lambda <2$; this proves~\eqref{aiuto2}.
We previously computed that
\[
\ttr{(ab^2)^2}-\ttr{(ab)^3} = \beta^4 -2\alpha\beta^3 +\beta^2 +2\alpha\beta +\alpha^2-4.
\]
Since $\ttr{ab}^3=\ttr{(ab)^3}+3\ttr{ab}$ and $\ttr{ab}=\beta^2-1$ we obtain
\begin{align*}
\ttr{ab}^3 &= \ttr{(ab^2)^2} -\beta^4 +2\alpha\beta^3 -\beta^2 -2\alpha\beta -\alpha^2+4 + 3\beta^2-3\\
&= \ttr{(ab^2)^2} -\beta^4 +2\alpha\beta^3 +2\beta^2 -2\alpha\beta -\alpha^2+1 \\
&= \ttr{(ab^2)^2} + (2\alpha-\beta)(\beta^3-\beta) +\beta^2 -\alpha^2+1\\
&\le \ttr{(ab^2)^2} +\beta^2 -\alpha^2+1\\
&\le \ttr{(ab^2)^2} +\ttr{ab},
\end{align*}
thus settling~\eqref{aiuto3}.
\end{proof}
\begin{proof}[Proof of Lemma~\ref{62}]
Continuing from~\eqref{eq5} (whose proof does not depend on the relative values of $\ttr{a}$, $\ttr{b}$, $\ttr{sb}$) and applying~\eqref{aiuto3}, we obtain
\begin{align*}
\ttr{ab^2wab(ab^2)^kab(ab^2)^hab}
&= \ttr{w(ab^2)^{k+h+1}}\ttr{ab}^3
-\ttr{wb(ab^2)^{k+h}}\ttr{ab}^2\\
& \qquad - \ttr{wab(ab^2)^kb(ab^2)^{h}}\ttr{ab}
- \ttr{bwab(ab^2)^kab(ab^2)^h}\\
&\le \ttr{w(ab^2)^{k+h+1}}\ttr{(ab^2)^2} \\
&\qquad + \ttr{w(ab^2)^{k+h+1}}\ttr{ab} - \ttr{wb(ab^2)^{k+h}}\ttr{ab}^2\\
& \qquad - \ttr{wab(ab^2)^kb(ab^2)^{h}}\ttr{ab} - \ttr{bwab(ab^2)^kab(ab^2)^h}\\
&= \ttr{w(ab^2)^{k+h+3}} + \ttr{w(ab^2)^{k+h-1}}\\
&\qquad + \ttr{w(ab^2)^{k+h+1}}\ttr{ab} - \ttr{wb(ab^2)^{k+h}}\ttr{ab}^2\\
& \qquad - \ttr{wab(ab^2)^kb(ab^2)^{h}}\ttr{ab} - \ttr{bwab(ab^2)^kab(ab^2)^h}.
\end{align*}
Thus we need to prove that
\begin{equation}\label{666}
\begin{split}
\ttr{w(ab^2)^{k+h-1}} &+ \ttr{w(ab^2)^{k+h+1}}\ttr{ab}\\
< \ttr{wb(ab^2)^{k+h}}\ttr{ab}^2 + \ttr{wab(ab^2)^k&b(ab^2)^{h}}\ttr{ab} + \ttr{bwab(ab^2)^kab(ab^2)^h}.
\end{split}
\end{equation}
We will need to switch the positions of some factors, and this will be accomplished by the formula
\begin{equation}\label{eq6}
\ttr{xyzw} = \ttr{xzyw} + \ttr{xz^{-1}yw} - \ttr{xyz^{-1}w},
\end{equation}
which moves $z$ to the left; of course, an analogous formula holds for moving to the right. Formula~\eqref{eq6} follows from Lemma~\ref{ref5}(2), since both $\ttr{xyzw} + \ttr{xyz^{-1}w}$ and
$\ttr{xzyw}+\ttr{xz^{-1}yw}$ equal $\ttr{xyw}\ttr{z}$.
In order to obtain~\eqref{666}, we work on the second summand of the second line. Moving $b$ to the left we get
\begin{align*}
\ttr{wab(ab^2)^kb(ab^2)^{h}}\ttr{ab}
&= \ttr{w(ab^2)^{k+h+1}}\ttr{ab} + \ttr{wa(ab^2)^{k+h}}\ttr{ab}\\
&\quad -\ttr{wab(ab^2)^{k-1}ab(ab^2)^{h}}\ttr{ab}\\
&= \ttr{w(ab^2)^{k+h+1}}\ttr{ab} + \ttr{wa(ab^2)^{k+h}}\ttr{ab} \\
&\qquad - \ttr{wab(ab^2)^{k-1}abab(ab^2)^{h}} - \ttr{wab(ab^2)^{k+h-1}}.
\end{align*}
Substituting this back into~\eqref{666}, the summand
$\ttr{w(ab^2)^{k+h+1}}\ttr{ab}$ simplifies.
The summand $\ttr{wab(ab^2)^{k-1}(ab)^2(ab^2)^{h}}$ is, by
Lemma~\ref{ref13}(2), always dominated by the last term
of~\eqref{666}, even when $k=h=0$.
Removing both of them we remain with the inequality
\begin{multline*}
\ttr{w(ab^2)^{k+h-1}} + \ttr{wab(ab^2)^{k+h-1}} \\
< \ttr{wb(ab^2)^{k+h}}\ttr{ab}\ttr{ab} +
\ttr{wa(ab^2)^{k+h}}\ttr{ab},
\end{multline*}
which holds by Lemma~\ref{ref13}(2). The case $k=h=0$ must be checked apart, but causes no problems.
\end{proof}
\begin{theorem}
Let\label{ref18} $A,B\in\SL_2\mathbb{Z}_{\ge0}$, and assume $\mathrm{tr}(A)<\mathrm{tr}(B)$ and $\mathrm{tr}(AB)>\mathrm{tr}(B^2)$. If $\mathrm{tr}((AB)^3)>\mathrm{tr}((AB^2)^2)$,
then $ab$ is the only optimal word; otherwise, so is $ab^2$
\end{theorem}
\begin{proof}
As in the proof of Theorem~\ref{ref16}, let $u$ be a word of length a multiple of~$6$ which is trace-maximizing among all words of the same length. The statement will result by proving that $u$ is a power of $ab$ (in case $\ttr{(ab)^3}>\ttr{(ab^2)^2}$), or of $ab^2$ (in case $\ttr{(ab)^3}<\ttr{(ab^2)^2}$, no equality being possible by Lemma~\ref{dichotomy}).
Since $\ttr{b^2}<\ttr{ab}$, at least one $a$ appears in $u$, but $a^2$ does not by Lemma \ref{4-5-6}.
We claim that the factor $b^3$ is also excluded.
Indeed, for every $w$ we have
\begin{multline*}
\ttr{wab^3}=\ttr{wab}\ttr{b^2}-\ttr{wab^{-1}}
\le\ttr{wab}\ttr{ab}-\ttr{wab}-\ttr{wab^{-1}}\\
=\ttr{w(ab)^2}+\ttr{w}-\ttr{wa}\ttr{b}<\ttr{w(ab)^2}.
\end{multline*}
Therefore, $u$ factors uniquely as a product of the syllables
$ab$ and $ab^2$.
Suppose that both syllables appear; since $\abs{u}$ is a multiple
of~$6$, $ab^2$ must appear at least two times, and $ab$ at least three times. If $ab^2\prec ab$ then Lemma~\ref{61} contradicts the maximality of $u$, and the same does Lemma~\ref{62} if $ab\prec ab^2$. Thus only one syllable appears in $u$, and the proof is complete.
\end{proof}
|
\section{Introduction}
The analysis of the behavior of museum visitors has a long-standing tradition~\cite{robinson1928book} and grows daily in importance as tourist flows increase and digital technologies get ubiquitous~\cite{falk2016book, yalowitz2009}.
The outstanding issue of visitors management demands for multidisciplinary skills connected to, among others, psychology, computer science, statistics, physics of complex systems as well as modeling and optimization theory.
Museums curators are expected to achieve three complex and seemingly
contradictory objectives: increasing the visitors number, enhancing the experience quality, and preserving the artworks~\cite{yoshimura2014}.
Accurately measuring and analyzing the visitors trajectories is an essential component towards these objectives and, specifically, when aiming at an efficient organization of the exhibits \cite{bourdeau2001, trondle2014}, the determination of adequate ticketing strategies, and also to verify if visitors experience complies with managers' intents~\cite{tzortzi2014}.
\medskip
A complete workflow enabling the full control of visitors in a museum consists of several challenging steps, that we here summarize.
\begin{description}[labelindent=\parindent,leftmargin=0cm]%
\item[Visitors tracking] the first goal is to understand the behavior of visitors in terms of paths followed in the museum. Not all museums have predefined paths and sometimes more than one choice is possible \cite{lanir2017}. Moreover, in large museums, it is rare that visitors see the whole exhibition \cite{piccialli2020b}.
A number of technologies exist for indoor tracking that is characterized by a trade-off between deployment complexity, invasiveness and accuracy.
Radio-based approaches, as considered in this work, enable room-level positioning accuracy: visitors trajectories are rendered into sequences of visited rooms and related permanence times.
At the price of more invasive and complex deployments, sometimes impossible in the context of cultural heritage, centimeter-level individual positioning can be also accomplished, e.g.\ via distributed grids of 3D scanners or video-cameras.
\\
Besides, psychological and sociological variables can be observed on side of paths, such as heart rate, skin conductance, emotional and aesthetic evaluations of specific artworks~\cite{kirchberg2015, trondle2014, kirchberg2014}, interactions with groupmates, degree of attention, boredom or fatigue. \\
Automatic systems could be complemented with manual activities, like paper-and-pencil annotations and questionnaires \cite{klein1993, mokatren2019, trondle2014}.
From questionnaires, one can estimate demographic related and museum visit related features \cite{mokatren2019} like age, gender, educational level, number of visits per year to museums, etc.
After the visit, one can measure the degree of satisfaction, the relationship between perceived and real time spent in the museum \cite{bollo2005}, etc.
\item[Behavior understanding]
a number of variables can be estimated from visitors trajectories: busy hours, movement patterns, length of visits, permanence times in each room, number of stops.
Two indicators are generally considered to quantify the importance of a specific exhibit, the \emph{attraction power} (relative amount of people who has stopped in front of an artwork during their visit) and the \emph{holding power} (average time spent in front of an artwork) \cite{lanir2017}.
\\
Clustering and AI-based algorithms can be used for inferring, from the whole trajectories dataset, the \emph{typical paths} or, equivalently, the \emph{typical individual behaviors} inside the museum.
Another interesting question regards the predictability of visitors behaviors \cite{cuomo2016, kuflik2012, martella2017}: can a person who starts visiting the museum in a certain manner be immediately labeled as a visitor of a certain type? \\
\emph{Social behavior} can be observed too. For example, one can wonder, e.g., if people belonging to the same group follow the same path or they split, or whether individuals are attracted or repelled by crowding.
\item[Museums digital twin]
once statistics about visitors trajectories and behavior are available, it is possible to create an algorithm capable of generating real-like visits paths in the museum \cite{balzotti2018}.
This is done by reproducing the movements of people from one room to another, duly determining their transition probability.
Moreover, herding behavior in social groups or the response to congestion and fatigue could be taken into account. A digital twin is able to reproduce virtual visitors moving in the museum with a realistic behavior, possibly in new (i.e.\ unexperienced) conditions. It is also possible to forecast the visitors flow from some initial conditions, like, e.g.\ the visitor inflow at a given time.
\item[Visitors flow optimization]
in order to use the museum digital twin as managing tool, curators and organizers have to identify relevant \emph{control variables} and \emph{objectives}: regarding the former, one can, e.g., regulate the entrance flows, limit the maximum occupancy of selected rooms, increase the number of entrances or exits, set a maximal duration of the visit. The ticket price can obviously be controlled too. \\
Regarding objectives instead, one can aim at maximizing the number of visitors, the pleasantness of the visit, the amount of information conveyed, or keeping the environmental parameters (e.g., temperature and humidity) in a given range, for best conservation of the collection.
\\
Once this is done, a museum digital twin can be profitably used to simulate different scenarios, aiming at matching the objectives while varying the control variables. Here optimization algorithms like gradient-based or PSO methods can be used to automatize the search for a solution.
\end{description}
\subsection{Relevant literature}\label{sec:refs}
The first step (Visitors tracking) is the one that has received the most attention in the literature, as it relates to pedestrian dynamics in general, i.e.\ beyond the museum context.
Focusing on (indoor) tracking systems, all kind of technologies have been exploited, such as
RFID \cite{lanir2017},
Wi-Fi \cite{georgievska2019, hong2018},
Bluetooth \cite{casolla2020, choi2017, versichele2012a, martella2017, oosterlinck2017, piccialli2020, piccialli2019a, pierdicca2019, piccialli2019b, versichele2012b, yoshimura2012, yoshimura2014},
video cameras \cite{lovreglio2018},
3D scanners \cite{corbetta2018, seer2014}.
An exhaustive review of these methods is out of the scope of the paper; we refer the interested reader to the papers \cite{gu2009, oosterlinck2017} for more references.
Different technologies require different degree of visitors involvement.
For example, video cameras, 3D scanners, Wi-Fi or Bluetooth mass scans require no collaboration, while Bluetooth-based apps and RFID tags usually require some degree of visitors interaction.
Measuring personal data like heart rate or skin conductance requires instead total involvement \cite{kirchberg2015, kirchberg2014}.
Moreover, convincing people to participate in an experiment, for example by downloading and installing a smartphone app, can be difficult and time-consuming \cite{pierdicca2019}.
Sometimes free tickets could yield a good incentive \cite{zancanaro2007}.
\medskip
The second step (Behavior understanding) has also been investigated in great detail in connection with museums. Regarding \emph{individual behavior}, the predominant idea is to classify visitors into four categories based on the way they interact with the artworks:
`Ants' (tend to follow a specific path and observe extensively almost all the exhibits);
`Butterflies' (do not follow a specific path but are guided by the physical orientation of the exhibits; stop frequently to acquire more information);
`Fish' (most of the time move around in the center of the room and usually avoid looking at exhibits details);
`Grasshoppers' (seem to have a specific preference for some pre-selected exhibits and focus their time on them, while tending to ignore the rest),
see, e.g.,~\cite{kuflik2012} or~\cite{veron1989book} for the origin of this taxonomy.
Regarding \emph{social behavior} instead, the idea is to label visitors in six categories based on how they interact with group mates:
`Doves' (interested in other visitors while ignoring the environment);
`Meerkats' (stand side by side, expressing great interest in the exhibits);
`Parrots' (share their attention between exhibits and group members);
`Geese' (advance together, however one visitor appears in the lead);
`Lone wolves' (enter the museum together and then separate);
`Penguins' (cross the space together while ignoring the exhibits),
see, e.g., \cite{dim2014, lanir2017}.
Clustering techniques (e.g., $k$-means, hierarchical clustering, sequence alignment) have been used to assign every visitor trajectory (spanning from room-scale to continuum) to one of the groups described above, or to some given typical movement patterns \cite{casolla2020, dim2014, versichele2012a, kuflik2012, mokatren2019, martella2017, oosterlinck2017, piccialli2019a, piccialli2019b, yoshimura2019, zancanaro2007}.
This enables one to quantify the percentage of visitors belonging to each group.
Note that typically the number of clusters is assigned \emph{a priori} and this can be an important limitation.
One crucial point for cluster investigation is the definition of a suitable \emph{metric}, to measure the distance between trajectories, and aggregate (cluster) trajectories close to each other.
Examples of such metrics devised at room-scale can be found in \cite{casolla2020, martella2017, piccialli2019a}.
In particular, \cite{casolla2020, piccialli2019a} propose a combination of well-known metrics defined in the space of characters strings (as trajectories can be suitably represented as sequences of characters), which is further corrected to take into account the differences in time of permanence in each room.
Regarding trajectory comparisons, let us mention also other two papers: \cite{yoshimura2019} compares measured trajectories with those coming from a random walk simulator in order to understand which kind of visitors exhibits stronger patterns.
The work~\cite{lanir2013} compares trajectories of visitors with and without audio-guides in order to measure the impact of the transmitted information.
\medskip
The third and fourth steps are also related to the rich pedestrian flow modeling literature:
if one considers the museum as a continuous space as in \cite{liakou2019}, one can refer to differential (agent-based, kinetic, fluid-dynamic) or nondifferential (discrete choice, cellular automata) models.
See, e.g., \cite{bellomo2011, cristiani2014book, duives2013, dong2020, eftimie2018, haghani2020, martinez-gil2017} for some reviews, books' chapters and books about this topic.
If one instead considers the museum as a graph -- where the nodes represent the rooms of the museum and the edges represent connections among rooms -- one can refer to some classical tools like transition matrices and deterministic/stochastic Markov chains with/without memory \cite{hong2018, piccialli2020, lees2016} in order to simulate a room-level walk in the museum (i.e.\ a trajectory on the graph).
Although many mathematical tools are available, examples of actual museums digital twins developed with the aim of reproducing, understanding and optimizing visitors behavior are largely missing.
This fact holds despite the fact that the path followed by visitors is evidently conditioned by the design of the exhibition galleries \cite{bourdeau2001, trondle2014}.
An interesting attempt can be found in \cite{guler2009thesis, guler2016book}: the author describes a museum simulator and uses it to show that changes in the layout design of an exhibition result in different visitor circulation patterns.
Unfortunately, that simulator can be hardly used in a museum with a very high density of artworks exposed, since it requires a complex calibration of many artwork-scale parameters which usually show a high variance between visitors.
See also \cite{balzotti2018} for a rudimentary simulator on graph and \cite{pluchino2014} for a simulator developed under the NetLogo software environment.
\subsection{Paper contributions}
In this paper, we perform an all-around investigation which includes contributions to all the four steps described above.
Covering the whole process allows us to reach an unprecedented level of understanding and control of the museum, which unleashes the capability of improving deep modifications to the ticketing strategy as well as to the museum access management.
Our results are based on real visitors data acquired in the Galleria Borghese museum (Rome, Italy).
In more details, the research unfolds along the following lines:
\begin{enumerate}[labelindent=0cm,wide]
\ite
We describe a cheap and easily reproducible data collection system, hinged on an IoT-based room-scale Lagrangian tracking system of the museum visitors.
Each visitor is provided with a portable Bluetooth Low Energy (BLE) beacon, whose signal is received by antennas (realized by means of common Raspberry Pi's) displaced in fixed positions within the museum rooms. We employ this system for an extended data collection campaign which provides the high statistics measurements employed in this work.
\ite
We employ and filter the Received Signal Strength Indicator (RSSI) of the beacons to reconstruct individual visitors trajectories.
Due to the restricted space and the numerous architectural and historical constraints, each beacon is often captured by multiple antennas at the same time.
Accurately reconstructing the trajectories in this settings defines a challenge \textit{per se}.
We propose a new machine learning approach which outperforms standard sliding window processing, especially, when it comes to estimating the correct time of permanence in rooms.
\ite
In order to get insights about visitors behavior, including the most common movement patterns, we analyze trajectories via statistical and clustering techniques.
Inspired by the Wasserstein distance\footnote{The Wasserstein distance was first introduced by Kantorovich in 1942 and then rediscovered many times. Nowadays, it is also known as $Lip'$-norm, earth mover's distance, $\bar d$-metric, Mallows distance. An important characterization is also given by the Kantorovich--Rubinstein duality theorem.}, we introduce a new \emph{ad hoc} trajectory clustering metric, which respects the geometrical properties of the museum. Our metric, in fact, builds upon the physical distance among rooms.
By using a hierarchical cluster analysis, only based on such metric (and with no \emph{a priori} hypothesis on the number of clusters nor on their size), we can unveil automatically hard-to-see movement patterns that go well beyond the standard animal-inspired classification (see Section \ref{sec:refs}).
As a by-product, we can also identify anomalous behaviors.
\ite
We employ statistical tools to build a probability transition matrix among museum rooms, which provides us with building blocks for a model capable of simulating \textit{in silico} the museum visits.
In particular, this enables us to forecast the path of visitors entering the museum from any room.
Unlike the simulator presented in~\cite{guler2009thesis, guler2016book}, our simulator leverages on the measured permanence time in each room on side of the probability of transition from one room to any other. This results in a tool easier to calibrate.
\ite
Finally, we employ the simulator to significantly increase the efficiency of the ticketing strategy and entrance/exit management. Our results suggest a way to enhance the museum fruition from both visitors and curators points of view, while keeping the numerous constraints within the limits.
\end{enumerate}
\subsection{Paper organization}
We present our methods and original contributions alongside our field activity at Galleria Borghese museum, our case study.
In Section \ref{sec:casestudy} we introduce Galleria Borghese and outline its floor plan and the current ticketing strategy.
In Section \ref{sec:trackingsystem} we describe our tracking system and the dataset we collected in the museum.
In Section \ref{sec:trajectoryreconstruction} we discuss our trajectory reconstruction methods.
In Section \ref{sec:trajectoryanalysisandclustering} we analyze the trajectories collected, fit their statistical observable with known distributions, and introduce our clustering approach.
In Section \ref{sec:model} we introduce the model which allows us to create a complete digital twin of the museum and simulate \emph{in silico} the visitors flow.
In Section \ref{sec:optimization} we employ the model to find optimal strategies for ticketing and museum management.
The discussion in Section \ref{sec:conclusions} closes the paper.
\section{Case study: the Galleria Borghese in Rome}\label{sec:casestudy}
The world-renowned Galleria Borghe\-se museum (Rome, Italy), is a relatively small, two-floor museum with 3 entrances and 21 exhibition areas.
Its sculptures and paintings attract visitors from all over the world, see Figure \ref{fig:splendoreGB}.
\begin{figure}[t]
\begin{center}
\textbf{a.} \includegraphics[width=0.438\linewidth]{GBsculture.jpg}
\textbf{b.} \includegraphics[width=0.47\linewidth]{GBpinacoteca.jpg}
\caption{The two largest rooms in Galleria Borghese. \textbf{a.} \emph{Ratto di Proserpina} located on the first floor. \textbf{b.} Main area on the second floor, part of the Pinacoteque. The map of the museum and the room names will be shown later in Figure \ref{fig:GBmap} and Table \ref{tab:matchroomsantennas}.}
\label{fig:splendoreGB}
\end{center}
\end{figure}
On the main floor, the exhibition area is circular, while on the second floor (Pinacoteque) it is U-shaped.
Rooms are numbered but no obligatory exhibition path is assigned, so many people do not visit the rooms in their natural order.
Moreover, the density of exhibits is so high that people often come back to already visited rooms multiple times to admire artworks missed during the previous passages.
Congestion is frequent in some rooms, like the one which hosts Caravaggio's paintings.
Audio-guides are available on-demand and guided tours are allowed but subject to quota (both in number and size).
To cope with the many historical, artistic and architectural constraints, museum curators established to schedule the visits: tickets must be booked in advance and give access to the museum for a slot of 2h.
Five slots per day are granted. The maximum number of visitors allowed in each slot is 360. Additionally, 30 tickets, called ``last-minute'', are sold 30 minutes after the beginning of each time slot.
People can also decide which floor to start the visit from, within some limits.
At the end of each time slot, people are invited to leave, and the museum empties completely.
Let us also note that many visitors enter without their smartphone since they must leave their personal bags in the wardrobe.
It is plain that Galleria Borghese has specific peculiarities which make it rather unique in the world. This means that not all aspects of the present study are directly applicable to other museums. Let us mention, in particular, the entrance system with quota and the fact that visitors often return to already visited rooms.
On the other hand, other aspects are easily generalizable. Mathematical and numerical methods presented here can be used whenever one has to track people moving in a built environment through a nonpredefined sequence of rooms. In addition, control and optimization technique are suitable whenever the flow of people can be controlled in some way, e.g.\ changing entrance doors and/or changing the visiting path dynamically.
\section{Data collection: IoT visitors tracking system}\label{sec:trackingsystem}
As many cultural heritage sites worldwide, Galleria Borghese is covered by frescoes and paintings, which heavily limit the possibility of displacing (electrical) devices. To cope with these historical and architectural constraints, we developed a noninvasive radio-based IoT measurement solution delivering room-level visitors trajectories.
\begin{figure}[t]
\centering
\textbf{a.} \includegraphics[width=0.36\linewidth]{GB-beacon.jpg}\hfill
\textbf{b.} \includegraphics[width=0.48\linewidth]{RPi+Beacon.jpg}
\caption{\textbf{a.} Sample visitor wearing the BLE beacon. \textbf{b.} Raspberry Pi used as Bluetooth antenna to receive beacon signals and measure their RSSI.}
\label{fig:trackingsystem}
\end{figure}
Figure \ref{fig:trackingsystem} shows the main components of the tracking system, which consists of:
\begin{description}
\item[Transmitters]
we gave a small BLE beacon to each visitor to track who was briefed about the experiment, see Figure \ref{fig:trackingsystem}\textbf a.
The beacon transmitted a signal at $+4$dB with iBeacon standard encoding \cite{newman2014}, which
carries a unique identifier (UUID).
%
\item[Fleet of receiving antennas]
we employed RaspberryPi 3B+ (RPi) as receivers. A RPi is a single-board computer with embedded Bluetooth and Wi-Fi modules, see Figure \ref{fig:trackingsystem}\textbf b.
RPi's were located along the museum in fixed positions, see Figure \ref{fig:GBmap}.
A Python code running on the RPi's was used to scan continuously the surrounding area listening for beacons signals. Each signal is stored as a tuple containing the beacon's identity, the RSSI and the timestamp of reception.
Every 5s a data packet was created (with only one occurrence of each beacon detected), it got signed with the RPi identity and posted on a central server via an internet connection.
%
\item[Central server]
the server received data packets from all RPi's and stored them in a SQL database along with the reception timestamp. Such couple of timestamps allows us to quantify the duration of the whole process.
\end{description}
The data presented in this paper come from a measurement campaign lasted between June and August 2019.
The central SQL server received $1,308,617$ records corresponding to $900$ visitors trajectories surveyed during $13$ 2h-long visit slots.
The percentage of tracked visitors w.r.t.\ the total number was about 1:5. %
As it usually happens, the vast majority of visitors came in groups (family, friends, guided tours, etc.).
In this case, apart from a few exceptions, we tracked only one member of each group, thus losing the ability to detect the interactions within social groups.
We have also tested that collecting dozens of beacons at the same time within a small physical space does not impact on the reliability of the system.
Figure \ref{F:RawBeacon} shows the history of a single beacon's RSSI (i.e., a single visitor) during a visit.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.75\linewidth]{RawBeacon.pdf}
\caption{Typical raw RSSI throughout a 120 minute visit as recorded by the 14 RPi's receiving antennas.
(Re-sampled every $\Delta t=10\,$s; RPi's antennas are distinguished by markers).
Inset: between minute 29 and 32 the visitor is detected by both RPi4 and RPi5, and the maximal RSSI strongly oscillates between the two. As a consequence, the signal strength is insufficient to associate unambiguously a visitor to a location.
}
\label{F:RawBeacon}
\end{center}
\end{figure}
The analysis of the raw data immediately confirms that RSSI signal suffers from high fluctuations (see also \cite{beder2012}). This means that a beacon fixed in the middle of a room is not received with a constant RSSI, and the RSSI of two equidistant beacons might not be the same.
In addition, we encountered other two important difficulties:
\begin{enumerate}
\item
A single beacon can be detected by multiple antennas at the same time. RSSI is used to resolve the ambiguity but high fluctuations make such task rather hard.
\item Some areas of the museum could not be covered at all (e.g., staircase between the two floors, due to lack of electrical outlets).
\end{enumerate}
Finally, let us also mention the possibility -- which we consider very rare -- that visitors wearing beacons could be influenced by the fact that they feel tracked, cf.\ \cite{trondle2014, yoshimura2014}.
In the next section, we describe how the raw RSSI data from the transmitters is processed to estimate individual trajectories.
\section{Trajectory reconstruction and filtering}\label{sec:trajectoryreconstruction}
We use the raw data collected by the tracking systems to reconstruct the sequence of visited rooms and the time of permanence in each room.
First of all, we achieve uniform temporal sampling of the signals through a re-sampling in bins of fixed time length $\Delta t$ = 10s (cf.\ Figure \ref{F:RawBeacon}). $\Delta t$ is to be tuned according to both the resolution needed and the signal granularity.
We employed a -120 dB threshold for those antennas not detecting the beacon.
The output of this procedure is a $A\times T$ matrix $\mathcal{R}$ for each beacon, where $A$ is the number of antennas and $T$ is the number of time bins (duration of the visit divided by $\Delta t$). In other words, for a given beacon, the element $\mathcal{R}_{a,t}$ is the RSSI of the signal received by $a$-th antenna in the $t$-th time bin.
In order to simplify our room-scale tracking, we have merged the 21 exhibition areas of the museum into $R=9$ (radio) rooms in which we deploy our $A=14$ receiving antennas (see Figure \ref{fig:GBmap} and Table \ref{tab:matchroomsantennas} for antennas positions and antenna-room assignments).
We remark that signal readings in a given room do not imply that the emitting beacon is located in the same room.
\begin{figure}[t]
\begin{center}
\def0.75\linewidth{0.75\linewidth}
\input{images/tracing_galleria_borghese_nomodel/tracing_galleria_borghese_nomodel.tex}
\caption{
Floor plan of Galleria Borghese.
Rooms (R) and receiving antennas (A) are reported (cf.\ Table~\ref{tab:matchroomsantennas} for room-antennas relations).
Red lines represent closed passages/doors.
Visitors admittance happens both at the main entrance on the first floor (E) and at the stairs (S) on either floor.
}
\label{fig:GBmap}
\end{center}
\end{figure}
\begin{table}[t]
\begin{center}
\begin{tabular}{ l l l }
\hline
Number & Room nickname & Antenna \\
\hline\hline
R1 & Paolina & A2 \\
R2 & David & A4 \\
R3 & Apollo e Dafne & A5 \\
R4 & Ratto di Proserpina & A1, A10 \\
R5 & Portico & A8, A9 \\
R6 & Enea e Anchise & A11 \\
R7 & Satiro su delfino & A7 \\
R8 & Caravaggio & A6 \\
R9 & Pinacoteque & A3, A12, A13, A14 \\
\hline
\end{tabular}
\end{center}
\caption{Match among rooms (R) and RPi antennas (A) in Galleria Borghese museum.}
\label{tab:matchroomsantennas}
\end{table}
The most natural way to reconstruct visitor trajectories is to compute the argmax of the RSSI history of each beacon:
at each time bin one retains the antenna that receives the highest RSSI.
Finally, one defines the current visitor location as the room associated to the antenna.
The result of this procedure is shown in Figure~\ref{F:TrElaboration}\textbf{ab}.
Unfortunately, when a visitor is at approximately equidistant from two or more RPi's, the maximal RSSI quickly bounces back and forth among the antennas. In signal terms, the visitor appears to perform extremely rapid and unrealistic room changes.
\begin{figure}[t]
\centering
\includegraphics[width=0.985\linewidth]{megamerge.pdf}
\caption{
A sample beacon RSSI elaborated by argmax (\textbf{a.}\ \& \textbf{b.}), sliding window (\textbf{c.}\ \& \textbf{d.}), and machine learning (\textbf{e.}\ \& \textbf{f.}) approaches.
The left column reports the max of RSSI for the argmax and sliding window approaches (antennas located in the same room are labeled with the same color but different markers), and the maximum among the rooms probabilities for the machine learning approach.
The right column reports the corresponding reconstructed trajectories as sequence of rooms.
Not-detected statuses are marked by green crosses ({\color{ForestGreen}$\times$}).
}
\label{F:TrElaboration}
\end{figure}
Building upon~\cite{centorrino2019}, we consider two data refinement methods: the first one relying on a sliding window approach and the second one based on a neural network.
\subsection{Sliding window approach}\label{ssec:SW}
The first method aims at smoothing the noise in the RSSI data by applying a low-pass filter, implemented as a \emph{weighted moving average}, and a \emph{normalization}.
RSSI's gathered close in time should have close values; besides, the closer the bins, the higher is the correlation.
In particular, we convolve the RSSI signals gathered by each antenna (i.e.\ the matrix rows $\mathcal{R}_{\cdot, t}$) with a (symmetric triangular) kernel with size $2\delta + 1$ and weights $w_{0},w_{1},\ldots,w_{2\delta}$.
In formulas, this approach generates a new matrix $\tilde \mathcal{R}$ defined as
\begin{equation}\label{E:scalarTildeR}
\tilde \mathcal{R}_{a, t} =
\sum_{d = t-\delta}^{t+\delta}
\mathcal{R}_{a, d} \cdot w_{\delta - t + d},
\quad
0 \leq a < A, \quad
\delta \leq t < T - \delta \ .
\end{equation}
Secondly, a normalization is applied across the signals acquired by the different antennas in order to make them comparable.
This produces a third matrix
$\bar \mathcal{R}$ as
\begin{equation}
\bar \mathcal{R}_{a, t} = \frac{\tilde \mathcal{R}_{a, t} - \mu_t}{\sigma_t}
, \qquad
0 \leq a < A
, \quad
\delta \leq t < T - \delta\ ,
\end{equation}
where $\mu_t$ and $\sigma_t$ are respectively the mean and the standard deviation of $\bar \mathcal{R}$ by time bin (i.e.\ by column, thus $\mu_t = \mu(\bar \mathcal{R}_{\cdot,t})$, $\sigma_t = \sigma(\bar \mathcal{R}_{\cdot,t})$).
Figure~\ref{F:TrElaboration}\textbf{cd} shows the result of this procedure.
\subsection{Machine learning approach}\label{MLapproach}
To improve performances of the sliding window method, we propose a trajectory reconstruction approach based on neural networks.
At any time bin $t$, we cast the localization of a visitor in one among the $R$ rooms as a classification problem. Our neural network processes the $\mathcal{R}$ matrix (in time windows) and returns the probability vector whose $r$-th component is the probability that the visitor is located in the room $r$.
\subsubsection{Building the neural network}
We consider a neural network made of $L + 1$ layers, with $L=2$. The data is injected in the first layer and flows ``forward'' in the network through the hidden layer to the output layer.
Each layer $\ell$ is built out of a different number $n_\ell$ of nodes
$a^{(\ell)}$, that represent the calculus units of the network, or artificial neurons,
where $a_j^{(\ell)}, 1 \leq j \leq n_\ell$ represents the $j$-th neuron of layer $\ell$.
The specific network that we employ, also known as Multi Layer Perceptron (MLP), is built as a complete weighted directed graph between the nodes within layer $0 \leq \ell < L$ and the nodes of the next layer $\ell+1$ (cf.\ Figure \ref{F:NeuralNetwork}).
A spare node with fixed value $1$ (\emph{bias}) and index $0$ is added to each layer but the last, that is $a^{(\ell)}_0 = 1, 0 \leq \ell < L$.
We denote by $\Theta_{s, d}^{(\ell)}$ the weight of the edge directed from the $s$-th node of the $\ell$-th layer to the $d$-th node of the $(\ell+1)$-th layer.
\begin{figure}[t]
\begin{center}
\scalebox{0.7}{
\input{images/NeuralNetwork.tex}
}
\caption{
The three layers ($L=2$) neural network employed to process the trajectories collected in Galleria Borghese. For each time bin, the neural network predicts the visitor position by considering the RSSI within the previous and the next minute.
The input layer is made of $(2\delta + 1) \times A = 182$ neurons, where $\delta = 6$ is the semi-amplitude (one-minute long) of the sliding time interval for each of the $A=14$ antennas.
The output layer is made of $R=10$ neurons, one for each room of the museum plus the ``out'' condition.
The single hidden layer is composed by $56 = 14 \times 4$ neurons, as a trade-off between the input and the output layer sizes.
}
\label{F:NeuralNetwork}
\end{center}
\end{figure}
\noindent We employ the sigmoid function
\begin{equation}
g : \mathbb R \to [0, 1] \ , \qquad g(x) = \frac{1}{1+e^{-x}}
\end{equation}
as activation function.
Hence, data propagate through the network as
\begin{equation}
a_{d}^{(\ell+1)} = g\left(\sum_{s=0}^{n_{\ell}} \Theta_{s, d}^{(\ell)} \cdot a_{s}^{(\ell)}\right), \quad 0 \leq d \leq n_{\ell+1}, \quad 0 \leq \ell < L \,
\end{equation}
being $\{a_{s}^{(0)},s=1,\ldots\}$ the input values.
Specifically, at time $t$ our input values are the $(2\delta +1)A$ values obtained by restricting the matrix $\mathcal{R}$ to the $\delta$ columns before and after $t$ (analogous notation to Section \ref{ssec:SW}), i.e. the column block $\mathcal{R}_{\cdot,t-\delta\cdots t+\delta}$. The network output are $R$ numbers in $[0,1]$ that we interpret -- after $L^1$ normalization -- as the instantaneous probability of being in the $r$-th room.
We train the network parameters via gradient descent such that the network output fits hand-annotated data.
In particular, we consider a dataset of $5427$ manually labelled input samples, $80\%$ of which are used to effectively set the weights, while the remaining $20\%$ is employed for testing (i.e. checking for generalization and absence of overfitting).
\subsubsection{Estimating a trajectory via neural networks}
Applying the neural network to the $\mathcal{R}$ matrix yields a $R \times T'$ ($T' = T-2\delta$) matrix $\mathcal P$ whose columns contain the probability of finding the considered visitor in room $r$ at time $t$.
Almost always, the network selects a room with an overall majority (probability $> 0.5$, see Figure~\ref{F:TrElaboration}\textbf e). However, it can
happen -- particularly during a room transition -- that no class reaches such a majority. We apply therefore an \textit{ad hoc} adjacency filter:
probabilities values smaller than a fixed threshold, $\chi = 0.15$, are removed from the candidates; then, unfeasible transitions are either penalised or removed (e.g.\ transitions that imply wall crossings or that cover more than two rooms).
After data re-normalization, in the unlikely event that no room is selected with an overall majority, we assume the visitor in the closest room or in the same room.
\subsubsection{Comparing with the sliding window approach}
Looking at Figure \ref{F:TrElaboration}\textbf{d} and Figure \ref{F:TrElaboration}\textbf{f} we observe two interesting features: the sliding window approach often prevents spikes from arising (see minutes $\sim 35, 53$), yielding cleaner trajectories that makes it easier to enumerate room transitions.
On the other hand, it tends to smooth trajectories excessively, ignoring fast transitions (see minute $\sim 52$).
The latter phenomenon is verified if we consider the bin-by-bin accuracy, i.e.\ the ratio between the correct predictions and the total number of samples analyzed.
The accuracy achieved over the test set by the neural network is in fact $0.858$ compared to $0.734$ obtained by the sliding windows approach.
Both of them however overcome results obtained via the argmax approach, which has an accuracy of $0.547$.
\subsection{Handling not-detected beacons}\label{sec:notdetected}
Due to (small) areas uncovered by antennas or because of random signal losses, it may happen that a beacon remains not detected. This is also what (correctly) happens before and after a visit or when the visitor leaves the museum during the visit, for example to reach the toilet.
Before performing statistical analyses, we amend for not-detected statuses whenever this can be done unambiguously.
Although we notice that such a process might require mainly museum-specific solutions, we report two corrections which we deem of general interest.
\begin{enumerate}
\item
If the ``blind period'' is less than 3 minutes (10 minutes for the Pinacoteque), and the visitor is detected in the same room before and after the blind period, the visitor is associated with that room for the whole period.
\item
If the blind period is less than 30 seconds, and the visitor is detected in two different rooms before and after the blind period, the visitor is supposed to be in one between the two rooms (at random).
\end{enumerate}
Whenever the overall not-detected status exceeds 25 minutes, we remove such a trajectory from the dataset.
Performing such pre-processing on the measurements collected during our field campaign, we obtain a dataset of $N=848$ trajectories, which will be the object of the analysis in the next sections.
\section{Trajectory analysis and clustering} \label{sec:trajectoryanalysisandclustering}
In this section we analyze the two datasets (including 848 trajectories) reconstructed via the two methods described before.
\subsection{Basic statistics}
We consider three basic illustrative statistics - that we also employ in Section \ref{sec:model} to calibrate our digital twin:
\begin{description}
\item[Time of Permanence]
we denote by ToP$(v,r)$ the total time spent by visitor $v\in\{1, \ldots, N\}$ in room $r\in\{1, \ldots, R\}$ during their visit.
\item[Returning visitors]
we denote by RET$(v,r)$ the number of times visitor $v$ stopped by room $r$.
\item[People per Room]
we denote by PpR$(r,t)$ the number of visitors in room $r\in\{1, \ldots, R\}$ during the time bin $t\in\{1, \ldots, T\}$.
\end{description}
Note that the first two indicators are Lagrangian, while the third is Eulerian.
To perform such statistics we employ the trajectories as reconstructed by the neural network (Section \ref{MLapproach}) since its accuracy is higher and spurious spikes do not affect the analysis.
\subsubsection{Size of the dataset}
Before using the datasets and the indicators introduced above, we need to check if the number of considered trajectories is enough to get useful information. To this end, we have compared ToP and PpR extracted from the whole dataset (848 trajectories) with ToP and PpR extracted from several random subsets of the datasets with variable size. We observed that the difference between partial and complete datasets stabilizes starting from 300 sampled trajectories, meaning that we would not have obtained significantly different results if we had tracked more than 300 trajectories.
\subsubsection{Time of Permanence} \label{sec:top}
For each room $r$, the distribution of $\{$ToP$(v,r)\}_{v\in\{1,\ldots,N\}}$ is well fit by a Weibull distribution with $r$-dependent parameters $\lambda$ and $k$.
The Weibull distribution performed best among the tested ones according to the Akaike Information Criterion. In Figure \ref{F:wei}, we report ToP empirical distributions and their Weibull fit for selected individual rooms as well as for the whole museum.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{TimeOfPermanences.pdf}
\caption{
Four ToP distributions and their Weibull fit.
\textbf{a.} \emph{Satiro su delfino} ($r_7$, $k=1.8$, $\lambda=17$).
\textbf{b.} \emph{Apollo e Dafne} ($r_3$, $k=2$, $\lambda=36$).
\textbf{c.} Pinacoteque ($r_9$, $k=2.8$, $\lambda=221$).
\textbf{d.} Whole museum ($k=4.1$, $\lambda=572$).
In the last case, the Weibull distribution does not fit correctly due to the forced exit after 2h. This problem will be solved later in Section \ref{sec:optim_no_bell}, by censoring the last 5 minutes of the visit.
Inset: related Quantile-Quantile plots that depict the Real vs.\ Weibull quantile relation.
}
\label{F:wei}
\end{figure}
The Weibull distribution is related to the ``time-to-failure'' of a system, which, in our context, is to be interpreted as the ``time-to-exit'' a room (more precisely as the ``time-to-exit-and-do-not-return'', since we consider the ToP as the total time spent in a room).
The parameter $\lambda$ (characteristic time of visit) gives information about the room holding power, while parameter $k$ (Weibull slope) characterizes the decision to leave the room.
In particular, for all rooms of Galleria Borghese $k>1$ holds, meaning that the exit (failure) rate increases with time ($k=1$ indicates that the exit rate is constant over time while $k<1$ indicates a decreasing-in-time exit rate).
In practice, we deem that visitors find all the rooms worthy of attention, and it happens rarely that visitors leave a room immediately.
In Section \ref{sec:model}, the survival and the hazard function associated to the Weibull distributions will be used as building blocks of the proposed museum digital twin.
\subsubsection{Returning visitors}
Galleria Borghese has a circular structure with no fixed or suggested path for visitors (cf.\ Section \ref{sec:casestudy}).
Besides, the density of artworks is so remarkably high that visitors easily miss a fraction of the pieces during the first passage of a room.
Therefore, we investigate the number of times a person visits the same room, on average.
In this analysis, we neglect quick returns (less than a minute of permanence).
We observe that each guest visits a room, on average, 1.3 times (1.5 times, for Room 8, \emph{Caravaggio}), while entrance rooms have 2.7 passages.
On the other hand, 25\% of the visitors skips at least one room (especially room 7, \emph{Satiro su delfino}). The time of permanence during the first passage by a room is generally the longest, in comparison to the next ones (this, however, does not hold for entry rooms).
The time of first return (time interval between the moment a visitor leaves a room and the moment they return) appears consistent throughout the museum rooms and is between 25 and 30 minutes.
Finally, we highlight that the occurrence of fast returns (less than 5 minutes), which in our case are about 10\% of all returns, could indicate that visitors frequently get lost or change the direction of visit (clockwise vs.\ counterclockwise, cf.\ museum map in Figure \ref{fig:GBmap}).
\subsubsection{People per room}
\label{sec:ppr}
The number of people per room, PpR, is probably the most relevant indicator as well as that of largest interest for museum curators, as it connects with safety (hyper-congestion), comfort, and attractiveness for the audience (under-used rooms could indicate scarce interest).
We calculate the PpR$(r,t)$ by counting the number of visitors of each turn who are in room $r$ in time bin $t$.
To amend for the fact that we gave beacons to a sample of visitors, and only to one member of each social group, we consistently replicate each trajectory $q$ times, where the integer $q$ is uniformly distributed between 1 and 6.
We compare the PpR time series of one room and of the whole museum with our simulations in Section \ref{sec:simulation-results}.
\subsection{Measuring the distance among trajectories: a Wasserstein-inspired metric}
Defining a suitable metric in the space of trajectories is an essential step to quantify how `close' (or `similar') are distinct paths followed by visitors.
We define a new, \emph{ad hoc}, metric inspired by the Wasserstein distance, which is usually employed to quantify the distance between two abstract measures or two density functions. Following the same ideas, we quantify how much it costs to transform one visitor trajectory, time bin-by-time bin, into another, until they are identical.
First of all, we model the map of the museum as a graph with $R$ nodes. Edges between pairs of nodes represent viable connections between rooms.
We also consider the possibility that the museum is organized in different \emph{wings}: wings are independent areas, which are so far from each other that it is natural to assume that visitors rarely visit the same wing twice (it is usually the case when a museum has multiple floors, comes in different buildings, or has multiple thematic areas).
For technical reason, we always assume that there exists a wing called `Out'. Visitors in this wing are waiting to enter or have already left the museum. Thus, we represent Galleria Borghese in three wings: main floor, Pinacoteque, and Out.
Second, we introduce a distance function \emph{on the graph}, i.e.\ between room-nodes.
We define the distance $\mathcal D(r^1,r^2)$ between room $r^1$ and room $r^2$ ($r^1,r^2 \in \{1, \ldots, R\}$) by
\begin{equation}\label{eq:defD}
\mathcal D(r^1,r^2):=
\left\{
\begin{array}{ll}
0, & r^1=r^2, \\
-\frac\alpha 2+\alpha\mathcal T_r(r^1,r^2)+\beta \mathcal T_w(r^1,r^2), & r^1\neq r^2,
\end{array}
\right.
\end{equation}
where $\mathcal T_r(r^1,r^2)$ is the minimum number of room transitions (i.e.\ graph edges to hop through) necessary to traverse the graph from room $r^1$ to room $r^2$, whereas $\mathcal T_w(r^1,r^2)$ is the number of wing transitions, and $\alpha, \ \beta >0$ are two parameters.
The term $-\frac\alpha 2$ is motivated by the need of decreasing the weight of short transitions (each room transition counts $\alpha$ but the first one which, instead, counts $\frac\alpha 2$), which often happen as many visitors stand still at the interface/door between two rooms, without actually moving in either direction.
Note that the distance $\mathcal D$ is not necessarily commutative. This can happen e.g.\ if a museum comes with some one-way room transitions.
Third, all trajectories are extended in order to have the same number of time bins.
We achieve this by fixing a maximal theoretical duration of the visit (2h in our case) and then exploiting the fictitious wing Out, where people are placed before and after the actual visit.
In this case, we have performed the analysis using the sliding window approach (Section \ref{ssec:SW}). This choice was driven by the need of lessening the amount of spikes, see, e.g., Figures \ref{F:TrElaboration}\textbf{d} and \ref{F:TrElaboration}\textbf{f} at minute $\sim 35$ where spikes may arise between rooms located on different floors, affecting the reliability of the measure.
We have now all the ingredients to define the distance between two trajectories $\textsc{t}^1,\textsc{t}^2\in\{1,\ldots,R\}^T$
\begin{equation}\label{eq:defW}
\mathcal W(\textsc{t}^1,\textsc{t}^2):=\sum_{t=1}^T \mathcal D(\textsc{t}^1_t,\textsc{t}^2_t),
\end{equation}
which represents the sum, time bin-by-time bin, of the distances between the rooms (according to \eqref{eq:defD}) occupied by the two visitors at each given instant.
As an example, in Figure \ref{F:likelihood} we report the distribution of the pairwise distances between all trajectories, computed using metric \eqref{eq:defW}.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.60\linewidth]{likelihood.pdf}
\caption{
Distribution of mutual distances between trajectories. The $x$-axis is normalized w.r.t.\ the longest measured distance.
The mean pairwise distance $\mu$ is reported in red while the shaded area denotes the range $\mu \pm \sigma$ ($\sigma$ being the standard deviation of the distribution).
}
\label{F:likelihood}
\end{center}
\end{figure}
Figure \ref{F:most-least-common} shows instead the most and least common trajectory. The higher the number of trajectories `close' to a given one, the more common the trajectory is. Hence, we report the single trajectory having the highest number of other trajectories within distance $\mu -\sigma$ (Figure \ref{F:most-least-common}\textbf{ab}, cf.\ definition of $\mu$ and $\sigma$ in Figure \ref{F:likelihood}) and the single trajectory having the least number of other trajectories within distance $\mu + \sigma$ (Figure \ref{F:most-least-common}\textbf{cd}).
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.9\linewidth]{most-least-common.pdf}
\caption{
\textbf{a.} Most common trajectory and, \textbf{b.}, distribution of the distances between such trajectory and all the others. The visitor performs a circular visit following the room numbering in the main floor, then they reach the Pinacoteque upstairs.
\textbf{c.}\ \& \textbf{d.} Analogous plot for the least common trajectory in our dataset. The visitor enters the museum via the Pinacoteque, then they visit the main floor twice, once clockwise and once counterclockwise.
$x$-axis in \textbf{b.}\ \& \textbf{d.}\ is normalized w.r.t.\ the longest measured distance among all the trajectories.
}
\label{F:most-least-common}
\end{center}
\end{figure}
Finally, we mention the capability of finding automatically members of social groups (it could happen that elements of the same social group went to the ticket office separately, thus were assigned more than one beacon). Indeed, two or more trajectories very close to one another likely belong to visitors in company. Figure \ref{fig:group-visitors_0} reports a sample of trajectories at different distance from a given reference trajectory.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.9\linewidth]{group-visitors.pdf}
\caption{
\textbf{a.} Sample of measured trajectories.
\textbf{b.} Distribution of the distances between the trajectory marked by ``blue plus signs'' ({\color{Cerulean}$+$}) in \textbf{a.} and all the others.
Distances are reported in percentage w.r.t.\ the longest distance measured.
Trajectories closer than $0.025\%$ (yellow bin) likely belong to the same group of visitors;
Trajectories closer than $0.15\%$ (red bins) are slightly time shifted;
In trajectories closer than $0.30\%$ (purple bins) relations are still identifiable;
Trajectories farther away than $0.30\%$ (green ones) are completely unrelated.
Trajectories in \textbf{a.} are random sampled from corresponding color percentile sets in \textbf{b.}
}
\label{fig:group-visitors_0}
\end{center}
\end{figure}
\subsection{Clustering algorithms}\label{sec:clustering-algorithms}
As we recalled in Section \ref{sec:refs}, clustering algorithms can be used for inferring, from the whole trajectories data set, the typical paths or, equivalently, the typical individual behaviors inside the museum.
Here we employ algorithms which do not require to define \emph{a priori} the number $k$ of clusters, nor to assign predefined reference trajectories around which clusters are agglomerated (as typically happens with, e.g., $k$-means approaches).
Moreover, we do not use the typical taxonomy (ant, butterfly, fish, grasshopper, cf.\ Section \ref{sec:refs}) to guide the clustering, aiming at other, possibly hybrid, behaviors.
To this end, we employ an \emph{agglomerative hierarchical clustering} (AHC) approach (see, e.g., \cite{duda2012pattern}).
These techniques consider a bottom-up cluster tree (dendrogram), that, step by step, gathers trajectories according to their mutual likelihood.
In the beginning, each trajectory is considered to be the only element of a distinct cluster. Then, at each iteration, the two closest clusters get merged into a single cluster.
The process is deterministic, unless we have two couples of clusters at exactly the same distance, and it always ends with one single cluster after $N-1$ steps, given $N$ the number of initial trajectories.
Note that cutting the dendrogram at the $\ell$-th layer from the tree leaves provides exactly $N-\ell$ clusters.
Finding an adequate cutting layer is an issue which we discuss in the following.
To measure the distance between two clusters, we leverage on \eqref{eq:defW}.
We consider, in particular, three common methods:
\begin{description}
\item[C-LINK]
in Complete Linkage, the distance between two clusters $\mathcal C^1$ and $\mathcal C^2$ is the maximum amongst the distances between all the trajectories within the two clusters:
\begin{equation}
\mathcal W(\mathcal C^1, \mathcal C^2) = \max\{\mathcal W(\textsc{t}^1, \textsc{t}^2) : \textsc{t}^1 \in \mathcal C^1, \textsc{t}^2 \in \mathcal C^2\}.
\end{equation}
%
\item[S-LINK]
in Single Linkage, the distance between two clusters $\mathcal C^1$ and $\mathcal C^2$ is the minimum amongst the distances between all the trajectories of the two clusters:
\begin{equation}
\mathcal W(\mathcal C^1, \mathcal C^2) = \min\{\mathcal W(\textsc{t}^1, \textsc{t}^2) : \textsc{t}^1 \in \mathcal C^1, \textsc{t}^2 \in \mathcal C^2\}.
\end{equation}
%
\item[UPGMC]
in Unweighted Pair Group Method with Mean Centroid, each cluster $\mathcal C$ is identified by a representative trajectory $\bar{\textsc{t}}_{\mathcal C}$, and the distance between two clusters is evaluated as the distance between representative trajectories:
\begin{equation}
\mathcal W(\mathcal C^1, \mathcal C^2) =
\mathcal W(\bar{\textsc{t}}_{\mathcal C^1},\bar{\textsc{t}}_{\mathcal C^2}).
\end{equation}
\end{description}
Determining a representative trajectory $\bar{\textsc{t}}_{\mathcal C}$ in a trajectory set $\mathcal C$ is useful in general, and mandatory to employ UPGMC.
To do so, we compute a mode among all the trajectories: for each time bin $t$, our representative trajectory reports the most visited room among the elements of $\mathcal C$:
\begin{equation}\label{E:centroid}
(\bar{\textsc{t}}_{\mathcal C})_t :=
\arg\max\limits_{r \in \{1,\ldots, R\}}\left\{
\sum_{\textsc t \in \mathcal C} \mathds{1}_{\textsc t_t = r}
\right\}\ ,
\qquad 0 < t \leq T\ .
\end{equation}
Note that the centroids found with a specific cut may also be employed to clusterize a different set of trajectories.
This also means that, if new trajectories are gathered, the same centroids may be used in order to get a clustering.
This may reveal that habits have been broken or new paths have been discovered.
\medskip
\emph{Cutting the dendrogram.}
In order to find the right cutting threshold for the dendrogram, we consider the number of the $p$-\emph{significant} clusters, i.e.\ the clusters with more than $p$ elements, while traversing the tree from the leaves to the root.
Having a high variation in the number of significant clusters in the proximity of the root often implies that clusters are unstable, i.e.\ they merge randomly in the process, preventing valuable interpretations. Having instead a very small number of significant clusters, say one or two, often means that each cluster contains very nonhomogeneous elements, thus resulting practically useless for categorization.
Figure \ref{F:dendrogram-density}\textbf a reports the number of 5- and 15-significant clusters as a function of the dendrogram depth, for the three methods described before.
C-LINK yields many small unstable clusters joining together, with no meaningful interpretation, towards the end of the process.
S-LINK offers, on the other hand, a poor set of typical clusters to which all the trajectories converge quickly throughout the clustering process.
Conversely, UPGMC leads to a relatively small amount of consistent stable clusters.
\begin{figure}[t]
\centering
\includegraphics[width=0.97\linewidth]{dendrogram-cutting.pdf}
\caption{
\textbf{a.} Number of 5- (filled markers) and 15- (empty markers) significant clusters as a function of the dendrogram depth for C-LINK, S-LINK and UPGMC methods.
\textbf{b.} Number of $p = 5, 15, 30, 50$ significant clusters obtained via UPGMC method.
The dendrogram is cut in correspondence to the plateau at depth $67$.}
\label{F:dendrogram-density}
\end{figure}
In particular, the UPGMC dendrogram shows a plateau around layer $\bar\ell\sim 67$, for many values of $p$, see Figure \ref{F:dendrogram-density}\textbf b. We adopt such a cutting layer since it ensures the maximum amount of highly significant clusters ($p = 30, 50$ have the last absolute maximum there) without trading-off too much information in smaller clusters.
\subsection{Clustering results} \label{sec:clustering-results}
We consider here the representative trajectories of each cluster obtained after a dendrogram cut at layer $\bar\ell$.
Although none of the representative trajectories strictly coincides with any among the trajectories observed, they all appear real (i.e.\ conform with a potential visit). This emphasizes that clusters indeed aggregate similar trajectories.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.95\linewidth]{cluster-trajectories.pdf}
\end{center}
\caption{
Four representative trajectories (centroids) of clusters joining respectively \textbf{a.} 16\%, \textbf{b.} 9\%, \textbf{c.} 4\%, \textbf{d.} 1\% of the trajectory data set.
Representative trajectories may show spikes (see, e.g., \textbf c., $\sim 75$ minute). According to \eqref{E:centroid}, this phenomenon arises whenever rooms have approximately the same number of visitors within the same interval of time.
}
\label{fig:clusters}
\end{figure}
Figure \ref{fig:clusters} shows four representative trajectories related to four clusters of different size.
The two most common patterns are related to visitors who follow the natural numbering of the rooms, starting or ending the visit in the Pinacoteque, which is visited once.
This identifies the most typical visit pattern for the curators.
Nevertheless, clustering investigation brings to light other, less expected, patterns: the one which does not include the visit at the Pinacoteque (possibly visitors who did not find the staircase) and patterns where the Pinacoteque comes amidst the visit.
Note that both patterns have been observed by the museum managers and are discouraged.
\emph{Filtering by clustering.} Clustering can be also used to detect unfeasible/unreal trajectories coming from system malfunctioning, since those trajectories tend to gather in a single cluster.
This powerful feature helps to design filters to clean up the data during the preprocessing phase.
\emph{Anomaly detection.} Trajectories which remain isolated in the last layers (close to the root of the tree) are, by definition, far from all the other centroids and therefore very atypical.
We claim that these trajectories are \emph{anomalies} detected during the process.
If they do not come from system malfunctioning, they belong to people who behave abnormally or suspiciously and deem additional checks.
Figure \ref{F:anomalies} shows some of the anomalies detected in our study.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.95\linewidth]{cluster-anomalies-2.pdf}
\end{center}
\caption{
Two anomalies detected.
\textbf{a.} A rare pattern where the Pinacoteque and main floor are both visited twice.
\textbf{b.} A strange pattern with many changes of direction (clockwise/counterclockwise).
}
\label{F:anomalies}
\end{figure}
\section{Model and calibration}\label{sec:model}
In this section, we develop a digital twin of the museum, i.e.\ an algorithm which is capable of generating new trajectories, (statistically) indistinguishable from measured ones.
In order to represent the complex visitor behaviour, we employ a stochastic approach based on Markov Chains (MC). We design our simulator to generate visiting paths with relevant observable features such as guests skipping one or more rooms and/or returning multiple times to the same room.
\medskip
The model is based on two important assumptions:
\begin{description}
\item[Visitors are independent from each other]
the decision to leave or remain in a room does not depend on the number of people in that room.
This assumption is certainly reasonable up to mild congestion levels. On the other hand, hyper-congestion has surely an impact on visitors choices, however, our current data collection seems still insufficient to quantify such a challenging aspect.
We suspect that congestion can either increase or decrease the ToP, depending on the perceived importance and fame of the room content.
\item[Social groups behave as one individual]
social groups visit the museum remaining together, i.e.\ following the same trajectory and thus spending the same time in each room.
This assumption, which is an important limitation, is consistent with the fact that beacons were given almost always to a single member of each social group.
Therefore, we are not capable of disentangling interactions and differences within social groups.
\end{description}
In a standard MC, the transition probability from a state (room) to the next depends only on the current state. However, in our context, it is an intuitive idea that the visitors choices depend, in some way, on the rooms that they have previously visited. Furthermore, since we are assuming that there is no predefined visit path, a standard MC creates a \emph{bounce phenomenon} among rooms, (i.e.\ $1 \to 2 \to 3 \to 2 \to 1 \to 2$), while the majority of paths are more regular (e.g.\ $1 \to 2 \to 3 \to 4$ or $4 \to 3 \to 2 \to 1$).
In order to simulate the complex visitor behavior, we introduce a \emph{memory} in the Markov Chain, to represent the visitors knowledge of the visited rooms. Moreover, we reasonably assume that visitors also remember the time spent in each room.
We use a nonhomogeneous transition matrix, which is time-dependent through a weight function $S$. Henceforth, we refer to this model as Time Varying Markov Model.
\subsection{Time Varying Markov Model (TVMM)} \label{sec:TVMM}
Since the museum comes with $R$ rooms, we consider a $R\times R$ transition matrix $\mathcal K$. Following the frequentist definition of probability, $\mathcal K_{r^1,r^2}$ is computed by first counting, from all the measured trajectories, the number of transitions from room $r^1$ to room $r^2$, where $r^1=r^2$ holds if the visitor remains in the same room.
\begin{equation}\label{eq:counter}
\mathcal K_{r^1,r^2}=\sum_{n=1}^{N} k_{n}(r^1,r^2),
\end{equation}
where $k_n(r^1,r^2)$ denotes the number of $r^1\rightarrow r^2$ transitions along the $n$-th trajectory of the data set.
The sum over columns of $\mathcal K_{r^1,r^2}$ represents the total time, in time bin, spent by all tracked visitors in room $r^2$.
If we normalize $\mathcal K$ by row, so that $\sum_{r^2} p_{r^1,r^2}=1$, we obtain a transition matrix $\mathcal M$ where the new element $p_{r^1,r^2}$ represents the probability to move from room $r^1$ to room $r^2$, see Figure \ref{F:TM}.
\begin{figure}[t]
\centering
\scalebox{0.7}{\input{images/tm-graph.tex}}
\caption{
Transition probabilities between rooms in Galleria Borghese (the probability to remain in the same room is not included).
We can see that the counterclockwise path is preferred and fast transitions from rooms 5, 2 to rooms 2, 4, respectively, exist.
The probability of leaving the museum is sampled as a hazard function $h(t)$.
}
\label{F:TM}
\end{figure}
In order to avoid the room bouncing phenomenon, we make the transition matrix $\mathcal M$ time-dependent. More precisely, we consider the matrix
\begin{equation}\label{eq:transition-matrix-intro}
\tilde{\mathcal M}_{r^1,r^2}(t)=\mathcal M_{r^1,r^2} \ S_{r^2}(t),\qquad r^1,r^2 \in \{1,\ldots,R\},
\end{equation}
where $S_r(t)$ is the survival function associated to ToP$(\cdot,r)$ via its Weibull fit parameters $(\lambda_r,k_r$),
\begin{equation}
S_r(t) = e^{-(t/\lambda_r)^{k_r}}.
\end{equation}
In other words, $S_r(t)$ quantifies the probability that a guest visits room $r$ for a time interval longer than $t$.
$S_r(t)$ is a decreasing function such that $S_r(0)=1$ and $S_r(t_{max})=0$, where $t_{max}$ is the largest measured ToP$(\cdot,r)$.
At each time step of the simulation, the function $S_r(t)$ must be updated on the basis of the time spent in each room, and the transition matrix $\tilde{\mathcal M}$ has to be normalized by rows in order to have a correct definition of the transition probability.
In the following, we detail the exceptions to the transition dynamics in \eqref{eq:transition-matrix-intro} to cope with the access and exit conditions.
\begin{description}
\item[Beginning of a visit]
in Galleria Borghese visitors enter all together at the beginning of the visit turn. However, due to some delay (ticket control, late arrival, queue at wardrobe), the entrance process is completed in about 20 minutes. We simulate these dynamics extracting the delay at random from the set of measured delay. In addition, we use another probability distribution function to assign the entrance room (we recall that Galleria Borghese has three entrances: \emph{Ratto di Proserpina} (room 4), \emph{Portico} (room 5), and Pinacoteque (room 9)).
\item[Conclusion of a visit]
the wing Out is conceptually different from the other wings. Therefore, we manage the exit time in a distinct manner. The exit is not controlled by $\tilde{\mathcal M}$, instead it is managed via the hazard function $h$ of the Weibull distribution with parameters $(\lambda_{*},k_{*})$ associated to the total time of visit, see Figure \ref{F:wei}\textbf d.
More precisely, at every time bin $t$ of the simulation, the exit probability is given by
\begin{equation}\label{eq:hazardexit}
P(r_{t+1}=\text{Out}\mid r_{t}=r_{exit})=h(t;k_{*},\lambda_{*}),
\end{equation}
where $r_{exit}$ is an exit room and
\begin{equation}
h(t;k_{*},\lambda_{*})=\frac{k_{*}}{\lambda_*^{k_{*}}}t^{k_{*}-1}.
\end{equation}
\end{description}
\subsection{Simulation results} \label{sec:simulation-results}
Before presenting the results of our model, we explain how we compare simulated trajectories with the measurements in our data set. We perform such comparison to quantify the accuracy of the simulation.
The observables of interest are the following:
\begin{description}
\item [ToP]
we evaluate the mean and the coefficient of variation of the ToP distributions, for both real and simulated visits.
\item [PpR]
we consider 100 statistically independent simulated turns, each including a total of 400 generated trajectories (similarly to Section \ref{sec:ppr}, simulated visits are replicated $q$ times, where $q$ is a uniform integer random variable between $1$ and $6$, to mimic social groups). We compute the PpR at each time bin as an ensemble average across such 100 realizations.
%
\item [Clusters]
we use the same clustering technique presented in Section \ref{sec:clustering-algorithms} to aggregate simulated trajectories.
This analysis aims to check if the most numerous cluster is sufficiently close to the measurements; this guarantees that the simulator creates a sufficient amount of plausible trajectories.
\end{description}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.9\linewidth]{simulated-trajectories.pdf}
\caption{Two simulated trajectories: \textbf{a.} A long trajectory which begins from the Pinacoteque (room 9) and then moves to the main floor according the room enumeration.
\textbf{b.} A short trajectory that begins from room 5, traverses the main floor according to the room enumeration, and finally reaches the Pinacoteque.
}
\label{fig:simulatedtrajectories}
\end{center}
\end{figure}
Figure \ref{fig:simulatedtrajectories} shows two simulated trajectories, which indeed share typical features with measurements: the Pinacoteque is visited once and the visit path follows the natural numbering of rooms. At times, people come back to rooms already visited, as in real life.
Table \ref{T:delta} compares the real and simulated ToP distributions, by considering the relative differences in ToP averages ($\mu$) and variation coefficient ($\mbox{VC}=\sigma/\mu$, $\sigma$ being the standard deviation of the ToP distribution), respectively
\begin{equation}
\delta \mu = \frac{\mu_{sim}}{\mu_{real}} - 1 \qquad \delta \mbox{VC} = \frac{\mbox{VC}_{sim}}{\mbox{VC}_{real}} - 1.
\end{equation}
The mean values of the distributions are well approximated, despite the simulations tend to slightly overestimate the ToP in the main floor and to underestimate it in the Pinacoteque. The $\delta$VC indicator highlights instead some differences between model and data: real visitors are more unpredictable than simulated ones, which yields negative $\delta$VC values. On the contrary, the dynamics in the Pinacoteque appears predictable and even more consistent than in simulations. This most likely relates to the fact that the Pinacoteque is the area with the weakest antenna coverage: amending not-detected data diminishes the variance of the measured ToP distribution.
\begin{table}[t]
\begin{center}
\begin{tabular}{c c c c c c c c c c c}
\hline
{\small Room } & {\small 1} & {\small 2} & {\small 3} & {\small 4} & {\small 5} & {\small 6} & {\small 7} & {\small 8} & {\small 9} & {\small Museum} \\
\hline\hline
{\small $\delta \mu$ } & $11\%$ & $10\%$ & $2\%$ & $8\%$ & $12\%$ & $-3\%$ & $-2\%$ & $7\%$ & $-8\%$ & $-3\%$ \\
{\small $\delta$VC } & $-28\%$ & $-10\%$ & $-15\%$ & $-15\%$ & $-20\%$ & $13\%$ & $12\%$ & $3\%$ & $31\%$ & $-11\%$ \\
\hline
\end{tabular}
\caption{
Relative error between mean and coefficient of variation of ToP distribution evaluated for real trajectories and simulated ones.
$\delta x= (x_{\mbox{\footnotesize sim}}/x_{\mbox{\footnotesize real}}-1)$, where $x$ is either $\mu$ or $VC$.
}
\label{T:delta}
\end{center}
\end{table}
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.9\linewidth]{comparePpR.pdf}
\caption{
Comparison of the average PpR of real visits (red line) and the ensemble-average PpR of simulated visits (blue line) in \textbf{a.} \emph{Ratto di Proserpina} and in \textbf{b.} the whole museum.
The shaded area corresponds to the interval $[\mu-\sigma,\mu+\sigma]$. We note that the blue line is almost entirely contained in the shaded area, as expected.
}
\label{fig:comparePpR}
\end{center}
\end{figure}
In Figure \ref{fig:comparePpR}, we compare measurements and simulations considering the PpR as a function of time. Simulations are reported in terms of ensemble statistics among 100 realizations, in particular we consider ensemble PpR average and ensemble PpR standard deviation.
In Figure \ref{fig:compareclusters} we finally report the representative trajectories of the two most numerous clusters obtained by gathering real and simulated trajectories.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.45\linewidth]{simulated-centroid.pdf}
\caption{
The two most numerous clusters obtained gathering real and simulated trajectories.
The real case joins $16\%$ of real trajectories, whereas the simulated one $18\%$.
We note that they share a number features, e.g., the ToP in each room, the total time of visit, the entry room (\emph{Portico}, room 5), and the final room (Pinacotque, room 9). The main difference is the behavior after completing the visit on the main floor. Real visitors come back counterclockwise, while simulated visitors keep walking clockwise.
This could be explained by the fact that many visitors ask for information in room 5 and are sent backwards to the staircases. The model does not include the interactions with the museum staff, hence cannot catch this feature.
}
\label{fig:compareclusters}
\end{center}
\end{figure}
\medskip
\section{Museum control and optimization} \label{sec:optimization}
We are now ready to employ the digital twin introduced in the previous section as a tool to improve the museum experience. More precisely, we simulate different scenarios and observe visitors behavior in virtual environments, aiming at supporting curators decisions.
In this regard, it is useful to remark that changing the ticketing strategy or the duties of security staff can require weeks of training in real life.
For our case study, we identify the following control variables and objectives. \\
\emph{Control variables:}
\begin{description}
\item[C1] considering that Galleria Borghese has three entrances (\emph{Ratto di Proserpina}, \emph{Portico}, Pinacoteque), museum managers can assign a certain percentage of visitors to each entrance (currently they are 15\%, 60\%, and 25\% respectively).
Operationally, such control can be implemented by introducing a tag (e.g.\ name or color) in the ticket which specifies the entrance.
\item[C2] the scheduled entry times in the museum can be tuned (currently visitors enter at 09:00, 11:00, 13:00, 15:00, and 17:00, after the museum empties).
\item[C3] the number of visitors allowed in each turn can be modified (currently it is 360 reserved in advance plus 30 last-minute).
\item[C4] the fixed duration of a visit turn can be either (C4a) modified or (C4b) totally removed (currently it is set to 2h slots).
\end{description}
\emph{Objectives:}
\begin{description}
\item[O1] keeping the PpR below a certain room-dependent threshold. Historically, our study began precisely to control the number of visitors in the Pinacoteque, which has a very low admittance limit for safety reasons.
\item[O2] keeping the PpR, in any room at any time, approximately constant. This would reduce strong variations of relative humidity which can damage the artworks \cite[Chapter 2]{camuffo1998book}.
\item[O3] decreasing the queue at the entrance.
\item[O4] increasing the number of visitors per day.
\end{description}
Note that O1, considering the emergency situation caused by the COVID-19 virus pandemic, can be employed to respect the imposed social distances legislation.
Among the many possibilities, we focused on two improvements: C1 aiming at O1, and C2, C3 \& C4b aiming at O1 \& O2.
\subsection{Entrance strategy optimization}
Keeping the existing conditions regarding the number of visitors and the time horizon, we explore the effects of a different visitor partition among the three entrances (C1).
We aim at a PpR as low as possible in all rooms (O1), especially in the Pinacoteque, which is the room with the most stringent safety constraints.
We fix an \emph{overcrowding threshold} for each room, representing a PpR limit the curators do not want to exceed.
Then, we pursue a brute force attack to the optimization problem, trying all the possible triplets $(E_1, E_2, E_3)\in [0,100]^3$, $\sum_{e=1}^3 E_e = 100$, which indicate the percentage of visitors starting the visit from each entrance $e=1$ (\emph{Ratto di Proserpina}), $e=2$ (\emph{Portico}), and $e=3$ (Pinacoteque).
Figure \ref{fig:entranceoptimization} shows the results of the optimization process evaluating the \emph{total time the PpR exceeds the overcrowding threshold} (ToT), for room 9 (Pinacoteque), room 8 (\emph{Caravaggio}), and for all the remaining rooms.
\begin{figure}[t]
\centering
\includegraphics[width=0.75\linewidth]{entrance-optimization.pdf}
\caption{
Total time duration in which the overcrowding threshold is exceeded (ToT) in room 9 (Pinacoteque), in room 8 (\emph{Caravaggio}) and in all other rooms (sum of each ToT is considered), for 13 triplets $(E_1, E_2, E_3)$.
We observe that the overall ToT exceeds at least 120 min over a day of visit regardless of the entrance system.
}
\label{fig:entranceoptimization}
\end{figure}
The best triplet for the Pinacoteque is $(20,60,20)$, while the best triplet for \emph{Caravaggio} is $(40,20,40)$: in fact, these configurations minimize the ToT in those rooms, respectively.
More in general, it is easy to see that optimal choices for one room do not necessarily mean optimality for others.
The solution currently employed by the museum, which is $(15,60,25)$, is almost optimal to reduce overcrowding in Pinacoteque and in the whole museum in general, but it sacrifices the pleasantness of the visit in some rooms of the main floor.
\subsection{Removing the finite time horizon of the visits}\label{sec:optim_no_bell}
The full elimination of the current finite time horizon allowed for the visits is a challenging improvement for the museum experience.
The idea is to keep the reservation mandatory, with entry interval fixed every 30, 60, or 120 minutes (C2), but, unlike the current setting, \emph{remove the requirement to leave after 2h} (C4b).
The immediate advantage is that the museum staff does not have to empty the museum at the end of the visit turn, thus saving about 5-7 minutes during which the museum remains completely empty (O2).
Moreover, this would also be a great advantage for the (few) visitors who want to stay for a very long time inside the museum.
Unfortunately, as it happens for every mathematical model, simulation results are reliable only in the conditions in which the simulator was developed and calibrated.
In our measurements, less than 1/4 of visitors are still inside the museum when the time limit is reached (and are forced to exit); for these a (negative) influence of the time limit certainly occurs.
Nevertheless, such influence possibly exists also for the other 3/4, that might have scheduled their visit according to the existing time constraints.
We attacked the problem by \emph{censoring} the Weibull distribution of the time of visit of the whole museum (cf.\ Section \ref{sec:top}). This statistical procedure allows us to deal with data set in which the event of interest is not observed during the study.
We obtain the new distribution as a maximum likelihood estimate censoring the last 5 minutes of visit, see Figure \ref{F:weibull_censored}.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{cum_distr_weibul.pdf}
\caption{
Cumulative hazard function associated to the Weibull distribution of the whole museum.
Empirical values are calculated with the Kaplan--Meier method.
\textbf{a.} Without censoring (cf.\ Figure \ref{F:wei}\textbf{d.}) and
\textbf{b.} after censoring the last 5 minutes of visit (new parameters are $k_{*}=3.5$ and $\lambda_{*}=596$).
This method allows us to get a better fit of the real distribution between 0 and 2h, i.e.\ the visit interval. The uncensored fit, instead, is negatively influenced by the forced exit.
}
\label{F:weibull_censored}
\end{figure}
We use the estimated parameters to modify the hazard function which controls the conclusion of the visit.
We simulated an entire day i.e.\ 9 a.m. -- 7 p.m., corresponding to the total time span of the 5 visit turns currently implemented.
This is necessary as after removing the time limit, visit turns overlap and the museum never empties.
Figure \ref{fig:removingbell} shows the result of the optimization process. The best strategy is to let 100 visitors (C3) enter from the main floor (C1) every 30 minutes (C2). These choices eliminate completely the peaks in the PpR indicator (congestion moments, O1) and the PpR remains stable with small fluctuations during the whole visit day (O2). Having the system approximately at this thermodynamic-like equilibrium greatly facilitates the management since it allows to calculate -- using the measured transition matrix -- the average number of people in each room from the number of visitors allowed (i.e.\ sold tickets).
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.75\linewidth]{ppr_vs_optim.pdf}
\caption{
PpR as a function of time in the current settings and considering the best entrance strategy. The comparison includes \textbf{a.} room 8 (\textit{Caravaggio}) and \textbf{b.} the whole museum.
}
\label{fig:removingbell}
\end{center}
\end{figure}
\section{Conclusions and future work} \label{sec:conclusions}
This study aimed at measuring, analyzing, modeling, and optimizing visitors behavior in museums and similar environments.
The practical goal was to provide suggestions to museum curators for efficiently managing visitors flows.
The implemented measurement system is sustainable for the museum, being economically viable and well accepted by visitors.
A free application to be installed on the smartphone could serve as a beacon as well, provided visitors find it useful (as an audio-guide, for example).
Employing Raspberry Pi's as fixed Bluetooth antennas appeared quite convenient and allowed the necessary development flexibility.
A major issue surely comes from the noisiness of the Bluetooth signal, which must be overcome by suitable data post-processing. The sliding window approach has proven to be more effective in measuring room transitions, while the machine learning approach performed better at estimating the permanence time in the various rooms.
From the trajectory analysis, we have identified some issues in the museum design and visit experience that can be considered by curators: for example, rooms of the same size have drastically different time of permanence, as it happens for \emph{Caravaggio} and \emph{Satiro su delfino}.
This suggests that the museum can benefit from a rearrangement of the artworks, although this is not always possible due to historical or architectural constraints.
In addition, rooms like \emph{Paolina} have an uneven distribution of visitors, being congested in the first half of the visit turn and under-used in the second half.
The museum simulator allowed us to propose the implementation of a new ticketing and entrance system.
The entry scheme identified is to let 100 people enter every 30 minutes from \emph{Portico} and \emph{Ratto di Proserpina} while eliminating the 2h time limit, thus reducing congestion and fluctuations of the number of people in each room.
\medskip
In the next future, we plan to further improve the model presented here. In particular, we aim at including the internal dynamics of \emph{social groups} (families, friends, guided tours), and at considering the impact of congestion on individual behavior. This is to lift the current statistical independence of simulated trajectories, thus increasing the level of complexity.
The impact of visitors on the local microclimate is also an outstanding issue to which we aim. On the basis of the present work and \cite{desantoli2016}, one can achieve a coupled model for predicting temperature, humidity, and crowding, on which basis one can program intelligent air conditioning systems.
\section*{Acknowledgements}
We would like to thank Sara Suriano, Massimiliano Adamo, Federico Ricci Tersenghi, Elisabetta Giani, and all the staff of Galleria Borghese for all their time and support during this project.
\section*{Funding}
Results presented in this paper are achieved under the project \textit{Management of flow of visitors inside the Galleria Borghese in Rome}, supported by Ministry of Cultural Heritage and Activities and Tourism, Galleria Borghese, and Istituto per le Applicazioni del Calcolo of National Research Council of Italy.
Project's Principal Investigators are Marina Minozzi (Galleria Borghese) and Roberto Natalini (IAC-CNR).
This work was also carried out within the research project ``SMARTOUR: Intelligent Platform for Tourism'' (No.\ SCN\_00166) funded by the Ministry of University and Research with the Regional Development Fund of European Union (PON Research and Competitiveness 2007-2013).
E. Cristiani also acknowledges the Italian Minister of Instruction, University and Research to support this research with funds coming from
the project entitled \textit{Innovative numerical methods for evolutionary partial differential equations and applications} (PRIN Project 2017, No.\ 2017KKJP4X).
A. Corbetta also acknowledges the support of the Talent Scheme (Veni) research programme, through project number 16771, which is financed by the Netherlands Organization for Scientific Research (NWO).
E. Cristiani is a member of the INdAM Research group GNCS.
\bibliographystyle{plain}
|
\section{Introduction}
\IEEEPARstart{T}{he} penetration of
new renewable sources of electrical energy is currently increasing in most electric power grids worldwide, as more and more traditional power plants are phased out.
A major concern is obviously that this substitution reduces the available inertia while it simultaneously induces
larger fluctuations in power generation~\cite{Ulb14}. Both changes may jeopardize grid stability, either individually or taken together.
A key issue is accordingly to evaluate how much power grids need to be adapted
to their resulting new modes of operation -- for instance through line extensions or deployment of resources providing ancillary services~\cite{AnnualEnOutlook}.
To ensure the stability of the grid and the safety of power supply, it is important to clarify the role of the generators dynamical parameters that will be affected by this transition, namely the
rotational
inertia and the frequency damping / droop control. Both are going to be globally reduced, moreover their geographical distribution will be modified. To try to identify
which of these dynamical parameters are most crucial, where they should be primarily deployed, as well as when and in what operating conditions they are most needed, analytical and
numerical works have investigated the robustness of electric power grids under external disturbances.
The response of power grids to external disturbances has been investigated through
quadratic performance metrics~\cite{Bam13,Dor14a,Teg15,Sia16,Poo17,Pag17,Gru18,Col19,Pag19b, Yan19},
eigenvalue damping ratios and frequency overshoots~\cite{Mev16}, rate of change of frequency\cite{Pag17,Guo18} or
disturbance wave propagation~\cite{Tam18,Hae18}. Except for numerical results~\cite{Mev16,Tam18,Hae18},
these works considered disturbances with infinitely short time scales, such
as white noise power fluctuations or instantaneous power injection changes~\cite{Bam13,Teg15,Sia16,Poo17,Gru18,Col19,Pag19b, Yan19},
or with infinitely long time scales, such as step changes in power injection~\cite{Pag17,Guo18}. All of the analytical works
relied on one of the two homogeneity assumptions that
inertia and frequency damping are the same everywhere, or that their ratio is.
These homogeneity assumptions
are not representative of real power grids, where in particular, consumer nodes are inertialess but with small, albeit finite frequency damping~\cite{Ber81,Mac08}.
Often, this inconsistency is circumvented by invoking a prior Kron reduction absorbing the inertialess nodes into an effective network.
One then measures the robustness of that reduced network, which may or may not be related to the robustness of the original one, because Kron reduction does not
capture the dynamics of the reduced, inertialess nodes.
To the best of our knowledge, Ref.~\cite{Pag19b} is the first work that tolerates deviations from homogeneity in an analytical calculation of a quadratic
performance metric. Its results suggest that grid robustness is crucially sensitive to
the geographic distribution of frequency control, while inertia has to be distributed rather evenly. This conclusion has to be revisited because,
first, Ref.~\cite{Pag19b} is based on an approximate method tolerating only small deviations from homogeneity and second, the only disturbances it considers
are long power losses.
In this manuscript, we investigate the response of power systems to colored noisy power fluctuations.
We quantify the response to these disturbances by a quadratic performance metric measuring the primary control effort
necessary to absorb the fault~\cite{Poo17}. Our analytical approach still relies on a homogeneity assumption.
Our results however emphasize the role played by the different time scales in the problem:
the performance metric depends on inertia only when the characteristic time scale of the disturbance is long compared to all other time scales in the system.
In that case we conjecture, and confirm numerically, that our analytical results also apply to heterogeneous systems.
In continental-size
high-voltage power grids, the network time scales are shorter than few seconds~\cite{Tyl18b}. Therefore, noise fluctuating on time scales of tens of second or
more induces transients whose amplitude, duration and oscillations are largely independent of inertia, and our analytical results directly apply to most
disturbances on realistic, inhomogeneous high voltage power grids.
The manuscript is organized as follows. Section \ref{notation} defines our mathematical notations.
Section \ref{section2} defines the power network model and gives an analytical expression for its linear response. Section \ref{section3} introduces performance metrics and gives analytical expressions for them. Of particular interest are the short- and long-correlation time asymptotics. In Sec.~\ref{section5}, we numerically confirm our theory on both the IEEE 118-Bus test case and the PanTaGruEl model of the synchronous grid of continental Europe. We discuss time scales in such
high-voltage power grids and show that the corresponding performance metric is given by the long noise correlation time asymptotic limit.
Our conclusions are given in section \ref{section6}.
\section{Mathematical Notation}\label{notation}
Given a vector $\bm v\in \mathbb{R}^n$, we denote its transpose by $\bm v^\top$. We write $\bm M = {\rm diag}\{m_i\}$
for the diagonal $n \times n$ matrix with $m_1,m_2,...,m_n \in \mathbb{R}$ on its diagonal. The $j$-th unit vector with a single nonzero component is
$(\hat{e}^j)_i=\delta_{ij}$\,. The scalar product of two vectors $\bm u, \bm v\in \mathbb{R}^n$ is written $\bm u^\top\bm v$ and the scalar product of a vector with itself
is $\bm v^2=\bm v^\top\bm v$. The statistical average of a random variable $x\in\mathbb{R}$ is $\overline{x}$\,. Finally, considering a diagonal matrix $\bm M$\,, we denote
its $p^{\rm th}$ power as $\bm M^p={\rm diag}\{m_i^p\}$\,.
\section{Power Grids and their Response to Fluctuating Power Injections}\label{section2}
\subsection{Swing dynamics near synchrony}
Transient dynamics in high-voltage power networks is commonly modelled by the swing equations which describe the dynamics of voltage angles assuming constant voltage amplitudes. In the lossless line approximation, appropriate to very high-voltages~\cite{Mac08},
they read
\begin{eqnarray}\label{eq:kuramoto}
m_i\,\dot{\omega}_i+d_i\,{\omega}_i=P_i-\sum_j b_{ij}\sin(\theta_i-\theta_j),
\end{eqnarray}
where each network node is labeled $i=1,...,n$ with a voltage angle $\theta_i$. Equation~\eqref{eq:kuramoto} is written
in a rotating frame, so that the frequency $\omega_i=\dot{\theta}_i$ refers to the deviation from the rated frequency of $50$ or $60$ Hz. Each node has inertia and damping control parameters $m_i$ and $d_i$ respectively, and an active power $P_i$ that is generated ($P_i>0$) or consumed ($P_i<0$). The coupling between node $i$ and $j$
is given by the susceptance $b_{ij} \ge 0$ of the corresponding power line. The operational state ${\bm \theta}^{(0)}$ is a synchronous
stationary solution to Eq.~(\ref{eq:kuramoto}).
Equation~(\ref{eq:kuramoto}) is governed by two sets of time scales. The first set is given by the ratio between inertia and damping coefficients $\gamma_i^{-1}\equiv m_i/d_i$. It corresponds to the local relaxation of synchronous machines. The second one is determined by the network characteristic time scales $d_i/\lambda_\alpha$
given by damping coefficients and the eigenvalues of the weighted Laplacian, see Eq.~(\ref{eq:laplacian}) below.
Depending on these two sets of time scales, perturbations are locally damped or spread across the network. In a synthetic synchronous grid of continental Europe
with constant damping and inertia corresponding to the average of their true values,
it has been found that
all modes are underdamped and propagate through the whole system with $d/\lambda_\alpha < \gamma^{-1}$, $\forall \alpha$~\cite{Tyl18b}.
We next investigate the response of the system to a time-dependent disturbance ${\bm{P}}(t) = {\bm{P}} + \delta {\bm{P}}(t)$ acting on
the operational state ${\bm \theta}^{(0)}$, following which angles become time-dependent,
${\bm{\theta}}(t) = {\bm{\theta}}^{(0)} + \delta {\bm{\theta}}(t)$. The small-signal response is governed by dynamical equations obtained after linearizing Eq.~(\ref{eq:kuramoto}) about
$\bm{\theta}^{(0)} $,
\begin{align}\label{eq:kuramoto_lin}
{\bm M}\,\dot{\bm \omega} + {\bm D}\, {\bm \omega} &= \delta {\bm P}(t) - {\mathbb L}({\bm\theta}^{(0)}) \, \delta {\bm \theta} \, ,
\end{align}
where we introduced inertia and damping matrices, ${\bm M}={\mathrm{diag}}\{m_i\}$ and ${\bm D}={\mathrm{diag}}\{d_i\}$ and the weighted Laplacian matrix ${\mathbb L}(\{ \theta_i^{(0)} \})$ with matrix elements
\begin{equation}\label{eq:laplacian}
{\mathbb L}_{ij} =
\left\{
\begin{array}{cc}
-b_{ij} \cos(\theta_i^{(0)} - \theta_j^{(0)}) \, , & i \ne j \, , \\
\sum_k b_{ik} \cos(\theta_i^{(0)} - \theta_k^{(0)}) \, , & i=j \, .
\end{array}
\right.
\end{equation}
This Laplacian is minus the stability matrix of the linearized dynamics, and
since we consider a stable synchronous state, it
is positive semidefinite, with a single vanishing eigenvalue $\lambda_1=0$ with eigenvector ${\bm u}_1=(1,1,...1)/\sqrt{n}$. All other eigenvalues are positive,
$\lambda_\alpha>0$, $\alpha=2,3,...n$.
Equation~(\ref{eq:kuramoto_lin}) can be integrated via a spectral decomposition provided either: (i) both $\bm M$ and $\bm D$ commute with $\mathbb{L}$, then Eq.~(\ref{eq:kuramoto_lin}) can be integrated in the eigenspace of $\mathbb{L}$, or (ii) ${{\bm M}^{-1}{\bm D}}=\gamma\,\mathbb{I}$\,, in which case Eq.~(\ref{eq:kuramoto_lin}) can be integrated in the eigenspace of ${\bm D^{-1/2}}\,\mathbb{L}\,{\bm D^{-1/2}}$.
\subsection{Analytical solution for constant damping-to-inertia ratio}
We consider the case (ii) above of constant inertia-to-damping ratio, $m_i/d_i=\gamma^{-1}$ $\forall i$\,. To calculate the response of the system,
we first perform a change of variable ${\delta\bm \varphi}={\bm D}^{1/2}\delta {\bm \theta}$
on Eq.~(\ref{eq:kuramoto_lin}). We obtain
\begin{align}\label{eq:kuramoto_lin3}
\gamma^{-1}\delta \ddot{{\bm \varphi}} + \delta \dot{{\bm \varphi}} &= {\bm D}^{-1/2}\delta {\bm P} - {\bm D}^{-1/2}\,\mathbb{L}\,{\bm D}^{-1/2}\, \delta {\bm \varphi} \, .
\end{align}
We choose this normalization with $\bm D$, rather than with $\bm M$ as proposed in Ref.~\cite{Pag17}, because it allows us to treat
the inertialess case with $\gamma^{-1}=0$, from which we will extrapolate
the realistic case where consumer nodes are inertialess and generator nodes have nonhomogeneous inertia.
Equation~(\ref{eq:kuramoto_lin3}) can be solved by expanding angle deviations as $\delta{\bm \varphi}(t)=\sum_{\alpha}c_\alpha(t)\,{{\bm u}_\alpha^{D}}$,
over the eigenvectors ${\bm u}_{\alpha}^{D}$ of $\mathbb{L}^D={\bm D}^{-1/2}\,\mathbb{L}\,{\bm D}^{-1/2}$.
The matrix $\mathbb{L}^D$ is no longer Laplacian but it still has a zero-mode ${\bm u}_1^D=(\sqrt{d_1},...,\sqrt{d_n})/\sqrt{\sum_i d_i}$. Angle shifts of $\delta\bm \varphi$ along ${\bm u}_1^D$ do not modify the synchronous state because ${\bm u}_1\propto{\bm D}^{-1/2}{\bm u}_1^D$. Note also that, by orthogonality with ${\bm u}_1^D$, eigenvector components
must satisfy $\sum_i \sqrt{d_i}u_{\alpha,i}^D=0$ for $\alpha\ge 2$.
\begin{prop}\label{prop1}
\textit{The general solution to Eq.~(\ref{eq:kuramoto_lin3}) reads}
\begin{equation}\label{eq:calpha2}
\begin{aligned}
\delta\varphi_i(t)&= \sum_\alpha \gamma e^{\frac{-\gamma-\Gamma_{\alpha}}{2}t}\int_0^{t}e^{{\Gamma_{\alpha}}t_1} \\
&\times \int_{0}^{t_1}[{\bm D}^{-1/2}\delta {\bm{P}}(t_2)]^\top {\bm{u}}_{\alpha}^De^{\frac{\gamma-\Gamma_{\alpha}}{2}t_2}{\rm d}t_2{\rm d}t_1 \;u_{\alpha,i}^D \;,
\end{aligned}
\end{equation}
\textit{with $\Gamma_\alpha=\sqrt{\gamma^2-4\lambda_\alpha^D\gamma}$ where $\lambda_\alpha^D$ is the eigenvalue associated with the eigenvector ${\bm u}_\alpha^D$ of $\mathbb{L}^D$.}
\end{prop}
\begin{IEEEproof}
We first expand angle deviations over
the eigenbasis of $\mathbb{L}^D$ as $\delta\varphi_i(t)=\sum_{\alpha}c_\alpha(t)\, u_{\alpha,i}^D$. From the orthogonality of the
eigenbasis, $({\bm{u}}_\alpha^D)^\top {\bm{u}}_\beta^D=\delta_{\alpha \beta}$, one straightforwardly rewrites Eq.~(\ref{eq:kuramoto_lin3}) as
\begin{eqnarray}\label{eq:ca}
\gamma^{-1}\ddot{c}_\alpha + \dot{c}_\alpha = ({\bm D}^{-1/2}\delta {\bm P})^\top {\bm{u}}_\alpha^D - \lambda_\alpha^D c_\alpha\;
\end{eqnarray}
for $\alpha=1,...,n$\,. The expansion coefficients $c_\alpha(t)$ can be read from
Eq.~(\ref{eq:calpha2}), and direct differentiation shows that they
solve this equation. This completes the proof.
\end{IEEEproof}
\section{Dynamical Parameters and Transient Excursions}\label{section3}
\subsection{Quantifying frequency excursions}\label{sec:perf}
To evaluate the global response of the system to an external disturbance, we use the following performance metric
\begin{align}\label{eq:c1}
\begin{split}
{\mathcal P}(T) &= T^{-1} \, \int_0^T \,({\bm \omega}^\top-\overline{{\bm \omega}}^\top){\bm D}({\bm \omega}-\overline{{\bm \omega}})\, {\rm d}t \, ,
\end{split}
\end{align}
because it measures the primary control effort and therefore has a physical meaning~\cite{Poo17}. The quantity
$\overline{{\bm \omega}}^\top=(\dot\Delta,\dot\Delta,...,\dot\Delta)$ with $\dot\Delta (t) = \sum_i d_i\omega_i(t)/\sum_i d_i$ gives the average frequency deviation over
all nodes in the system.
Because synchronous states are defined modulo any
homogeneous angle shift, the transformation $\theta_i^{(0)} \rightarrow \theta_i^{(0)} + C$ does not change the
synchronous state. Accordingly only phase and frequency shifts with $\sum_i \delta \theta_i(t)=0$ and $\sum_i d_i \omega_i(t)=0$ matter. This is included
in ${\mathcal P}$ by subtracting the average $\overline{{\bm \omega}}$. That ${\mathcal P}(T)$ is a performance metric is easily understood:
low values indicate that the system absorbs the perturbation with little fluctuations, while large values
indicate a large transient excursion around the initial synchronous state.
Using the above change of variables, $\delta\dot{{\bm{\varphi}}}={\bm D}^{1/2}{\bm \omega}=\sum_{\alpha}\dot{c}_\alpha(t)\, {\bm u}_{\alpha}^D$\,,
Eq.~(\ref{eq:c1}) becomes,
\begin{align}\label{eq:perf2}
\begin{split}
\mathcal{P}(T)&=T^{-1} \sum_{\alpha \ge 2}\int_0^T\dot{c}_\alpha^2(t){\rm d}t \; .
\end{split}
\end{align}
This can be calculated using the explicit expression for $c_\alpha(t)$ from Eq.~(\ref{eq:calpha2}), once a perturbation $\delta {\bm P}(t)$ is given.\\
\begin{prop}\label{prop2}
\textit{Consider a noisy disturbance acting on $N_n$ of the network nodes. The noise ensemble is Gaussian and defined by its
vanishing first moments, $\overline{\delta P_i(t)}=0$, and its second moments $\overline{\delta P_i(t)}\overline{\delta P_j(t')}=\delta_{ij}\delta P_{0i}^2\exp[-|t-t'|/\tau_0]$ with the noise correlation time $\tau_0$. The performance metric $\overline{{\mathcal P}^\infty}$ for primary control effort averaged over this noise ensemble is given by}
\begin{align}\label{eq:P2}
\begin{split}
\overline{\mathcal{P}^\infty}&=\sum_{\alpha\ge 2}\frac{\sum_{i\in N_n} \delta P_{0i}^2{u_{\alpha,i}^D}^2d_i^{-1}}{\lambda_\alpha^D\tau_0+1+\gamma^{-1}\tau_0^{-1}} \; .
\end{split}
\end{align}
\end{prop}
\begin{cor}
\textit{In the limit of short noise correlation time, $\tau_0\ll \gamma^{-1},{\lambda_\alpha^D}^{-1}$ one has}
\begin{align}
\begin{split}\overline{\mathcal{P}^\infty}&=\tau_0 {\sum_{i\in N_n} \delta P_{0i}^2 \, ( 1/m_i - 1 / \sum_j m_j )}\; .\label{eq:P2s}
\end{split}
\end{align}
\end{cor}
\begin{cor}
\textit{In the opposite asymptotic limit, $\tau_0\gg \gamma^{-1},{\lambda_\alpha^D}^{-1}$ one has}
\begin{align}
\begin{split}
\overline{\mathcal{P}^\infty}&=\tau_0^{-1}\sum_{\alpha\ge 2}\frac{\sum_{i\in N_n} \delta P_{0i}^2{u_{\alpha,i}^D}^2d_i^{-1}}{{\lambda_\alpha^D}} \; .\label{eq:P2l}
\end{split}
\end{align}
\end{cor}
\begin{IEEEproof} Inserting the time derivative of Eq.~\eqref{eq:calpha2} into Eq.~(\ref{eq:perf2}) and taking the average over the noise ensemble
with $\overline{\delta P_i(t)}\overline{\delta P_j(t')}=\delta_{ij}\delta P_{0i}^2\exp[-|t-t'|/\tau_0]$
gives Eq.~\eqref{eq:P2}, with few straightforwardly calculated exponential integrals. The two asymptotic results \eqref{eq:P2s} and \eqref{eq:P2l}
are easily obtained by a Taylor expansion, keeping only the first non-vanishing term. To obtain Eq.~\eqref{eq:P2s},
we also used $\sum_{\alpha\ge 2}{u_{\alpha,i}^D}^2 = \sum_{\alpha\ge 1}{u_{\alpha,i}^D}^2 - {u_{\alpha,1}^D}^2 =1-d_i/\sum_i d_i$.\\
\end{IEEEproof}
{\textit{Remark 1: The short correlation time asymptotic of Eq.~\eqref{eq:P2s} agrees with the result of~\cite{Poo17} obtained for either single-pulsed or
averaged white-noise perturbations.}}
{\textit{Remark 2: The noise correlators are defined either as time averages, $\overline{\delta P_i(t)}\overline{\delta P_j(t')} = {\rm lim}_{\tau \rightarrow \infty} \tau^{-1}
\int_0^\tau \delta P_i(t+\tau') \delta P_j(t'+\tau') {\rm d}\tau$ or as averages over different noise sequences.
}}
{\textit{Remark 3: Finite-time corrections to Eqs.~\eqref{eq:P2}--\eqref{eq:P2l} disappear
as ${\cal O}(1/T)$ as $T \rightarrow \infty$.
}}\\
\begin{prop}\label{prop3}
Under the same assumptions as Proposition~\ref{prop2}, the variance ${\rm var}\left[{\mathcal P}(T)\right]$ of the performance metric for primary control effort
over the noise ensemble vanishes as $\sim T^{-1} + {\cal O}(1/T^2)$ as $T \rightarrow \infty$.
\end{prop}
The proof proceeds through direct calculation of ${\rm var}\left[{\mathcal P}(T)\right]$. It
is too long to fit in this article and here we only sketch it.
From Eq.~\eqref{eq:perf2} one has
\begin{align}\label{eq:varP}
\begin{split}
{\rm var}\left[{\mathcal P}(T) \right]&=
T^{-2} \sum_{\alpha,\beta \ge 2}\iint_0^T \overline{\dot{c}_\alpha^2(t)\dot{c}_\beta^2(t')}{\rm d}t {\rm d}t' - \overline{\mathcal{P}(T)}^2 \, .
\end{split}
\end{align}
From Eq.~\eqref{eq:calpha2},
each $\dot{c}_{\alpha,\beta}$ contains a noise term $\delta {\bm P}$. The noise average in the first term on the right-hand side of
Eq.~\eqref{eq:varP} therefore consists in pairings of four noise terms.
There are three such contributions. The first one pairs the two $\delta {\bm P}$'s in $\dot{c}_{\alpha}^2(t)$
and the two $\delta {\bm P}$'s in $\dot{c}_{\beta}^2(t)$. This contribution is cancelled by a similar pairing in $\overline{\mathcal{P}(T)}^2$. The other two contributions
pair $\delta {\bm P}$'s across indices $\alpha$ and $\beta$ and accordingly, they constrain the values that $t$ and $t'$ can take with respect to one another,
$|t-t'| \lesssim \tau_0$. Accordingly, the double time integral in Eq.~\eqref{eq:varP} gives a contribution $\sim T \tau_0$, instead of $\sim T^2$, resulting in
${\rm var}\left[{\mathcal P}(T)\right] \sim T^{-1}$.
{\textit{Remark 4: Proposition~\ref{prop3} means that for specific noisy disturbances satisfying the
assumption of Proposition~\ref{prop2} and for long enough observation times $T \gg \tau_0$,
$\mathcal{P}^\infty = \overline{\mathcal{P}^\infty} + {\cal O}(T^{-1/2})$. The statistical average is therefore representative of specific noise disturbances for
sufficiently long observation time. The validity of Proposition~\ref{prop3} is illustrated numerically in Fig.~\ref{fig2} (c).}}
The two asymptotic limits of large and small $\tau_0$ are particularly interesting as they shed light on the influence of dynamical parameters,
in particular on the interplay between local disturbance absorption by inertia and long-range propagation through low-lying network modes.
First, in the short correlation time limit given by Eq.~(\ref{eq:P2s}), $\overline{\mathcal{P}^\infty}$ explicitly depends on inertia but not on the
coupling network. This reflects the fact that, in the white-noise limit, the perturbation remains local and is easily absorbed, if
there is enough inertia. Second,
Eq.~(\ref{eq:P2l}) shows that, in the long correlation time limit, $\overline{\mathcal{P}^\infty}$ does not depend on
inertia. This suggests that changing inertia in any direction
will not change $\overline{\mathcal{P}^\infty}$ in the limit of long noise correlation time. This is a conjecture since Eq.~(\ref{eq:P2l}) has been derived
under the assumption of constant damping-to-inertia ratio, $\gamma=d_i/m_i$. Below we numerically confirm this conjecture.
Simultaneously, Eq.~(\ref{eq:P2l}) also shows that, in the long correlation time limit, $\overline{\mathcal{P}^\infty}$
is determined by the structure of the coupling network, with the modes with smallest eigenvalues having the largest influence. Those modes
are extended over the whole network in large power grids, as is illustrated in Fig.~\ref{fig3} (b). Accordingly, in the limit of long noise correlation time,
the disturbance is able to propagate over large distances in the network, and inertia has little influence on this large-scale propagation.
\begin{figure*}
\centering
\includegraphics[width=0.9\textwidth]{fig1.pdf}
\caption{Comparison between numerical calculations obtained by time-evolving Eq.~(\ref{eq:kuramoto}) and the theoretical result of Eq.~(\ref{eq:P2l}) for the primary control effort,
for short correlation time $\gamma\tau_0 = 4\times 10^{-3}$ (a) and long correlation time $\gamma\tau_0 = 40$ (b)
in the IEEE 118-Bus test case. Averages are made over $10$ noisy sequences and standard deviations are shown by barely visible vertical line. Blue crosses correspond to constant damping-to-inertia ratio, $d_i/m_i=\gamma=0.4s^{-1}$, while orange symbols correspond to inhomogeneous,
non-vanishing inertia on generator nodes (squares) and inertialess consumer nodes (crosses). On generation nodes one has $\gamma_i\tau_0\in [30 , 60]$\,. (c) Ratio between standard deviation and average of the primary control effort for the IEEE 118-Bus test case. For long enough $T$, the ratio scales as $T^{-1/2}$, confirming Prop.~\ref{prop3}. Averages are made over $40$ different noise sequences.}\label{fig2}
\end{figure*}
{\textit{Remark 5: It is important to realize that
fluctuations of renewable energy sources occur on time scales that are large compared to the intrinsic time scales of the power system~\cite{Mei11}. As a matter of fact, time scales in the synchronous grid of continental Europe have been found to
satisfy $\gamma^{-1} \simeq 2.5 s$ and ${\lambda_\alpha^D}^{-1} \lesssim 0.5 s$~\cite{Tyl18b}.
Provided the above conjecture is corroborated, Eqs.~\eqref{eq:P2}--\eqref{eq:P2l} suggest that fluctuations from new renewables excite network modes and
are efficiently absorbed by optimizing the distribution of damping with little regard for inertia. This conjecture is numerically confirmed below.}
{\textit{Remark 6: Similar conclusions as in Eqs.~\eqref{eq:P2}--\eqref{eq:P2l} regarding local inertia absorption vs. large-scale mode propagation
are obtained in the case of step disturbances corresponding to sudden power losses, as a function of their duration $\tau_0$.}
{\textit{Remark 7: The lossless line approximation used in this paper does not account for ohmic dissipation. We expect that the latter enhances mode damping
and accordingly undermines disturbance propagation in the case of noise with long correlation time, but that it affects only marginally our result for short correlation time.
Investigations beyond the lossless line approximation would be very welcome but lie beyond the scope of the present paper.}
\section{Numerical Simulations}\label{section5}
\subsection{Dynamical parameters for simulations}\label{section5a}
We consider two different cases: (i) cases with homogeneous damping-to-inertia ratio $d_i/m_i=\gamma$; (ii) realistic heterogeneous cases, using the nonlinear swing dynamics of Eq.~(\ref{eq:kuramoto}).
In the homogeneous case, we used $d_i = \alpha |P_i^{(0)}|/\omega_0$, where $P_i^{(0)}$ is the produced/consumed power at nominal frequency $\omega_0$,
and set $m_i=\gamma^{-1}d_i$.
In the heterogeneous case, the inertia parameter vanishes on consumer nodes and is $m_i = 2H_i |P_i^{(0)}|/\omega_0$ on generator nodes,
where $H_i$ depends on the type of generator~\cite{Mac08}.
Damping is given by Eq.~(5.24) and Table 4.3 in Ref.~\cite{Mac08} for generators and by
$d_i = \alpha |P_i^{(0)}|/\omega_0$ for consumer nodes. In all
cases we use $\omega_0=2\pi\times 50Hz$. For the IEEE-118 Bus test case discussed in Section \ref{sec:ieee118}, $\alpha=1.5$ and $H_i=5s$. For the PanTaGruEl European model discussed in Section \ref{sec:panta}), $\alpha=1.5$ and $H_i$ varies according to the generator type as described in Ref.~\cite{Pag19c}.
\subsection{IEEE 118-Bus test case}\label{sec:ieee118}
The main prediction from Eqs.~\eqref{eq:P2}-\eqref{eq:P2l} is that, for noise correlation $\tau_0$ that is longer than the
other characteristic time scales in the system, the performance metric $\overline{\mathcal{P}^\infty}$ does not depend on inertia parameters. This was conjectured from
Eq.~\eqref{eq:P2l} where inertia does not appear.
Fig.~\ref{fig2} shows performance metric obtained from individual noisy disturbance of a single node, repeating the operation for all nodes in the IEEE 118-Bus test case.
As disturbance, we take Gaussian noise with the same first two moments as in Proposition~\ref{prop2}.
First, the primary control effort $\overline{\mathcal{P}^\infty}$ is calculated for a constant damping-to-inertia ratio (blue crosses), then for a distribution of dynamical parameters where generator nodes have inertia while consumer nodes do not (orange squares and crosses). As predicted by Eq.~(\ref{eq:P2l}), for long correlation time of the noise [panel (b), $\gamma \tau_0=40$],
the two distributions of dynamical parameters give the same $\overline{\mathcal{P}^\infty}$, corroborating our conjecture that it does not depend on inertia. We have found (but do not show) that
the performance metric only depends on the damping distribution in that case.
For short noise correlation time, on the other hand,
Eq.~(\ref{eq:P2s}) explicitly depends on the damping-to-inertia ratio $d_i/m_i=\gamma$\,, and we expect that the numerical data will differ from the theoretical
prediction once this ratio is no longer constant. This is confirmed in
Fig.~\ref{fig2} [panel (a), $\gamma \tau_0=4\times 10^{-3}$] where with $d_i/m_i=\gamma$, numerical data fall on the theory (blue crosses). However, once
$d_i/m_i$ is no longer constant, numerical data and theoretical prediction differ significantly (orange symbols).
Quite interestingly, we found that for noisy perturbations on generator nodes
with inertia, the theory still gives a remarkably accurate estimate for the primary control effort $\mathcal{P}^\infty$. An understanding of this remarkable agreement would be highly welcome, particularly since it could justify dynamic performance analysis on Kron reduced networks.
\subsection{Time scales in high-voltage electric power grids}
We have shown that the primary control effort against fluctuating disturbances in the form of colored noise
behaves very differently depending on the position of the noise correlation time relative to the characteristic time scales in the system.
Furthermore the primary control effort is captured by our theory even for inhomogeneous dynamical parameters, when the noise correlation time is long enough.
It is therefore desirable to identify
what regime applies to a realistic high-voltage power grid subjected to fluctuating sources of power, in particular those generated by
new renewable sources of energy.
To that end, we consider in the next paragraph a realistic model of the synchronous grid of continental Europe~\cite{Tyl18b,Pag19c},
with the following time scales
\begin{subequations}\label{eq:time}
\begin{align}
{\lambda_\alpha^D}^{-1}&< 0.5s\;, \; {\text{for $\alpha=2,...,n$}\,,}\\
\gamma^{-1}&=\frac{\langle m_i \rangle}{\langle d_i \rangle}= 2.5s\, ,
\end{align}
\end{subequations}
where $\langle \;\ldots\; \rangle$ means that we take the average over all nodes in the grid. It is commonly accepted that
power fluctuations from renewable energy sources such as wind turbines or photovoltaic panels
fluctuate on time scales that are larger than both time scales in Eq.~(\ref{eq:time})~\cite{Mei11}. Therefore, the asymptotic limit
of large noise correlation time, corresponding to Eq.~(\ref{eq:P2l}) applies, and we expect that the primary control effort as measured by
Eq.~\eqref{eq:c1} is influenced only by damping, and not by inertia. Numerical results to be presented in the next paragraph corroborate this expectation.
As a side-remark we note that when the perturbation corresponds to the sudden disconnection of a power generator,
controls usually try to reconnect the bus several times quickly after the fault (typically within few AC cycles).
The typical time scale for such a perturbation is then less than several if not all system's time scales, and inertia obviously matters to absorb such sudden faults, as it is predicted by Eq.~(\ref{eq:P2s}).
\subsection{The PanTaGruEl European model}\label{sec:panta}
To further illustrate the influence of dynamical parameters on the primary control effort,
we numerically compute Eq.~(\ref{eq:c1}) on the large-scale PanTaGruEl model of the European high-voltage transmission grid~\cite{Pag19c}.
The model is shown on Fig.~\ref{fig3} (b). It has 3809 nodes and 4944 lines. More details can be found in Ref.~\cite{Pag19c}.
We considered three different cases, (i) a homogeneous situation that corresponds to today's grid in terms of its global amount of inertia, with $\gamma\tau_0=4$ and $\gamma=d_i/m_i$ constant,
(ii) a homogeneous situation where the inertia is reduced by a factor $10$\,, i.e. $\gamma\tau_0=40$, and $\gamma=d_i/m_i$ constant,
and (iii) a realistic situation with inhomogeneous damping parameters and where inertia vanishes on consumer nodes and is inhomogeneous on production nodes
as described in Section~\ref{section5a} and Ref.~\cite{Pag19c}, with
$\tau_0$ larger than all other time scales in the system.
Remarkably, Fig.~\ref{fig3} (a) shows that the primary control effort for all cases is well predicted by Eq.~(\ref{eq:P2l}). This confirms our main finding
that, for fluctuations with a correlation time longer than any other characteristic time scale in the system, inertia does not affect the primary control effort, Eq.~(\ref{eq:P2l}).
Quite surprisingly, an overall reduction of the total available inertia by a factor of 10 does not affect the primary control
effort, Eq.~\eqref{eq:c1}.
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{fig2.pdf}
\caption{(a) Comparison between numerical calculations obtained by time-evolving Eq.~(\ref{eq:kuramoto}) and the theoretical result of Eq.~(\ref{eq:P2l}) for the PanTaGruEl model of the synchronous grid
of continental Europe~\cite{Pag19c} shown on panel (b).
Three cases are considered. The first two are grids with $\gamma=d_i/m_i$ constant,
with today's average inertia (orange crosses) and an inertia reduced by a factor of 10 (blue crosses). The third case corresponds to a realistic situations
as discussed in the text, with $\gamma_i\tau_0\in [ 20,1600 ]$ on generation nodes.
In all cases, $\tau_0$ is the longest time scale,
consequently, the inertia-independent theoretical prediction of Eq.~\eqref{eq:P2l} accurately captures all numerical data. Averages are made over $10$ noisy sequences. (b) Network eigenmodes of $\mathbb{L}^D$
with the first two non-vanishing eigenvalues. These slow modes are extended over the whole network, with higher amplitudes on peripheral nodes.
Disturbances on the four buses highlighted in light blue correspond to the smallest primary control effort. These buses
lay in the center of the network where the slow modes have small amplitudes.
Disturbances on the four buses in dark red, on the other hand, have largest primary control effort. They are located at the periphery of the network
where the slow modes have large amplitudes~\cite{Tyl18b}. This shows that large primary control effort for noise with large correlation time correspond
to excitations of slow network modes, which in their turn propagate the disturbance over large distances in the network.}\label{fig3}
\end{figure*}
\section{Conclusion}\label{section6}
With the ongoing energy transition resulting in strongly increased penetrations of new renewable sources of electrical energy, a question of crucial importance is how
grid stability will evolve, given the resulting reduction of globally available rotational inertia and enhanced power fluctuations. We have shown that reduced inertia may pose
problems only for perturbation occurring/fluctuating on very short time scales, shorter than all other characteristic time scales in the system. In continental-size transmission grids,
these time scales are shorter than few seconds, consequently, power fluctuations from new renewables will not affect grid stability per se.
Inertia is of course important to absorb
sudden faults occurring on very short time scales such as line faults or disconnection/reconnection of large power plants and so forth. Simultaneously, our result
of Eq.~\eqref{eq:P2s} indicates that the resulting primary control effort is independent of the grid topology. Accordingly,
optimal inertia distribution needs to follow the distribution of potential faults, for instance being larger in regions with higher density of generators. A similar conclusion was
drawn in Refs.~\cite{Poo17} and \cite{Pag19b}.
|
\section{Introduction}
In this paper we will give new proofs for the following theorems of Schrijver. The first theorem is about perfect matchings of regular bipartite graphs.
\begin{Th}[Schrijver \cite{Sch2}] \label{matching}
Let $G=(A,B,E)$ be $d$--regular bipartite graph on $2n$ vertices. Let $\prm(G)$ denote the number of perfect matchings of $G$. Then
$$\prm(G)\geq \bigg( \frac{(d-1)^{d-1}}{d^{d-2}} \bigg)^{n}.$$
\end{Th}
The next theorem is about Eulerian orientations of regular graphs. Recall that an orientation of a graph $G$ is Eulerian if the in-degree and out-degree are equal at each vertex. (In particular, the degree of a vertex must be even.)
\begin{Th}[Schrijver \cite{Sch1}] \label{orientation}
Let $G$ be a graph, where the degree of the vertex $v$ is $d_v$. Suppose that $d_v$ is even for each $v$. Let $\e(G)$ denote the number of Eulerian orientations of the graph $G$. Then
$$\e(G)\geq \prod_{v\in V(G)}\frac{\binom{d_v}{d_v/2}}{2^{d_v/2}}.$$
\end{Th}
The main goal of this paper is to show that these two theorems have a common generalization. To spell out this generalization we will count the number of orientations in a graph with prescribed in-degree sequence.
\begin{Def}
Let $\underline{r}=(r_v)_{v\in V(G)}\in \mathbb{Z}^{V(G)}$. Let $\e_{\underline{r}}(G)$ denote the number of those orientations of the graph $G$, where the in-degree of the vertex $v$ is $r_v$.
\end{Def}
Observe that if we have a $d$--regular bipartite graph $G=(A,B,E)$, then the perfect matchings are in bijection with those orientations of the graph, where the in-degree of the vertices in $A$ is $1$, and is $d-1$ in case of the vertices of $B$. Indeed, simply orient each edge of a perfect matching towards $A$, and every other edge towards $B$. Clearly, if we have such an orientation, then the edges oriented towards $A$ form a perfect matching.
\bigskip
The following theorem might look technical, but it easily implies both Theorem~\ref{matching} and \ref{orientation}.
\begin{Th} \label{general}
Let $G=(V,E)$ be a graph with degree $d_v$ at vertex $v$, and let
$\e_{\underline{r}}(G)$ denote the number of orientations of the graph $G$, where the in-degree of the vertex $v$ is $r_v$. Then
$$\e_{\underline{r}}(G)\geq \prod_{v\in V(G)}\binom{d_v}{r_v}\bigg(\frac{r_v}{d_v}\bigg)^{r_v}\bigg(\frac{d_v-r_v}{d_v}\bigg)^{d_v-r_v}\cdot \inf_{x_u>0}\frac{\prod_{(u,v)\in E(G)}(x_u+x_v)}{\prod_{u\in V(G)}x_u^{r_u}}.$$
\end{Th}
We remark that the role of the multivariate polynomial $P_G(\underline{x}):=\prod_{(u,v)\in E(G)}(x_u+x_v)$ in the theorem comes from the fact that
$$\prod_{(u,v)\in E(G)}(x_u+x_v)=\sum_{\underline{r}}\e_{\underline{r}}(G)\prod_{u\in V(G)}x_u^{r_u}.$$
In other words, this theorem is about how to give a lower bound on a coefficient of a multivariate polynomial in terms of the polynomial. It turns out that this lower bound is possible, because the polynomial $P_G(\underline{x})$ is a real stable polynomial. The definition of real stability is the following.
\begin{Def}
A multivariate polynomial $P(x_1,\dots ,x_n)$ with complex coefficients is stable if $P(z_1,\dots ,z_n)\neq 0$ whenever $\mathrm{Im}(z_i)>0$ for $i=1,\dots ,n$. A polynomial is called real stable if it is stable and its coefficients are real.
\end{Def}
Note that a univariate polynomial with real coefficients is stable if and only if it is real-rooted. So real stability is a generalization of real-rootedness for multivariate polynomials.
We remark that Gurvits \cite{Gur1} already gave a proof of Theorem~\ref{matching} using real stable polynomials. He used the polynomial
$$Q(\underline{x})=\prod_{v\in B}\bigg(\sum_{u\in A}x_u\bigg)$$
in his proof. In this case the coefficient of $\prod_{u\in A}x_u$ is exactly $\prm(G)$. In fact, we will follow exactly the strategy of Gurvits. This strategy is based on two concepts, the real stability and the capacity of a polynomial. The latter was invented by Gurvits himself. We will review these concepts in the next section.
\bigskip
\noindent \textbf{What is new in this paper?} The proof
of Theorem~\ref{coefficient theorem} is new, the theorem itself appeared in \cite{Gur3} in a slightly different form.
The use of the polynomial $P_G(\underline{x})$ in these proofs also seems to be new, although variants of this polynomial appeared in \cite{Gur3}, but never exactly this one. Proving Theorem~\ref{orientation} via stable polynomials is also new. On the other hand, the general strategy is not new at all. In fact, one of our main goals is to advertise this theory, so this paper can be considered as a mini survey.
\bigskip
\noindent \textbf{This paper is organized as follows.} In the next section we collect the basic facts about real stable polynomials and capacity. In Section 3 we prove Theorem~\ref{general} and derive Theorem~\ref{matching} and \ref{orientation} from it. In Section 4 we collected pointers to the literature.
\section{Real stability and capacity}
In this section we review the basic properties of real stability and capacity. To keep this paper self-contained we will prove every result that we use apart from the Hermite-Sylvester criterion.
\subsection{Stability} Recall that a multivariate polynomial $P(x_1,\dots ,x_n)$ is real stable if it has real coefficients, and $P(z_1,\dots ,z_n)\neq 0$ whenever $\mathrm{Im}(z_i)>0$ for $i=1,\dots ,n$. We have seen that real stability is a generalization of univariate real-rooted polynomials. As the next lemma shows there is a more direct connection between the two concepts.
\begin{Lemma} \label{stable-real-rooted} A multivariate polynomial $P(z_1,\dots ,z_n)\in \R[z_1,\dots ,z_n]$ is stable if and only if for all $\underline{v}=(v_1,\dots, v_n)\in \R^n$ and $\underline{u}=(u_1,\dots ,u_n)\in \R^n_{>0}$, the univariate polynomial $g(t)=P(v_1+tu_1,\dots ,v_n+tu_n)$ is real-rooted.
\end{Lemma}
\begin{proof} Let $\mathbb{H}=\{z \ |\ \mathrm{Im}(z)>0\}$. First suppose that $g(t)$ is not real rooted, then it has a root $a+bi\in \C$, where $b\neq 0$. Since the coefficients of $g$ are real, $a\pm bi$ are both zeros of $g$. So we can assume that $b>0$. But then $\mathrm{Im}(v_j+u_j(a+bi))=bu_j>0$ shows that the numbers $z_j=v_j+u_j(a+bi) \in \HH$ for $j=1,\dots ,n$ and $P(z_1,\dots ,z_n)=0$, thus $P$ is not stable.
Next suppose that $P$ is not stable, it has some zero $(z_1,\dots ,z_n)\in \HH^n$. Let $z_j=a_j+b_ji$. Then $b_j>0$. Then the polynomial
$g(t)=P(a_1+tb_1,\dots ,a_n+tb_n)$ satisfies that $(a_1,\dots ,a_n)\in \R^n$ and $(b_1,\dots ,b_n)\in \R^n_{>0}$ and has a non-real zero, namely $i$.
\end{proof}
We will also need the following lemma.
\begin{Lemma} \label{real-rooted-convergence}
Given a polynomial $P(x)$ and suppose that $(P_n(x))_n$ is a sequence of real-rooted polynomials such that
$\lim_{n\to \infty}P_n(x)=P(x)$ coefficientwise. Then $P(x)$ is real-rooted.
\end{Lemma}
Lemma~\ref{real-rooted-convergence} is a simple consequence of the well-known Hermite-Sylvester criterion. For a short, simple proof see that paper \cite{Nat}.
\begin{Lemma}[Hermite-Sylvester criterion] \label{HS-criterion}
Let $P(x)$ be a non-constant polynomial of degree $d$ with real coefficients and $\lambda_1,\dots ,\lambda_d$ not necessarily distinct zeros. Let $m_k=\sum_{j=1}^d\lambda_j^k$. Then $P(x)$ is real-rooted if and only if the $d\times d$ matrix $(m_{i+j-2})_{i,j=1,\dots ,d}$ is positive semi-definite.
\end{Lemma}
Note that $m_k$ can be computed from the coefficients by the Newton-Girard identities, so they are continuous functions of the coefficients. Hence the Hermite-Sylvester criterion implies Lemma~\ref{real-rooted-convergence}.
Next we collect some operations that preserve stability. A general theory of stability preserver operations is developed by Borcea and Br\"and\'en \cite{BoBr}. We also recommend the paper of Choe, Oxley, Sokal and Wagner \cite{COSW} for a comprehensive list of operations that preserves stability.
\begin{Th}
Let $P(x_1,\dots ,x_n)$ be a real stable polynomial. Suppose that the degree of $x_1$ in $P$ is $d$. Then the following hold true.\\
(a) The polynomial $x_1^dP(-1/x_1,x_2,\dots x_n)$ is real stable.\\
(b) If $a\in \mathbb{R}$, then $P(a,x_2,\dots ,x_n)$ is real stable or the constant $0$ polynomial.\\
(c) The polynomial $\frac{\partial}{\partial x_1}P$ is real stable or the constant $0$ polynomial.\\
\end{Th}
\begin{proof}
As before let $\mathbb{H}=\{z \ |\ \mathrm{Im}(z)>0\}$. Then the first claim is trivial since $z\mapsto \frac{-1}{z}$ maps $\mathbb{H}$ to $\mathbb{H}$.
\medskip
Next we prove part (b). By Lemma~\ref{stable-real-rooted} the polynomial $P(a+\varepsilon t,v_2+tu_2,\dots ,v_n+tu_n)$ is real-rooted for every $\varepsilon,u_2,\dots u_n\in \R_{>0}$ and $a,v_2,\dots ,v_n\in \R$. Let $\varepsilon \to 0$, then by Lemma~\ref{real-rooted-convergence} we get that $P(a,v_2+tu_2,\dots ,v_n+tu_n)$ is real-rooted or the constant $0$ function for every $u_2,\dots ,u_n\in \R_{>0}$ and $v_2,\dots ,v_n$. Now using the other direction of Lemma~\ref{stable-real-rooted} we get that $P(a,x_2,\dots ,x_n)$ is a real stable polynomial.
Next we prove part (c). Let
$$P(x_1,\dots ,x_n)=\sum_{k=0}^dP_k(x_2,\dots ,x_n)x_1^k.$$
First we show that $P_d(x_2,\dots ,x_n)$ is a real stable polynomial. By part (a)
$$R(x_1,\dots ,x_n):=x_1^dP(-1/x_1,x_2,\dots x_n)=\sum_{k=0}^dP_k(x_2,\dots ,x_n)(-1)^kx_1^{d-k}$$
is real stable. Then $R(0,x_2,\dots ,x_n)=(-1)^dP_d(x_2,\dots ,x_n)$ is real stable, and so $P_d(x_2,\dots ,x_n)$ is real stable.
Now let $Q=\frac{\partial}{\partial x_1}P$, and let $\underline{a}=(a_1,a_2,\dots, a_n)\in \mathbb{H}^n$. We show that if $Q\not\equiv 0$, then $\mathrm{Im}\left(\frac{Q(\underline{a})}{P(\underline{a})}\right)<0$. Note that here we use that $P$ is stable, so $P(\underline{a})\neq 0$, and we can divide with it. If $d=0$ then $Q\equiv 0$. We can assume that $d\geq 1$. Let
$$g(x):=P(x,a_2,\dots ,a_n)=\sum_{k=0}^dP_k(a_2,\dots ,a_n)x^k.$$
Note that $P_d(a_2,\dots ,a_n)\neq 0$ since $P_d$ is real-stable. So $g(x)$ has degree $d\geq 1$. Then
$g(x)=c\prod_{i=1}^d(x-\rho_i)$, and we have
$$\frac{g'(x)}{g(x)}=\sum_{i=1}^d\frac{1}{x-\rho_i}.$$
Note that $\mathrm{Im}(\rho_i)\leq 0$, otherwise $P(\rho_i,a_2,\dots ,a_n)=0$ would yield a zero in $\mathbb{H}^n$. Hence
$$\mathrm{Im}\left(\frac{Q(\underline{a})}{P(\underline{a})}\right)=\mathrm{Im} \left(\frac{g'(a_1)}{g(a_1)}\right)=\mathrm{Im} \left(
\sum_{i=1}^d\frac{1}{a_1-\rho_i}\right)<0.$$
In particular, this shows that $Q(\underline{a})\neq 0$. Hence $Q$ is stable. (Remark: we essentially repeated the proof of Gauss--Lucas theorem that asserts that the zeros of the derivative of a polynomial lie in the convex hull of the zeros of the polynomial.)
\medskip
\end{proof}
\subsection{Capacity of a polynomial} In this section we introduce the concept capacity.
\begin{Def}[Gurvits \cite{Gur1}]
Let $P(x_1,\dots ,x_n)$ be a multivariate polynomial with non-negative coefficients. Let $\underline{\alpha}=(\alpha_1,\dots ,\alpha_n)$ be a non-negative vector. Then the $\underline{\alpha}$-capacity of the polynomial $P(x_1,\dots ,x_n)$ is
$$\mathrm{cap}_{\underline{\alpha}}(P)=\inf_{x_1,\dots ,x_n>0}\frac{P(x_1,\dots ,x_n)}{\prod_{i=1}^nx_i^{\alpha_i}}.$$
\end{Def}
Note that if the numbers $\alpha_i$ are integers, then the capacity is an upper bound for the coefficient of the term $\prod_{i=1}^nx_i^{\alpha_i}$.
So in Theorem~\ref{general} we had $\mathrm{cap}_{\underline{r}}(P_G)$ in the statement.
In the definition of capacity we never used that $P$ is real stable, but it turns out that this concept is especially useful when we study stable polynomials. The main reason for this phenomenon is that one can often govern the capacity for stability preserver operators. A general theory of capacity preserver linear operators was developed in the paper of Leake and Gurvits \cite{LeGu}. The following theorem is a special case of their theory.
\begin{Th} \label{capacity preserving}
Let $P(x_1,\dots ,x_n)$ be a real stable polynomial with non-negative coefficients. Suppose that the degree of $x_1$ in $P$ is $d$. Let
$$Q=\frac{1}{r!}\bigg( \frac{\partial^r}{\partial x_1^r}P\bigg)\bigg|_{x_1=0}.$$
In other words, if we expand $P$ as a polynomial of $x_1$, then $Q$ is the coefficient of $x_1^r$. \\
Let $\underline{\alpha}=(\alpha_1,\dots ,\alpha_n)$, where $\alpha_1=r$, and $\underline{\alpha}'=(\alpha_2,\dots ,\alpha_n)$. Then
$$\mathrm{cap}_{\underline{\alpha}'}(Q)\geq \binom{d}{r}\bigg(\frac{r}{d}\bigg)^{r}\bigg(\frac{d-r}{d}\bigg)^{d-r}\mathrm{cap}_{\underline{\alpha}}(P).$$
\end{Th}
An immediate corollary of Theorem~\ref{capacity preserving} is the following theorem.
\begin{Th}[Coefficient lemma for stable polynomials with non-negative coefficients] \label{coefficient theorem}
Let $P(x_1,\dots ,x_n)$ be a real stable polynomial with non-negative coefficients. Suppose that the degree of $x_i$ in $P$ is at most $d_i$ for $i=1,\dots ,n$. Let $\underline{r}=(r_1,\dots ,r_n)$ and $a_{\underline{r}}$ be the coefficient of $\prod_{i=1}^nx_i^{r_i}$ in $P$. Then
$$a_{\underline{r}}\geq \prod_{i=1}^n\binom{d_i}{r_i}\bigg(\frac{r_i}{d_i}\bigg)^{r_i}\bigg(\frac{d_i-r_i}{d_i}\bigg)^{d_i-r_i}\mathrm{cap}_{\underline{r}}(P).$$
\end{Th}
If we apply Theorem~\ref{coefficient theorem} to the real stable polynomial $P_G(\underline{x})=\prod_{(u,v)\in E(G)}(x_u+x_v)$ we get Theorem~\ref{orientation}.
The key lemma to prove the above theorems is the following.
\begin{Lemma} \label{main lemma}
Let $p(z)=\sum_{k=0}^da_kz^k$ be a real-rooted polynomial with non-negative coefficients. Then
$$a_r\geq \binom{d}{r}\bigg(\frac{r}{d}\bigg)^r\bigg(\frac{d-r}{d}\bigg)^{d-r}\inf_{t>0}\frac{p(t)}{t^r}.$$
\end{Lemma}
We remark that this lemma for $r=1$ was the main ingredient of the proof of Gurvits for Theorem~\ref{matching}.
We will derive Lemma~\ref{main lemma} from the following statement.
\begin{Lemma} \label{main lemma2}
Let $p(z)=\sum_{k=0}^da_kz^k$ be a real-rooted polynomial with non-negative coefficients. Suppose that $p(1)=1$ and $p'(1)=r$ is an integer. Then
$$a_r\geq \binom{d}{r}\bigg(\frac{r}{d}\bigg)^r\bigg(\frac{d-r}{d}\bigg)^{d-r}.$$
\end{Lemma}
We remark that Lemma~\ref{main lemma2} is a special case of the following theorem of Hoeffding \cite{Hoe}. Nevertheless to keep our paper self-contained we will give a proof of Lemma~\ref{main lemma2}. The intuitive meaning of Hoeffding's theorem is that among probability distributions coming from real-rooted polynomials and fixed expected value the binomial distribution is the least concentrated around its expected value.
\begin{Th}[Hoeffding \cite{Hoe}] \label{Hoeffding} Let $p(z)=\sum_{k=0}^dp_kz^k$ be a real-rooted polynomial with $p_k\geq 0$, and $p(1)=1$, that is, $\sum_{k=0}^dp_k=1$. Let $s$ be defined by the equation $\sum_{k=0}^dkp_k=ds$.
Suppose that for non-negative integers $b$ and $c$ we have $b\leq ds\leq c$. Then
$$\sum_{k=b}^cp_k\geq \sum_{k=b}^c\binom{d}{k}s^k(1-s)^{d-k}.$$
\end{Th}
It is easy to see that Lemma~\ref{main lemma2} is a special case of Lemma~\ref{main lemma}, but as the following proof shows they are actually equivalent statements.
\begin{proof}[Proof of Lemma~\ref{main lemma} from Lemma~\ref{main lemma2}.]
Suppose that $p(z)=\sum_{k=m}^Ma_kz^k$, where $a_m,a_M>0$.
If $r<m$ or $r>M$, then $\inf_{t>0} \frac{p(t)}{t^r}=0$ so the claim is true in this case. If $r=m$, then $\inf_{t>0} \frac{p(t)}{t^r}=a_m$ so the claim is again true. If $r=M$, then $\inf_{t>0} \frac{p(t)}{t^r}=a_M$ so we are again done.
Thus we can assume that $m<r<M$.
Observe that $\frac{tp'(t)}{p(t)}$ is monotone increasing, $\frac{tp'(t)}{p(t)}\big|_{t=0}=m$ and $\lim_{t\to \infty} \frac{tp'(t)}{p(t)}=M$. So we can choose $t_r>0$ in such a way that $\frac{t_rp'(t_r)}{p(t_r)}=r$. Let us consider the probability distribution $q_j=\frac{a_jt_r^j}{p(t_r)}$. Then $\sum_k kq_k=r$, and $\sum_{j=0}^dq_jz^j$ is still a real-rooted polynomial.
Next let us apply Lemma~\ref{main lemma2}. Then
$$\frac{a_rt_r^r}{p(t_r)}=q_r\geq \binom{d}{r}\left(\frac{r}{d}\right)^r\left(\frac{d-r}{d}\right)^{d-r}.$$
In other words,
$$a_r\geq \binom{d}{r}\left(\frac{r}{d}\right)^r\left(\frac{d-r}{d}\right)^{d-r}\frac{p(t_r)}{t_r^r}\geq \binom{d}{r}\left(\frac{r}{d}\right)^r\left(\frac{d-r}{d}\right)^{d-r}\inf_{t>0}\frac{p(t)}{t^r}.$$
\end{proof}
Before we prove Lemma~\ref{main lemma2} we show that Lemma~\ref{main lemma} indeed implies Thorem~\ref{capacity preserving}.
\begin{proof}[Proof of Theorem~\ref{capacity preserving}]
For fixed $a_2,\dots ,a_n$ consider the polynomial
$$g(x)=P(x,a_2,\dots ,a_{n}).$$
Clearly, $\frac{1}{r!}\frac{d^r}{dx^r}g(x)\Big|_{x=0}=Q(a_2,\dots ,a_{n})$. Since $P$ is stable, and we substituted $a_i\in \mathbb{R}$ into it, $g(x)$ is stable. In other words, it is real-rooted. Thus we can use Theorem~\ref{main lemma}:
$$a_r \geq \binom{d}{r}\bigg(\frac{r}{d}\bigg)^r\bigg(\frac{d-r}{d}\Bigg)^{d-r}\inf_{x>0}\frac{g(x)}{x^r}.$$
Hence we have
\begin{align*}
\frac{Q(a_2,\dots ,a_{n})}{\prod_{i=2}^{n}a_i^{\alpha_i}}=\frac{\frac{1}{r!}\frac{d^r}{dx^r}g(x)\Big|_{x=0}}{\prod_{i=2}^{n}a_i^{\alpha_i}}&\geq \binom{d}{r}\bigg(\frac{r}{d}\bigg)^r\bigg(\frac{d-r}{d}\bigg)^{d-r}\frac{1}{\prod_{i=2}^{n}a_i^{\alpha_i}}\inf_{x>0}\frac{g(x)}{x^r}\\
&= \binom{d}{r}\bigg(\frac{r}{d}\bigg)^r\bigg(\frac{d-r}{d}\bigg)^{d-r}\inf_{x>0} \frac{P(x,a_2,\dots ,a_{n})}{x^r\cdot \prod_{i=2}^{n} a_i^{\alpha_i}}\\
&\geq \binom{d}{r}\bigg(\frac{r}{d}\bigg)^r\bigg(\frac{d-r}{d}\bigg)^{d-r}\mathrm{cap}_{\underline{\alpha}}(P).
\end{align*}
Taking infimum on the left side we get that
$\mathrm{cap}_{\underline{\alpha}'}(Q)\geq \binom{d}{r}\big(\frac{r}{d}\big)^r\big(\frac{d-r}{d}\big)^{d-r}\mathrm{cap}_{\underline{\alpha}}(P)$.
\end{proof}
\subsection{Proof of Lemma~\ref{main lemma2}}
In this section we prove Lemma~\ref{main lemma2}. The condition on the non-negativity of the coefficients and $p(1)=1$ implies that $p(z)$ can be written as follows:
$$p(z)=\prod_{i=1}^d(1-\alpha_i+\alpha_iz),$$
where $0\leq \alpha_i\leq 1$. Indeed, as all coefficients are non-negative, there can be no positive roots, thus using $p(1)=1$ the polynomial can be rewritten in the following way:
$$p(z)=a_n\prod_{i=1}^d(z+\rho_i)=\prod_{i=1}^d\left(\frac{z+\rho_i}{1+\rho_i}\right)=\prod_{i=1}^d(1-\alpha_i+\alpha_iz),$$
where $\alpha_i=\frac{1}{1+\rho_i}$.
Note that
$$r=p'(1)=\sum_{i=1}^d \alpha_i\prod_{j\not=i}(\alpha_j+1-\alpha_j)=\sum_{i=1}^d \alpha_i.$$
Consider the domain
$$D_{d,r}=\left\{(\alpha_1,\dots ,\alpha_d)\in \mathbb{R}^d\ \Big|\ 0\leq \alpha_i\leq 1\ (i=1,\dots ,d),\ \sum_{i=1}^d\alpha_i=r\right\}.$$
Clearly, the coefficient $a_r$ of $z^r$ in $p(z)$ can be expressed as
$$a_{r}=\sum_{\substack{K\subset[n]\\ |K|=r}}\prod_{j\in K}\alpha_j\prod_{j\not\in K}(1-\alpha_j).$$
So let us introduce the function
$$f_{d,r}(x_1,\ldots,x_d)=\displaystyle\sum_{\substack{K\subset[n]\\ |K|=r}}\prod_{j\in K}x_j\prod_{j\not\in K}(1-x_j).$$
Clearly, the statement of Lemma~\ref{main lemma2} is equivalent with
$$\min_{\underline{x}\in D_{d,r}}f_{d,r}(\underline{x})=f_{d,r}\left(\frac{r}{d},\dots ,\frac{r}{d}\right)=\binom{d}{r}\bigg(\frac{r}{d}\bigg)^r\bigg(\frac{d-r}{d}\bigg)^{d-r}.$$
We will prove this statement by induction on $d$. The case $d=1$ and $r=0$ or $r=1$ is trivial. In general, the case $r=0$ or $r=d$ is trivial since $D_{d,r}$ consists of only one point in this case. So we can always assume that $0<r<d$. First we prove that the statement is true for the boundary $\partial D_{d,r}$. Then we will prove that if $\underline{x}\neq \left(\frac{r}{d},\dots ,\frac{r}{d}\right)$, then we either have a point $\underline{y}\in \partial D_{d,r}$ for which $f_{d,r}(\underline{x})\geq f_{d,r}(\underline{y})$ or there exists an $\underline{x}'\in D_{d,r}$ such that $f_{d,r}(\underline{x})> f_{d,r}(\underline{x}')$.
The compactness of $D_{d,r}$ then implies that $\min_{\underline{x}\in D_{d,r}}f_{d,r}(\underline{x})=f_{d,r}\left(\frac{r}{d},\dots ,\frac{r}{d}\right)$.
So let us first prove that for $\underline{x}\in \partial D_{d,r}$ we have $f_{d,r}(\underline{x})\geq f_{d,r}\left(\frac{r}{d},\dots ,\frac{r}{d}\right)$.
Clearly, if $\underline{x}\in \partial D_{d,r}$ then one of its coordinates is $0$ or $1$. If we delete this coordinate, then the obtained vector $\underline{x}'$ is in $D_{d-1,r}$ in the first case, and in $D_{d-1,r-1}$ in the second case. Furthermore, $f_{d,r}(\underline{x})=f_{d-1,r}(\underline{x}')$ in the first case, and $f_{d,r}(\underline{x})=f_{d-1,r-1}(\underline{x}')$ in the second case. By induction we know that we have
$$f_{d-1,r}(\underline{x}')\geq \binom{d-1}{r}\bigg(\frac{r}{d-1}\bigg)^r\bigg(\frac{d-1-r}{d-1}\bigg)^{d-1-r}$$
for $\underline{x}'\in D_{d-1,r}$, and
$$f_{d-1,r-1}(\underline{x}')\geq \binom{d-1}{r-1}\bigg(\frac{r-1}{d-1}\bigg)^{r-1}\bigg(\frac{d-r}{d-1}\bigg)^{d-r}$$
for $\underline{x}'\in D_{d-1,r-1}$. Let us introduce the function $\ell_{n,k}=x^k(1-x)^{n-k}$. It is easy to see that it takes its maximum at the value $x=k/n$ in the interval $[0,1]$ as its derivative is $(k-nx)x^{k-1}(1-x)^{n-k-1}$. Hence
\begin{align*}
\binom{d}{r}\bigg(\frac{r}{d}\bigg)^r\bigg(\frac{d-r}{d}\bigg)^{d-r}&=\binom{d-1}{r}\bigg(\frac{r}{d}\bigg)^r\bigg(\frac{d-r}{d}\bigg)^{d-r-1}\\
&=\binom{d-1}{r}\ell_{d-1,r}\left(\frac{r}{d}\right)\\
&<\binom{d-1}{r}\ell_{d-1,r}\left(\frac{r}{d-1}\right)\\
&=\binom{d-1}{r}\bigg(\frac{r}{d-1}\bigg)^r\bigg(\frac{d-1-r}{d-1}\bigg)^{d-1-r}
\end{align*}
and
\begin{align*}
\binom{d}{r}\bigg(\frac{r}{d}\bigg)^r\bigg(\frac{d-r}{d}\bigg)^{d-r}&=\binom{d-1}{r-1}\bigg(\frac{r}{d}\bigg)^{r-1}\bigg(\frac{d-r}{d}\bigg)^{d-r}\\
&=\binom{d-1}{r-1}\ell_{d-1,r-1}\left(\frac{r}{d}\right)\\
&<\binom{d-1}{r-1}\ell_{d-1,r-1}\left(\frac{r-1}{d-1}\right)\\
&=\binom{d-1}{r-1}\bigg(\frac{r-1}{d-1}\bigg)^{r-1}\bigg(\frac{d-r}{d-1}\bigg)^{d-r}.
\end{align*}
Hence for $\underline{x}\in \partial D_{d,r}$ we have $f_{d,r}(\underline{x})\geq f_{d,r}\left(\frac{r}{d},\dots ,\frac{r}{d}\right)$.
Next we show that if $\underline{x}\neq \left(\frac{r}{d},\dots ,\frac{r}{d}\right)$, then either
we have a point $\underline{y}\in \partial D_{d,r}$ for which $f_{d,r}(\underline{x})\geq f_{d,r}(\underline{y})$ or
there exists an $\underline{x}'\in D_{d,r}$ such that $f_{d,r}(\underline{x})> f_{d,r}(\underline{x}')$. Since $\underline{x}\neq \left(\frac{r}{d},\dots ,\frac{r}{d}\right)$ there exists $i$ and $j$ such that $x_i\neq x_j$. Let $x_i+x_j=u$ and
$$\prod_{k\neq i,j}(1-x_k+x_kz)=\sum_{m=0}^{d-2}b_mz^m.$$
Then $f_{d,r}(\underline{x})$ is the coefficient of $z^r$ in the polynomial
$$\left(\sum_{m=0}^{d-2}b_mz^m\right)(1-x_i+zx_i)(1-(u-x_i)+z(u-x_i)).$$
A little computation shows that
$$f_{d,r}(\underline{x})=x_i(u-x_i)(b_{r-2}-2b_{r-1}+b_r)+(rb_{r-1}+(1-r)b_r).$$
If $b_{r-2}-2b_{r-1}+b_r<0$, then we get a strictly smaller value for $f_{d,r}(\underline{x})$ if we replace $(x_i,x_j)$ with $(u/2,u/2)$. If $b_{r-2}-2b_{r-1}+b_r\geq 0$, then we can replace $(x_i,x_j)$ with $(0,u)$ or $(1,u-1)$ depending on $u\in [0,1]$ or $[1,2]$ yielding a boundary point $\underline{y}$ for which $f_{d,r}(\underline{x})\geq f_{d,r}(\underline{y})$. This completes the proof.
\section{Capacity of the polynomial $P_G$}
In this section we prove Theorems~\ref{matching} and \ref{orientation} by computing the capacity of $P_G(\underline{x})=\prod_{(u,v)\in E}(x_u+x_v)$ with respect to various vectors $\underline{\alpha}$.
\begin{Lemma} \label{lemma-matching} Let $G=(A,B,E)$ be a $d$--regular bipartite graph on $2n$ vertices. Let $\underline{\alpha}$ be the vector that takes value $1$ at a vertex $u\in A$, and value $d-1$ at a vertex $v\in B$. Then
$$\capp_{\underline{\alpha}}(P_G) = \frac{d^{nd}}{(d-1)^{n(d-1)}}.$$
\end{Lemma}
\begin{proof} For sake of convenience let us denote the variables by $x_u$ if $u\in A$, and $y_v$ if $v\in B$. Then
$$\capp_{\underline{\alpha}} (P_G)=\inf_{x_u,y_v>0 \atop u\in A, v\in B} \frac{ \prod_{(u,v)\in E(G)}(x_u+y_v)}{\prod_{u\in A}x_u \cdot \prod_{v\in B}y_v^{d-1}}=\inf_{x_u,y_v>0 \atop u\in A, v\in B}\prod_{(u,v)\in E(G)}\frac{x_u+y_v}{x_u^{1/d}y_v^{(d-1)/d}}.$$
Note that
$$x_u+y_v=\frac{1}{d}(dx_u)+\frac{d-1}{d}\left(\frac{dy_v}{d-1}\right)\geq (dx_u)^{1/d}\left(\frac{dy_v}{d-1}\right)^{(d-1)/d}$$
by weighted arithmetic-geometric mean inequality. In other words,
$$\frac{x_u+y_v}{x_u^{1/d}y_v^{(d-1)/d}}\geq \frac{d}{(d-1)^{(d-1)/d}}.$$
Hence
$$\capp_{\underline{\alpha}} (P_G)\geq \left(\frac{d}{(d-1)^{(d-1)/d}}\right)^{nd}=\frac{d^{nd}}{(d-1)^{n(d-1)}}.$$
Observe that if $x_u=1$ and $y_v=d-1$ for all $u\in A$ and $v\in B$ then this bound is sharp.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{matching}]
As before let $\underline{\alpha}$ be the vector that takes value $1$ at a vertex $u\in A$, and values $d-1$ at a vertex $v\in B$.
By Theorem~\ref{general} we have
$$\prm(G)\geq \prod_{v\in A}\binom{d}{1}\bigg(\frac{1}{d}\bigg)^{1}\bigg(\frac{d-1}{d}\bigg)^{d-1}\cdot \prod_{v\in B}\binom{d}{d-1}\bigg(\frac{d-1}{d}\bigg)^{d-1}\bigg(\frac{1}{d}\bigg)^{1}\cdot \mathrm{cap}_{\underline{\alpha}}(P_G).$$
Then by Lemma~\ref{lemma-matching} we have
$$\prm(G)\geq \left(d\bigg(\frac{d-1}{d}\bigg)^{d-1}\right)^{2n}\frac{d^{nd}}{(d-1)^{n(d-1)}}=\bigg( \frac{(d-1)^{d-1}}{d^{d-2}} \bigg)^{n}.$$
\end{proof}
\begin{Lemma} \label{lemma-orientation} Let $\underline{\alpha}=(d_1/2,\dots ,d_n/2)$, where $d_j$ is the degree of the vertex $j$. Then
$\capp_{\underline{\alpha}} (P_G) = 2^{e(G)}$,
where $e(G)$ is the number of edges.
\end{Lemma}
\begin{proof}
First we will prove that $\capp_{\underline{\alpha}}(P_G) \leq 2^{e(G)}$.
Indeed, if we substitute $\underline{1}$ as $\underline{x}$, then we get
$$\frac{P_G(1,1, \dots 1)}{\prod_{v\in V(G)}1^{d_v/2}}=\prod_{(i,j)\in E(G)}(1+1)=2^{e(G)}.$$
Next we will prove the other direction: $\capp _{\underline{\alpha}}(P_G) \geq 2^{e(G)}$.
Using that $x_i+x_j\geq 2\sqrt{x_ix_j}$ we get that
$$\frac{\prod_{(u,v)\in E}(x_u+x_v)}{\prod_{v\in V(G)}x_v^{d_v/2}} \geq \frac{\prod_{(i,j)\in E(G)} 2\sqrt{x_ix_j}}{\prod_{v\in V(G)}x_v^{d_v/2}}=2^{e(G)}.$$
Thus $\capp_{\underline{\alpha}} (P_G) = 2^{e(G)}$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{orientation}]
As before let $\underline{\alpha}=(d_1/2,\dots ,d_n/2)$, where $d_j$ is the degree of the vertex $j$.
By Theorem~\ref{general} we have
$$\e(G)\geq \prod_{v\in V}\binom{d_v}{d_v/2}\bigg(\frac{d_v/2}{d_v}\bigg)^{d_v/2}\bigg(\frac{d_v/2}{d_v}\bigg)^{d_v/2}\cdot \mathrm{cap}_{\underline{\alpha}}(P_G).$$
Then by Lemma~\ref{lemma-orientation} we have
$$\e(G)\geq \prod_{v\in V}\frac{\binom{d_v}{d_v/2}}{2^{d_v}}\cdot 2^{e(G)}=\prod_{v\in V}\frac{\binom{d_v}{d_v/2}}{2^{d_v/2}}.$$
\end{proof}
\begin{Rem}
For $d$--regular graphs Las Vergnas \cite{Ver1} improved Theorem~\ref{orientation} as follows:
$$\e(G)\geq \frac{2^d}{\binom{d}{d/2}}\left(\frac{\binom{d}{d/2}}{2^{d/2}}\right)^n.$$
(At Proposition 5.3 of \cite{Ver1} there is a typo as it is pointed out in another paper of Las Vergnas in the footnote of the first page of \cite{Ver2}.) This strengthening can be obtained by our method too: all we have to note that we only need to apply Theorem~\ref{capacity preserving} to $n-1$ variables corresponding to vertices of $P_G(\underline{x})$: at the very end we should get a polynomial of the form $cx_n^{d_n/2}$ since throughout the process we get homogeneous polynomials. For this univariate polynomial there is no need to apply Theorem~\ref{capacity preserving} once more, so we get a $2^d/\binom{d}{d/2}$ improvement. This argument applies to non-regular graphs too thereby saving a factor $2^{\Delta}/\binom{\Delta}{\Delta/2}$, where $\Delta$ is the largest degree.
\end{Rem}
\section{Beyond this paper}
In this section we collected some pointers to the literature.
Stable polynomials have a huge literature.
If someone is interested in a comprehensive introduction to the this theory, then the paper of Choe, Oxley, Sokal and Wagner \cite{COSW} or Wagner's survey \cite{Wag} might be a good choice. Another excellent survey of the area is the paper of Vishnoi \cite{Vis}. For capacity preserver operations the paper of Gurvits and Leake \cite{LeGu} gives a treatment that is both very general and very readable. Statements on capacity often boils down to some statement about coefficients of univariate real-rooted polynomials often with a probabilistic flavour like Hoeffding's theorem \cite{Hoe}. A good source of such inequalities and results is Pitman's survey \cite{Pit}. The coefficient lemma, Theorem~\ref{coefficient theorem}, already appeared in the paper \cite{Gur3}. Interestingly this paper also considers various versions of the polynomial $P_G(\underline{x})=\prod_{(u,v)\in E(G)}(x_u+x_v)$, but never exactly this form. The coefficient lemma could have been easily derived from the work of Gurvits and Leake \cite{LeGu} too.
There are many different proofs and generalizations of Theorem~\ref{matching}. The original proof of Schrijver \cite{Sch2} is elementary, but involved. The first proof based on stable polynomials and capacity is due to Gurvits \cite{Gur1}, his proof is simplified in the paper of Laurent and Schrijver \cite{LaSc}. Another proof based on the theory of graph covers is given by Csikv\'ari \cite{Csik}. (The relationship between the theory of graph covers and the theory of stable polynomials is not yet well-understood.) Theorem~\ref{matching} has a very natural generalization for permanents of non-negative matrices. This generalization was derived by Gurvits \cite{Gur2} from the original paper of Schrijver \cite{Sch2}. Subsequently, Anari and Oveis-Gharan \cite{A-OG} and Straszak and Vishnoi \cite{StVi} gave a proof that only relies on the theory of stable polynomials. Another possible generalization considers counting matchings of fixed size in bipartite graphs instead of perfect matchings. This question was treated in the papers Csikv\'ari \cite{Csik}, Lelarge \cite{Lela} and Gurvits and Leake \cite{LeGu} (this last one uses only stable polynomials).
Concerning Eulerian orientations, Theorem~\ref{orientation} has many different proofs either. The original proof of Schrijver is very elegant and simple. Las Vergnas \cite{Ver1} gave another proof building on the theory of Martin's polynomial that gives a slightly stronger result. Borb\'enyi and Csikv\'ari \cite{BoCs} gave a proof using gauge transformation. The proof presented here is the first one using stable polynomials, but we remark that this result could have been easliy deduced from the paper of Straszak and Vishnoi \cite{StVi} too that uses stable polynomials.
Both Theorems~\ref{matching} and \ref{orientation} can be interpreted as a correlation inequality. If we divide by $2^{e(G)}=\prod_{v\in V(G)} 2^{d_v/2}$ in Theorem~\ref{orientation}, then the the right hand side of the inequality
$$\frac{\e(G)}{2^{e(G)}}\geq \prod_{v\in V(G)}\frac{\binom{d_v}{d_v/2}}{2^{d_v}}$$
is the probability that a random orientation is Eulerian, while on the left hand side the term $\frac{\binom{d_v}{d_v/2}}{2^{d_v}}$ is the probability that a random orientation is balanced at vertex $v$. So this inequality can be interpreted as a positive correlation inequality. Similarly, if $G=(A,B,E)$ is a $d$--regular bipartite graph, then we can consider the probability space where for each vertex $u$ in $A$ we pick exactly one of the edges incident to $u$ uniformly at random. This way we picked $n$ edges. For a vertex $v\in B$ let $E_v$ be the event that we picked exactly one of the edges incident to $v$. Then $\mathbb{P}(\cap_{v\in B}E_v)=\frac{\prm(G)}{d^n}$ while $\mathbb{P}(E_v)=d\cdot \frac{1}{d}\left(1-\frac{1}{d}\right)^{d-1}=\left(\frac{d-1}{d}\right)^{d-1}$. Hence Theorem~\ref{matching} is equivalent with $\mathbb{P}(\cap_{v\in B}E_v)\geq \prod_{v\in B}\mathbb{P}(E_v)$. It seems that many results on capacity are indeed fuelled by positive correlation inequalities, often in a very disguised way. This connection is the most explicit in the paper \cite{StVi} where the authors introduce the iterated positive correlation property and connect it with stable polynomials. Surprisingly it is also possible to build out a theory of negative correlation based on stable polynomials, for details see the paper \cite{Pem}.
One might wonder whether there is a deeper connection between Theorems~\ref{matching} and \ref{orientation}. It turns out that both theorems fall into a pattern that is about the so-called Bethe approximation. Hans Bethe was a Nobel-prize laurate physicist. Among many other achievements he introduced the concept that is now known as Bethe--approximation. Originally this was a highly heuristic concept that approximates well quantities coming from counting objects with local constraints like perfect matchings and Eulerian orientations. For a long time it was overlooked by the mathematics community, then Bethe approximation showed up in two different lines of research. In the work of Dembo, Montanari and their coauthors \cite{DeMa,DeMa2,DMS,DMSS} about graph limit theory of sparse graphs, an appropriate version of Bethe approximation played the role of the limit value of certain graph parameters. Another line of research emerged from the work of Gurvits establishing inequalities between a graph parameter and its Bethe approximation. In many cases Bethe approxiation turns out to be a lower bound for the corresponding graph parameter. Theorem~\ref{matching} and \ref{orientation} belong to this line of research. This area almost exclusively relies on stable polynomials \cite{A-OG,StVi} and graph covers \cite{Csik,Ruo,Lela,Von}.
\bigskip
\noindent \textbf{Acknowledgment.} The first author thanks Jonathan Leake for the discussions on the topic of this paper.
The authors are very grateful to the anonymous referee for his/her suggestions that greatly improved the paper.
|
\section{Introduction}
Hundreds of ultra-diffuse galaxies (UDGs) have been recently discovered in the Coma cluster. They are typically spheroidal in shape with central surface brightness $\mu_{g,0} = 24-26$~mag~arcsec$^{-2}$ and have large effective radii (of a few kpc) for their stellar masses \citep{vdokkum2015, koda2015}. Most UDGs adhere to the low mass end of the red sequence, indicating that they are passively evolving cluster members \citep{yagi2016, zaritsky2019}. UDGs are remarkable for their potentially high dark matter fractions ($\gtrsim98\%$), whose large gravitational potentials may protect the diffuse stellar component from tidal disruptions by the cluster environment \citep{vdokkum2016, koda2015}.
Although UDGs were first reported by \citet{binggeli1985} in the Virgo Cluster, their low surface brightness and low concentration makes them difficult to detect, and it is only recently that large populations have been discovered. These larger studies began with the Dragonfly (DF) Telephoto Array by \cite{vdokkum2015}, which found 47 UDGs in the Coma Cluster. This DF database constitutes the sample that is investigated in this paper, but subsequently more UDGs have been discovered in Coma. Prompted by the DF study, \cite{koda2015} found 854 UDGs in the $\sim4.1^{\circ}\times4.1^{\circ}$ field of view (FOV) of archival Subaru Prime Focus Camera data. Since the DF FOV is about twice the size, one expects about 1,000 UDGs in its field, provided equally sensitive data. \cite{yagi2016} extended the analysis with the Subaru data to specify the UDG parameters. \cite{zaritsky2019} further expanded the search area using the DECam Legacy Survey fields in a 10$^{\circ}$ radius around Coma, finding about 300 UDGs when using a more conservative threshold for the effective radius, $r_{\text{eff}} \gtrsim 2.5$~kpc. This sample constitutes the SMUDGes survey (Systematically Measuring Ultra-Diffuse Galaxies). Beyond Coma, UDGs have been identified in other galaxy clusters and in the field.
The Coma UDGs are indeed different from other galaxies. \cite{vdokkum2017} found that Coma UDGs have $\sim$7 times more globular clusters (GCs) than galaxies of same luminosity, while \cite{danieli2018} found that the S\'ersic index of faint large galaxies decreases with magnitude (opposite to that of bright, large galaxies). Meanwhile, the stellar populations appear to be old and metal poor \citep{gu2018}, with little to no star formation activity \citep{singh2019}.
One Coma UDG, DF44, has received particular attention because its large number of GCs and stellar velocity dispersion imply a Milky Way-like dynamical mass of $10^{12} M_{\odot}$, whereas its stellar mass of $3\times 10^8 M_{\odot}$ is consistent with a dwarf galaxy \citep{vdokkum2015,vdokkum2016,vdokkum2017,gu2018}. This implies an extreme dark matter fraction exceeding 99\%, which challenges traditional galaxy models, especially as DF44 may be representative of larger UDGs.
The origin of UDGs remains unclear. \cite{jiang2019} conducted smoothed-particle hydrodynamic simulations of both field and cluster UDGs, concluding that cluster UDGs become quiescent through ram-pressure stripping of cold gas by the intracluster medium and become diffuse through tidal distortions near pericenter. They find that the star-formation rate (SFR) and $B-R$ color gradient depend on the distance to the cluster center. In contrast, they find that field UDGs may occur through gas depletion by supernova feedback, which predicts that only galaxies of a certain mass range become UDGs in the field. Meanwhile, \cite{liao2019} used the Auriga cosmological magneto-hydrodynamical simulations to investigate the formation of UDGs of Milky Way-sized galaxies. They studied the morphology and visual properties of the UDGs, such as sizes, central surface brightnesses, S\'ersic indicies, colors, spatial distribution, and abundance. Their simulations suggest that field UDGs form in ``high-spin'' dark matter halos and that supernova feedback is less important, but that half of UDGs become diffuse through tidal interactions.
Ultraviolet (UV) and X-ray observations can provide crucial clues to UDG formation. From the perspective of galaxy assembly, the UV band is sensitive to even low levels of star formation, while X-rays trace X-ray binaries (which scale with the stellar population) and can show whether UDGs exist in hot halos of gas, as expected for relatively massive galaxies \citep{white&frenk1991}. X-rays can also identify even weakly accreting supermassive black holes, which have yet to be discovered in UDGs and whose absence in Milky Way-sized galaxies would be surprising. \cite{singh2019} recently studied the UV properties of SMUDGes UDGs using GALEX data, finding few detections and concluding that there is no star formation. Meanwhile, \cite{kovacs2019} reported X-ray non-detections for isolated UDGs using \textit{XMM-Newton} survey fields that cover Subaru survey fields \citep{greco2018}.
Here we examine the UV and X-ray properties of the Coma DF survey galaxies \citep{vdokkum2015} with GALEX and the \textit{Chandra} X-ray Observatory. We restrict ourselves to this sample because the archival UV and X-ray coverage and sensitivity is highest within $\sim$1~Mpc of the cluster core. We also report on new observations of DF44 with the \textit{Neil Gehrels Swift Observatory} (\textit{Swift}) and \textit{XMM-Newton} (XMM). These deep data enable us to sensitively search for ongoing or recent star formation, X-ray binaries, hot gas, and an AGN. DF44 is presently the best Coma UDG target for X-ray observations because of its potentially large dynamical mass (predicting a hot halo) and its large projected distance from the Coma cluster center (1.8~Mpc from the Coma core in projection and separated by 900~km~s$^{-1}$ from the systemic velocity), which reduces the background from the hot intracluster medium and makes it more likely that the galaxy can hold onto its hot gas.
In the remainder of this paper, we describe data we used, then present the archival UV and X-ray data analysis. This is followed by the study of DF44, and we close by discussing our results and summarizing our conclusions.
\begin{table*}
\centering
\caption{Galaxies and archival UV and X-ray data used in this paper. \label{tab:datasources}}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}} lcccclrrlcr}
\hline
{Name} & R.A. & Dec. & $r_{\text{eff}}$ & $R_{\text{Coma}}$ & GALEX tile & $t_{\text{FUV}}$ & $t_{\text{NUV}}$ & \textit{Chandra} & Chip & $t_{\text{X-ray}}$ \\
& & & & & & & & obsID & & \\
& [J2000] & [J2000] & [kpc] & [Mpc] & & [ks] & [ks] & & [ks] \\
\hline
DF1 & 12:59:14.1 & 29:07:16 & 3.1 & 2.0 & GI1\_039006\_Coma\_MOS06 & 1.7 & 3.5 & $\cdots$ & $\cdots$ & $\cdots$\\
DF2 & 12:59:09.5 & 29:00:25 & 2.1 & 1.8 & GI1\_039006\_Coma\_MOS06 & 1.7 & 3.5 & $\cdots$ & $\cdots$ & $\cdots$\\
DF3 & 13:02:16.5 & 28:57:17 & 2.9 & 1.9 & GI1\_039006\_Coma\_MOS06 & 1.7 & 2.6 & $\cdots$ & $\cdots$ & $\cdots$\\
DF4 & 13:02:33.4 & 28:34:51 & 3.9 & 1.5 & GI1\_039008\_Coma\_MOS08 & 1.7 & 2.6 & $\cdots$ & $\cdots$ & $\cdots$\\
DF5 & 12:55:10.5 & 28:33:32 & 1.8 & 2.0 & GI1\_039003\_Coma\_MOS03 & 1.7 & 2.5 & $\cdots$ & $\cdots$ & $\cdots$\\
DF6 & 12:56:29.7 & 28:26:40 & 4.4 & 1.5 & GI1\_039003\_Coma\_MOS03 & 1.7 & 2.5 & $\cdots$ & $\cdots$ & $\cdots$\\
DF7 & 12:57:01.7 & 28:23:25 & 4.3 & 1.3 & GI1\_039003\_Coma\_MOS03 & 1.7 & 2.5 & $\cdots$ & $\cdots$ & $\cdots$\\
DF8 & 13:01:30.4 & 28:22:28 & 4.4 & 0.9 & GI5\_025001\_COMA & 18.6 & 25.7 & $\cdots$ & $\cdots$ & $\cdots$\\
DF9 & 12:56:22.8 & 28:19:53 & 2.8 & 1.5 & GI1\_039003\_Coma\_MOS03 & 1.7 & 2.5 & $\cdots$ & $\cdots$ & $\cdots$\\
DF10 & 12:59:16.3 & 28:17:51 & 2.4 & 0.6 & GI5\_025001\_COMA & 18.6 & 25.7 & $\cdots$ & $\cdots$ & $\cdots$\\
DF11 & 13:02:25.5 & 28:13:58 & 2.1 & 1.1 & GI5\_025001\_COMA & 18.6 & 25.7 & $\cdots$ & $\cdots$ & $\cdots$\\
DF12 & 13:00:09.1 & 28:08:27 & 2.6 & 0.3 & GI5\_025001\_COMA & 18.6 & 25.7 & 18235 & ACIS-S & 30 \\
DF13 & 13:01:56.2 & 28:07:23 & 2.2 & 0.9 & GI5\_025001\_COMA & 18.6 & 25.7 & $\cdots$ & $\cdots$ & $\cdots$\\
DF14 & 12:58:07.8 & 27:54:46 & 3.8 & 0.7 & GI5\_025001\_COMA & 18.6 & 25.7 & 18236, 19909, & ACIS-I & 30 \\
& & & & & & & & 19910 & & \\
DF15 & 12:58:16.3 & 27:53:29 & 4.0 & 0.6 & GI5\_025001\_COMA & 18.6 & 25.7 & 18236, 19909, & ACIS-I & 30 \\
& & & & & & & & 19910 & & \\
DF16 & 12:56:52.4 & 27:52:29 & 1.5 & 1.1 & COMA\_SPEC\_A & 1.3 & 2.8 & $\cdots$ & $\cdots$ & $\cdots$\\
DF17 & 13:01:58.3 & 27:50:11 & 4.4 & 0.9 & GI5\_025001\_COMA & 18.6 & 25.7 & 10921 & ACIS-S & 5 \\
DF18 & 12:59:09.3 & 27:49:48 & 2.8 & 0.4 & GI5\_025001\_COMA & 18.6 & 25.7 & 13994, 14411 & ACIS-I & 110 \\
DF19 & 13:04:05.1 & 27:48:05 & 4.4 & 1.7 & GI1\_039009\_Coma\_MOS09 & 1.7 & 1.7 & $\cdots$ & $\cdots$ & $\cdots$\\
DF20 & 13:00:18.9 & 27:48:06 & 2.3 & 0.4 & GI5\_025001\_COMA & 18.6 & 25.7 & 2941, 13993, & ACIS-I & 303 \\
& & & & & & & & 14410, 13995, & & \\
& & & & & & & & 14406, 14415 & & \\
DF21 & 13:02:04.1 & 27:47:55 & 1.5 & 0.9 & GI5\_025001\_COMA & 18.6 & 25.7 & 10921 & ACIS-S & 5 \\
DF22 & 13:02:57.8 & 27:47:25 & 2.1 & 1.3 & GI1\_039009\_Coma\_MOS09 & 1.7 & 1.7 & $\cdots$ & $\cdots$ & $\cdots$\\
DF23 & 12:59:23.8 & 27:47:27 & 2.3 & 0.4 & GI5\_025001\_COMA & 18.6 & 25.7 & 13993, 14410, & ACIS-I & 233 \\
& & & & & & & & 13994, 14411 & & \\
DF24 & 12:56:28.9 & 27:46:19 & 1.8 & 1.3 & COMA\_SPEC\_A & 1.3 & 2.8 & $\cdots$ & $\cdots$ & $\cdots$\\
DF25 & 12:59:48.7 & 27:46:39 & 4.4 & 0.3 & GI5\_025001\_COMA & 18.6 & 25.7 & 13993, 14410, & ACIS-I & 479 \\
& & & & & & & & 13994, 14411, & & \\
& & & & & & & & 13995, 14406, & & \\
& & & & & & & & 14415, 13996 & & \\
DF26 & 13:00:20.6 & 27:47:13 & 3.3 & 0.4 & GI5\_025001\_COMA & 18.6 & 25.7 & $\cdots$ & $\cdots$ & $\cdots$\\
DF27 & 12:58:57.3 & 27:44:39 & $\cdots$ & 0.5 & GI5\_025001\_COMA & 18.6 & 25.7 & 18237, 19911, & ACIS-I & 30 \\
& & & & & & & & 19912 & & \\
DF28 & 12:59:30.4 & 27:44:50 & 2.7 & 0.4 & GI5\_025001\_COMA & 18.6 & 25.7 & 18237, 19911, & ACIS-S & 30 \\
& & & & & & & & 19912 & & \\
DF29 & 12:58:05.0 & 27:43:59 & 3.1 & 0.8 & GI5\_025001\_COMA & 18.6 & 25.7 & $\cdots$ & $\cdots$ & $\cdots$\\
DF30 & 12:53:15.1 & 27:41:15 & 3.2 & 2.6 & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$\\
DF31 & 12:55:06.2 & 27:37:27 & 2.5 & 1.9 & COMA\_SPEC\_A & 1.3 & 2.8 & $\cdots$ & $\cdots$ & $\cdots$\\
DF32 & 12:56:28.4 & 27:37:06 & 2.8 & 1.4 & GI2\_046001\_COMA3 & 29.9 & 31.1 & $\cdots$ & $\cdots$ & $\cdots$\\
DF33 & 12:55:30.1 & 27:34:50 & 1.9 & 1.8 & GI2\_046001\_COMA3 & 29.9 & 31.1 & $\cdots$ & $\cdots$ & $\cdots$\\
DF34 & 12:56:12.9 & 27:32:52 & 3.4 & 1.6 & GI2\_046001\_COMA3 & 29.9 & 31.1 & $\cdots$ & $\cdots$ & $\cdots$\\
DF35 & 13:00:35.7 & 27:29:51 & 2.7 & 0.9 & GI5\_025001\_COMA & 18.6 & 25.7 & $\cdots$ & $\cdots$ & $\cdots$\\
DF36 & 12:55:55.4 & 27:27:36 & 2.6 & 1.8 & GI2\_046001\_COMA3 & 29.9 & 31.1 & $\cdots$ & $\cdots$ & $\cdots$\\
DF37 & 12:59:23.6 & 27:21:22 & 1.5 & 1.1 & GI2\_046001\_COMA3 & 29.9 & 31.1 & $\cdots$ & $\cdots$ & $\cdots$\\
DF38 & 13:02:00.1 & 27:19:51 & 1.8 & 1.4 & GI1\_039009\_Coma\_MOS09 & 1.7 & 1.7 & $\cdots$ & $\cdots$ & $\cdots$\\
DF39 & 12:58:10.4 & 27:19:11 & 4.0 & 1.3 & GI2\_046001\_COMA3 & 29.9 & 31.1 & 4724 & ACIS-I & 60 \\
DF40 & 12:58:01.1 & 27:11:26 & 2.9 & 1.5 & GI2\_046001\_COMA3 & 29.9 & 31.1 & $\cdots$ & $\cdots$ & $\cdots$\\
DF41 & 12:57:19.0 & 27:05:56 & 3.4 & 1.8 & GI2\_046001\_COMA3 & 29.9 & 31.1 & $\cdots$ & $\cdots$ & $\cdots$\\
DF42 & 13:01:19.1 & 27:03:15 & 2.9 & 1.7 & GI5\_063019\_A1656\_FIELD2 & 1.5 & 1.6 & $\cdots$ & $\cdots$ & $\cdots$\\
DF43 & 12:54:51.4 & 26:59:46 & 1.5 & 2.6 & GI2\_046001\_COMA3 & 29.9 & 31.1 & $\cdots$ & $\cdots$ & $\cdots$\\
DF44 & 13:00:58.0 & 26:58:35 & 4.6 & 1.8 & GI5\_063019\_A1656\_FIELD2 & 1.5 & 4.2 & $\cdots$ & $\cdots$ & $\cdots$\\
DF45 & 12:53:53.7 & 26:56:48 & 1.9 & 2.9 & GA\_DDO154 & 4.2 & 4.2 & $\cdots$ & $\cdots$ & $\cdots$\\
DF46 & 13:00:47.3 & 26:46:59 & 3.4 & 2.1 & GI5\_063019\_A1656\_FIELD2 & 1.5 & 2.6 & $\cdots$ & $\cdots$ & $\cdots$\\
DF47 & 12:55:48.1 & 26:33:53 & 4.2 & 2.9 & GI1\_039004\_Coma\_MOS04 & 1.7 & 2.5 & $\cdots$ & $\cdots$ & $\cdots$\\
\hline
\multicolumn{11}{p{0.9\textwidth}}{Columns. (1) Dragonfly ID from \cite{vdokkum2015}, (2) Right Ascension, (3) Declination, (4) Effective radius, (5) projected distance from Coma center, (6) GALEX observation used, (7-8) NUV and FUV exposure times, (9) \textit{Chandra} observations used, (10) \textit{Chandra} detector used, (11) \textit{Chandra} exposure time.}
\end{tabular*}
\end{table*}
\section{Observations and Data Processing}
The Coma DF galaxies, archival UV observations, and archival X-ray observations used in this paper are given in Table~\ref{tab:datasources}. The measurements and upper limits from these data are given in Table~\ref{tab:udgoptuvxray}.
\subsection{GALEX}
Most of the DF survey footprint is covered by deep GALEX NUV and FUV exposures, and they tend to be the brighter and less confused sources within the Coma cluster, which motivates our use of the DF sample. Nevertheless, none of the DF galaxies are in the GALEX source catalog, so we performed aperture photometry with Photutils v0.5 \citep{bradley2017} on calibrated images retrieved from MAST\footnote{The Mikulski Archive for Space Telescopes; https://archive.stsci.edu}. For each individual galaxy, we centered an aperture {at the DF catalog coordinates with a radius equal to the optical half-light radius and apply the aperture correction.} We subtracted the local background after masking point sources (detected at $2.5\sigma$) and sigma-clipping the background (removing deviations at $\sigma > 2.5$). We visually inspected each source to verify this procedure. {Out of the 47 DF targets, 7 were rejected in the FUV channel and 8 were rejected in the NUV channel due to additional sources in the aperture.} We used the same images for stacking by clipping the image to a region several times the effective radius in size and subtracted the estimated background.
At a 3$\sigma$ threshold, we detected {3/40 DF UDGs in the FUV channel and 6/39 in the NUV} (Table~\ref{tab:udgoptuvxray}). This is {inconsistent with} \citet{singh2019}, who did not detect any DF UDGs in Coma and only 2/110 of the SMUDGes UDGs within the Coma virial radius were detected. They also used Photutils on the same data, with the main difference being that they measured the sky background within the same aperture as the UDG from the GALEX-provided sky background image. They verified that this heavily smoothed background map was not contaminated by the UDG itself. However, their method of identifying detections is different: whereas we use the 3$\sigma$ threshold based on the sigma-clipped background, they classify sources with $>2\sigma$ as candidate detections, then visually inspect the data to determine whether the candidate is consistent with the UDG position or affected by contamination.
Most of our UV-detected sources are marginal detections ($3< \sigma < 5$), so on a case-by-case basis it is often unclear whether the source is mis-centered, contaminated, etc. However, the stacked image of all the sources (Figure~\ref{fig:fullstackims}) shows a very strong signal from the NUV channel, which gives us confidence that at least some Coma DF UDGs do have UV emission; DF44 is separately described below.
\subsection{SDSS}
The UV-optical color is of interest to determine the stellar populations of UDGs, so we used \textit{ugriz} data from the Sloan Digital Sky Survey (SDSS) Data Release 12 \citep{alam2015}. Following the same procedure as with GALEX, we find the following detection fractions: {\textit{u} = 7/40, \textit{g} = 25/41, \textit{r} = 32/41, \textit{i} = 24/40, \textit{z} = 8/41.} For reference, \cite{york2000} provides 5$\sigma$ point-source detection limits for 1" seeing at an airmass of 1.4 of 22.3, 24.4, 24.1, 22.3, and 20.8 mags in the \textit{u, g, r, i} and \textit{z} filters, respectively, but with somewhat lower sensitivities for extended sources. The magnitudes and limits are reported in Table~\ref{tab:udgoptuvxray}.
\subsection{\textit{Chandra}}
Most of the DF galaxies are in regions where intracluster X-ray emission from Coma is a major or dominant source of ``background'' for potential UDG emission from X-ray binaries or AGNs. This background is irreducible, so high angular resolution provides the best opportunity to detect UDGs in the X-rays and we used archival \textit{Chandra} data to search for UDG emission. 12 UDGs lie in \textit{Chandra} exposures. We obtained the archival data and re-processed it with the \textit{Chandra} Interactive Analysis of Observations (CIAO) v.4.8.2 software, following the standard pipeline.
We filtered event files to 2-7~keV to optimize our sensitivity to nonthermal emission and merged overlapping files when the detector was the same and the observational parameters were similar, then performed aperture photometry using CIAO tools and local (annular) background. Unlike with the optical data, we used a single aperture size with $r=15^{\prime\prime}$ because of the stochastic distribution of bright X-ray binaries. No sources were detected, with 0.3-10~keV 3$\sigma$ upper limits of 1$-$7~$\times~10^{40}$~erg~s$^{-1}$ (Table~\ref{tab:udgoptuvxray}), assuming a power law with $\Gamma=1.8$ and a Galactic absorbing column from the Leiden-Argentine-Bonn \citep{kalberla05} survey\footnote{\url{https://heasarc.gsfc.nasa.gov/cgi-bin/Tools/w3nh/w3nh.pl}} of $7.98\times10^{19}$~cm$^{-2}$.
\begin{table*}
\centering
\caption{Dragonfly Coma X-ray Measurements. \label{tab:xray}}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}} lccccccccc}
\hline
{Name} & $R_{\text{Coma}}$ & {Instrument} & {$t_{\text{exp}}$} & {Counts} & {$N_{\text{pix}}$} & $\bar{B}$ & Net Rate & 3$\sigma$ Sens. & $L_X$\\
& [Mpc] & & [ks] & & & [Counts/pix] & [$10^{-4}$ Counts/s] & [$10^{-4}$ Counts/s] & [$10^{40}$ erg~s$^{-1}$] \\
\hline
DF12 & 0.3 & ACIS-S & 29.7 & 50 & 2922 & 0.014 & 2.8 & 6.5 & $<$3.2 \\
DF14 & 0.7 & ACIS-I & 27.4 & 31 & 2922 & 0.010 & 0.4 & 6.0 & $<$3.4 \\
DF15 & 0.6 & ACIS-I & 27.4 & 25 & 2922 & 0.010 & -1.7 & 6.0 & $<$3.4 \\
DF17 & 0.9 & ACIS-S & 5.0 & 7 & 2922 & 0.002 & 4.1 & 13.4 & $<$6.6 \\
DF18 & 0.4 & ACIS-I & 187.1 & 493 & 2922 & 0.161 & 1.2 & 3.5 & $<$2.0 \\
DF20 & 0.4 & ACIS-I & 158.8 & 310 & 2922 & 0.094 & 2.2 & 3.1 & $<$1.5 \\
DF21 & 0.9 & ACIS-S & 5.0 & 6 & 2922 & 0.002 & 2.1 & 13.4 & $<$7.6 \\
DF23 & 0.4 & ACIS-I & 43.5 & 155 & 2922 & 0.053 & 0.03 & 8.6 & $<$4.2 \\
DF25 & 0.3 & ACIS-I & 235.2 & 583 & 2922 & 0.203 & -0.4 & 3.1 & $<$1.8 \\
DF27 & 0.5 & ACIS-S & 265.9 & 531 & 2922 & 0.166 & 1.7 & 2.5 & $<$1.4 \\
DF28 & 0.4 & ACIS-S & 29.7 & 47 & 2922 & 0.016 & -0.2 & 7.0 & $<$4.0 \\
DF39 & 1.3 & ACIS-S & 59.7 & 28 & 2922 & 0.008 & 0.8 & 2.4 & $<$1.2 \\
DF44 & 1.8 & EPIC & 166 & 143 & 143 & 1.23 & 0.02 & 0.5 & $<$0.01 \\
\hline
\multicolumn{10}{p{0.9\textwidth}}{Columns. (1) Dragonfly ID from \cite{vdokkum2015}, (2) projected distance from Coma cluster center, (3) Instrument (only DF44 is \textit{XMM-Newton}), (4) Vignetting-corrected exposure time at DF position, (5) Counts in aperture (2-7~keV for \textit{Chandra} observations, 0.4-2~keV for DF44), (6) Pixels in aperture, (7) Average background measured from annulus, (8) Net count rate, (9) 3$\sigma$ sensitivity count rate, (10) Limiting 0.3-10~keV $L_X$ (no galaxy was detected), based on a power law and the measurement bandpass.}\\
\multicolumn{10}{p{0.9\textwidth}}{Notes. Source fluxes were measured in circular apertures and backgrounds in surrounding annuli (see text).}
\end{tabular*}
\end{table*}
\subsection{DF44 \textit{Swift} Data}
We obtained about 90~ks of new \textit{Swift} UV and Optical Telescope (UVOT) and X-ray Telescope (XRT) data for DF44. The UVOT data are in the UVW2 (1928\AA) and UVW1 (2600\AA) bands (40~ks and 20~ks of good time, respectively), while the XRT data covers 0.3-10~keV. The UVOT data are accumulated as snapshots, so we combined them into a merged exposure after reduction using the standard pipeline, except to remove a diffuse artifact (described below). We also specified parameters in the pipeline to standardize the pixel scale, file format, and astrometric solution (known to $<0.1^{\prime\prime}$). Following the pipeline reduction, but before merging, we applied corrections based on the large-scale sensitivity maps (produced through the pipeline but not applied).
Scattered light from the filter window creates persistent patterns on UVOT images. Since the UVOT operates by taking multiple snapshots for each observation and the pattern is static in detector coordinates \citep{breeveld11}, stacking images to improve sensitivity also stacks these artifacts with various offsets, which reduces the final sensitivity of the combined image. To mitigate this, we used templates for the scattered light in the UVW1 and UVW2 filters created by \citet{hk14} to subtract the artifact from each exposure. We masked the sources in each exposure down to a threshold of $3\sigma$ above the background, then used least-squares fitting to determine the appropriate scale factor for the template, which was then subtracted from the image. This technique can remove up to 95\% of the artifact, but can fail in crowded fields or with very short exposures, so the scale factors were verified by visual inspection (e.g., over-subtraction leads to a clear negative imprint of the features on the image). In cases where no solution was found, the image was not added to the stack {(about 33\% of the total UVW1 exposure time and 32\% of the total UVW2 exposure time)}. For comparison, we also stacked the images with no correction.
We then used aperture photometry to measure the UVW1 and UVW2 magnitudes for DF44 using the UVOT software. The UVW1 flux densities measured when correcting for the diffuse artifact and combining images without this correction are both $F_{\lambda} = 9\pm2\times 10^{-18}$~erg~s$^{-1}$~cm$^{-2}$~\AA$^{-1}$. The UVW2 corrected and uncorrected fluxes are $F_{\lambda} = 6\pm3\times 10^{-18}$~erg~s$^{-1}$~cm$^{-2}$~\AA$^{-1}$ and $3\pm2\times 10^{-18}$~erg~s$^{-1}$~cm$^{-2}$~\AA$^{-1}$. These values are consistent with each other.
Finally, the fluxes must be corrected for a red transmission leak. In the UVW1 filter there is a small but non-negligible response between 3000-5500\AA, so that a significant fraction of the measured UVW1 flux can be optical light in red sources. DF44 appears to have optical flux densities several times higher than the UV flux densities, so this cannot be ignored.
We used the SDSS $u$ (2980-4130\AA) and $g$ (3630-5680\AA) filters along with the UVW1 transmission curve to estimate the optical contamination, based on the following expression (using fluxes rather than flux densities):
\begin{equation*}
F_{\textrm{leak,SDSS}} = \bigg(\frac{\int_{\lambda_{\textrm{min,SDSS}}}^{\lambda_{\textrm{max,SDSS}}}T_{UVW1}}{\int_{\lambda_{\textrm{min},UVW1}}^{\lambda_{\textrm{max},UVW1}}T_{UVW1}}\bigg) \times F_{\textrm{SDSS}}[\textrm{erg}~\textrm{s}^{-1} \textrm{cm}^{-2}].
\end{equation*}
where $F_{\textrm{leak,SDSS}}$ is the flux measured in the UVW1 band from only that flux seen in the \textit{u} or \textit{g} SDSS filter, and can be subtracted from the measured UVW1 flux to obtain the true UV flux, $F_{\textrm{corr},UVW1} = F_{UVW1} - F_{\textrm{leak},u} - F_{\textrm{leak},g}$. We then measured the effective wavelength of the ``corrected'' filter as 2505\AA. We estimate that the optical contribution to the DF44 UVW1 flux is 14\%, so its true UV flux near 2500\AA\ is $7.2\times10^{-18}$ erg~s$^{-1}$cm$^{-2}$\AA$^{-1}$; 14\% is comparable to the flux uncertainty. After subtraction, the UVW1 flux density is closer to that in the NUV, which is expected based on the overlap of the transmission curves. The effective wavelengths, flux densities, uncertainties, exposure times, and conversion factors for all relevant filters used for DF44 (including GALEX and SDSS) are reported in Table \ref{tab:df44phot}.
The XRT data were obtained in photon-counting mode and were reduced using the HEASoft XRT analysis tools\footnote{\url{https://swift.gsfc.nasa.gov/analysis/}}. The total good time was 88~ks, and we combined the exposures in \textit{Xselect}\footnote{\url{https://heasarc.nasa.gov/ftools/xselect/}} and created a corrected exposure map. For analysis we used the stacked images in both the 0.3-2~keV and 0.3-10~keV bandpasses. No source is detected at the position of DF44, with a 0.3-10~keV upper limit of $L_X < 4.1\times 10^{39}$~erg~s$^{-1}$, assuming a power-law with $\Gamma=1.8$ and Galactic absorption.
\begin{table*}
\centering
\caption{Dragonfly Coma UDG UV, Optical, and X-ray measurements. \label{tab:udgoptuvxray}}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}} lccccccccc}
\hline
{Name} & $R_{\text{Coma}}$ & $m_{\text{FUV}}$ & $m_{\text{NUV}}$ & $m_u$ & $m_g$ & $m_r$ & $m_i$ & $m_z$ & $L_X$ \\
& [Mpc]& [mag] & [mag] & [mag] & [mag] & [mag] & [mag] & [mag] & [erg s$^{-1}$] \\
\hline
DF1 & 2.0 & $>$24.47 & $>$24.29 & $>$20.90 & 20.22$\pm$0.15 & 19.94$\pm$0.20 & 19.76$\pm$0.26 & $>$19.41 & $\cdots$ \\
DF2 & 1.8 & $>$25.77 & $>$24.26 & $>$21.06 & $>$22.35 & $>$21.76 & $>$21.13 & $>$19.87 & $\cdots$ \\
DF3 & 1.9 & $>$25.33 & $>$24.11 & $>$20.47 & 20.28$\pm$0.17 & 19.59$\pm$0.14 & $>$20.71 & $>$19.33 & $\cdots$ \\
DF4 & 1.5 & $>$24.85 & $>$23.50 & $>$20.68 & $>$21.27 & 19.78$\pm$0.21 & $>$20.53 & 18.21$\pm$0.31 & $\cdots$ \\
DF5 & 2.0 & $>$26.08 & $>$24.28 & $>$21.29 & $>$22.08 & 20.55$\pm$0.21 & 20.24$\pm$0.32 & $>$19.85 & $\cdots$ \\
DF6 & 1.5 & $>$24.12 & $>$23.39 & $\cdots$ & $>$21.44 & 19.81$\pm$0.24 & $>$20.32 & $>$18.86 & $\cdots$ \\
DF7 & 1.3 & $>$24.58 & 22.47$\pm$0.30 & 19.06$\pm$0.21 & 19.16$\pm$0.09 & 18.42$\pm$0.07 & 18.31$\pm$0.09 & $>$18.87 & $\cdots$ \\
DF8 & 0.9 & $>$24.69 & $>$24.34 & $>$20.28 & 20.20$\pm$0.22 & 19.35$\pm$0.16 & $>$20.34 & $>$18.88 & $\cdots$ \\
DF9 & 1.5 & $>$24.97 & $>$23.98 & 19.81$\pm$0.28 & 20.33$\pm$0.17 & $>$21.43 & 19.83$\pm$0.28 & $>$19.36 & $\cdots$ \\
DF10 & 0.6 & $>$25.44 & 24.48$\pm$0.31 & $>$20.97 & 20.93$\pm$0.25 & 19.49$\pm$0.11 & 19.37$\pm$0.16 & 18.45$\pm$0.25 & $\cdots$ \\
DF12 & 0.3 & $>$25.41 & $>$25.97 & $>$21.05 & 20.96$\pm$0.22 & 20.31$\pm$0.22 & 20.11$\pm$0.29 & $>$19.56 & $<3.2\times10^{40}$ \\
DF13 & 0.9 & $>$25.51 & $>$25.39 & $>$21.23 & $>$21.1 & $>$21.75 & $>$21.23 & $>$19.81 & $\cdots$ \\
DF14 & 0.7 & $>$24.72 & $>$24.32 & $>$20.52 & 20.23$\pm$0.20 & $>$21.11 & 19.61$\pm$0.31 & 18.18$\pm$0.30 & $<3.4\times10^{40}$\\
DF15 & 0.6 & $>$24.81 & $>$24.61 & $>$20.65 & 19.74$\pm$0.14 & 19.48$\pm$0.16 & 19.40$\pm$0.27 & 17.61$\pm$0.18 & $<3.4\times10^{40}$ \\
DF16 & 1.1 & $>$25.33 & $>$25.27 & $>$21.62 & $>$22.69 & 21.14$\pm$0.28 & $\cdots$ & $\cdots$ & $\cdots$ \\
DF17 & 0.9 & $>$24.79 & $>$24.32 & $>$20.52 & 19.16$\pm$0.09 & 19.66$\pm$0.21 & 18.66$\pm$0.16 & 18.12$\pm$0.33 & $<6.6\times10^{40}$ \\
DF18 & 0.4 & 24.01$\pm$0.28 & $>$24.99 & $>$20.97 & $>$21.92 & 20.38$\pm$0.27 & $>$20.70 & $>$19.37 & $<2.0\times10^{40}$ \\
DF19 & 1.7 & $>$24.64 & $>$23.38 & $>$20.40 & $>$21.44 & $>$20.95 & $>$20.24 & $>$18.91 & $\cdots$ \\
DF20 & 0.4 & $>$25.23 & $>$24.89 & $>$21.22 & $>$22.13 & 20.80$\pm$0.32 & $>$20.96 & $>$19.66 & $<1.5\times10^{40}$ \\
DF21 & 0.9 & $>$26.11 & $>$26.09 & $>$21.60 & 20.50$\pm$0.13 & 20.13$\pm$0.12 & 19.90$\pm$0.17 & $>$20.01 & $<7.6\times10^{40}$ \\
DF22 & 1.3 & $\cdots$ & $>$24.44 & $>$21.39 & $>$21.1 & 20.24$\pm$0.18 & 19.88$\pm$0.22 & $>$19.73 & $\cdots$ \\
DF23 & 0.4 & $>$25.37 & 24.40$\pm$0.32 & $>$21.25 & 20.49$\pm$0.16 & 20.54$\pm$0.25 & 19.93$\pm$0.26 & $>$19.69 & $<4.2\times10^{40}$ \\
DF24 & 1.3 & $>$24.73 & $>$24.37 & $>$21.53 & $>$22.48 & $>$21.94 & $>$21.30 & $>$20.09 & $\cdots$ \\
DF25 & 0.3 & $>$24.70 & 23.12$\pm$0.16 & $>$20.55 & 20.10$\pm$0.20 & 19.94$\pm$0.27 & 19.32$\pm$0.29 & $>$18.92 & $<1.8\times10^{40}$ \\
DF26 & 0.4 & $>$24.78 & $>$24.94 & $>$20.81 & 19.46$\pm$0.10 & 18.69$\pm$0.07 & 18.44$\pm$0.10 & 18.13$\pm$0.25 & $\cdots$ \\
DF27 & 0.5 & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $<1.4\times10^{40}$ \\
DF28 & 0.4 & $>$25.40 & 23.55$\pm$0.16 & $>$21.02 & 20.24$\pm$0.14 & 19.32$\pm$0.09 & 19.48$\pm$0.18 & $>$19.62 & $<4.0\times10^{40}$ \\
DF29 & 0.8 & 23.94$\pm$0.26 & $>$25.07 & 19.96$\pm$0.28 & 20.39$\pm$0.18 & 19.59$\pm$0.14 & 19.59$\pm$0.23 & $>$19.50 & $\cdots$ \\
DF31 & 1.9 & $\cdots$ & $>$23.95 & $>$20.75 & 20.79$\pm$0.23 & 20.26$\pm$0.23 & 19.80$\pm$0.24 & $>$19.15 & $\cdots$ \\
DF32 & 1.4 & $>$25.49 & $\cdots$ & $>$20.65 & 20.26$\pm$0.16 & 19.44$\pm$0.12 & 19.87$\pm$0.28 & $>$19.03 & $\cdots$ \\
DF33 & 1.8 & $>$26.15 & $\cdots$ & $>$21.18 & $>$22.37 & $>$21.88 & $>$21.28 & $>$19.48 & $\cdots$ \\
DF34 & 1.6 & $>$25.25 & $\cdots$ & $>$20.57 & $>$21.75 & 20.30$\pm$0.29 & $>$20.68 & $>$18.85 & $\cdots$ \\
DF36 & 1.8 & $>$25.54 & $\cdots$ & $>$20.81 & 20.54$\pm$0.19 & 20.00$\pm$0.19 & $>$20.84 & $>$19.24 & $\cdots$ \\
DF38 & 1.4 & $>$25.97 & $>$24.27 & $>$21.23 & 20.75$\pm$0.17 & 20.03$\pm$0.14 & $>$21.24 & $>$19.57 & $\cdots$ \\
DF39 & 1.3 & 23.97$\pm$0.23 & 23.43$\pm$0.17 & 19.55$\pm$0.27 & 20.59$\pm$0.26 & 19.40$\pm$0.14 & 19.04$\pm$0.16 & $>$18.76 & $<1.2\times10^{40}$ \\
DF40 & 1.5 & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ & $\cdots$ \\
DF41 & 1.8 & $>$25.05 & $>$24.81 & $>$20.69 & 20.52$\pm$0.22 & 19.67$\pm$0.17 & 19.34$\pm$0.20 & 18.03$\pm$0.26 & $\cdots$ \\
DF42 & 1.7 & $>$24.55 & $>$23.86 & $>$20.84 & $>$21.90 & 19.91$\pm$0.18 & 19.56$\pm$0.21 & $>$19.24 & $\cdots$ \\
DF43 & 2.6 & $\cdots$ & $>$26.42 & $>$21.75 & $>$22.72 & $>$22.19 & $>$21.65 & $>$20.14 & $\cdots$ \\
DF44 & 1.8 & $>$24.54 & $>$23.78 & 19.31$\pm$0.26 & 19.21$\pm$0.10 & 18.75$\pm$0.10 & 18.51$\pm$0.14 & $>$18.76 & $<1.1\times 10^{38}$ \\
DF45 & 2.9 & $>$25.26 & $>$24.49 & $>$21.56 & 20.51$\pm$0.13 & 20.41$\pm$0.18 & 19.91$\pm$0.17 & $>$19.95 & $\cdots$ \\
DF46 & 2.1 & $>$24.84 & $>$23.50 & $>$20.91 & $>$21.84 & 19.59$\pm$0.15 & $>$20.67 & 18.24$\pm$0.26 & $\cdots$ \\
DF47 & 2.9 & $>$24.96 & $>$23.53 & 19.84$\pm$0.32 & 20.31$\pm$0.21 & $>$21.05 & 19.12$\pm$0.19 & $>$19.03 & $\cdots$ \\
\hline
\multicolumn{10}{p{0.9\textwidth}}{Columns. (1) Dragonfly ID from \cite{vdokkum2015}, (2) projected distance from Coma cluster center, (3-7) AB magnitudes or 3$\sigma$ upper limits, (8) 3$\sigma$ upper limits to the 0.3-10~keV X-ray luminosity.} \\
\multicolumn{10}{p{0.9\textwidth}}{Notes. DF11, DF27, DF30, DF35, and DF37 were excluded from the UV analysis due to contamination or artifacts. Individual filters in several other galaxies were excluded for the same reasons.}
\end{tabular*}
\end{table*}
\subsection{DF44 \textit{XMM-Newton} Data}
\begin{figure*}
\includegraphics[width=\linewidth]{figures/full_stacks_rev1.pdf}
\caption{Full stacks of all DF data with filter, magnitude, source significance, and effective exposure time labeled. The images have been convolved with a Gaussian kernel of $\sigma=3$ pixels. The dashed black circle in each filter is 12" in radius.}
\label{fig:fullstackims}
\end{figure*}
\begin{figure}
\includegraphics[width=0.95\linewidth]{figures/undetected_DF.pdf}
\caption{The stacked NUV image of all individually undetected DF galaxies shows a clear detection (3$\sigma$ contour above background shown in white). The apparent magnitude measured from within an aperture about twice the average effective radius is highly significant, motivating the use of shallower stacks in our analysis.}
\label{fig:undetected_nuv}
\end{figure}
We obtained about 167~ks of new XMM observations, which were split between December 23rd, 2017 (OBS-ID:0800580101) and January 4th, 2018 (OBS-ID:0800580201). We used the EPIC MOS1, MOS2, and pn cameras. The XMM data were processed using standard methods with the Science Analysis System (SAS v16.0.0)\footnote{\url{https://www.cosmos.esa.int/web/xmm-newton/what-is-sas}}. We used the SAS tools {\tt emchain} and {\tt epchain} to register events for the MOS and pn cameras (we used events with pattern$\le$4), then measured the count rate over the full arrays in bins of 200~s to identify particle background flares, which were filtered out. After processing, the good time intervals (GTIs) for the three cameras in both observing periods are 66 ks and 46 ks in MOS1, 70 ks and 54 ks in MOS2, and 60 ks and 37 ks in pn, respectively.
The quiescent particle background was subtracted based on the unexposed array corners with the Extended-SAS software. We then filtered the images to the 0.4-2~keV bandpass for source detection, which maximizes the sensitivity to the expected sources (hot gas or LMXBs). However, our results are consistent with measurements made from a 0.3-10~keV bandpass. No source is detected at the position of DF44, and the 0.3-10~keV, 3$\sigma$ upper limit of $L_X< 1.0\times 10^{38}$~erg~s$^{-1}$ is much lower than the XRT value ($L_X < 4.1\times 10^{39}$~erg~s$^{-1}$).
DF44 is on the outskirts of the Coma cluster, just south of the mosaic produced by \citet{Briel2001}. In the XMM field we do not find strong cluster emission, and instead find that such emission can only be detected when integrating over large areas and that its surface brightness is below that of the other backgrounds and foregrounds. We find that the background around DF44 is essentially flat (when accounting for vignetting and the extended wings of other nearby sources) on a scale of several arcmin, so we adopt an annular region of width 10~arcsec for the background. By fitting a linear gradient to the 0.4--2.0~keV image, we estimate that the uncertainty in the upper limit quoted above is of order 1\% due to the intracluster medium, i.e., smaller than other uncertainties.
\section{Ultraviolet Fluxes and Colors}
A minority of galaxies were detected in either the NUV or FUV channels, and the detected magnitudes or upper limits are consistent with quiescent galaxies. To improve the sensitivity, we stacked the data in both GALEX and the \textit{ugriz} SDSS bands to measure the average SED (Figure~\ref{fig:fullstackims}; for presentation, each stack was convolved with a Gaussian kernel of $\sigma=3$ pixels). The SED is consistent with the DF UDGs being quiescent, red galaxies. However, in contrast to \citet{singh2019}, who argued that stacking was not likely to be useful for ``red'' UDGs, we detect some DF UDGs individually and more through stacks. Figure~\ref{fig:undetected_nuv} shows the stacked NUV galaxies that are \textit{not} individually detected. There is clearly a strong detection (the white contour is 3$\sigma$ above background).
\subsection{Correlation with Distance}
\begin{figure*}
\includegraphics[width=\linewidth]{figures/NUV_R_runningstack.pdf}
\caption{Scatter plots of apparent magnitudes from running stack samples for NUV and \textit{r} data as a function of the weighted average stack distance from the cluster center. All scatters were fit with \texttt{linmix}. N denotes the number of galaxies in each stack, $n_{\mathrm{det}}$ denotes the number of points with S/N $>$ 3 which are plotted by squares with error bars, $n_{\mathrm{lim}}$ denotes the number of points fit as $3\sigma$ upper limits which are plotted by inverted triangles, and $\sigma_{\mathrm{m}}$ denotes the slope significance. Note $\sigma_{\mathrm{m}}$ is different from the $\sigma$ indicating the Gaussian scatter in the regression. The lighter dashed grey lines indicate the slope uncertainty.}
\label{fig:NUV_R_runningstack}
\end{figure*}
Motivated by the \cite{jiang2019} scenario in which ram-pressure stripping quenches star formation and the good signal in the full stacks, we also stacked sub-samples of the data to search for a correlation between UV luminosity and distance from the cluster core. First, we created a ``running'' stack in which $N$ galaxies were stacked at a time, with the first point combining the $N$ galaxies closest to the core, and the next point excluding the closest galaxy and including the next farthest not already in the stack. Since the stacked images have different exposure times, the mean distance for each point is weighted by the exposures of each constituent.
Running stacks with $N=3$ and 5 were created for the FUV, NUV, and SDSS $ugriz$ bands. Of these, the FUV, $u$, and $z$ bands lack the signal for useful fitting. We fitted the remainder using the {hierarchical Bayesian linear regression model} \citep[{\tt linmix} from][]{Kelly2007}, which has the form $m = \alpha + \beta R \pm \sigma$, where $m$ is the magnitude, $\alpha$ is the intercept, $\beta$ is the slope, $R$ is the projected radius from the Coma core in Mpc, and $\sigma$ is a Gaussian scatter term. {Upper limits are included through standard approaches to censored data.} The significance of each parameter is estimated from the posterior distribution produced by {\tt linmix}. {The $g$ and $i$ band fits are similar to $r$ with less significance, so only the fits to $r$ are presented.} The NUV and $r$ fits are summarized in Table~\ref{tab:stackfits}. The fits (especially to the $N=5$ stacks) hint at correlations with distance, but none are inconsistent with zero slope at the $3\sigma$ level.
\begin{figure}
\includegraphics[width=\linewidth]{figures/NUV_R_binstack.pdf}
\caption{{ Scatter plots of apparent NUV (\textit{top}) and \textit{r} (\textit{bottom}) magnitudes from galaxy stacks of 0.5 Mpc bins with a similar layout as Figure \ref{fig:NUV_R_runningstack}. The numbers next to the data points denote the number of galaxies within each stack and triangles indicate upper limits. Both scatters were fit using \texttt{linmix}.}}
\label{fig:NUV_R_binstack}
\end{figure}
In addition to the running stacks, we measured the fluxes from stacks based on distance (i.e., with a non-uniform number of galaxies per stack). We used bins 0.5~Mpc wide, as shown in Figure~\ref{fig:NUV_R_binstack}. In all but one NUV bin we have a clear detection, so we applied the same regression as to the running stacks. We find a significant correlation of $r$-band magnitude with distance but not NUV for this binning. The FUV is too faint to search for a correlation, and the signal-to-noise ratio in $g$, $i$, and $z$ stacks is considerably lower than for $r$, leading to best-fit slopes in agreement with $r$ but less significant.
The running stacks have an advantage over the distance-binned stack in that the distance-binned stack can have many more galaxies per data point at one distance than another. Since there is certainly scatter in the magnitudes at any radius, the formal uncertainties on the magnitudes in the distance-binned stack could be too low. However, the running stacks also have a disadvantage in that the central data points are each used $n=3$ or 5 times, whereas the outermost point only contributes to a single stack. The outermost galaxies provide a lever arm on any correlation and are important. We note that the reported slopes for each stack method are consistent with each other, supporting the prospect of a weak correlation of $r$ magnitude with distance and significant galaxy-to-galaxy scatter.
\subsection{UV-Optical Colors}
\begin{figure}
\includegraphics[width=\linewidth]{figures/dfcmrs.pdf}
\caption{Color-magnitude comparisons of NUV$-$\textit{r} \textit{(top)} and \textit{g}$-$\textit{r} \textit{(bottom)} against \textit{r} magnitudes for the DF sample. The red circle denotes DF44's position in the plot. Only galaxies with S/N $>$ 3 are shown.}
\label{fig:dfcmrcomparisons}
\end{figure}
The NUV detections suggest that the UDGs have recently formed stars. \citet{kaviraj2007} studied UV-optical colors of early-type galaxies and found that a color index of NUV$-r < 5.5$~mag indicates star formation within the past Gyr, with the younger stars representing $\sim$1\% of the galaxy's stellar mass (implying a mean SFR$< 10^{-3} M_{\odot}$~yr$^{-1}$). This is consistent with a similar analysis by \cite{Schawinski2007} that identified NUV$-r < 5.4$~mag as indicating recent star formation. Although this criterion was applied to early-type galaxies and not UDGs, one would expect UV-optical colors to reflect starlight and the UV-optical color-magnitude diagram for the UDGs (Figure~\ref{fig:dfcmrcomparisons}) is consistent with early-type galaxies \citep[see Figure~4 in][]{kaviraj2007}. Moreover, most of the Coma DF UDGs are part of the ``red'' family of UDGs \citep{zaritsky2019}. Notably, all of the UDGs fall below the NUV$-r < 5.5$~mag cutoff, so they have likely formed some stars in the past Gyr. This is also true for the average NUV$-r$ color for the full stack of about 4.7~mag.
Meanwhile, the mean FUV magnitude $m_{\text{FUV}} = 25.84\pm0.28$ from a stack of all Coma UDGs indicates that current star formation is much lower. This corresponds to a FUV luminosity $\nu L_{\nu} \approx 4\times 10^{39}$~erg~s$^{-1}$. \citet{Calzetti2013} suggest a conversion factor of SFR~$=3.0\times 10^{-47} \times \nu L_{\nu} M_{\odot}$~yr$^{-1}$, yielding SFR~$\approx 1.2\times 10^{-7} M_{\odot}$~yr$^{-1}$. There is clearly variation well in excess of the 0.28~mag uncertainty, with some galaxies detected with $m_{\text{FUV}} < 24$ (Table~\ref{tab:udgoptuvxray}). Nevertheless, the typical UDG is not forming stars rapidly enough to explain its NUV$-r$ color in the steady state.
\begin{table}
\centering
\caption{Stacked NUV and $r$ Fit Parameters.\label{tab:stackfits}}
\begin{tabular*}{0.5\textwidth}{@{\extracolsep{\fill}}ccccc}
\hline
{Band} & {N} & {Slope} & {Intercept} & {Significance} \\
& & & [mag] & [$\sigma$] \\
\hline
Running & & & & \\
\hline
$r$ & 3 & $0.13\pm0.12$ & $19.8\pm0.2$ & 1.1\\
$r$ & 5 & $0.14\pm0.07$ & $19.8\pm0.1$ & 2.0 \\
NUV & 3 & $0.12\pm0.26$ & $24.1\pm0.4$ & 0.5 \\
NUV & 5 & $0.31\pm0.25$ & $24.0\pm0.3$ & 1.2 \\
\hline
Distance & & & & \\
\hline
$r$ & bin & $0.19\pm0.05$ & $19.8\pm0.1$ & 3.8 \\
NUV & bin & $0.09\pm0.30$ & $24.1\pm0.8$ & 0.3 \\
\hline
\multicolumn{5}{p{0.4\textwidth}}{Notes. The values given are from a linear fit to the data shown in Figure~\ref{fig:NUV_R_runningstack} and Figure~\ref{fig:NUV_R_binstack}.}
\end{tabular*}
\end{table}
\section{\textit{Chandra} Limits}
\begin{figure}
\includegraphics[width=\linewidth]{figures/chandra_stack.pdf}
\vspace{-0.2cm}
\caption{{Stacked 2-7~keV \textit{Chandra} image for 12 DF UDGs in the Coma cluster (in units of counts per 100~ks, left) and the exposure map in units of 1000~s (right). The galaxies and observations included are listed in Table~\ref{tab:datasources}. There is no detection within the 15~arcsec aperture shown in green, and the 3$\sigma$ upper limit (reported in the 0.3-10~keV bandpass) is $L_X < 2\times 10^{39}$~erg~s$^{-1}$.}}
\label{fig:chandrastack}
\end{figure}
The \textit{Chandra} upper limits range from 0.3-10~keV $L_X < 1-7\times 10^{40}$~erg~s$^{-1}$ for the 12 DF galaxies in the field of view. When stacking the data {(Figure~\ref{fig:chandrastack})}, we obtain a limit of $L_X < 2\times 10^{39}$~erg~s$^{-1}$ for the average galaxy, with the caveat that the sample size is small and the detectors are different. {In addition, the exposure times are very different.} Hence, we prefer the individual limits. This is a worse constraint than the recent \citet{kovacs2019} result for isolated UDG candidates as they stacked more galaxies. Assuming that their galaxies are at an average distance of 65~Mpc they report an average $L_X < 3.1\times 10^{37}$~erg~s$^{-1}$. Another reason for their higher sensitivity is that we are limited by the cluster background.
Recently, \citet{kovacs2020} extended that analysis to Coma itself with an investigation of a much larger sample of UDGs from \citet{yagi2016}. They find at most two off-nuclear X-ray sources among Coma UDGs, which is fully consistent with our analysis that used much of the same data but restricted analysis to the DF sample.
The individual upper limits rule out any strong AGN activity. Considering that UDGs tend to have stellar masses consistent with dwarf galaxies ($M_* < 10^9 M_{\odot}$), and thus would be expected to host smaller central black holes if they follow a similar relation to high surface brightness galaxies, we may not be sensitive to weak Seyfert-like accretion. The sample size is too small to use the AGN frequency to test whether UDGs obey scaling relations measured in other systems, but if future surveys find few or no AGNs at luminosities near $10^{40}$~erg~s$^{-1}$, then it would indicate that any massive black holes in UDGs are more likely to scale with the stellar population rather than the (potentially much larger) dynamical mass. This is because the same luminosity probes lower Eddington ratios in larger black holes, as described in \citet{miller2012}.
These limits also far exceed the contributions from bright LMXBs, except perhaps the brightest ultraluminous X-ray sources (ULXs). ULXs are often associated with star formation, and so we do not expect to see any in the Coma UDGs. In the next section, we discuss the constraints on LMXBs in DF44 in more detail.
\section{Dragonfly 44}
\begin{figure}
\includegraphics[width=\linewidth]{figures/DF44_R_UV_Xray.pdf}
\caption{Optical (HST F606W), UV (\textit{Swift} UVW2), and X-ray (\textit{XMM-Newton} 0.4-2 keV) images of DF44. The weak detection in the UV indicates little ongoing star formation, while the non-detection in the X-rays at $10^{38}$~erg~s$^{-1}$ indicates a lack of hot gas and low XRB activity. The circle is the same scale in each panel, and background was measured in annular regions outside the source region.}
\label{fig:DF44_images}
\end{figure}
DF44 merits special consideration because it has some of the strongest evidence for a large dynamical mass and the deepest data {(Figure~\ref{fig:DF44_images})}. When using an elliptical aperture based on the \textit{HST} optical image (as opposed to a circular aperture based on $r_{\text{eff}}$, in which it was not detected), DF44 was detected in the NUV channel in a shallow archival GALEX exposure, {but it is detected in a circular aperture for both the \textit{Swift} and SDSS bands}. We used the UV data to calculate the SFR based on the relation from \cite{rozagonzalez2002}:
\begin{equation*}
\textrm{SFR}(UV)[\textrm{M}_{\odot}\textrm{yr}^{-1}] = 1.4\times10^{-28}L_{\nu}[\textrm{erg}~\textrm{s}^{-1} \textrm{Hz}^{-1}]
\end{equation*}
which is applicable within 1500-2400\AA\ where the considered integrated spectrum is nearly flat in star-forming regions (Table~\ref{tab:df44phot}). The deep UVW2 data (Figure~\ref{fig:DF44_images}) provide the strongest constraint, with SFR $=6\times 10^{-4} M_{\odot}$~yr$^{-1}$. This is significantly below the value for UVW1 (after correcting for the red leak) and NUV values of SFR~$\approx 2\times 10^{-3}$. As we noted above, the NUV light may not trace very young stars, and the difference between the NUV and UVW2 indicates that the NUV light traces stars formed in the past Gyr but not in the past few Myr. Indeed, even some of the UVW2 light likely comes from these older stars, so this SFR is an upper bound.
Nevertheless, if we take SFR $=6\times 10^{-4} M_{\odot}$~yr$^{-1}$ as the steady-state value, in 1~Gyr DF44 would produce $6\times 10^5 M_{\odot}$ of new stars, which is only 0.2\% of its stellar mass ($3\times 10^8 M_{\odot})$. The NUV$-r$ color places DF44 (Figure~\ref{fig:dfcmrcomparisons}) firmly in the regime where there has been star formation in the past Gyr comprising $\sim$1\% of the stellar mass \citep{kaviraj2007}, suggesting that the SFR was higher in the past. Notably, DF44 is photometrically indistinct from the general DF population. This situation is consistent with the sample's NUV$-r$ colors and the finding by \cite{koda2015} that the Coma UDGs are not currently forming stars (based on H$\alpha$ luminosities).
\begin{table*}
\caption{Optical and ultraviolet photometric information for DF44.\label{tab:df44phot}}
\begin{tabular*}{0.7\textwidth}{@{\extracolsep{\fill}}lccccc}
\hline
{Filter} & {$\lambda_{\text{eff}}$} & {$F_{\lambda}(\times10^{-18})$} & {$t_{\text{exp}}$} & {Conv. Factor} & {SFR} \\
& {[\AA]} & {[erg~s$^{-1}$~cm$^{-2}$~\AA$^{-1}$]} & {[s]} & & {[$M_{\odot}$~yr$^{-1}$]}\\
\hline
FUV & 1516 & $<14$ & 1488 & (1) $1.40\times10^{-15}$ & $\cdots$ \\
UVW2 & 1928 & $2.8\pm1.3$ & 40876 & (1) $5.98\times10^{-16}$ & $6\times10^{-4}$\\
NUV & 2267 & $6.8\pm2.6$ & 1639 & (1) $2.06\times10^{-16}$ & $2\times10^{-3}$\\
UVW1 & 2505 & $7.2\pm1.4$ & 30134 & (1) $4.21\times10^{-16}$ & $3\times10^{-3}$\\
SDSS \textit{u} & 3580 & $37\pm4.1$ & 54.0 & (2) $8.63\times10^{-18}$ & $\cdots$ \\
SDSS \textit{g} & 4754 & $42\pm3.6$ & 54.0 & (2) $4.96\times10^{-18}$ & $\cdots$ \\
SDSS \textit{r} & 6166 & $42\pm3.3$ & 54.0 & (2) $2.86\times10^{-18}$ & $\cdots$ \\
\hline
\multicolumn{6}{p{0.6\textwidth}}{Conversion factors from standard calibration: (1) counts~s$^{-1}$ to $F_{\lambda}$ ; (2) nMgy to $F_{\lambda}$.}
\end{tabular*}
\end{table*}
Compared to the \textit{Chandra} upper limits on the X-ray luminosity for Coma DF galaxies, DF44 has a much tighter {0.3-10~keV, 3$\sigma$ upper limit} of $L_X< 1.0\times 10^{38}$~erg~s$^{-1}$, {although it is individually undetected (E. Miller 2019, private communication)}. This is only a few times higher than the \cite{kovacs2019} measurement from a stack of isolated UDG candidates, and it is sensitive enough to constrain the X-ray binary population, low level nuclear activity, and the presence of hot gas. In particular, it rules out the presence of any nuclear activity that could suppress star formation. This is not surprising, since most local early-type galaxies with similar stellar mass are not detected at this sensitivity \citep{gallo2010,miller2012}. However, we cannot rule out very low level or highly radiatively inefficient accretion, since Sgr~A* has a quiescent luminosity of a few $10^{33}$~erg~s$^{-1}$ \citep{baganoff2003}.
The brightest non-ULX LMXBs have {$0.3-10$~keV} luminosities exceeding $10^{38}$~erg~s$^{-1}$, so it is worth examining whether the X-ray non-detection for DF44 constrains its LMXB population. The total LMXB luminosity is known to scale with the stellar mass \citep{gilfanov2004}, and using the \cite{lehmer2010} scaling of $L_{X,\text{LMXB}} = 9\times 10^{28} \times M_*$, the average LMXB luminosity for a DF44-sized galaxy is $3\times 10^{37}$~erg~s$^{-1}$, which is consistent with the upper limit from XMM. However, LMXBs are Poisson distributed according to the X-ray luminosity function \citep{gilfanov2004}, which is a power law in the regime of $L_X \sim 10^{38}$~erg~s$^{-1}$. Thus, there is a finite chance of detecting LMXBs above a given luminosity for any galaxy. To estimate this probability, we adopt the formalism of \cite{foord2017} and \cite{lee2019}, which uses the \cite{lehmer2010} scaling and luminosity function to estimate the average \textit{number} of LMXBs and treats it as the mean of a Poisson distribution. Then, galaxies are simulated drawing from this distribution and randomly assigning a luminosity weighted by the luminosity function. With a large number of draws from the sample, one can calculate the likelihood of detecting a total LMXB luminosity above some threshold. For DF44 and the $10^{38}$~erg~s$^{-1}$ limit, we find a 95\% chance of no detection. This amounts to a stronger statement that the non-detection of LMXBs is expected from the stellar mass.
However, it is not clear whether LMXBs scale with stellar mass or GCs, since a large fraction of LMXBs may form in GCs before migrating into the rest of the galaxy. In the Milky Way galaxy, GCs host about 10\% of the active LMXB population, while at the same time accounting for less than 1\%\ of the light, implying that LMXB formation in GCs must be hundreds of times more efficient than in the field \citep{clark1975}. This has long been attributed to enhanced dynamical formation mechanisms in the dense cores of GCs \citep{pooley2003}. Since the number of GCs and stellar mass both typically scale with dynamical mass, for most galaxies one cannot tell the difference. However, one piece of evidence in favor of a large dynamical mass in DF44 is its large number ($74\pm14$) of GCs identified in deep Gemini images \citep{vdokkum2016,vdokkum2017}. This number is consistent with a Milky Way-like dynamical mass, and much higher than expected for the stellar mass.
For nearby galaxies, the odds of detecting a GC LMXB above $10^{38}$~erg~s$^{-1}$ are $\sim$5-10\%, with a mild dependence on morphological type. More specifically, the number of GC LMXBs brighter than $10^{38}$ erg~s$^{-1}$ is estimated by \cite{sarazin03} to be $1.5\times 10^{-7}$ per optical luminosity of GCs (or a 1-2\% chance detection for a $10^5$ solar mass GC). {While the actual number likely depends on other GC properties as well (such as color and concentration), overall, the percentage of GCs containing a bright LMXB ranges between a few up to 10\% \citep{sivakoff07}}.
Additionally, \cite{peacock2016} find that the XLF of GC XRBs is flatter than in the field, implying that the chances of detecting a bright LMXB are somewhat higher in GCs than they are in the field (this is not necessarily true at lower X-ray luminosities, where the XLF slope(/s) is(/are) flatter and the chance of detection depends more strongly on the actual slope value). In DF44, if we adopt a conservative value of a 5\% chance detection per GC and consider a total of 74 GCs, then we infer a 2\% probability of detecting \textit{no} GC LMXB in our data. The probability drops to 0.04\% if we assume a 10\% detection chance for a single GC. Overall, these numbers indicate that the GCs in DF44 may be fairly inefficient at producing high luminosity LMXBs. However, as shown by \cite{kundu2007}, GC metallicity is likely to play a major role in these estimates (with metal-rich GCs being three times as likely to host LMXBs).
Finally, we estimated the sensitivity to a hot gas halo with $kT=0.2$~keV and $Z=0.3 Z_{\odot}$, which is based on the parameters measured around the Milky Way \citep{matt_miller15}. We used an elliptical aperture with semi-major and semi-minor axes $a=9^{\prime\prime}$ and $b=6^{\prime\prime}$, respectively, and a position angle matched to DF44 based on the HST image. The XMM non-detection limits the emission measure to EM $< 2.5\times 10^{-7}$~cm$^{-3}$ for these parameters, which implies a mass $M_{\text{hot}} < 6\times 10^6 M_{\odot}$ in the region, assuming an ellipsoidal, axisymmetric volume with $b=c=6^{\prime\prime}$ at $d\approx 100$~Mpc. This rules out a significant hot halo: assuming that there is this much gas and that it is distributed following the same $\beta$ model measured around the Milky Way \citep{matt_miller15}, with $\beta \approx 0.5$ and a core radius of 3-4~kpc, the limit to the mass within the virial radius is $M_{\text{hot}}(r<r_{\text{vir}}) < 10^8 M_{\odot}$. It is more likely that there is no extended hot halo, and that any hot gas in the region is essentially contiguous with the intracluster medium.
\section{Discussion}
The UV and X-ray properties of DF44 and other Coma DF UDGs offer clues to their formation. The NUV$-r$ colors (Figure~\ref{fig:dfcmrcomparisons}) suggest that at least 1\% of the stellar mass has formed in the past Gyr \citep{kaviraj2007}, but their current SFR is much smaller (near $10^{-7} M_{\odot}$~yr$^{-1}$ on average and less than $6\times 10^{-4} M_{\odot}$~yr$^{-1}$ in DF44). This would lead to a buildup of less than $<$~0.2\% in stellar mass over 1~Gyr, which is consistent with \citet{koda2015}. It is possible that the NUV does not trace young stars, but instead hot horizontal branch stars \citep{OConnell1999}, which may be responsible for the ``UV upturn'' seen in elliptical galaxies \citep{Yi2011}. However, the sample-averaged FUV$-$NUV magnitude of 1.15 is inconsistent with the UV upturn \citep{Yi2011} and the colors of most of the detected UDGs are too blue (NUV$-r \lesssim 4$~mag) to explain apart from residual star formation.
The discrepancy between ongoing and recent star formation indicates that the galaxies were recently quenched within the past 1~Gyr. One possible explanation is ram-pressure stripping \citep{jiang2019}, {which is based in part on their report of a $B-R$ color gradient. If so, we would also expect to see a color gradient in NUV$-r$, since galaxies that have recently passed through the Coma core should be quenched. Meanwhile, assuming that UDGs have a similar velocity dispersion to the brighter Coma members, it is possible that galaxies at the outskirts have yet to pass through the core.} For example, DF44 is offset from the Coma systemic velocity by 900~km~s$^{-1}$ and lies at a projected distance of 1.8~Mpc from the Coma core, placing a lower limit of 3~Gyr on its last possible core passage.
We find a weak gradient in NUV$-r$, but only because the average $r$ band magnitudes increase by $\approx$~0.5~mag from 0 to 2.5~Mpc, corresponding to a decrease in stellar mass of about a factor of two. There is no significant correlation between NUV flux and distance from the core. One reason may be that estimating the distance from the core by projected radius can be unreliable, and another is that, depending on when the UDGs formed, some fraction of UDGs at the outskirts have already passed through the core. We also do not find a significant correlation between $g$ or $i$ and distance from the core, but this is because of the considerably higher number of undetected galaxies in these bands. Regardless, the $r$ band correlation with projected radius provides tentative support to the \citet{jiang2019} scenario, which in turn explains our more robust result that complete quenching occurred in the past Gyr.
If UDGs have dark-matter fractions exceeding 99\%, then multiple Coma UDGs should be massive enough to retain hot halos from their formation \cite{white&frenk1991}. DF44 is sufficiently far from the cluster core (1.8~Mpc in projection) that it could have a hot corona if it did not pass through the denser regions of the intracluster medium. The lack of such a halo may indicate that it was previously stripped (hot gas is easier to strip than cold gas), which is consistent with the UV data, that it is not dynamically massive \citep{kovacs2019}, or that the hot halo failed to form. In any case, the absence of a halo implies that there is no hidden reservoir of material that could fuel future star formation, and the current lack of star formation and low stellar mass imply that no X-ray corona will be built up through stellar feedback. Meanwhile, using a larger sample, \citet{kovacs2020} argue that the lack of X-ray point sources in Coma UDGs implies that they are dwarf galaxies, which is consistent with our stacked X-ray upper limits and the expected number of X-ray binaries from the stellar mass (as opposed to GCs).
The only galaxy with relatively tight limits on LMXB emission is DF44, and the limit of $10^{38}$~erg~s$^{-1}$ is fully consistent with the expectation from its low stellar mass. However, if most LMXBs are formed in GCs, the lack of detection in DF44 is surprising. This suggests that DF44 GCs were inefficient at producing LMXBs, but deeper data would be needed to fully explore this issue. Future surveys with high-resolution X-ray optics could determine whether UDG GCs are generally inefficient at forming LMXBs, as multiple Coma UDGs have many more GCs than expected for their stellar masses.
\section{Summary}
We investigated DF UDGs in the Coma cluster using new and archival X-ray and UV data. Our principal findings include:
\begin{enumerate}
\item We clearly detect the UDGs in the FUV and NUV bands when stacking all the data. {7.5\% of galaxies (3/40) are individually detected in the FUV, and about 15\% (6/39) in the NUV}, although we note that the fields do not have uniform sensitivity in either band. The NUV$-r$ colors imply that at least 1\% of the stellar mass has formed within the past Gyr, but the average FUV magnitude implies that the SFR has since fallen by several orders of magnitude.
\item There is a weak correlation between $r$-band magnitude and distance from the cluster core that implies that the stellar mass decreases by a factor of $\sim$2 from 0 to 2.5~Mpc. On the other hand, no correlation is seen in the NUV. Together, these imply that recent star formation has been slightly more efficient in the cluster outskirts, but with galaxy-to-galaxy scatter.
\item There is no significant AGN activity in the 13 galaxies for which there are X-ray data. DF44 also has no bright X-ray binaries, which is expected if LMXBs track stellar mass but not if they track GCs, of which DF44 has many. The lack of X-ray detections is similar to what is found in isolated UDGs \citep{kovacs2019} and consistent with a wider study of Coma UDGs \citep{kovacs2020}.
\item DF44 has little or no hot gas bound to its potential ($M_{\text{hot}} < 6\times 10^6 M_{\odot}$ within the XMM aperture, or $<10^8 M_{\odot}$ within the virial radius), whereas one would expect a hot halo for a dynamical mass near $10^{12} M_{\odot}$.
\end{enumerate}
Overall, the drop in SFR and the correlation of $r$ magnitude with projected distance from the Coma core is consistent with quenching by ram-pressure stripping, as suggested by \citet{jiang2019}. Ram-pressure stripping would also explain the loss of the hot gaseous halo even if DF44 is dynamically massive. However, deeper UV data are required to confirm this hypothesis, as any correlation between NUV flux and projected distance is too weak to find in existing data.
\section{Acknowledgments}
The authors thank the reviewer for helpful comments and suggestions that improved the manuscript. C.~L. and E.~G. gratefully acknowledge support for this work under the NASA award \#80NSSC18K0376.
This work is based on observations obtained with \textit{XMM-Newton}, an ESA science mission with instruments and contributions directly funded by ESA Member States and NASA.
This research has made use of data and/or software provided by the High Energy Astrophysics Science Archive Research Center (HEASARC), which is a service of the Astrophysics Science Division at NASA/GSFC and the High Energy Astrophysics Division of the Smithsonian Astrophysical Observatory.
Funding for SDSS-III has been provided by the Alfred P. Sloan Foundation, the Participating Institutions, the National Science Foundation, and the U.S. Department of Energy Office of Science. The SDSS-III web site is http://www.sdss3.org/.
SDSS-III is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS-III Collaboration including the University of Arizona, the Brazilian Participation Group, Brookhaven National Laboratory, Carnegie Mellon University, University of Florida, the French Participation Group, the German Participation Group, Harvard University, the Instituto de Astrof\'isica de Canarias, the Michigan State/Notre Dame/JINA Participation Group, Johns Hopkins University, Lawrence Berkeley National Laboratory, Max Planck Institute for Astrophysics, Max Planck Institute for Extraterrestrial Physics, New Mexico State University, New York University, Ohio State University, Pennsylvania State University, University of Portsmouth, Princeton University, the Spanish Participation Group, University of Tokyo, University of Utah, Vanderbilt University, University of Virginia, University of Washington, and Yale University.
\bibliographystyle{mnras}
|
\section{Introduction}
Knowledge tracing is a fundamental task in computer-aided educational system, which can benefit both learning and teaching~\cite{anderson2014engaging,burns2014intelligent,jones2004ubiquitous}. A typical knowledge tracing process is as follows: when a question is posted, a learner reads its text and applies knowledge to solve it. After getting a bunch of learners' interactive items, corresponding knowledge sequences and interactive sequences of exercises are extracted to train a KT model which could predict the hidden knowledge states of a learner. However, there are two main problems shown in the traditional knowledge tracing workflow: (1) Exercise representation loss problem: traditional workflow simplifies the exercising records as knowledge sequences, which ignores the difficulty and semantic information contained in exercises. In other words, existing methods have an information loss in the process of representing exercises. (2) Insufficient diagnosis problem: it is also difficult to provide specific learning suggestions based on the diagnosis results. Specifically, suppose we know a learner has a weak knowledge (e.g., $Coordinate\ Calculation$), it is hard to decide which one fits him better (e.g., $e_{7}$ or $e_{8}$ in Fig.~\ref{fig:inspiration}), since mappings between knowledge and exercises is too broad~\cite{hontangas2000choice,zhao2018automatically}.
In the literature, there are many efforts in knowledge tracing. Existing methods can be divided into two tracks: traditional knowledge based and exercise based track. The traditional knowledge based method converts learners'exercising sequences into knowledge sequences without considering the text information of the exercises. The most popular one is Bayesian Knowledge Tracing (BKT)~\cite{BKT}, which updates learner's knowledge state by a Hidden Markov Model. Deep learning methods like Deep Knowledge Tracing (DKT) model the learning process as a recurrent neural network~\cite{DKT}. Dynamic Key-Value Memory Network (DKVMN) enhances the ability of the recurrent neural network by introducing two memory matrices to represent knowledge and learner's mastery for each knowledge respectively~\cite{DKVMN}. Graph-based Knowledge Tracing (GKT) combines knowledge tracing with a graph neural network~\cite{GKT}. It encodes a learner's hidden knowledge state as embeddings of graph nodes and updates the state in a knowledge graph. Those models have been proved effective but still have limitations. Most existing methods face the problem of exercise representation loss since they have not taken the texts of exercises into consideration. For the exercise based track, to the best of our knowledge, Exercise Enhanced Knowledge Tracing (EKT) is the first method to incorporate features of exercise text into the knowledge tracing model~\cite{EKT}. However, EKT extracts features of text by feeding the text of exercise directly into a bi-directional LSTM network~\cite{memeory}, which fails to consider the latent hierarchical graph nature of exercises and brings in extra noise from embeddings of text.
Our insight of solving the exercise representation loss problem and insufficient diagnosis problem hinges upon the idea of fully exploring latent hierarchical graph relations between exercises. Incorporating hierarchical relations between exercises could not only increase the accuracy of learner's performance prediction, but also enhance the interpretability of knowledge tracing. Fig.~\ref{fig:inspiration} clearly illustrates how hierarchical relations affect knowledge diagnosis results and the advantages of our approach compared to traditional knowledge tracing methods. As researchers have demonstrated the effectiveness of exploiting prerequisite relations into KT tasks\cite{chen2018prerequisite}, we decompose the hierarchical graph relations between exercises into direct support and indirect support relations. The intuition of support relations is that they represent different types of exercise learning dependencies, which could act as constraints for a knowledge tracing task. Moreover, to learn more fine-grained representation of exercises and avoid single exercise noises, inspired by math word solving technique~\cite{zhang2019gap,fuchs2004enhancing}, we introduce the concept of problem schema to summarize a group of similar exercises with similar solutions. The relation of two exercises is indirect support only if they belong to the same problem schema. It is worth mentioning that, we assume each exercise only has one main knowledge and one problem schema. Considering that exercises belonging to different knowledge could have similar solutions, and exercises with the same knowledge may also belong to different problem schemas due to their difficulty difference, we assume the relations between knowledge and problem schemas are many-to-many (as shown in Fig.~\ref{fig:relation}(a)).
\iffalse
Fig.~\ref{fig:inspiration} illustrates why we need mining the relationship between exercises and how hierarchical relations affect knowledge diagnosis results. As researchers have demonstrated the effectiveness of exploiting prerequisite relations into KT tasks\cite{chen2018prerequisite}, we decompose the hierarchical graph nature of exercises into direct support and indirect support relations, where direct support refers to those exercises whose solutions have direct support relations and indirect support refers to those exercises with same problem schema. The intuition for the exercise support is that it represents exercise learning dependencies, which could act as constraints on the knowledge tracing task. Besides, to learn more fine-grained features of exercises and avoid single exercise noises, inspired by math word solving technique~\cite{zhang2019gap,fuchs2004enhancing}, we introduce the concept of problem schema to summarize a group of similar exercises with similar solutions. Two exercises are indirect supported if they belong to the same problem schema. An outstanding property is that two exercises belong to the same problem schema only if the learner gives similar feedback to both exercises since their difficulty, semantics and solutions are also similar. As shown in Fig.~\ref{fig:inspiration}, if we could take those prior exercises support relation into account (i.e. $e_{2}$ direct support $e_{7}$,$e_{3}$ indirect support $e_{8} $), the diagnosis results will be different. Besides, by incorporating the features of problem schemas, we also propose the K\&S diagnosis method, which could not only deduce the mastery of each knowledge but also the mastery of each problem schema (as shown in Fig.~\ref{fig:first_concept}).
\fi
\iffalse
The above analysis shows the prospects of introducing prior exercises support relations into KT task. However, it may also bring in the following three problems. Firstly, the definition of the direct support and indirect support relations require lots of expert time, and manual labeled mappings are hard to generalize to other datasets. Secondly, it remains unexplored that what is the most effective way to map an exercise to a problem schema. Thirdly, the relationship between problem schemas and knowledge is complex. In particular, as Fig.~\ref{fig:first_concept} shows, two exercises (e.g., $e_{3}$ and $e_{4}$) with the same problem schema may contain different knowledge since they may contain multiple knowledge distribution. Here, we assume each exercise contains a distribution of knowledge while it has one major knowledge like most knowledge tracing methods.Similarly, two exercises (e.g., $e_{1}$ and $e_{2}$) with the same knowledge can also belong to different problem schemas due to their difficulty difference.
\fi
The above analysis shows the prospects of introducing prior exercises support relations into a KT task. However, it may also bring in the following problems. First, the direct support relations between exercises can be defined in multiple ways, but which is the most suitable one for KT tasks remains unknown. Second, the definitions of problem schema and indirect support relations require understanding of exercises from a semantic perspective. How to automatically understand and represent the information is still challenging. Third, the hierarchical exercise relations contain different levels of features of exercises, how to organically combine these different levels of features is still worth exploring. Last, after we encode the information of hierarchical relations, we also hope that the model can always draw on past key information during current prediction. As shown in Fig.~\ref{fig:inspiration}, during making predictions for $e_{7}$, our model needs easily look back to important historical information such as the learner's wrong answer in $e_{2}$ or the relation between $e_{2}$ and $e_{7}$.
\iffalse
However, it may also bring in the following three problems. Firstly, the definition of the direct support and indirect support relations require lots of expert time, and manual labeled mappings are hard to generalize to other datasets. Secondly, it remains unexplored that what is the most effective way to map an exercise to a problem schema. Thirdly, the relationship between problem schemas and knowledge is complex. In particular, as Fig.~\ref{fig:first_concept} shows, two exercises (e.g., $e_{3}$ and $e_{4}$) with the same problem schema may contain different knowledge since they may contain multiple knowledge distribution. Here, we assume each exercise contains a distribution of knowledge while it has one major knowledge like most knowledge tracing methods.Similarly, two exercises (e.g., $e_{1}$ and $e_{2}$) with the same knowledge can also belong to different problem schemas due to their difficulty difference.
\fi
\begin{figure}[htb]
\centering
\includegraphics[width=\linewidth]{figures/schema_know_diagnosis_re.pdf}
\caption{The left image shows the relations between knowledge and problem schema while the right image shows the diagnosis results of K\&S diagnosis matrix at time $\emph{t}$}
\label{fig:relation}
\end{figure}
To address above challenges, we propose a hierarchical graph knowledge tracing framework named HGKT, which unifies the strengths of hierarchical graph neural network and recurrent sequence model with attention to enhance the performance of knowledge tracing. To summarize, the main contributions of this paper are as follows:
\iffalse
In HGKT, we present several data-driven methods to model direct support relations, and also introduce a way to semantically model the indirect support relations. Besides, we build a hierarchical exercise graph (HEG) consisting above two relations and leverage attention mechanisms to improve the accuracy of the KT task. To make a more detailed and convincing diagnosis results, a To summarize, the main contributions of this paper are as follows:
\fi
\begin{itemize}
\item We introduce the concept of hierarchical exercise graph consisting of direct and indirect support relations between exercises, which could serve as learning constraints for a knowledge tracing task. We also present several data-driven methods to model direct support relations and introduce a way to semantically model the indirect support relations.
\item We propose the concept of problem schema for exercises and explore a novel approach called hierarchical graph neural network to learn a refined representation for problem schema. This approach could help to address the exercise representation loss problem.
\item We present two attention mechanisms that can highlight the important states of learners and fully leverage the information learned in HEG.
\item To make the diagnosis results detailed and convincing, we put forward a knowledge\&schema (K\&S) diagnosis matrix that can trace the mastery of both knowledge and problem schema at the same time (as shown in Fig.~\ref{fig:relation}(b)), which can also contribute to solving the insufficient diagnosis problem.
\end{itemize}
\begin{figure*}[htb]
\centering
\includegraphics[scale=0.415]{diagram/framework_wz1.pdf}
\caption{Overview of HGKT implementation. When predicting the performance of a leaner, System 1 generates problem schema embeddings fusing exercise hierarchical relations by HGNN and feed this information to System 2. System 2 incorporating graph and sequential information and updates learner's knowledge mastery at each time.}
\label{fig:HGKT_framework}
\end{figure*}
\section{Problem Definition}
We formulate the KT task in this paper as follows. The goal of knowledge tracing is to estimate learners’ hidden knowledge states based on learners’ exercise data. Suppose we have a learner set $\mathcal P$, an exercise set $\mathcal E$, a knowledge set $\mathcal K$ and a problem schema set $\mathcal S$. Each exercise associates with a knowledge $k \in \mathcal K$ and a problem schema $s \in \mathcal S$. We further record exercise sequence for each learner $\emph{l}$ as $\mathcal R_{l} =\{(e_{1},r_{1}),(e_{2},r_{2})…(e_{m},r_{m})\}$, where $e_{i} \in \mathcal E$ and $r_{i}\in \{0,1\}$. Here $r_{i}=1$ indicates a correct answer and $r_{i}=0$ indicates an incorrect one. To clearly explain the framework of HGKT, we first briefly introduce the definition of HEG.
\subsubsection{Hierarchical Exercise Graph (HEG)} In an educational system, HEG is a hierarchical graph consisting of different levels of graphs that store prior direct and indirect support relations between exercises. Specifically, the HEG used in HGKT has two graphs : the bottom one is called direct support graph while the top one is called indirect support graph. Each node in the bottom graph corresponds to an exercise in our database while each node in top graph corresponds to a problem schema. The graph links in the bottom graph model the direct support relations while the relations between those two graphs model the indirect support relations. Formally, we represent a HEG graph as $(A,F,S_{e})$, where $A \in \{0,1\}^{\left|\mathcal E\right|*\left|\mathcal E\right|}$ is an adjacency matrix of the bottom graph, $F \in \mathbb{R}^{\left|\mathcal E\right|*t}$ is the node feature matrix assuming each node has $\emph{t}$ features and $S_{e} \in \mathbb{R}^{\left|\mathcal E\right|*\left|\mathcal S\right|}$ is an assignment matrix which denotes an assignment of each node from the bottom graph to a node in the top graph.
\iffalse
\newtheorem{ddd1}{Knowledge Tracing Problem}
\begin{definition}
\textbf{Knowledge Tracing Task (KT Task)}. In an intelligent learning system, we assume that each learner contains a hidden temporal knowledge state for each concept at a time step, which denotes the mastery of each concept. Given an exercise interaction sequence for a learner $\emph{s}$ before time $\emph{t}$, the goal of a KT task is to find a method (i.e., train a model) to predict the learner's answer $r_{t+1}$ on the exercise $q_{t+1}$ at time $\emph{t+1}$. After the model has been trained, we could use it to infer the learner's hidden knowledge state at any time.
\end{definition}
\fi
\section{HGKT Framework}
\subsection{Framework Overview}
The entire structure of the framework is shown in Fig.~\ref{fig:HGKT_framework}. Here, System 1 aims to learn hierarchical graph structure of exercises by a hierarchical graph neural network (HGNN) and provides those resources to System 2. System 2 then conducts the sequential processing, makes performance prediction and collects clues for System 1 to better refine the embeddings of each problem schema. To start training, we first need to generate the HEG from exercise database and interaction history. Thus, we introduce several methods to construct the direct support graph in section~\ref{section:s4.2} and to extract indirect support relations from a semantic perspective in section~\ref{section:psrl}. After the construction of HEG, exercising records and the problem schema embeddings are combined in a recurrent neural network (RNN) to predict the learner's performance. It is worth mentioning that HGNN and RNN are trained in an end-to-end fashion.
\iffalse
First of all, many successes have demonstrated BERT's ability to understand the semantic information behind the text recently~\cite{Bert}, thus it is used to encode all exercises in the database to obtain the semantic representation. Next, we extract hierarchical graph relations from those embedding, which are combined to generate the hierarchical exercise graph. The third step is to learn the representation of problem schema by a hierarchical graph neural network (HGNN). Finally, the problem schema embedding and exercising records are used in a recurrent neural network (RNN) to predict the learner's performance. It is worth mentioning that the HGNN and RNN are trained in an end-to-end fashion.
\fi
\subsection{Direct Support Graph Construction}\label{section:s4.2}
HGKT can leverage prior hierarchical relations among exercises for a KT task as shown in Fig.~\ref{fig:inspiration}. However, the hierarchical relations of exercises are not explicitly provided in most cases. In this section, we first introduce several methods to explore the direct support relations among exercises, which are further used for learning the representations of problem schema. Many predecessors have proven the effectiveness of introducing graph structures into a KT task ~\cite{GKT,wang2019deep}, thus we propose several similarity rules based graph structures for exercises in section \ref{section:Knowledge-based}, \ref{section:Bert-Sim}, \ref{section:Exercise Transition}. Moreover, based on Bayesian statistical inference, we also present a method which could leverage the prior exercise support relations from the exercise interaction sequences in the training set and construct the direct support graph. To modeling the prior exercise relations as constraints, we first define the following properties about exercise support relations:
\noindent{$\textbf{\emph Property\ 1}$}: We use $Sup(e_{1} \rightarrow e_{2})$ to denote the support degree of two exercises $e_{1}$ and $e_{2}$. $R_{e_{i}}$ and $W_{e_{i}}$ denote an event that whether a learner give right or wrong answer to $e_{i}$. A large value of $Sup(e_{1} \rightarrow e_{2})$ indicates the solutions of $e_{1}$ and $e_{2}$ have a strong support, which means if we know that a learner did $e_{1}$ wrong, then he has a high probability of doing $e_{2}$ wrong. In addition, if the learner is known to have done $e_{2}$ correctly, then the probability of him doing correct $e_{1}$ is also high.
\begin{align*}
P(R_{e_{1}}|R_{e_{2}}) > P(R_{e_{1}}|R_{e_{2}}, W_{e_{2}}),P(W_{e_{2}}|W_{e_{1}}) > P(W_{e_{2}}|R_{e_{1}}, W_{e_{1}}),\\
if \ Sup(e_{1} \to e_{2})>0. \numberthis \label{eq:p1}
\end{align*}
\noindent{$\textbf{\emph Property\ 2}$}: Conversely, if $Sup(e_{1} \rightarrow e_{2})$ is small, it means that there is no prior support relations between the content and the solutions of the two exercises. In other words, the performance of learners for the two exercises is two unrelated events. Thus, we could conduct the following formulas:
\begin{align*}
P(R_{e_{1}}|R_{e_{2}}) = P(R_{e_{1}}|R_{e_{2}}, W_{e_{2}}),P(W_{e_{2}}|W_{e_{1}}) = P(W_{e_{2}}|R_{e_{1}}, W_{e_{1}}),\\
if \ Sup(e_{1} \to e_{2})=0. \numberthis \label{eq:p2}
\end{align*}
\iffalse
\begin{equation}
P(S{e_{j}} =0|\delta{(e_{i})=0}) > P(\delta{(e_{i})=1}),
if \ Sup(e_{1} \textrightarrow e_{2}) > 0,
\end{equation}
\fi
Based on the above reasoning, we construct the value of support between exercises as follows. Here, $Count((e_{i},e_{j})=(r_{i},r_{j}))$ counts the number of exercise sequences that reply $e_{i}$ with an answer $r_{i}$ before $e_{j}$ with an answer $r_{j}$. Besides, to prevent the denominator from being too small, we introduced the laplacian smoothing parameter $\lambda_{p} = 0.01$~\cite{field1988laplacian} in Eq.(\ref{eq:p4}) and Eq.(\ref{eq:p41}).
\begin{equation}
P(R_{e_{1}}|R_{e_{2}}) = \frac{Count((e_{2},e_{1})=(1,1)) + \lambda_{p}}{\sum_{r_{1}=0}^{1}Count((e_{2},e_{1})=(1,r_{1})) + \lambda_{p}}, \numberthis \label{eq:p4}
\end{equation}
\begin{equation}
P(R_{e_{1}}|R_{e_{2}}, W_{e_{2}}) =\frac{\sum_{r_{2}=0}^{1}Count((e_{2},e_{1})=(r_{2},1)) + \lambda_{p}}{\sum_{r_{2}=0}^{1}{\sum_{r_{1}=0}^{1}}Count((e_{2},e_{1})=(r_{2},r_{1})) + \lambda_{p}}. \numberthis \label{eq:p41}
\end{equation}
Similarly, we could also estimate the probability of $P(W_{e_{2}}|W_{e_{1}})$ and $P(W_{e_{2}}|R_{e_{1}}, W_{e_{1}})$. The value of support is defined as the sum of the following two components. Here, $max $ function is used to ensure the non-negativity of support value.
\iffalse
\begin{equation}
\frac{P(R_{e_{1}}|R_{e_{2}})}{P(R_{e_{1}})} = \frac{Count((e_{2},e_{1})=(1,1)) + \lambda}{\sum_{r_{1} \in \{0,1\}}Count((e_{2},e_{1})=(1,r_{1})) + \lambda} \\
, \numberthis \label{eq:p4}
\end{equation}
\begin{equation}
\frac{P(W_{e_{2}}|W_{e_{1}})}{P(W_{e_{2}})} = \frac{Count((e_{1},e_{2})=(0,0)) + \lambda}{\sum_{r_{2} \in \{0,1\}}Count((e_{1},e_{2})=(0,r_{2})) + \lambda} \\
, \numberthis \label{eq:p5}
\end{equation}
\fi
\begin{align*}
Sup(e_{1} \to e_{2}) = max(0, \ln{\frac{P(R_{e_{1}}|R_{e_{2}})}{P(R_{e_{1}}|R_{e_{2}},W_{e_{2}})}}) \ \\
+\ max(0,\ln{\frac{P(W_{e_{2}}|W_{e_{1}})}{P(W_{e_{2}}|R_{e_{1}}, W_{e_{1}})}}).
\numberthis \label{eq:p3}
\end{align*}
\subsubsection{Knowledge-based Method}\label{section:Knowledge-based} generates a densely connected graph, where $A_{i,j}$ is \emph{1} if two different exercises $e_{i}$ and $e_{j}$ contain same knowledge; else
it is \emph{0}.
\subsubsection{Bert-Sim Method}\label{section:Bert-Sim} generates a graph by the cosine similarity of two exercises' BERT embedding. Here $A_{i,j}$ is \emph{1} if the similarity between two different exercises is larger than hyperparameter $\omega$; else, it is \emph{0}.
\subsubsection{Exercise Transition Method}\label{section:Exercise Transition} generates a graph whose adjacency matrix is a transition probability matrix, where $A_{i,j}$ is \emph{1} if $\frac{n_{i,j}}{\sum_{k=1}^{\left|\mathcal E\right|}n_{i,k}} > \omega$; else, it is \emph{0}. Here $n_{i,j}$ represents the number of times exercise $\emph{j}$ is answered immediately after exercise $\emph{i}$ is answered.
\subsubsection{Exercise Support Method}\label{section:Exercise Support} generates a graph
by bayesian statistical inference, where $A_{i,j}$ is \emph{1} if $Sup(e_{i},e_{j})
> \omega$; else, it is \emph{0}.
\subsection{Problem Schema Representation Learning} \label{section:psrl}
In this section, we first describe methods to explore the indirect support relations between exercises. The hierarchical relations extracted are used to compose the HEG. The goal of System 1 is to learn representations of problem schema for each exercise, thus we also propose a way to fuse those hierarchical relations.
The essence of mining indirect support relations is to find the corresponding problem schema for each exercise (as shown in Fig.~\ref{fig:inspiration}), which can be transformed into a problem of unsupervised clustering of exercises. Considering the semantic nature of problem schema, we use BERT~\cite{Bert} to encode all exercises in the database to obtain its semantic representation as many successes have demonstrated BERT's ability to understand the semantic information behind texts. Besides, to better obtain the clustering results with multiple levels to accommodate the cognitive nature of students at different levels, we adopt the hierarchical clustering~\cite{johnson1967hierarchical} to cluster the BERT embeddings of exercises. Hierarchical clustering is an unsupervised cluster analysis method using agglomerative or divisive strategies to build a hierarchy of clusters. We could set different clustering threshold $\lambda$ to get different levels of clustering results, which could be used to find the most suitable level for problem schema. Moreover, in order to better combine the graph structures and the clustering results of the exercises, motivated by the assignment matrix proposed in DiffPool~\cite{ying2018hierarchical}, we propose an exercise assignment matrix $S_{e}$ that can provide an assignment of each exercise node in the direct support graph to a problem schema node in the indirect support graph. Here, each row of the $S_{e}$ corresponds to one of exercise at bottom graph, and each column of $S_{e}$ corresponds to one of problem schema at top graph. It is worth mentioning that the assignment matrix in DiffPool is learned by a separate GNN model which is computational expensive and hard to control hierarchical relations in our cases. Thus, we use hierarchical clustering to generate the assignment matrix to store indirect support relations.
After extracting the hierarchical relations from exercises, we now introduce the detailed strategie of fusing graph information in HEG. Here we propose a hierarchical graph neural networks (HGNN) to propagate semantic features of exercises into problem schema embedding in the HEG (as shown in Fig.~\ref{fig:HGKT_architectures}). The HGNN consists of convolutional layers and pooling layers~\cite{kipf2016semi,lee2019self,scarselli2008graph}. The key intuition of HGNN is that we stack a direct support graph with exercise information and a indirect support graph with problem schema information, and leverage assignment matrix to assign each node from direct support graph to indirect support graph. Formally, for a given HEG = $(A,F,S_{\lambda})$, the HGNN propagates features using following equations. First of all, we build two GNN networks named $GNN_{exer}$ and $GNN_{sche}$ whose parameters are used to update exercise embedding and problem schema embedding correspondingly. The node feature matrix for the first $\emph{k}$ layers is one-hot embedding of each exercise $F$ and a direct support graph $A_{e}$. Here $H_{e}$ and $H_{s}$ correspond to the embedding of exercises and problem schemas. Note that $H^{0}=F$. At $\emph{k-th}$ layer, as shown in Eq.~\ref{eq:aa}, we utilize a pool operation to coarsen the direct support graph to get a smaller indirect support graph $A_{s}$. The linear transformation introduced in Eq.~\ref{eq:sh} aggregates the exercise representations to correspond problem schema representations. At last, $GNN_{sche}$ updates each embedding of problem schema and sends this information to sequence processing stage of HGKT.
\iffalse
First of all, many successes have demonstrated BERT's ability to understand the semantic information behind the text recently~\cite{Bert}, thus it is used to encode all exercises in the database to obtain the semantic representation. Next, we extract hierarchical graph relations from those embedding, which are combined to generate the hierarchical exercise graph. The third step is to learn the representation of problem schema by a hierarchical graph neural network (HGNN). Finally, the problem schema embedding and exercising records are used in a recurrent neural network (RNN) to predict the learner's performance. It is worth mentioning that the HGNN and RNN are trained in an end-to-end fashion.
Considering the hierarchical nature of the exercise data, we propose a hierarchical graph neural networks (HGNN) to propagate semantic features of exercises into problem schema embedding in the HEG (as shown in Fig.~\ref{fig:HGKT_architectures}). The HGNN consists of convolutional layers and pooling layers~\cite{kipf2016semi,lee2019self,scarselli2008graph}. The key intuition of HGNN is that we stack a layer with exercise information and a layer with problem schema information and learn a method to assign each node from the exercise layer to the problem schema layer.
Motivated by assignment matrix proposed in DiffPool~\cite{ying2018hierarchical} which has been proved to be an efficient way to capture the hierarchical community structure of a graph, we propose an exercise assignment matrix $S_{e}$ that can provide an assignment of each exercise node in the bottom layer to a problem schema node in the top layer. Here, each row of an $S_{e}$ corresponds to one of exercises at the bottom layer, and each column of $S_{e}$ corresponds to one of problem schemas at the top layer. The assignment matrix in DiffPool is learned by a separate GNN model which is computational expensive and hard to control its hierarchical relation. Thus, in our paper, we adopt hierarchical clustering~\cite{johnson1967hierarchical} to explore the latent structure of exercise and then generate an assignment matrix based on its clustering results. Specifically, for each clustering threshold $\lambda$, we could generate a clustering assignment matrix $S_{\lambda}$ where $S_{\lambda(i,j)}$ is 1 if exercise $\emph{i}$ belong to cluster $\emph{j}$, else it is 0.
Now we formally introduce structures of HGNN. For a given HEG = $(A,F,S_{\lambda})$, the HGNN propagates features using following equations. First of all, we build two GNN networks named $GNN_{exer}$ and $GNN_{sche}$ whose parameters are used to update exercise embedding and problem schema embedding correspondingly. The node feature matrix for the first $\emph{k}$ layers is one-hot embedding of each exercise $F$ and an exercise graph $A_{e}$. Here $H_{e}$ and $H_{s}$ correspond to the embedding of exercises and problem schemas. Note that $H^{0}=F$. At $\emph{k-th}$ layer, we coarsen the input exercise graph to get a smaller problem schema graph $A_{s}$ using Eq.~\ref{eq:aa}. The linear transformation introduced in Eq.~\ref{eq:sh} aggregates the exercise representations to correspond problem schema representations. At last, $GNN_{sche}$ updates each embedding of problem schema and sends this information to sequence processing stage of HGKT.
\fi
\begin{equation}
H_{e}^{(l+1)} = GNN_{exer}(A_{e},H_{e}^{(l)}), l<k,
\label{eq:exer}
\end{equation}
\begin{equation}
A_{s} = S_{\lambda}^\mathrm{T}A_{e}S_{\lambda}, l=k,
\label{eq:aa}
\end{equation}
\begin{equation}
H_{s}^{(l+1)} = S_{\lambda}^\mathrm{T}H_{e}^{(l)}, l=k,
\label{eq:sh}
\end{equation}
\begin{equation}
H_{s}^{(l+1)} = GNN_{sche}(A_{s},H_{s}^{(l)}), l>k.
\label{eq:sche}
\end{equation}
\subsection{Sequence Modeling Process}
This section mainly focuses on combining the representations of problem schemas
with sequence information.
\begin{figure}[htb]
\centering
\includegraphics[width=1.0\linewidth]{diagram/architecture1.pdf}
\caption{The architecture of HGKT}
\label{fig:HGKT_architectures}
\end{figure}
\subsubsection{\textbf{Sequence Propagation}} The overall input for the sequence processing part is exercise interaction sequences. Each exercise interaction contains three components : knowledge $v_{t}$, exercise result $r_{t}$, problem schema $s_{t}$. Here, $v_{t}$ is a one-hot embedding of $\left|\mathcal K\right|$ distinct knowledge. $r_{t}$ is a binary value indicating whether the learner answers an exercise correctly. $s_{t}$ is an embedding of problem schema generated in HEG. At each time step $\emph{t}$, to differentiate the exercise contribution to its corresponding knowledge hidden state, the input for the sequence model is a joint embedding $x_{t}$ of $(v_{t},r_{t},s_{t})$. In the propagation stage, we process $x_{t+1}$ and the previous hidden state $h_{t}$ using RNN network to get current learner's hidden state $h_{t+1}$ as shown in Eq.~\ref{lstm}. Here we use LSTM as a variant of RNN since it can better preserve long-term dependency in the exercise sequences~\cite{memeory}. Eq.~\ref{cur} shows the prediction of mastery for each problem schema $m_{t+1}^{cur} \in \mathbb{R}^{\left|\mathcal S\right|}$ at time \emph{t+1}. $\{W_{1},b_{1}\}$ are the parameters.
\begin{equation}{\label{lstm}}
h_{t+1},c_{t+1} = LSTM(x_{t+1},h_{t},c_{t};\theta_{t+1}),
\end{equation}
\begin{equation}{\label{cur}}
m_{t+1}^{cur}= ReLU(W_{1} \cdot h_{t+1}+b_{1}).
\end{equation}
\subsubsection{\textbf{Attention Mechanism}} The HGKT leverages two kinds of attention mechanism, called sequence attention and schema attention, to enhance effects of typical states in the history of modeling. Learners may perform similarly on exercises with same problem schemas, thus we use a sequence attention shown in Eq.~\ref{seqa} to refer to previous results of similar exercises. Here, we assume the attention of historical problem schema mastery $m_{t+1}^{att}$ is a weighted sum aggregation of previous mastery state. However, different from the attention used in ~\cite{EKT}, we set an attention window limit $\lambda_{\beta}$ in HGKT for the following two reasons: (1) If the length of sequence attention is not limited, the computational cost would be intensive when the exercise sequence is extremely long. (2) Experiments prove that recent memory has a greater impact on knowledge tracing results than long past memories, which is consistent with educational psychology since learners start losing the memory of learned knowledge over time~\cite{ebbinghaus2013memory}.
\begin{equation}{\label{seqa}}
m_{t+1}^{att}= \sum_{i=max(t-\lambda_{\beta},0)}^{t}\beta_{i}m_{i}^{cur},
\beta_{i}=\cos(s_{t+1},s_{i}).
\end{equation}
\iffalse
\begin{figure*}[htb]
\centering
\includegraphics[scale=0.5]{figures/test_stage}
\caption{The illustration of testing stage and K\&S diagnosis matrix}
\label{fig:test_stage}
\end{figure*}
\fi
The schema attention aims to focus learner's attention on a given problem schema with $\alpha_{t} \in \mathbb{R}^{\left|\mathcal S\right|}$, which is the similarity to other problem schemas. As shown in Eq.~\ref{eq:schea}, we propose an external memory $M_{sc} \in \mathbb{R}^{k*\left|\mathcal S\right|}$ collected from embeddings in the final layer of indirect support graph. Each column of $M_{sc}$ corresponds to an embedding of problem schema. Here, \emph{k} is the embedding dimension in HEG. The intuition for Eq.~\ref{eq:schea} is that the answer of an exercise would relate to exercises with similar problem schema, thus we could focus our attention on a certain problem schema. Notice that the value of memory $M_{sc}$ changes over time during training.
\begin{equation}
m_{t+1}^{f} = \alpha_{t+1}^{T}m_{t+1}^{cur},
\alpha_{t+1}=Softmax(s_{t+1}^TM_{sc}).
\label{eq:schea}
\end{equation}
To summarize, the states to predict the learner's performance at time \emph{t+1} consists of three components : current knowledge mastery $m_{t+1}^{cur}$, historical related knowledge mastery $m_{t+1}^{att}$ and focus problem schema mastery $m_{t+1}^{f}$. As Eq.~\ref{eq:pred} shows, those states are concatenated together to predict a final result. $\{W_{2},b_{2}\}$ are the parameters.
\begin{equation}
\widetilde{r_{t+1}}= \sigma(W_{2} \cdot
[m_{t+1}^{att},m_{t+1}^{cur},m_{t+1}^{f}]+b_{2}).
\label{eq:pred}
\end{equation}
\subsubsection{\textbf{Model Learning}} The goal of training is the negative log likelihood of the observed sequence of learner response. During training, both parameters, i.e., parameters in $GNN_{exer}$ and $GNN_{sche}$, parameters in sequence propagation $\{W_{1},b_{1},W_{2},b_{2}\}$ are updated. The loss for a learner is shown in Eq.~\ref{eq:loss}. Specifically, the loss for a response log is defined as the cross entropy between the real answer $r_{t}$ at time $\emph{t}$ and predicted score $\widetilde{r_{t}}$ on exercise. The objective function is minimized using the Adam optimization~\cite{kingma2014adam}. More implementation details will be introduced in the experiments section.
\begin{equation}
loss =
-\sum_{t}(r_{t}log\widetilde{r_{t}}+(1-r_{t})log(1-\widetilde{r_{t}})).
\label{eq:loss}
\end{equation}
\subsection{Prediction Output of HGKT}
After modeling the exercising process of each learner from step \emph{1} to \emph{t}, we now introduce our strategies for predicting his performance on the next exercise $e_{t+1}$. Moreover, to diagnose a learner's learning process in detail, we introduce a K\&S diagnosis matrix to fully trace the mastery of both knowledge and problem schema dynamically.
\subsubsection{\textbf{K\&S Diagnosis Matrix}}
Unlike the traditional knowledge tracing method, the input for HGKT in testing stage is the knowledge and problem schema of an exercise. Thus, we could trace the transition mastery of knowledge, problem schema, or their combination in HGKT. Specifically, at each time step $\emph{t}$, we can predict the performance $r_{t}\label{key}$ for each combination of $(k_{i},s_{j})$ ($k_{i} \in \mathcal K, s_{j} \in \mathcal S$). Thus, we could use those results to generate a matrix $R_{t}^{ks}$ called K\&S diagnosis matrix whose vertical axis represents different knowledge and horizontal axis represents different problem schema. Next, the knowledge mastery $R_{t}^k$ at time $\emph{t}$ is calculated by the weighted aggregation of mastery of each problem schema as shown in Eq.~\ref{eq:d1}. Here $q_{i,j}$ denotes the number of exercises containing knowledge $k_{i}$ and problem schema $s_{j}$. Similarly, we could calculate the mastery of each problem schema $R_{t}^s$.
\begin{equation}
q_{i,j} = \left|{\{(e_{(k_{i},s_{j})}\mid k_{i}\in \mathcal K,s_{j}\in \mathcal
S)}\}\right|
\label{eq:d}
\end{equation}
\begin{equation}
R_{t,i}^k = R_{t,i}^{ks}d_{i}^{k},\ d_{i,j}^{k} =
\frac{q_{i,j}}{\sum_{j}q_{i,j}}
\label{eq:d1}
\end{equation}
\begin{equation}
R_{t,j}^s = R_{t,j}^{ks}d_{j}^{s},\ d_{i,j}^{s} =
\frac{q_{i,j}}{\sum_{i}q_{i,j}}.
\label{eq:d2}
\end{equation}
\subsubsection{\textbf{Interpretability of Problem Schema}} The introduction of problem schema can effectively improve the effect of models in predicting learners' performance. However, the interpretability of problem schema is unknown. Based on above challenges, we propose an unsupervised schema summarization algorithm. The core idea of this algorithm is to utilize TextRank~\cite{textrank} to extract meaningful condition descriptions and objective descriptions accordingly and use them to form a description for an exercise cluster based on certain rules. Table~\ref{tab:Summarization} shows an example of the summarization of a group of exercises. More details about the algorithm would be introduced in Appendix.
\section{Experiments}
In this section, we evaluate the performance of the proposed framework HGKT from following three aspects: (1) The prediction performance of HGKT against baselines. (2) The analysis of effectiveness of hierarchical exercise graph. (3) The effectiveness of sequence attention and schema attention mechanism used in HGKT.
\subsection{Experiment Setup}
\subsubsection{\textbf{Dataset and Preprocessing}} Since there is no open dataset which could provide exercising records with text information, our experimental dataset is derived from a large real-world online education system. The Aixuexi online system \footnote{www.aixuexi.com} could track exercising records of learners. Each record of a learner includes learner id, lecture id, class type, question information, answer time, learners' answers and correction results. The question information includes the question id, question content, question type and the primary knowledge contained in the question. To avoid data distribution differences, we use the data obtained after 2018. In the data preprocessing stage, we group the exercising records with learner id and sort the records according to the response time. In total, we get 91,449,914 answer records from 132,179 learners. The detailed statistics of dataset are shown in Table~\ref{tab:statistic}.
\iffalse
\begin{figure}[htb]%
\centering
\subfloat[Distribution of exercise records
]{{\includegraphics[width=3.5cm]{number_of_exercising_records}}}%
\qquad
\subfloat[Distribution of content length
]{{\includegraphics[width=3.5cm]{content_length}}}%
\caption{The distribution of Aixuexi dataset
\label{fig:distribution}%
\end{figure}
\fi
\begin{figure*}[htb]
\centering
\includegraphics[width=\linewidth]{figures/all_auc3.pdf}
\caption{The results of learner performance prediction under four metrics.}
\label{fig:all_model_auc}
\end{figure*}
\iffalse
\renewcommand{\arraystretch}{1.2}
\begin{table*}[]\tiny%
\caption{Comparison of prediction performance.}
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{clllllllll}
\toprule[0.6pt]
\multicolumn{1}{l}{} & Ratio & BKT & DKT & DKVMN &
EKTA & GKT & HGKT-B & HGKT-S & HGKT \\
\hline
\multirow{4}{*}{AUC} & 60\% & 0.8372 & 0.8372 & 0.8372
&
0.8372 & 0.8372 & 0.8372 & 0.8372 & 0.8372 \\
& 70\% & 0.8372 & 0.8372 & 0.8372 & 0.8372 & 0.8372 &
0.8372
& 0.8372 & 0.8372 \\
& 80\% & 0.8372 & 0.8372 & 0.8372 & 0.8372 & 0.8372 &
0.8372
& 0.8372 & 0.8372 \\
& 90\% & 0.8372 & 0.8372 & 0.8372 & 0.8372 & 0.8372 &
0.8372
& 0.8372 & 0.8372 \\
\hline
\multirow{4}{*}{ACC} & 60\% & 0.8372 & 0.8372 & 0.8372
&
0.8372 & 0.8372 & 0.8372 & 0.8372 & 0.8372 \\
& 70\% & 0.8372 & 0.8372 & 0.8372 & 0.8372 & 0.8372 &
\textbf{0.8372}
& 0.8372 & 0.8372 \\
& 80\% & 0.8372 & 0.8372 & 0.8372 & 0.8372 & 0.8372 &
0.8372
& 0.8372 & 0.8372 \\
& 90\% & 0.8372 & 0.8372 & 0.8372 & 0.8372 & 0.8372 &
0.8372
& 0.8372 & 0.8372 \\
\hline
\multirow{4}{*}{MAE} & 60\% & 0.8372 & 0.8372 & 0.8372
&
0.8372 & 0.8372 & 0.8372 & 0.8372 & 0.8372 \\
& 70\% & 0.8372 & 0.8372 & 0.8372 & 0.8372 & 0.8372 &
0.8372
& 0.8372 & 0.8372 \\
& 80\% & 0.8372 & 0.8372 & 0.8372 & 0.8372 & 0.8372 &
0.8372
& 0.8372 & 0.8372 \\
& 90\% & 0.8372 & 0.8372 & 0.8372 & 0.8372 & 0.8372 &
0.8372
& 0.8372 & 0.8372 \\
\hline
\multirow{4}{*}{RMSE} & 60\% & 0.8372 & 0.8372 & 0.8372
&
0.8372 & 0.8372 & 0.8372 & 0.8372 & 0.8372 \\
& 70\% & 0.8372 & 0.8372 & 0.8372 & 0.8372 & 0.8372 &
0.8372
& 0.8372 & 0.8372 \\
& 80\% & 0.8372 & 0.8372 & 0.8372 & 0.8372 & 0.8372 &
0.8372
& 0.8372 & 0.8372 \\
& 90\% & 0.8372 & 0.8372 & 0.8372 & 0.8372 & 0.8372 &
0.8372
& 0.8372 & 0.8372 \\
\bottomrule[0.8pt]
\label{tab:auc}
\end{tabular}%
}
\end{table*}
\fi
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\begin{table}[]
\caption{An Example of Schema Summarization}
\centering
\resizebox{0.45\textwidth}{!}{%
\begin{tabular}{|p{9cm}|}
\hline
\textbf{Exercise Cluster} \\ \hline
\emph{Exercise 1}: If the ratio of lengths of three sides of a triangle is
$2:3:4$, and its circumference is 18, the shortest side length is?
\\
\emph{Exercise 2}: Given ratio of lengths of triangle sides is 2:4:4 and
circumference is 20, what is the shortest side length? \\
\emph{Exercise 3}: If we know that ratio of lengths of three sides of a
triangle is 3:4:5, and circumference of the triangle is 24, find
the shortest side length? \\
... \\
\hline
\textbf{Condition keyphrases} \\
\hline
\multicolumn{1}{|c|}{ratio of lengths, circumference} \\
\hline
\textbf{Objective keyphrases} \\ \hline
\multicolumn{1}{|c|}{shortest side length} \\ \hline
\textbf{Problem Schema} \\ \hline
\multicolumn{1}{|c|}{Given ratio of lengths, circumference, find
the shortest side length?} \\ \hline
\end{tabular}%
\label{tab:Summarization}
}
\end{table}
\begin{table}
\centering
\caption{The statistics of the Aixuexi dataset.}
\label{tab:statistic}
\begin{tabular}{c|c}
\hline
Statistics & Value\\
\hline
{number of learners} & {$132,179$}\\
number of response logs & {$9,144,914$}\\
{number of exercises correctly answered} &
{$6,747,435$}\\
average of exercises in response log & {$69$}\\
number of different exercises& {$11,410$}\\
number of knowledge related& {$490$}\\
total number of knowledge & {$672$}\\
\hline
\end{tabular}
\end{table}
\iffalse
\subsubsection{\textbf{Implementation Details}} We conduct extensive experiments to find the best parameters for HGKT models. To extract the semantic representation for exercises, we utilize a public BERT-Base tool~\cite{Bert} to convert each exercise into an embedding vector with 768 dimensions. In problem schema representation learning part, we use transition graph to construct the exercise graph. Besides, the best graph structure for HEG is three graph convolutional layers followed by a graph pooling layer and a graph convolutional layer. The size of graph hidden embedding is 64, the size of problem schema embedding is 30. The most suitable clustering threshold for HEG is 9 and we get 1136 problem schemas. In the sequence propagate stage, we set the attention window size to 20.
\fi
\subsubsection{\textbf{Implementation Details}} To set up the training of HGKT, we first need to use unsupervised data to generate the HEG. We leverage the exercising interaction data in the training set to construct the direct support graph and all exercise data involved in our experiment to construct the indirect support relations. Because the construction of the former one requires the learner's exercising results while the latter one only needs the exercise information. Specifically, we utilize a public BERT-Base tool \footnote{https://github.com/hanxiao/bert-as-service} without any fine-tune to convert each exercise into an embedding vector with 768 dimensions, and the hierarchical clustering tool \footnote{https://docs.scipy.org/doc/scipy/reference/cluster.hierarchy.html} during the indirect support relations construction process. The framework of HGKT is joint training after the construction of HEG. During the testing stage, we use the same HEG to make predictions.
We conduct extensive experiments to find the best parameters for HGKT. During HEG construction process, we use the exercise support method to construct a direct support graph and set the clustering threshold to 9 then get 1136 problem schemas to construct the indirect support graph. In HGNN, we get best results with three graph convolutional layers in $GNN_{exer}$ and one graph convolutional layer in $GNN_{sche}$. The exercise embedding size in HGNN is set to 64 and schema embedding is set to 30. In the sequence propagate stage, we set the attention window size to 20 and LSTM hidden embedding size to 200. During training, we use Adam optimizer with the learning rate 0.01 and set mini-batches as 32. We also use dropout \cite{srivastava2014dropout} with probability 0.5 to prevent overfitting.
\iffalse
We will publish our code after the acceptance of our paper.
\fi
\subsubsection{\textbf{Comparison Baselines}} \label{sec:baselines} To demonstrate the effectiveness of our framework, we compare HGKT with following state-of-art methods. Those methods are chosen from three aspects: (1) Traditional educational models: Bayesian Knowledge Tracing (BKT) (2) Deep learning models: Deep Knowledge Tracing (DKT), Dynamic Key-Value Memory Networks (DKVMN), Exercise-aware Knowledge Tracing (EKT), Graph Knowledge Tracing (GKT) (3) Variants of HGKT : HGKT-B, HGKT-S
\begin{itemize}
\item \textbf{BKT} is a traditional knowledge tracing model which is based
on
hidden Markov model. The knowledge states for each concept are a set of
binary variables~\cite{BKT}.
\item \textbf{DKT} uses a recurrent neural network
to model the learners'exercising process~\cite{DKT}. It models
question sequences as knowledge sequences. We follow the hyperparameters
mentioned in ~\cite{DKT}, where the size of the hidden layer is 200 and we
use a GRU for the RNN.
\item \textbf{DKVMN} is a deep learning
method using a key matrix to store knowledge
representation and a value matrix for each learner to
update the mastery for each concept~\cite{DKVMN}. In the experiment,
we get best results when we set the size of the memory slot is 30, the
size of embedding of key matrix is 50 and the size of embedding of value
matrix is 200.
\item \textbf{GKT} is a GNN-based knowledge tracing method which
reformulate
the knowledge tracing task as a time-series node-level classification
problem in the GNN~\cite{GKT}. In experiments, we get the best results
using the
knowledge graph generated by transition-graph method.
\item \textbf{EKTA} uses a bidirectional
LSTM to encode exercise text and leverage an attention mechanism to enhance
the prediction accuracy~\cite{EKT}. In experiments, we get best results when
the shape of its hidden state is 672 and 30.
\item \textbf{HGKT-B} is a variant of HGKT framework. Here, in the modeling
process, we do not use problem schema embedding generated by HEG. Instead,
we use BERT to encode text features.
\item \textbf{HGKT-S} is another variant of HGKT framework. In the modeling
process, we use the
one-hot of problem schema generated by hierarchical clustering to replace
embedding of problem schema generated by HEG.
\end{itemize}
\subsubsection{\textbf{Evaluation Setting}} We first compare the overall performance of HGKT with baseline models. In the experiment, we randomly select $60\%$,$70\%$,$80\%$,$90\%$ of the exercise records as training data, and the remaining as testing data. To qualify the effectiveness of the model, we compare the results of experiments from both regression and classification perspectives~\cite{fogarty2005case,kuang2018stable,wu2015cognitive}. We use open-source platform to implement most comparison baseline models and searched the hyperparameters to find the best performance for each model. Our framework for HGKT is implemented by Torch on a Linux server with eight Intel Xeon Skylake 6133 (2.5 GHz) CPUs and four Tesla V100 GPUs. For all models, we test different parameters and experiments five times, respectively, and finally take the average of metrics as evaluation metrics.
\subsection{Comparison}
Fig.~\ref{fig:all_model_auc} shows the overall comparing results on this task. The results indicate that HGKT performs better than other baseline models. Besides, we draw several conclusions from the results: (1) All our proposed HGKT based models perform better than other baseline models. The result clearly demonstrates that HGKT framework could make full use of the information contained in exercising records and exercise texts, benefiting the prediction performance. (2) The result of HGKT-B outperforms EKTA, which shows that BERT could better extract the representation of text than a Bi-LSTM neural network. (3) The HGKT-S has better performance than HGKT-B, which proves that compared to directly using the text embedding of each exercise, summarizing a bunch of similar exercises with problem schemas would bring in less noise. (4) The HGKT performs better than HGKT-S, which reveals that the introduction of HEG for problem schema representation learning is also effective. The reason may be that in HEG, the representation of each exercise is propagated through a HGNN, so that problem schema embedding could aggregate the information of a certain kind of exercises.
\subsection{Analysis}
To understand the effects of the various components and hyperparameters in HGKT, we conduct ablation studies and parameter analysis on the Aixuexi dataset.
\subsubsection{\textbf{Ablation Study}} We investigate the effects of different support relations and attention mechanisms in our model. The results of ablation study are shown in Table \ref{tab:ablation_study}. To conduct a fair comparison experiment, we propose methods that have the same framework but have different components with HGKT. The first method in Table \ref{tab:ablation_study} adopts the same framework with HGKT but has no problem schema embeddings learning part. The second method does not contain hierarchical information of exercises and only leverages the bottom layer of HEG to learn the embeddings of each exercise. The third method only uses the information of indirect support relations, which is the one-hot of hierarchical clustering results with a dense layer to learn embeddings of each problem schema. The fourth method drops the hierarchical graph neural network part and directly merges the information learned by previous two methods. All of above methods have no attention component and those learned embeddings are used to replace the embeddings of problem schemas in HGKT framework. The fifth method only lacks the attention components compared to the HGKT. From the result, we observe that HGKT with both support relations and attention mechanisms performs the best. We also note that having direct support relations and indirect support relations still outperform the implementation without either relations. More interestingly, we also find that merging two support relations directly performs worse than using hierarchical graph neural network. We postulate that a possible reason for the inferior performance may be due to the extra noise introduced by single exercise. The HGNN part in HGKT acts as a coarse-grained structure extractor which also averages the features of a group of similar exercises, thus reduces the noise of exercises.
\begin{table}[]
\caption{Model performance over different graph structure in HGKT. $\emph{GS}$ denotes the structure of graph. $B-i\_T-j$ indicates \emph{i} graph convolutional layers in $GNN_{exer}$ and \emph{j} graph convolutional layers in $GNN_{sche}$.}\label{tab:graph_number}
\begin{tabular}{l|cc|l|cc}
\hline
GS & AUC (\%) & ACC (\%) & GS & AUC (\%) & ACC (\%) \\ \hline
B-1\_T-1 & 81.4 & 81.5 & B-2\_T-2 & 82.5 & 82.4 \\
B-1\_T-2 & 82.3 & 82.4 & B-2\_T-3 & 81.7 & 81.6 \\
B-1\_T-3 & 81.8 & 81.9 & B-3\_T-1 & \textBF{82.8} & \textBF{82.6} \\
B-2\_T-1 & 82.4 & 82.4 & B-3\_T-2 & 82.4 & 82.4 \\ \hline
\end{tabular}
\end{table}
\begin{table}[]
\caption{Ablation study of HGKT}\label{tab:ablation_study}
\begin{tabular}{cc|cc}
\hline
methods & attention & AUC (\%) & ACC (\%) \\ \hline
no support relations & no & 78.8 & 78.9 \\
only direct support & no & 79.3 & 79.5 \\
only indirect support & no & 80.4 & 80.2 \\
merge two support relations & no & 80.3 & 80.5 \\
both support relations & no & 81.2 & 81.1 \\ \hline
(HGKT) both support relations & yes & \textBF{82.8} & \textBF{82.6} \\ \hline
\end{tabular}
\end{table}
\subsubsection{\textbf{Graph Structure Analysis}} In HGKT, there are three factors that may affect the structure of hierarchical exercise graph: different direct support graph construction methods, the problem schema clustering level and the number of layers used in HGNN.
In section~\ref{section:s4.2}, we propose several methods to explore different direct support relations of exercises. Those methods decide different graph structures of direct support graph in HEG. Fig.~\ref{fig:graph_auc}(a) shows the knowledge tracing AUC of different methods. From the figure, we find that bayesian-based graph construction method outperforms other methods. A potential reason could be due to that the bayesian-based method could employ the information in exercise interactions history rather just use the exercise features like other methods. We also test impacts of different edge-to-node ratio for different Graph (as shown in Fig.~\ref{fig:graph_auc}(b)). The result shows the effect of graph convolution is the best when the ratio of edge-to-node is about \emph{3}-\emph{4}.
The problem schema clustering level affects the number of nodes in an indirect support graph. As shown in Fig.~\ref{fig:level_auc}(a), the number of problem schemas is more than \emph{3430} when the threshold is \emph{5}, which results in intensive computational cost, and the generated problem schemas are not typical enough to represent a group of exercises. Thus, we test the overall performance of HGKT with thresholds between \emph{5} to \emph{20}. Fig.~\ref{fig:level_auc}(b) shows that the best AUC occurs when clustering threshold is 9 and the number of problem schema is 1136. Besides, from the overall trend of the curve, we could infer that there is a most suitable division of exercises for KT task, which means that using problem schema to represent a group of similar exercises is reasonable.
The numbers of graph convolutional layers used in the $GNN_{exer}$ and $GNN_{sche}$ are also tuneable hyperparameters. Thus, we investigate the effect of the numbers of graph convolutional layers on our model’s performance. Table \ref{tab:graph_number} shows the result of experiments. We find that $B-3\_T-1$ version achieves the best result for the KT task. The reason may be that the $B-3\_T-1$ structure could achieve the optimal capacity of information aggregation.
\begin{figure}[htb]%
\centering
\subfloat[AUC between Graph
Methods]{{\includegraphics[width=3.5cm]{figures/graph_compare} }}%
\qquad
\subfloat[AUC between edge-to-node
Ratio]{{\includegraphics[width=3.5cm]{figures/graph_length_sigir} }}%
\caption{The effectiveness of different graph methods
\label{fig:graph_auc}%
\end{figure}
\begin{figure}[htb]%
\centering
\subfloat[The Statistics of Number of Problem Schema
]{{\includegraphics[width=3.5cm]{figures/exercise_num1}}}%
\qquad
\subfloat[AUC between Different Threshold
]{{\includegraphics[width=3.5cm]{figures/level_test1}}}%
\caption{Performance over different clustering values in HGKT
\label{fig:level_auc}%
\end{figure}
\begin{figure}[!ht]
\centering
\subfloat[AUC between different Attention Methods
]{{\includegraphics[width=3.5cm]{figures/attention_compare}}}%
\qquad
\subfloat[AUC between different Attention Window
Size]{{\includegraphics[width=3.5cm]{figures/window_test1}}}%
\caption{The effectiveness of attention
\label{fig:attention_auc}%
\end{figure}
\subsubsection{\textbf{Effectiveness of Attention}} As we have clarified in Attention Mechanism section, the HGKT with attention could enhance the memory ability of the networks. The model could observe following information to predict the performance of a learner for the exercise: the performance of similar exercises in the past, learner's current knowledge state, learner's current focus. To confirm the effectiveness of those two attention mechanisms, we set up the following two experiments. Firstly, we compare the HGKT with the following three models: HGKT-noa, HGKT-sche and HGKT-seq. HGKT-noa is a no-attention version of HGKT. Similarly, HGKT-sche and HGKT-seq are variants of HGKT which only contain schema attention and sequence attention respectively. Fig.~\ref{fig:attention_auc}(a) shows the comparison results of them. From the figure, we could infer that HGKT outperforms other comparison methods. Besides, both HGKT-sche and HGKT-seq perform better than HGKT-noa, which proves that both attention mechanisms could contribute to the HGKT model. Moreover, we also conduct experiments to see the impacts of different window size of sequence attention. In the experiment, we set length of window sizes from 10 to 50 to see which size could better preserve the memory useful for KT. As Fig.~\ref{fig:attention_auc}(b) shows, the best parameter for window size is 20, which reveals that the results of nearly 20 exercises can best reflect their current learning status. This value provides a great reference for setting the number of exercises in learners' proficiency test.
\begin{figure}[!ht]
\centering
\includegraphics[width=8cm]{figures/case_study_all.pdf}
\caption{Each row of the diagnosis matrix corresponds to a related knowledge and each column corresponds to a related problem schema.}
\label{fig:application}
\end{figure}
\iffalse
\renewcommand{\arraystretch}{1.2}
\begin{table*}[]
\caption{P-value Comparison}
\centering
\resizebox{\textwidth}{!}{%
\begin{tabular}{p{2.5cm}|llllllllll}
\hline
Binary Threshold & 0.1 & 0.2 & 0.3 &
0.4 &
0.5 & 0.6 & 0.7 & 0.8 & 0.9 \\
\hline
With-PS method & 1.18E-164 & 5.62E-107 & 4.63E-53 &
3.45E-15 &
0.97 & 2.46E-16 & 1.44E-61 & 3.63E-141 & 8.94E-273 \\
No-PS method & 7.47E-118 & 1.91E-104 & 1.22E-87 &
1.53E-67 &
2.94E-40 & 3.92E-20 & 4.60E-03 & 1.35E-12 & 8.51E-126 \\
\hline
\end{tabular}%
}
\label{p_compare}
\end{table*}
\fi
\iffalse
\begin{figure}[htb]
\centering
\includegraphics[width=0.8\linewidth]{figures/scatter_all.eps}
\caption{The distribution comparison between With-PS method and No-PS method}
\label{fig:dis_compare}
\end{figure}
\fi
\section{Case Study}
The introduction of problem schema and HEG can make diagnosis more accurate and improve the interpretability of diagnosis results. Moreover, the diagnosis results generated by HGKT could be easily applied to the following two applications: (1) Learner Diagnosis Report~\cite{ragnemalm1995student,wang2019interpretable} (2) Adaptive Question Recommendation~\cite{chen2018recommendation,liu2019exploiting}. To make deep analysis about above statements, we visualize the predicted knowledge mastery of a learner during the exercising process in Fig.~\ref{fig:application}.
From the figure we could more intuitively understand the advantages of the HGKT compared to traditional knowledge tracing diagnosis (TKT). When $t=\emph{8}$, the learner did $e_{12}$ wrong, the TKT diagnosis result shows that the color of the knowledge related to the $k_{7}$ (e.g., $k_{29}$) became darker. However, the reasons for the poor mastery of related knowledge remain unknown. In HGKT, it can be found that the transition of learners' knowledge mastery is strongly correlated with the local HEG diagram in Fig.~\ref{fig:application}(b). Specifically, the learner's mistakes in $e_{12}$ led to the poor mastery of the knowledge $k_{7}$, $k_{23}$ and $k_{29}$, which demonstrates that incorporating the direct and indirect support graph do affect the results of the diagnosis. Moreover, the HEG information could also explain the change of knowledge mastery in a way, so the learner diagnosis report could be more reliable. The QA part in the figure shows the improvement of HGKT's interpretability by comparing the explanations of HGKT and TKT for several questions. By K\&S diagnosis matrix method, learners can not only know the underlying reasons for not mastering certain knowledge (e.g., $k_{1}$, $k_{1}$), but also enhance the level of mastery of the knowledge through more targeted and accurate training advice, which reveals that HGKT framework can better be applied to adaptive question recommendation scenario.
\iffalse
\begin{table}[!ht]
\centering
\resizebox{0.45\textwidth}{!}{%
\begin{tabular}{|l|l|}
\hline
\multirow{11}{*}{\begin{tabular}[c]{@{}l@{}}14 Proportion\\ 18
Addition and Subtraction Positive Decimals\\ 22 Addition
Whole
Numbers\\ 20 Addition and Subtraction Integers\\ 24
Addition and
Subtraction Fractions\\ 26 Equation Solving Two or Fewer
Steps\\ 31
Circumference\\ 44 Square Root\\ 49 Complementary and
Supplementary
Angles\\ 110 Quadratic Formula to Solve Quadratic
Equation\end{tabular}} &
\multirow{11}{*}{\begin{tabular}[c]{@{}l@{}}3 Probability of Two
\\ 5 Median\\ 7 Mode\\ 8 Mean\\ 9 Range\\ 10 Venn Diagram\\
12
Circle Graph\end{tabular}} \\
& \\
& \\
& \\
& \\
& \\
& \\
& \\
& \\
& \\
& \\ \hline
\end{tabular}%
}
\end{table}
\fi
\section{Related Work}\label{2}
\iffalse
\subsubsection{Knowledge Tracing} In the literature, there are many efforts in knowledge tracing. Existing methods can be divided into two tracks: traditional and deep learning track. The most popular traditional approach is Bayesian Knowledge Tracing (BKT)~\cite{BKT}, which updates learner's latent knowledge state by a Hidden Markov Model. Deep learning methods like Deep Knowledge Tracing (DKT) model the learning process as a recurrent neural network~\cite{DKT}. Dynamic Key-Value Memory Network (DKVMN) enhances the ability of the recurrent neural network by introducing two memory matrices to represent knowledge and learner's mastery for each knowledge respectively~\cite{DKVMN}. Graph-based Knowledge Tracing (GKT) combines knowledge tracing with a graph neural network~\cite{GKT}. It encodes a learner's hidden knowledge state as embedding of graph nodes and updates the state in a knowledge graph. Those models have been proved effective but still have limitations. Most existing methods face the problem of exercise representation loss since they have not taken the texts of exercises into consideration.
\fi
\iffalse
\subsubsection{Cognitive Diagnosis} In the domain of intelligent education, cognitive diagnosis is an essential task combining cognitive discipline and psychometric measurement. It uses computers as primary research tools to explore the cognitive processes of participants. Generally, cognitive diagnostic models (CDM)~\cite{leighton2007cognitive} could be categorized into two tracks: traditional models and deep learning models. Among them, item response theory (IRT)~\cite{IRT}, as a typical traditional diagnosis model, considers each learner as a factor and leverages a logistic-like function to model the diagnosis process. Unlike the IRT model, which locates a learner's latent trait on a continuum and does not consider the mastery difference for different knowledge, DINA (deterministic inputs, noisy “and” gate)~\cite{de2009dina} encodes each learner as a binary vector which denotes whether he has mastered the knowledge concepts or not. Compared to traditional methods, recent deep learning methods could achieve better prediction results since they could capture more complex relations between learners and exercises. In this track, researchers proposed a general Neural Cognitive Diagnosis (NeuralCD) framework~\cite{wang2019interpretable} that could automatically generate item response functions through deep learning techniques. However, the interpretability of deep learning diagnostic models still has a large gap compared to traditional cognitive diagnostic models.
\fi
\iffalse
\subsubsection{Problem Schema} The concept of problem schema was first proposed in automatic solving for mathematical word problems (MWPs) field~\cite{zhang2019gap}. The early approaches to math word problems are rule-based systems based on hand engineering~\cite{slagle1965experiments,fletcher1985understanding,bakman2007robust,yuhui2010frame}. WORD-PRO~\cite{fletcher1985understanding} solves one-step arithmetic problems by defining four problem schemas: change-in, change-out, combine and compare. ROBUST~\cite{bakman2007robust} further expands them into six distinct categories and Yun et al. develops the usage of problem schema to solve multi-step math problem~\cite{yuhui2010frame}. To summarize, the purpose of introducing problem schema to MWPs is to classify a group of exercises sharing same solution. However, in KT tasks, a crucial consideration is that whether learners' performance are same in solving exercises with the same problem schema. Thus, as long as two exercises share similar solution or similar text representation, they could be considered as sharing the same problem schema. Besides, compared to manual defined schemas in MWPs, we use text clustering methods to generate clusters for similar exercises since it is more applicable for those datasets with more diversified types.
\fi
\iffalse
\subsection{BERT}
BERT is the latest natural language processing (NLP) pre-training method proposed by Google, which trains the general language model on large text corpus. After pre-training, the model can be applied to the downstream NLP tasks such as text classification and reading comprehension. Compared to previous word embedding method word2vec~\cite{goldberg2014word2vec} and ELMo~\cite{peters2018deep}, BERT’s innovation is that it uses bidirectional transformer for language models, which can extract deeper contextual semantic features.
\fi
\subsubsection{Graph Neural Networks} Graph neural networks (GNN) that can model graph structured data such as social network data~\cite{hamilton2017inductive,kipf2016semi,velivckovic2017graph} or knowledge graph data~\cite{seyler2017knowledge,fan2014transition,zhang2019long} has attracted great attention recently. GNN learns node representation by transforming, propagating and aggregating node features and has been proved effective for most graph-structured data~\cite{gilmer2017neural,hamilton2017inductive}. However, the ability of current GNN architectures is limited since their graphs are inherently flat as they only propagate information across the edges of the graph and could not infer and aggregate the information in a hierarchical way~\cite{scarselli2008graph,wu2020comprehensive}. DiffPool~\cite{ying2018hierarchical} is a differentiable graph pooling method, which is first proposed to learn interpretable hierarchical representations of graph. However, it needs to train a separate cluster assignment matrix which is computation expensive. Thus, we present a variant of DiffPool to learn node representations in our paper.
\iffalse
\subsection{Text Clustering}
The purpose of text clustering is to group a set of objects in a manner where objects in the same cluster are more similar. NLP-based method usually involves three aspects~\cite{karypis2000comparison,zhao2004empirical,zhao2002evaluation,zamir1997fast}. Firstly, a suitable distance measure need to be applied to identify the proximity of two feature vectors. Secondly, text clustering requires a criterion function to get the best possible clusters and stop further processing. Thirdly, we need to propose an algorithm to optimize the criterion function. Here, the algorithms proposed could be divided into a wide variety of different types such as agglomerative clustering algorithms, partitioning algorithms, and standard parametric modeling based methods~\cite{aggarwal2012survey}. There are many commonly used methods such as K-means, Hierarchical Clustering and DBSCAN~\cite{macqueen1967some,johnson1967hierarchical,ester1996density}.
\fi
\section{Conclusion}
In this article, we prove the importance of hierarchical relations between exercises for KT tasks. To make full use of text information in exercises, we propose a novel knowledge tracing framework HGKT which leverages the advantages of hierarchical exercise graph and sequence model with attention to enhance the ability of knowledge tracing. Besides, we come up with the concept of a K\&S diagnosis matrix that could trace the mastery of both knowledge and problem schema, which has been proved more effective and useful in industrial applications than the traditional knowledge tracing methods. Moreover, we construct a large-scale knowledge tracing dataset containing exercise text information and conduct extensive experiments to show the effectiveness and interpretability of our proposed models.
\iffalse
\begin{acks}
To Robert, for the bagels and explaining CMYK and color spaces.
\end{acks}
\fi
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
A variety of deep learning (DL) models and algorithms have been proposed and successfully solved a great many of applications fields such as computer vision, natural language processing, data analysis etc. Due to its great success, recently, leveraging machine learning (ML) and deep learning approaches to support decision-making problem has attracted huge attentions in operations research community. Unlike regular ML approaches, DL methods are capable of dealing with intrinsic and potential features that are hidden behind the complex data. In data-driven optimization frameworks, the uncertainty is modeled based on complex data which may has great impact on the optimization solutions. The inaccurate parameters that are derived from complex data may lead the optimizations model sub-optimal or even infeasible. In this sense, data-driven optimization could be benefit from utilizing DL tools.
Recently, decision-making under uncertainty has been applied in various fields such as intelligent transportation, network optimization, scheduling problems, supply chain management etc. In this work, we focus on stochastic programming (SP) technique which is aiming to find the optimal solution that maximize / minimize the expected value of objective function while satisfying all the scenarios that are obtained from uncertain parameters. Conventionally, SP assumes that the probability distribution of uncertain parameters is known from perfect knowledge. In reality, however, it is difficult even impossible to obtain the accurate probability distribution from complex data. It is worth noting that the inaccurate probability distribution may lead the optimization solution to be sub-optimal even infeasible, therefore, it is quite necessary to integrate ML / DL approaches to improve the solution quality of SP.
In this paper, we consider using SP to solve time-series decision-making problems, which has been extensively studied and widely applied in intelligent transportation domain. As discussed in the previous section, probability distribution is required by SP, however, the existing deep learning approaches for time-series predication such as long-short term memory (LSTM)\cite{hochreiter1997long} and gated recurrent unit (GRU)\cite{cho2014learning, chung2014empirical}, both of them return single unit (scalar) output as the predicted results which cannot be used as the parameters for SP. To fill the gap, we devise an innovative deep neural network which involves GRU and mixture density network (MDN) \cite{bishop1994mixture} called GRU-MDN for the time-series probability distribution prediction. Further, we propose a novel data-driven dynamic stochastic programming framework that integrates GRU-MDN along with SP to solve time-series decision-making problems under uncertainty. To best of our knowledge, this is the first work that combine DL and SP for time-series decision-making problems.
The contributions in this work are two-folded: (1) a novel GRU-MDN deep neural network is devised to predict probability distribution of time-series data, (2) stochastic programming is seamlessly integrated with GRU-MDN to formulate relevant problems. The remainder of this paper is organized as follows, the details of data-driven dynamic stochastic programming framework is discussed in section \ref{S:2}, next we apply the framework using a toy example which is in section \ref{S:3}, the conclusions and future work is summarized in section \ref{S:4}.
\section{GRU-MDN stochastic programming framework} \label{S:2}
To make data-driven SP that is capable of solving time-series problem, we propose a GRU-base mixture density network called GRU-MDN. The framework involves three components, GRU is in charge of predicting customer demands, GMM focuses on the probability distribution that are based on the outcomes from GRU, the SP is in charge of modeling uncertainty. We will elaborate our framework in this section.
\subsection{Gated Recurrent Unit}
Unlike the single and simple building block in RNNs, LSTM uses forgetting and gating mechanisms to select and filter information that is necessary for future computation. Figure \ref{fig:gru} shows the detail gating mechanism of GRU.
\begin{figure}
\centering
\includegraphics[scale = 0.3]{images/GRU.png}
\caption{Gated Recurrent Unit}
\label{fig:gru}
\end{figure}
The state of GRU for each time step $t$ is given by the following equations.
\begin{equation}
\begin{array}{l}
z_{t}=\sigma_{g}\left(W_{z} x_{t}+U_{z} h_{t-1}+b_{z}\right) \\
r_{t}=\sigma_{g}\left(W_{r} x_{t}+U_{r} h_{t-1}+b_{r}\right) \\
h_{t}=z_{t} \odot h_{t-1}+\left(1-z_{t}\right) \odot \phi_{h}\left(W_{h} x_{t}+U_{h}\left(r_{t} \odot h_{t-1}\right)+b_{h}\right)
\end{array}
\end{equation}
where $x_{t}$ is the input vector, $h_{t}$ is the output vecotr, $z_{t}$ is the update gate vector, $r_{t}$ is the reset gate vector, $W,U,b$ are the parameter matrices and vector. $\sigma_{g}$ and $\phi_{h}$ are activation functions in sigmoid and hyperbolic types, respectively.
\subsection{Mixture Density Network}
MDN is a variant of a neural network whose output is probability distribution(s) rather than single unit for most of neural networks. The basic idea of MDN is \textit{combining a deep neural network (DNN) and a (group of) mixture of distributions}. Actually, most of the modern DNN architectures such as CNN, RNN and LSTM can be extended to become the special MDNs. The DNN provides the parameters for multiple distributions, which are then mixed by some weights. Also, These weights are provided by the DNN. Notice that the true distribution of the input data can be any type which cannot be described by the parametric methods (e.g. Gaussian distribution, Poisson distribution). Actually, the non-parametric method such as kernel density estimation (KDE) is able to handle with the arbitrary probability distribution learning problem. However, KDE method cannot be directly applied for probability distribution storage in MDN since KDE is a non-parametric approach which implies that it may contain infinite parameters that cannot be store by finite number of neurons. Therefore, the semi-parametric approach - Gaussian mixture model (GMM) is adopted to overcome the problem in MDN, which is formulated as follows.
$$p(\boldsymbol{X \big| \theta})=\sum_{i=1}^{K} w_{i} \mathcal{N}\left(\boldsymbol{X \big| \mu}_{i}, \boldsymbol{\Sigma}_{i}\right)$$
where $\theta = (W, \mu, \Sigma)$, $W = (w_{1}, w_{2}, \cdots, w_{K})$, $\mu = (\mu_{1}, \mu_{2}, \cdots, \mu_{K})$ and $\Sigma = (\Sigma_{1}, \Sigma_{2}, \cdots, \Sigma_{K})$. $K$ is the number of Gaussian distributions. Generally, GMM can be considered as a group of Gaussian distributions with different weights, where the $i-th$ Gaussian is determined by weight $w_{i}$, means $\mu_{i}$ and covariance matrix $\Sigma_{i}$ (variance for univariate Gaussian). Then the predicted probability distribution can be represented using GMM by adjusting the parameter $\theta$. In this work, we use expectation maximization (EM)\cite{dempster1977maximum} algorithm to determine the parameter of GMM, which is summarized as follows.
\begin{algorithm}
\caption{EM}\label{euclid}
\label{al}
\hspace*{\algorithmicindent}
\textbf{Input: The GMM $\sum_{k=1}^{K} \pi_{k} \mathcal{N}\left(\boldsymbol{x} | \boldsymbol{\mu}_{k}, \mathbf{\sigma}_{k}\right)$}
\\
\hspace*{\algorithmicindent}
\textbf{Output: GMM}
\begin{algorithmic}[1]
\State Initialize $\mu_{j}, \Sigma_{j}$ and $\pi_{j}$, $j=1, \cdots, K$
\State E-step. Compute $$
\gamma_{n j}=\frac{\pi_{j} \mathcal{N}\left(\xi_{n} | \mu_{j}, \Sigma_{j}\right)}{\sum_{i=1}^{K} \pi_{i} \mathcal{N}\left(\xi_{n} | \mu_{i}, \Sigma_{i}\right)}
$$
\State M-step. Re-estimate $$
\mu_{j}^{\mathrm{new}}=\frac{1}{N_{j}} \sum_{n=1}^{N} \gamma_{n j} \xi_{n}, \Sigma_{j}^{\mathrm{new}}=\frac{1}{N_{j}} \sum_{n=1}^{N} \gamma_{n j}\left(\xi_{n}-\mu_{j}^{\mathrm{new}}\right)\left(\xi_{n}-\mu_{j}^{\mathrm{new}}\right)^{\top}, \pi_{j}^{\mathrm{new}}=\frac{N_{j}}{N}
$$
where
$$
N_{j}=\sum_{n=1}^{N} \gamma_{n j}
$$
\State Check whether the convergence is satisfied. If not, return to step 2.
\end{algorithmic}
\end{algorithm}
The structure of MDN can be shown in the figure\ref{fig:mdn}.
\begin{figure}
\centering
\includegraphics[scale = 0.2]{images/MDN.png}
\caption{Architecture of Mixture Density Network}
\label{fig:mdn}
\end{figure}
One of the significant contribution of this work is to combine GRU and MDN, our proposed architecture of GRU-MDN can be shown in figure \ref{fig:GRU-MDN}.
\begin{figure}
\centering
\includegraphics[scale=0.4]{images/GRU-MDN.png}
\caption{The Structure of GRU-MDN}
\label{fig:GRU-MDN}
\end{figure}
\begin{comment}
\begin{algorithm}
\caption{Monte Carlo Sampling from GMM}\label{euclid}
\label{al}
\hspace*{\algorithmicindent}
\textbf{Input: The GMM $\sum_{k=1}^{K} \pi_{k} \mathcal{N}\left(\boldsymbol{x} | \boldsymbol{\mu}_{k}, \mathbf{\sigma}_{k}\right)$, number of data points $N$}
\\
\hspace*{\algorithmicindent}
\textbf{Output: the Monte Carlo sampling result}
\begin{algorithmic}[1]
\State $k \gets 0$
\State Repeat steps 1) and 2) until you have the desired amount of samples from the mixture distribution
\State Return the sampling result
\end{algorithmic}
\end{algorithm}
\end{comment}
\subsection{Data-driven dynamic stochastic programming framework}
Finally, we come to summarize our data-driven SP framework which is displayed in the figure \ref{fig:framework}.
\begin{figure}
\centering
\includegraphics[scale = 0.4]{images/framework.png}
\caption{GRU-MDN SP Framework}
\label{fig:framework}
\end{figure}
There are four major components involving in the proposed framework. Specifically, the time-series data is the input of the framework, then the predicted time-series probability distribution is obtained via GUR-MDN, the distribution is input as the parameters of SP, after that various model reformulation and decomposition algorithms can be applied to solve the SP model.
\section{Case Study}\label{S:3}
To validate our data-driven dynamic SP framework, we investigate the car-sharing relocation problem (CSRP) which is referred in \cite{li2020ddksp}. Specifically, we will use the same stochastic programming model and data sets.
\subsection{Experimental Setting}
We use real data from New York taxi trip record data set from January 2017 to June 2019\footnote{https://www1.nyc.gov/site/tlc/about/tlc-trip-record-data.page}. The entire data set is split into training set (from January 2017 to March 2019), and testing set (from April 2019 to June 2019). We compare our proposed approach with data-driven deterministic optimization model where the demand is obtained from the typical time-series prediction approach - LSTM.
The GRU-MDN models are implemented using Python 3.7 + tensorflow 2.1 under the platform CUDA 10.2 GPU, 16GB RAM, Ubuntu 18.04, the mathematical models are solved by Gurobi \footnote{https://www.gurobi.com/downloads/gurobi-software/} 9.0 academic version using Python 3.7 under the platform Intel i7 CPU, 32GB RAM, Windows 10.
\subsection{Experimental Validation}
In order to compare the optimization performance of GRU-MDN and LSTM. We select the similar deep neural network stricture which is shown in Table \ref{table:structure of deep neural network}.
\begin{table}[]
\centering
\caption{The structures of two deep neural network}
\begin{tabular}{|c|c|c|c|}
\hline
& Input Layer & Hidden Layers & Output Layer \\ \hline
GRU-MDN & ws = 10 & (256, 128) & (3*3, 1) \\ \hline
LSTM & ws = 10 & (256, 128) & 1 \\ \hline
\end{tabular}
\label{table:structure of deep neural network}
\end{table}
We select the window size equals 10 as the input layer for both GRU-MDN and LSTM. There are two hidden layers for each neural network with the number of neurons 256 and 128, respectively. Since GRU-MDN returns a probability distribution, we use GMM as the output which involves 3 Gaussian.
It is worth noting that although the structures of both deep neural network are quite similar, the output of GRU-MDN is a predicted time-series distribution, which integrates a two-stage stochastic programming model, while the output of LSTM is a predicted time-series value, which integrates a deterministic model. Then we use the first-stage solutions that are obtained from GRU-MDN and LSTM to test on the testing set (91 days from April 1, 2019 to June 30, 2019). The comparison of the experiment results is shown in Table \ref{table:GRU-MDN vs. LSTM}.
\begin{table}[]
\centering
\caption{Comparison between GRU-MDN Stochastic Programming Model and LSTM Deterministic Model}
\begin{tabular}{|c|c|c|c|}
\hline
& Average Revenue & Average Cost & Average Moving \\ \hline
GRU-MDN & \$947552.6 & \$415601.5 & 231.4615 \\ \hline
LSTM & \$921922.4 & \$438014.2 & 248.7143 \\ \hline
\end{tabular}
\label{table:GRU-MDN vs. LSTM}
\end{table}
The experiment results show that GRU-MDN with SP model is able to yield more average revenue with relatively lower average cost compared to LSTM with deterministic model. Additionally, the moving average of GRU-MDN is 6.94\% lower than LSTM.
\section{Concluding Remarks}\label{S:4}
In this work, we developed a practical data-driven dynamic stochastic programming framework for time-series problem. The approach integrates a GRU-MDN deep neural network along with a two-stage stochastic programming model. Our proposed methodology provides a very efficient framework for data-driven dynamic SP technique. Furthermore, the framework does not apply in the discussed example only. Actually, as a potential extension, the framework can be applied in a number of different applications by replacing the components. For instance, the component of SP can be replaced by distributionally robust optimization \cite{delage2010distributionally} (DRO) which relies on the ambiguity set that contains a family of probability distributions. We believe that the framework is capable of dealing with DRO modeling problems by minor modifications on GRU-MDN. Additionally, different model decomposition algorithms such as L-shape, column generation can be adopted for model solving according to the characteristics of mathematical models.
Although the proposed framework utilize the historical data to solve the time-series decision-making under uncertainty, it does not consider the prior probability distribution which may be quite informative for SP. We believe that using prior probability distribution information may improve the SP solution in a very effective way, therefore, we will investigate Bayesian learning with SP for data-driven SP framework in our future work.
\section{Experiment Validation}
\subsection{Experiment Setting}
To simulate the profitability of demand optimization, we used taxi operation data from 2006 to 2018 in New York City. The LSTM models are implemented using Python 3.7 under the platform CUDA 10.2 GPU, 16GB RAM, Ubuntu 18.04. The entire data set contains more than 22 million taxi trip records, because huge amount of data set is always beneficial to get better performance using deep learning models. To model LSTM, we split the total data set by train and validation data set. We first separate a test set by selecting the data after 2016-06-01. The training and validation data set are further separated from the data before 2016-6-1 so that the time order is preserved. The features for training models includes latitudes and longitudes for both pickup and drop-off locations, as well as pickup time stamp and number of passengers. These features are necessary for a taxi dispatching system when a passenger call for a demand. First, we hope to build predictive models to estimate the taxi duration trip. In the historical training data, the taxi trip duration is known and therefore the target variable for this task. Although the online APIs such as google map can estimate the taxi trip duration by calculate the distance between the pick-up and drop-off coordinates, it is not automation processing for a dispatching system, and moreover, the taxi trip duration depends on many other factors such as the traffic jam affected by time stamps, the road conditions and so on. The GRU-MDN are implemented using Python 3.7 under the platform CUDA 10.2 GPU, 16GB RAM, Ubuntu 18.04, the mathematical models are solved by Gurobi \footnote{https://www.gurobi.com/downloads/gurobi-software/} 9.0 academic version under the platform Intel i7, 32GB RAM, Windows 10.
The charing station information can be found from \footnote{https://data.ny.gov/Energy-Environment/Electric-Vehicle-Charging-Stations-in-New-York/7rrd-248n}
The EVDP can be formulated as follows.
$$
\text{maximize} \sum_{t \in T} \sum_{v \in V} \sum_{q \in Q} x_{v q t} d_{q}-\lambda \sum_{t \in T} \sum_{v \in V} t w_{v t}
$$
\subsection{Performance Measures}
\section{Related work (literature review)}
\subsection{Gaussian Mixture Model}
The standard GMM can be stated as follows.
$$
p(\boldsymbol{x})=\sum_{k=1}^{K} \pi_{k} \mathcal{N}\left(\boldsymbol{x} | \boldsymbol{\mu}_{k}, \mathbf{\Sigma}_{k}\right)
$$
Given a GMM, the goal is to maximzie the likelihood function with respect to the parameters.
The Expectation-Maximization (EM) algorithm \cite{} is used in maximum likelihood estimation where the problem involves two sets of random variables of which one, X, is observable and the other, Z, is hidden. The goal of the algorithm is to find the parameter vector $\Phi$. The algorithm basically involves two steps: Expectation step and Maximization step which can be summarized as follows.
\par
E-step : $\mathcal{Q}\left(\Phi | \Phi^{l}\right)=E\left[\mathcal{L}_{c}(\Phi | \mathcal{X}, \mathcal{Z}) | \mathcal{X}, \Phi^{l}\right]$
\par
M-step : $\Phi^{l+1}=\arg \max _{\Phi} \mathcal{Q}\left(\Phi | \Phi^{l}\right)$
\par
asdf
\par
asdf
\par
asdf
\begin{algorithm}
\caption{EM algorithm}\label{euclid}
\label{al}
\hspace*{\algorithmicindent}
\textbf{Input: data $\mathcal{X}$, number of Gaussian $\mathcal{K}$}
\\
\hspace*{\algorithmicindent}
\textbf{Output: the parameters of GMM ($\pi$, $\mu$, $\sigma^{2}$)}
\begin{algorithmic}[1]
\State Initialize the means $\mu_{k}$, variances $\sigma^{2}_{k}$ and mixing coefficients $\pi_{k}$, and evaluate the initial value of the log likelihood.
\State $\boldsymbol{E-step}$
$$
\gamma\left(z_{n k}\right)=\frac{\pi_{k} \mathcal{N}\left(\mathbf{x}_{n} | \boldsymbol{\mu}_{k}, \mathbf{\Sigma}_{k}\right)}{\sum_{j=1}^{K} \pi_{j} \mathcal{N}\left(\mathbf{x}_{n} | \boldsymbol{\mu}_{j}, \mathbf{\Sigma}_{j}\right)}
$$
\State $\boldsymbol{M-step}$
$$
\boldsymbol{\mu}_{k}^{\mathrm{new}} =\frac{1}{N_{k}} \sum_{n=1}^{N} \gamma\left(z_{n k}\right) \mathbf{x}_{n}
$$
$$
\mathbf{\Sigma}_{k}^{\mathrm{new}} =\frac{1}{N_{k}} \sum_{n=1}^{N} \gamma\left(z_{n k}\right)\left(\mathbf{x}_{n}-\boldsymbol{\mu}_{k}^{\mathrm{new}}\right)\left(\mathbf{x}_{n}-\boldsymbol{\mu}_{k}^{\mathrm{new}}\right)^{\mathrm{T}}
$$
$$
\pi_{k}^{\mathrm{new}} =\frac{N_{k}}{N}
$$
\State Evaluate the log likelihood
$$
\ln p(\mathbf{X} | \boldsymbol{\mu}, \mathbf{\Sigma}, \boldsymbol{\pi})=\sum_{n=1}^{N} \ln \left\{\sum_{k=1}^{K} \pi_{k} \mathcal{N}\left(\mathbf{x}_{n} | \boldsymbol{\mu}_{k}, \boldsymbol{\Sigma}_{k}\right)\right\}
$$
\end{algorithmic}
\end{algorithm}
\section{Proof of Proposition \ref{p1}}
\begin{proof}
The SUBP can be treated as min-cost flow problem on the directed graph $\mathscr{N}$
\end{proof}
\section{}
\bibliographystyle{model1-num-names}
|
\section{Introduction}
Neural networks are mathematical functions that map some set of input values to output values \citep{Goodfellow2016}. Neural network models belong to the class of representation learning methods that automatically discover the underlying representations of data. A neural network model is composed of multiple processing layers that each transforms the representation at one level into a more abstract representation starting from the raw input \citep{LeCun2015}. As such, a very complex function can be learned if we combine enough transformations. Such transformations are obtained by stacking nonlinear modules \citep{LeCun2015,Goodfellow2016}. Neural networks are also known to be universal approximators which means that regardless of the function we want to learn, a large enough neural networks can represent such a function \citep{hornik1989multilayer}. However, learning the desired
function using neural networks is challenging and there is no guarantee that we can find the right parameters for the neural networks \citep{Goodfellow2016}.
Similar to most machine learning methods for prediction, neural networks usually produce point predictions without any information about how far point predictions are from the ground truth response variables. Because point predictions produced by neural networks do not assess the prediction error from the same data used to generate point predictions, neural networks are lacking in inferential capability from a statistical standpoint. In this paper, we develop prediction intervals based on neural network point predictions that produce a range of values including an unknown continuous univariate response with any specified level of confidence.
Following the setup in \citep{zhang2019random}, $(X,Y) \in \mathbb{R}^p \times \mathbb{R} $ denote the predictor-response pair randomly sampled from some distribution $\mathbb{G}$, where $p$ is the number of predictors and $Y$ is a continuous univariate response. We develop a prediction interval for the observation $(X,Y)$ denoted as $I_\alpha(X,C_n)$ that will cover the true response variable with the probability $1-\alpha$, where $C_n$ is a training set including observations $(X_1,Y_1),...,(X_n,Y_n) \overset{\text{iid}}{\sim} \mathbb{G} $ and $(X,Y)$ is independent of the training set $C_n$.
Several approaches have been proposed to construct prediction intervals for neural networks. For example, \cite{hwang1997prediction} proposed an asymptotic approach to construct prediction intervals for neural networks. They estimated the asymptotic variance of the neural network predictions and used the $1-\alpha/2$ quantile of a $t$-distribution to create prediction intervals. \cite{de1998prediction} constructed prediction intervals for neural networks based on the asymptotic variance of the estimated parameters of the neural networks. \cite{khosravi2010lower} proposed a lower upper bound estimation method to construct two outputs for a neural network model for estimating the prediction interval bounds. \cite{kivaranovic2019adaptive} proposed a distribution-free split conformal prediction interval for neural networks. They designed a prediction interval network which had three outputs to estimate the median and the lower and upper bounds of prediction intervals.
The conformal prediction interval framework is a general method to construct prediction intervals and provides distribution-free predictive inference \citep{vovk2005algorithmic}. Many prediction interval methods have been proposed based on conformal inference. For example, \cite{lei2018distribution} proposed a distribution-free predictive inference for regression leading to split conformal (SC) prediction intervals. However, the SC method may not always have a good performance, especially when the sample size is small because the SC approach uses only half of the data to train the regression function, which may not be sufficient. In this creative component, we propose a $k$-fold prediction interval method to construct prediction intervals based on $k$-fold cross validation. This method tends to produce narrower prediction intervals compared to SC intervals while maintaining the same coverage probability. Our experimental results suggested that the proposed $k$-fold prediction interval method is effective and especially advantageous when the number of training observations is limited.
The remainder of this paper is organized as follows. Section 2 describes the methodology. Section 3 presents the simulation study. Section 4 explains the data analysis results. Finally, we conclude the paper in section 5.
\section{$k$-fold Conformal Prediction Intervals}
The conformal prediction interval framework is a general approach for efficiently constructing prediction intervals \citep{vovk2005algorithmic}. To decrease the computational cost of the full conformal method, Lei et al. proposed split conformal prediction intervals, which are considerably computationally more efficient than the full conformal method \citep{lei2018distribution}. The SC prediction interval algorithm includes the following steps:
\begin{enumerate}
\item Randomly split $\{1,...,n\}$ into two equal-sized subsets $L1$ and $L2$.
\item Train a regression function from $\{(X_i,Y_i): i \in L_1 \}$ to estimate the mean function denoted as $\hat{m}_{n/2}(X)$.
\item For $i \in L_2$, compute the prediction error $D_i=Y_i-\hat{Y_i}$, where $\hat{Y_i}=\hat{m}_{n/2}(X_i)$.
\item Construct the prediction interval with coverage probability $1-\alpha$ for $Y$ as $[\hat{Y}-D_{[n/2,\alpha/2]}, \hat{Y}+D_{[n/2,\alpha/2]}]$, where $D_{[n/2,\eta]}$ is the $\eta$ quantile of the empirical distribution of $D_1,..., D_{n/2}$.
\end{enumerate}
Although SC method generates reliable prediction intervals, it may not always have good performance, especially when the sample size is small. The SC method uses half of the data to train the regression function which may not always be sufficient. All observations do not get a chance to contribute to the construction of empirical distribution of errors. In this paper, we propose a new method to construct prediction intervals based on $k$-fold cross validation called $k$-fold conformal prediction interval. The $k$-fold conformal prediction interval algorithm includes the following steps:
\begin{enumerate}
\item Randomly split $\{1,...,n\}$ into $k$ equal-sized subsets denoted as $L_1$, $L_2$,..., $L_k$.
\item For each $L_j$ where $j\in\{1,...,k\}$ do the following:
\begin{enumerate}
\item Train a regression function from $\{(X_i,Y_i): i \in \bigcup\limits_{r=1 }^{k} L_{r}, \; r\neq j \}$ to estimate the mean function denoted as $\hat{m}_{j}(X)$.
\item For $i \in L_j$, compute the prediction error $D_i=Y_i-\hat{Y_i}$, where $\hat{Y_i}=\hat{m}_{j}(X_i)$.
\end{enumerate}
\item Construct the prediction interval with coverage probability $1-\alpha$ for $Y$ as $[\hat{Y}-D_{[n,\alpha/2]}, \hat{Y}+D_{[n,\alpha/2]}]$, where $D_{[n,\eta]}$ is the $\eta$ quantile of the empirical distribution of $D_1,..., D_{n}$.
\end{enumerate}
The proposed $k$-fold conformal prediction interval requires estimation of $k$ regression functions which results in the empirical distribution of errors based on the all training data observations. Thus, the proposed method's computational cost is on the order of $k$ times that of the SC method. Each training set is larger too, so there could be added expense.
\section{Simulation Study}
To evaluate the finite-sample performance of the proposed approach, we conducted a simulation study to compare our proposed method to the SC method with respect to coverage rate and interval width performance measures. Data are simulated from an additive error model: $Y=m(X)+\epsilon$, where $X=(X_1,...,X_p)$ with $p=10$ and $X \sim \mathcal{N}(0,\varSigma_p)$, where $\varSigma_p$ is an AR(1) covariance matrix with $\rho=0.6$ and diagonal values equal to 1. We considered three factors, namely the distribution of the error terms, the choice of mean function $m()$, and the number of training observations $n$. Following \citep{zhang2019random}, we considered the following factorial design for these three factors:\\
\begin{itemize}
\item \textbf{Mean functions}:
\begin{enumerate}
\item linear: $m_1(x)=x_1+x_2$
\item nonlinear: $m_2(x)=2\exp(-|x_1|-|x_2|))$
\item nonlinear with interaction: $m_3(x)=2\exp(-|x_1|-|x_2|))+x_1x_2$
\end{enumerate}
\item \textbf{Distributions of errors}:
\begin{enumerate}
\item homoscedastic: $\epsilon \sim \mathcal{N}(0,1)$
\item heavy-tailed: $\epsilon \sim \frac{t_3}{\sqrt{3}}$, where $t_3$ is a $t$-distribution with 3 degrees of freedom.
\item heteroscedastic: $\epsilon \sim \mathcal{N}(0,\frac{1}{2}+\frac{1}{2} \frac{|m(X)|}{E|m(X)|})$
\end{enumerate}
\item \textbf{Training Sample sizes}: $n=500, 2500$, and $5000$
\end{itemize}
The full-factorial design has 27 different simulation scenarios. The following hyperparameter were used to train the neural network model. The neural network model has 2 fully connected layers with 15 neurons in each layer. We investigated different activation functions, such as ReLU and tanh, and found that ReLU had the best overall performance. Only results for ReLU are reported here. All weights were initialized with the Xavier method \citep{Glorot2010}. We used stochastic gradient descent (SGD) with a mini-batch size of 32. The Adam optimizer \citep{Kingma2014} with learning rate of 0.03\% was used to minimize the loss function. The model was trained for 20,000 iterations.\\
\subsection{Evaluation of Coverage Rates and Interval Widths}
The nominal coverage level was set at 0.9 for the construction of all prediction intervals in the simulation study. 50 datasets were simulated for each simulation scenario. We also generated 500 test samples independently from the joint distribution of $(X,Y)$ for each simulation scenario. We defined the coverage rate as the percentage of response values contained in their corresponding prediction intervals for the test data. We estimated the coverage rate by mean of coverage rates obtained from 50 simulated datasets for each simulation scenario. To evaluate the effect of $k$ on the performance of prediction intervals, we considered three different $k$ values, namely 2, 5 and 10. Figure \ref{fig:coverage} compares the coverage rate estimates of the SC method, 2-fold prediction interval method, 5-fold prediction interval method, and 10-fold prediction interval method. The white circle in each boxplot is the average of the 50 coverage rate estimates for each simulation scenario. The estimates of the coverage rates for k5 and k10 are closer to 0.9 (the nominal level) especially for the larger sample sizes compared to the SC and k2 methods. The SC and k2 methods tend to over-cover the response values based on coverage rates especially when the sample is small because these methods use only half of the training data to train the mean function which may be insufficient when $n$ is small. Thus, the prediction errors would be larger resulting in a wider prediction intervals and over-coverage. The SC method uses only half of the training data to find the prediction errors for the other half of the training data which makes it similar to the k2 method except that k2 method finds prediction errors for all training data. As such, SC and k2 methods have very similar boxplots for the coverage rates. As the sample size increases, the coverage estimates of all methods become more concentrated around the nominal level due to having adequate information to estimate the mean functions using neural networks. The results suggest that all competing methods showed stable behavior in terms of the coverage rate estimates across all factors including the mean functions and the choice of error distributions.
\begin{figure}[H]
\centering
\includegraphics[scale=0.38]{Coverage_NN_cor.png}
\caption{Boxplots of the coverage rate estimates of the split conformal method (SC), 2-fold prediction interval method (k2), 5-fold prediction interval method (k5), and 10-fold prediction interval method (k10). The white circle in each boxplot is the average of the 50 coverage rate estimates for each simulation scenarios. The dashed red lines show the nominal coverage level which is set to be 0.9 in our study.}\label{fig:coverage}
\end{figure}
To evaluate the prediction interval widths, we averaged the 500 test cases' interval widths for each simulated dataset. To better compare the proposed $k$-fold prediction interval method with the SC method, we computed the ratio of the SC interval width to the $k$-fold prediction interval width. Figure \ref{fig:width} shows the $log_2$ ratios of the interval widths. As shown in Figure \ref{fig:width}, interval width decreases as the sample size increases due to availability of enough training data to estimate the mean functions well. The SC and k2 produce intervals of approximately the same width as indicated by log ratios close to zero. However, the SC method tends to have a slightly smaller interval width, especially when the sample size is small. Results demonstrate that k5 and k10 prediction intervals are smaller than intervals constructed by the SC method. The k5 and k10 prediction interval methods have a comparable performance in terms of prediction interval widths, which indicates that increasing $k$ in the $k$-fold prediction interval method does not always improve the performance. This indicates an opportunity to choose $k$ to obtain narrow prediction intervals while maintaining low computational costs. The log ratios show that the k5 and k10 methods have the biggest advantages over SC intervals, in terms of width, when the training sample size is small.
\begin{figure}[H]
\centering
\includegraphics[scale=0.38]{interval_width_N_cor_new.PNG}
\caption{Boxplots of the $log_2$ ratios of split conformal (SC) interval widths to 2-fold prediction interval (k2) widths, 5-fold prediction interval (k5) widths, and 10-fold prediction interval (k10) widths. The white circle in each boxplot is the average of the 50 $log_2$ interval width ratios for each simulation scenario.}\label{fig:width}
\end{figure}
\section{Data Analysis}
To evaluate the performance of our proposed prediction interval method on real-world datasets, we selected 10 datasets from UC Irvine Machine Learning Repository website which are summarized in the Table \ref{tab:summary}.
\begin{table}[H]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
No.& Name of dataset & Number of predictors & Number of observations\\
\hline
1&Power Plant&4&9,568\\
\hline
2&Facebook Metrics & 18& 500\\
\hline
3& Parkinsons Telemonitoring& 21& 5,875\\
\hline
4& Bodyfat& 13& 252\\
\hline
5& Residential Building& 106 & 372\\
\hline
6 & Real Estate Valuation & 5& 414\\
\hline
7& Wine Quality & 11 & 4898\\
\hline
8& Aquatic Toxicity &8&546\\
\hline
9& Fish Toxicity & 6 &908\\
\hline
10&Energy Efficiency &8 & 768\\
\hline
\end{tabular}\caption{The summary of real datasets.}\label{tab:summary}
\end{table}
To obtain the data analysis results in this section, we used the following hyperparameters for neural networks. The neural networks model has 2 fully connected layers with 10 neurons in each layer. As in the simulation described in the section 3, ReLU activation functions were used and all weights were initialized with Xavier method \citep{Glorot2010}. We used stochastic gradient descent (SGD) with a mini-batch size of 16. The Adam optimizer \citep{Kingma2014} with learning rate of 0.03\% was used to minimize the loss function. The model was trained for 25,000 iterations.\\
To estimate the coverage rates and interval widths, we used 5-fold cross validation which was repeated 20 times for each dataset. Since the simulation study suggested that the SC method tends to perform better compared to the 2-fold prediction interval method (k2), we did not use the k2 method in this section. We employed the 5-fold prediction interval method (k5) rather than the 10-fold prediction interval method (k10) to decrease the computational cost. Figure \ref{fig:coverage_real} compares the coverage rate estimates of the SC method and the k5 method. The white circle in each boxplot is the average of the 20 coverage estimates that resulted from the cross validation procedure.
As shown in Figure \ref{fig:coverage_real}, the k5 prediction interval has higher average coverage than the SC method except for the Fish dataset. The results also indicate under coverage for some of the datasets. To evaluate the prediction interval widths, we averaged the 20 interval width estimates for each dataset. To better compare the proposed 5-fold prediction interval method with the SC method, we computed the ratio of the SC interval width to the width of the 5-fold prediction interval method. Figure \ref{fig:width_real} shows the $log_2$ ratios of the interval widths. The results indicate that the k5 method tends to have a smaller interval width than the SC method especially for the datasets with the smaller sample sizes since the $log_2$ ratios of the average SC interval widths to the average k5 widths are highest when the sample sizes are small. The results also reveal that the k5 method and SC method have a comparable performance for datasets with sufficiently large sample size. The SC method had a smaller interval width compared to the k5 method for the Power Plant dataset, but this was due to under coverage.
\begin{figure}[H]
\centering
\includegraphics[scale=0.32]{real_coverage.PNG}
\caption{Boxplots of the coverage rate estimates of split conformal method (SC) and 5-fold prediction interval method (k5). The white circle in each boxplot is the average of the 20 coverage rate estimates resulting from the cross validation procedure. The dashed red lines show the nominal coverage level which is set to be 0.9 in our study.}\label{fig:coverage_real}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[scale=0.32]{real_width.PNG}
\caption{Scatter plot of the $log_2$ ratios of the average split conformal (SC) interval widths to the average 5-fold prediction interval (k5) widths.}\label{fig:width_real}
\end{figure}
\section{Conclusion}
In this paper, we presented a conformal prediction interval method for neural networks using cross validation. The proposed method uses $k$-fold cross validation on the training data to estimate the empirical error distribution of the errors. Then, the quantiles of the empirical error distribution are used to construct prediction intervals for test data. Our experimental results indicate that the $k$-fold prediction interval method can efficiently construct prediction intervals for neural networks that compare favorably with intervals produced by the split conformal method. Our results suggest that the proposed method is well-suited for datasets with a small number of observations. We also found that the performance of the proposed model is somewhat sensitive to the choice of $k$. As such, it is important to tune $k$ in the proposed method to get narrow prediction intervals while maintaining low computational cost.
\bibliographystyle{plainnat}
|
\section{Introduction and model}
\label{sec:introduction}
A future much greater reliance on renewable energy means that there is
likely to be corresponding much greater need for storage in order to
keep electricity systems balanced---see~\cite{Strbacetal,NSPN}. The
optimal operation of energy storage for such balancing may be
considered from the viewpoint of the provider
(see~\cite{SDJW,CZ,DEKM,GTL,PADS,WEITZEL2018582} and the references
therein), or from that of the system operator, who is seeking to
schedule given storage resources so as to balance the system as far as
possible. The latter problem has only received significant attention
relatively recently, and then mostly for the problem of scheduling
initially full stores so as to cover periods of continuous energy
shortfall---see, e.g., \cite{NGECR18} for practical applications in
the context of the GB energy system, \cite{KHAN201839, Sioshansi2014,
ZHOU201512} for dynamic programming and simulation approaches,
\cite{ETA-pscc,ETA-tosg, ETA-tops, CZ2018} for underlying theory and
\cite{Zhu2019} for a hybrid approach that uses an analytical discharge
policy and a recharging policy based on machine learning.
The present paper considers the problem of optimally scheduling
heterogeneous storage resources---characterised by different
capacities, input/output rate constraints, and round-trip
efficiencies---over extended periods of time in which there are both
periods of energy shortfall to be met from storage and periods of
energy surplus available to recharge storage. Our main objective is
the minimisation of unserved energy demand over any given time
horizon, or, in a stochastic environment, the expectation of
this---often referred to as \emph{expected-energy-not-served} (EENS).
We work primarily in a deterministic environment. Progress may be
made using strong Lagrangian or other closely related
techniques---see, e.g.,~\cite{Whi} and, for an application of this
approach to the present problem, see~\cite{CZ2018}. However, in a
deterministic environment optimal solutions are typically far from
unique, and the optimal solution obtained by such techniques as above
is often such that at any point in time the optimal decision has some
dependence on the future evolution of the external supply-demand
balance process. The approach of the present paper does not require
the use of such machinery and, for the problems considered in the
paper, it turns out that the optimal (or near optimal) policy is
generally \emph{non-anticipatory}, i.e.\ the optimal decision at each
successive time is independent of the evolution of the external
supply-demand balance subsequent to that time, so that results
continue to hold in a stochastic environment, both for the
minimisation of the unserved energy demand and for the minimisation of
any quantile of the distribution of the unserved energy demand.
We thus study a given set~$\S$ of energy stores. At each
time~$t\in[0,\infty)$, the volume of usable energy (after accounting
for any output losses) within each store $i\in\S$ is given by $E_i(t)$,
where the latter is subject to the \emph{capacity constraint}
\begin{equation}
\label{eq:1}
0 \le E_i(t) \le \overline E_i.
\end{equation}
A \emph{policy} for the use of the stores is a specification, for each
time~$t\ge0$, of the rate~$r_i(t)$, at which each store $i\in\S$
serves energy at time~$t$, where positive values of $r_i(t)$
correspond to the store discharging, and negative values of $r_i(t)$
correspond to the store charging. Hence, in particular,
\begin{equation}
\label{eq:2}
E_i(t)=E_i(0)-\int_0^tr_i(u)\,du,
\quad\text{ for all $t\ge0$.}
\end{equation}
For each store $i\in\S$, the rates~$r_i(t)$ are further required to
satisfy the power or \emph{rate constraints}
\begin{equation}
\label{eq:3}
-P'_i \le r_i(t) \le P_i,
\quad\text{ for all $t\ge0$,}
\end{equation}
for given constants $P'_i\ge0$ and $P_i>0$. Finally, each store
$i\in\S$ is assumed to have a \emph{round-trip efficiency}
$0<\eta_i\le1$, so that, at any time~$t$ such that the store is
charging (i.e.\ $r_i(t)<0$) the rate at which energy must be supplied
to the store from some external source is given by $-r_i(t)/\eta_i$
(recall that the level of energy within a store is measured as that
which it may usefully output).
The stores are used to assist in managing some given \emph{demand
process} $(d(t), t\ge0)$, defined for all times~$t\ge0$, positive
values of which correspond to an external energy demand to be met
(perhaps partially) from the stores, and negative values of which
correspond to an external energy surplus which may be used to recharge
the stores.
In Section~\ref{sec:pure-discharge-model} we study the case in which
the demand process $(d(t),\,t\in[0,T])$ is nonnegative over some given
time interval $[0,T]$ and is to be served as far as possible over that
interval by the stores, subject to the constraint that the latter may
only discharge for $t\in[0,T]$. In particular this may be appropriate
to the situation in which storage is used to cover continuous periods
of what would otherwise be energy shortfall (e.g.\ periods of daily
peak demand), but may readily be fully recharged between such periods.
We show that there is a policy which minimises the unserved energy
demand
\begin{equation}
\label{eq:4}
\int_0^T\max\biggl(d(t) - \sum_{i\in\S}r_i(t),\,0\biggr)\,dt,
\end{equation}
and in which the rates~$r_i(t)$, $i\in\S$, at each time~$t$ depend
only on $d(t)$ and the energy in each store at time~$t$. This policy
therefore continues to be optimal in a stochastic environment. The
results in this section gather together---and, in considering
arbitrary time intervals, extend---results collectively obtained in
\cite{Nash1978, ETA-pscc, ETA-tosg, ETA-tops, CZ2018}, but are now
unified and presented with considerably simpler proofs, laying a
necessary foundation for subsequent sections.
In Section~\ref{sec:disch-with-cross} we continue to assume a
nonnegative demand process over some time interval $[0,T]$, but allow
that individual stores may both charge and discharge over that
interval, typically corresponding to the situation is which
\emph{cross-charging} between stores is allowed. We show that such
cross-charging may occasionally be helpful, but give results which
identify common situations in which it is not. In particular, we show
that cross-charging cannot be helpful when, as discussed in the
preceding paragraph, storage may be fully recharged between periods of
external energy shortfall and in which the energy shortfall during
such periods is \emph{unimodal}, increasing to a maximum and
thereafter decreasing.
Finally, in Section~\ref{sec:charging-discharging} we consider the
general case in which the demand process may be both positive and
negative, where negative values have the interpretation given above.
We study the situation in which the stores have a common round-trip
efficiency, and use earlier results to identify heuristic policies for
the (near) optimal management of the storage, and to derive conditions
under which they are truly optimal.
\section{Pure discharge model}
\label{sec:pure-discharge-model}
In this section we take the demand process $(d(t), t\ge0)$ to be
nonnegative over some time interval of interest. Without loss of
generality---by, if necessary, redefining the demand process to be
zero outside it---we may take this time interval to consist of the
entire positive half-line, so that $d(t)\ge0$ for all $t\ge0$. We
assume that, throughout this time interval, each store~$i\in\S$ may
only discharge, so that the rate constraint~\eqref{eq:3} is replaced
by
\begin{equation}
\label{eq:5}
0 \le r_i(t) \le P_i,
\quad\text{ for all $t\ge0$,}
\end{equation}
and that
\begin{equation}
\label{eq:6}
\sum_{i\in\S}r_i(t) \le d(t),
\qquad t \ge 0.
\end{equation}
The energy~$E_t(t)$ in each store~$i$ at each time~$t\ge0$ is then as
given by~\eqref{eq:2} and is a (weakly) decreasing function of~$t$
which we continue to require to satisfy~\eqref{eq:1} (though the
second inequality in~\eqref{eq:1} plays no part in the analysis of
this section). Our objective is to choose rate processes
$(r_i(t),\,t\ge0)$ for all stores~$i\in\S$, satisfying the above
constraints, with the objective of either satisfying~\eqref{eq:6} with
equality for all~$t$ in some interval $[0,T]$ where $T$ is as large as
possible, or else that of minimising the \emph{unserved energy
demand}, given by~\eqref{eq:4}, over any given time interval $[0,T]$
(where we allow also $T=\infty$).
Under any given policy for the use of the stores in~$\S$, the
\emph{further} capabilities of the energy contained within that set at
and subsequent to any given time~$t\ge0$ are defined by the rate
constraints~$P_i$ and by the residual stored energies
$(E_i(t),\,i\in\S)$ at time~$t$. It is helpful to have some efficient
way of representing this \emph{residual stored-energy configuration}.
This should be sufficient to characterise the set of residual demand
processes $(d(t'),\,t'\ge t)$ which may be fully served at and
subsequent to the time~$t$. For any such time~$t$, define the
(residual) \emph{burst-power profile} of the stored-energy
configuration at that time as the necessarily decreasing function
$s^t(u)$ of $u$ given by
\begin{equation}
\label{eq:7}
s^t(u) = \sum_{i\in\S\colon E_i(t)/P_i \ge u} P_i
\end{equation}
(see also Figure~\ref{fig:lgtf}).
For each store~$i\in\S$, we refer to the quantity $E_i(t)/P_i$ as the
\emph{(residual) discharge-duration} of the store~$i$ at the time~$t$.
This is the length of further time for which the store~$i$ could
supply energy if it did so at its maximum rate. Thus the integral of
$s^t(u')$ from~$0$ to any time~$u$ is the maximum amount of further
energy which can possibly be supplied by the stores between the
times~$t$ and $t+u$. Define also the \emph{energy-power transform}
$\epts{e}{t}{p}$ \cite{ETA-tosg} of the \emph{burst-power profile} at
the time~$t$ to be given by
\begin{equation}
\label{eq:8}
\epts{e}{t}{p} = \int_0^\infty \max(0, s^t(u) - p)\,du, \qquad p \ge 0.
\end{equation}
This has the interpretation that, for each~$t$ and for each $p\ge0$,
the quantity $\epts{e}{t}{p}$ would be the energy supplied above the
reference output~$p$ if all stores output at their maximum rates for
as long as possible (i.e.\ until empty) subsequent to the time~$t$
(again see Figure~\ref{fig:lgtf}).
Note that, for each time~$t$, the burst-power profile (function
of~$u$) given by $s^t(u)$ is recoverable from the energy-power
transform (function of~$p$) given by $\epts{e}{t}{p}$.
Observe that $\epts{e}{t}{p}$ is a (weakly) decreasing function of
both $p$ and $t$. For each $t\ge0$, the quantity $\epts{e}{t}{0}$ is
the total energy in the stores at time~$t$, and $\epts{e}{t}{p}$ is a
convex, piecewise linear, (weakly) decreasing function of $p$ which is
zero for all $p\ge\sum_{i\in\S:E_i(t)>0}P_i$. The quantity
$\sum_{i\in\S:E_i(t)>0}P_i$ is also the maximum rate at which demand
which may be served by the stored-energy configuration at time~$t$,
and is of course decreasing in~$t$.
For each~$T>0$ and for each~$t\in[0,T]$, define also the
\emph{energy-power transform} $\eptd{e}{t,T}{p}$ of the \emph{demand
process} on the interval $[t,T]$ to be given by
\begin{equation}
\label{eq:9}
\eptd{e}{t,T}{p} = \int_t^T\max(0, d(u) - p)\,du, \qquad p \ge 0.
\end{equation}
We allow also $T=\infty$, and write $\eptd{e}{t}{p}$ for
$\eptd{e}{t,\infty}{p}$. The quantity $\eptd{e}{t,T}{p}$ has the
interpretation that it would be the unserved energy demand over the
interval $[t,T]$ demand if power were supplied at a constant rate~$p$
during that interval. For each~$T$ and for each $p\ge0$, the function
$\eptd{e}{t,T}{p}$ is (weakly) decreasing in~$t$ and, for each
$t\ge0$, the function $\eptd{e}{t,T}{p}$ is convex and (weakly)
decreasing in~$p$ and is zero for all $p\ge\max_{u\ge t}d(u)$.
We shall say that the (residual) stored-energy configuration
$(E_i(t), i\in\S)$ at any time~$t$ is \emph{balanced} at time~$t$ if
and only if $E_i(t)/P_i$ is constant over all stores $i\in\S$. If the
stored energy configuration is balanced at time~$t$, then it may be
kept balanced at all subsequent times by always serving energy from
each individual store~$i\in\S$ at a rate which is proportional
to~$P_i$. Thus, in this case and under such a policy, the stores may
fully serve any residual demand process $(d(u), u\ge t)$ such that
\begin{equation}
\label{eq:10}
d(u) \le \hat P, \text{ for all $u\ge t$,}
\qquad\text{and}\qquad
\int_t^\infty d(u)\,du \le \hat E(t),
\end{equation}
where $\hat P=\sum_{i\in\S}P_i$ and $\hat E(t)=\sum_{i\in\S}E_i(t)$.
Since, under \emph{any} policy, these conditions are clearly also
necessary in order that the stores, balanced at time~$t$ as above, may
fully serve a given residual demand process subsequent to time~$t$, it
follows that the above policy for the subsequent use of a balanced
energy configuration is optimal in terms of its ability to satisfy any
requested demand. Indeed the balanced store configuration at time~$t$
has the same subsequent energy-serving capability as a single store
with the same total energy content $\hat E(t)$ and a maximum output
rate of $\hat P$.
When the residual stored-energy configuration $(E_i(t), i\in\S)$ at
time~$t$ is balanced as above, the corresponding energy-power
transform $\epts{e}{t}{p}$ decreases \emph{linearly} in $p$ from
$\hat E(t)$ when $p=0$ to zero when $p=\hat P$ (and is zero
thereafter). The conditions~\eqref{eq:10} on the residual demand
process $(d(u), u\ge t)$ are equivalent to
$\eptd{e}{t}{0}\le\hat E(t)$ and $\eptd{e}{t}{\hat P}=0$, where
$\eptd{e}{t}{p}$ is the corresponding energy-power transform of that
process on $[0,\infty)$. Since the latter is convex, it follows that
the residual demand at and subsequent to time~$t$ may be completely
served if and only if $\eptd{e}{t}{p}\le\epts{e}{t}{p}$ for all
$p\ge0$, and is then served by keeping the residual stored-energy
configuration balanced subsequent to time~$t$. It is a consequence of
Theorem~\ref{theorem:1} below that, under a suitable policy for the
use of the stores, this result extends to general stored-energy
configurations.
Suppose that, under a given policy for the use of the set of
stores~$\S$, the total rate at which energy is to be served at each
time~$t\ge0$ is given by $\bar d(t)\le d(t)$. We shall say that such
a policy is a \emph{greatest-discharge-duration-first} policy if, at
each successive time~$t$, the rates at which the individual stores
$i\in\S$ serve energy is given by prioritising the use of the stores
in descending order of their discharge-durations $E_i(t)/P_i$. More
exactly, under this policy at each time~$t$ the stores are grouped
according to their current discharge-durations~$E_i(t)/P_i$ (so that
stores belong to the same group if and only if their
discharge-durations are equal); sufficient groups are then selected in
descending order of their discharge-durations such that, using the
stores within each group at their maximum rates (i.e.\ each store~$i$
within a selected group serves energy at a rate~$P_i$), the total rate
at which energy is served is the required~$\bar d(t)$; however, in
order to meet precisely the total rate~$\bar d(t)$, each store in the
\emph{last} group thus selected may only require to serve energy at
some fractional rate~$\lambda P_i$ for some constant~$\lambda$ such
that $0<\lambda\le1$. (See Figure~\ref{fig:lgtf}.) As time~$t$
increases, any such greatest-discharge-duration-first policy gradually
equalises over stores the discharge-durations~$E_i(t)/P_i$, thus
pushing the residual stored energy configuration towards a balanced
state as defined above. Additionally, under this policy, once the
discharge-durations within any set of stores have become equal they
remain equal for all subsequent times. Thus, when groups of stores as
defined above coalesce they remain coalesced, and further the (weak)
ordering of stores by their discharge-durations remains unchanged over
time.
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.6]{lgtf}
\caption{Greatest-discharge-duration-first policy for five stores.
The shaded areas correspond to the residual energies $E_i(t)$
within the stores at time~$t$. Stores are selected in descending
order of $E_i(t)/P_i$. (Thus the total shaded area is the area
under the burst-power profile at time~$t$, while the shaded area
above any level~$p$ defines the energy-power transform
$\epts{e}{p}{t}$ at time~$t$.) When $\bar d(t)$ is the energy to
be served, $\S_1$, $S_2$ and $\S_3$ are the sets of stores which
are fully, partially, or not utilised at time~$t$.}
\label{fig:lgtf}
\end{figure}
We shall say that a policy is \emph{greedy} if, at each successive
time~$t>0$, it serves as much as possible of the demand~$d(t)$ at that
time, i.e.\ if, under this policy,
\begin{displaymath}
\sum_{i\in\S}r_i(t) = \min\left(d(t),\,\sum_{i\in\S: E_i(t)>0}P_i\right),
\quad\text{for all $t\ge0$.}
\end{displaymath}
Note that there is a unique \emph{greedy
greatest-discharge-duration-first} (GGDDF) policy. This policy was
independently proposed in \cite{Nash1978} in the context of water
management, and in \cite{ETA-pscc, CZ2018} for the current context of
energy storage. A discrete-time algorithm implementing this policy
for a piecewise constant demand process is given in
\cite{ETA-tops}. Finally, we note that \cite{Zhu2019} introduced a
closely related discharge policy for the class of \emph{constant}
demand signals ($d(t)=d$, for $t \in [0,T]$ and $d(t)=0$ for $t > T$),
where $d$ (but not $T$) is known to the dispatcher in advance. For
this limited class of signals, the proposed policy is optimal in the
sense that it minimises unserved energy demand. It will be shown below
that the GGDDF\ policy is optimal in the same sense, but for a much
broader class of demand signals that includes constant signals as a
special case.
Suppose now that it may not be possible to serve the entire demand
process $(d(t),\,t\ge0)$ and that our objective is the minimisation of
the total unserved energy demand~\eqref{eq:4} over some time interval
$[0,T]$. Theorem~\ref{theorem:1} below is central to the rest of the
paper. It gathers together and provides a unified, economical and
accessible treatment of results previously obtained by various
authors. In particular the optimality of the GGDDF\ policy is
established in \cite{Nash1978, ETA-pscc, CZ2018},
while the implied necessary and sufficient condition for this policy
to be able to serve a given demand process is derived by
\cite{ETA-tosg, CZ2018}, and the expression for
the unserved energy demand under this policy is given by
\cite{ETA-tops}. In considering arbitrary time
intervals, the theorem also provides a very modest extension of these
results.
\begin{theorem}
\label{theorem:1}
For the given demand function $(d(t),\,t\ge0)$ and the given initial
energy configuration $(E_i(0),\,i\in\S)$, the total unserved energy
demand~(\ref{eq:4}) over any time interval $[0,T]$ is minimised by
the GGDDF\ policy, and this minimum is given by
\begin{equation}
\label{eq:11}
\max_{p\ge0}\bigl(\eptd{e}{0,T}{p} - \epts{e}{0}{p}\bigr).
\end{equation}
\end{theorem}
\begin{proof}
We show first that the use of the GGDDF\ policy results in unserved
energy as given by~\eqref{eq:11}. For each~$t\ge0$, let
$\bar d(t) \le d(t)$ be the demand actually served at time~$t$ under
the GGDDF\ policy. Similarly, for each such~$t$, let $E_i(t)$ be
the energy remaining in each store~$i\in\S$ at time~$t$ under the
GGDDF\ policy and let $\epts{e}{t}{p}$, $p\ge0$, be the
corresponding energy-power transform. Suppose that, at time~$t$,
stores in $\S$ are ranked in descending order of their
discharge-durations $E_i(t)/P_i$ (as is required for the
implementation of the GGDDF\ policy). Let $\pi(t)$ be the set of
values of~$p$ such that $p = \sum_{i=1}^j P_i$
for some~$j$ such that $E_{j+1}(t)/P_{j+1} < E_j(t)/P_j$, and
include also in the set~$\pi(t)$ the values $p=0$ and (if no store is
empty) $p=\sum_{i\in\S}P_i$. (See the example of
Figure~\ref{fig:epts} below.) It follows from the earlier
observation that once, under the GGDDF\ policy, the
discharge-durations of any two stores have become equal they remain
equal thereafter, that
\begin{equation}
\label{eq:12}
\pi(u) \supseteq \pi(t),
\quad \text{ for all $0 \le u \le t$.}
\end{equation}
Further, from~\eqref{eq:8} and from the definition of the GGDDF\
policy, it is readily checked that, under this policy, at each
time~$t$ and for all $p\ge0$, the derivative with respect to~$t$ of
the energy-power transform $\epts{e}{t}{p}$ of the residual energy
configuration satisfies
\begin{equation}
\label{eq:27}
\frac{d}{dt}(\epts{e}{t}{p}) \le \min(0, p - \bar d(t)),
\end{equation}
with equality for all~$p\in\pi(t)$. The latter equality is easily
seen, and the general result~(\ref{eq:27}) follows from the
observation that the above derivative varies linearly in~$p$ between
successive points of the set $\pi(t)$ (again see the example of
Figure~\ref{fig:epts}), while the function $\min(0,p - \bar d(t))$
is concave in~$p$. (Indeed we have equality in~(\ref{eq:27}) for
all values of~$p$ other than for those lying within that interval
which is bounded by two consecutive points of the set~$\pi(t)$ and
within which $\bar d(t)$ lies.) Similarly, from~\eqref{eq:9}, at
each time~$t$ and for all $p\ge0$, the derivative with respect
to~$t$ of the energy-power transform $\eptd{e}{t,T}{p}$ of the
residual demand is given by
\begin{equation}
\label{eq:28}
\frac{d}{dt}(\epts{e}{t}{p}) = \min(0, p - d(t)).
\end{equation}
Thus, from~(\ref{eq:27}) and~(\ref{eq:28}), under the GGDDF\
policy, at each time~$t$ and for all $p\ge0$,
\begin{equation}
\label{eq:13}
\frac{d}{dt}(\epts{e}{t}{p} - \eptd{e}{t,T}{p}) \le
\begin{cases}
d(t) - \bar d(t), & \qquad p < \bar d(t),\\
d(t) - p, & \qquad \bar d(t) \le p \le d(t),\\
0, & \qquad p > d(t),
\end{cases}
\end{equation}
with equality for all $p\in\pi(t)$.
It now follows from~\eqref{eq:13} that, under the GGDDF\ policy and
for all $p\ge0$, the unserved energy demand over the interval
$[0,T]$ is given by
\begin{align}
\int_0^T(d(t) - \bar d(t))\,dt
& \ge \epts{e}{T}{p}
- \epts{e}{0}{p} + \eptd{e}{0,T}{p} \label{eq:15}\\
& \ge \eptd{e}{0,T}{p} - \epts{e}{0}{p}, \label{eq:16}
\end{align}
where \eqref{eq:16} follows since necessarily
$\epts{ e}{T}{p} \ge 0$. It follows that the unserved energy demand
over the time interval $[0,T]$ is greater than or equal to
the expression given by (\ref{eq:11}). To prove equality, define
$\hat p = \min\{p\ge0 \colon \epts{ e}{T}{p} = 0\}$.
Note that $\hat p$ necessarily exists and that
$\hat p \in \pi(T)$; the latter follows, for example, from the linearity
of $\epts{e}{T}{p}$ between adjacent points of $\pi(T)$. Observe
that $\hat p \le \bar d(t)$ for all $t\in[0,T]$ such that
$\bar d(t) < d(t)$ (i.e.\ under the GGDDF\ policy there is unserved
demand at time~$t$); this follows since, for any such~$t$,
necessarily $\epts{ e}{t}{\bar d(t)}=0$ and so also
$\epts{ e}{T}{\bar d(t)}=0$. Observe also that, by~\eqref{eq:12},
$\hat p \in \pi(t)$ for all $t\in[0,T]$. It follows from the above
two observations, and by using~\eqref{eq:13}, that~\eqref{eq:15}
holds with equality for $p=\hat p$. Since also
$\epts{e}{T}{\hat p}=0$, the relation~\eqref{eq:16} also holds with
equality for $p = \hat p$. Hence the expression~(\ref{eq:11}) also
provides an upper bound on the unserved energy demand over the
interval $[0,T]$ under the GGDDF\ policy.
(Note that, combining this with the
earlier lower bound, it follows (i) that the quantity
$\eptd{e}{0,T}{p} - \epts{e}{0}{p}$ is maximised
for $p=\hat{p}$, and (ii) that when the demand function
can be fully served, we have $\eptd{e}{0,T}{\hat{p}} = \epts{e}{0}{\hat{p}}$.)
It now follows that, under the GGDDF\ policy,
the unserved energy demand is as given by~\eqref{eq:11}.
We now show that any nonnegative demand process $(d(t),\,t\ge0)$
which may be completely served over the interval $[0,T]$ by
\emph{some} policy, may be completely served over that interval by
the use of the GGDDF\ policy. To do this, it is sufficient to show
that the condition
\begin{equation}
\label{eq:17}
\epts{e}{0}{p} \ge \eptd{e}{0,T}{p}
\qquad\text{for all $p \ge 0$}.
\end{equation}
is necessary (as well as sufficient) for the demand process
$(d(t),\,t\ge0)$ to be capable of being completely served over the
interval $[0,T]$. For any $p\ge0$, let $\tau_p$ be the of the set
of times~$t$ within the interval $[0,T]$ such that $d(t)\ge p$. For
the demand process $(d(t),\,t\ge0)$ to be completely served over the
interval $[0,T]$, it is necessary that
\begin{equation}
\label{eq:18}
\int_0^{m(\tau_p)} s^0(u)\,du
\ge
\int_{\tau_p} d(u)\,du,
\end{equation}
where $m(\tau_p)$ is the total length of the set of times $\tau_p$.
That this is so follows since, from~\eqref{eq:7}, the integral on
the left side of~\eqref{eq:18} is the maximum amount of energy which
is capable of being served within any set of times of total length
$m(\tau_p)$. The relation~\eqref{eq:18} in turn implies that
\begin{displaymath}
\int_0^{m(\tau_p)} \max(0,s^0(u) - p)\,du
\ge
\int_{\tau_p} \max(0,d(u) - p)\,du,
\end{displaymath}
since $d(u)\ge p$ on the set~$\tau_p$. However, this is simply the
condition~\eqref{eq:17}.
Finally, to complete the proof of the theorem, suppose that the
nonnegative demand process $(d(t),\,t\ge0)$ is not necessarily
completely served over the interval $[0,T]$ by any policy. Consider
any policy which minimises the unserved energy demand over the
interval $[0,T]$ and let $(\hat d(t),\,t\ge0)$, with
$\hat d(t) \le d(t)$ for all~$t$ and $d(t)=0$ for $t>T$, be the
process of such demand as \emph{is} served over $[0,T]$ under that
policy. Then, by the result of the preceding paragraph (with
$\hat d(t)$ replacing $d(t)$ for all~$t$), the process
$(\hat d(t),\,t\ge0)$ may also be completely served over the
interval $[0,T]$ by the use of the GGDDF\ policy. The GGDDF\
policy therefore also minimises the unserved energy demand over that
interval associated with the original process $(d(t),\,t\ge0)$.
\end{proof}
Theorem~\ref{theorem:1} has the following immediate corollary, which
is fundamental in establishing the energy-power transform of a
stored-energy configuration as containing all the information as to
which future demand processes may be completely served.
\begin{corollary}
\label{corollary:1}
Any given demand process $(d(t), t\ge0)$ may be completely served
over any interval $[0,T]$ (i.e.\ the unserved energy
demand~(\ref{eq:4}) is zero) by a given energy configuration with
(initial) energy-power transform $(\epts{e}{0}{p},\,p\ge0)$ if and
only if $\epts{e}{0}{p} \ge \eptd{e}{0,T}{p}$ for all $p \ge 0$.
Under this condition the demand process is completely served by the
use of the GGDDF\ policy \cite{ETA-tosg}.
\end{corollary}
\begin{remark}\label{remark:1}
It follows from Corollary~\ref{corollary:1} and the properties of
the function $\epts{e}{0}{p}$ noted earlier that, for a given total
volume of stored energy $\sum_{i\in\S}E_i(0)$ at time~$0$, the set
of future demand processes which may be completely served is
maximised when the stored-energy configuration at time~$0$ is
\emph{balanced} as defined above, so that the corresponding
energy-power transform $\epts{e}{0}{p}$ decreases linearly in~$p$.
\end{remark}
The GGDDF\ policy has the important property of being
\emph{non-anticipatory}---as defined in
Section~\ref{sec:introduction}. It follows that the GGDDF\ policy
remains feasible within a stochastic environment, i.e.\ when, at each
successive time~$t$, the demand function $d(t')$ is known for times
$t'\le t$, but not necessarily for $t'>t$. Since, by
Theorem~\ref{theorem:1}. the GGDDF\ policy thus minimises
unserved energy demand for all possible evolutions of the
demand function, we have the following
further corollary to that theorem.
\begin{corollary}
\label{corollary:2}
Suppose that (in a stochastic environment) the objective for the
optimal serving of energy is the minimisation of the expectation of
the unserved energy demand~(\ref{eq:4}), or the
minimisation of any quantile of the distribution of
the unserved energy demand. Then this objective is
achieved by the use of the unique GGDDF\ policy \cite{ETA-tops,
CZ2018}.
\end{corollary}
\begin{corollary}
\label{corollary:2b}
Suppose that a possibly stochastic demand signal $(d(t),t\ge 0)$
cannot be completely served. In this case, consideration of the
truncated signals on $t \in [0,\tilde{T}]$, for $\tilde{T}\ge0$
shows that the GGDDF\ policy maximises the time until the storage
fleet is first unable to serve all demand \cite{ETA-pscc}.
\end{corollary}
\begin{example}\label{ex:ggddf}
We illustrate various features of the GGDDF\ and other policies
with a simple example, which is adapted to present needs from one
given by~\cite{CZ2018} and which is reasonably realistic in its
dimensioning. Consider a time interval $[0,4]$ and a demand process
$(d(t),\,t\in[0,4])$ given by
\begin{displaymath}
d(t) =
\begin{cases}
200, & \quad 0 \le t \le 2,\\
500, & \quad 2 < t \le 3,\\
100, & \quad 3 < t \le 4.
\end{cases}
\end{displaymath}
Consider also $5$ stores, initially full, each with the same rate
constraint $P_i=100$ and with values of $E_i$ given by $100$, $150$,
$200$, $200$, $250$ for $i = 1,\dots,5$. These stores are to be
used to serve as much as possible of the above demand process. If
time is measured in hours, power in MW (and so energy in MWh), then
this example might correspond to a modest level of shortfall over a
four-hour period in a country such as Great Britain, with the stores
corresponding to four moderately large batteries. It is readily
verified that the GGDDF\ policy empties all five stores over the
time period~$[0,3]$, serving all the demand during that time period,
and none of the demand during the remaining time period $[3,4]$,
thereby leaving (minimised) unserved energy demand of $100$. This
is as predicted by Theorem~\ref{theorem:1}---see
Figure~\ref{fig:epts}, which plots the energy-power transforms
$\epts{e}{0}{p}$ and $\eptd{e}{0}{p}$. Various other policies also
empty all the stores and hence minimise unserved energy demand. One
such is the GGDDF\ policy applied to the time-reversed demand
process, which serves all the demand except that occurring during
the period $[0,0.5]$ (during which none is served). A further such
policy is that which uses the greatest-discharge-duration-first
policy, not greedily, but rather to serve all demand in excess of a
given level~$k$. For $k=25$ this policy just serves all the demand
in excess of that level, again emptying the stores. However,
neither of the above two policies, viewed as algorithms, could be
implemented in a stochastic environment, as in each case the
decisions to be made at each successive time require a knowledge of
the demand process over the entire time period $[0,4]$.
\begin{figure}[!ht]
\centering
\vspace{-4ex}
\includegraphics[scale=0.7]{epts}
\vspace{-2ex}
\caption{Example~\ref{ex:ggddf}: \emph{energy-power transforms}
$\epts{e}{0}{p}$ and $\eptd{e}{0}{p}$ at time~$0$ of the
burst-power profile (of the initial stored-energy configuration)
and of the demand process $(d(t),\,t\ge0)$. The set $\pi(0)$
defined in the proof of Theorem~\ref{theorem:1} is given by
$\{0,100,300,400,500\}$ and is indicated on the horizontal axis.
Both $\epts{e}{0}{p}$ and $\eptd{e}{0}{p}$ are convex decreasing
functions of $p$ and the function $\epts{e}{0}{p}$ is linear
between successive points of the set $\pi(0)$. The value of the
expression~(\ref{eq:11}) is equal to 100
(the maximum being attained at $p=0$), which, from
Theorem~\ref{theorem:1} is the minimum unserved energy demand
over all policies.}
\label{fig:epts}
\end{figure}
Finally, the heuristic greedy policy studied in~\cite{ESDT} arranges
the stores in some order and completely prioritises the use of
earlier stores (with respect to that order) over later ones. It may
be checked that, with respect to the arrangement of the stores in
either ascending or descending order of capacity, the suggested
policy fails to empty the stores and hence fails to minimise
unserved energy demand.
\end{example}
\section{Cross-charging of stores}
\label{sec:disch-with-cross}
We continue to study the situation in which the demand process
$(d(t),\,t\ge0)$ is nonnegative over some time interval, which we
again take, without loss of generality, to be the entire positive
half-line, so that $d(t)\ge0$ for all $t\ge0$. We now allow that
stores may be charging as well as discharging, so that, for each
$i\in\S$, the rates $r_i(t)$ satisfy the rate
constraints~\eqref{eq:3}. However, we require also that the net
energy supplied by the stores in $\S$ is positive for all $t\ge0$ and
is used to satisfy, again as far as possible in some suitable sense,
the demand process $(d(t),\,t\ge0)$. Hence we require
\begin{equation}
\label{eq:19}
0 \le \sum_{i\in\S:r_i(t)\ge0}r_i(t) +
\sum_{i\in\S:r_i(t)<0}r_i(t)/\eta_i \le d(t),
\qquad t \ge 0,
\end{equation}
where, as previously defined, $0<\eta_i\le1$ is the round-trip
efficiency of each store $i\in\S$. This corresponds to the situation
in which stores may supply energy to each other---which we refer to as
\emph{cross-charging}---but in which no external energy is available
for the charging of stores. Our main aim in this section is to show
that, while such cross-charging may often assist in serving an
external demand, it is also possible to identify circumstances, of
considerable importance in practical applications, in which it does
not. We give first a simple example in which cross-charging is
helpful.
\begin{example}
\label{example:cc}
Consider two stores with capacity and rate constraints~\eqref{eq:1}
and~\eqref{eq:3} given by $(\overline E_1,P_1)=(2,2)$,
$(\overline E_2,P_2)=(4,1)$ and $P_i=P'_i$ for $i=1,2$, and assume
further that each store~$i$ has a round-trip efficiency~$\eta_i=1$.
We take the demand process $(d(t),\,t\ge0)$ to be given by $d(t)=3$
for $t\in[0,1]$ and $t\in[3,4]$ and $d(t)=0$ otherwise. Finally, we
assume that the two stores are full at time~$0$. Then it is
straightforward to see the only way in which the demand signal can
be completely served for all $t\ge0$ is to fully empty store~$1$ and
use one unit of energy from store~$2$ during the time period
$[0,1]$, to fully recharge the store~$1$ from store~$2$ during the
time period $[1,3]$, and then to fully discharge both stores during
the time period $[3,4]$.
\end{example}
In the above example there are initial and final periods of high
demand, requiring service from both stores, separated by a period of
low demand during which the low capacity store may be recharged from
the high capacity store. This is the typical situation in which
cross-charging may be useful. However, for round-trip efficiencies
which are less than one, such cross-charging is inherently wasteful of
energy. The following theorem is now fundamental.
\begin{theorem}
\label{theorem:2}
Suppose that the demand process $(d(t),t\ge0)$ is (weakly) decreasing
for all~$t\ge0$, and is such that it may be served, possibly with
cross-charging and subject to the conditions~(\ref{eq:19}), by the
given stored-energy configuration. Then it may also be served
without cross-charging (i.e.\ with $r_i(t)\ge0$ for all $i$ and for
all $t$) by the use of the GGDDF\ policy.
\end{theorem}
\begin{proof}
We prove the implication of the theorem by proving the
contrapositive: if the GGDDF\ policy (which does not permit
cross-charging) cannot serve the demand process $d(t)$, then neither
can any other policy, including those that make use of
cross-charging.
Clearly, if $d(0)>\sum_{i \in \S: E_i(0)>0} P_i$ (so that there is
insufficient power at time $t=0$) or if
$\int_0^{\infty} d(t)\mathrm{d}t > \sum_{i \in \S} E_i(0)$ (i.e.\
the total energy in all the stores is insufficient to meet the
demand process), then no policy is able to serve the demand process.
Hence, we consider only the remaining cases where the GGDDF\ policy
fails to fully serve the demand. In these cases, there must be a
first failure time $t=t^*$, characterised by the condition
\begin{equation}
\label{eq:29}
\sum_{i \in \S^*} P_i < d(t^*),
\end{equation}
where $\S^* = \{i \in \S: E_i(t^*)>0\}$ is the set of stores that are
not empty at time~$t^*$.
The GGDDF\ policy has the property that it preserves through time
the ordering of discharge-durations $E_i(t)/P_i$, except for
equalisations. Since the stores in $\S^*$ have the highest values
of $E_i(t^*)/P_i$ at time~$t^*$ (the other stores being empty at
that time), this must also have been the case for all times $t<t^*$.
Since, further, the demand process $(d(t),t\ge0)$ is (weakly)
decreasing, it now follows from~(\ref{eq:29}) that,
under the GGDDF\ policy, all the stores in the set~$\S^*$ have
served energy at their maximum rates (i.e.\ $r_i(t) = P_i$ for all
$i\in\S^*$) for all times~$t\in[0,t^*]$. Since also the stores
$i\notin\S^*$ are empty at time~$t^*$, it now follows that at
time~$t^*$ the GGDDF\ policy has extracted the maximum possible
amount of energy from the entire set of stores~$\S$ over
$t\in[0,t^*]$. Hence, under any policy (with or without
cross-charging) which succeeded in serving all demand over the time
interval $[0,t^*]$, the stores $i\notin\S^*$ would also be empty at
time~$t^*$, so that that policy would also fail at that time.
\end{proof}
Theorem~\ref{theorem:2} has the following companion result.
\begin{theorem}
\label{theorem:3}
Suppose that the demand function $(d(t),t\in[0,T])$ is \emph{(weakly)
increasing} on $[0,T]$, and that all stores are full at time~$0$.
Then if $(d(t),t\in[0,T])$ can be served, it can be served without
cross-charging (i.e.\ with $r_i(t)\ge0$ for all $i$ and for all
$t$), and with at least as much energy remaining in each store at
the final time~$T$.
\end{theorem}
\begin{proof}
We first consider the case where every store $i\in\S$ has a
round-trip efficiency $\eta_i=1$, so that there is no loss of energy
in cross-charging. Here the result can be deduced from
Theorem~\ref{theorem:2} by an argument involving time and space
reversal. Consider any policy for the use of the stores, possibly
involving cross-charging, which serves the given demand process
$(d(t),t\in[0,T])$. For each store $i\in\S$, let $E_i(t)$ be the
corresponding level of store~$i$ at each time~$t$; define a new
sequence of storage levels $(E^*_i(t),t\in[0,T])$ on the
interval~$[0,T]$ for the store~$i\in\S$ by
\begin{equation}
\label{eq:21}
E^*_i(t) = \overline E_i - E_i(T-t),
\qquad t \in [0,T].
\end{equation}
The set of such sequences over all $i\in\S$ corresponds to the use
of the stores, with the same input and output rate constraints and
again with no loss of energy in cross-charging, to serve fully a
demand process $(d^*(t),t\in[0,T])$ given by $d^*(t) = d(T-t)$ for
all $t$, with the initial level of each store~$i$ being given
(from~\eqref{eq:21}) by
$E^*_i(0) = \overline E_i - E_i(T) \le \overline E_i$ and with the
final level of every store being given by~$E^*_i(T) = 0$ (again
from~\eqref{eq:21} since, by the hypothesis of the theorem,
$E_i(0)=\overline E_i$ for all~$i$). Further, this pattern of use
of the stores involves cross-charging if and only if the use of the
original set of sequences $(E_i(t),t\in[0,T])$, for $i\in\S$, to
serve the original demand process $(d(t),t\in[0,T])$ similarly
involves cross-charging. Since the original demand process is
increasing, the demand process $(d^*(t),t\in[0,T])$ is decreasing
and so, by Theorem~\ref{theorem:2}, it may also be served fully,
without cross-charging, by a modified set of sequences of store
levels $(\hat E^*_i(t),t\in[0,T])$, $i\in\S$, with
$\hat E^*_i(0) = E^*_i(0)$ and $\hat E^*_i(T) = E^*_i(T) = 0$ for
all~$i$ (this last result following since the sets of processes
$(E^*_i(t),t\in[0,T])$, $i\in\S$, and $(\hat E^*_i(t),t\in[0,T])$,
$i\in\S$, both start at the same set of levels, serve the same total
volume of energy over the period $[0,T]$, and, since the former set
of processes fully empties the set of stores, so also must the
latter). Finally, transforming back in time and space, it follows
that set of sequences of store levels $(\hat E_i(t),t\in[0,T])$,
$i\in\S$, given by
\begin{displaymath}
\hat E_i(t) = \overline E_i - \hat E^*_i(T-t),
\qquad t \in [0,T],
\end{displaymath}
fully serves the original demand process without cross-charging, and
that $\hat E_i(0) = \overline E_i$ and $\hat E_i(T) = E_i(T)$ for
all $i\in\S$, as required, and indeed so that in this case the
modified process, with cross-charging eliminated, leaves exactly the
same volume of energy in each store at the final time~$T$.
We now consider the general case $\eta_i\le1$ for all $i\in\S$.
Again consider any policy for the use of the stores, possibly
involving cross-charging, which serves the given demand process
$(d(t),t\in[0,T])$. In particular the rates~$r_i(t)$ associated
with this policy satisfy~(\ref{eq:19}) with the second inequality in
that expression replaced with equality. As usual, we denote by
$E_i(t)$ the energy level in each store $i\in\S$ at each
time~$t\ge0$. Consider also a modified model in which the stores
and demand process remain the same, except only that the round-trip
efficiencies~$\eta_i$ are all replaced by one. It is clear that we
may choose a policy (possibly including cross-charging), i.e.\ a set
of rate functions $(\hat r_i(t),t\ge0),\,i\in\S$, for this modified
model in which again all demand is served, i.e.\
\begin{equation}
\label{eq:30}
\sum_{i\in\S}\hat r_i(t) = d(t), \qquad t\ge0,
\end{equation}
and in which, for each time~$t\ge0$, the corresponding energy levels
$\hat E_i(t)$ in the stores satisfy
\begin{equation}
\label{eq:25}
\hat E_i(t) \ge E_i(t), \qquad i\in\S.
\end{equation}
To see this observe that we may, inductively over time, choose the
rates $\hat r_i(t)$, and hence the store levels $\hat E_i(t)$ as
follows: at each time~$t$, for those $i$ such that $r_i(t) < 0$, set
$\hat r_i(t) = r_i(t)$ unless $E_i(t) = \overline E_i$ (store~$i$ is
full) in which case set $\hat r_i(t) = 0$; similarly, at each
time~$t$, for those~$i$ such that $r_i(t) \ge 0$, choose
$0 \le \hat r_i(t) \le r_i(t)$ and such that equation~(\ref{eq:30})
is satisfied for the modified model. Arguing inductively, it is
easy to see that, at each successive point in time, this modified
model preserves all rate and capacity constraints together with the
relation~(\ref{eq:25}). (Informally we might think of the modified
model as corresponding to the idea that the same external demand is
notionally served at the same rates from the same stores as in the
unmodified model, while the perfect round-trip efficiencies of the
stores in the modified model enable cross-charging to be used to
ensure that, at all times, the relation~(\ref{eq:30}) is satisfied
and all store levels are greater than or equal to their levels in
the unmodified model.) By the result of the theorem already
proved for the case of perfect round trip efficiency, the above
policy may now be further modified so as to serve the same demand
process while eliminating cross-charging and leaving the final
volume of energy (at time~$T$) in each store $i\in\S$ unchanged by
this elimination. Since no cross-charging is now taking place, this
further modified policy now also serves the demand in the original
system with $\eta_i\le1$ for all~$i$, again with the same final
volume of energy in each store at time~$T$, and this volume is
therefore greater than or equal to that which was present in the
original model when cross-charging was used.
\end{proof}
The need, in Theorem~\ref{theorem:3}, for some condition such as the
requirement that all stores are full at time~$0$ is shown by
Example~\ref{example:cc2} below.
\begin{example}
\label{example:cc2}
Consider two stores with capacity and rate constraints~\eqref{eq:1}
and~\eqref{eq:3} given by $(\overline E_1,P_1)=(2,1)$,
$(\overline E_2,P_2)=(1,1)$, with $P_i=P'_i$ for $i=1,2$, and with
round-trip efficiencies~$\eta_i=1$ for $i=1,2$. Consider a time
horizon $T=2$ and let the demand process $(d(t),\,t\in[0,2])$ to be
given by $d(t)=0$ for $t\in[0,1]$ and $d(t)=2$ for $t\in[1,2]$.
Finally, assume that, at time~$0$, store~$1$ is full while store~$2$
is empty. Then it is easy to see the only way in which the demand
signal can be completely served for all $t\in[0,2]$ is to fully
charge store~$2$ from store~$1$ during the time period $[0,1]$ and
to then fully discharge both stores during the time period $[1,2]$.
Hence in this case the need for cross-charging cannot be dispensed
with.
\end{example}
Theorems~\ref{theorem:2} and \ref{theorem:3} have the following
corollary.
\begin{corollary}
\label{corollary:3}
Suppose that the demand process $(d(t),t\in[0,T])$ is (weakly)
increasing on $[0,T']$ and (weakly) decreasing on $[T',T]$ for some
$0\le T'\le T$, and that all stores are full at time~$0$. Then if
$(d(t),t\in[0,T])$ can be served, it can be served without
cross-charging, and by the use of the GGDDF\ policy.
\end{corollary}
\begin{proof}
This is an application of Theorem~\ref{theorem:3} for the period
$[0,T']$---including the result that any cross-charging may be
eliminated in this period without reducing the volume of energy in
each store at time~$T'$---followed by the use of
Theorem~\ref{theorem:2} for the period $[T',T]$ (since clearly any
increase in the volume of energy in each store at the time~$T'$ due
to the elimination of cross-charging in $[0,T']$ means that we may
continue to serve the given demand process on the interval $[T',T]$.
Finally, since no cross-charging is necessary, it follows from
Theorem~\ref{theorem:1} that the demand $(d(t),t\in[0,T])$ can be
served by the use of the GGDDF\ policy.
\end{proof}
The following further corollary is an immediate application of
Theorem~\ref{theorem:1} and Corollary~\ref{corollary:3}.
\begin{corollary}
\label{corollary:4}
Under the conditions of Corollary~\ref{corollary:3}, and in a
possibly stochastic environment, the minimisation of the expectation
of the unserved energy demand~(\ref{eq:4}), or the minimisation of
any quantile of the distribution of the unserved energy demand, is
achieved by the use of the GGDDF\ policy.
\end{corollary}
An important application of the above result is to the frequently
occurring case where stores may be fully recharged overnight, and
there is a single period of shortfall during the day which is
\emph{unimodal} in the sense that it is monotonic increasing and then
decreasing as in the statement of Corollary~\ref{corollary:3}. One
may assume that there is no surplus energy available for charging any
of the stores during this period. Then, in a possibly stochastic
environment, the expectation of the unserved energy
demand~(\ref{eq:4}) is minimised by the use of the unique GGDDF\
policy.
Finally in this section, and for completeness, Theorem~\ref{theorem:4}
below gives a useful variation of Theorem~\ref{theorem:3}. As should
be clear from the statement of Theorem~\ref{theorem:3} itself, the
condition of that theorem that ``all stores are full at time~$0$'' may
be relaxed subject to the additional restriction of some minimum level
on the demand function. Theorem~\ref{theorem:4} makes this idea
precise. It does of course also have corollaries analogous to
Corollaries~\ref{corollary:3} and~\ref{corollary:4} above.
\begin{theorem}
\label{theorem:4}
Consider any initial stored-energy configuration $(E_i(0), i\in\S)$.
Let $\S'=\{i \in \S: E_i(0) < \overline{E}_i\}$ and define
$u_0=\min_{i \in \S'}(E_i(0)/P_i)$, with $u_0=\infty$ if all stores
in $\S$ are initially full. Let $\S_1$ be the set of stores~$i$
such that $\overline E_i/P_i \le u_0$. (Note that from the
definition of $u_0$ all the stores in the set~$\S_1$ are necessarily
full.) Define also $\S_2=\S\setminus\S_1$ and let
$k=\sum_{i\in\S_2}P_i$. Suppose that the demand function
$(d(t),t\in[0,T])$ is (weakly) increasing, further satisfies
$d(0)\ge k$, and may be served possibly with the use of
cross-charging. Then it may also be served without cross-charging.
\end{theorem}
\begin{proof}
Consider a modified set of store capacities $(\hat E_i, i\in\S)$
given by
\begin{equation}
\label{eq:22}
\hat E_i =
\begin{cases}
\overline E_i, & \quad i \in \S_1,\\
E_i(0) + T' P_i, & \quad i \in \S_2,
\end{cases}
\end{equation}
where the constant~$T'=\max_{i\in\S_2}(\overline E_i - E_i(0))/P_i$.
Observe that these modified store capacities are all at least as
great as the original capacities. Let the demand function
$(d(t),t\in[0,T])$ be as in the statement of the theorem. Extend
the time interval $[0,T]$ to $[-T',T]$ and consider the demand
function $(\hat d(t),t\in[-T',T])$ on this latter interval given by
\begin{equation}
\label{eq:23}
\hat d(t) =
\begin{cases}
k, & \quad t \in [-T',0],\\
d(t) & \quad t \in [0,T].
\end{cases}
\end{equation}
If the modified set of stores are considered to be full at
time~$-T'$ then their energy content is sufficient to serve the
demand function $(\hat d(t),t\in[-T',T])$, again possibly with the
use of cross-charging. To see this, observe that the stores in the
set~$\S_2$ may be utilised at their full rates to directly serve the
demand $\hat d(t)=k$ on the interval $[-T',0]$ (as would be the case
with the use of the GGDDF\ algorithm on that interval); at time~$0$
the remaining energy in each store~$i$ is then the original energy
content $E_i(0)$ of that store in the initial energy configuration
of the theorem; the demand $\hat d(t)=d(t)$ on the remaining
interval $[-T',T]$ may now be served as in the statement of the
theorem (possibly with cross-charging, since the modified stores are
all at least as large as the original stores). Since the demand
function $(\hat d(t),t\in[-T',T])$ is, by construction, increasing
on $[-T',T]$, it follows from Theorem~\ref{theorem:3} that this
demand function may also be served by the modified stores without
cross-charging, and in particular by the use of the GGDDF\ algorithm.
Since the residual energy content of these stores at time~$0$ is
again simply the original energy content of the original stores at
time~$0$, the conclusion of the theorem now follows.
\end{proof}
The need, in Theorem~\ref{theorem:4}, for some condition such as the
requirement that the demand function $(d(t),\,t\ge0)$ has some minimum
level at time~$0$ is again shown by the earlier
Example~\ref{example:cc2}. The issue here is essentially the same as
in Theorem~\ref{theorem:3}: informally, for a weakly increasing demand
function, it is the combination of an initially sufficiently low level
of demand and spare capacity in the stores at time~$0$, which enables
cross-charging to assist in fully serving the demand function where,
in the absence of such cross-charging, this might not be possible.
\section{Charging and discharging}
\label{sec:charging-discharging}
We now allow that, for every time $t\ge0$, both the demand $d(t)$ and
the rates~$r_i(t)$, $i\in\S$, may be arbitrary (in particular may take
either sign) subject to the constraints given
by~\eqref{eq:1}--\eqref{eq:3} and the condition
\begin{equation}
\label{eq:26}
\min(d(t),0) \le \sum_{i\in\S:r_i(t)\ge0}r_i(t) +
\frac{1}{\eta}\sum_{i\in\S:r_i(t)<0}r_i(t) \le \max(d(t),0),
\qquad t \ge 0,
\end{equation}
where, throughout this section, we assume that all stores $i\in\S$
have the same round-trip efficiency $\eta_i=\eta$. (For some
discussion of the case where the stores have different round-trip
efficiencies, see below and also Section~\ref{sec:conc}.) Thus, when
the demand $d(t)\ge0$ the stores collectively serve energy (possibly
with cross-charging) to assist in meeting that demand and when
$d(t)<0$, corresponding to a surplus of energy external to the stores,
some of that surplus may be supplied to the stores (again possibly
with cross-charging). Our objective continues to be to manage the
stores so as to minimise, over those times~$t$ such that $d(t)\ge0$,
the long-term unserved energy demand given by~\eqref{eq:4}, or, in a
stochastic environment, the expectation of this quantity.
For any continuous period of time $[T,T']$ over which $d(t)\le0$
(energy may be supplied to charge stores), there is now a theory which
is fully analogous to that developed in
Section~\ref{sec:pure-discharge-model} for discharging stores. In
particular we may define the \emph{(residual) charge-duration} of any
store $i\in\S$ at any time~$t$ as $(\overline{E}_i - E_i(t))/P'_i$
(the time that would be required to fully charge the store if this was
done at the maximum possible rate). Further the GGDDF\ policy for
discharging-only which is optimal in the sense of
Theorem~\ref{theorem:1} is replaced by the analogous \emph{greedy
greatest-charge-duration-first} (GGCDF) policy, which, in the
absence of cross-charging, is similarly optimal for attempting to
accept as much charge as possible over the interval $[T,T']$. (Note
that this would no longer be true in the absence of a common
round-trip efficiency for the stores. While one might, in this case,
continue to formulate a result analogous to Theorem~\ref{theorem:1} by
redefining store capacities and input rates in terms of external
energy input rather than output, the results of the present section
would not then continue to hold.)
In many circumstances a reasonable \emph{heuristic} policy for the
management of the stores is given by the use of the GGDDF\ policy to
serve as much of the demand process as possible during periods when
that process is positive, and the use of the GGCDF\ policy to charge
the stores as rapidly as possible during periods when the demand
process is negative. This policy again has the attractive property,
discussed in Section~\ref{sec:pure-discharge-model}, of being
non-anticipatory, so that it continues to be fully implementable in a
stochastic environment. As in Section~\ref{sec:disch-with-cross},
this policy may be expected to work particularly well when continuous
periods of storage discharge are separated by lengthy periods
providing ample time for recharging. However, in general, it need not
always be \emph{optimal} for the minimisation of long-term unserved
energy. The reason for this is that the GGDDF\ policy attempts to
equalise as quickly as possible the discharge-durations of the
stores, while the GGCDF\ policy attempts to equalise as quickly as
possible the corresponding charge-duration times, and these two
objectives in general conflict with each other. However, we do have
the following theorem.
\begin{theorem}
\label{theorem:5}
Suppose that a set of stores~$\S$ is such that $\overline E_i/P_i$
is the same for all~$i\in\S$, that $P'_i=\alpha P_i$ for
all~$i\in\S$ for some $\alpha>0$, and that the stores in~$\S$ have a
common round-trip efficiency $\eta$. Suppose also that $E_i(0)/P_i$
is the same for all~$i\in\S$.
Then an optimal policy for the minimisation of the expectation of
the unserved energy demand~(\ref{eq:4}) over any subsequent time
period $[0,T]$ is given by the use of the GGDDF\ policy at those
times~$t$ such that $d(t)$ is positive, and the use of the GGCDF\
policy at those times~$t$ such that $d(t)$ is negative.
\end{theorem}
\begin{proof}
Under the conditions of the theorem, at any time~$t$, balance with
respect to charging ($E_i(t)/P_i$ constant over $i\in\S$) is
equivalent to balance with respect to discharging
($(\overline E_i - E_i(t))/P'_i$ constant over $i\in\S$). The
initial stored-energy configuration is balanced, and so, since the
stores have a common round-trip efficiency~$\eta$, the
GGDDF/GGCDF\ policy maintains a balanced stored-energy
configuration for all $t\ge0$. It now follows from the results of
Section~\ref{sec:pure-discharge-model}---see in particular
Remark~\ref{remark:1}---that any demand process $(d(t),\,t\ge0)$
which may be completely managed (i.e. completely served at times~$t$
such that $d(t)\ge0$ and completely utilised for charging at
times~$t$ such that $d(t)<0$) under some policy may similarly be
completely managed by the use of the GGDDF/GGCDF\
policy. The conclusion of the theorem now follows as for
Corollary~\ref{corollary:2}.
\end{proof}
\begin{remark}\label{remark:2}
The essence of Theorem~\ref{theorem:5} is that, under the conditions
on device parameters given by the theorem, once stores become
balanced with respect to their initial energy configurations, they
remain so thereafter and may then be used with the same flexibility
as a single large store. It follows as in
Section~\ref{sec:pure-discharge-model} that the GGDDF/GGCDF\
policy drives any initial energy configuration towards such a
balanced state. However, cross-charging may speed such convergence.
\end{remark}
\section{Concluding remarks}
\label{sec:conc}
The present paper has considered the optimal or near optimal
scheduling of heterogeneous storage resources for the ongoing
balancing of electricity supply and demand. The results, which hold
in both deterministic and stochastic environments, are particularly
applicable when storage is used to cover periods of energy shortfall,
such as periods of daily peak demand, and may be completely or mostly
recharged between such periods. However, in future years storage may
also be used to balance electricity systems over much longer
timescales, such as between summer and winter. The simultaneous
existence of multiple timescales, together with the physical
characteristics of such storage types as are available, is likely to
lead to a very heterogeneous storage fleet, in which, in particular,
round-trip efficiencies vary considerably. The optimal dimensioning
and control of such storage presents significant further research
challenges.
Another topic of ongoing research is to extend the storage dispatch
methodology beyond a scenario with a central dispatcher and develop,
for example, hierarchical schemes for aggregation and disaggregation
of energy units.
\section*{Acknowledgements}
\label{sec:acknowledgements}
The authors are grateful to the Isaac Newton Institute for
Mathematical Sciences in Cambridge and to National Grid plc for their
support for the programme in which much of the current work was
carried out. They also thank Andy Philpott for some helpful
literature discussion. Finally, the authors are especially grateful
to the reviewers for their most careful and thoughtful readings of the
paper and helpful comments and suggestions.
|
\section{Introduction }
\subsection{Statement of main results}
\label{intro}
Consider an elliptic curve in the form of Weierstrass
\begin{equation}\label{W}
u^2=4x^3-g_2x-g_3,\quad g_2^3-27g_3^2\neq 0.
\end{equation}
{\em Lam\'e equation} (algebraic version) is the second order differential
equation
\begin{equation}\label{lame1}
\left(\left( u\frac{ d}{dx}\right)^2-m(m+1)x-\lambda\right)w=0,
\end{equation}
with respect to a function $w(x)$, where $m\geq 0$ is an integer. The more familiar form of the Lam\'e equation is
\begin{equation}\label{lame2}
\frac{d^2W}{dz^2}-\left(m(m+1)\wp(z)+\lambda\right) W=0,
\end{equation}
which is obtained from (\ref{lame1}) by the change of the
independent variable $x=\wp(z),\; u=\wp'(z)$, $W(z)=w(\wp(z))$.
Here $\wp$ is the Weierstrass
function of the lattice $\Lambda$ with invariants
$$g_2=60\sum_{\omega\in\Lambda\backslash\{0\}}\omega^{-4},
\quad g_3=140\sum_{\omega\in\Lambda\backslash\{0\}}\omega^{-6}.$$
Changing the variables in (\ref{lame1}) to $x_1=kx,\; w_1(x_1)=w(x_1/k)$,
we obtain a new equation (\ref{lame1}) with parameters
\begin{equation}\label{scaling}
(k\lambda,k^2g_2,k^3g_3),\quad k\in\mathbf{C}^*.
\end{equation}
Two equations obtained by such a change of the variables
are called {\em equivalent.}
The set of equivalence
classes is the {\em moduli space for Lam\'e equations}, $\mathrm{Lame}_m$.
The quotient of $\mathbf{C}^3\backslash\{0\}$ by the $\mathbf{C}^*$ action
(\ref{scaling}) is the weighted projective space $\mathbf{CP}(1,2,3)$, see
for example
\cite{H}, and
$\mathrm{Lame}_m$ is obtained from it by removing the curve $g_2^3-27g_3^2=0$.
It parametrizes projective structures on tori with one conic singularity
with angle $2\pi(2m+1)$. For a general
discussion of projective structures
with conic singularities we refer to \cite{Loray}.
$\mathrm{Lame}_m$ has two singularities corresponding to two points with non-trivial
stabilizers: $(0,1,0)$ and $(0,0,1)$, which are associated with
square and hexagonal
tori in (\ref{W}).
The homogeneous function
$$J=\frac{g_2^3}{g_2^3-27g_3^2}$$
is called the {\em absolute invariant} of the elliptic curve (\ref{W});
it defines a map $\pi_m:\mathrm{Lame}_m\to\mathbf{C}_J$ which is called the
{\em forgetful map}.
A {\em Lam\'e function} is a non-trivial solution $w$ of (\ref{lame1}) with
the property that $w^2$ is a polynomial in $x$. It is easy to see
that the degree of this polynomial must be $m$.
For given
$g_2,g_3$, such solutions exist if and only if the {\em accessory parameter}
$\lambda$ satisfies
an algebraic equation
\begin{equation}
\label{F}
F_m(\lambda,g_2,g_3)=0,
\end{equation}
where $F_m$ is a polynomial in three variables, monic with
respect to $\lambda$,
which is invariant under the $\mathbf{C}^*$ action (\ref{scaling}).
For each $\lambda$ satisfying (\ref{F}), the Lam\'e function
is unique up to a constant factor, \cite[23.41]{WW}.
Two Lam\'e functions
are {\em equivalent} if $w(x)=cw_1(kx),\; c,k\in\mathbf{C}^*$, and the set of equivalence
classes is the {\em moduli space for Lam\'e functions} $\mathbf{L}_m\subset\mathrm{Lame}_m$.
We consider this space $\mathbf{L}_m$ as an abstract Riemann surface which is obtained
by taking the quotient of the surface
$$\{(\lambda,g_2,g_3)\in\mathbf{C}^3:F_m(\lambda,g_2,g_3)=0,\; g_2^3-27g_3^2\neq 0\}$$
by the $\mathbf{C}^*$ action~(\ref{scaling}).
It is the normalization of the curve defined by equation (\ref{F})
in $\mathbf{CP}(1,2,3)\backslash\{ g_2^3-27g_3^2=0\}$.
Since $F_m$ in (\ref{F}) is a monic polynomial, the forgetful map $\mathbf{L}_m\to\mathbf{C}_J$
is proper.
Equations (\ref{F}) for $m\leq 8$ are
explicitly written
in Table 3 in \cite{Maier}, see also Appendix to the present paper.
Maier calls $F_m$ the
{\em Lam\'e spectral polynomials}.
Lam\'e equation and Lam\'e functions
have long history going back to the work of Gabriel Lam\'e
\cite{Lame1,Lame2}, and they have been intensively studied ever since,
because of their
importance for mathematical physics. Good reference
for the classical work is \cite{WW},
and a modern survey is contained in
the first three sections of \cite{Maier}.
Most of the classical work on Lam\'e functions was
concentrated on the real case with positive discriminant
($g_2,g_3,\lambda$ are real and $g_2^3-27g_3^2>0$), and we are not aware of any
systematic study of
general properties of $F_m$ and~$\mathbf{L}_m$.
In this paper we determine the topology of the Riemann surfaces $\mathbf{L}_m$.
\def\mathbf{N}{\mathbf{N}}
To state our main result,
we recall the notion of a $2$-dimensional orbifold
\cite{Milnor},\cite{Thurston}.
It is a compact Riemann surface $S$ with a function
$n:S\to\mathbf{N}_{\geq 1}\cup\{\infty\}$,
where $\mathbf{N}_{\geq 1}$ is
the set of positive integers, such that $n(x)=1$ for all points, except
finitely many {\em orbifold points} of order $n(x)>1$.
Points with $n(x)=\infty$ are interpreted as
punctures.
For example, the moduli space for tori $\mathbf{C}_J$ has a natural orbifold
structure with $S=\mathbf{\overline{C}}$ and
three orbifold points:
$n(0)=3$, $n(1)=2$, and $n(\infty)=\infty$. $J=0$ corresponds to
the hexagonal torus, and $J=1$ to the square torus.
The orbifold Euler characteristic is defined as
$$\chi^O=\chi(S)-\sum_{x\in S} \left(1-1/n(x)\right),$$
where $\chi(S)=2-2g(S)$ is the ordinary Euler characteristic of
the underlying compact surface,
and $g$ is its genus. Here we follow \cite{Thurston}.
Notice that in \cite{EMP} a different definition
is used: the Euler characteristic in \cite{EMP} is smaller
by a factor of $2$.
A ramified covering $f:S_1\to S_2$ is called an {\em orbifold map}
if $n(f(x))$ divides $n(x)\deg_x(f)$ for all $x\in S_1$, and an
{\em orbifold covering} if
\begin{equation}\label{ocov}
n(f(x))=n(x)\deg_x f\quad\mbox{for all}\quad x\in S_1.
\end{equation}
If $f$ is an orbifold covering, the Riemann-Hurwitz formula
gives
\begin{equation}\label{rh}
\chi^O(S_1)=\deg(f)\chi^O(S_2).
\end{equation}
We introduce the following functions of non-negative integers $m$.
\begin{eqnarray}\label{degFI}
d_m^I&:=&\left\{\begin{array}{ll}m/2+1,& m\equiv0\;
(\mathrm{mod\,} 2)\\
(m-1)/2,& m\equiv 1\; (\mathrm{mod\,} 2),\end{array}\right.\\
\label{degFII}
d_m^{II}&:=&3\lceil m/2\rceil.
\end{eqnarray}
It is easy to see that $d_m^I$ and $d_m^{II}$ always have
opposite parity: when $m\in\{0,3\}\;(\mathrm{mod\,} 4),$ $d_m^I$ is odd
and $d_m^{II}$ is even; when $m\in\{1,2\}\;(\mathrm{mod\,} 4),$ $d_m^I$ is even
and $d_m^{II}$ is odd.
\begin{eqnarray}\label{epsilon0}
\epsilon_0&=&\left\{\begin{array}{ll}0,&\mbox{if}\; m\equiv1\;(\mathrm{mod\,} 3),\\
1&\mbox{otherwise,}\end{array}\right.\\ \label{epsilon2}
\epsilon_1&=&\left\{\begin{array}{ll}0,&\mbox{if}\; m\in\{1,2\}\;(\mathrm{mod\,} 4),\\
1&\mbox{otherwise.}\end{array}\right.
\end{eqnarray}
One can restate the definitions of $\epsilon_j$ as follows:
$\epsilon_0=0$ if and only if $d_m^I$ is divisible by $3$,
and $\epsilon_1=0$ if and only if $d_m^I$ is even.
Our main result is the following.
\begin{theorem}\label{theorem1}
When $m\geq 2$, $\mathbf{L}_m$ is a Riemann surface consisting
of two connected components which we call $\mathbf{L}_m^I$
and $\mathbf{L}_m^{II}$, while
$\mathbf{L}_0$ and $\mathbf{L}_1$ are connected: $\mathbf{L}_0=\mathbf{L}_0^I$ and
$\mathbf{L}_1=\mathbf{L}_1^{II}$.
The Riemann surface $\mathbf{L}_m$ has a natural orbifold structure with $\epsilon_0$
orbifold points of order 3 on $\mathbf{L}_m^I$, and one orbifold point
of order 2 which belongs to $\mathbf{L}_m^I$ when $\epsilon_1=1$ and
to $\mathbf{L}_m^{II}$ otherwise.
The component $\mathbf{L}_m^I$ has $d_m^I$ punctures, and the component $\mathbf{L}_m^{II}$ has
$2d_m^{II}/3=2\lceil m/2\rceil$
punctures.
The restrictions of the forgetful map to these components are orbifold maps
and their degrees are
$d_m^I$ and $d_m^{II}$.
The orbifold Euler characteristics are
\begin{equation}\label{euler}
\chi^O(\mathbf{L}_m^I)=-(d_m^I)^2/6\quad\mbox{and}\quad\chi^O(\mathbf{L}_m^{II})=
-(d_m^{II})^2/18.
\end{equation}
\end{theorem}
{\bf Remark.}
Ordinary Euler characteristics $\chi$ are
obtained from the $\chi^O$ by adding the {\em orbifold corrections}
which in our case are
\begin{equation}\label{ocorr}
E^I=(4\epsilon_0+3\epsilon_1)/6\quad\mbox{and}\quad
E^{II}=(1-\epsilon_1)/2.
\end{equation}
Euler characteristics can be expressed as functions of $m$
rather than $d$, see Appendix.
It is well known that for $m\geq 2$ each polynomial $F_m$ factors
into four factors in
$\mathbf{C}[\lambda,e_1,e_2,e_3]$, where $e_i$ are related to $g_2,g_3$ by the equation
$$4x^3-g_2x-g_3=4(x-e_1)(x-e_2)(x-e_3),$$
see for example \cite[Thm. 2]{T} for an explicit statement of this.
However there is no discussion of irreducibility of these four factors
in the literature. Our theorem says that for $m\geq 2$, $F_m$ has exactly
two irreducible factors in $\mathbf{C}[\lambda,g_2,g_3]$ and implies that
the four factors of $F_m$ in $\mathbf{C}[\lambda,e_1,e_2,e_3]$ are irreducible.
Theorem \ref{theorem1} implies
the formulas for the genera of $\mathbf{L}_m^K,\; K\in\{ I,II\}$ in terms of $m$
or $d_m^K$ which are given in the Appendix.
We give several applications of Theorem \ref{theorem1}.
As a first application, we prove that the two irreducible
components of the surface $F_m(\lambda,g_2,g_3)=0$ in (\ref{F})
have no singularities in $\mathbf{C}^3$ except the lines $(0,t,0)$ and $(0,0,t)$.
To obtain a non-singular curve in $\mathbf{CP}^2$ parametrizing $\mathbf{L}_m$, we use
Legendre's family of elliptic curves
\begin{equation}\label{le}
v^2=z(z-1)(z-a),\quad a\in\mathbf{C}_a:=\mathbf{C}\backslash\{0,1\}.
\end{equation}
For the $J$-invariant of this curve we have
\begin{equation}\label{Ja}
J=\psi(a)=\frac{4}{27}\frac{(a^2-a+1)^3}{a^2(1-a)^2}.
\end{equation}
If $\mathbf{\overline{C}}_J$ is considered as an orbifold with $n(0)=3$, $n(1)=2$,
$n(\infty)=\infty$,
and in $\mathbf{\overline{C}}_a$ we set
$n(a)=\infty$ for $a\in\{0,1,\infty\}$, and $n(a)=1$ otherwise,
then $a\mapsto \psi(a)$ is an orbifold
covering.
The form of the Lam\'e equation corresponding to (\ref{le}) is
\begin{equation}\label{legendre}
Py''+\frac{1}{2}P'y'-((m(m+1)z+B)y=0,\quad P(z)=4z(z-1)(z-a),
\end{equation}
where the accessory parameter
$B$ is an affine function of $\lambda$ (see Section \ref{nons} for the
details of this transformation). Lam\'e functions correspond to non-trivial
solutions $y$ of (\ref{legendre}) such that
$y^2$ is a polynomial in $z$ of degree $m$. For such a solution to exist, a
polynomial equation
\begin{equation}\label{H}
H_m(B,a)=0
\end{equation}
must be satisfied. The Riemann surface defined by this equation
will be denoted by $\mathbf{H}_m$. It is the normalization of
the algebraic curve $H_m(B,a)=0$ in $\mathbf{C}_a\times\mathbf{C}_B$.
For $m\geq 2$, we will show that
it consists of four irreducible components,
$\mathbf{H}_m^j,\; j\in\{0,1,2,3\}.$
These components are defined as follows: $\mathbf{H}_m^0$
corresponds to polynomial solutions $y$ when $m$ is even and
to solutions of the form $y=Q\sqrt{P}$ when $m$ is odd. The other
three components $j\in\{1,2,3\}$ correspond to solutions
$$Q(z)\sqrt{z},\quad Q(z)\sqrt{z-1},\quad Q(z)\sqrt{z-a},\quad
\mbox{when}\quad m\quad\mbox{is odd},$$
and
$$Q(z)\sqrt{(z-1)(z-a)},\quad Q(z)\sqrt{z(z-a)},\quad Q(z)\sqrt{z(z-1)},$$
when $m$ is even,
where $Q$ is a polynomial. The forgetful maps $\sigma_m^j:\mathbf{H}_m^j\to\mathbf{C}_a$
are defined by $(B,a)\mapsto a$.
The polynomial $H_m$ is a product of four
factors $H_m^j$, and we have ramified coverings
$\Psi_m^0:\mathbf{H}_m^0\to \mathbf{L}_m^I,$
and $\Psi_m^j:\mathbf{H}_m^j\to\mathbf{L}_m^{II}$, $j\in\{1,2,3\}$
such that
$$\pi_m^K\circ\Psi_m^j=\psi\circ\sigma_m^j,$$
where $\pi_m^K:\mathbf{L}_m^K\to\mathbf{C}_J$ and $\sigma_m^j:\mathbf{H}_m^j\to\mathbf{C}_a$
are the forgetful maps, and $\psi$ is the
function (\ref{Ja}).
We will show that these maps $\Psi_m^j$, are
orbifold coverings with respect to the appropriate orbifold structures
defined on some compactifications of $\mathbf{L}_m^K$ and $\mathbf{H}_m^j$.
This will permit us to compute the genera of
components of $\mathbf{H}_m$ via the
Riemann--Hurwitz formula. Once the genera and degrees are known
one can conclude that these curves are non-singular by the
``genus--degree formula'':
\begin{equation}\label{gd}
g\leq(d-1)(d-2)/2,
\end{equation}
where we have equality only for non-singular curves.
We consider compactifications $\overline{\mathbf{H}}_m^j$ obtained
from $\mathbf{H}_m^j$ by filling in the punctures. Equivalently,
$\overline{\mathbf{H}}_m^j$ is the normalization of the algebraic curve
obtained as the projective closure of the zero set of equation
$H_m^j=0$.
\begin{theorem}\label{theorem2}
The maps $\overline{\mathbf{H}}_m^j\to\mathbf{CP}^2$ for $j\in\{0,\ldots,3\}$
are non-singular embeddings, in particular $\overline{\mathbf{H}}_m^j$
are irreducible. The
degrees of the ramified coverings $\Psi_m^j$
are $6$ for $j=0$ and $2$ for $j\in\{1,2,3\}$.
\end{theorem}
\def\mathbf{Q}{\mathbf{Q}}
So we have interesting sequences of non-singular planar curves
$\overline{\mathbf{H}}^j_m$
defined over $\mathbf{Q}$ for which
degrees and genera have been explicitly determined. Only a few
such examples of high genus are known to the authors,
see~\cite{Bries}, \cite{MO}.
One can deduce Theorem~\ref{theorem1} from Theorem~\ref{theorem2}:
if we know that $\overline{\mathbf{H}}_m^j$ are non-singular,
we can find their genera from the equality in (\ref{gd}) and obtain all
topological characteristics of $\mathbf{L}_m$ using the orbifold coverings
$\Psi_m^j$.
\begin{cor}\label{corollary1}
All singularities of
irreducible components of
surfaces (\ref{F}) are contained in the
lines $(0,t,0)$ and $(0,0,t)$.
\end{cor}
Computation for $m\leq 6$ shows that the only singularities
of surfaces (\ref{F}) are the zeros of the discriminant, but we do
not prove this in this paper.
These results allow us to prove a conjecture of Maier about degrees
of Cohn's polynomials \cite[Conj. 3.1(ii)]{Maier}. We recall the definition.
Let $F_m=F_m^IF_m^{II}$ be the irreducible
factorization. Let $D_m^K$ be the discriminant of $F_m^K$ with respect to
$\lambda$. Then $D_m^K(g_2,g_3)$ is quasi-homogeneous, that is
the curves $D_m^K(g_2,g_3)=0$ are
invariant under the scaling transformations
(\ref{scaling}). Therefore, the equations $D_m^K(g_2,g_3)=0$ can be
rewritten as $C_m^K(J)=0$, and these $C_m^K$ are called
{\em Cohn's polynomials}.
\begin{cor}\label{corollary2} (Maier's conjecture)
$\deg C_m^I=\lfloor ((d_m^I)^2-d_m^I+4)/6\rfloor$ and
$\deg C_m^{II}=d_m^{II}(d_m^{II}-1)/2$, where $d_m^K=\deg_\lambda F_m^K,$
as in (\ref{degFI}), (\ref{degFII}).
\end{cor}
Our second application, and the original motivation of this work
is
the problem of describing degeneration of
metrics of constant positive curvature with conic
singularities which recently attracted substantial attention,
\cite{L1,L2,L4,EMP,L3,LW,MP1,MP2}. Let $S$ be a compact surface,
and $\alpha_1,\ldots,\alpha_n$ positive
numbers. Consider Riemannian metrics on $S$
with $n$ conic singularities with the angles $2\pi\alpha_j$.
Each such metric defines a conformal structure on S with $n$ marked points,
so we have the forgetful map assigning this conformal structure to the metric.
The goal is understanding the space of such metrics and the properties
of the forgetful map.
In this paper
we restrict ourselves to the case when $S$ is a torus with one singularity
with the angle $2\pi\alpha$, where $\alpha=2m+1$ is an odd integer.
Following \cite{MP2} we denote the set of all such metrics by
$\mathrm{Sph}_{1,1}(\alpha)$. One can embed $\mathrm{Sph}_{1,1}(\alpha)$ into $\mathrm{Lame}_m$:
the image of this embedding consists of those Lam\'e equations
whose projective monodromy is {\em unitarizable}
(that is conjugate to a subgroup of $PSU(2).$)
We have the forgetful map
$\mathrm{Sph}_{1,1}(\alpha)\to \mathbf{C}_J$
which assigns to each metric its conformal class. It is known that
when $\alpha>1$ is not an odd integer, this
forgetful map is proper and surjective \cite{BT}, \cite{MP2},
and $\mathrm{Sph}_{1,1}(\alpha)$
is properly embedded in $\mathrm{Lame}_m$.
This is not the case for odd integers $\alpha$, the fact discovered
in \cite{LW} (see also \cite{BE} for a shorter proof of the
main result of \cite{LW}).
As the conformal class varies, a spherical metric can degenerate.
Let us define the set $\mathrm{LW}_m\in\mathrm{Lame}_m$ consisting
of all Lam\'e equations whose projective monodromy consists
of collinear translations (by the periods of the integral (\ref{Ab})
below). In Section~\ref{mono} we show that
\begin{theorem}\label{LWm}
\begin{equation}\label{lw}
\partial\mathrm{Sph}_{1,1}(2m+1)=\mathrm{LW}_m,
\end{equation}
where the boundary is with respect to $\mathrm{Lame}_m$.
\end{theorem}
Then we describe the set $\mathrm{LW}_m$.
\begin{theorem}\label{theorem4} The set $\mathrm{LW}_m$ consists of $m(m+1)/2$
curves. These curves and their projections on $\mathbf{C}_J$ are smooth
real analytic curves
(images of intervals under analytic functions with non-vanishing derivatives).
\end{theorem}
We propose to call projections of the curves $\mathrm{LW}_m$ to $\mathbf{C}_J$
Lin--Wang curves.
They can be seen in
the pictures in \cite{LW,BE} and our Figs.~12, 13 for $m=1$;
in \cite{L3,L4} and our Figs, 14, 15 for $m=2$, and our Figs.~16--18 for $m=3$.
Large part of the papers \cite{L1}, \cite{L2} and \cite{L3} is
dedicated to analytic study of these curves for small $m$;
here we propose a different,
geometric description of them.
\subsection{Description of the method}\label{description}
Our main tool in this paper is
a new geometric interpretation of Lam\'e functions and their moduli
space $\mathbf{L}_m$.
Lam\'e functions correspond to what we call
{\em translation structures} on tori with one conic singularity
with the angle $2\pi(2m+1)$ and $m$ simple poles.
(``Simple pole'' refers to the developing map; its differential
has double poles).
Let $(S,O)$ be an elliptic curve, that is $S$ is a torus with a marked
point
$O\in S$, and
$m\geq 0$ an integer. Translation structures we are talking about
can be identified with Abelian differentials of
the second kind $g(z)dz$ on $S$
with
single zero of multiplicity $2m$ at $O$, and $m$ double poles, subject
to the condition that {\em all residues vanish}. Two translation structures
are {\em equivalent} if the differentials differ by a non-zero
constant factor.
We refer to a survey \cite{Z} of translation structures. Structures
considered in this survey have no poles and correspond to Abelian
differentials of the first kind.
To explain the name ``translation structure'',
consider the Abelian integral
\begin{equation}\label{Ab}
f(z)=\int_{z_0}^z g(\zeta)d\zeta.
\end{equation}
This is a multi-valued function on $S$ with a single critical
point at $O$, and the monodromy of $f$ consists of
translations by the periods of the integral coming from
the fundamental group of $S$. This function $f$
is a developing map of a singular
flat structure on $S$: it has one conic singularity
with the angle $2\pi(2m+1)$ at $O$ and $m$ simple poles; the monodromy of
this structure consists of translations, and the local monodromy at
all points is trivial.
\begin{prop}\label{trans-structures} The correspondence
$w\mapsto\Omega= dx/(uw^2)$ is a bijection
between the space of Lam\'e functions and the
space of triples $(S,O,\Omega)$,
where $S$ is a torus, $O\in S$ a point, and $\Omega$ is
an Abelian differential on $S$, which switches sign under
the conformal involution, and has a single zero of
multiplicity $2m$ at the point $O$,
and $m$ double poles with vanishing residues.
This bijection defines a biholomorphic map between $\mathbf{L}_m$ and
the moduli space of Abelian differentials of considered type, up to scaling.
\end{prop}
\vspace{.1in}
One can pull back the {\em flat} metric on $\mathbf{C}$ via $f$ and obtain a flat
metric on the torus with one conic singularity at $O$ with the angle $2\pi(2m+1)$
and $m$ simple ``poles''. A {\em pole} of a flat metric is a point which
has a punctured neighborhood
isometric to $\{ z\in\mathbf{C}:|z|>R\}$ with Euclidean metric,
for some $R>0$.
We call our torus equipped with this metric
a {\em flat singular torus}. Two flat singular tori are
considered {\em equivalent} if there is an orientation-preserving
diffeomorphism between them
multiplying the metric by a non-zero constant.
To study flat singular tori
we cut each of them into two congruent flat singular triangles.
{\em Congruent} means ``related by an {\em orientation-preserving} isometry''.
\begin{definition}
{\em A} flat singular triangle (FT) {\em is a closed disk
with three marked boundary points which are called} corners,
{\em equipped with a flat
metric with conic singularities at the corners and possibly
simple poles inside the disk or on the open boundary arcs (}sides{\em)
between adjacent corners,
and such that the sides
are geodesic. A side passing through a pole
must be ``unbroken'' at this pole: in the chart $\{ z\in\mathbf{C}:|z|>R\}$
it corresponds to two rays of the} the same {\em line.}
\end{definition}
Alternatively an FT can be described as a triple
$(D,\{ a_j\},f)$, where $D$ is a closed disk in the complex plane,
$(a_1,a_2,a_3)$
three distinct boundary points, and $f$ a locally univalent meromorphic
function on $D\backslash\{ a_j\}$ with conic singularities at
$a_j$, which means
$$f(z)=f(a_j)+(z-a_j)^{\alpha_j}h_j(z),$$
where $\alpha_j>0$, and $h_j$ is analytic near $a_j$, $h_j(a_j)\neq 0$,
and such that the images of the sides
$f([a_j,a_{j+1}])\subset \ell_j\cup\{\infty\}$ where $\ell_j$
are three straight lines (not necessarily distinct)
in the complex plane.
Two such triples $(D,\{ a_j\},f)$ and $(D',\{ a_j^\prime\}, g)$
are {\em equivalent} if there is
a conformal\footnote{We assume that the complex plane has
the standard orientation, and conformal maps preserve it
otherwise we call them anti-conformal.}
homeomorphism $\phi:D\to D'$, $\phi(a_j)=a_j^\prime$,
and complex constants $c_1,c_2$ such that
$f=c_1g\circ\phi+c_2$.
These two definitions of FT are equivalent: for a given triple, we pull back
the Euclidean metric from $\mathbf{C}$ via $f$ and obtain a metric on $D$
satisfying the first definition. In the opposite direction, given such
a metric we obtain $f$ as its developing map.
It is easy to see that the sum of the interior angles $\pi\alpha_j$ of a flat
singular
triangle is an odd integer multiple of $\pi$. If one
angle is an integer multiple of $\pi$,
then all three of them
are
integer multiples of $\pi$.
An FT can be visualized by making a picture of its image in the plane under
the developing map $f$. Such a picture consists of three lines
(not necessarily distinct), three (pairwise distinct) points
of their intersections
$f(a_j)$,
and a marking of the angles. See Figs. 1a, 1b, 1d, 1e, 2a-f.
We mark the angles by little arcs near the images of the corners,
and write $a_j$ instead of $f(a_j)$ in the pictures.
Two such pictures define
the same FT if they are related by a complex affine map.
An FT is called {\em balanced} if its
angles $\pi\alpha_1,\pi\alpha_2, \pi\alpha_3$ satisfy the
triangle inequalities
\begin{equation}\label{balanced}
\alpha_i\leq\alpha_j+\alpha_k,
\end{equation}
for all permutations of $(i,j,k)$.
For example, a triangle with angle sum $\pi$
in Fig.~1a is balanced if and only if all its angles are $\leq\pi/2$.
A triangle with angle sum $3\pi$
in Fig.~1b is balanced if and only if the largest angle
is $\leq 3\pi/2$.
All flat singular triangles
whose angles are integer multiples of $\pi$ are balanced.
A balanced triangle is called
{\em marginal} if we have equality in (\ref{balanced})
for at least one permutation $(i,j,k)$.
Otherwise it is called {\em strictly balanced}.
\vspace{.1in}
{\em We abbreviate the expression ``balanced flat singular triangle''
as BFT.}
\vspace{.1in}
Our main technical result is the following
\vspace{.1in}
\noindent
\begin{theorem}\label{theorem3}
Every flat singular torus has a
decomposition into two congruent BFT.
When the triangles are strictly balanced,
this decomposition is unique up to a cyclic permutation of
the corner labels.
If they are marginal, there are at most two such decompositions:
a marginal triangle and its reflection define the same torus.
\end{theorem}
This is similar to Theorem 1.3 in \cite{EMP} for spherical tori
with one singularity. Theorem \ref{theorem3} gives a parametrization
of our moduli space $\mathbf{L}_m$ by a simpler moduli space $\mathbf{T}_m$ for
BFT's with the sum of the angles $\pi(2m+1)$. This last space admits a nice
partition into open cells which is used to prove Theorem~\ref{theorem1}.
The plan of the paper is the following. In section \ref{slam} we recall basic
facts about Lam\'e equations and Lam\'e functions, and explain
the connection between Lam\'e functions and translation structures.
In section \ref{sBFT} we discuss BFT
and define
a map $\Phi=\Phi_m:\mathbf{T}_m\to \mathbf{L}_m$.
Explicit local coordinates on $\mathbf{T}_m$
are described in Section~\ref{complexanalytic}, and we show that
$\Phi$ is complex analytic and proper. The proof of properness
is based on the study of geodesics on flat singular tori.
In section \ref{sth3} we prove the first part of Theorem \ref{theorem3},
surjectivity of $\Phi$.
In section \ref{spaces} we develop
a classification of BFT and explicitly describe a partition
of $\mathbf{T}_m$ into open cells.
We show that $\mathbf{T}_m$ consists of two connected components,
and that $\Phi$ is in fact $3$-to-$1$ on
the subset of strictly balanced triangles, and $6$-to-$1$ on
the subset of marginal triangles. Factoring $\mathbf{T}_m$ by an appropriate equivalence
relation we obtain a space $\mathbf{T}^*_m$ and show that the induced map
$\Phi^*_m:\mathbf{T}^*_m\to\mathbf{L}_m$ is injective, thus completing the proof
of Theorem~\ref{theorem3}.
In section \ref{ident} we analyze the natural
partition of $\mathbf{T}_m$
and prove Theorem~\ref{theorem1}.
In section \ref{nons} we prove Theorem \ref{theorem2}
and its two corollaries. This is based
on Lemma~\ref{lemma2} which is proved in
Section~\ref{tarasov}.
The last two sections are devoted to spherical metrics:
in Section \ref{mono} we discuss the monodromy of Lam\'e equations, and
in Section \ref{linwang} we
produce equations and pictures of Lin--Wang curves, enumerate them,
and show that they are smooth and real analytic.
\vspace{.1in}
{\em Since we refer to some figures many times, in different places,
all figures are collected at the end of the paper, after the
reference list.}
\vspace{.1in}
We thank Walter Bergweiler, Robert Maier and Vitaly
Tarasov for useful discussions. Walter Bergweiler produced Figs.~12--18
and a Maple program generating polynomials $F_m$.
Vitaly Tarasov proved Lemma~\ref{lemma2}. We also thank
Eduardo Chavez Heredia for bringing \cite{ST} to our attention.
\section{Lam\'e equations, Lam\'e functions and
\newline
translation structures}\label{slam}
We use the form (\ref{lame2}) of the Lam\'e equation.
Every solution $W$ of (\ref{lame2}) is meromorphic in the $z$-plane.
Indeed, by the existence theorem for linear ODE, the singularities of
solutions belong to the lattice $\Lambda$,
and plugging a power series for $W(z)$ at $0$ shows that
there are two linearly independent meromorphic solutions.
\vspace{.1in}
{\em Proof of Proposition~\ref{trans-structures}.}
\vspace{.1in}
We start with a Lam\'e function and assign to it a translation structure.
Let $W$ be a meromorphic
solution of (\ref{lame2}) whose square is even and $\Lambda$-periodic. Then
\begin{equation}\label{period}
W(z+\omega)=\pm W(z),\quad \omega\in\Lambda.
\end{equation}
Now
\begin{equation}\label{w2}
W_1(z):=W(z)\int^z W^{-2}(\zeta)d\zeta
\end{equation}
is another solution of (\ref{lame1}), linearly independent of $W$,
which can be seen by direct computation, so it must be also
meromorphic, since every solution of (\ref{lame2}) is meromorphic.
It follows that all residues of $W^{-2}(\zeta)d\zeta$ vanish.
Now the ratio of two solutions $f=W_1/W$ is an Abelian integral
of the second kind,
and as a ratio of two solutions of a second order linear equation,
it also satisfies the Schwarz equation
\begin{equation}\label{schwa}
\frac{f'''}{f'}-\frac{3}{2}\left(\frac{f''}{f'}\right)^2=
-2\left( m(m+1)\wp+\lambda\right),
\end{equation}
thus all critical points of $f$ are in $\Lambda$
and $f$ is $(2m+1)$-to-$1$ at these points.
So every Lam\'e function defines a translation structure of the desired type.
The differential
$W^{-2}(z)dz$ descends via the map
$$z\mapsto (\wp(z),\wp^\prime(z))=(x,u),$$
to the differential $dx/(w^2u)$ on the curve (\ref{W})
as stated in
Proposition~\ref{trans-structures}.
Conversely, suppose that a translation structure of the described type is
given and its developing map is
\begin{equation}\label{abel}
f(z)=\int^z g(\zeta)d\zeta
\end{equation}
is given. Then $g$ is an even elliptic function with $m$ double poles
per period parallelogram, vanishing residues,
and zeros of order $2m$ at the points of $\Lambda$.
So $g=W^{-2}$ for some meromorphic function $W$ satisfying
(\ref{period}). Now we define $W_1$ by (\ref{w2}),
and a direct computation shows that
$$WW_1^\prime-W'W_1=1.$$
This means that $W$ and $W_1$ are two linearly independent solutions
of some equation $W''=PW$, where $P$ is an elliptic function
with periods $\Lambda$. As the only pole
of $P$ can occur at a critical point of $f$, we conclude that poles
of $P$ must be at the points of $\Lambda$, and a simple calculation
with power series at $0$ shows that
$$P(z)=m(m+1)z^{-2}+O(1),\quad z\to 0,$$
so $P(z)=m(m+1)\wp(z)+\lambda$ with some $\lambda\in\mathbf{C}$,
and thus $W$ is a Lam\'e function.
We recall that two Lam\'e functions $W_1$ and $W_2$ are equivalent if
$W_1(z)=cW_2(kz)$ for some $c$ and $k$ in $\mathbf{C}^*$.
Translation structures are equivalent
if their developing maps $f_1,f_2$ are related by post-composition
with an affine map: $f_1(z)=af_2(kz)+b$, $a,k\in\mathbf{C}^*,\; b\in\mathbf{C}$.
We proved that {\em equivalence classes of degree $m$ Lam\'e functions
are in one-to-one correspondence with classes of translation
structures on the torus with one conic
singularity with the angle $2\pi(2m+1)$.}
It is clear that this correspondence is continuous and holomorphic,
therefore it is biholomorphic.\hfill$\Box$
\vspace{.1in}
Let us recall how the polynomial $F_m$
is computed (another method is described in Section 6).
It is convenient to use the algebraic form of the Lam\'e equation
(\ref{lame1}) which can be also written as
\begin{equation}\label{newlame}
w''+\frac{1}{2}\left(\sum_{j=1}^3\frac{1}{\zeta-e_j}\right)w'=
\frac{m(m+1)\zeta+\lambda}{4(\zeta-e_1)(\zeta-e_2)(\zeta-e_3)}w,
\end{equation}
where $w=y\circ\wp$.
Since the only singularities in $\mathbf{C}$ of
this equation are $e_j$, and the local exponents
at these singularities are $\{0,1/2\}$, each Lam\'e function can be written
as
\begin{equation}\label{lfunction}
w(\zeta)=c\prod_{j=1}^3(\zeta-e_j)^{k_j/2}\prod_{j=1}^n(\zeta-\zeta_j),
\end{equation}
where $k_j\in\{0,1\}$ and
\begin{equation}\label{num}
\sum_{j=1}^3 k_j+2n=m.
\end{equation}
Plugging $\zeta=\zeta_k$ into (\ref{newlame}) we obtain after simple
calculations (see, for example, \cite[23.21]{WW})
the following system of equations for $\zeta_j$
\begin{equation}\label{bete}
2\sum_{j:j\neq k}^n\frac{1}{\zeta_k-\zeta_j}+
\sum_{j=1}^3\frac{k_j+1/2}{\zeta_k-e_j}=0,\quad 1\leq k\leq n.
\end{equation}
This system of equations determines Lam\'e functions. According to a
theorem of Heine and Stieltjes \cite[23.46]{WW}, system (\ref{bete}) has
at most $n+1$ solutions and exactly $n+1$ for generic $e_j$. Moreover,
it has exactly $n+1$ solutions when all $e_j$ are real.
System (\ref{bete}) is a very special case of the Bethe ansatz equations
which frequently occur in mathematical physics and in the study
of metrics with conic singularities.
Lam\'e functions are classified according to the values of $k_j$ in (\ref{lfunction}):
traditionally the number
$$1+\sum_{j=1}^3k_j\in\{1,2,3,4\}$$
is called the {\em kind} of a Lam\'e function \cite{WW}.
Using (\ref{num}) and Stieltjes theorem
we conclude that the total number of Lam\'e functions
for a given generic lattice is $2m+1$, and this is the degree
of the Lam\'e spectral polynomial $F_m$ with respect to $\lambda$.
The number of Lam\'e functions is exactly $2m+1=d_m^I+d_m^{II}$
for lattices with
real $e_j$, or equivalently with real $g_j$ with $g_2^3-27g_3^2>0$.
It is easy to see that for even $m\geq 2$ there exist Lam\'e functions
of the first and third kind; we denote the corresponding subsets by $\mathbf{L}_m^I$
and $\mathbf{L}_m^{II}$. Similarly, when $m\geq 3$ is odd, we define $\mathbf{L}_m^I$ as the
set of Lam\'e functions of the fourth kind, and $\mathbf{L}_m^{II}$
as the set of Lam\'e functions of the second kind.
This explains why
$\mathbf{L}_m$ has {\em at least two components} when $m\geq 2$.
The more difficult result, which is a part of Theorem~\ref{theorem1},
is that these components are in fact irreducible.
\vspace{.1in}
{\em Connection with spin structures.}
\vspace{.1in}
Connected components of moduli spaces of Riemann surfaces endowed
with nonzero holomorphic differentials were classified by
Kontsevich and Zorich \cite{KZ}; the case
of meromorphic differentials was treated by Boissy \cite{Bo}.
A consequence of Theorem 4.1 in \cite{Bo} is that the moduli space of triples
$(S,O,\Omega)$, where $(S,O)$ is an elliptic curve
and $\Omega$ is
a meromorphic differential on $S$ with a zero at $O$ of order $2m$
and $m$ poles $q_1,\dots,q_m$ of order $2$ (and arbitrary residues),
has exactly two connected components.
Moreover, such components are distinguished by the
{\it{spin invariant}} (already defined in \cite[Sect. 2.2]{KZ}).
In our particular case, the spin invariant of $(S,O,\Omega)$ is {\it{odd}}
if there exists a function on $S$ with simple
poles at $q_1,\dots,q_m$ and a zero of order $m$
at $O$, and it is {\it{even}} if such function does not exist.
\begin{prop}
A Lam\'e function $w$ is in $\mathbf{L}_m^I$ if the spin
invariant of the corresponding translation surface is odd,
and $w$ is in $\mathbf{L}_m^{II}$ if the spin invariant is even.
\end{prop}
{\em Proof.}
Let $(S,O,\Omega)$ be the translation surface associated
to the Lam\'e function $w$.
In particular, $\Omega=\varphi/w^2$,
where $\varphi$ is a nonzero holomorphic
differential on $S$, see Proposition~\ref{trans-structures}
Then
the spin invariant of $(S,O,\Omega)$ is odd
if and only if $w$ is a well-defined function on $S$,
which happens if and only
if $w=Q(x)$ or $w=Q(x)u$ for a suitable $Q\in\mathbb{C}[x]$.
Hence, the spin invariant
is odd if and only if $w$
is of type $I$.\hfill$\Box$
As a consequence of \cite[Thm. 4.1]{Bo} we obtain
\begin{cor}
The space $\mathbf{L}_m$ of Lam\'e functions
is the disjoint union of the subset
$\mathbf{L}_m^I$
of Lam\'e functions of type $I$ and of the subset
$\mathbf{L}_m^{II}$
of Lam\'e functions of type $II$.
\end{cor}
The techniques of \cite{Bo}
do not allow to study connected components
of moduli of meromorphic differentials with vanishing residues.
So it does not follow from \cite{Bo}
that $\mathbf{L}_m$ has exactly two connected components, and so that
$\mathbf{L}_m^I$ and $\mathbf{L}_m^{II}$ are connected.
\section{Balanced flat singular triangles}\label{sBFT}
In this section we consider balanced flat singular triangles (BFT)
with marked corners $a_1,a_2,a_3$,
enumerated according to the positive orientation of the boundary
(so that the region is on the left when we trace the boundary).
We denote the interior angles at these corners by $\pi\alpha_1,\pi\alpha_2,
\pi\alpha_3$. As we already noticed, the sum of the angles is
an odd multiple of $\pi$, more precisely
\begin{equation}\label{suma}
\alpha_1+\alpha_2+\alpha_3=4n+2k+1=2m+1,
\end{equation}
where $n$ is the number of interior poles, and $k$ is the number
of boundary poles.
To prove (\ref{suma}) we recall the argument used in the
proof of the Schwarz--Christoffel formula. Consider the developing map $f$ defined
in the upper half-plane with corners at $(a_1,a_2,a_3)=(0,1,\infty)$.
Since the monodromy of $f$ is affine, $f''/f'$ must
extend to the complex plane as
a rational function whose poles are symmetric with respect to
the real line, and
we have
$$\frac{f''}{f'}(z)\sim\frac{\alpha_j-1}{z-a_j},\quad z\to a_j,
\quad j\in\{1,2\},\quad
\frac{f''}{f'}(z)\sim-\frac{\alpha_3+1}{z},\quad z\to\infty,$$
and
$$\frac{f''}{f'}(z)\sim -\frac{2}{z-z_j},\quad z\to z_j$$
at the poles. So (\ref{suma}) follows by the Residue Theorem.
Fig.~1a shows a triangle with $n=k=0$. In Fig.~1b $n=0, k=1$.
Fig.~1d shows a triangle with $n=1, k=0$ (left) and two triangles with
$n=0, k=2$ (right).
Fig.~1e shows three triangles with $n=1, k=0$. Fig.~2
shows all
types of triangles with sum of the angles $5\pi$ ($m=2$).
We also recall that {\em either none of the $\alpha_j$ or all of them
are integers.}
\begin{prop}\label{prop1}
An FT with non-integer $\alpha_j$ is completely
determined by the angles, and any positive angles $\pi\alpha_j$ where
$\alpha_j$ are not integers and their sum is odd can occur.
For integer angles, the necessary and sufficient condition of
existence of FT is that the sum of $\alpha_j$ is odd
and (\ref{balanced}) is satisfied. For any such angles,
there are three
one-parametric families of FT's.
\end{prop}
A similar result for spherical triangles was proved in \cite{E}.
\vspace{.1in}
{\em Proof.} The first statement is essentially due to Klein \cite{Klein}.
If our triangle is modeled
on the upper half-plane $D=\overline{H}$ with vertices $(a_1,a_2,a_3)=(0,1,\infty)$
then the developing map $f:\overline{H}\to\mathbf{\overline{C}}$ satisfies
the Schwarz equation with three singularities $(0,1,\infty)$
and $\alpha_j$
determine this equation completely.
When the $\alpha_j$ are not integers,
the monodromy representation
corresponding to this equation is non-trivial, and there is only
one choice, up to an affine transformation, of a solution
with affine monodromy.
If all $\alpha_j$ are integers, then the monodromy representation
is trivial, and the developing map $f$ extends
from its domain $D$ to the whole Riemann sphere and we obtain
a rational
function with three critical points $a_1$, $a_2$ and $a_3$.
The images of all three sides under the developing map $f$ belong the same
line $\ell$. Preimage $f^{-1}(\ell)\cap \overline{D}$ is called the {\em net}
of $f$ (see \cite{EG}). The net is a cell decomposition of $D$
with three vertices $a_1,a_2,a_3$. The $1$-cells of this decomposition
are disjoint chords of the disk $D$ and three arcs of $\partial D$.
The number of chords in
the interior of $D$ which are adjacent to $a_j$ is $\alpha_j-1\geq 0$.
If $m_j$ is the number of chords from $a_i$ to $a_k$, then
$\alpha_j-1=m_i+m_k$, and these three equations have unique non-negative
solution $(m_1,m_2,m_3)$ if and only if the integers $\alpha_j\geq 1$ satisfy
(\ref{balanced}) and their sum is odd.
Thus the angles determine the net. Once the net
is given, the developing map can be recovered from the values $f(a_1),f(a_2),
f(a_3)$ which in the considered case belong to a line.
It is clear that one of the $2$-faces of a net is a triangle while
the others are digons.
Suppose that the triangular face is mapped by
$f$ onto the upper half-plane. Then we have three possible
orderings of $f(a_j)$ on the real line. By scaling we may arrange
that $\min_j f(a_j)=0$ and $\max_j f(a_j)=1$, then the intermediate
point of the three $f(a_j)$ serves as a parameter. So
the set of triangles with given integer angles is parametrized by three
intervals.
\hfill$\Box$
\vspace{.1in}
Examples of nets of triangles are shown in Fig.~1c, where the stars
mark the location of the poles of $f$. These three nets correspond to
triangles with the angles $(2\pi,3\pi,4\pi)$.
In Fig.~1d, three
triangles with the angles $(\pi,2\pi,2\pi)$ are shown together with their nets.
\vspace{.1in}
Let us define two types of triangles which we call {\em primitive}.
A primitive triangle of type $A$ has all angles in $(0,\pi)$ and their
sum is $\pi$ (Fig.~1a),
and the primitive triangle of type $B$ has one angle in $[\pi,2\pi)$,
two others in $(0,\pi]$, and the sum of the angles is $3\pi$ (Fig.~1b).
\begin{prop}\label{prop2}
Every BFT can be obtained from one of the two primitive
triangles $A$ or $B$ by gluing half-planes
to the sides.
\end{prop}
{\em Proof.}
This is essentially due to Klein \cite{Klein}, but we sketch a proof. Consider the
case of non-integer angles.
Let $\alpha_i=\{\alpha_i\}+\[\alpha_i\],$ be the decomposition into
fractional and integer parts.
\vspace{.1in}
\noindent
a) If $\sum\[\alpha_i\]$ is even, then $\sum\{\alpha_i\}$ is odd.
Since the last
sum belongs to $(0,3)$, we must have
\begin{equation}\label{sumone}
\sum\{\alpha_i\}=1.
\end{equation}
Since our triangle is balanced, and we have (\ref{sumone}), we obtain
for all permutations $(i,j,k)$ of $(1,2,3)$:
\begin{eqnarray*}
\[\alpha_i\]&=&\alpha_i-\{\alpha_i\}\leq\alpha_j+\alpha_k-\{\alpha_i\}=
\[\alpha_j\]+\[\alpha_k\]+\{\alpha_j\}+\{\alpha_k\}-\{\alpha_i\}\\
&<&
\[\alpha_j\]+\[\alpha_k\]+1.
\end{eqnarray*}
So $\[\alpha_i\]\leq \[\alpha_j\]+\[\alpha_k\]$. It follows that
the following
quantities are non-negative
integers:
$$x_i=(\[\alpha_j\]+\[\alpha_k\]-\[\alpha_i\])/2,$$
and we have $\[\alpha_i\]=x_j+x_k,$ for all permutations $(i,j,k)$
of $(1,2,3)$.
So we can take a triangle with the angles $\{\alpha_j\}$ of the type A
and glue $x_j$ half-planes to the side opposite to $a_j$, for each $j$.
The resulting triangle has the
same angles as our original triangle, so it is the same triangle by
Proposition~\ref{prop1}.
\vspace{.1in}
\noindent
b) If $\sum\[\alpha_i\]$ is odd, then we decrease one of the
integer parts (for example, the largest one) by $1$, and increase
the corresponding fractional part by $1$. So we set for some $i$
$$p_i=\[\alpha_i\]-1,\quad\alpha_i^{\prime}=\{\alpha_i\}+1,$$
leaving other numbers unchanged ($\alpha_j^\prime=\{\alpha_j\},\;
p_j=\[\alpha_j\]$, and similarly for $k$).
Now we have
\begin{equation}\label{b1}
\alpha_i^{\prime}\in(1,2),\quad\alpha_j^{\prime}\in(0,1),\quad
\alpha_k^{\prime}\in(0,1),
\end{equation}
and since $\sum\{\alpha_\ell\}$ is even and less than $3$ in this case,
it must be $2$, and
thus
\begin{equation}\label{sumthree}
\sum\alpha_\ell^{\prime}=3.
\end{equation}
Now, since our triangle is balanced, and using (\ref{sumthree}) and (\ref{b1})
we obtain
$$p_i=\alpha_i-\alpha_i^{\prime}\leq
\alpha_j+\alpha_k-\alpha_i^{\prime}=p_j+p_k
+\alpha_j^{\prime}+\alpha_k^{\prime}-\alpha_i^{\prime}
\leq p_j+p_k+1,$$
and since the sum of $p_\ell$ is even, we conclude that
$p_i\leq p_j+p_k.$
So we can define $x_i,x_j,x_k$ as in part a) and conclude that our triangle
is obtained from a triangle with the angles $\alpha_\ell^\prime$ of
the type B, by gluing $x_\ell$ half-planes to the side opposite to $a_\ell$.
Triangles with
integer angles can be considered using their nets introduced
in the proof of Proposition~\ref{prop1}.
A net is a chord diagram in a disk with three vertices
on the circle. Evidently each net has one triangular face,
and the rest of the faces are digons. Triangular face corresponds to
the primitive triangle and digons are half-planes.
Since the angle sum is
odd, only case b) can occur, and the primitive triangle
is of type B with the angles $(\pi,\pi,\pi)$.
Thus the primitive triangle in this case
is just a half-plane with three marked boundary points.\hfill$\Box$
\begin{cor}\label{corpol}
Each side of a BFT contains at most one pole, and
the developing map sends each side $(a_i,a_{i+1})$ injectively
either to the interval
\newline
$(f(a_i),f(a_{i+1}))$, or to the complement
of this interval on the line in $\mathbf{C}\cup\{\infty\}$ containing it.\hfill$\Box$
\end{cor}
\noindent
{\bf Remarks.} 1. By analyzing the proof of Proposition~\ref{prop2} one
can obtain the following criterion: the side opposite to $a_i$
contains a pole if and only if
$$\lceil(\[\alpha_j\]+\[\alpha_k\]-\[\alpha_i\]-1)/2\rceil\quad\mbox{is odd}.$$
\noindent
2. It follows from this Corollary and from (\ref{suma})
that for each $m\geq 2$ we have exactly two possibilities
for the number of sides with a pole. When $m$ is even,
we have either none or 2 sides with poles. If $m$ is odd, we have
ether one or three sides with poles. This shows that
the set $\mathbf{T}_m$
of all BFT's with the angle sum $\pi(2m+1),\; m\geq 2$ must have at least
four connected components. We show in Section \ref{spaces} that there
are exactly four.
\vspace{.1in}
\noindent
3. The decomposition into a primitive triangle and half-planes
stated in Proposition~\ref{prop2} is canonical when $\sum\[\alpha_j\]$
is even, but not canonical when it is odd.
In the latter case, we can
obtain from one to three such different decompositions, depending
on the number of positive $\[\alpha_j\]$.
\vspace{.1in}
\noindent
4. The primitive triangle $T'$ obtained from a balanced triangle $T$
may be unbalanced. In this case, there is always at least one half-plane
in $T$ glued to the side of $T'$ which is opposite to the largest
angle of $T'$. Indeed, let $\alpha_j^\prime$ be the angles of $T'$,
and $\alpha_1$ is the largest one. Then the angles of
$T$ are $\alpha_i=\alpha_i^\prime+x_j+x_k$, where $(i,j,k)$
is a permutation of $(1,2,3)$, and $x_j$ are the numbers of half-planes
glued to the sides opposite to $\alpha_j^\prime$, $j=1,2,3$. Since
$T$ is balanced,
$$\alpha_1^\prime+x_2+x_3=
\alpha_1\leq\alpha_2+\alpha_3=\alpha_2^\prime+\alpha_3^\prime+x_2+x_3+2x_1,$$
so $\alpha_1^\prime\leq \alpha_2^\prime+\alpha_3^\prime+2x_1$. Thus if
$T'$
is unbalanced, we have $x_1>0$.
\vspace{.1in}
\noindent
{\em Construction of the map $\Phi$.}
\vspace{.1in}
Let $\mathbf{T}_m$ be the set of all balanced triangles with the sum of the angles
$\pi(2m+1)$.
For every $T\in\mathbf{T}_m$ we define a singular flat torus
$\Phi(T)$ in the following way. We take two copies of $T$
and identify each pair of corresponding sides
by the {\em orientation-reversing} isometry.
Thus all corners of both copies are glued into one point,
and the sides are glued into three simple loops on the torus based
at this point and otherwise disjoint.
Notice that the resulting torus has an orientation-preserving
isometric involution which interchanges the two triangles.
The four fixed points of this involution are: the conic singularity
(corresponding
to the vertices of the triangle) and the ``midpoints of the sides'' which
are points of order $2$ on the elliptic curve. If
a side is unbounded, its midpoint is a pole.
With such gluing we obtain a flat singular torus
with one singularity
with the angle $2\pi(2m+1)$. There are $m$ simple poles on the
torus coming from the poles of the metric on the triangle.
An interior pole of $T$ gives two poles on $\Phi(T)$, while
a pole on a side of $T$ gives one pole on $\Phi(T)$.
Let $T$ be a BFT with corners $a_1,a_2,a_3$ and developing map $f$.
Let $b_j=(f(a_i)+f(a_k))/2$, where $(i,j,k)$ is a permutation of $(1,2,3)$.
We define affine maps $s_j$ to be rotations by $\pi$ about $b_j$.
Consider the group generated by $s_1,s_2,s_3$. It contains a subgroup
$G$ of index $2$ consisting of translations;
elements of $G$ are products of even numbers
of $s_j$. The following proposition is evident:
\begin{prop}\label{prop2.5}
The monodromy group of the developing map of the flat singular torus
$\Phi(T)$ is~$G$.\hfill$\Box$
\end{prop}
In particular, the monodromy consists of collinear translations
if and only if the angles of the triangle are integer multiples of $\pi$.
Next we address the question when two different BFT's can give
the same (isometric up to a constant factor) tori.
This can happen in at least two ways:
\vspace{.06in}
\noindent
1. The triangles are obtained by a cyclic
permutation of the corners.
\vspace{.1in}
\noindent
2. Some pairs of marginal triangles define the same torus.
\vspace{.1in}
More precisely, we have
\begin{prop}\label{prop3}
For a marginal triangle $T$
with the angles $(\alpha_1,\alpha_2,\alpha_3)$ where
$\alpha_1=\alpha_2+\alpha_3$, and the triangle $T^*$
with the angles $(\alpha_1,\alpha_3,\alpha_2)$ the tori
$\Phi(T)$ and $\Phi(T^*)$ are congruent.
\end{prop}
$T$ and $T^*$ are related by a reflection, an orientation-reversing
isometry.
\vspace{.1in}
{\em Proof of Proposition~\ref{prop3}.} First we notice that a marginal triangle cannot
have integer angles (since the angle sum is odd),
so it is completely determined
by the angles. We use Klein's decomposition described in the proof of
Proposition~\ref{prop2}. Our triangle is obtained from a primitive triangle by gluing
half-planes to the sides.
We claim that for a marginal triangle,
no half-planes are glued to the side opposite to the larger angle.
Indeed, let $\alpha_j'$ be the angles of the primitive triangle,
and $\alpha_j=\alpha_j^\prime+p_j=\alpha_j^\prime+x_i+x_k$,
where $p_j,x_i,x_k$ are non-negative integers,
Then $\alpha_1=\alpha_2+\alpha_3$
implies
\begin{equation}\label{ma}
\alpha_1^\prime=\alpha_2^\prime+\alpha_3^\prime+2x_1,
\end{equation}
and we obtain that $x_1=0$ since
$\alpha_1<2,\; \alpha_2^\prime>0,\alpha_3^\prime>0.$
This proves the claim.
We also conclude from (\ref{ma}) with $x_1=0$ that the primitive triangle $T'$
with the angles $\alpha_1^\prime,\alpha_2^\prime,\alpha_3^\prime$
corresponding to our triangle $T$ is also marginal, with the larger angle
$\alpha_1.$ If $T'$ is of the type A, then $\alpha_1^\prime=1/2$.
If $T'$ is of the type B, then $\alpha_1^\prime=3\pi/2$.
Gluing together two copies of $T'$ along the side $(a_3,a_1)$
we obtain either a rectangle in the plane, or a complement of
a rectangle on the Riemann sphere.
It is clear that rectangles obtained from
$T$ and $T^*$ are congruent, and
the corresponding tori $\Phi(T)$ and $\Phi(T^*)$ are also congruent,
see Fig.~3.\hfill$\Box$
\vspace{.1in}
We give two examples illustrating Theorem~\ref{theorem3} in the simplest cases.
\vspace{.1in}
\noindent
{\bf Example 1.} Consider a triangle $T$ shown in Fig.~1a, whose angle sum is
$\pi$. It is balanced iff all angles are at most $\pi/2$. Gluing two congruent copies
of such a triangle along a common side,
we obtain a parallelogram in the plane. Identifying the opposite sides
of this parallelogram by translations we obtain a flat non-singular
torus $\Phi(T)$ ($m=0$).
Now consider a flat non-singular torus $L$ with a marked point $O$.
Let $f:\mathbf{C}\to L,\; f(0)=O$ be the universal cover. Then there is a
lattice $\Lambda\subset\mathbf{C}$ such that $L=\mathbf{C}/\Lambda$. A fundamental region
$D$ of this lattice can be taken in the form of a parallelogram which can
be normalized so that the shorter side is $[0,1]$. Let $[0,\tau]$ be the
longer side. It is well known that $\tau$ can be always chosen
in the fundamental region of the modular group
$$G=\{\tau:|\tau|\geq 1,|\mathrm{Re}\,\tau|\leq 1/2\}.$$
To achieve this one can normalize so that the shortest non-zero element
of $\Lambda$ is $1$, then the shortest non-real element of $\Lambda$
is $\tau$.
For $\tau\in G$, each diagonal of $D$
breaks $D$ into a pair
of congruent triangles, such that at least one pair consists of balanced
triangles. Both pairs consists of balanced triangles if and only
if $D$ is a rectangle, in which case the triangles of different
pairs are marginal and are
reflections of each other. It is easy to see
that at least one diagonal breaks $D$ into balanced triangles
if and only if $\tau\in G$. This proves Theorem~\ref{theorem3} for
$m=0$. The proof in the general case is a generalization
of this argument. \hfill$\Box$
\vspace{.1in}
\noindent
{\bf Example 2.} Consider a triangle $T$ in Fig.~1b.
The angle sum is $3\pi$ ($m=1$). Rotating $T$
by $\pi$ about the point $(a_1+a_3)/2$ we obtain a
congruent triangle $T'$. Gluing $T$ and $T'$ along the side which
contains pole, we obtain the ``exterior parallelogram'' $Q$.
Identifying the opposite sides of $Q$ by translations, we obtain the
flat singular torus $\Phi(T)$ with the angle $6\pi$
at the singularity. It will be proved later that every flat singular
torus with the angle $6\pi$ can be obtained from an exterior
parallelogram by identifying the opposite sides by translations.
We claim that every exterior parallelogram $Q$ can be obtained
by gluing two {\em balanced} triangles of the type B. Indeed,
consider the diagonals of the parallelogram $\mathbf{C}\backslash Q$, and extend
them to $Q$. Each extension of a diagonal breaks $Q$ into two
triangles of the type B, and it is easy to see that only for
one diagonal these triangles are balanced, unless our $Q$
is an exterior rectangle as in Fig.~3. When $Q$ is an exterior rectangle,
our triangles are marginal, and they are related
by reflection as in Proposition~\ref{prop3}.
\hfill$\Box$
\vspace{.1in}
Now we define an {\em equivalence relation} $\sim$ on $\mathbf{T}_m$,
in accordance with statements 1,~2 before Proposition \ref{prop3}.
Two BFT's $T$ and $T'$
are equivalent if either there is a congruence between them
cyclically permuting the corners, or if they are marginal
and related by a reflection as described Proposition
\ref{prop3}. Then our map $\Phi$ is well defined on the equivalence
classes, and we prove in Section~\ref{sth3} that
the induced map
$$\Phi_m^*:\mathbf{T}_m^*\,\longrightarrow\mathbf{L}_m,\quad \mathbf{T}_m^*=\mathbf{T}_m/\sim$$
is bijective. We denote by $\Phi^*$ the map defined
by this formula on the disjoint union $\mathbf{T}^*$ of $\mathbf{T}_m^*,\; m\geq 0$,
and mapping it to the disjoint union $\mathbf{L}$ of $\mathbf{L}_m$.
\section{Complex analytic coordinates on $\mathbf{T}_m$}
\label{complexanalytic}
We introduce a complex analytic structure on the set of
BFT. We define functions on the set of BFT:
$$\phi_{i,j,k}=\frac{f(a_i)-f(a_j)}{f(a_k)-f(a_j)}.$$
These functions are locally
injective and we use them as complex coordinates on $\mathbf{T}_m^*$.
The correspondence maps between charts are linear-fractional.
So they define
a complex analytic structure on $\mathbf{T}_m^*$ (and even a projective structure).
To show that the map $\Phi_m^*$ is complex analytic, we recall
Proposition~\ref{prop2.5} which implies that $\phi_{i,j,k}$
are ratios of periods of the differential in
Proposition~\ref{trans-structures}. It is clear
that periods and their ratios are analytic on $\mathbf{L}_m$ (which can
be identified with a space of proportionality classes of differentials).
We refer to a much more general statement of this kind in
\cite[Corollary 2.3]{Ba}.
So $\phi_{i,j,k}$ are local complex analytic
coordinates on $\mathbf{T}_m^*$, and we have
\begin{prop}\label{prop4}
The map $\Phi_m^*:\mathbf{T}_m^*\to \mathbf{L}_m$
is complex analytic.\hfill$\Box$
\end{prop}
Next we prove
\begin{prop}\label{prop5}
The maps $\Phi_m^*$ are proper.
\end{prop}
{\em Proof.} We are going to show that a degenerating sequence
of triangles gives a degenerating sequence of tori. First we clarify
the notions of degenerating sequences.
For every flat singular torus $L$ with one conic singularity at $O$,
we define the set $G$ of all simple geodesics loops based at $O$.
Some of these loops may pass through a pole and thus
have infinite length.
The {\em systole} $s_1$ is the minimal length of all elements of $G$.
The {\em second systole} $s_2\geq s_1$ is the minimal length of
all elements of $G$ whose homotopy class is not a multiple
of the class of some element of $G$ of length $s_1$.
Since we identify tori with proportional metrics,
only the ratio $s_2/s_1$ is defined as a function on $\mathbf{L}_m$.
It is clear that $s_2/s_1$ is bounded on every compact
subset of $\mathbf{L}_m$.
{\em If $s_2/s_1\to\infty$ for a sequence of tori in $\mathbf{L}_m$, then
this sequence diverges in $\mathbf{L}_m$, and
we say that tori of this sequence degenerate.}
Proposition~\ref{prop2} defines for every balanced $T$ a primitive
triangle $T'$ which is of type A or B (Fig.~1a,b).
When the sum of $\lfloor\alpha_j\rfloor$ is even,
then $T'$ is of type $A$ and it is defined
uniquely. When this sum is odd, we may have up to three choices for $T'$;
they are of type $B$
(see Remark~3 after Corollary~\ref{corpol}).
We pick one of them, as we did
in the proof of Proposition~\ref{prop2}.
This primitive triangle $T'$ may be unbalanced. If this is the case,
then there is at least one half-plane in $T$ attached
to the side of $T'$
opposite to the largest angle of $T'$ (Remark 4 after Corollary~\ref{corpol}).
Then we denote by $T''$ the union (more precisely the result of gluing) of
$T'$ with this half-plane. When $T'$ is balanced we set $T''=T'$.
We call $T''$ the {\em balanced extension} of $T'$.
There are $4$ types of $T''$: balanced of type $A$ or $B$
and balanced extensions of unbalanced $A$ and $B$. We call
these last two types $A''$ and $B''$.
The numbers $|f(a_i)-f(a_j)|$ are all the same for
$T, T'$ and $T''$. Consider the ratios
$$\frac{f(a_i)-f(a_j)}{f(a_k)-f(a_j)}$$
It is easy to see that if a sequence of triangles in $\mathbf{T}_m$ leaves
every compact in $\mathbf{T}_m$ then one of these ratios tends to infinity.
We call triangles of such a sequence {\em degenerate},
and notice that {\em $T,T'$ and $T''$ degenerate simultaneously.}
(Unbalanced triangle of type $A$
can degenerate in a different way:
when a vertex tends to an interior point of the opposite side. But
its balanced extension does not degenerate in this case.)
We claim that $s_1,s_2$ for $\Phi(T'')$ are the same as for
$\Phi(T)$. Indeed, attaching $n$ half-planes to a side of
a triangle $T''$ results
in attaching a digon $D$ (with poles)
with angles $2\pi n$ at at its two corners to the torus $\Phi(T)$.
Every curve in such a digon with endpoints at $\partial D$ is
at least as long as the segment between its endpoints.
On the other hand, there are four
types of $T''$ of which $A$ and $B''$ have all sides bounded,
while the complementary segment of the unbounded side of $A''$ or $B$
is at least $s_2$ for $T''$.
It remains to consider the tori $\Phi(T'')$, and to show that
when $T''$ degenerates then $s_2/s_1\to\infty$ for these tori.
There are 4 cases to consider:
\vspace{.06in}
\noindent
a) $T'$ is balanced, then $T''=T'$. Then $T'$ is of type A or B,
and balanced,
and this is essentially the cases of Examples 1, 2 in the end of Section 3.
\vspace{.1in}
\noindent
b) $T'$ is unbalanced, $T''$ is the balanced extension
of $T'$ of the type $A$ or $B$.
One has to find the first and second systole of such tori $\Phi(T'')$.
We denote $A''$ and $B''$ the classes of triangles consisting
of balanced extensions
of unbalanced triangles of type
$A$ and $B$.
\vspace{.1in}
We are interested in the length spectrum of the set of geodesic loops
in $\Phi(T_1)$ where $T_1$ is a balanced triangle of one of the types
$A,B,A'',B''$.
To find it we describe all geodesic loops of finite length
in $G$ for each type.
\begin{lemma}\label{u1}
For $T_1$ of type $A$, all elements of $G$ have finite length,
and they are in bijective
correspondence with indivisible elements\footnote{An element of $\mathbf{Z}^2$
is called indivisible
if it is not an integer multiple of any other element.}
of the lattice generated by
the sides of $T_1$.
So $s_1$ and $s_2$ are the smallest and the second smallest
lengths of sides of $T_1$.
For $T_1$ of type $B$, there are two elements in $G$ of finite length.
They correspond to the two sides of $T_1$ of finite length.
For $T_1$ type $A''$, let $a_3$ be the corner opposite to the side
of infinite length, and denote $a=f(a_2)-f(a_1),\; b=f(a_3)-f(a_1)$
Then the length spectrum is
\begin{equation}\label{spectrum}
\{|na+b|:n\in\mathbf{Z}\}
\end{equation}
and $\{ s_1,s_2\}=\{|b|,|a-b|\}.$
For $T_1$ of type $B''$,
there are exactly three elements of $G$ of
finite length. They correspond to the sides of $T_1$.
\end{lemma}
{\em Proof.} We look at all geodesics starting at
a vertex of $T_1$ with all possible slopes.
If $T_1$ is of type $A$, then the torus $\Phi(T_1)$
is a flat non-singular torus from Example~1
in Section~\ref{sBFT}.
Elements
of $G$ are in bijective correspondence with segments whose endpoints
are elements of the lattice and which contain no lattice points in their
interiors.
For $T_1$ of type $B$, the torus $\Phi(T_1)$ can be represented
as an exterior parallelogram $Q$ whose opposite sides
are identified by translations. Any geodesic starting from a vertex
which is not a side of $Q$, visits the pole and thus
has infinite length. The set $G$ contains two elements of finite length
(corresponding to two pairs of parallel sides of $Q$).
Let $T_1$ be the balanced extension of an unbalanced triangle
$T_2$ with angle sum $\pi$, and $f$ the developing map of $T_1$.
We normalize so that $f(T_2)=(0,1,b),$ so that $a=1$,
and $|b|\leq 1.$
The torus $\Phi(T_1)$
is partitioned into two triangles
congruent to $T_2$ and two half-planes. Suppose that a
geodesic starting from
a vertex $v$ is not a side of $T_2$. Then it visits the regions of
our partition, one after another. Once the geodesic enters a half-plane
it must stay there until it hits a pole, so the length of such geodesic is
infinite. A geodesic in $G$ of finite length must cross
the two finite sides of $T_1$ alternatively, and its length is given by
(\ref{spectrum}). See Fig.~4, which shows the images
under the developing map:
$f(T_1)$ is dark, the image of one congruent copy of $f(T_1)$
is grey, and images other congruent copies of $T_2$ are white triangles.
Images of several geodesics issued from one corner are shown:
the dotted lines are images of geodesics of
infinite length which contain poles, and dashed lines are images of
some
geodesics in $G$ of finite length. Since $|b|\leq 1$, the smallest
and the second smallest elements of (\ref{spectrum})
are $|b|$ and $|1-b|$.
Now consider a torus $\Phi(T_1)$ where $T_1$ is of type $B''$,
that is $T_1$ is a balanced extension of some unbalanced $T_2$ of type $B$.
Then $T_1$ is an exterior triangle (exterior of a bounded
triangle with angle sum $\pi$).
The torus $\Phi(T_1)$
is obtained by gluing two copies of exterior
triangles $T_1$. Every geodesic on this torus
which does not correspond to a side of
$B''$ passes through a pole.\hfill$\Box$
We conclude from Lemma~\ref{u1}:
\vspace{.1in}
\noindent
{\em For a balanced triangle $T$, the first and second
systoles of $\Phi(T)$ are two of the three numbers
$|f(a_i)-f(a_j)|$.}
\vspace{.1in}
It follows that
when a balanced triangle $T$ degenerates, then
the torus $\Phi(T)$ also degenerates, therefore the map $\Phi$ is proper,
and thus $\Phi^*$ is proper as well.
This completes the proof of
Proposition~\ref{prop5}.\hfill$\Box$
\section{Surjectivity of $\Phi^*$}\label{sth3}
We recall that $\mathbf{T}_m$ is the set of BFT with the angle sum $\pi(2m+1)$,
and $\mathbf{T}^*_m=\mathbf{T}_m/\sim$ is the quotient by the following equivalence relation:
(i) we identify triangles obtained from each other by a cyclic permutation
of the three vertices, and
(ii) we identify pairs of triangles described in Proposition \ref{prop3}.
In this section we prove that the map $\Phi^*_m:\mathbf{T}_m^*\to\mathbf{L}_m$
is surjective establishing the first part of
Theorem \ref{theorem3}. Injectivity will be proved in Section~\ref{spaces}.
\vspace{.1in}
{\em Proof of Surjectivity of $\Phi^*$.}
\vspace{.1in}
The plan of the proof is the following: for a given flat singular torus $L$
we find two special geodesic loops whose complement
is a quadrilateral $Q$. Then we construct cell decompositions $C_2,C_4$
of $Q$, and reassembling certain cells of $C_4$ we obtain
a decomposition of $L$ into two congruent balanced triangles.
Let $L$ be a torus with the singular point $O$.
Consider the germ at $O$ of the developing map $f:L\to\mathbf{\overline{C}}$, $f(O)=0$.
Let $g:\mathbf{C}\to L$ be a universal covering with $g(0)=O$. Then the
composition\footnote{Many authors call this $F$ a developing map.}
$$F=f\circ g$$
has a meromorphic continuation
to the whole plane. This meromorphic function
satisfies
\begin{equation}\label{period3}
F(z+\omega)=F(z)+\eta,\quad \omega\in\Lambda.
\end{equation}
Here $\omega\mapsto\eta(\omega)$ is a group homomorphism $\Lambda\to\mathbf{C}$,
and there are
two possibilities:
\vspace{.1in}
\noindent
{\em a) Generic case.} The image of $\Lambda$ is
another lattice $\Lambda'\subset \mathbf{C}$, of rank $2$, and $F:\Lambda\to\Lambda'$
is an isomorphism, or
\vspace{.1in}
\noindent
{\em b) Degenerate case.} The image of $\Lambda$ belongs to a line through the
origin.
\vspace{.1in}
The pull back the flat metric via $F$
has the length element
$$ds=|F'(z)||dz|.$$
This metric has conic singularities at the critical points
of $F$ which are the points of
$\Lambda$, and some poles.
Let $\gamma_1$ be a shortest curve among all curves from $0$ to some point
$\omega\in\Lambda\backslash\{0\}$. We denote its endpoint other than $0$
by $\omega_1$. It is clear that $\gamma_1$ is a simple curve; $F(\gamma_1)$
is a segment $[0,\eta_1]$, where $\eta_1\in\Lambda'\backslash\{0\}$,
and the map $F:\gamma_1\to[0,\eta_1]$ is a homeomorphism.
Let $\gamma_2$ be a shortest of all curves from $0$ to some point
$\omega_2\in \Lambda\backslash\{\mathbf{Z}\omega_1\}$. The following lemma implies
that
$\gamma_1$ and $\gamma_2$ are disjoint except their common endpoint
at $0$.
\begin{lemma}\label{lemma31}
The curves $g(\gamma_1)$ and $g(\gamma_2)$ in $L$ intersect only at $O$.
\end{lemma}
{\em Proof.} Suppose that this is not so, and let $p\neq O$
be a point of intersection. Since $p$ is not a conic singularity,
and our curves are geodesic, they must make a non-zero angle st $p$.
It follows that the ratio of the periods of the differential $df$
over $g(\gamma_1)$ and $g(\gamma_2)$ is not real.
Now we construct a loop $\Gamma$ in $L$ based at $O$ which is shorter than
$g(\gamma_2)$ and whose homology class is not a multiple of $g(\gamma_1)$.
The point $p$ breaks $g(\gamma_1)$ into two arcs, and we denote the shorter
of these arcs by $I_1$. Similarly $I_2$ is the shorter of the two
arcs into which $p$ breaks $g(\gamma_2)$. Let $\Gamma$ be the concatenation
of $I_1$ and $I_2$. From our observations on the periods of $df$ we conclude
that
$$\int_\Gamma df\neq 0,$$
therefore $\Gamma$ is non-trivial. Here we used that $df$ has no residues
at the poles. Moreover, this integral cannot be a real multiple of
the integral over $g(\gamma_1)$, so $\Gamma$ is not a
multiple of $g(\gamma_1)$. Finally, the length of $\Gamma$ is at most
the length of $g(\gamma_2)$, but $\Gamma$ can be shortened since it
has a non-zero angle at $p$, so we obtain a contradiction.\hfill$\Box$
\vspace{.1in}
The loops $g(\gamma_1)$ and $g(\gamma_2)$ cut the torus into
a quadrilateral. Preimage of this quadrilateral under $g$ is a quadrilateral
in the plane
bounded by $\gamma_1,\gamma_2$ and their shifts $\gamma_1+\omega_2$
and $\gamma_2+\omega_1$. From Lemma~\ref{lemma31}
we conclude that all four curves are pairwise
disjoint except their endpoints.
Thus we obtain a Jordan quadrilateral that will be called
$Q$ (the boundary is included).
Since the curves $g(\gamma_1)$ and $g(\gamma_2)$ have intersection index
$\pm1$, they generate the fundamental group of the torus, and it
follows that $Q$ does not contain other lattice points except
$0,\omega_1,\omega_2$ and $\omega_1+\omega_2$.
The image $F(\partial Q)$ consists of $4$ straight segments
which form a parallelogram in the plane in the non-degenerate case.
In the degenerate case these $4$ segments belong to the same line.
Next we study
\vspace{.1in}
{\em Topology of the map $F:Q\to\mathbf{\overline{C}}$.}
\vspace{.1in}
The following argument is purely topological, so we consider
an arbitrary Jordan quadrilateral $Q$ in the plane
(a closed disk with four
distinct marked boundary points
$a_1,\ldots,a_4$, which we call {\em corners},
enumerated
according to the standard orientation). The boundary arcs $(a_i,a_{i+1})$,
where $i$ is a residue modulo 4,
are called the {\em sides}.
Let $F:Q\to\mathbf{\overline{C}}$ be a continuous function which is a local homeomorphism
on the complement of the corners,
and topologically holomorphic\footnote{Topologically equivalent to
$z\mapsto z^{n_i}$, $1\leq i\leq 4$.}
at the corners.
About the boundary behavior we make
one of the two assumptions:
\vspace{.1in}
\noindent
a) Generic case:
$F(\partial Q)$ is a Jordan curve $\gamma$ and $F:\partial Q\to\gamma$
is a homeomorphism, or
\vspace{.1in}
\noindent
b) Degenerate case: the restrictions of $F$ to the sides
are homeomorphisms onto the image of each side, and
these images are segments of the same straight line $\ell$ in $\mathbf{C}$.
The images of opposite sides have equal length.
\vspace{.1in}
We want to obtain a topological description
of possible partitions of $Q$
by $F^{-1}(\gamma)$ in case a) and by $F^{-1}(\ell)$ in case b).
First we address the generic case a). Consider the cell decomposition $C_1$
of the Riemann sphere which has two $2$-cells:
the interior $I$ and the exterior $E$
of
$\gamma$ (Fig.~5d).
The $0$-cells are $F(a_j)$ and $1$-cells are the four arcs
into which $F(a_j)$ divide $\gamma$.
We assign the labels to $0$- and $1$-cells by the
following rules: $F(a_j)$ has label $j$; the arc $(F(a_j),F(a_{j+1}))$
has label $j$.
Now consider the preimage $C_2=F^{-1}(C_1)$ in $Q$.
Our assumptions about $F$ imply that $C_2$ is a finite cell decomposition
of $Q$.
It is called the {\em net} of $F$.
Closures of the cells of $C_2$ are mapped onto the closures
of the cells of $C_1$ homeomorphically,
and we label cells of $C_2$ by their images.
Since $F$ is a local homeomorphism
on $Q\backslash\{ a_j\}$, the $1$-skeleton of $C_2$ consists of simple curves
which can meet only at the corners. We call the intersections
of these curves with the interior of $Q$ {\em arcs} and define the
length of an arc as the number of $1$-cells that it contains. An example
of the cell
decomposition $C_2$ is shown in Fig.~5a, where the black dots are $0$-cells.
The faces of $C_2$ are quadrilaterals, and we classify them as follows:
A face is called {\em lateral} if its boundary consists of one arc
of length $1$ and one arc of length $3$, both arcs having
as endpoints two adjacent corners of $Q$.
A face is called {\em diagonal} if its boundary consists of two arcs
of length $2$ both having as endpoints two opposite corners of $Q$.
A face is called {\em triangular} if its boundary consists of two arcs
of length $1$ and one arc of length $2$, arcs of length $1$ connecting
pairs of adjacent corners, while the arc of length 2
connects opposite corners.
A face is called {\em quadrilateral} if its boundary consists of $4$
arcs of length $1$, each connecting a pair of adjacent corners of $Q$.
Fig.~5a contains 8 lateral, 1 diagonal and 2 triangular faces.
Let us show that
this classification exhausts all possibilities
for the faces of $C_2$.
A face of $C_2$ cannot have all $4$ boundary vertices
in the interior of $Q$, since then there would be an adjacent face which
is not simply connected. Neither a face of $C_2$ can have two vertices
at the same corner, because the restriction of $f$ on the boundary
of a face is a homeomorphism onto $\gamma$. A face of $C_2$ cannot have
only one
vertex at a corner, because if this were the case, an adjacent $2$-cell
will have all its 4 boundary edges the same as the original face, which is
impossible.
\begin{lemma}\label{lemma32}
Under the assumption a) there are the
following possibilities:
(i) The net contains one quadrilateral face and some (possibly none)
lateral
faces.
(ii) The net contains at least one diagonal face, two triangular faces
and several (possibly none) lateral faces.
All diagonal faces share the same
opposite corners on their boundaries.
\end{lemma}
This lemma and its proof are illustrated in Fig.~5. In Fig.~5b
case (i) is illustrated ($C_2$ is shown with bold lines).
Fig.~5a is an example of case (ii).
\vspace{.1in}
{\em Proof.} Notice that lateral faces come in pairs, so the number
of lateral faces
sharing two given corners $a_i,a_{i+1}$ on their boundaries must be even.
So
the innermost arc in $Q$, connecting $(a_i,a_{i+1})$, has length $1$.
Removing all lateral faces, we obtain a smaller quadrilateral
$Q'$, and a cell decomposition $C_3$ of it which has no lateral faces.
The restriction of $f$ to $Q'$ satisfies the
same conditions as $f$ on $Q$: the boundary $\partial Q$ is mapped
onto $\gamma$ homeomorphically.
If $C_3$ consists of a single face, we are in case (i).
If $C_3$ contains a diagonal face, suppose it has $a_1$ and $a_3$ on
the boundary. Then all diagonal faces must have $a_1$ and $a_3$
on their boundaries. Removing all of them, we obtain two triangular faces,
so we are in case (ii).
If $C_2$ contains no diagonal faces, then there are no
triangular faces. Indeed, suppose that the cell decomposition
of $Q'$ consists of just two triangles.
The $1$-cells on the boundary of each triangle have $4$ distinct labels,
and two of these $1$-cells are in the common boundary of these two
triangles.
But the $1$-cells on the boundary of $Q'$ also have $4$ distinct labels,
and this is evidently impossible.
Thus if there are no diagonal faces in $Q'$, then there are also no
triangular faces, and we are in case (i).\hfill$\Box$
\vspace{.1in}
{\em Transformation of the
cell decomposition $C_2$ into another cell decomposition
$C_4$ of $Q$.}
\vspace{.1in}
The edges of $C_4$ are defined as follows. First, they are
arcs of length $1$ of $C_2$. Then we discard all arcs of length at least $2$,
and add new edges by the following rules:
Suppose that $C_2$ has a diagonal face $G$.
We recall that cells of $C_2$ are labeled by their images in $C_1$.
If $G$ is a cell of $C_2$ is labeled $I$, we
draw the {\em diagonal}: the $F$-preimage in $G$ of that diagonal of
the parallelogram $I$ which has two corners of $Q$ as its extremities.
If $G$ is labeled $E$,
we use one of the two
{\em exterior diagonals} of the parallelogram $I$. An exterior
diagonal is the complement to a diagonal in the line
which contains this diagonal.
In Fig.~5b the added diagonal is red (dotted), and the discarded arcs are grey.
If $C_2$ has no diagonal faces, then it has one quadrilateral face.
If this quadrilateral face is labeled $I$,
we break this quadrilateral face by the preimage of the
{\em shorter} diagonal of the
parallelogram $I$. If the quadrilateral face of $C_2$ is labeled $E$
we break this quadrilateral face by the preimage of the
exterior diagonal of $I$ which connects the two vertices of
this parallelogram with the larger exterior corner (Fig.~5g).
Partition of a quadrilateral face is shown if Fig.~5c.
By these rules, we obtain a cell decomposition $C_4$ of $Q$ which has
no interior vertices. This decomposition contains two congruent triangles
and a number of digons. Each digon is mapped to the sphere
with a cut along a segment. We break it into two digons by the $F$-preimage
of the complement of this segment in the line that contains it.
These lines are shown as red/dotted lines in Fig.~5b,c.
After these cuts are made, the number of digons in every ``bunch''
becomes even.
Adding half of them to the adjacent side of the triangle
we obtain a decomposition of our torus into
two triangles.
The final decomposition of $Q$ into two primitive triangles
and digons isometric to half-planes
is shown in red/dashed and bold black lines in Figs.~5b,c in two cases:
4b) when $C_2$ has a diagonal face, and 4c) when it does not.
Now we show that these two triangles are balanced.
We refer to the decomposition of a singular triangle described
in Proposition~\ref{prop2}.
Gluing any numbers of half-planes to the sides of a balanced triangle
results in a balanced triangle. Primitive triangles are
balanced in the following cases. Primitive triangle of the type A
is balanced if all angles are less than $\pi/2$. If the greater angle
is $>\pi/2$ and at least one half-plane is glued to the opposite side,
the resulting triangle is balanced. If the cell decomposition $C_2$
contains a diagonal face, this implies that at least one half-plane
was glued opposite the largest angle of the triangular face.
If the triangular face
is of the type A, then its longest side is the diagonal, so
the largest angle is opposite to it.
If this triangle is of type B, then it is
balanced (a triangle of this type is always balanced).
If there was no diagonal in $C_2$, then we obtained triangular faces of
$C_4$ by drawing either the smaller diagonal in a parallelogram,
or the exterior diagonal in its exterior which has endpoints
at the bigger
exterior angles. In both cases the triangle is balanced.
\vspace{.1in}
So we obtained a partition of $Q$ into two balanced triangles.
We can re-assemble it by moving digons adjacent to a side of $Q$
to the opposite side to make the two balanced triangles congruent.
This completes the proof in the non-degenerate case.
\vspace{.1in}
Now we consider the degenerate case. $F:Q\to\mathbf{\overline{C}}$ maps the sides
of $Q$ into a line $\ell$, and we assume without loss of generality
that $\ell=\mathbf{R}\cup\{\infty\}$. The images of sides occupy some segment
$(a,b)\in\mathbf{R}$, where $a<b$. It is evident that $a$ and $b$ are the
images of two opposite corners of $Q$. Without loss of generality,
these corners are $a_1$ and $a_3$.
The preimage
$F^{-1}(\mathbf{R})$ defines a cell decomposition $C_5$ of $Q$.
It is exactly of the same type as nets studied in \cite{EG}:
they consist of simple curves with endpoints at the corners and
disjoint interiors,
and each curve is mapped homeomorphically onto its image.
\begin{lemma}\label{lemma33} Under these assumptions $C_5$ contains
a curve from $a_1$ to $a_3$. So the faces of $C_5$ are two triangles
and several (possibly none) digons.
\end{lemma}
{\em Proof.} Any component of $F^{-1}((\mathbf{R}\backslash[a,b])\cup\{\infty\})$
must be a curve from $a_1$ to $a_3$ in the interior of $Q$. This proves
the lemma.
\vspace{.1in}
Each digon of $C_5$ is mapped by $F$ to $\mathbf{\overline{C}}$ with a cut
(bounded or
containing $\infty$). We partition digons into two halves by complements
of these cuts to the $\mathbf{R}\cup\{\infty\}$. Then we split these half-planes
in each ``bunch'' into two equal parts and add them to the corresponding
sides of triangular faces. This defines a decomposition of
our torus into two triangles. That these triangles are balanced
is proved in the same way as in the non-degenerate case.
\hfill$\Box$
This completes the proof of surjectivity of $\Phi^*$.\hfill$\Box$
\vspace{.1in}
\section{The spaces $\mathbf{A}_m$, $\mathbf{T}_m$ and $\mathbf{T}_m^*$}\label{spaces}
\subsection{Connected components}
To visualize Proposition~\ref{prop1}, we introduce
the {\em space of angles} $\mathbf{A}_m$. In the intersection
of the plane
\begin{equation}\label{plane}
P=\{\alpha\in\mathbf{R}^2:\alpha_1+\alpha_2+\alpha_3=2m+1\},
\end{equation}
with the open first octant in $\mathbf{R}^3$ (Fig.~9) we consider the triangle
$\Delta_m$ defined by the inequalities
$$0<\alpha_j\leq\alpha_i+\alpha_k\quad\mbox{for all permutations}\quad (i,j,k);$$
it is shaded in Fig.~9.
The vertices of $\Delta_m$ are
$$(m+1/2,m+1/2,0), \quad (m+1/2,0,m+1/2), \quad (0,m+1/2,m+1/2).$$
Notice that the vertices do
not belong to $\Delta_m$ but the sides do belong, so $\Delta_m$
is neither open nor closed.
To obtain $\mathbf{A}_m$ we remove from $\Delta_m$ all lines where some $\alpha_j$
is an integer, and add all points where
all three $\alpha_j$ are integers.
The intersections of lines $\alpha_j=k$ with
$\Delta_m$ will be called {\em segments}. A segment is called even
or odd depending on the parity of $k$. There are three
families of parallel segments, each containing $m$ segments. Spaces of
angles for $m=1,\ldots,5$ are shown in Figs.~6, 7.
The set $\mathbf{A}_m$ has a natural partition
into open topological disks (faces) open intervals (edges)
and points (vertices): the faces are components
of the interior of $\mathbf{A}_m$ (they are triangles or quadrilaterals), the vertices
are the points where all $\alpha_j$ are integers, and the edges
are open intervals in $\mathbf{A}_m\cap\partial\Delta_m$. The set of
vertices of $\mathbf{A}_m$ will be denoted by $\mathbf{V}_m$.
We have a natural projection
\begin{equation}\label{phi}
\varphi:\mathbf{T}_m\to\mathbf{A}_m
\end{equation}
which to every
balanced triangle with angles $(\pi\alpha_1,\pi\alpha_2,\pi\alpha_3)$
puts into correspondence the point
$(\alpha_1,\alpha_2,\alpha_3)\in\mathbf{A}_m\subset\mathbf{R}^3$. It follows
from Proposition~\ref{prop1} that this correspondence maps
the part of $\mathbf{T}_m$ where $\alpha_j$ are not integers bijectively
onto $\mathbf{A}_m\backslash\mathbf{V}_m$.
Triangles with integer angles
are mapped to $\mathbf{V}_m$
and the preimage of each point in $\mathbf{V}_m$
consists of three intervals.
This induces a partition of $\mathbf{T}_m$: the faces of $\mathbf{T}_m$ are $\varphi$-preimages
of the faces of $\mathbf{A}_m$, the edges are of two types: interior edges
which are mapped by $\varphi$ to the vertices of $\mathbf{A}_m$ and boundary edges
which are mapped bijectively onto intervals of $\mathbf{A}_m\cap\partial \Delta_m$.
There
are no vertices in this partition of $\mathbf{T}_m$.
The faces of $\mathbf{T}_m$ are adjacent when their images in $\mathbf{A}_m$
share a boundary vertex and
{\em their angles at this vertex are vertical\,\footnote{Opposite angles
among the four angles made by crossing of two lines.}.}
Notice that the map $\phi$ switches the orientation when one passes through
any interior edge of $\mathbf{T}_m$ from a face to an adjacent face.
This can be seen from the explicit formula for the angles in terms of
the conformal coordinate $z=\phi_{i,j,k}$ introduced in Section~\ref{complexanalytic}:
in the chart where $f(a_1)=0,\; f(a_2)=1,\; f(a_3)=z=x+iy$ we have
$$\alpha_1=p+\arctan(y/x),\quad \alpha_2=q+\arctan(y/(1-x)),$$
where $p,q$ are integers. Assuming that $x\in(0,1)$ we compute the Jacobian
and see that it switches sign simultaneously with $y$.
\vspace{.1in}
\noindent
{\bf Remark.} Gluing of two $2$-cells along their common
boundary $1$-cell corresponding to a vertex of $\mathbf{A}_m$
reverses the natural
orientation of these $2$-cells induced from the
$(\alpha_1,\alpha_2)$-plane. Nevertheless,
it is easy to check that the surface $\mathbf{T}_m$ is orientable. To
do this one paints the $2$-cells of $\mathbf{T}_m$ into two colors,
so that each two $2$-cells with a common vertex have different
colors. It is clear that such a coloring is possible, see Fig.~8.
\vspace{.1in}
To study the surface $\mathbf{T}_m$ we introduce the graph $\Gamma_m$, which will
be called the {\em nerve}.
Examples of these graphs are shown in figures 6, 7.
Their vertices correspond to $2$-cells of $\mathbf{T}_m$ (or faces of $\mathbf{A}_m$)
and two vertices of $\Gamma_m$ are
connected by an edge if the corresponding two $2$-cells of $\mathbf{T}_m$ share
an edge or, which is the same, if the corresponding $2$-cells
of $\mathbf{A}_m$ share a vertex and their angles at this vertex are vertical.
Then we find
\begin{prop}\label{prop6}
When $m\geq 2$, the graph $\Gamma_m$ has
$4$ connected components. Exactly one of them, $\Gamma_m^\prime$ is
invariant under the order $3$ rotation about the center
of $\mathbf{A}_m$. Three others are permuted by this rotation.
$\Gamma_1$ has only three components, permuted by
the rotation. $\Gamma_0$ consists of one vertex only.\hfill$\Box$
\end{prop}
In Fig.~6, $\Gamma_m^\prime$ is blue/dotted, while in Fig.~7
one of the three components permuted by the order $3$ rotation
is red, any of these three components is called $\Gamma_m^{\prime\prime}$
(they are isomorphic graphs embedded in the plane).
We give first a geometric sketch which makes our proposition evident,
and then a more formal proof.
\vspace{.1in}
{\em Sketch of a proof of Proposition~\ref{prop6}.} Let us consider the plane
$P$ which is defined in (\ref{plane}).
Intersections of $P$ with the planes $\{\alpha_j=\mbox{integer}\}$ break $P$
into triangles.
Connecting the centers of pairs of triangles which share a vertex
and whose angles at this vertex are vertical, we obtain four honeycomb
structures $X_j$ with disjoint vertices. See Fig.~8 which shows two
of these honeycombs.
Choosing one vertex of one honeycomb as a center, we see that this honeycomb
is invariant under rotation by $120^o$ about this vertex,
while the other three are permuted.
\vspace{.1in}
{\em Proof of Proposition~\ref{prop6}.}
\vspace{.1in}
Let $\Delta_m^\prime$ be the intersection of the plane
$$\alpha_1+\alpha_2+\alpha_3=2m+1$$
with the closed first octant
$\alpha_j\geq 0,\; 1\leq j\leq 3.$
The {\em segments}
$$\{(\alpha_1,\alpha_2,\alpha_3)\in\Delta_m^\prime:\alpha_j=k\},
\quad 1\leq j\leq 3,\quad 0\leq k\leq 2m$$
divide $\Delta_m^\prime$ into open triangles which we call {\em faces}
of $\Delta_m^\prime$. There are three families of
these segments, depending on the value of $j$ which were discussed
in the beginning of this section.
We recall that a segment with even/odd $k$ is
called {\em even/odd}.
Since $\alpha_1+\alpha_2+\alpha_3$ is odd, among three segments intersecting
at an integer vertex, either one or all of them are odd. This implies
that a face has either one or three sides on even segments.
We classify faces of $\Delta_m^\prime$ into four types:
\vspace{.1in}
\noindent
type $I$, if all three sides of the triangular face belong to even segments,
\vspace{.1in}
\noindent
type $II_j,\; j\in\{1,2,3\}$
if one side belongs to an even segment of family $j$, while two
other sides belong to odd segments of the other two families.
\vspace{.1in}
Two faces sharing an integer vertex are called {\em vertical}
if their
angles at this vertex are vertical (opposite).
We claim that vertical faces are of the same type.
Indeed if the two segments bounding the vertical angles are both
even, then each of our two faces must have all three sides on even segments,
thus both faces are of type $I$. If exactly one of the segments
is even, and belongs to family $j$, then both faces are of the type
$II_j$. If both segments are odd, then the sides of our faces
opposite to the considered vertex are even and parallel, so they
are in the same family and our two faces are in the same family.
This proves the claim.
Now we claim that faces of the same type cannot have a common side.
If two faces have a common side on an even segment then one of them is
of type $I$ and another is of type $II$. If the common side is
on an odd segment then both faces are of type $II$ and their sides
on even segments
are not parallel. Thus they belong to different types.
This proves the claim.
Our next claim is that the closure of the union of faces of the same type
is connected. Consider faces of type $I$. The closure of their union
consists of the faces themselves and all even segments. It is clear
that the union of all even segments is connected.
The proof for other types is similar: the closure of the
union of faces of type $II_j$ consists of the faces themselves,
the even segments of family $j$,
and odd segments of two other families. The union of these
segments is connected.
If we restrict now to $\mathbf{A}_m$ and consider the union of those faces
of a single type which intersect $\mathbf{A}_m$ and their vertices,
this union is still connected. Indeed if two faces of one family
share a vertex and both intersect $\mathbf{A}_m$ then this common vertex belongs
to $\mathbf{A}_m$.
This proves Proposition~\ref{prop6}.\hfill$\Box$
\vspace{.1in}
Now consider the map $\varphi:\mathbf{T}_m\to\mathbf{A}_m$. Component $\mathbf{L}_m^I$ of $\mathbf{T}_m$
consists of preimages of faces of type $I$ of $\mathbf{A}_m$ and common
edges of pairs of these preimages that project to the vertices of
faces of type $I$ of $\mathbf{A}_m$.
Similarly for $i\in\{1,2,3\}$, components $II_i$ of $\mathbf{T}_m$,
consist of preimages of faces of type $II_i$ and common edges of pairs of
these preimages which project to the vertices of faces of type $II_i$
of $\mathbf{A}_m.$ So we obtain
\begin{cor}
When $m\geq 2$, $\mathbf{T}_m$ consists of four connected components
\newline
$I,II_1,II_2,II_3$.
Cyclic permutation of vertices preserves component $I$ and permutes
components $II_i$. As a consequence, $\mathbf{T}_m^*$ has two components, $I$ and $II$.
These components
are distinguished by the number of sides with poles:
When $m$ is even,
$\mathbf{L}_m^I$ consists of BFT with no poles on the sides, and
$\mathbf{L}_m^{II}$ consists of BFT with two poles on the sides.
When $m$ is odd, $\mathbf{L}_m^I$ consists of BFT with 3 poles on the sides,
and
$\mathbf{L}_m^{II}$ consists of BFT with one pole on
the side. \hfill$\Box$
\end{cor}
\noindent
{\bf Example.}
Figure 2 shows all types of BFT for $m=2$ (angle sum $5\pi$).
Triangles of types a), b), c) belong to $\mathbf{L}_m^I$.
Suppose for example that triangle a) is deformed so that the top
vertex moves towards the opposite (horizontal) side. Eventually we obtain
a triangle b) with the angles $2\pi,2\pi,\pi$. If the middle vertex of b)
continues
moving downwards, we obtain triangle c). Its developing map
is $2$-to-$1$ onto the darkly shaded region and $1$-to-$1$ onto the lightly
shaded region. There are three types of such triangles c) if the
vertices are labeled, but only one type with unlabeled vertices.
Triangles of types d), e), f) belong to $\mathbf{L}_m^{II}$. It is easy to
visualize how they are deformed to each other.
Triangles a), b), c) have one pole inside, while triangles d), e), f)
have 2 poles, one on each of the two unbounded sides, and the third side
is free of poles. Fig.~2 should be compared with Figs.~6, 7, $m=2$:
the set $\mathbf{A}_2$ shows the location of all these triangles
in the parameter space.
\subsection{Proof of injectivity of $\Phi^*$}
We established in Section~\ref{complexanalytic}
that $\Phi_m^*:\mathbf{T}_m^*\to\mathbf{L}_m$
is a proper holomorphic map between punctured Riemann surfaces. So to prove injectivity
it is sufficient to show that every component of $\mathbf{T}^*_m$ contains an interval $I$
such that for every $T\in I$,
$T$ is the unique $\Phi^*$-preimage of $\Phi^*(T)$.
Assume that $m\geq 1$. Then every component contains a triangle
with integer angles.
Let $T$ be a triangle with integer angles.
According to Proposition~\ref{prop2.5}
the monodromy group of $\Phi^*(T)$ is a subgroup of a line,
so for the torus $\Phi^*(T)$ alternative
b) (degenerate case) holds in the proof of surjectivity.
It follows that any $\Phi^*$-preimage of $\Phi^*(T)$ also
has all integer angles (see Proposition~\ref{prop2.5}).
Recall that a triangle with integer angles is obtained from a triangle
in Fig.~1b in which $a_1,a_2,a_3$ belong to the same line, by gluing
half-planes to the sides (see Proposition~\ref{prop2}).
Let us normalize so that $a_3=0$, $a_1=1$, and we choose our interval $I$
so that
$a_2:=a\in(0,1/2)$. Then the following properties
of $\Phi^*(T)$ are evident:
The shortest non-trivial loop
$\gamma_1$ based
at the singularity
has length $a_2$. We define the orientation
of this loop by orienting $(0,a)$ from $a$ to $0$
on the boundary of the reduced triangle $T'$.
We recall that the reduced triangle $T'$ is the upper half-plane
with corners at $0,a,1$.
So the parameter $a$ is uniquely defined by $\Phi^*(T)$,
as the shortest length of a loop based at $O$ on $\Phi^*(T)$.
Now we define $\gamma_2$ as the loop corresponding to
$(a,1)$ in $\partial T'$, oriented from $a$ to $1$.
This loop $\gamma_2$ is characterized as the shortest loop
whose class does not belong to $\mathbf{Z}\gamma_1$.
Each side of $T'$
defines a homotopy class of loops based at $O$. Two of them
are $\gamma_1$ and $\gamma_2$.
Suppose that $m_1$ half-planes were glued to the side $(0,a)$,
and $m_2$ half-planes were glued to the side $(a,1)$.
Then the torus $\Phi^*(T)$ contains $m_1+1$ disjoint (except the base point)
geodesic loops in the class $[\gamma_1]$, and $m_2+1$ disjoint
geodesic loops in the class $[\gamma_2]$.
This implies that the angles $\pi\alpha_i$ of $T$ are defined by
the properties of the torus $\Phi^*(T)$, namely $\alpha_i=1+m_j+m_k$.
This proves injectivity of
the map $\Phi^*$ and completes the proof of Theorem~\ref{theorem3}.
\section{Euler characteristics of components of $\mathbf{L}_m$
and completion of the proof of Theorem~\ref{theorem1}}\label{ident}
Theorem~\ref{theorem3} reduces the study of topology of $\mathbf{L}_m$ to the study of
topology of $\mathbf{T}_m^*$.
It is convenient to use the nerves
$\Gamma_m^\prime$ and $\Gamma_m^{\prime\prime}$
introduced in Section~\ref{spaces}.
First we recall that $\Gamma_m/\mathbf{Z}_3$ consists of $2$ components. One of them
comes from the component $\Gamma_m^\prime$ which is invariant with
respect to the $\mathbf{Z}_3$ action.
This is our component $\mathbf{L}_m^I$. Component $\mathbf{L}_m^{II}$
comes from
the three components of $\Gamma_m^{\prime\prime}$ which are permuted by the $\mathbf{Z}_3$ action.
See
Figs. 6, 7.
\vspace{.1in}
\noindent
{\em Computation of the Euler characteristic for component $\mathbf{L}_m^I$.}
\vspace{.1in}
Let $\Gamma_m^\prime$ be the component of $\Gamma_m$ which is invariant
with respect to the $\mathbf{Z}_3$ action.
The numbers $\epsilon_0,\epsilon_1$ are defined in (\ref{epsilon0})
and (\ref{epsilon2}).
We add to them
$$\epsilon_2=\left\{\begin{array}{ll}0,&\mbox{if}\; m\equiv 1\; (\mathrm{mod\,} 2)\\
1,&\mbox{if}\; m\equiv 0\; (\mathrm{mod\,} 2).\end{array}\right.$$
and interpret these numbers in terms of
$\mathbf{A}_m$:
\vspace{.1in}
$\epsilon_0=1$ if the center of $\mathbf{A}_m$ belongs to a $2$-cell.
Equivalent condition is that a vertex of $\Gamma_m^\prime$
is fixed by the $\mathbf{Z}_3$
action.
\vspace{.1in}
$\epsilon_1=1$ if there is a vertex of $\Gamma_m^\prime$ representing
a face which has the
middle of the side of $\Delta_m$ on the boundary.
\vspace{.1in}
$\epsilon_2=1$ if there is a vertex of $\Gamma_m^\prime$ representing
a face which has a corner
of $\Delta_m$ on the boundary.
\vspace{.1in}
\vspace{.1in}
We introduce further notation:
$V_1,V_2,V_3$ are the numbers of vertices of $\Gamma_m^\prime$ of
degrees $1,2,3$, and $V$ is the total number of vertices.
$E$ is the number of edges of $\Gamma_m^\prime$.
Taking into account all identifications on $\Gamma_m^\prime$, we obtain
the following formula for the Euler characteristic:
\begin{equation}\label{chiI}
\chi(\mathbf{L}_m^I)=
V_3/3+V_1/6+V_2/6-E/3+2e_0/3+(\epsilon_1/2-\epsilon_2)/2.
\end{equation}
To explain this formula, we compute the contributions to the
(ordinary) Euler characteristic of $\mathbf{L}_m^I$.
The group $\mathbf{Z}_3$ fixes the center of $\mathbf{A}_m$ which belongs to $\Gamma_m^\prime$
if and only if $\epsilon_0=1$, and acts freely on the set of remaining
$3$-valent vertices, so the number of $2$-cells in $\mathbf{L}_m^I$ corresponding to
$3$-valent vertices is $V_3/3+2\epsilon_0/3$.
Also, the group $\mathbf{Z}_3$ acts freely on
the set of edges, and thus there are $E/3$
corresponding $1$-cells in $\mathbf{L}_m^I$.
For a $1$-valent or $2$-valent vertex $v$ of $\Gamma_m^\prime$, its
class under the equivalence relation consists of
\begin{itemize}
\item[a)]
6 elements, if $v$ is neither a corner of $\Delta_m$ nor a midpoint
of a side of $\Delta_m$;
\item[b)]
3 elements, if $v$ is a corner of $\Delta_m$;
\item[c)]
3 elements, if $v$ is the midpoint of a side of $\Delta_m$.
\end{itemize}
If we have $V^a$ vertices of type a), such vertices correspond to
$V^a/6$ $2$-cells in $\mathbf{L}_m^I$.`The number $V^b$ of vertices
of type b) can be $0$ or $3$: the latter case occurs if and only if
$\epsilon_2=1$, in which case such vertices correspond to a punctured disk in
$\mathbf{L}_m^I$. In both cases, vertices of type b) contribute $V_b/6-\epsilon_2/2=0$
to the Euler characteristic. Similarly, the number $V^c$ of vertices
of type c) can be $0$ or $3$: the latter case occurs
if and only if $\epsilon_1=1$, in which case such vertices correspond to a disk
in $\mathbf{L}_m^I$. In both cases, vertices of type c)
contribute $V^c/6+\epsilon_1/2=1$
to the Euler characteristic. Adding these contributions, we obtain (\ref{chiI}).
It is easy to see that for odd $m$
$$E=3(m^2-1)/8,\quad V_1=3(m-1)/2,\quad V_2=0,\quad V=(m^2+4m-5)/4,$$
and $V_3$ can be computed by the formula $V_3=V-V_1-V_2$.
This gives the formula for $\chi(\mathbf{L}_m^I)$ when $m$ is odd.
When $m$ is even, we have
$$E=3(m^2+2m)/8,\quad V_1=3,\quad V_2=3(m-1)/2,\quad V=(m/2+1)^2,$$
and again $V_3=V-V_1-V_2$. This gives the formula for $\chi(\mathbf{L}_m^I)$ when $m$
is even. The resulting formulas for $\chi$ in terms of
$m$ are written in the Appendix. Expressing $m$ in terms
of $d_m^I$ in (\ref{degFI})
and subtracting the orbifold correction we obtain
$\chi^O(\mathbf{L}_m^I)=-(d_m^I)^2/6.$
\vspace{.1in}
{\em Computation of the Euler characteristic of component $\mathbf{L}_m^{II}$}
\vspace{.1in}
Let $\Gamma_m^{\prime\prime}$ be one of the three components of $\Gamma_m$
which are permuted by the $\mathbf{Z}_3$ action. We use the following notation
$E$ is the number of edges of $\Gamma_m^{\prime\prime}$
$V_1,V_2,V_3$ and $V$ are the numbers of vertices of $\Gamma_m^{\prime\prime}$
of orders $1,2,3$ and the total number of vertices.
$\epsilon_1$ and $\epsilon_2$ have the same meaning as before.
A computation analogous to that for $\mathbf{L}_m^I$ gives
$$\chi(\mathbf{L}_m^{II})=V_3+V_1/2+V_2/2-E+(\epsilon_2-\epsilon_1)/2,$$
as vertices of type b) (respectively, of type c)) belong to
$\mathbf{L}_m^{II}$ if and only if $\epsilon_2=0$ (respectively $\epsilon_1=0$).
When $m$ is odd,
$$E=(3m^2+4m+1)/8,\quad V_1=(m+3)/2,\quad V_2=m-1,\quad V=(m^2+4m+3)/4.$$
When $m$ is even,
$$E=(3m^2+2m)/8,\quad V_1=m,\quad V_2=m/2,\quad V=(m/2+1)^2-1.$$
This gives the formulas for $\chi(\mathbf{L}_m^{II})$
in terms of $m$ and $\epsilon_1$
(written in the Appendix). Expressing $m$ in terms of $d_m^{II}$
from (\ref{degFII}) and subtracting the orbifold correction
we obtain $\chi^O(\mathbf{L}_m^{II})=-(d_m^{II})^2/18$.
\vspace{.1in}
{\em Computation of the number of punctures.}
\vspace{.1in}
Consider a small simple loop around a puncture of $\mathbf{T}^*_m$. This loop projects
to a contour in $\mathbf{A}_m$ which goes close to
the lines $\alpha_j=k$, switching the side at each integer point.
For component $\mathbf{L}_m^I$, the contour goes near lines with the same even $k$,
and $j=1,2,3$ and closes. See Fig.~10. So there is a $1-1$ correspondence between these
contours and triples of segments (one in each family) with even $k$.
So there are $\lfloor m/2\rfloor$ of such loops. In addition,
when $m$ is even there is a puncture corresponding to the vertices
of $\Delta_m$. Thus the total number of punctures on $\mathbf{L}_m^I$
is $m/2+1$ when $m$ is even and $(m-1)/2$ when $m$ is odd.
In other words, the number of punctures on Component $\mathbf{L}_m^I$ equals
\begin{equation}\label{holesI}
h_m^I:=d_m^I,
\end{equation}
where $d_m^I$ was defined in (\ref{degFI}).
For component $\mathbf{L}_m^{II}$, the computation is similar, see Fig.~11.
Each contour goes either near an even segment, in which case it closes
after describing three segments, one of each family. If a contour accompanies
an odd segment, it ends on the other side of the odd segment after
describing three segments. So the total number of contours is
$m$ when $m$ is even and $m+1$ when $m$ is odd (the extra puncture for odd $m$
coming from the corners of $\Delta_m$), in other words
\begin{equation}\label{holesII}
h_m^{II}:=2\lceil m/2\rceil=2d_m^{II}/3.
\end{equation}
\begin{center} Component $\mathbf{L}_m^I$
\vspace{.1in}\nopagebreak
\nopagebreak
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|}\hline
$m$&$\epsilon_0$&$\epsilon_1$&$\epsilon_2$&$V_1$&$V_2$&$V_3$&$E$&$V$&$\chi$&$h$&$g$&$d$\\
2&1&0&1&3&0&1&3&4&0&2&0&2\\
3&1&1&0&3&0&1&3&4&1&1&0&1\\
4&0&1&1&3&3&3&9&9&-1&3&0&3\\
5&1&0&0&6&0&4&9&10&0&2&0&2\\
6&1&0&1&3&6&7&18&16&-2&4&0&4\\
7&0&1&0&9&0&9&18&18&-1&3&0&3\\
8&1&1&1&3&9&13&30&25&-3&5&0&5\\
9&1&0&0&12&0&16&30&28&-2&4&0&4\\
10&0&0&1&3&12&21&45&36&-6&6&1&6\\
11&1&1&0&15&0&25&45&40&-3&5&0&5\\
12&1&1&1&3&15&31&63&49&-7&7&1&7\\
13&0&0&0&18&0&36&63&54&-6&6&1&6
\end{tabular}
\end{center}
\vspace{.1in}
\begin{center} Component $\mathbf{L}_m^{II}$\nopagebreak
\vspace{.1in}\nopagebreak
\nopagebreak
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|}\hline
$m$&$\epsilon_1$&$\epsilon_2$&$V_1$&$V_2$&$V_3$&$E$&$V$&$\chi$&$h$&$g$&$d/3$\\
1&0&0&2&0&0&1&2&0&2&0&1\\
2&0&1&2&1&0&2&3&0&2&0&1\\
3&1&0&3&2&1&5&6&-2&4&0&2\\
4&1&1&4&2&2&7&8&-2&4&0&2\\
5&0&0&4&4&4&12&12&-4&6&0&3\\
6&0&1&6&3&6&15&15&-4&6&0&3\\
7&1&0&5&6&9&22&20&-8&8&1&4\\
8&1&1&8&4&12&26&24&-8&8&1&4\\
9&0&0&6&8&16&35&30&-12&10&2&5\\
10&0&1&10&5&20&40&35&-12&10&2&5\\
11&1&0&7&10&25&51&42&-18&12&4&6\\
12&1&1&12&6&30&57&48&-18&12&4&6
\end{tabular}
\end{center}
We include two tables for $1\leq m\leq 13$.
Notation, besides that already introduced is: $g$ for the genus,
$h$ for the number of punctures, $d$ for the degree of the forgetful map
as in (\ref{degFI}), (\ref{degFII}).
The formulas for the degrees follow from \cite[sections 23.21-23.24]{WW}.
\vspace{.1in}
{\bf Remarks.} There is an alternative method of counting the punctures,
based on the description on compactifications of the spaces
of Abelian differentials in \cite{Ba,Ba2}. In recent preprints
\cite{M1}, \cite{M2} a general method of computation
of Euler's characteristics
for spaces of Abelian differentials with prescribed multiplicities
of zeros and poles is developed. However, our results do not
follow from the results stated in these preprints, mainly because
of the additional condition that residues vanish.
\vspace{.1in}
{\em Orbifold points.}
\vspace{.1in}
By definition, an orbifold point in $\mathbf{L}_m$ is a point which corresponds
to a flat singular torus with a non-trivial automorphism.
An automorphism here means an orientation-preserving isometry.
The trivial automorphism is the involution which exists on every flat singular
torus. There are two types of tori with non-trivial automorphisms:
hexagonal ones with an automorphism
of order $3$, and square ones, with non-trivial automorphism of order $4$.
In the representation of tori as $\Phi^*(T)$, hexagonal tori correspond to
triangles whose all angles are equal, while square tori correspond to
marginal triangles whose two smaller angles are equal.
So in the space of angles $\mathbf{A}_m$, the hexagonal torus arises from the center
of $\Delta_m$ when this center belongs to $\mathbf{A}_m$, and the square torus
corresponds to the middles of the sides of $\Delta_m$. In Figs.~6, 7,
these points are denoted by little circles in the center of the picture,
and little black triangles in the middles of the sides.
In the next section we will use the following
\begin{prop}\label{special}
In the Lam\'e equation (\ref{lame1}) or (\ref{lame2})
corresponding to a hexagonal
or square torus
(in the metric sense), the accessory parameter $\lambda$ is
equal to $0$ (see
the text after (\ref{scaling})).
\end{prop}
{\em Proof.} Since a metric automorphism is also a conformal automorphism,
it corresponds to an automorphism of
the Lam\'e equation, that is to a fixed point of
transformation (\ref{scaling}). For both fixed points we have $\lambda=0$.
\hfill$\Box$
\section{Theorem \ref{theorem2} and Maier's conjecture}\label{nons}
To prove Theorem~\ref{theorem2} and its corollaries we first state
the exact relation between $\mathbf{L}_m$ and $\mathbf{H}_m$.
A {\em marked} elliptic curve is an elliptic curve on which the three
points of (exact) order $2$ are labeled. Legendre's family (\ref{le})
parametrizes marked elliptic curves: the labels are $0,1,a$.
The permutation group $S_3$ acts on the space of marked elliptic curves
by permuting the labels. Explicitly, the orbit of $a$ under this
action is
\begin{equation}\label{cross}
a,\quad 1-a,\quad 1/a,\quad 1-1/a,\quad 1/(1-a),\quad a/(a-1).
\end{equation}
This action lifts to the moduli space $\mathbf{C}\times\mathbf{C}_a$ of Lam\'e equations
in the form of Legendre: the generators $a\mapsto 1-a$ and $a\mapsto1/a$
lift to
$$(B,a)\mapsto\left(-B-m(m+1),\; 1-a\right),\quad\mbox{and}$$
$$(B,a)\mapsto\left(B/a,\;1/a\right).$$
To obtains these two transformations,
one changes the independent variable in
(\ref{legendre})
$z\mapsto 1-z$ and $z\mapsto z/a$, respectively.
Taking the quotient of the space $\mathbf{C}\times\mathbf{C}_a$ of equations
(\ref{legendre}) by this $S_3$ action
we obtain an orbifold covering $\Psi_m$ of degree $6$
from the moduli space
of equations (\ref{legendre}) to the moduli space $\mathrm{Lame}_m$,
such that the following diagram is commutative:
\begin{equation}\label{cd}
\begin{CD}
\mathbf{H}_m^j @>{\displaystyle\Psi_m^j}>> \mathbf{L}_m^K\\
@VV{\displaystyle\sigma_m} V @VV{\displaystyle\pi_m} V\\
\mathbf{C}_a @>{\displaystyle\psi}>> \mathbf{C}_J
\end{CD}
\end{equation}
Here $\Psi_m^j$ are restrictions of $\Psi_m$ on $\mathbf{H}_m^j$, and $K=I$
for $j=0$, $K=II$ for $j\in\{1,2,3\}$.
(We have not proved yet that $\mathbf{H}_m^j$ are irreducible; this will be
done only in the end of this section).
The explicit expression of $\psi$ is in (\ref{Ja}). To obtain an explicit
expression of $\Psi_m$ we change the independent variable $z$ in the equation
(\ref{legendre}) to $z-(1+a)/3$. Then we easily obtain $\Psi_m=(R_1,R_2,R_3)$
modulo scaling (\ref{scaling}), where
\begin{eqnarray}\label{R}
\lambda&=&R_1(B,a):=B+m(m+1)(a+1)/3,\\
\nonumber
g_2&=&R_2(B,a):=4(a^2-a+1)/3,\\
\nonumber
g_3&=&R_3(B,a):=8(a^3-3a^2/2-3a/2+1)/27.
\end{eqnarray}
We define compact Riemann surfaces $\overline{\mathbf{L}}_m^K$, $\overline{\mathbf{H}}_m^j$,
$\mathbf{\overline{C}}_J$ and $\mathbf{\overline{C}}_a$ by filling the punctures.
Later we will endow them
with orbifold structures.
The forgetful maps $\pi_m,\sigma_m$ and maps $\psi,\; \Psi$ extend uniquely to these compactifications.
\def\overline{\L}{\overline{\mathbf{L}}}
\def\overline{\H}{\overline{\mathbf{H}}}
\begin{definition}\label{def3} A point $x\in\overline{\L}_m$ is called special if
$\pi_m(x)\in\{0,1,\infty\}$. A point $x\in\overline{\H}_m$ is called special
if
$$\sigma_m(x)\in\{0,1,\infty,2,1/2,-1,(1\pm i\sqrt{3})/2\}.$$
\end{definition}
We will later show (in the proof of Corollary 1.1 in this section)
that $\Psi_m^j:\mathbf{H}_m^j\to\mathbf{L}_m^j$ are orbifold coverings,
so the maps $\Psi_m^j:\overline{\H}_m^j\to\overline{\L}_m^K$, as maps between Riemann
surfaces, can be ramified only at special
points.
Next we study ramification properties of forgetful maps at the special
points. For this we need two lemmas,
the first one is classical, see for example
\cite[Ch. II, \S 1, Thm 1]{GK}:
\begin{lemma}\label{lemma1}
Let $A=(a_{i,j})$ be an $n\times n$ matrix with
$a_{i,i+1}>0,\; 1\leq i\leq n-1,$ and $a_{i,i-1}>0,\;2\leq i\leq n,$
the rest of the entries are zeros. Then all roots of the characteristic
polynomial are real and simple. The characteristic polynomial
is either even or odd, in other words it has the form
$\lambda^kP(\lambda^2),$ where $k\in\{0,1\}$, and $P$ is a real
polynomial.\hfill$\Box$
\end{lemma}
The second lemma was communicated to us by V. Tarasov; it is inspired by
\cite[Prop. 3]{ST}:
\begin{lemma}\label{lemma2}
Let $A=(a_{i,j})$ be an $n\times n$ matrix with
$a_{i,i+1}>0,\; 1\leq i\leq n-1$ and $a_{i,i-2}>0,\; 3\leq i\leq n$,
the rest of the entries are zeros. Then all roots of the characteristic
polynomial, except possibly $0$, are simple and their arguments
are of the form $2\pi k/3,\; k\in\{0,1,2\}$. In fact this characteristic
polynomial has the form $\lambda^kP(\lambda^3)$, where $k\in\{0,1,2\}$
and $P$ is a real polynomial with all roots positive.
\end{lemma}
A proof of Lemma~\ref{lemma2} will be given in the next section.
The following proposition lists ramification of
forgetful maps over special points. We use the word ``ramification''
in the sense of maps between Riemann surfaces, not orbifolds.
\begin{prop}\label{prop-ramJ}
\noindent
{\bf 1.} Ramification of $\pi_m^K$ over special
points is the following:
\noindent
Over $J=0$ there are $\lfloor d_m^K/3\rfloor$ triple points,
and one additional point $x$ when $d$ is not divisible by $3$.
This additional point $x$ is the orbifold point of order $3$,
and $\pi_m^K$ has $x$ as a double point when $d\equiv 2\; (\mathrm{mod\,} 3),$
and a simple point when $d\equiv1\; (\mathrm{mod\,} 3)$.
\noindent
Over $J=1$ there are $\lfloor d_m^K/2\rfloor$ double points,
and one simple point when $d_m^K$ is odd. This simple point is
the orbifold point of order $2$.
\noindent
Over $J=\infty$ there are $d_m^{II}/3$ double points when $K=II$. The rest
$d_m^{II}/3$ points are simple. For $K=I$ all points over $\infty$
are simple.
\vspace{.1in}
\noindent
{\bf 2.} Ramification of $\sigma_m^j$ over special points
is the following:
Over each $a=1/2\pm i\sqrt{3}/2$,
there is one double point when $d_m^j\equiv 2\; (\mathrm{mod\,} 3).$
There is no other ramification over special points.
\end{prop}
{\em Proof.} For component $\mathbf{L}_m^I$ with even $m$ and $J=0$, we consider
polynomial solutions $Q$ of equation (\ref{lame1}) with
$g_2=0,\; g_3=1$, that is
$$(4x^3-1)Q''+6x^2Q'-m(m+1)xQ=\lambda Q.$$
The matrix of the linear operator in the left-hand side in the
basis of monomials has the form as in Lemma \ref{lemma2}.
Therefore the characteristic polynomial of this
matrix has the form $\lambda^kP(\lambda^3)$.
This has a root of multiplicity $2$ at $0$ when $d\equiv 2\; (\mathrm{mod\,} 2)$.
Other roots come in triples, each triple
lies on the same orbit under the $\mathbf{C}^*$ action (\ref{scaling}),
so we have $\lfloor d_m^K/3\rfloor$
triple points.
Similar considerations apply to other special points.
For component $\mathbf{L}_m^I$ with odd $m$ and $J=0$, we consider
solutions of (\ref{lame1})
of the form $\sqrt{4x^3-g_2x-g_3}\, Q(x)$, where $Q$ is a polynomial.
The equation for $Q$ becomes
$$(4x^3-1)Q''+18x^2Q'+\left(12-m(m+1)\right)x Q=\lambda Q,$$
and this leads to a matrix of the same form
described in Lemma~\ref{lemma2},
so the same argument as in the case of even $m$ applies.
For component $\mathbf{L}_m^I$ with even $m$ and $J=1$, we set $g_2=1,g_3=0$, and
obtain
$$(4x^3-x)Q''+(6x^2-1/2)Q'-m(m+1)xQ=\lambda Q$$
which leads to a matrix described in Lemma~\ref{lemma1}.
The characteristic polynomial is of
the form $\lambda^kP(\lambda^2),\; k\in\{0,1\}$
which has one simple root $\lambda=0$ when $k=1$ and other roots come
in pairs which are on the same orbit under the $\mathbf{C}^*$ action.
For component $\mathbf{L}_m^I$ with odd $m$ and $J=1$ we obtain the equation
$$(4x^3-x)Q''+(18x^2-3/2)Q'+\left(12-m(m+1)\right)x Q=\lambda Q$$
which again leads to a matrix described in Lemma~\ref{lemma1}.
The conclusion
is similar.
For component $\mathbf{L}_m^{II}$ we use the Legendre's
form of Lam\'e equation (\ref{legendre}).
When $m$ is odd, and $J=1$, we plug the solution of the form
$\sqrt{z}\, Q(z)$ and obtain
\begin{eqnarray*}
&&4z(z-1)(z-a)Q''+(10z^2+8z(1+a)+6a)Q'\\
&-&\left((m^2-m-2)z+1+a+B\right)Q=0.
\end{eqnarray*}
When $m$ is even, and $j=1$, we plug the solution of the form
$\sqrt{z(z-1)}\, Q(z)$ and obtain
\begin{eqnarray*}
&&4z(z-1)(z-a)Q''+(14z^2-(12a+8)z+6a)Q'\\
&-&\left((m^2+m-6)z+B+4a+1\right)Q=0.
\end{eqnarray*}
Both these equations lead to Jacobi matrices as in Lemma~\ref{lemma1}.
To study ramification at the punctures, we use again Legendre's form.
Take, for example, $a=0$.
The matrix of the operator in the left-hand side of
(\ref{lame2}) is triangular, with distinct eigenvalues.
So $\sigma_m$ is unramified at a point $x$ with $\sigma_m(x)=0$.
Now we have $\deg_0\psi=2$, so by (\ref{cd})
$$\deg_{\Psi_m(x)}(\pi_m)\cdot\deg_x\Psi_m=2,$$
thus each multiple is either $1$ or $2$. But we know the total number
of points in $\mathbf{L}_m^K$ over $J=\infty$ (punctures) and this implies
the statement of Proposition~\ref{prop-ramJ} for $J=\infty$.
That $\pi_m$ is an orbifold map follows from the identification
of the orbifold points in $\mathbf{L}_m^K$ in Proposition~\ref{special}.\hfill$\Box$
\vspace{.1in}
The difference between $\pi_m$ and $\sigma_m$
is that there is no $\mathbf{C}^*$ action in the second case.
Proposition~\ref{prop-ramJ} together with relation (\ref{cd}) and known
ramification of $\psi$ allows us to define the orbifold structure
on the compactified spaces, so that the $\psi$ and $\Psi_m$ extend
to
orbifold coverings of these compactifications.
For what follows we define compactifications of our orbifolds:
$$\mathbf{\overline{C}}_J=\mathbf{\overline{C}}(0(3),1(2),\infty(2)),\quad \mathbf{\overline{C}}_a=\mathbf{\overline{C}}.$$
Then $a\mapsto J=\psi(a)$ which is defined in (\ref{Ja}) is an orbifold
covering. Then we define ${\overline{\mathbf{L}}}_m^K$ by adding the punctures
$x, \;\pi_m^K(x)=\infty$, and defining $n(x)=1$ if $\deg_x(\pi_m^K)=2$
and $n(x)=2$ when $\deg_x(\pi_m^K)=1$.
Finally we define $\mathbf{\overline{C}}_a$ as the Riemann sphere with $n(a)=1$ for all $a$,
and define ${\overline{\mathbf{H}}}_m^j$ as $\mathbf{H}_m^j$ with filled punctures.
The orbifold structure on ${\overline{\mathbf{H}}}_m^j$ is trivial: $n(x)\equiv 1.$
With these definitions Theorem~\ref{theorem1} gives:
\begin{equation}\label{chiO}
\chi^O({\overline{\mathbf{L}}}_m^K)=\chi^O(\mathbf{L}_m^K)+d_m^K/2.
\end{equation}
\begin{prop}\label{total-ram} The following diagram is commutative:
\begin{equation}\label{cd1}
\begin{CD}
\overline{\mathbf{H}}_m^j @>{\displaystyle\Psi_m^j}>> \overline{\mathbf{L}}_m^K\\
@VV{\displaystyle\sigma_m} V @VV{\displaystyle\pi_m} V\\
\mathbf{\overline{C}}_a @>{\displaystyle\psi}>> \mathbf{\overline{C}}_J
\end{CD}
\end{equation}
Here all four spaces are orbifolds, with orbifold functions just
defined,
the horizontal arrows
are orbifold coverings, and vertical arrows are maps of orbifolds.
We have
$$\deg\Psi_m^0=6\quad\mbox{and}\quad\deg\Psi_m^j=2,\quad j\in\{1,2,3\}.$$
Furthermore,
\begin{equation}\label{chiH0}
\chi({\overline{\mathbf{H}}}_m^0)=\chi^O({\overline{\mathbf{H}}}_m^0)=6\chi^O({\overline{\mathbf{L}}}_m^I),
\end{equation}
\begin{equation}\label{chiHj}
\chi(\overline{\mathbf{H}_m^j})=\chi^O({\overline{\mathbf{H}}}_m^j)=
2\chi^O({\overline{\mathbf{L}}}_m^{II}).
\end{equation}
\end{prop}
{\em Proof.} Since in the diagram (\ref{cd}) the horizontal arrows
are orbifold coverings and vertical arrows are orbifold maps,
it remains to check the points over $J=\infty$ and
over $a\in\{0,1,\infty\}$.
That $\psi:\mathbf{\overline{C}}_a\to\mathbf{\overline{C}}_J$ is an orbifold covering is well known
and follows from the explicit formula (\ref{Ja}).
Let $x\in\mathbf{H}_m^j$, $\sigma_m(x)\in\{0,1,\infty\}$.
By Proposition~\ref{prop-ramJ}, $\deg_x(\sigma_m)=1$ and we know that
$\deg_{\sigma(x)}(\psi)=2.$ Therefore,
$$\deg_x(\Psi_m^j)\cdot\deg_{\Psi_m^j(x)}(\pi_m)=2,$$
thus $\deg_x\Psi_m^j$ is either $1$ or $2$, and the definition
of $n(\Psi_m^j(x))$ ensures that $\Psi_m^j$ is an orbifold
covering.
That the vertical arrows are orbifold maps follows from
Proposition~\ref{prop-ramJ}. Formulas (\ref{chiH0}), (\ref{chiHj})
follow from (\ref{rh}).\hfill$\Box$
Now we are ready to prove Theorem~\ref{theorem2}.
\begin{prop}\label{generaH}
The Riemann surfaces $\overline{\H}_m^j$ are connected,
and their images in $\mathbf{CP}^2$ are non-singular.
\end{prop}
{\em Proof.}
Using (\ref{rh}), (\ref{chiO}) and (\ref{chiH0}), we obtain
\begin{eqnarray*}
2-\chi({\overline{\mathbf{H}}}_m^0)&=&2-6\chi^O({\overline{\mathbf{L}}}_m^I)
=2-6\chi^O(\mathbf{L}_m^I)-3d_m^I\\
&=&2+(d_m^I)^2-3d_m^I
=(d_m^I-1)(d_m^I-2).
\end{eqnarray*}
Similarly, using (\ref{rh}), (\ref{chiO}) and (\ref{chiHj}),
we obtain
\begin{eqnarray*}
2-\chi({\overline{\mathbf{H}}}_m^j)&=&2-2\chi^O({\overline{\mathbf{L}}}_m^j)
=2-2\chi^O(\mathbf{L}_m^j)-d_m^{II}\\
&=&2+(d_m^{II})^2/9-d_m^{II}
=(d_m^{II}/3-1)(d_m^{II}/3-2),\quad j\in\{1,2,3\}.
\end{eqnarray*}
Therefore, in any case we have
\begin{equation}\label{aa}
2-\chi(\overline{\H}_m^j)=(\deg\overline{\H}_m^j-1)(\deg\overline{\H}_m^j-2).
\end{equation}
Suppose that for some $j$ and $m$,
$\overline{\H}_m^j$ has $N$ irreducible components of degrees $d_k$ genera $g_k$
and degrees $d_k$ for $1\leq k\leq N$.
Then
\begin{equation}\label{bb}
\deg\overline{\H}_m^j=\sum_{k=1}^Nd_k,\quad\chi(\overline{\H}_m^j)=\sum_{k=1}^N\chi_k,
\end{equation}
\begin{equation}\label{cc}
\chi_k=2-2g_k,\quad\mbox{and}\quad 2g_k\leq(d_k-1)(d_k-2);
\end{equation}
the last inequality follows from (\ref{gd}). Substituting the expressions
$\deg\overline{\H}_m^j$ and $\chi(\overline{\H}_m^j)$ from (\ref{bb}) to (\ref{aa})
and using (\ref{cc}) we obtain after simple manipulation
$$\left(\sum_{k=1}^Nd_k\right)^2\leq\sum_{k=1}^Nd_k^2;$$
since all $d_k\geq 1$, this is
possible only when $N=1$. Thus $\overline{\H}_m^j$ is irreducible.
Then from (\ref{aa}) we obtain its genus,
$$g(\overline{\H}_m^k)=(2-\chi(\overline{\H}_m^j)/2=(\deg\overline{\H}_m^j-1)(\deg\overline{\H}_m^j-2)/2,$$
so it is non-singular since it satisfies (\ref{gd})
with equality.\hfill $\Box$
This proposition
completes
the proof of Theorem~\ref{theorem2}.
\vspace{.1in}
{\em Proof of Corollary \ref{corollary1}.}
\vspace{.1in}
Consider the map $R:\mathbf{H}_m\to\{ F_m(\lambda,g_2,g_3)=0\}$
defined in (\ref{R}). We will show that it is transversal
to the orbits
of the $\mathbf{C}^*$ action (\ref{scaling})
at non-special points.
A trajectory of restriction of this action onto the $(g_2,g_3)$ plane
has the form $(g_2,g_3)=(t^2,ct^3),\, t\in\mathbf{C}^*$, so the tangent vector is
$(2t,3ct^2)$ which is parallel to $(2/g_3,3/g_2)=(2/R_3,3/R_2)$.
If the vectors $(2/R_3,3/R_2)$ and $(R_2^\prime,R_3^\prime)$ are
collinear, we must have
$$S:=R_2R_3(3R_2^\prime/R_2-2R_3^\prime/R_3)=0.$$
But an explicit computation shows that
$$S=-16a(a-1)/3,$$
which can be zero only at the special points.
Therefore the maps $\Psi_m^j$ are ramified only at the special points.
Diagram (\ref{cd1}) is clear from the definition. \hfill$\Box$
\vspace{.1in}
\noindent
{\em Proof of Corollary \ref{corollary2}.}
\vspace{.1in}
We compute the ramification of the forgetful map $\pi$ and then
make the correction for special points.
The usual (not orbifold) Euler characteristic of the compactification
of $\mathbf{L}_m^{I}$ is $\chi(\overline{\L}_m^I)=\chi(\mathbf{L}_m^I)+h$, where $h$ equals
the number of punctures. So from Theorem~\ref{theorem1} for $\mathbf{L}_m^I$
and Riemann--Hurwitz formula
the total ramification of $\pi$ is
\begin{eqnarray*}
&&2d-\chi(\overline{\L}_m^I)=2d-\chi(\mathbf{L}_m^I)-h=d-\chi(\mathbf{L}_m^I)\\
&=&d+d^2/6-(4\epsilon_0+3\epsilon_1)/6
=\lfloor (d^2-d+4)/6\rfloor+2\lfloor d/3\rfloor+\lfloor d/2\rfloor,
\end{eqnarray*}
where $d=\deg_\lambda F_m^{I}$. The first summand is the degree of
the Cohn polynomial, and the other two reflect the additional
ramification over the special points $0$ and $1$ (Proposition~\ref{prop-ramJ}).
Indeed, since the only singularities of
the surface $F_m^I(\lambda,g_2,g_3)=0$ lie over $g_2=0$ and $g_3=0$,
the zeros of the Cohn polynomial at all points $J\in\mathbf{C}\backslash\{0,1\}$
come from ramification points of $\pi$. For $J=0$, our curve
has the form $\lambda^kP(\lambda^3)$, where $P$ has only simple zeros,
so only $\lambda=0$ is a multiple zero when $k=2$. Other ramification
points of $\pi$ over $J=0$ do not contribute to zeros of Cohn's polynomial.
Similarly, $J=1$ is not a zero of Cohn's polynomial.
For $\mathbf{L}_m^{II}$ the total ramification is
$$d+d^2/18-(1-\epsilon_1)/2=(d/3)(d/3-1)/2+2d/3+\lfloor d/2\rfloor.$$
where $d=\deg_\lambda F_m^{II}.$
Again, the first summand corresponds to the degree of the Cohn polynomial
while the other two reflect additional ramification over the special points
(Proposition~\ref{prop-ramJ}).\hfill$\Box$
\vspace{.1in}
Next we briefly describe an alternative approach to our main results.
The following remarks are not necessary for understanding the
rest of the paper.
\vspace{.1in}
\noindent
{\bf Remarks on parametrization of $\mathbf{H}_m$ by a space of triangles}
\vspace{.1in}
In the beginning of the previous section we mentioned
that $\mathbf{H}_m$ represents the space of marked singular tori.
In view of the above interpretation of $\mathbf{H}_m$,
we can construct a natural lift of the isomorphism
$\Phi^*:\mathbf{T}_m^*\to\mathbf{L}_m$
to an isomorphism $\widehat{\Phi}^*:\widehat{\mathbf{T}}_m^*\rightarrow\mathbf{H}_m$ that makes
the following diagram commutative:
$$
\begin{CD}
\widehat{\mathbf{T}}_m^* @>{\displaystyle \Sigma} >> \mathbf{T}_m^*\\
@VV{\displaystyle\widehat{\Phi}^*} V @ VV{\displaystyle\Phi^*} V\\
\mathbf{H}_m @ >{\displaystyle\Psi} >> \mathbf{L}_m
\end{CD}
$$
Here $\widehat{\mathbf{T}}^*_m$ is a suitable space of flat singular triangles.
The point is that one could fully describe the
topology of the open Riemann surface $\widehat{\mathbf{T}}^*_m$
and so of $\mathbf{H}_m$ in a direct way,
and from that deduce the topology of $\mathbf{L}_m$
by taking the quotient by a natural $S_3$-action described below.
Recall that in a balanced flat singular triangle
$(D,\{a_i\},f)\in \mathbf{T}_m$,
the cyclic order of the corners $(a_1,a_2,a_3)$ on $\partial D$
matches the orientation induced by the developing map $f$,
and that we coordinatize
$\mathbf{T}_m$ using the functions $\phi_{i,j,k}$.
Denote by $-\mathbf{T}_m$ the space of balanced
flat singular triangles $(D,\{a_i\},f)$
for which the cyclic order $(a_1,a_3,a_2)$
matches the orientation induced by $f$, and coordinatize $-\mathbf{T}_m$
using the complex conjugates $\bar{\phi}_{i,j,k}$.
Moreover, let
$\widehat{\mathbf{T}}_m$ be the disjoint union of $\mathbf{T}_m$ and $-\mathbf{T}_m$.
The permutation group
$S_3$ acts
on $\widehat{\mathbf{T}}_m$ by relabeling the corners.
By identifying each marginal triangle
$(D,\{a_i\},f)\in\widehat{\mathbf{T}}_m$
with its conjugate $(D,\{a_i\},\bar{f})$, we obtain a space $\widehat{\mathbf{T}}_m^*$.
It is immediate that the charts defined above
for $\widehat{\mathbf{T}}_m$ induce
a complex structure on $\widehat{\mathbf{T}}_m^*$, and that the $S_3$-action
descends to $\widehat{\mathbf{T}}_m^*$.
The quotient space $\widehat{\mathbf{T}}_m^*/S_3$ is naturally identified
with $\mathbf{T}_m^*$ and such identification induces the map $\Sigma$.
Moreover, to a triangle $T$ in $\widehat{\mathbf{T}}_m$
we can associate the torus $\Phi(T)$ with the marking $t(T)=(t_1,t_2,t_3)$,
where $t_1,t_2,t_3$ are the midpoints of
$[a_1,a_2]$, $[a_1,a_3]$, $[a_2,a_3]$ respectively.
Thus we can define $\widehat{\Phi}^*(T)=(\Phi(T),t(T))$.
Since the diagram is manifestly commutative,
the map $\widehat{\Phi}^*:\widehat{\mathbf{T}}_m\to\mathbf{H}_m$
is an isomorphism of Riemann surfaces
(with trivial orbifold structure on both surfaces),
and $\Sigma$ is an orbifold cover.
Connected components of $\mathbf{H}_m$ can be
studied by analyzing $\widehat{\mathbf{T}}^*_m$, instead of exploring
the orbifold cover $\Psi_m$ as we did in the previous section.
Similarly to what we did with $\mathbf{T}_m$,
we can construct a nerve graph $\widehat{\Gamma}_m$
analogous to $\Gamma_m$. It consists of the disjoint union of
two isomorphic components: $\Gamma_m$ associated
to $\mathbf{T}_m$ and $-\Gamma_m$ associated to $-\mathbf{T}_m$.
Note that, if $v$ is a {\em lateral vertex} of $\Gamma_m$,
namely a vertex corresponding to a face of $\mathbf{A}_m$
adjacent to a boundary edge in $\partial\Delta_m$, then
the preimage of $v$ and the preimage of $-v$ belong
to the same connected component
of $\widehat{\mathbf{T}}^*_m$.
Since each component of $\Gamma_m$ has a lateral
vertex, it follows that every component
of $\Gamma_m$ exactly corresponds to
a component in $\widehat{\mathbf{T}}^*_m$,
which thus has four connected components.
Similarly to what was done in Section~\ref{ident}
one can also compute the genera of $\mathbf{H}_m^j$
using the parametrization $\widehat{\Phi}:\widehat{\mathbf{T}}_m^*\to\mathbf{H}_m$
and obtain an alternative proof of our results about
$\mathbf{H}_m$ in this section. Once the number of components, their
genera and non-singularity are established for $\mathbf{H}_m$, one can
obtain Theorem ~\ref{theorem1} via (\ref{cd1}).
\section{Proof of Lemma~\ref{lemma2}}\label{tarasov}
In this section, we give a proof of Lemma~\ref{lemma2}
which is due to V. Tarasov.
It is inspired by an argument from \cite[Proposition 3]{ST}
which was brought
to our attention by Eduardo Chavez Heredia from the University of Bristol.
The proof is a generalization of the classical arguments, going back
to Ch. Sturm, which are used in the proof of Lemma~\ref{lemma1}.
Let $D=\{ d_{i,j}\}$ be an $n\times n$ matrix
with entries
\begin{eqnarray*}
d_{i,i}=s,&&i=1,\dots, n,\\
d_{i,i+1}=a_i,&& i=1,\dots, n-1,\\
d_{i,i-2} = b_i,&& i=3,\dots, n.
\end{eqnarray*}
and consider the principal minors
$$D_k = \det(d_{ij})_{i,j=1,\dots, k}$$
They satisfy the recurrences
$$D_{k+3} = sD_{k+2} + c_k D_k,$$
where
$$c_k=a_{k+1}a_{k+2}b_{k+3}>0,$$
with the initial conditions
$$D_0=1,\quad D_1=s ,\quad D_2=s^2.$$
Then
$$D_{3j}(s)=P_j(s^3),\quad D_{3j+1}(s)=sQ_j(s^3),\quad D_{3j+2}(s)=
s^2R_j(s^3)$$
where the polynomials $P_j,Q_j,R_j$ satisfy
the recurrences
\begin{equation}\label{11t}
P_{j+1}=sR_j+A_jP_j,
\end{equation}
\begin{equation}\label{12}
Q_{j+1}=P_{j+1}+B_j Q_j,
\end{equation}
\begin{equation}\label{13}
R_{j+1}= Q_{j+1}+C_j R_j,
\end{equation}
where
$$A_j = c_{3j},\quad B_j=c_{3j+1},\quad C_j=c_{3j+2}$$
and the initial conditions
\begin{equation}\label{init}
P_0 = Q_0 = R_0 = 1.
\end{equation}
Lemma~9.2 follows from
\begin{prop}\label{prop-tarasov}
Let us define polynomials $P_j(s), Q_j(s), R_j(s)$ by recurrences
(\ref{11t}), (\ref{12}), (\ref{13}) where all
$A_j,B_j,C_j$ are strictly positive, and initial conditions (\ref{init}).
Then all polynomials $P_j, Q_j$ and $R_j$, $j=1,\dots$ are monic,
have degree $j$ and positive coefficients, and
all their roots are negative and simple. Moreover, if
$p_{j1}>\ldots>p_{jj}$, $q_{j1}>\ldots>q_{jj}$, $r_{j1}>\ldots>r_{jj}$,
are respective roots of the polynomials $P_j,Q_j,R_j$, then
$\;p_{jk}>q_{jk}>r_{jk}$ for all $\;k=1,\dots,j$, and $\;r_{jk}>p_{j,k+1}$
for all $\;k=1,\dots,j-1$.
\end{prop}
\noindent
{\em Proof.} We prove this by induction.
It is evident that our polynomials are monic and have positive coefficients.
Therefore their real roots are negative. To find the number of real
roots and to show that they are interlacent we look at the signs
of our polynomials at the roots of other polynomials using
(\ref{11t})--(\ref{13}).
\smallskip
The base of induction is given by $j=1$ and is clear.
The induction procedure is as follows.
\smallskip
By (\ref{11t}) and the induction assumption,
$$
P_{j+1}(0)\,P_{j+1}(p_{j1})\,=\,p_{j1} A_j P_j(0)\,R_j(p_{j1})<0\,,
$$
$$
P_{j+1}(r_{jk})\,P_{j+1}(p_{j,k+1})\,=\,
p_{j,k+1} A_j P_j(r_{jk})\,R_j(p_{j,k+1})<0\,,
$$
for all $k=1,\dots,j-1$, and
$$
(-1)^j P_{j+1}(r_{jj})=(-1)^j A_j P_j(r_{jj})>0\,.
$$
This implies that $P_{j+1}$ has roots $p_{j+1,k}$, $\,k=1,\dots, j+1$,
located as follows:
\begin{equation}
\label{pr}
0>p_{j+1,1}>p_{j1}\,,\quad
r_{j,k-1}>p_{j+1,k}>p_{jk}\,,\;\;k=2,\dots,j\,,\quad r_{jj}>p_{j+1,j+1}\,.
\end{equation}
Thus all roots of $P_{j+1}$ are negative and simple.
\smallskip
By the induction assumption and (\ref{pr}),
\begin{equation}
\label{pq}
p_{j+1,1}>q_{j1}\,,\quad
q_{j,k-1}>p_{j+1,k}>q_{jk}\,,\;\;k=2,\dots,j\,,\quad q_{jj}>p_{j+1,j+1}\,.
\end{equation}
Then by (\ref{12}) and (\ref{pq}),
$$
Q_{j+1}(p_{j+1,k})\,Q_{j+1}(q_{jk})\,=\,B_j Q_j(p_{j+1,k})\,P_{j+1}(q_{jk})<0\,,
$$
for all $k=1,\dots,j$, and
$$
(-1)^j Q_{j+1}(p_{j+1,j+1})=(-1)^j B_j Q_j(p_{j+1,j+1})>0\,.
$$
This implies that $Q_{j+1}$ has roots $q_{j+1,k}$, $\,k=1,\dots, j+1$,
located as follows:
\begin{equation}
\label{qp}
p_{j+1,k}>q_{j+1,k}>q_{jk}\,,\;\;k=1,\dots,j\,,\quad
p_{j+1,j+1}>q_{j+1,j+1}\,.
\end{equation}
Thus all roots of $Q_{j+1}$ are negative and simple.
\smallskip
By the induction assumption, (\ref{qp}), and (\ref{pr}),
\begin{equation}
\label{qr}
q_{j+1,1}>r_{j1}\,,\quad
r_{j,k-1}>q_{j+1,k}>r_{jk}\,,\;\;k=2,\dots,j\,,\quad r_{jj}>q_{j+1,j+1}\,.
\end{equation}
Then by (\ref{13}) and (\ref{qr}),
$$
R_{j+1}(q_{j+1,k})\,R_{j+1}(r_{jk})\,=\,C_j R_j(q_{j+1,k})\,Q_{j+1}(r_{jk})<0\,,
$$
for all $k=1,\dots,j$, and
$$
(-1)^j R_{j+1}(q_{j+1,j+1})=(-1)^j C_j R_j(q_{j+1,j+1})>0\,.
$$
This implies that $R_{j+1}$ has roots $r_{j+1,k}$, $\,k=1,\dots, j+1$,
located as follows:
\begin{equation}
\label{rq}
q_{j+1,k}>r_{j+1,k}>r_{jk}\,,\;\;k=1,\dots,j\,,\quad
q_{j+1,j+1}>r_{j+1,j+1}\,.
\end{equation}
Thus all roots of $R_{j+1}$ are negative and simple.
\smallskip
The inequalities $\;p_{j+1,k}>q_{j+1,k}>r_{j+1,k}$ for all $\;k=1,\dots,j+1$,
follow from (\ref{qp}) and (\ref{rq}), and the inequalities
$\;r_{j+1,k}>p_{j+1,k+1}$ for all $\;k=1,\dots,j$, follow from
(\ref{rq}) and (\ref{pr}). This completes the proof.\hfill$\Box$
\section{Projective monodromy of the Lam\'e equation}\label{mono}
In this section we prove Theorem~\ref{LWm}.
Lam\'e equation with integer $m$ has trivial local monodromy about the origin.
Since the fundamental group of the torus is $\mathbf{Z}^2$, the projective
monodromy
is represented by a pair of commuting elements $PSL(2,\mathbf{C})$.
So we investigate the set of pairs of commuting elements of $PSL(2,\mathbf{C})$
modulo conjugation. Every such pair $(A,B)$ can be
conjugated to one of the following forms:
\begin{equation}
\label{diag}
(z\mapsto \mu_1z,\;z\mapsto\mu_2z),\quad(\mu_1,\mu_2)\in (\mathbf{C}^*)^2,
\end{equation}
\begin{equation}\label{transl}
(z\mapsto z+a_1,\; z\mapsto z+a_2),\quad(a_1,a_2)\in\mathbf{C}^2,
\end{equation}
or
\begin{equation}\label{klein}
(z\mapsto -z,\; z\mapsto 1/z).
\end{equation}
It is proved in \cite[Theorem 2.2]{L1} that the third possibility (\ref{klein})
cannot happen for
the projective monodromy of Lam\'e equations with integer $m$.
Notice that $PSL(2,\mathbf{C})$ representations
(\ref{diag}) and (\ref{transl}) can be lifted to $SL(2,\mathbf{C})$. The
pair $(A,B)=(\mathrm{id},\mathrm{id})$ is also excluded.
Conjugacy classes of pairs of the form (\ref{diag}) are parametrized by
$(\mathbf{C}^*)^2$. Two pairs of the form (\ref{transl}) are conjugate iff
$(a_1:a_2)=(a_1^\prime:a_2^\prime)$, so they are parametrized
by the projective line $\mathbf{CP}^1$ and one point $(0,0)$.
Suppose that a sequence of pairs $(A,B)$ of type (\ref{diag}) converges
in $PSL(2,\mathbf{C})$
to a pair of type (\ref{transl}).
To figure out how $(a_1,a_2)$ in (\ref{transl}) are related to $(\mu_1,\mu_2)$
we consider
commuting pairs of linear-fractional transformations $(\phi_1,\phi_2)$,
$$\phi_j(z)=\frac{(1+f_j)z+(a_j+p_j)}{q_jz+(1+g_j)},\quad j\in\{1,2\},$$
where $f_j,g_j,p_j,q_j$ are small numbers,
and $a_j$ are constants not simultaneously equal to $0$.
The condition that matrices representing $\phi_j$ have determinant $1$ gives
\begin{equation}\label{11}
f_j+g_j\equiv a_jq_j,
\end{equation}
where $\equiv$ means that we neglected the terms of
order $2$ and higher.
The condition that $\phi_1$ and $\phi_2$ commute implies by comparing
the diagonal elements of the product matrices
\begin{equation}\label{comm}
a_1q_2\equiv a_2q_1.
\end{equation}
Now it follows from (\ref{11}) and (\ref{comm}) that
\begin{equation}
\label{cond1}
(f_1+g_1):(f_2+g_2)\to (a_1^2:a_2^2).
\end{equation}
Similar equation can be obtained when the eigenvalue of
one or both limit matrices is $-1$.
In other words, if $A$ and $B$ are $SL(2,\mathbf{C})$ matrices
representing $\phi_1$ and $\phi_2$ tending to parabolic $\phi_1^*$
and $\phi_2^*$, then we have
\begin{equation}\label{cond2}
\lim\frac{\mathrm{tr}\, ^2 A-4}{\mathrm{tr}\, ^2 B-4}=(a_1^2:a_2^2),
\end{equation}
where $(a_1:a_2)$ is the ``ratio invariant'' of the pair of commuting
parabolic transformations.
So we obtain that the space of projective monodromy representations
for Lam\'e equations is the blow up of $(\mathbf{C}^*)^2$ at the point $(1,1)$.
Monodromies of unitarizable Lam\'e equations form the real torus
$$\{(\mu_1,\mu_2)\in(\mathbf{C}^*)^2:|\mu_1|=|\mu_2|=1\}\backslash(1,1),$$
and the boundary of this real torus in the blow up is the real projective line.
Since $A$ and $B$ are elliptic,
the left hand side of (\ref{cond2}) is positive, so the ratio
$(a_1:a_2)$ is real.
Thus we obtain
\begin{prop}\label{real}
Abelian integrals
arising as limits of developing maps of spherical metrics
have real period ratios.
\end{prop}
So we have
$$\partial\,\mathrm{Sph}_{1,1}(2m+1)\subset\mathrm{LW}_m.$$
Let us prove the opposite inclusion.
We recall that $\mathrm{Lame}_m$ is biholomorphic to $\mathbf{CP}(1,2,3)$,
and each point of $\mathrm{Lame}_m$ corresponds to a Schwarz equation (\ref{schwa})
whose solution is a developing map of a translation structure
on a torus with one conic singularity with angle $2\pi(2m+1)$.
Let us fix $L\in\mathrm{LW}_m$. We want to prove that there exists
a sequence $L_n\in\mathrm{Lame}_m$, $L_n\to L$ such that $L_n$
have unitarizable monodromy.
Since $L\in\mathrm{LW}_m$, is obtained as $\Phi(T)$ from some BFT,
we can normalize $T$ so that
the images of corners under the developing map are
$$(-\delta,k\delta,\delta),$$
where $k\in(-1,1)$ depends on the similarity class of $T$ and thus
is fixed, while $\delta>0$ is arbitrary.
We glue a congruent copy $-T$ to $T$ by the side
with endpoints $\{-\delta,\delta\}$. Notice that the gluing map $z\mapsto-z$
is the orientation-reversing isometry of this side, with respect to
both flat and spherical metric. The result of this gluing is a
``parallelogram'' $Q$, which can be represented by
$(\Delta,\{ a_j\},f)$, where $\Delta$ in the unit disk, $a_1,a_2,a_3,a_4$
are boundary points (corners), and $f$ is a meromorphic function,
a developing map of $Q$. This developing map is a local homeomorphism
at all points of $\overline\Delta\backslash\{ a_j\}$ while at the corners
it behaves like an integer power. The torus $\Phi(T)$
is obtained by gluing the
opposite sides of $Q$, so we have
\begin{equation}\label{71}
f(p(z))=t_i\circ f(z),\quad z\in\partial\Delta, \quad i=1,\ldots,4,
\end{equation}
where $p(z)$ is an involution of $\partial\Delta$ defining the gluing
of opposite sides, $t_j$ are translations, and $i$ depends
on the side to which $z$ belongs. Subscript $i$ has to
be interpreted similarly in all following formulas (\ref{72}), (\ref{75}), (\ref{73}).
Each of these translations $t_i$
is a product of two Euclidean rotations by $\pi$ about the ``middles'' of
the sides (see Proposition~\ref{prop2.5}).
Notice that $T$ can be also considered as a spherical triangle, since its
sides belong to the real line, which is both spherical and Euclidean geodesic.
So we can produce from $T$ a spherical torus with one conic singularity
with the same angle as in $\Phi(T)$, by identifying the opposite sides
of $Q$ by spherical isometries $s_i,\; 1\leq i\leq 4$. Similarly to $t_i$,
each $s_i$ is a product of two spherical rotations by $\pi$ about
the ``spherical middles'' of the sides.
Notice the following: {\em let $a,b$ be two points in $[-\delta,\delta]$.
There exists a unique Euclidean rotation $e$ by angle $\pi$,
and a unique spherical
rotation $s$ by angle $\pi$ which interchange $a$ and $b$.
Moreover, when $\delta$ is
small, then $e$ and $s$ are close to each other
in the sense that $s=e\circ\phi$, where
$\phi$ is a linear fractional homeomorphism of $R$ which
is $\epsilon$-bi-Lipschitz with respect to the spherical metric, where
$\epsilon=\epsilon(\delta)\to 1$ as $\delta\to 0$.}
We call such homeomorphisms $\phi$
{\em close to the identity.} Using this statement,
we obtain that
\begin{equation}\label{72}
s_i=t_i\circ\phi_i,
\end{equation}
where $\phi_j$ are close to the identity. Now we define homeomorphisms
$\psi_i$ of the sides of $\Delta$ and the involution
$q$ of the unit circle $\partial\Delta$ by
\begin{equation}\label{75}
f\circ\psi_i=\phi_i^{-1}\circ f,\quad q:=\psi_i^{-1}\circ p,
\end{equation}
and the function $g:\partial\Delta\to\mathbf{R}$,
\begin{equation}\label{73}
g=f\circ\psi_i.
\end{equation}
It is easy to see that $\psi_i$ are close to the
identity: the only points where $f$ is not bi-Lipschitz are
the corners, where $f$ behaves like $z^n$, and conjugating
a diffeomorphism which is $C^1$-close to the identity
by $z^m$ result in a diffeomorphism which is $C^1$ close to the identity.
Now we use the theorem from \cite[Thms. 5.3, 5.17]{Tu} which says that
every $\epsilon$-bi-Lipschitz homeomorphism of the unit circle
with $\epsilon$ sufficiently close to $1$ has an $\epsilon_1$-bi-Lipschitz
extension to a homeomorphism of the unit disk. So we extend $\psi$ and
define $g$ in $\Delta$ by (\ref{73}). This extended $g$
satisfies
$$g(q(z))=s_i\circ g(z),\quad z\in\partial\Delta,$$
which is analogous to (\ref{71}), and follows from (\ref{71}), (\ref{72}),
(\ref{75}) and (\ref{73}).
So $g$ can be considered
as a developing
map of a spherical metric on a torus obtained by identifying the opposite
sides
by the involution $q$. The map $g$ is not conformal with respect to
the standard conformal structure on $\Delta$ but conformal in a new
structure defined by $\psi$. This new structure is close to the standard
one in the sense of the theory of quasiconformal mappings \cite{Ahlfors}.
Indeed, a bi-Lipschitz map with constant close to $1$ is quasiconformal
with dilatation close to $1$. Let $L_\delta$ be the spherical torus
obtained by this construction, and let $F$ and $G_\delta$ be the
lifts the developing maps
of the tori $\Phi(T)$ and $L_\delta$ to the universal cover.
Then $F$ is a meromorphic function, and $G_\delta$
a quasiregular one, conformal
with respect to the non-standard conformal structure in $\mathbf{C}$.
We normalize $F$ and $G_\delta$ so that the preimages
of the conic singularities
form lattices $\Lambda_F$ and $\Lambda_{G_\delta}$,
with primitive periods $1$ and $\omega_F,\omega_{G_\delta}$.
To compare $F$ and $G_\delta$ we find a homeomorphism $\eta_\delta:\mathbf{C}\to\mathbf{C}$
such that
$F_\delta:=G_\delta\circ\eta_\delta$ is a meromorphic function,
and $\eta_\delta$ is normalized by $\eta_\delta(0)=0,\;
\eta_\delta(1)=1$. This homeomorphism is obtained as a normalized solution
of the Beltrami equation \cite[Chap. V]{Ahlfors},
and since the quasiconformal dilatation of $\eta_\delta$ is small,
$\eta_\delta\to\mathrm{id},$ as $\delta\to 0$.
We conclude that the lattices of poles $\Lambda_{F_\delta}$
and $\Lambda_F$ are close to each other
uniformly on compacts,
and $F_\delta\to F$ as $\delta\to 0$ uniformly
(with respect to the spherical
metric) on compact subsets of the plane. This means that when $\delta\to 0$,
the lattice invariants $g_j(\Lambda_{F_\delta})\to g_j(\Lambda_F),\;
j=2,3,$ and the Schwarzian of $F_\delta$ converges
to the Schwarzian of $F$. Since the Schwarzians
of $F$ and $F_\delta$ are both of the form
$$m(m+1)\wp(z,\Lambda)+\lambda,$$
where $\lambda=\lambda(F)$ or $\lambda=\lambda(F_\delta)$,
we conclude that $\lambda(F_\delta)\to\lambda(F)$ as $\delta\to 0$,
and this proves
that $L_\delta\to L$ in $\mathrm{Lame}_m$. This completes the proof of
Theorem~\ref{LWm}.
\section{Lin--Wang curves}\label{linwang}
Proposition~\ref{real} gives
the following
characterization of Lin--Wang curves: {\em $J\in\pi_m(\mathrm{LW}_m)$ if
some Lam\'e equation with invariant $J$ has translational monodromy
which belongs to a straight line.}
In particular, there is a flat singular
torus corresponding to $(J,\lambda)$, and this torus is of the
form $\Phi^*(T)$ where $T$ is a BFT with integer angles.
\vspace{.1in}
{\em Proof of Theorem~\ref{theorem4}.}
\vspace{.1in}
The set $\mathrm{LW}_m$ is the image
of the set of triangles with angles integer multiples of $\pi$ under
the map $\Phi_m^*$. Triangles with angles integer multiples of $\pi$
form straight intervals in the local coordinates $\phi_{i,j,k}$ introduced
in Section~\ref{complexanalytic}, and the map $\Phi^*_m$ is biholomorphic.
We conclude that the set $\mathrm{LW}_m$ consists of analytic (non-singular)
curves. There are three such curves corresponding to each integer point
in $\mathbf{A}_m$. So all together we have $m(m+1)/2$ Lin--Wang curves.
\hfill$\Box$
\vspace{.1in}
Proposition~\ref{real} shows how to find
equations of Lin--Wang curves. They are curves in $\mathbf{L}_m$, the
moduli space of Lam\'e functions,
where the ratio of periods of the integral (\ref{abel}) is real.
According to \cite{EMP}, to each triple of integers satisfying
the triangle inequalities corresponds a component of the space
$\mathrm{Sph}_{1,1}(2m+1)$. This component is parametrized by an open triangle,
and the sides of this triangle correspond to three interior edges of $\mathbf{T}_m$
which are mapped by $\phi$ in (\ref{phi})
to integer points of $\mathbf{A}_m$ by the map $\phi$
in (\ref{phi}). These three edges parametrize
Lin--Wang curves by the map $\pi_m\circ\Phi_m^*$.
Thus each integer point in $\mathbf{A}_m$ corresponds to a component
of the moduli space of $\mathrm{Sph}_{1,1}(2m+1)$ of spherical tori. The boundary of
this component consists of one or three Lin--Wang curves: when the integer
point is the center of $\mathbf{A}_m$, which happens when $m\equiv1\, (\mathrm{mod\,} 3)$,
there is one curve, otherwise there are three of them. If there is one
curve, it belongs to $\mathbf{L}_m^{II}$, as it happens for $m=1$ and $m=4$.
For integer points other than the center of $\mathbf{A}_m$ we have three curves which can be all
on component $\mathbf{L}_m^{II}$, or one of them can be on $\mathbf{L}_m^I$ and two on
$\mathbf{L}_m^{II}$.
Figs.~6, 7 and similar figures for other $m$ permits to determine this
for every integer point on $\mathbf{A}_m$.
We give explicit formulas for Lin--Wang curves for $1\leq m\leq 3$.
We recall that the developing
map is given by (\ref{Ab}). Since working with
elliptic functions
is easier than with Abelian integrals (this was the primary reason
for introducing elliptic functions), we pass to the universal covering.
In what follows, $g$ is the integrand in (\ref{Ab}). It is an even
elliptic function
with a single zero of multiplicity $2m$ at the origin, and double poles
with vanishing residues. The general form of such function is
$$g(z)=c_0+\sum_{j=1}^mc_j\wp(z-a_j).$$
By Abel's theorem, $2(a_1+\ldots+a_m)\equiv 0$, and we want to choose
$c_j$ so that $g$ and its first $2m-1$ derivatives vanish at the origin.
So
$$c_0=-\sum_{j=1}^mc_j\wp(a_j),$$
and for the rest of $c_j$ we have a system of equations
$$\sum_{j=1}^mc_j\wp^{(k)}(-a_j)=0,\quad 1\leq k\leq 2m-1,$$
which has a non-trivial solution if the matrix of this system has rank
at most $m-1$.
Once $g$ is found, we are interested in the ratio of periods
of the integral (\ref{Ab}). Lin--Wang curves make the locus of points where
this ratio is real. Below we give the results of computation for $1\leq m\leq3$.
We use the standard notation of the theory of elliptic functions \cite{A}:
$\wp(\omega_j)=e_j,\; 1\leq j\leq 3$, so $\omega_j$ are {\em half-periods},
$\zeta$ is the Weierstrass zeta function, $\zeta'=-\wp$, satisfying
$$\zeta(z+\omega_j)=\zeta(z)+\eta_j.$$
\vspace{.1in}
\noindent
{\em Case $m=1$}. $g(z)=\wp(z-\omega_j)-e_j,\; j\in\{1,2,3\}$. In this case,
$\mathbf{L}_1=\mathbf{L}_1^{II}$.
The equation of Lin--Wang curves is
$$\mathrm{Im}\, \frac{\eta_1+\omega_1e_j}{\eta_2+\omega_2e_j}=0,\quad j\in\{1,2,3\}.$$
This defines three curves in the fundamental region of the modular group
in the $\tau$-half-plane, Fig.~12,
which correspond to one curve in the $J$-plane
(Fig.~13).
\vspace{.1in}
\noindent
{\em Case $m=2$}. For $\mathbf{L}_2^I$, we obtain
$$g(z)=\wp(z+a)+\wp(z-a)-2\wp(a),\quad\mbox{where}\quad \wp(a)=\sqrt{g_2/12}.$$
The equation of the Lin--Wang curve is
$$\mathrm{Im}\, \frac{\eta_1+\omega_1\sqrt{g_2/12}}{\eta_2+\omega_2\sqrt{g_2/12}}=0.$$
For $\mathbf{L}_2^{II}$, we have
$$g(z)=\wp^{\prime\prime}(\omega_j)\left(\wp(z+\omega_k)-e_k\right)-
\wp^{\prime\prime}(\omega_k)\left(\wp(z+\omega_j)-e_j\right),$$
and the equation of Lin--Wang curve is
$$\mathrm{Im}\, \frac{\omega_1(e_j\wp''(\omega_k)-e_k\wp''(\omega_j))+
\eta_1(\wp''(\omega_k)-\wp''(\omega_j))}{\omega_2(e_j\wp''(\omega_k)-e_k\wp''(\omega_j))+\eta_2(\wp''(\omega_k)-\wp''(\omega_j))}=0.$$
These curves in the $\tau$-half-plane are shown in Fig.~14 and their images
in the $J$-plane are in Fig.~15. These are the three curves bounding a single
triangle
which is the moduli space $\mathrm{Sph}_{1,1}(5)$.
One of these curves, (which has a loop
in Fig. 12) belongs to $\mathbf{L}_2^I$, other two belong to $\mathbf{L}_2^{II}$.
Shading in Fig.~15 is the {\em hypothetical} projection
of a component of $\mathrm{Sph}_{1,1}(5)$ by the forgetful map.
We do not know whether the restriction of the
forgetful map on $\mathrm{Sph}_{1,1}(2m+1)$
is open. So we don't know that the boundary of this projection
is contained in Lin--Wang curves.
\vspace{.1in}
\noindent
{\em Case $m=3$}. For $\mathbf{L}_3^I$ we obtain:
$$g(z)=c_0+c_1\wp(z+\omega_1)+c_2\wp(z+\omega_2)+c_3\wp(z+\omega_3),$$
where
$$c_0=-\sum_{j=1}^3c_je_j,$$
$$c_k=(6e_{k+2}^2-g_2/2)(6e_{k+1}^2-g_2/2)(e_{k+2}-e_{k+1}),\quad k\in\{1,2,3\},$$
where the subscripts are understood as residues $\mathrm{mod\,} 3$, but we use $3$
instead of $0$ to prevent the confusion with previous formula.
Setting $B=c_1+c_2+c_3$, the equation of the Lin-Wang curve is
$$\mathrm{Im}\, \frac{\omega_1c_0-\eta_1B}{\omega_2c_0-\eta_2B}=0.$$
This curve in the $\tau$-plane is shown
in Fig.~16.
For component $\mathbf{L}_3^{II}$ we introduce the notation for $k\in\{1,2,3\}$:
$$P_k^\pm=-e_k/5\pm\sqrt{3(5g_2/4-3e_k^2)},$$
$$c_{1,k}^\pm=-2\frac{6(P_k^\pm)^2-g_2/2}{6e_k^2-g_2/2},$$
$$c_{0,k}^\pm=-c_{1,k}^\pm e_k-2P_k^\pm.$$
Then
$$g(z)=c_{0,k}^\pm+c_{1,k}^\pm\wp(z+\omega_k)+\wp(z+a_k^\pm)+\wp(z-a_k^\pm),$$
where $\pm a_k^\pm$ are solutions of the equation $\wp(z)=P_k^\pm$.
Then we have six Lin--Wang curves
$$\mathrm{Im}\, \frac{c_{0,k}^\pm\omega_1-(c_{1,k}^\pm+2)\eta_1}{c_{0,k}^\pm\omega_2-
(c_{1,k}^\pm+2)\eta_2}=0.$$
These curves in the $\tau$-half-plane are shown in Fig. 17, and their image
in the $J$-plane in Fig. 18, where the curve from $\mathbf{L}_3^I$
is also included (it is the one which looks like a vertical line in
the middle). Fig. 18 shows in the right-hand side
the detail which looks like
a tripod in the left-hand side. Figure 17 contains 16 curves which give $5$
images in Fig. 18. Three of these $5$ curves in Fig. 18
constitute the full boundary
of one triangle of the moduli space for spherical tori, and the remaining
three curves in Fig. 18, including that one curve which comes from Fig. 16
constitute the boundary of the second triangle in the moduli space for
spherical tori.
\newpage
\noindent
\section*{Appendix. List of formulas}
\vspace{.2in}
Polynomials $F_m^I$:
\vspace{.2in}
\begin{tabular}{r|l}
2&$ \lambda^2-3g_2$\\ \\
3&$ \lambda$\\ \\
4&$ \lambda^3-52g_2\lambda+560g_3$\\ \\
5&$ \lambda^2-27g_2$\\ \\
6&$ \lambda^4-294g_2\lambda^2+7776g_3\lambda+3465 g_2^2$\\ \\
7&$ \lambda^3-196g_2\lambda+2288g_3$\\ \\
8&$ \lambda^5-1044g_2\lambda^3+48816g_3\lambda^2+112320g_2^2\lambda
-4665600g_2g_3$\\ \\
9&$ \lambda^4-774g_2\lambda^2+21600g_3\lambda+41769 g_2^2.$
\end{tabular}
\vspace{.2in}
Polynomials $F_m^{II}$:
\vspace{.2in}
\begin{tabular}{r|l}
1&$ 4\lambda^3-g_2\lambda-g_3$\\ \\
2&$ 4\lambda^2-9g_2\lambda+27 g_3$\\ \\
3&$ 16\lambda^6-504g_2\lambda^4+2376g_3\lambda^3+4185g_2^2\lambda^2-
36450g_2g_3\lambda-3375g_2^3+91125g_3^2$\\ \\
4&$ 16\lambda^6-1016g_2\lambda^4+8200g_3\lambda^3+10297g_2^2\lambda^2
-41650g_2g_3\lambda$\\
&$-27783g_2^3-42875g_3^2.$
\end{tabular}
\newpage
Polynomials $H_m^0$:
\vspace{.2in}
\begin{tabular}{r|l}
2&$ B^2+4(a+1)B+12a$\\ \\
3&$ B+4(a+1)$\\ \\
4&$ B^3+20(a+1)B^2+(64a^2+336a+64)B+640(a^2+a)$\\ \\
5&$ B^2+20(a+1)B+64(a^2+1)$\\ \\
6&$ B^4+56(a+1)B^3+(784a^2+2744a+784)B^2$\\
&$+(2304a^3+29472a^2+29472a+2304)B$\\
&$+48384a^3+152208a^2+48384a$
\end{tabular}
\vspace{.2in}
Polynomials $H_m^1$:
\vspace{.2in}
\begin{tabular}{r|l}
1&$ B+a+1$\\ \\
2&$ B+4a+1$\\ \\
3&$ B^2+10(a+1)B+9a^2+78a+9$\\ \\
4&$ B^2+(20a+10)B+64a^2+136a+9$\\ \\
5&$ B^3+35(a+1)B^2+(259a^2+1046a+259)B$\\
&$+225a^3+5235(a^2+a)+225$\\ \\
6&$ B^3+(56a+35)B^2+(784a^2+1568a+259)B$\\
&$+2304a^3+13008a^2+7464a+225$
\end{tabular}
\vspace{.2in}
Degrees of forgetful maps:
\begin{eqnarray}
d_m^I&:=&\left\{\begin{array}{ll}m/2+1,& m\equiv0\;
(\mathrm{mod\,} 2)\\
(m-1)/2,& m\equiv 1\; (\mathrm{mod\,} 2),\end{array}\right.\\
d_m^{II}&:=&3\lceil m/2\rceil.
\end{eqnarray}
\newpage
Euler characteristics:
$$\chi(\mathbf{L}_m^I)=\left\{\begin{array}{ll}\displaystyle
-\frac{(m+2)^2}{24}+\frac{4\epsilon_0+3\epsilon_1}{6},&m\equiv 0\;(\mathrm{mod\,} 2),\\ \\
\displaystyle
-\frac{(m-1)^2}{24}+\frac{4\epsilon_0+3\epsilon_1}{6},&m\equiv 1\;(\mathrm{mod\,} 2).
\end{array}\right.$$
$$\chi(\mathbf{L}_m^{II})=\left\{\begin{array}{ll}\displaystyle
-\frac{m^2}{8}+\frac{1-\epsilon_1}{2},& m\equiv0\;(\mathrm{mod\,} 2),\\ \\
\displaystyle
-\frac{(m+1)^2}{8}+\frac{1-\epsilon_1}{2},& m\equiv 1\;(\mathrm{mod\,} 2).
\end{array}\right.,$$
where
\begin{eqnarray*}
\epsilon_0&=&\left\{\begin{array}{ll}0,&\mbox{if}\; m\equiv1\;(\mathrm{mod\,} 3),\\
1&\mbox{otherwise,}\end{array}\right.\\ \\
\epsilon_1&=&\left\{\begin{array}{ll}0,&\mbox{if}\; m\in\{1,2\}\;(\mathrm{mod\,} 4),\\
1&\mbox{otherwise.}\end{array}\right.
\end{eqnarray*}
Numbers of punctures: $h_m^I=d^I_m,\; h_m^{II}=2d^{II}_m/3.$
Genera in terms of $d_m^K$:
\begin{equation}\label{gI}
g(\mathbf{L}^I_m)=1+\frac{(d_m^I)^2}{12}-\frac{d_m^I}{2}-
\frac{4\epsilon_0+3\epsilon_1}{12},\end{equation}
\begin{equation}\label{gII}
g(\mathbf{L}^{II}_m)=1+\frac{(d_m^{II})^2}{36}-\frac{d_m^{II}}{3}-
\frac{1-\epsilon_1}{4}.
\end{equation}
Genera in terms of $m$:
\begin{equation}\label{gIe}
g(\mathbf{L}_m^I)=\frac{m^2-8m+28}{48}-\frac{4\epsilon_0+3\epsilon_1}{12},
\quad m\equiv0\; (\mathrm{mod\,} 2),
\end{equation}
\begin{equation}\label{gIo}
g(\mathbf{L}_m^I)=\frac{m^2-14m+61}{48}-\frac{4\epsilon_0+3\epsilon_1}{12},
\quad m\equiv1\; (\mathrm{mod\,} 2),
\end{equation}
\begin{equation}\label{gIIe}
g(\mathbf{L}_m^{II})=\frac{m^2-8m+16}{16}-\frac{1-\epsilon_1}{4},\quad
m\equiv0\; (\mathrm{mod\,} 2),
\end{equation}
\begin{equation}\label{gIIo}
g(\mathbf{L}_m^{II})=\frac{m^2-6m+9}{16}-\frac{1-\epsilon_1}{4},
\quad m\equiv 1\; (\mathrm{mod\,} 2).
\end{equation}
Degrees of Cohn's polynomials
$$\deg C_m^I=\left\lfloor \frac{(d_m^I)^2-d_m^I+4)}{6}\right\rfloor,$$
$$\deg C_m^{II}=\frac{d_m^{II}(d_m^{II}-1)}{2}.$$
|
\section{Introduction}
\subsection{Background}
In the April 1986 issue of \emph{The American Mathematical Monthly}, Richard Guy~\cite{Guy86} presented an open question, attributed to Boguslav Tomaszewski:
\begin{question}
Consider $n$ real numbers $a_1,a_2,\ldots,a_n$ such that $\sum_i a_i^2=1$. Of the $2^n$ expressions $|\epsilon_1 a_1 + \ldots + \epsilon_n a_n|$, with $\epsilon_i=\pm 1$, $1 \leq i \leq n$, can there be more with value $>1$ than with value $\leq 1$?
\end{question}
In the first paper studying the problem, Holzman and Kleitman~\cite{HK92} presented several equivalent formulations:
\begin{itemize}
\item \emph{Sum partitions:} Let $a_1,\ldots,a_n$ be real numbers with $\sum_i a_i^2=1$. Is it true that in at least half of the partitions of $\sum a_i$ into two sums, the sums differ by at most $1$?
\item \emph{Chebyshev-type inequality:} Let $X=\sum_i a_i x_i$, where $\{x_i\}$ are uniformly distributed in $\{-1,1\}$ and independent. Is it true that $\Pr[|X| \leq \sqrt{\var(X)}] \geq 1/2$? Note that Chebyshev's inequality yields a lower bound of $0$ for this probability.
\item \emph{A ball and a cube:} Consider an $n$-dimensional ball and a smallest $n$-dimensional cube containing it. Is it true that for any pair of parallel supporting hyperplanes of the ball, at least half the vertices of the cube lie between (or on) the two hyperplanes?
\end{itemize}
In the sequel, we use the probabilistic notation $X=\sum a_i x_i$, normalizing such that $\sum a_i^2=1$. Holzman and Kleitman~\cite{HK92} proved the lower bound $\Pr[|X|<1] \geq 3/8$, which is tight for $X=\frac{1}{2}(x_1+x_2+x_3+x_4)$, and were not able to prove a stronger lower bound for $\Pr[|X| \leq 1]$. They conjectured that $\Pr[|X|\leq 1] \geq 1/2$, which would be tight, e.g., for $X=a_1 x_1+a_2 x_2$ with any $|a_1|,|a_2|<1$. This conjecture became known as \emph{Tomaszewski's conjecture} or \emph{Tomaszewski's problem}.
As is suggested by its equivalent formulations, Tomaszewski's problem and its variants naturally appear in diverse fields, including probability theory~\cite{BD15,Pinelis12}, geometric analysis~\cite{KR20}, optimization and operation research~\cite{BNR02,So09}, statistics~\cite{Pinelis94}, and theoretical computer science~\cite{DDS16,DG20,Tan12}. It became well-known, and was mentioned in lists of open problems in various fields (e.g.,~\cite{F+14,Hiriart09}).
A number of works obtained partial results toward Tomaszewski's conjecture. Ben-Tal et al.~\cite{BNR02} (who were not aware of the previous work on the conjecture and arrived to it independently, from applications to optimization) proved that $\Pr[|X|\leq 1] \geq 1/3$. Shnurnikov~\cite{Shnurnikov12} improved their lower bound to $0.36$. Boppana and Holzman~\cite{BH17} were the first to cross the $3/8$ barrier, proving a lower bound of $0.406$, and Boppana et al.~\cite{BHZ20} further improved the lower bound to $0.428$. Very recently, Dvo{\v{r}}{\'{a}}k et al.~\cite{DHT20} proved a lower bound of $0.46$, which is the best currently known bound.
Several other authors proved the conjecture in special cases: Bentkus and Dzinzalieta~\cite{BD15} proved it in the case $\max |a_i| \leq 0.16$, Hendriks and van Zuijlen~\cite{HZ17a} proved it for $n \leq 9$ (a proof-sketch in that case, using different methods, was presented earlier by von Heymann~\cite{Heymann12}), van Zuijlen~\cite{vanZuijlen11} proved it when all $a_i$'s are equal, and Toufar~\cite{Toufar18} extended his result to the case where all $a_i$'s but one are equal. In another direction, De et al.~\cite{DDS16} presented an algorithm that allows approximating $\min \{\Pr[|X|\leq 1]|X=\sum a_i x_i\}$ up to an additive error of $\epsilon$. Unfortunately, the complexity of the algorithm is $\exp(O(1/\epsilon^6))$.
\subsection{Our results}
In this paper we prove Tomaszewski's conjecture.
\begin{theorem}\label{thm:main}
Let $X=\sum_{i=1}^n a_i x_i$, where $\sum_{i=1}^{n} a_i^2=1$ and $\{x_i\}$ are independent and uniformly distributed in $\{-1,1\}$. Then
\begin{equation}\label{eq:main}
\Pr[|X|\leq 1] \geq 1/2.
\end{equation}
\end{theorem}
Our result may be interpreted within the general context of \emph{tail bounds for Rademacher sums}. A Rademacher sum is a random variable $X=\sum a_i x_i$, where the $x_i$'s are i.i.d. Rademacher random variables (i.e., are uniformly distributed in $\{-1,1\}$). Estimates on Rademacher sums were studied in numerous papers, both for their own sake (e.g.,~\cite{BGH01,Veraar08}) and for the sake of applications to statistics~\cite{Efron69} and to optimization~\cite{BNR02}. A main direction in this study is obtaining upper and lower bounds on the tail probability $\Pr[|X| > t]$, aiming at showing that the tail of any such $X$ behaves `similarly' to the tail of a Gaussian random variable with the same variance
(see, e.g.,~\cite{BD15,Eaton70,Pinelis12} for upper bounds and~\cite{HK94,MS90,Oles96} for lower bounds).
One of the upper bounds on the tail, proved by Dzindzalieta~\cite{Dzindzalieta14a}, asserts that
\[
\Pr[X>t] \leq \frac{1}{4} + \frac{1}{8}(1-\sqrt{2-\frac{2}{t^2}}), \qquad \forall 1<t \leq \sqrt{2}.
\]
when $\var(X)=1$.
Theorem~\ref{thm:main} improves over this result, showing that
\begin{equation}
\Pr[X>t] \leq 1/4, \qquad \forall t \geq 1,
\end{equation}
which is tight for all $t<\sqrt{2}$, due to $X=\frac{1}{\sqrt{2}}x_1+\frac{1}{\sqrt{2}}x_2$.
\medskip Furthermore, as was observed by Dzindzalieta~\cite{Dzindzalieta14a}, Theorem~\ref{thm:main} implies the more general:
\begin{equation}\label{eq:dzin-intro1}
\Pr[|X|<t] \geq \Pr[|X|>1/t], \qquad \forall t>0,
\end{equation}
and notably, an improvement of~\eqref{eq:main},
$
\Pr[|X| < 1] + \frac{1}{2}\Pr[|X|=1] \geq \frac{1}{2},
$
being tight for $X= 1\cdot x_1$.
\subsection{Our tools}
The proof of Theorem~\ref{thm:main} uses four main tools.
\paragraph{A new local concentration inequality for Rademacher sums.} In~\cite{KK17}, the authors introduced several local concentration inequalities that allow comparing the probabilities $\Pr[X \in I]$ and $\Pr[X \in J]$, for segments (or rays) $I,J$. We enhance some techniques of~\cite{KK17} and prove:
\begin{theorem}[segment comparison]\label{thm:seg_compare_intro}
Let $X = \sum_{i=1}^{n} a_{i} x_{i}$ be a Rademacher sum, and write $M=\max_i |a_i|$. For any $A, B, C, D \in \mathbb{R}$ with $|A| \leq \min(B,C)$, $2M \leq C-A$ and $D-C + \min(2M, D-B) \leq B-A$, we have
\begin{equation}
\Pr[X \in (C,D)] \leq \Pr[X \in (A,B)].
\end{equation}
\end{theorem}
We prove Theorem~\ref{thm:seg_compare_intro} by constructing an explicit injection.
\paragraph{An improved Berry-Esseen inequality for Rademacher sums.} The classical Berry-Esseen theorem~(\cite{Berry41,Esseen42}) allows approximating a sum of independent random variables by a Gaussian. If $X=\sum a_i x_i$ where all $a_i$'s are `sufficiently small', this allows deducing~\eqref{eq:main}, since for a Gaussian $Z\sim N(0,1)$ we have $\Pr[|Z|\leq 1] \approx 0.68$. Specifically, as was observed by Bentkus and Dzindzalieta~\cite{BD15}, Tomaszewski's conjecture in the case $\max |a_i| \leq 0.16$ follows from the Berry-Esseen bound.
We show that for Rademacher sums, improved Berry-Esseen type bounds (i.e., a tighter approximation by a Gaussian) can be obtained.
We prove these new bounds via a method proposed by Prawitz~\cite{pra72} which employs the characteristic functions of $X$ and of a Gaussian to bound the difference between the distributions. Our key observation here is that in the case of Rademacher sums, Prawitz' method can be refined, yielding significantly stronger bounds. Then, we use these bounds to deduce Theorem~\ref{thm:main} in the range $\max |a_i| \leq 0.31$ (compared to $0.16$ of~\cite{BD15}) and in cases where the second-largest or the third-largest among the $a_i$'s is `sufficiently small'. We conjecture that these bounds can be improved further; see Section~\ref{sec:open}.
\paragraph{A `semi-inductive' approach, using a stopping time argument.} In their proof of the lower bound $\Pr[|X|\leq 1] \geq 1/3$, Ben-Tal et al.\cite{BNR02} introduced a `stopping time' argument, which treats $X$ as a sum $X=X'+X''$ where $X'=\sum_{i=1}^k a_i x_i$ and $X''=\sum_{i=k+1}^n a_i x_i$, and shows that if the partial sum $X'$ is `a little less than $1$', then the final sum $X$ has a decent chance of remaining less than $1$ in absolute value. Variants of this argument (which was attributed in~\cite{BNR02} to P. Van der Wal) were used in all subsequent proofs of lower bounds for Tomaszewski's problem (i.e.,~\cite{BHZ20,BH17,DHT20,Shnurnikov12}).
In our proof, we also employ a `stopping time' argument, but in a somewhat different manner. We use it to show that the statement of Theorem~\ref{thm:main} regarding $X=\sum_{i=1}^n a_i x_i$, with $a_1+a_2\geq 1$, follows from the assertion of Theorem~\ref{thm:main} for $Z=\sum_{i=1}^{m} b_i x_i$, with properly chosen $m<n$ and $\{b_i\}$. However, as it might be that $b_1+b_2 < 1$, this argument is not really inductive, and is completely useless, unless we can prove~\eqref{eq:main} in the case $a_1+a_2<1$ by different tools.
The semi-inductive approach is used once again, when treating the case $a_1 \in [0.31,0.387] \wedge a_1+a_2+a_3 \geq 1$ of Theorem~\ref{thm:main}.
\paragraph{A refinement of Chebyshev's inequality.} We make repeated use of the following, rather standard, refinement of Chebyshev's inequality:
\begin{proposition}\label{prop:Intro-Chebyshev}
Let $X$ be a symmetric (around $0$) random variable with $\var(X) = 1$, and let $c_0,\ldots,c_n,d_1,\ldots,d_m \in \mathbb{R}$ be such that
\[
0 = c_0 \leq c_1 \leq \ldots \leq c_n = 1 = d_0 \leq d_1 \leq \ldots \leq d_m.
\]
Then
\begin{equation}
\sum_{i=0}^{n-1} (1-c_i^2) \Pr[X \in (c_i, c_{i+1}]] \geq \sum_{i=1}^{m} (d_i^2-d_{i-1}^2) \Pr[X \geq d_i].
\end{equation}
\end{proposition}
\paragraph{Proof outline.} In the proof of Theorem~\ref{thm:main}, we consider several cases, according to the sizes of the $a_i$'s, and prove the assertion in each of them separately using the tools described above. In particular, the case $a_1+a_2 \geq 1$ is covered by the aforementioned semi-inductive argument, the case $\max |a_i| \leq 0.31$ is proved by the refined Berry-Esseen inequality, and the cases `in the middle' are proved via various combinations of the segment comparison argument and the refined Chebyshev inequality, sometimes using also the refined Berry-Esseen bound.
Unfortunately, this part of the proof requires somewhat grueling computations, including two cases in which a light computer-aided check is applied. For the sake of readability, we divide the proofs into their `essential' part and their `calculation' part, and relay the calculations to the appendices.
\paragraph{Organization of the paper.}
In Section~\ref{sec:preliminaries} we introduce notation, a basic lemma, and a more detailed outline of the proof. The segment comparison argument (i.e., Theorem~\ref{thm:seg_compare_intro}) is presented in Section~\ref{sec:comparison}. In Section~\ref{sec:BE} we present the refined Berry-Esseen bounds. The semi-inductive argument for the case $a_1+a_2\geq 1$ is given in Section~\ref{sec:induction}. The rest of the cases are presented in Sections~\ref{sec:055}--\ref{sec:remaining} and Appendices~\ref{sec:31big}--\ref{sec:0.31<a_1<0.387}. We conclude with several open problems in Section~\ref{sec:open}.
\section{Preliminaries and Structure of the Proof}
\label{sec:preliminaries}
This section presents notation, a basic lemma, and the structure of the proof in more detail.
\subsection{Notation}
\paragraph{Standard notation.} $\phantom{}$
\medskip For $n \in \mathbb{N}$, $[n]$ denotes $\{1,2,\ldots,n\}$.
\medskip For $A,B,C \in \mathbb{R}$, $A+[B,C]$ denotes the segment $[A+C,B+C]$.
\medskip The shorthand $A_1,A_2,\ldots,A_m=x_1,x_2,\ldots,x_m$ denotes the $m$ equalities $A_i=x_i$, $i\in [m]$.
\medskip The shorthands LHS and RHS denote the left hand side and the right hand side of an equation (or an inequality).
\paragraph{Setting.} Throughout the paper, $X$ denotes a normalized Rademacher sum, that is, $X=\sum_{i=1}^n a_i x_i$, where $\sum_{i=1}^{n} a_i^2=1$ and $\{x_i\}$ are independent and uniformly distributed in $\{-1,1\}$. Without loss of generality, we always assume
\[
a_1 \geq a_2 \geq \ldots \geq a_n > 0.
\]
Note that sometimes, we pass from $X$ to an auxiliary random variable, which we denote by $X'=\sum_{i=1}^m b_i x_i$. In such cases, nothing is assumed on the $\{b_i\}$, unless stated otherwise explicitly.
\paragraph{Precision.} The paper contains explicit real numbers, which we present in decimal expansion. When we write these, we mean the exact value we write; we never write a rounded value and mean `a close' number. We make such roundings by writing, e.g., $\pi=3.1416 \pm 10^{-5}$.
\paragraph{Notation for segments.} We use the somewhat non-standard notation
\begin{equation}\label{eq-def:segment}
\hpr{X}{a}{b} =
\begin{cases}
\Pr[X \in [a,b]] - \Pr[X=a]/2 - \Pr[X=b]/2, & a<b \\
0, & \mrm{Otherwise}
\end{cases}.
\end{equation}
Note that we have
\begin{equation}\label{eq:additivity}
A \leq B \leq C \implies \hpr{X}{A}{C} = \hpr{X}{A}{B} + \hpr{X}{B}{C}.
\end{equation}
Similarly, for $a\leq b$ we denote $\Pr[X \in \langle a,b]] = \Pr[X \in [a,b]]-\Pr[X=a]/2$, and $\Pr[X \in [a,b \rangle]=\Pr[X \in [a,b]] - \Pr[X=b]/2$.
\subsection{A basic lemma -- elimination of variables}
The following lemma allows eliminating several $x_i$'s, by taking into account each possible value of these $x_i$'s separately. The lemma exchanges Tomaszewski's conjecture~\eqref{eq:main} by inequalities which are occasionally easier to approach.
\begin{lemma}\label{lem:nm}
Let $X = \sum_{i=1}^{n} a_i x_i$ be a Rademacher sum with $\var(X) = 1$ and let $m < n$. Write
\begin{equation}\label{Eq:Elimination1}
\sigma = \sqrt{1-\sum_{i=1}^{m} a_i^2} \qquad \mbox{and} \qquad X' = \sum_{i=m+1}^{n} a_i' x_i, \qquad \mbox{with} \qquad a_i' = \frac{a_i}{\sigma}.
\end{equation}
Tomaszewski's assertion~\eqref{eq:main} is equivalent to
\begin{equation}\label{Eq:Elimination2}
\sum_{j=0}^{2^m-1} \Pr[X' > T_j] \leq 2^{m-2},
\end{equation}
where $\{T_j\} = (1 \pm a_1 \pm \cdots \pm a_m)/\sigma$ ranges over all $2^m$ options.
\end{lemma}
\begin{proof}
As $X$ is symmetric,~\eqref{eq:main} is equivalent to $\Pr[X > 1] \leq 1/4$. By the law of total probability,
$
2^{m} \Pr[X > 1] = \sum_{j} \Pr[X' > T_j],
$
implying the lemma.
\end{proof}
Lemma~\ref{lem:nm} is simple yet useful. For example, applying the lemma with $m=1$, one can see that~\eqref{eq:main} is equivalent to the inequality
\begin{equation}\label{eq:nm1}
\Pr[X' \in \li\langle 0, t\right] ] \geq \Pr[X' > 1/t],
\end{equation}
where
\begin{equation}\label{Eq:Prelim-Eliminate-one}
\sigma = \sqrt{1-a_1^2}, \qquad X'=\sum_{i=2}^n \frac{a_i}{\sigma}x_i, \qquad \mbox{and} \qquad t=\sqrt{\frac{1-a_1}{1+a_1}}.
\end{equation}
This readily implies the aforementioned observation of Dzindzalieta~\cite{Dzindzalieta14a} that Tomaszewski's conjecture is equivalent to the more general inequality~\eqref{eq:dzin-intro1}.
\subsection{Hard cases for the proof}\label{ssec:hard-cases}
There are several classes of Rademacher sums which are `hard to handle' with our tools. These classes motivate the partition into cases used in the proof.
One of the obstacles we have to overcome, is the difficulty in distinguishing between the probabilities $\Pr[|X| \leq 1]$ and $\Pr[|X| < 1]$. This obstacle appeared in previous works as well, and is probably the reason for which the `barrier' of $3/8$ (which is the tight lower bound for $\Pr[|X|<1]$) was not beaten for almost 25 years, until the work of Boppana and Holzman~\cite{BH17}.
As a result, `hard' cases for our proof are not only tightness examples for the conjecture, but also $X$'s for which $\Pr[|X|<1]$ is small.
We list three such examples of Rademacher sums:
\begin{enumerate}
\item $X=(x_1+x_2)/\sqrt{2}$, and more generally, $X=\sum a_i x_i$ with $\var(X)=1$ and
\begin{equation}\label{Eq:Aux-Hard-case}
a_1 + \min_{x \in \left\{-1,1\right\}^{[n]\setminus \{1\}}} \left|\sum_{i=2}^{n} a_i x_i\right| > 1.
\end{equation}
\item $X = \frac{1}{2}\sum_{i=1}^{4} x_i$.
\item $X = \frac{1}{3}\sum_{i=1}^{9} x_i$.
\end{enumerate}
Rademacher sums $X$ that belong to the first class are tightness examples for Theorem~\ref{thm:main}, in the sense that $\Pr[|X|\leq 1]=1/2$. These are the only tightness examples we are aware of.
The two latter classes are not tightness cases of the conjecture, but rather satisfy $\Pr[|X| < 1] < 1/2$, and inevitably complicate the proof. For example, in the third case, $\Pr[|X| < 1]=\frac{63}{128} < 0.493$, which demonstrates that the improved Berry-Esseen bound we prove in Section~\ref{sec:BE} is almost optimal. Indeed, while the bound implies that $\Pr[|X| < 1] \geq 1/2$ holds whenever $\max_i |a_i| \leq 0.31$, the example shows that this assertion fails when $\max |a_i|=1/3$.
\subsection{Structure of the proof}
The proof of Theorem~\ref{thm:main} is split into seven cases, which we shortly overview:
\paragraph{Case~1: $a_1 \leq 0.31$.} This case is covered by the improved Berry-Esseen bound for Rademacher sums presented in Section~\ref{sec:BE}.
\paragraph{Case~2: $a_1 + a_2 \geq 1$.} This case is covered by the semi-inductive argument (which uses the `stopping time' method), and is presented in Section~\ref{sec:induction}.
\paragraph{Case~3: $a_1 \geq 0.55$ and $a_1 + a_2 \leq 1$.} The proof in this case combines the `segment comparison' argument (i.e., Theorem~\ref{thm:seg_compare_intro}) with Chebyshev's inequality. The proof in this case is a simple example of the proof strategy in some the following cases, and so we slightly detail about it.
By Lemma~\ref{lem:nm}, applied with $m=1$, in order to verify $\Pr[|X| \leq 1] \geq 1/2$, it suffices to show
\begin{equation}\label{Eq:Prelim-Aux1}
\Pr[|X'| \leq t] > \Pr[|X'| > 1/t],
\end{equation}
where $X',t$ are as defined in~\eqref{Eq:Prelim-Eliminate-one}. Since $a_2'$ -- the largest weight of $X'$ -- satisfies $a_2' \leq t$ (which follows from the condition $a_1+a_2 \leq 1$), Theorem~\ref{thm:seg_compare_intro} implies that $\Pr[|X'| \leq t] = \Omega(t)$. As Chebyshev's inequality yields $\Pr[|X'| > 1/t] \leq t^2$,~\eqref{Eq:Prelim-Aux1} follows if $t$ is sufficiently small, that is, if $a_1$ is sufficiently large. This argument applies when $a_1 \geq 0.55$, and is presented in Section~\ref{sec:055}.
\paragraph{Case~4: $a_1 \in [0.5,0.55]$ and $a_1 + a_2 \leq 1$.} The proof in this case splits according to whether $a_2$ is `large' or `small'. If $a_2$ is small, then the method of the previous case is sufficient. If $a_2$ is large, we eliminate two variables (by applying Lemma~\ref{lem:nm} with $m=2$) and prove the resulting inequality using Theorem~\ref{thm:seg_compare_intro} and the refined Chebyshev inequality (i.e., Proposition~\ref{prop:Intro-Chebyshev}). This case is presented in Section~\ref{sec:0.5<a_1<0.55}.
\paragraph{Case~5: $a_1 \in [0.31, 0.5]$ and $a_1 + a_2 + a_3 \leq 1$.} The proof in this case splits according to the sizes of $a_2$ and $a_3$.
If either of $a_2$ or $a_3$ is sufficiently small, then after elimination of one or two variables (respectively), the problem is reduced to a probabilistic inequality concerning a Rademacher sum with `sufficiently small' weights, and follows from the improved Berry-Esseen inequality.
Otherwise, we eliminate three variables and prove the assertion using Theorem~\ref{thm:seg_compare_intro} and Proposition~\ref{prop:Intro-Chebyshev}.
This case is demonstrated in Section~\ref{sec:remaining} and treated in detail in Appendix~\ref{sec:31big}.
\paragraph{Case~6: $a_1 \in [0.387, 0.5]$ and $a_1 + a_2 + a_3 \geq 1$.} The proof method in this case is superficially similar to that of the previous case. However, a subtle difference in the details makes this case simpler to handle. The proof in this case is demonstrated in Section~\ref{sec:remaining} and treated in detail in Appendix~\ref{sec:39geq}.
\paragraph{Case~7: $a_1 \in [0.31, 0.387]$ and $a_1 + a_2 + a_3 \geq 1$.} The proof in this case splits according to whether there exists some $k\geq 4$ with a `medium-sized' $a_k$.
If there is no such $a_k$, then the weights are partitioned into `large' ones and `small' ones. If there are at most four large weights, then by eliminating the four variables with largest weight, we reduce the problem to a probabilistic inequality concerning a Rademacher sum with `sufficiently small' weights that can be handled easily. Otherwise, we eliminate five variables and show that the assertion follows from Proposition~\ref{prop:Intro-Chebyshev} and a `light' semi-inductive argument. The latter part is however somewhat cumbersome, as after eliminating five variables, we have to deal with $2^5=32$ summands simultaneously.
In the case where there exists a medium-sized $a_k$, we use an explicit bijection to prove a special segment comparison lemma that holds for it, which in turn allows deducing the assertion using Theorem~\ref{thm:seg_compare_intro} and Proposition~\ref{prop:Intro-Chebyshev}. The proof in this case is demonstrated in Section~\ref{sec:remaining} and treated in detail in Appendix~\ref{sec:0.31<a_1<0.387}.
\medskip \noindent Combination of the seven cases with induction over $n$ (where the inductive assumption is used in Cases 2 and 7), completes the proof of Theorem~\ref{thm:main}.
\section{Local Concentration Inequalities for Rademacher Sums}\label{sec:comparison}
In this section we present two concentration inequalities, that allow comparing the probabilities $\Pr[X \in I]$ and $\Pr[X \in J]$, where $X$ is a Rademacher sum and $I,J$ are segments or rays. Both results are used extensively throughout the paper. The first is a local concentration inequality:
\begin{theorem}[segment comparison]\label{thm:seg_compare}
Let $X = \sum_{i=1}^{n} a_{i} x_{i}$, and write $M=\max_i |a_i|$. For all $A, B, C, D \in \mathbb{R}$ such that
\begin{equation}\label{Eq:Comparison_condition_general}
|A| \leq \min(B,C), \qquad 2M \leq C-A, \qquad \mbox{and} \qquad D-C + \min(2M, D-B) \leq B-A,
\end{equation}
one has
\begin{equation}\label{eq:seg_compare}
\Pr[X \in \hseg{C}{D}] \leq \Pr[X \in \hseg{A}{B}].
\end{equation}
\end{theorem}
The assertion of Theorem~\ref{thm:seg_compare} holds for other types of segments as well; see Appendix~\ref{app:sub:other-types}.
\medskip We note that inequalities of the same type were obtained by the authors in~\cite{KK17}, where they were used to obtain an alternative proof of another local tail inequality due to Devroye and Lugosi~\cite{DL08} and to study analytic properties of linear threshold functions. While the inequalities in~\cite{KK17} are qualitative (i.e., of the form $\Pr[X \in \hseg{C}{D}] \leq c \Pr[X \in \hseg{A}{B}]$ for some non-optimal constant $c$), for our purposes here a more exact inequality is required. Such an inequality is given in Theorem~\ref{thm:seg_compare}, which strictly supersedes~\cite[Lemma 3.1]{KK17}.
\medskip The usefulness of `segment comparison' in the proof of Tomaszewski's conjecture is apparent, as the conjecture itself can be rephrased as a segment comparison inequality:
\begin{equation}\label{Eq:Toma-aux}
\hpr{X}{1}{\infty} \leq \hpr{X}{0}{1}.
\end{equation}
However, Theorem~\ref{thm:seg_compare} alone is not sufficient for our needs since it allows deducing $\Pr[X \in I] \leq \Pr[X \in J]$ only for segments $I,J$ that satisfy, in particular, $|I| \leq |J|$.
\medskip
The second result we present is a simple-yet-powerful generalization of the classical Chebyshev's inequality, which allows handling cases where $|I|>|J|$. For example, it enables to deduce $\hpr{X}{0}{1} \geq \hpr{X}{\sqrt{2}}{\infty}$, which reminds of~\eqref{Eq:Toma-aux} but is of course much weaker.
\begin{lemma}\label{lem:cheby}
Let $X$ be a symmetric (around $0$) random variable with $\var(X) = 1$, and let
\[
0 = c_0 \leq c_1 \leq \ldots \leq c_n = 1 = d_0 \leq d_1 \leq \ldots \leq d_m \leq d_{m+1}=\infty.
\]
Then
\begin{equation}\label{eq:cheby_our}
\sum_{i=0}^{n-1} (1-c_i^2) \hpr{X}{c_i}{c_{i+1}} \geq \sum_{i=1}^{m} (d_i^2-d_{i-1}^2) \Pr[X \geq d_i],
\end{equation}
and similarly,
\begin{equation}\label{eq:cheby_our2}
\sum_{i=0}^{n-1} (1-c_i^2) \hpr{X}{c_i}{c_{i+1}} \geq \sum_{i=1}^{m} (d_i^2-1) \hpr{X}{d_i}{d_{i+1}}.
\end{equation}
\end{lemma}
While we did not find this result in the literature, it is presumably known or even folklore. We note that a variant of Lemma~\ref{lem:cheby} was used in the recent work of Dvo{\v{r}}{\'{a}}k et al.~\cite{DHT20}
\medskip Neither Theorem~\ref{thm:seg_compare} nor Lemma~\ref{lem:cheby} is sufficient for tackling Tomaszewski's conjecture. However, a combination of these tools allows proving the conjecture in several significant cases.
\paragraph{Organization.}
The proof of Theorem~\ref{thm:seg_compare} uses an explicit injection, which maps the event on the left hand side of~\eqref{eq:seg_compare} to the event corresponding to the right hand side. We present several auxiliary bijections that will be used to construct our injection in Section~\ref{ssec:inj-maps}, and then we prove Theorem~\ref{thm:seg_compare}
in Section~\ref{ssec:seg-compare}. The simple and standard proof of Lemma~\ref{lem:cheby} is presented in Appendix~\ref{ssec:cheby}.
\subsection{Auxiliary bijections}\label{ssec:inj-maps}
\newcommand{\mathrm{PF}}{\mathrm{PF}}
\newcommand{\mathrm{SF}}{\mathrm{SF}}
\newcommand{\mathrm{RF}}{\mathrm{RF}}
\newcommand{\mathrm{Domain}}{\mathrm{Domain}}
In this subsection we present three bijections on the discrete cube $\{-1,1\}^n$ (i.e., bijections from the discrete cube to itself) that satisfy certain desired properties. The first two bijections were already introduced in~\cite{KK17} and are presented here for the sake of completeness
\subsubsection{Prefix flip}
\begin{lemma}[Prefix flip]\label{lem:prefix-flip}
Let $a = (a_1, \ldots, a_n)$ be a sequence of positive real numbers, let $M=\max_i a_i$ and let $Q \geq 0$. There exists a bijection $\func{\mathrm{PF}_{a, Q}}{\left\{-1,1\right\}^n}{\left\{-1,1\right\}^n}$ such that for any $v \in \left\{-1,1\right\}^n$ with
\begin{equation}\label{eq:pf-assumption}
X(v)\mathrel{\mathop:}= \sum_{i=1}^{n} a_i v_i \geq Q/2,
\end{equation}
the image $w=\mathrm{PF}_{a, Q}(v)$ satisfies
\begin{equation}\label{Eq:Comp-Aux1}
X(w)=\sum_{i=1}^{n} a_i w_i \in (X(v)-Q-2M,X(v)-Q].
\end{equation}
\end{lemma}
\paragraph{Construction of $\mathrm{PF}_{a, Q}$.} Let $v \in \left\{-1,1\right\}^n$ satisfy~\eqref{eq:pf-assumption}. Let $k>0$ be minimal so that $\sum_{j=1}^{k} a_j v_j \geq Q/2$. Define $w=\mathrm{PF}_{a, Q}(v)$ by $w_i = -v_i$ for $i \leq k$, and $w_i = v_i$ for $i > k$.
\medskip \noindent We prove that $\mathrm{PF}_{a, Q}$ satisfies the requirements of the lemma in Appendix~\ref{ssec:prefix-flip}.
We call the function $\mathrm{PF}_{a, Q}$ a \emph{prefix flip}.
\subsubsection{Single coordinate flip}
\begin{lemma}[Single coordinate flip]\label{lem:single-flip}
Let $a = (a_1, \ldots, a_n)$ be a sequence of positive real numbers and let $M=\max_i a_i$ and $m = \min_i a_i$. There exists a bijection $\func{\mathrm{SF}_{a}}{\left\{-1,1\right\}^n}{\left\{-1,1\right\}^n}$ such that for any $v \in \left\{-1,1\right\}^n$ with
\begin{equation}\label{eq:sf-assumption}
X(v)\mathrel{\mathop:}= \sum_{i=1}^{n} a_i v_i > 0,
\end{equation}
the image $w=\mathrm{SF}_{a}(v)$ is obtained from $v$ by flipping a single coordinate from $1$ to $(-1)$, and in particular, satisfies
\[
X(w)=\sum_{i=1}^{n} a_i w_i \in [X(v)-2M,X(v)-2m].
\]
\end{lemma}
\paragraph{Construction of $\mathrm{SF}_{a}$.} Let $v \in \left\{-1,1\right\}^n$ satisfy~\eqref{eq:sf-assumption}. Further assume that $a_1 \geq \ldots \geq a_n > 0$. Let $k > 0$ be the minimal value that maximizes the quantity $\sum_{j=1}^{k} v_j$. Define $w=\mathrm{SF}_{a}(v)$ by $w_i = v_i$ for all $i \neq k$, and $w_k = -v_k$.
\medskip \noindent We prove that $\mathrm{SF}_{a}$ satisfies the requirements of the lemma in Appendix~\ref{ssec:single-flip}. We call the function $\mathrm{SF}_{a}$ a \emph{single coordinate flip}.
\subsubsection{Recursive flip}
\begin{lemma}[Recursive flip]\label{lem:recursive-flip}
Let $a = (a_1, \ldots, a_n)$ be a sequence of positive real numbers and let $M=\max_i a_i$ and $m = \min_i a_i$. There exists a bijection $\func{\mathrm{RF}_{a}}{\left\{-1,1\right\}^n}{\left\{-1,1\right\}^n}$ such that any $v \in \left\{-1,1\right\}^n$ and its image $w = \mathrm{RF}_a(v)$ satisfy:
\medskip\indent \textbf{1)} If $X(v) > 0$ then $X(w) \in [X(v)-2M, X(v)-2m]$;
\medskip\indent \textbf{2)} If $X(v) \leq 0$, then either $w=-v$ (and $X(w)=-X(v)$) or $X(w) \in [X(v) -2M, 0)$,
\medskip \noindent where $X(u) \mathrel{\mathop:}= \sum_{i=1}^{n} a_i u_i$ for $u \in \left\{-1,1\right\}^n$.
\end{lemma}
\paragraph{Construction of $\mathrm{RF}_a$.}
Define the auxiliary injection $\func{F}{\set{v\in\left\{-1,1\right\}^n}{X(v)>0}}{\left\{-1,1\right\}^n}$ by
\[
F(v)=-\mathrm{SF}_{a}(v),
\]
and note that $F$ is not defined for $\set{v}{X(v) \leq 0}$.
Define $\mathrm{RF}_{a}:\left\{-1,1\right\}^n \to \left\{-1,1\right\}^n$ by
\[
\mathrm{RF}_{a}(v) =
\begin{cases}
\phantom{-}\mathrm{SF}_{a}(v), & \quad X(v)>0 \\
-(F^{-1})^{k}(v), & \quad X(v)\leq 0
\end{cases},
\]
where $k\geq 0$ is minimal such that $(F^{-1})^{k+1}(v)$ does not exist, i.e., $(F^{-1})^{k}(v) \notin \mathrm{Image}(F)$.
\medskip
Informally, after defining $\mathrm{RF}_a(v)$ for all $\set{v}{X(v)>0}$ using a single coordinate flip, we would like to define $\mathrm{RF}_a(v)$ for any other $v$ as a simple negation: $\mathrm{RF}_a(v)=-v$. However, this may breach the injectivity if there exists some $w$ such that $X(w)>0$ and $\mathrm{SF}_a(w)=-v$. In such a case, we define $v'=\mathrm{SF}_a^{-1}(-v)$ and check whether $-v'$ is `vacant' (i.e., does not collide with any $\mathrm{SF}_a(z)$). If it is vacant, we set $\mathrm{RF}_a(v)=-v'$; otherwise, we continue applying $\mathrm{SF}_a^{-1}$ and negating until we reach a vacant value. The auxiliary function $F$ combines application of $\mathrm{SF}_a$ with negation, and hence, in each step we apply $F^{-1}$, as stated in the definition. We note that this construction is reminiscent of the way of constructing a bijection from two injections, used in the classical proof of the Cantor-Schr\"oder-Bernstein theorem by J. K\"onig.
\medskip
We call the function $\mathrm{RF}_{a}$ a \emph{recursive flip}. A concrete example that demonstrates the way $\mathrm{RF}_a$ works is presented in Appendix~\ref{app:misc:RF}, and the proof that $\mathrm{RF}_{a}$ satisfies the assertion of the lemma is presented in Appendix~\ref{ssec:recursive-flip}.
\subsection{Proof of Theorem~\ref{thm:seg_compare}}\label{ssec:seg-compare}
The proof of Theorem~\ref{thm:seg_compare} is split into two different arguments, corresponding to how the condition $D-C + \min(D-B, 2M) \leq B-A$ in~\eqref{Eq:Comparison_condition_general} is realized -- either as $D-C+2M \leq B-A$, or as $D-C + (D-B) \leq B-A$. We formulate these two cases as two separate lemmas.
\begin{lemma}\label{lem:seg-compare1}
Let $X = \sum_{i=1}^{n} a_{i} x_{i}$ be a Rademacher sum, and write $M=\max_i a_i$. For any $A, B, C, D \in \mathbb{R}$ such that
\begin{equation}\label{Eq:Aux-Compare-proof1}
\min(|A|, |B|) \leq C \qquad \mbox{and} \qquad D-C+2M \leq B-A,
\end{equation}
we have
\begin{equation}\label{eq:seg-compare1}
\Pr[X \in \hseg{C}{D}] \leq \Pr[X \in \hseg{A}{B}].
\end{equation}
\end{lemma}
\begin{lemma}\label{lem:seg-compare2}
Let $X = \sum_{i=1}^{n} a_{i} x_{i}$ be a Rademacher sum, and write $M=\max_i a_i$. For any $A, B, C, D \in \mathbb{R}$ such that
\begin{equation}\label{Eq:Aux-Compare-proof2}
|A| \leq C \qquad \mbox{and} \qquad 2\max(M, D-B) \leq C-A,
\end{equation}
we have
\begin{equation}\label{eq:seg-compare2}
\Pr[X \in \hseg{C}{D}] \leq \Pr[X \in \hseg{A}{B}].
\end{equation}
\end{lemma}
The combination of Lemmas~\ref{lem:seg-compare1} and~\ref{lem:seg-compare2} immediately implies Theorem~\ref{thm:seg_compare}. Notice that the assumptions in both lemmas are slightly weaker than the assumptions in Theorem~\ref{thm:seg_compare}; this weakening of the assumptions will be needed in the sequel.
\begin{remark}
The assertions of Lemmas~\ref{lem:seg-compare1} and~\ref{lem:seg-compare2} hold for other types of segments as well; see Appendix~\ref{app:sub:other-types}.
\end{remark}
\paragraph{Notation.}
Given a Rademacher sum $X = \sum_{i=1}^{n} a_i x_i$ and a quadruple $A,B,C,D\in \mathbb{R}$ that satisfies either~\eqref{Eq:Aux-Compare-proof1} or~\eqref{Eq:Aux-Compare-proof2} (so that either Lemma~\ref{lem:seg-compare1} or Lemma~\ref{lem:seg-compare2} can be applied), or $D \leq C$, we write $\hseg{C}{D} \prec_X \hseg{A}{B}$. Using this notation, the lemmas can be rewritten as the deduction
\[
\hseg{C}{D} \prec_X \hseg{A}{B} \qquad \implies \qquad \hpr{X}{C}{D} \leq \hpr{X}{A}{B}.
\]
\subsubsection{Proof of Lemma~\ref{lem:seg-compare1}}
\begin{proof}[Proof of Lemma~\ref{lem:seg-compare1}]
We first show that we may assume $|A| \leq |B|$, so that~\eqref{Eq:Aux-Compare-proof1} is upgraded to
\begin{equation}\label{Eq:Aux-Compare-proof3}
|A| \leq C \qquad \mbox{and} \qquad D-C+2M \leq B-A.
\end{equation}
\medskip\noindent\textbf{Upgrading to~\eqref{Eq:Aux-Compare-proof3}.}
If $A > B$, Lemma~\ref{lem:seg-compare1} is vacant. Otherwise, if $|A| \leq |B|$ we are done. In the remaining case, we exchange $A,B,C,D$ by $-B, -A, C, D$, so that~\eqref{Eq:Aux-Compare-proof3} (and hence~\eqref{Eq:Aux-Compare-proof1}) are satisfied by the new quadruple, and conclude~\eqref{Eq:Aux-Compare-proof1} by noting $\hpr{X}{A}{B} = \hpr{X}{-B}{-A}$.
\medskip\noindent\textbf{Proving~\eqref{eq:seg-compare1}.}
We assume $A,B,C,D$ satisfy~\eqref{Eq:Aux-Compare-proof3} and prove~\eqref{eq:seg-compare1}
We may also assume $Q \mathrel{\mathop:}= D-B>0$, as otherwise $D\leq B$, and $[C,D] \subseteq [A,B]$ implies~\eqref{eq:seg-compare1}.
Consider the prefix flip map $\mathrm{PF}_{a,Q}$. For any $v \in \left\{-1,1\right\}^n$ with $X(v) \in [C,D]$, we have $X(v) \geq Q/2$, since by~\eqref{Eq:Aux-Compare-proof3},
\[
Q =D-B \leq D-B + 2M \leq C-A \leq 2C.
\]
Hence, by Lemma~\ref{lem:prefix-flip}, for any such $v$, the image $w=\mathrm{PF}_{a, Q}(v)$ satisfies
\[
X(w) \in (X(v)-Q-2M, X(v)-Q].
\]
As $C \leq X(v) \leq D$ and $Q=D-B$, this implies $X(w) \in (A, B]$. Finally, since $X(w)=B$ may occur only if $X(v)=D$, the injectivity of $\mathrm{PF}_{a, Q}$ implies
\begin{equation}\label{eq:semi-compare}
\Pr[X \in \left[C,D\ri\rangle] \leq \Pr[X \in \left(A, B\ri\rangle],
\end{equation}
which is even slightly stronger than the assertion~\eqref{eq:seg-compare1}. This completes the proof.
\end{proof}
\subsubsection{Proof of Lemma~\ref{lem:seg-compare2}}
The proof of the lemma uses an explicit injection which we hereby describe.
\paragraph{Notation.}
Let $X(x)=\sum_{i=1}^n a_i x_i$ be a Rademacher sum, and $A,B,C,D,M$ be real numbers that satisfy the conditions of Lemma~\ref{lem:seg-compare2}. Note that we may assume $D>B$, as otherwise, the assertion of the lemma holds trivially.
We partition the coefficients $\{a_i\}$ into `large' and `small' ones. Let
\[
L = \left\{ \given{i}{a_{i} \geq (D-B)/2} \right\}, \qquad S = [n] \setminus L.
\]
Note that this partition depends only on the fixed parameters $\{a_i\}$ and $A,B,C,D$.
In addition, for a subset $I \subseteq [n]$ and for $x \in \left\{-1,1\right\}^{n}$, we write $x_I \mathrel{\mathop:}= \restrict{x}{I}$ and $a_I \mathrel{\mathop:}= \restrict{a}{I}$, and correspondingly, $a_I \cdot x_I = \sum_{i \in I} a_{i} x_{i}$.
\theoremstyle{definition}
\newtheorem*{definition*}{Definition}
\begin{definition*}
Let $v \in \left\{-1,1\right\}^n$ be such that $X(v) \in [C,D]$. We define $u=f(v) \in \left\{-1,1\right\}^n$ in two steps. First, we denote
\[
w=g(v)=(\mathrm{RF}_{a_L}(v_L), v_S) \qquad \mbox{and} \qquad Q_{v} = (D-B)-(X(v)-X(w)).
\]
Then, we set
\[
u =f(v)=
\begin{cases}
w, & \quad Q_v \leq 0, \\
(w_L, \mathrm{PF}_{a_S, Q_{v}}(w_S)), & \quad Q_v>0.
\end{cases}
\]
\end{definition*}
\paragraph{Motivation.} We prove Lemma~\ref{lem:seg-compare2} by showing that $f\colon v\mapsto u$ injectively maps any $v$ with $X(v) \in [C,D]$ to some $u$ with $X(u) \in [A,B]$.
The single-coordinate-flip map of the large coordinates, $\mathrm{SF}_{a_L}$, seemingly has the same property, that is, maps any $v'$ with $X(v') \in [C,D]$ to some $w'$ with $X(w') \in [A,B]$.
However, $\mathrm{SF}_{a_L}$ might fail to flip a single coordinate when $a_L \cdot v'_{L} \leq 0$. In such a case, we have $a_S \cdot v'_{S} \geq C \geq (D-B)/2$, and a prefix-flip map $\mathrm{PF}_{a_S, D-B}\colon v' \mapsto u'$ does satisfy $X(u') \in [A,B]$.
This reasoning, of mapping $v'$ into either $w'$ or $u'$, proves that
\[
\hpr{X}{C}{D} \leq 2\cdot\hpr{X}{A}{B},
\]
as it results in a 2-to-1 map (each of $v'\mapsto w'$ and $v'\mapsto u'$ is injective).
We wish to show the stronger inequality $\hpr{X}{C}{D} \leq \hpr{X}{A}{B}$. For this, we construct $f$ in two steps. At the first step we apply on the large coordinates the recursive-flip $\mathrm{RF}_{a_L}$ (which is closely related to $\mathrm{SF}_{a_L}$) to obtain $w$. Then, at the second step, conditioned on data available in $w_L$, we choose whether to apply a prefix flip on the small coordinates or not. This results in a single 1-to-1 map. The usage of $\mathrm{RF}_{a_L}$ instead of $\mathrm{SF}_{a_L}$ is important in order for $w_L$ to have guaranteed properties, even when $a_L \cdot v'_L \leq 0$.
\begin{proof}[Proof of Lemma~\ref{lem:seg-compare2}]
Let $X = \sum_{i=1}^{n} a_i x_i$ and $A,B,C,D,M$ be as in the statement of the lemma, and define the function $f$ on the set $\set{v}{X(v) \in [C,D]}$ as described above.
\paragraph{Why is $f$ injective?}
To compute the inverse map $u \mapsto v$, first consider $u_L$ and apply the inverse recursive flip $\mathrm{RF}_{a_L}^{-1}$ to recover $v_L$. Then, compute $Q_{v}$ (which depends only on $v_L$ and not on $v_S$). Now, it is clear from the definition of $f$ that we have
\[
v =
\begin{cases}
(\mathrm{RF}_{a_L}^{-1}(u_L), u_S), & \quad Q_v \leq 0, \\
(\mathrm{RF}_{a_L}^{-1}(u_L), \mathrm{PF}_{a_S, Q_{v}}^{-1}(u_S)), & \quad Q_v>0.
\end{cases}
\]
\paragraph{Why is $f$ into $[A,B]$?}
Let $v$ satisfy $X(v) \in [C,D]$. We want to show that $u=f(v)$ satisfies $X(u) \in [A,B]$.
We consider two cases.
\paragraph{Case~1: $a_L\cdot v_L > 0$.} In this case, since for all $i \in L$ we have $ (D-B)/2 \leq a_i \leq M$, the first property of the recursive flip (presented in Lemma~\ref{lem:recursive-flip}) implies
\begin{equation}\label{Eq:Aux-Compare-proof4}
X(w) \in [X(v) - 2M, X(v) - (D-B)].
\end{equation}
In particular, we have $Q_{v} = (D-B)-(X(v)-X(w)) \in [D-B-2M,0]$, and thus, by the definition of $f$, we set $u=f(v)=w$. By~\eqref{Eq:Aux-Compare-proof4}, we have
\[
X(u) = X(w) \in [X(v) - 2M, X(v) - (D-B)] \subseteq [C-2M,D-(D-B)] \subseteq [A,B],
\]
where the last inclusion uses the assumption $2M \leq C-A$.
\paragraph{Case~2: $a_L\cdot v_L \leq 0$.} In this case, the second property of $\mathrm{RF}_{a_L}$ (in Lemma~\ref{lem:recursive-flip}) implies
\begin{equation}\label{Eq:Aux-Compare-proof4.5}
a_L\cdot w_L \in [a_L\cdot v_L-2M, -a_L\cdot v_L],
\end{equation}
and hence,
\[
Q_v=(D-B)-(X(v)-X(w) \in (D-B) - [2 a_L \cdot v_L, 2M] = [D-B-2M, D-B-2a_L \cdot v_L].
\]
We further subdivide this case into two sub-cases.
\paragraph{Case~2a: $Q_v\leq 0$.} In this case, we have $Q_v \in [D-B-2M,0]$, and hence,
\[
X(v)-X(w) = (D-B)-Q_v \in [D-B,2M].
\]
As $X(v) \in [C,D]$, this implies
\[
X(w) \in [C-2M,D-(D-B)] \subseteq [A,B],
\]
where the last inclusion follows from the assumption $2M \leq C-A$.
By the definition of the function $f$, we have $u=w$, and so, $X(u)=X(w) \in [A,B]$.
\paragraph{Case~2b: $Q_v> 0$.} In this case, by the definition of $f$, we have $u=f(v)=(w_L, \mathrm{PF}_{a_S, Q_{v}}(w_S))$, and hence, we would like to apply
Lemma~\ref{lem:prefix-flip} to the function $\mathrm{PF}_{a_S,Q_v}:w_S \mapsto u_S$. To this end, we have to prove
\begin{equation}\label{Eq:Aux-Compare-proof5}
a_S \cdot w_S \geq Q_{v}/2.
\end{equation}
To prove~\eqref{Eq:Aux-Compare-proof5}, note that the assumptions $2(D-B) \leq C-A$ and $|A| \leq C$ imply
\begin{equation}\label{Eq:Aux-Compare-proof6}
2C \geq C-A \geq 2(D-B) \geq D-B.
\end{equation}
In addition, by~\eqref{Eq:Aux-Compare-proof4.5} we have
\begin{equation}\label{Eq:Aux-Compare-proof7}
X(v)-X(w)=a_L \cdot v_L -a_L \cdot w_L \geq 2a_L \cdot v_L.
\end{equation}
Since $a_S \cdot w_S = a_S \cdot v_S$ and $a_S \cdot v_S + a_L \cdot v_L = X(v) \in [C,D]$,~\eqref{Eq:Aux-Compare-proof6} and~\eqref{Eq:Aux-Compare-proof7} imply
\[
a_S \cdot w_S = X(v)- a_L \cdot v_L \geq C - a_L\cdot v_L \geq (D-B - (X(v)-X(w)))/2=Q_v/2,
\]
proving~\eqref{Eq:Aux-Compare-proof5}.
\medskip As stated above,~\eqref{Eq:Aux-Compare-proof5} allows us to apply Lemma~\ref{lem:prefix-flip} to the function $\mathrm{PF}_{a_S,Q_v}:w_S \mapsto u_S$. Since $M'=\max_{i \in S} a_i \leq (D-B)/2$, the lemma implies
\begin{align*}
X(u)-X(w) &= a_S \cdot u_S - a_S \cdot w_S \in (-2M'-Q_{v}, -Q_{v}] \\
&\subseteq ((X(v)-X(w)) - 2(D-B), (X(v)-X(w)) - (D-B)] \\
&= (X(v)-X(w)) + (-2(D-B), -(D-B)].
\end{align*}
Therefore,
\[
X(u) = X(v) + (X(w)-X(v)) + (X(u)-X(w)) \in X(v) + (- 2(D-B), -(D-B)].
\]
Since $X(v) \in [C,D]$, we have
\[
X(u) \in (C-2(D-B),D-(D-B)] \subseteq (A,B],
\]
where the last inclusion follows from the assumption $2(D-B)\leq C-A$.
\paragraph{What happens to the endpoints?}
Notice that in all cases, $X(u)=A$ may hold only if $X(v)=C$, and $X(u)=B$ may hold only if $X(v)=D$. Hence, the assertion of the lemma:
\[
\Pr[X \in \hseg{C}{D}] \leq \Pr[X \in \hseg{A}{B}].
\]
follows from the injectivity of $f\colon v \mapsto u$.
\end{proof}
\section{Improved Berry-Esseen Type Inequalities for Rademacher Sums: Proving Theorem~\ref{thm:main} for \tops{$a_1 \leq 0.31$}}\label{sec:BE}
A natural approach toward proving Tomaszewski's conjecture in the case where all the coefficients $a_i$ are small, is using the classical Berry-Esseen theorem~(\cite{Berry41,Esseen42}), which allows approximating a sum of independent random variables by a Gaussian (i.e., a normally distributed random variable).
\begin{theorem}[Berry-Esseen]
Let $X_1,X_2,\ldots,X_n$ be independent random variables, such that $\forall i:\mathbb{E}[|X_i|^3]<\infty$. Let $X=\sum_{i=1}^n X_i$. Then for all $x$,
\begin{equation}\label{Eq:Aux-BE1}
\left|\Pr\left[\frac{X}{\sqrt{\var(X)}} \leq x\right]-\Pr[Z \leq x]\right| \leq C \cdot \frac{\sum_{i=1}^n \mathbb{E}|X_i|^3}{\var(X)^{3/2}},
\end{equation}
where $Z \sim N(0,1)$ is a standard Gaussian and $C$ is an absolute constant.
\end{theorem}
\noindent For a Rademacher sum $X=\sum a_i x_i$ with $\var(X)=1$, the theorem yields
\[
\left|\Pr[X \leq x]-\Pr[Z \leq x]\right| \leq C \cdot \sum_i a_i^3 \leq C \cdot \max_i |a_i| \sum_i a_i^2 = C \cdot \max_i |a_i|,
\]
and consequently,
\begin{equation}\label{Eq:Aux-BE2}
\left|\Pr[|X| \leq 1]-\Pr[|Z| \leq 1]\right| \leq 2C \cdot \max_i |a_i|.
\end{equation}
The best currently known upper bound on the constant $C$ in~\eqref{Eq:Aux-BE1} is $C \leq 0.56$, obtained in~\cite{Shevtsova10}. Plugging it into~\eqref{Eq:Aux-BE2} and noting that
$\Pr[|Z|\leq 1] \geq 0.682$,~\eqref{Eq:Aux-BE2} implies Tomaszewski's conjecture in the range $\max_i |a_i| \leq 0.162$, as was noted by Bentkus and Dzindzalieta~\cite{BD15}.
Using merely the general form of the Berry-Esseen theorem, this result cannot be improved much. Indeed, it was shown by Esseen~\cite{Esseen56} that the constant $C$ in~\eqref{Eq:Aux-BE1} satisfies $C>0.409$, and hence, the best one can hope for by plugging an improved $C$ into~\eqref{Eq:Aux-BE2} is extending the range to $\max_i |a_i| \leq 0.182/0.818 \leq 0.223$.
\medskip In this section we show that refined Berry-Esseen type bounds can be obtained in the special case where $X$ is a Rademacher sum.
Our starting point is a smoothing inequality of Prawitz~\cite{pra72} which allows obtaining bounds on the cumulative distribution function of a random variable $X$ (i.e., $\Pr[X \leq x]$), given partial knowledge of its characteristic function $\varphi_X(t)=\mathbb{E}[e^{itX}]$. Prawitz' inequality has many applications (see Section~\ref{sec:sub:Prawitz}). In particular, Prawitz himself suggested using his inequality to bound $|\Pr[X < x]-\Pr[Z < x]|$, and Shevtsova~\cite{Shevtsova10} used his strategy to prove the best currently known bound on the constant $C$ in the Berry-Esseen theorem.
When $X$ is a Rademacher sum, the strategy of Prawitz can be refined, yielding significantly better bounds than in the general case. Specifically, we obtain the following technical result.
\begin{proposition}\label{prop:Prawitz-Rademacher}
Let $X = \sum_i a_i x_i$ be a Rademacher sum with $a_1 \geq a_2 \geq \ldots \geq a_n>0$ and $\var(X)=1$. Then for any $T>0$, $q \in [0,1]$, and $x \in \mathbb{R}$ we have
\begin{align}\label{Eq:Our-Prawitz}
\begin{split}
\Pr[Z < x] - \Pr[X < x] &\leq
\int_{0}^{q} \left|k(u,x,T)\right| g(Tu) \mrm{d}{u} +
\int_{q}^{1} \left|k(u,x,T)\right| h(Tu) \mrm{d}{u} + \\
&+ \int_{0}^{q} k(u,x,T) \exp(-(Tu)^2/2) \mrm{d}{u} +
\int_{0}^{x} \frac{1}{\sqrt{2\pi}}\exp(-u^2/2) \mrm{d}{u},
\end{split}
\end{align}
where $k(u,x,T) = \frac{(1-u)\sin(\pi u - T u x)}{\sin(\pi u)} - \frac{\sin(T u x)}{\pi}$,
\[
g(v) =
\begin{cases}
\exp(-v^{2}/2) - \cos(a_1 v) ^ {1/a_1^2}, & a_1 v \leq \frac{\pi}{2}\\
\exp(-v^{2}/2)+1, & \mrm{otherwise}
\end{cases}
, \quad h(v) =
\begin{cases}
\exp(-v^{2}/2), & a_1 v \leq \theta\\
(-\cos(a_1 v))^{1/a_1^2}, & \theta \leq a_1 v \leq \pi\\
1, & \mrm{otherwise}
\end{cases},
\]
$Z \sim N(0,1)$ is a standard Gaussian and $\theta=1.778 \pm 10^{-4}$ is the unique root of the function $y \mapsto \exp(-y^2/2)+\cos(y)$ in the interval $[0,\pi]$.
\end{proposition}
While in this paper we bound ourselves to proving Proposition~\ref{prop:Prawitz-Rademacher}, it appears that one can obtain more general estimates about Rademacher sums using similar strategies; see Section~\ref{sec:open}.
\medskip Our first application of Proposition~\ref{prop:Prawitz-Rademacher} validates Tomaszewski's conjecture in the range $\max_i a_i \leq 0.31$ (compared to $0.16$ that can be obtained by the general Berry-Esseen bound):
\begin{proposition}\label{lem:be_toma}
Let $X = \sum_i a_i x_i$ be a Rademacher sum with $\max_i a_i \leq 0.31$ and $\var(X)=1$. Then
\begin{equation}\label{eq:small-toma2}
\Pr[X < 1] \geq \Pr[Z < 1] - 0.09115,
\end{equation}
and consequently,
\begin{equation}\label{eq:small-toma}
\Pr[X < 1] \geq 0.7501, \qquad \Pr[|X| < 1] \geq 0.5002.
\end{equation}
\end{proposition}
Proposition~\ref{lem:be_toma} follows from Proposition~\ref{prop:Prawitz-Rademacher} directly, by substitution of suitable parameters. We note that the condition $\max_i a_i \leq 0.31$ cannot be relaxed significantly. Indeed, this is demonstrated by $X = \frac{1}{3} \sum_{i=1}^{9} x_{i}$ having $\max_i a_i =1/3$ and $\Pr[|X| < 1] = \frac{63}{128} < 0.493$.
\medskip Our second application of Proposition~\ref{prop:Prawitz-Rademacher} is a concrete estimate which we shall use in the proof of Tomaszewski's conjecture in the range $\max_i a_i \in (0.31, 0.5)$.
\begin{proposition}\label{prop:ad-hoc2}
Let $X = \sum_i a_i x_i$ be a Rademacher sum with $\max_i a_i \leq 0.22$ and $\var(X) = 1$. Then, for any $x \geq 0$, we have
\begin{equation}\label{eq:ad-hoc2}
\Pr[X \leq x] \geq \Pr[Z \leq x] - \max(0.084, \Pr[|Z| \leq a_1]/2),
\end{equation}
where $Z \sim N(0,1)$ is a standard Gaussian variable.
\end{proposition}
\paragraph{Organization.}
In Section~\ref{sec:sub:Prawitz} we describe the inequality of Prawitz~\cite{pra72} and apply it to Rademacher sums, proving Proposition~\ref{prop:Prawitz-Rademacher}. In Section~\ref{sec:sub:BE-applications} we prove Propositions~\ref{lem:be_toma} and~\ref{prop:ad-hoc2}.
In Appendix~\ref{app:no-numeric} we show how to practically evaluate the bound~\eqref{Eq:Our-Prawitz} to a required precision.
\subsection{The smoothing inequality of Prawitz applied to Rademacher sums}
\label{sec:sub:Prawitz}
\subsubsection{Prawitz' inequality}
In~\cite{pra72}, H\r{a}kan Prawitz proposed a way for bounding the cumulative distribution function of a random variable $X$, in terms of partial information on its characteristic function $\varphi_X(t)=\mathbb{E}[e^{itX}]$. The main result of~\cite{pra72} (specifically,~\cite[(1b)]{pra72}) reads:
\begin{theorem}\label{thm:prawitz}
Let $X$ be a real-valued random variable, and assume that the characteristic function $\varphi_X(t)=\be[\exp(i t X)]$ is given for $|t| \leq T$. Then for any $x \in \mathbb{R}$,
\begin{equation}\label{eq:prawitz1}
\Pr[X < x] \geq \frac{1}{2} - v.p.\int_{-T}^{T} e^{-ixu} \frac{1}{T} K(-u/T) \varphi_X(u) \mrm{d}{u},
\end{equation}
where
\begin{equation}\label{eq:Ku}
K(u) = \frac{1-|u|}{2} + \frac{i}{2} \left( (1-|u|) \cot(\pi u) + \frac{\mrm{sgn}(u)}{\pi} \right).
\end{equation}
\end{theorem}
The notation $v.p. \int_{-T}^{T}$, which is an abbreviation for \emph{valeur principale} integral, has the meaning $\lim_{\epsilon \to 0^{+}} \left[ \int_{-T}^{-\epsilon} + \int_{\epsilon}^{T} \right]$. It is required, as $K(u) = \frac{i}{2u\pi}+O(1)$ around $u=0$, and a naive integral would diverge. Once using $v.p. \int_{-T}^{T} = \lim_{\epsilon \to {0}} \left[ \int_{-T}^{-\epsilon} + \int_{\epsilon}^{T} \right]$, the integral in~\eqref{eq:prawitz1} converges, as $K$ satisfies $K(-u) + K(u) = O(1)$, and is multiplied by a function which is $C+O(u)$ around $u=0$.
\medskip
Theorem~\ref{thm:prawitz} has numerous applications in probability theory (e.g.,~\cite{BG97,GZ14}) and in statistics (e.g.,~\cite{BGZ97,BG02,HW04}). In particular, already in~\cite[Sec.~10]{pra72}, Prawitz suggested using his method for obtaining approximation by a Gaussian variable in terms of distance between characteristic functions. A few decades later, this strategy was used by Shevtsova (\cite{Shevtsova10}, see also~\cite{KS10}) to obtain the best currently known bound on the constant $C$ in the Berry-Esseen theorem.
For more information on Theorem~\ref{thm:prawitz} and its applications, see the survey~\cite{Bobkov16}.
\subsubsection{A refined inequality for Rademacher sums}
For a Rademacher sum $X = \sum_{i=1}^{n} a_{i} x_{i}$, the characteristic function $\varphi_X$ has the convenient form \begin{equation}\label{Eq:Aux-Prawitz1}
\varphi_{X}(t) = \mathbb{E}[e^{itX}] =\prod_{i=1}^{n} \cos(a_{i} t).
\end{equation}
Since~\eqref{Eq:Aux-Prawitz1} is defined for any $t \in \mathbb{R}$, for any $T>0$ we can substitute in~\eqref{eq:prawitz1} $u/T$ in place of $u$:
\begin{equation}\label{eq:prawitz}
\Pr[X < x] \geq \frac{1}{2} - v.p.\int_{-1}^{1} e^{-ixuT} K(-u) \varphi_X(uT) \mrm{d}{u}.
\end{equation}
Due to the symmetry of $X$ around $0$, we have $\varphi_{X}(-u) = \varphi_{X}(u) \in \mathbb{R}$. This, together with $\overline{K(-u)}=K(u)$, implies
\[
\Pr[X < x] \geq \frac{1}{2} - 2\int_{0}^{1} \mrm{Re} \left( e^{ixuT} K(u) \right) \varphi_{X}(uT) \mrm{d}{u}.
\]
Hence, for $Z \sim N(0,1)$, and for any $T > 0$, $q \in [0,1]$, we have
\begin{equation}\label{eq:prawitz_renew}
\begin{aligned}
\Pr[Z < x]- \Pr[X < x]
\leq
& \left(\Pr[Z < x] - \frac{1}{2} \right)
+ \int_{0}^{1} 2\mrm{Re} \left( e^{iTux} K(u) \right) \varphi_{X}(uT) \mrm{d}{u} = \\
=
& \int_{0}^{q} 2\mrm{Re} \left( e^{iTux} K(u) \right) \left( \varphi_{X}(uT) - \varphi_{Z}(uT) \right) \mrm{d}{u} + \\
& +\int_{q}^{1} 2\mrm{Re} \left( e^{iTux} K(u) \right) \varphi_{X}(uT) \mrm{d}{u} + \\
& +\int_{0}^{q} 2 \mrm{Re} \left( e^{iTux} K(u) \right) \varphi_{Z}(uT) \mrm{d}{u} + \\
& +\left( \Pr[Z < x] - \frac{1}{2} \right) \leq \\
\leq
(S_{1}=) & \int_{0}^{q} \left| 2\mrm{Re} \left( e^{iTux} K(u) \right) \right| \left| \varphi_{X}(uT) - \varphi_{Z}(uT) \right| \mrm{d}{u} + \\
+(S_{2}=) & \int_{q}^{1} \left| 2\mrm{Re} \left( e^{iTux} K(u) \right) \right| \left| \varphi_{X}(uT) \right| \mrm{d}{u} + \\
+(S_{3}=) & \int_{0}^{q} 2 \mrm{Re} \left( e^{iTux} K(u) \right) \varphi_{Z}(uT) \mrm{d}{u} + \\
+(S_{4}=) & \left( \Pr[Z < x] - \frac{1}{2} \right).
\end{aligned}
\end{equation}
Notice that given $x,T,$ and $q$, the quantities $S_{3}, S_{4}$ are two constants which are, in principle, easy to compute, as $\varphi_{Z}(v)=\exp(-v^2/2)$. In the following, we obtain bounds on $S_{1}$ and $S_{2}$, and deduce Proposition~\ref{prop:Prawitz-Rademacher}.
\paragraph{Bounding $\left| \varphi_{X}(uT) \right|$.} We claim that
\begin{equation}\label{eq:be_fx}
\left| \varphi_{X}(v) \right| \leq
\begin{cases}
\exp(-v^{2}/2), & 0\leq a_1 v \leq \theta\\
(-\cos(a_1 v))^{1/a_1^2}, & \theta \leq a_1 v \leq \pi\\
1, & \mrm{Otherwise}
\end{cases},
\end{equation}
where $\theta$ is the unique root of
\[
\exp(-x^2/2)+\cos(x)=0
\]
in the interval $[0,\pi]$. Its numerical value is $1.778\pm 10^{-4}$.
\medskip \noindent
To see this, recall that $\varphi_{X}(v) = \prod_{i} \cos(a_i v)$. Clearly, $|\varphi_{X}(v)| \leq 1$. Moreover, it is easy to check by differentiation that as long as $a_i v \leq \theta$, we have $|\cos(a_i v)| \leq \exp(-(a_i v)^2/2)$. As $a_1=\max_i a_i$, it follows that if $a_1 v \leq \theta$, then
\begin{equation}\label{eq:aux-prawitz}
|\varphi_{X}(v)| = |\prod_{i} \cos(a_i v)| \leq \exp(-\sum_i a_i^2 v^2 / 2) = \exp(-v^2/2).
\end{equation}
To handle the remaining case, $a_1 v \in (\theta, \pi]$, note that
\begin{equation}\label{Eq:Aux-Prawitz2}
|\varphi_X(v)| = \exp\left( \sum_i \log|\cos(a_i v)| \right) = \exp\left( \sum_i a_i^2 \frac{\log|\cos(a_i v)|}{a_i^2} \right)
\leq \exp\left( \max_i \frac{\log|\cos(a_i v)|}{a_i^2} \right),
\end{equation}
where the ultimate inequality holds since $\sum a_i^2 = 1$. The right hand side is maximized at $i=1$. Indeed,
\begin{itemize}
\item For $i$ with $a_i v \leq \theta$, by the previous case we have $\log|\cos(a_i v)|/a_i^2 \leq -v^2/2$.
\item For $i$ with $a_i v \in [\theta, \pi]$, as the function $\psi: a \mapsto \log|\cos(a)|/a^2$ increases in the range $a \in [\theta, \pi]$, we get $\psi(a_i v) \leq \psi(a_1 v)$.
\end{itemize}
Since $a_1=\max_i a_i$ satisfies $a_1 v \in (\theta, \pi]$, a combination of the two cases gives
\[
\forall i: \qquad \log|\cos(a_i v)|/a_i^2 \leq \max(-v^2/2, \log|\cos(a_1 v)|/a_1^2) = \log|\cos(a_1 v)|/a_1^2,
\]
yielding $\varphi_X(v) \leq |\cos(a_1 v)|^{1/a_1^2}$, as we claimed.
\paragraph{Bounding $\left| \varphi_{X}(uT) - \varphi_{Z}(uT) \right|$.} We claim that
\begin{equation}\label{eq:be_fxfz}
\left| \varphi_{X}(v) - \varphi_{Z}(v) \right| \leq
\begin{cases}
\exp(-v^{2}/2) - \cos(a_1 v) ^ {1/a_1^2}, & 0\leq a_1 v \leq \frac{\pi}{2}\\
\exp(-v^{2}/2)+1, & \mrm{Otherwise}
\end{cases}.
\end{equation}
To see this, recall that $\varphi_{Z}(v) = \exp(-v^2/2)$, and hence, for any $v \in \mathbb{R}$,
\[
\left| \varphi_{X}(v) - \varphi_{Z}(v) \right| \leq \exp(-v^2/2)+1.
\]
In the case $0 \leq a_1 v \leq \pi / 2$, on the one hand, similarly to~\eqref{eq:aux-prawitz},
\[
\varphi_X(v) = \prod_i \cos(a_i v) \leq \exp(-\sum_i a_i^2 v^2/2) = \exp(-v^2/2)=\varphi_Z(v).
\]
On the other hand, by virtue of~\eqref{Eq:Aux-Prawitz2} and $\cos(a_i v) \geq 0$,
\[
\prod_i \cos(a_i v) \geq \exp(\min_i (\log(\cos(a_i v)) / a_i^2)).
\]
As the function $a \mapsto \log(\cos(a))/a^2$ decreases in the range $a \in (0, \pi/2)$, one has that for all $i$, $\log(\cos(a_i v)) / a_i^2 \geq \log(\cos(a_1 v)) / a_1^2$, and thus, \[
\varphi_Z(v) \geq \varphi_X(v)=\prod_i \cos(a_i v) \geq \exp(\min_i (\log(\cos(a_i v)) / a_i^2)) = \cos(a_1 v)^{1/a_1^2},
\]
yielding~\eqref{eq:be_fxfz}.
\paragraph{Simplifying $2\mrm{Re} \left( e^{iTux} K(u) \right)$.}
We claim that for all $u \in (0,1)$,
\begin{equation}\label{eq:k_def}
k(u,x,T) \mathrel{\mathop:}= 2\mrm{Re} \left( e^{iTux} K(u) \right) = \frac{(1-u)\sin(\pi u - T u x)}{\sin(\pi u)} - \frac{\sin(T u x)}{\pi}.
\end{equation}
To verify~\eqref{eq:k_def}, notice that by definition, $K(u) = \frac{1-|u|}{2} + \frac{i}{2} \left( (1-|u|) \cot(\pi u) + \frac{\mrm{sgn}(u)}{\pi} \right)$ and $\exp(iTux) = \cos(Tux)+i\sin(Tux)$, and thus, for $u \in (0,1)$ we have
\[
k(u,x,T) = \frac{2\cos(Tux)}{2}(1-u) - \frac{2\sin(Tux)}{2}((1-u)\cot(\pi u)+1/\pi).
\]
Substituting $\cot(\pi u) = \cos(\pi u) / \sin(\pi u)$, we get
\[
k(u,x,T) = (1-u)\left(\cos(Tux) - \frac{\sin(Tux)\cos(\pi u)}{\sin(\pi u)}\right) - \sin(Tux)/\pi.
\]
Using the identity $\sin(\alpha-\beta)=\sin(\alpha)\cos(\beta)-\sin(\beta)\cos(\alpha)$, we derive~\eqref{eq:k_def}.
\paragraph{Combining the bounds.} Substituting the bounds~\eqref{eq:be_fx} and~\eqref{eq:be_fxfz} and the simplification~\eqref{eq:k_def} into~\eqref{eq:prawitz_renew}, we obtain~\eqref{Eq:Our-Prawitz}, namely, the assertion of Proposition~\ref{prop:Prawitz-Rademacher}.
\subsection{Applications of the refined Berry-Esseen type inequalities}
\label{sec:sub:BE-applications}
\subsubsection{Tomaszewski's conjecture for \tops{$a_1 \leq 0.31$}}
We prove Proposition~\ref{lem:be_toma}, which implies Tomaszewski's conjecture in the range $a_1 \leq 0.31$.
\begin{proof}[Proof of Proposition~\ref{lem:be_toma}]
Consider first the case $a_1=0.31$. Applying Proposition~\ref{prop:Prawitz-Rademacher} with $a_1=0.31$, $x=1$, $T=10$ and $q=0.4$, we obtain
\begin{equation}\label{eq:0.31}
\Pr[Z < 1] - \Pr[X < 1] \leq 0.09114 \pm 10^{-5} \leq 0.09115.
\end{equation}
Consequently, $\Pr[X < 1] \geq \Pr[Z < 1] - 0.09115 > 0.7501$, and since $X$ is a symmetric random variable,
\[
\Pr[|X| < 1] = 2\Pr[X < 1] - 1 \geq 0.5002,
\]
as asserted.
To handle the case $a_1 < 0.31$, note that the bound in the right hand side of~\eqref{Eq:Our-Prawitz} is increasing in $a_1$, and thus, an application of Proposition~\ref{prop:Prawitz-Rademacher} with $a_1<0.31$ and the same values $x,T,q$ as above, leads to a stronger lower bound on $\Pr[|X| < 1]$. This completes the proof.
\end{proof}
\subsubsection{A Berry-Esseen type inequality for Rademacher sums with \tops{$a_1 \leq 0.22$}}
To prove Proposition~\ref{prop:ad-hoc2}, we use the following lemma, whose proof is given in Appendix~\ref{ssec:ad-hoc}.
\begin{lemma}
\label{lem:ad-hoc}
Let $X = \sum_i a_i x_i$ be a Rademacher sum with $\max_i a_i \leq 0.22$ and $\var(X) = 1$. Then for every $x \geq 0.35$, we have
\begin{equation}\label{eq:ad-hoc}
\Pr[X \leq x] \geq \Pr[Z \leq x] - 0.084,
\end{equation}
where $Z \sim N(0,1)$ is a standard Gaussian variable.
\end{lemma}
The proof of Lemma~\ref{lem:ad-hoc} proceeds by applying Proposition~\ref{prop:Prawitz-Rademacher} to $X$, with suitably chosen parameters $T,q$, and obtaining a slightly stronger version of~\eqref{eq:ad-hoc} for a finite set of $x$'s. Choosing this set of $x$'s fine enough,~\eqref{eq:ad-hoc} follows for all $x \geq 0.35$, by the monotonicity of $x\mapsto \Pr[X \leq x]$.
\medskip
Proposition~\ref{prop:ad-hoc2} follows from Lemma~\ref{lem:ad-hoc} and Lemma~\ref{lem:seg-compare2}.
\begin{proof}[Proof of Proposition~\ref{prop:ad-hoc2}]
The proof is split according to the value of $x$.
For $x \in [0, a_1)$ we have
\[
\Pr[X \leq x] \geq 1/2 \geq \Pr[Z \leq x] - \Pr[|Z| \leq a_1]/2,
\]
implying~\eqref{eq:ad-hoc2}. For $x \in [0, 0.2]$, we have $\Pr[Z \leq x] \leq 0.58$, and thus, $\Pr[X \leq x] \geq 1/2 \geq \Pr[Z \leq x] - 0.084$, as asserted. For $x \geq 0.35$, the assertion~\eqref{eq:ad-hoc2} follows directly from Lemma~\ref{lem:ad-hoc}.
\medskip Hence, it is left to prove the assertion for $x \in [\max(a_1, 0.2), 0.35]$. We show that in this range, $\Pr[X \leq x] \geq \Pr[Z \leq x] - 0.084$.
Indeed, applying Lemma~\ref{lem:seg-compare2} to $X$, with the parameters $A,B,C,D,M = -x, x, x+\epsilon, 2x+\epsilon/2, a_1$ and letting $\epsilon\to 0^{+}$, we obtain
$
\Pr[|X| \leq x] \geq \Pr[X \in (x, 2x]],
$
and thus,
\[
\Pr[X \in \langle 0,x ] ] \geq \frac{1}{3} \Pr[X \in \langle 0,2x] ].
\]
(Note that the parameters $A,B,C,D,M$ satisfy the assumptions of Lemma~\ref{lem:seg-compare2} since $x \geq a_1$.)
Applying Lemma~\ref{lem:ad-hoc} with the parameter $2x$ (which can be done, as by assumption, $2x \geq 0.4> 0.35$), and using the symmetry of $X$, we get
\[
\Pr[X \leq x] \geq 1/2 + \frac{1}{3}\Pr[X \in \li\langle 0, 2x\right]] \geq \frac{1}{2} + \frac{1}{3}(\Pr[Z \leq 2x] - 0.584).
\]
Therefore, in order to complete the proof it is sufficient to show that
\begin{equation}\label{eq:ad-aux}
\frac{1}{2} + \frac{1}{3}(\Pr[Z \leq 2x] - 0.584) \geq \Pr[Z \leq x] - 0.084,
\end{equation}
The inequality~\eqref{eq:ad-aux} indeed holds for all $x \in [0.2, 0.35]$. To see this, note that the function $x \mapsto \Pr[Z \leq 2x] / 3 - \Pr[Z \leq x]$ is decreasing in $[0.2,0.35]$, and thus, it suffices to verify~\eqref{eq:ad-aux} for $x=0.35$. At $x=0.35$, the inequality holds, completing the proof.
\end{proof}
\section{Theorem~\ref{thm:main} for \tops{$a_1 + a_2 \geq 1$}, via a Semi Inductive Argument}\label{sec:induction}
In this section we prove the following result.
\begin{proposition}\label{prop:semi-inductive}
Let $3 \leq n \in \mathbb{N}$. Assume that for any Rademacher sum $Z=\sum_{i=1}^m a_i x_i$ with $m < n$ and $\var(Z)=1$, we have $\Pr[|Z|\leq 1] \geq 1/2$.
Let $X=\sum_{i=1}^n a_i x_i$ be a Rademacher sum, such that $\var(X)=1$ and $a_1+a_2 \geq 1$. Then $\Pr[|X|\leq 1] \geq 1/2$.
\end{proposition}
There is no restriction in assuming $n \geq 3$, as the assertion of Theorem~\ref{thm:main} (namely, $\Pr[|X|\leq 1] \geq 1/2$) holds trivially for Rademacher sums with $n \leq 2$. Proposition~\ref{prop:semi-inductive} is only `semi-inductive' in the sense that it assumes that Theorem~\ref{thm:main} holds for all $m <n$, with no restriction on the $a_i$'s, and deduces Theorem~\ref{thm:main} for $m=n$ only in the case $a_1+a_2 \geq 1$. Hence, it can be used only if we resolve the case $a_1+a_2<1$ by a different argument, as we do in the following sections.
The proof relies on a `stopping time' argument that is reminiscent of the stopping time argument of Ben-Tal et al.~\cite{BNR02} used in all recent works on Tomaszewski's conjecture.
\paragraph{Elimination of two variables.} We start with the following variant of Lemma~\ref{lem:nm}.
\begin{lemma}\label{cor:nm2}
Let $X = \sum_{i=1}^{n} a_{i} x_i = a_1 x_1 + a_2 x_2 + \sigma X'$ with $\var(X) = \var(X') = 1$ (so that $\sigma=\sqrt{1-a_1^2-a_2^2}$). The assertion
\[
\Pr[|X| \leq 1] \geq 1/2
\]
is equivalent to the following inequality involving $X'$:
\begin{equation}\label{eq:gt1_ineq}
\Pr\left[ X' \in [L_{1}, L_{2}] \right] \geq \Pr \left[ X' > R_{1} \right] + \Pr \left[ X' > R_{2} \right],
\end{equation}
where
\[
L_1, L_2 = \frac{a_1+a_2-1}{\sigma}, \frac{1-a_1+a_2}{\sigma},
\]
and
\[
R_1, R_2 = \frac{1+a_1-a_2}{\sigma}, \frac{1+a_1+a_2}{\sigma}.
\]
\end{lemma}
\begin{proof}
The assertion follows immediately from Lemma~\ref{lem:nm} with $m=2$, since $\Pr[X' \in [L_1, L_2]] = 1 - \Pr[X' > -L_1] - \Pr[X' > L_2]$.
\end{proof}
\noindent Note that if $a_1 \geq a_2$ and $a_1+a_2 \geq 1$ then the parameters $L_1,L_2,R_1,R_2$ satisfy
\begin{equation}\label{Eq:Aux-Induction1}
0 \leq L_1 < L_2 \leq R_1 < R_2.
\end{equation}
This sequence of inequalities will be used several times in the sequel.
\paragraph{The semi-inductive stopping time argument}
Let $X = \sum_{i=1}^{n} a_i x_i$ (with $n \geq 3$) be a Rademacher sum with $\var(X)=1$, and write $X = a_1 x_1 + a_2 x_2 + \sigma X'$, as in Lemma~\ref{cor:nm2}. For $i \geq 3$, write $a_i' = a_i / \sigma$ so that $X' = \sum_{i=3}^{n} a_i' x_i$. By Lemma~\ref{cor:nm2}, in order to deduce $\Pr[|X|\leq 1]\geq 1/2$, it suffices to verify~\eqref{eq:gt1_ineq}.
Write $X'$ as the sum of two random variables $X'=Y'+Z'$, as follows. Let the random variable $k$ to be the minimal index with $\sum_{i=3}^{k} a_{i}'x_{i} \geq L_{1}$, and set
\[
Y' \mathrel{\mathop:}= \sum_{i=3}^{k} a_{i}'x_{i}, \qquad \mbox{and} \qquad Z' \mathrel{\mathop:}= \sum_{i=k+1}^n a_{i}' x_{i}.
\]
If no such index exists, let $k=n$ and set $Z'=0$.
\medskip By lemma~\ref{cor:nm2}, in order to prove $\Pr[|X| \leq 1] \geq 1/2$, it suffices to prove~\eqref{eq:gt1_ineq}, that is
\begin{equation}\label{eq:gt11}
\Pr\left[ Y'+Z' \in [L_{1}, L_{2}] \right] \geq \Pr \left[ Y'+Z' > R_{1} \right] + \Pr \left[ Y'+Z' > R_{2} \right].
\end{equation}
We shall show that~\eqref{eq:gt11} holds even if we condition on any possible value of $k,Y'$. This is clearly sufficient, due to the law of total probability.
For any specific assignment of $k$ and $Y'$, the above inequality is a probabilistic inequality involving the random variable $Z'$. We consider two cases:
\begin{itemize}
\item \emph{Case~1: $Z' \equiv 0$.} We show that in this case,~\eqref{eq:gt11} holds as its right hand side is $0$.
\item \emph{Case~2: $Z' \not \equiv 0$.} In this case, we show that~\eqref{eq:gt11} follows from the inequality~\eqref{eq:gt1_ineq} applied to the Rademacher sum $Z'/\sqrt{\var(Z')}$. To show that the latter inequality holds, we note that when $k$ is fixed, $Z'$ is a Rademacher sum on the $n-k$ variables $x_{k+1},\ldots,x_n$. Hence, by applying Lemma~\ref{cor:nm2} in the inverse direction, we may infer~\eqref{eq:gt1_ineq} for $Z'$ from the assertion $\Pr[|Z|\leq 1]\geq 1/2$ for an appropriate Rademacher sum $Z$ on $n-k+2$ variables, which holds due to the inductive hypothesis, since $k \geq 3$.
\end{itemize}
\noindent For the proof of~\eqref{eq:gt11}, we observe the following relation between $Y'$ and $\var(Z' | Y')$.
\begin{claim}
\label{claim:aux-induction}
Let $X',Y',Z'$ be as defined above, assume $Y' \geq L_1$ and let $s=(1-\sum_{i=3}^k (a'_i)^2)^{1/2}$. We have
\begin{equation}\label{eq:gt1_sy}
Y'(Y' - L_{1}) \leq 1 - s^2.
\end{equation}
\end{claim}
\begin{proof}
By the definition of $k$, we have $Y'-L_{1} \leq a_{k}' \leq a_{k-1} \leq \ldots \leq a_3'$. Thus,
\[
1-s^{2} = \sum_{i=3}^{k} (a'_{i})^{2} \geq \min_{3 \leq i \leq k} \left\{ a_{i}'\right\} \sum_{i=3}^{k} a_{i}' \geq (Y'-L_{1})
\sum_{i=3}^{k} a'_{i} x_{i} = (Y'-L_{1}) Y',
\]
as asserted.
\end{proof}
\paragraph{The case $Z' \equiv 0$.} We observe that in this `singular' case, the right hand side of~\eqref{eq:gt11} is zero, and hence the inequality trivially holds.
Indeed, $Z' \equiv 0$ occurs in one of two cases:
\begin{itemize}
\item There does not exist $k$ such that $Y' = \sum_{i=3}^{k} a_{i}'x_{i} \geq L_{1}$. In this case, $Y'+Z'=\sum_{i=3}^{n} a_{i}'x_{i} < L_1$, and thus, the right hand side of~\eqref{eq:gt11} is clearly equal to zero by~\eqref{Eq:Aux-Induction1}.
\item The minimal $k$ such that $Y' \mathrel{\mathop:}= \sum_{i=3}^{k} a_{i}'x_{i} \geq L_{1}$ is $k=n$.
By~\eqref{Eq:Aux-Induction1}, in order to show that the r.h.s. of~\eqref{eq:gt11} is equal to zero, it is sufficient to prove that $Y'=Y'+Z' < R_1$. To see this, observe that by Claim~\ref{claim:aux-induction}, $Y'(Y'-L_1) \leq 1$, and hence $Y' < R_1$ follows from $R_1(R_1 - L_1) > 1$ (recall $0 \leq L_1 \leq R_1$ by~\eqref{Eq:Aux-Induction1}). This latter inequality reads as $2(1+a_1-a_2)(1-a_2)/\sigma^2 > 1$, where $\sigma^2=1-a_1^2-a_2^2$ and $a_1 \geq a_2$, and follows by
\[
2(1+a_1-a_2)(1-a_2)=\sigma^2+(a_1-a_2)(2+a_1-a_2)+(2a_2-1)^2/2+1/2 > \sigma^2.
\]
\end{itemize}
\paragraph{The case $Z' \not \equiv 0$.} Denote $s = (\var(Z'))^{1/2} = (\sum_{i=k+1}^n (a'_i)^2)^{1/2}$, so that $\frac{1}{s} Z'$ is a Rademacher sum with variance~1 on $n-k$ variables. We have to prove~\eqref{eq:gt11} which reads as
\begin{equation}\label{eq:gt1_need}
\Pr\left[ \frac{Z'}{s} \in \left[ \frac{L_{1}-Y'}{s}, \frac{L_{2}-Y'}{s}\right] \right] \geq \Pr \left[ \frac{Z'}{s} > \frac{R_{1}-Y'}{s} \right] + \Pr \left[ \frac{Z'}{s} > \frac{R_{2}-Y'}{s} \right].
\end{equation}
(Note that we assume $k,Y'$ are fixed, and hence, the probabilities in~\eqref{eq:gt1_need} depend only on $Z'$.)
We would like to deduce~\eqref{eq:gt1_need} from the assertion $\Pr[|Z|\leq 1]\geq 1/2$ for an auxiliary Rademacher sum $Z$ on $n-k+2$ variables, which holds due to the inductive hypothesis (since $n-k+2<n$).
\medskip To this end, we pick $b_1, b_2 \in \mathbb{R}$ such that
\begin{equation}\label{Eq:Aux-Induction3}
\frac{L_{1}-Y'}{s}, \frac{L_{2}-Y'}{s} =\mathrel{\mathop:} L_{1}', L_{2}' = \frac{b_1+b_2-1}{\sigma'}, \frac{1-b_1+b_2}{\sigma'},
\end{equation}
where $\sigma' = \sqrt{1-b_1^2 -b_2^{2}}$. (Concrete values of the possibly negative $b_1, b_2$ are given below.) We define a Rademacher sum $Z = b_1 z_1 + b_2 z_2 + \frac{\sigma'}{s} Z'$, let
\[
R_{1}', R_{2}' = \frac{1+b_1-b_2}{\sigma'}, \frac{1+b_1+b_2}{\sigma'},
\]
and show that $R'_1$ and $R'_2$ satisfy
\begin{equation}\label{eq:gt1_RR}
R_{1}' \leq \frac{R_{1}-Y'}{s} \qquad \mbox{and} \qquad R_{2}' \leq \frac{R_{2}-Y'}{s}.
\end{equation}
As $Z$ is a Rademacher sum with variance~1 on $n-k+2<n$ variables, the induction hypothesis together with Lemma~\ref{cor:nm2} yields
\[
\Pr\left[\frac{Z'}{s} \in [L'_{1}, L'_{2}] \right] \geq \Pr \left[\frac{Z'}{s} > R'_{1} \right] + \Pr \left[\frac{Z'}{s} > R'_{2} \right],
\]
which implies~\eqref{eq:gt1_need} via~\eqref{eq:gt1_RR}. So, it is only left to show that $Z$ is well defined and that~\eqref{eq:gt1_RR} holds.
\paragraph{Why is $Z$ well-defined?}
To show that there exist $b_1,b_2 \in \mathbb{R}$ that satisfy~\eqref{Eq:Aux-Induction3}, let
\[
b_1 = \frac{2 + 2L_1' L_2'}{2 + (L'_1)^2 + (L'_2)^2} \qquad \mbox{and} \qquad b_2 = \frac{(L'_2)^2 - (L'_1)^2}{2 + (L'_1)^2 + (L'_2)^2}.
\]
A direct computation shows that
\begin{equation}\label{Eq:Aux-Induction4}
\sigma' = \sqrt{1-b_1^2-b_2^2} = \frac{2(L'_2-L'_1)}{2+(L'_1)^2+(L'_2)^2},
\end{equation}
and in particular, $b_1^2+b_2^2 < 1$ (via~\eqref{Eq:Aux-Induction1} and the definition of $L_1', L_2'$ in~\eqref{Eq:Aux-Induction3}).
A further direct computation shows that the right equality in~\eqref{Eq:Aux-Induction3} holds as well.
\paragraph{Proving~\eqref{eq:gt1_RR}.}
To verify the two inequalities in~\eqref{eq:gt1_RR}, notice that by adding $L_{2}'$ to both sides of the first, and $-L_1'$ to both sides of the second, and recalling the definitions of $L_1,L_2,R_1,R_2$ and $L_1', L_2', R_1', R_2'$, these inequalities are respectively equivalent to:
\[
\frac{s}{\sigma'} + Y' \leq \frac{1}{\sigma} \qquad \mbox{and} \qquad \frac{s}{\sigma'} \leq \frac{1}{\sigma}.
\]
As $Y' \geq L_1 \geq 0$ by~\eqref{Eq:Aux-Induction1}, the former inequality clearly implies the latter; hence, we focus only on it. Note that similarly to~\eqref{Eq:Aux-Induction4}, we have
\[
\sigma = \frac{2(L_2-L_1)}{2 + L_1^2 + L_2^2}.
\]
In addition, by~\eqref{Eq:Aux-Induction3}, we have $L'_2-L'_1 = (L_2-L_1)/s$. Using this and substituting the values of $\sigma,\sigma'$, the inequality we seek to prove $\frac{s}{\sigma'} + Y' \leq \frac{1}{\sigma}$, reads as
\[
\frac{s^{2}(2+L_1'^2+L_2'^2)}{2(L_2-L_1)} + Y' \leq \frac{(2+L_1^2+L_2^2)}{2(L_2-L_1)}.
\]
Multiplying by $2(L_2-L_1)$ and substituting the values of $L_1', L_2'$ from~\eqref{Eq:Aux-Induction3}, we reduce to showing
\[
2s^2 + (L_1-Y')^2 + (L_2-Y')^2 + 2Y'(L_2-L_1) \leq 2 + L_1^2 + L_2^2,
\]
Since $s^2 \leq 1-Y'(Y' - L_{1})$ by Claim~\ref{claim:aux-induction}, it is sufficient to prove that
\[
2 - 2Y'(Y'-L_1) + (L_1-Y')^2 + (L_2-Y')^2 + 2Y'(L_2-L_1) \leq 2+L_1^2+L_2^2.
\]
Simplifying this inequality, one sees it is equivalent to the inequality $2Y' \cdot L_{1} \geq 0$, which indeed holds since by~\eqref{Eq:Aux-Induction1}, we have $0 \leq L_{1} \leq Y'$. This completes the proof of Proposition~\ref{prop:semi-inductive}.
\section{Theorem~\ref{thm:main} for \tops{$a_1 \geq 0.55$}, \tops{$a_1+a_2 < 1$}}\label{sec:055}
In the previous sections we handled Theorem~\ref{thm:main} in the cases $a_1 \leq 0.31$ (Section~\ref{sec:sub:BE-applications}), and $a_1+a_2 \geq 1$ (Section~\ref{sec:induction}, semi-inductively). In this section we handle the case $(a_1 \geq 0.55) \wedge (a_1 + a_2 < 1)$.
\paragraph{Elimination step.} Let $X = \sum_{i=1}^{n} a_i x_i$ be a Rademacher sum with $\var(X)=1$, $a_1 \geq (1+\sqrt{8})/7$, and $a_1+a_2 < 1$. (Note that $1+\sqrt{8}/7 \leq 0.55$.)
We want to prove $\Pr[|X| \leq 1] \geq 1/2$.
\medskip By the case $m=1$ of Lemma~\ref{lem:nm}, it is sufficient to prove that
\begin{equation}\label{Eq:Aux6.1}
\Pr[X' \in \li\langle 0, t\right] ] \geq \Pr[X' > 1/t],
\end{equation}
where
\[
\sigma = \sqrt{1-a_1^2}, \qquad a'_i = \frac{a_i}{\sigma}, \qquad X'=\sum_{i=2}^n a'_i x_i, \qquad \mbox{and} \qquad t=\sqrt{\frac{1-a_1}{1+a_1}}=\frac{1-a_1}{\sigma}
\]
(see~\eqref{eq:nm1} and~\eqref{Eq:Prelim-Eliminate-one} above). As we assume $a_1 + a_2 < 1$, we have
\begin{equation}\label{Eq:Aux6.2}
\forall i\colon a_i' < t.
\end{equation}
\paragraph{Segment comparison step.}
\begin{claim}\label{Cl:Aux6.1}
Let $X',t$ be as defined above. Then:
\begin{enumerate}
\item[(a)] $\hpr{X'}{t}{2t} \leq 2\hpr{X'}{0}{t},$ and
\item[(b)] $\forall k \geq 2 \colon \hpr{X'}{kt}{(k+1)t} \leq 4\hpr{X'}{0}{t}$.
\end{enumerate}
\end{claim}
\begin{proof}$\phantom{}$
\noindent (a) Applying Lemma~\ref{lem:seg-compare2} to $X'$, with the parameters $A,B,C,D,M=-t,t,t,2t,a_2'$, we get
\[
\hpr{X'}{t}{2t} \leq \hpr{X'}{-t}{t}.
\]
(Note that $2M\leq C-A$ follows from~\eqref{Eq:Aux6.2}). By the symmetry of $X'$, this implies~(a).
\medskip \noindent
(b) Applying Lemma~\ref{lem:seg-compare1} to $X'$, with the parameters $A,B,C,D,M=-t, 2t, kt, (k+1)t, a_2'$, we get $\hpr{X'}{kt}{(k+1)t} \leq \hpr{X'}{-t}{2t}$. Using (a) as
\[
\hpr{X'}{-t}{2t} = \hpr{X'}{-t}{t} + \hpr{X'}{t}{2t} \leq 4 \hpr{X'}{0}{t},
\]
the assertion (b) follows.
\end{proof}
\paragraph{Chebyshev-type inequality step.} Applying the inequality~\eqref{eq:cheby_our} to $X'$, with
\[
c_0, c_1, c_2, c_3, \ldots, c_{\lceil 1/t \rceil} = 0, t, 2t, 3t, \ldots, 1 \qquad \mbox{and} \qquad d_0, d_1 = 1, 1/t,
\]
we obtain
\[
\sum_{k=0}^{\lceil 1/t \rceil-1} (1-(k t)^2) \hpr{X'}{k t}{(k+1)t} \geq \Big(\frac{1}{t^2}-1 \Big) \Pr[X' \geq 1/t].
\]
By Claim~\ref{Cl:Aux6.1}, this implies
\[
\Big( 1 + 2(1-t^2) + 4\sum_{k=2}^{\lceil 1/t \rceil -1} \left( 1-(kt)^2\right) \Big) \hpr{X'}{0}{t}
\geq
\Big( \frac{1}{t^2} - 1 \Big) \Pr[X' \geq 1/t].
\]
Dividing both sides by $1/t^2-1$, we obtain an inequality of the form
\[
C_{t} \cdot \hpr{X'}{0}{t} \geq \Pr[X' \geq 1/t].
\]
This inequality implies~\eqref{Eq:Aux6.1}, provided $C_{t} \leq 1$. Hence, it is left to verify:
\begin{equation}\label{eq:055fin}
C_{t} \stackrel{?}{\leq} 1, \quad \text{with} \quad C_{t} = \frac{1 + 2(1-t^2) + 4\sum_{k=2}^{\lceil 1/t \rceil-1} \left( 1-(kt)^2\right)}{(1/t^2) - 1}, \quad 0 < t \leq \sqrt{1-2^{-1/2}}.
\end{equation}
(Note that the condition $0< t \leq (1-2^{-1/2})^{1/2}$ follows from the assumption $a_1 \in \left[ (1+\sqrt{8})/7, 1\right)$, via $t=\sqrt{(1-a_1)/(1+a_1)}$. This is the only place where the assumption $a_1 \geq 0.55$ is used.) The derivation of~\eqref{eq:055fin} is presented in Appendix~\ref{app:055fin}.
\section{Theorem~\ref{thm:main} for \tops{$0.5 \leq a_1 \leq 0.55$}, \tops{$a_1 + a_2 < 1$}}\label{sec:0.5<a_1<0.55}
The proof of Theorem~\ref{thm:main} in this range is split into two cases: Small $a_2$ and large $a_2$. The threshold between the cases depends on $a_1$, being $(a_1-3+\sqrt{25+10a_1-63a_1^2})/8$.
\subsection{The case \tops{$a_2 \leq (a_1-3+\sqrt{25+10a_1-63a_1^2})/8$}}
Let $X = \sum a_i x_i$ be a Rademacher sum with $\var(X)=1$, $a_1 \in [0.5, 0.55]$, $a_1+a_2 < 1$, and $a_2 \leq (a_1-3+\sqrt{25+10a_1-63a_1^2})/8$. The proof that $\Pr[|X| \leq 1] \geq 1/2$ is almost identical to the argument of Section~\ref{sec:055}, as we explain below.
\paragraph{Elimination step.} Like in Section~\ref{sec:055}, we note that it is sufficient to prove $\Pr[X' \in \li\langle 0, t\right] ] \geq \Pr[X' > 1/t]$, where $X',t$ are defined as in~\eqref{Eq:Aux6.1}.
\paragraph{Segment comparison step.} Instead of Claim~\ref{Cl:Aux6.1}, we use the following comparisons:
\begin{equation}\label{eq:05rep}
\hpr{X'}{t}{3t/2} \leq \hpr{X'}{0}{t} \quad \mbox{and} \quad \hpr{X'}{3t/2}{1} \leq \hpr{X'}{0}{t}.
\end{equation}
The first inequality follows from Lemma~\ref{lem:seg-compare2}, applied to $X'$ with the parameters $A,B,C,D,M = 0, t, t, 3t/2, a_2/\sigma$, and the second inequality follows from Lemma~\ref{lem:seg-compare1}, applied to $X'$ with the parameters $A,B,C,D,M = 0, t, 3t/2, 1, a_2/\sigma$. To show that Lemmas~\ref{lem:seg-compare2} and~\ref{lem:seg-compare1} indeed can be applied (i.e., that the assumptions of the lemmas are satisfied), it is sufficient to verify:
\begin{equation}\label{eq:05saux}
\begin{gathered}
\frac{2a_2}{\sigma} \stackrel{?}{\leq} t, \qquad \quad (1 - 3t/2) + \frac{2a_2}{\sigma} \stackrel{?}{\leq} t\\
\text{with:}\qquad a_1 \in [0.5, 0.55],\quad\sigma = \sqrt{1-a_1^2},\quad\\ t=\sqrt{\frac{1-a_1}{1+a_1}},\quad a_2 \leq \frac{a_1-3+\sqrt{25+10a_1-63a_1^2}}{8}.
\end{gathered}
\end{equation}
These inequalities are proved in Appendix~\ref{app:05saux}.
\paragraph{Chebyshev-type inequality step.} Note that as $a_1 \geq 1/2$, we have $t=\sqrt{(1-a_1)/(1+a_1)} \leq 1/\sqrt{3}< 2/3$.
Applying the inequality~\eqref{eq:cheby_our} to $X'$, with
\[
c_0, c_1, c_2, c_3 = 0, t, 3t/2, 1 \qquad \mbox{and} \qquad d_0, d_1 = 1, 1/t,
\]
and using~\eqref{eq:05rep}, we obtain
\[
(1 + (1-t^2) + (1-(3t/2)^2)) \hpr{X'}{0}{t} \geq \left(\frac{1}{t^2}-1\right) \Pr[X' \geq 1/t].
\]
Hence, the task of deducing~\eqref{Eq:Aux6.1} boils down to verifying:
\begin{equation}\label{eq:05rep2}
C_t' \stackrel{?}{\leq} 1, \qquad \text{with:}\qquad C_t' = \frac{1 + (1-t^2) + (1-(3t/2)^2)}{(1/t^2)-1},\quad t^2 \in (0, 1/3].
\end{equation}
This verification is done in Appendix~\ref{app:05fin2}.
\subsection{The case \tops{$a_2 \geq (a_1-3+\sqrt{25+10a_1-63a_1^2})/8$}}
Let $X = \sum a_i x_i$ be a Rademacher sum with $\var(X)=1$, $a_1 \in [0.5, 0.55]$, $a_1+a_2 < 1$, and $a_2 \geq (a_1-3+\sqrt{25+10a_1-63a_1^2})/8$. The proof of $\Pr[|X| \leq 1] \geq 1/2$ is similar to the above strategy, but this time, two variables are eliminated.
\paragraph{Elimination step.} By Lemma~\ref{cor:nm2}, it it sufficient to show that
\begin{equation}\label{eq:05need}
\hpr{X'}{0}{-L_1} + \hpr{X'}{0}{L_2} \geq \Pr[X' > R_1] + \Pr[X' > R_2],
\end{equation}
where $X', \sigma, L_1, L_2, R_1, R_2$ are as defined in Lemma~\ref{cor:nm2}. (But this time, since $a_1+a_2<1$ we have $L_1<0$). Note that unlike the previous subsection, $X'$ depends on $n-2$ variables.
\paragraph{Auxiliary estimates.} We use several auxiliary estimates on $L_2,R_1,R_2$:
\begin{equation}\label{eq:05fin}
\begin{gathered}
L_2 \stackrel{?}{\geq} \frac{2}{3}, \qquad R_1 \stackrel{?}{\geq} \max(\sqrt{3 - L_2^2},\sqrt{2}), \qquad R_2 \stackrel{?}{\geq} \max(\sqrt{5 - 2L_2^2},\sqrt{3}),\\
\text{with:}\qquad L_2 = \frac{1-a_1+a_2}{\sigma},\quad R_1 = \frac{1+a_1-a_2}{\sigma},\quad R_2 = \frac{1+a_1+a_2}{\sigma}, \\
\sigma = \sqrt{1-a_1^2-a_2^2},\quad
a_1 \in [0.5, 0.55],\quad
a_1+a_2 < 1,\\
a_2 \geq \left( a_1-3+\sqrt{25+10a_1-63a_1^2}\right)/8.
\end{gathered}
\end{equation}
These inequalities are proved in Appendix~\ref{app:05fin}.
\paragraph{The sub-case $L_2 \geq 1$.} Applying the Chebyshev-type inequality~\eqref{eq:cheby_our} to $X'$, with $c_0,c_1=0,1$ and $d_1,d_2,d_3=1,\sqrt{2},\sqrt{3}$, we obtain
\[
\hpr{X'}{0}{1} \geq \Pr[X' \geq \sqrt{2}] + \Pr[X' \geq \sqrt{3}],
\]
which implies~\eqref{eq:05need} via~\eqref{eq:05fin}.
\paragraph{The sub-case $L_2 < 1$.} Applying Lemma~\ref{lem:seg-compare2} to $X'$, with the parameters $A,B,C,D,M = 0, L_2, L_2, 1, a_{3}/\sigma$, we obtain
\begin{equation}\label{Eq:Aux7.1}
\hpr{X'}{L_2}{1} \leq \hpr{X'}{0}{L_2}.
\end{equation}
Notice that the assumptions of Lemma~\ref{lem:seg-compare2} are satisfied, as $2(D-B)=2(1-L_2)\leq L_2=C-A$ by~\eqref{eq:05fin}, and
\[
2M \leq \frac{2a_2}{\sigma} < \frac{1-a_1+a_2}{\sigma} = L_2 =C-A,
\]
by the assumption $a_1+a_2 < 1$.
\medskip Applying the Chebyshev-type inequality~\eqref{eq:cheby_our} to $X'$, with $c_0,c_1,c_2=0,L_2,1$ and $d_1,d_2,d_3=1,\sqrt{3-L_2^2},\sqrt{5-2L_2^2}$, we obtain
\begin{equation}\label{eq:05cheby}
\begin{split}
\hpr{X'}{0}{L_2} &+ (1-L_2^2)\hpr{X'}{L_2}{1} \geq \\
&\geq (2-L_2^2)\Pr\left[X'\geq \sqrt{3-L_2^{2}}\right] + (2-L_2^2)\Pr\left[X' \geq \sqrt{5-2L_2^{2}}\right].
\end{split}
\end{equation}
By~\eqref{Eq:Aux7.1}, this implies
\[
\hpr{X'}{0}{L_2} \geq \Pr\left[X'\geq \sqrt{3-L_2^{2}}\right] + \Pr\left[X' \geq \sqrt{5-2L_2^{2}}\right],
\]
which, in turn, implies~\eqref{eq:05need} via~\eqref{eq:05fin}. This completes the proof.
\section{The remaining case: $a_1 \in (0.31, 0.5)$}\label{sec:remaining}
Our proof is most involved in this range, although there are no remarkable tightness examples with $a_1 \in (0.31, 0.5)$ (except for $X = \frac{1}{3} \sum_{i=1}^9 x_i$, being the `lightest' appearing in Section~\ref{ssec:hard-cases}).
The full proof in this range is deferred to appendices~\ref{sec:31big},~\ref{sec:39geq}, and~\ref{sec:0.31<a_1<0.387}, corresponding to three main subcases. The general structure of the proof is to condition on the values of a few largest weights (say, $a_1, a_2, a_3$), usually through elimination (Lemma~\ref{lem:nm}), and to prove Tomaszewski's assertion~\eqref{eq:main} regardless of the values we condition on.
While the careful proof is somewhat cumbersome, it was plotted by considering several specific fixings of the large weights (say, $a_1,a_2,a_3 = 0.4, 0.3, 0.2$) and proving the assertion~\eqref{eq:main} under these fixings. Then, the proof was generalized to capture any such fixing.
We demonstrate the proof by presenting several (not fully) representative such fixings, and proving~\eqref{eq:main} under them.
\subsection{Subcase $a_1 + a_2 + a_3 \leq 1$}\label{ssec:rem-case1}
The following example concisely demonstrates all the steps in the proof of this case (given in Appendix~\ref{sec:31big}).
Assume
\[
a_1, a_2, a_3 = 0.36, 0.2, 0.15,
\]
so that 3-elimination (Lemma~\ref{lem:nm}) reduces us to proving the following inequality for all Rademacher sums $X'$ with $\var(X')=1$, whose largest weight is $\leq 0.15/(1-0.36^2-0.2^2-0.15^2)^{1/2} < 0.17$:
\begin{equation}\label{eq:demo1}
\begin{gathered}
\hpr{X'}{0}{0.32} + \hpr{X'}{0}{0.65} + \hpr{X'}{0}{0.76} + \hpr{X'}{0}{1.1}
\\ \geq \\
\Pr[X' > 1.12] + \Pr[X' > 1.45] + \Pr[X' > 1.56] + \Pr[X' > 1.9]
\end{gathered}
\end{equation}
Using Chebyshev's inequality~\eqref{eq:cheby_our}, we lower bound an expression similar to the LHS of~\eqref{eq:demo1}:
\begin{equation}\label{eq:demo1-cheby}
\begin{gathered}
\hpr{X'}{0}{0.32} + 0.9\hpr{X'}{0.32}{0.65} +\\
0.6\hpr{X'}{0.65}{0.76} + 0.5\hpr{X'}{0.76}{1}
\\ \geq \\
(0.25\Pr[X' > 1.12] + 0.25\Pr[X' > 1.23]) + 0.5\Pr[X' > 1.42] +\\
(0.4\Pr[X' > 1.56] + 0.1\Pr[X' > 1.6]) + 0.5\Pr[X' > 1.75].
\end{gathered}
\end{equation}
Can we deduce~\eqref{eq:demo1} from~\eqref{eq:demo1-cheby}? Not immediately. If we multiply~\eqref{eq:demo1-cheby} by $2$ and see what is `missing' in order to deduce~\eqref{eq:demo1}, we get an inequality weaker than:
\[
\hpr{X'}{0}{0.32} + \hpr{X'}{0}{0.65} \geq 0.5\hpr{X'}{1.12}{1.23} + 0.2\hpr{X'}{1.56}{1.6}.
\]
Recall that the largest weight of $X'$ is at most $0.17$, so using segment comparison (specifically, Lemma~\ref{lem:seg-compare1}), we can prove this inequality by showing $\hpr{X'}{1.12}{1.23} \leq \hpr{X'}{0}{0.65}$ and $\hpr{X'}{1.56}{1.6} \leq \hpr{X'}{0}{0.65}$.
\subsection{Subcase $a_1 + a_2 + a_3 \geq 1$ and $a_1 \in (0.387, 0.5)$}\label{ssec:rem-case2}
The proof in this range is similar to that of Section~\ref{ssec:rem-case1}, but the details are much simpler. We again give a demonstration that captures the essence of the proof (given in Appendix~\ref{sec:39geq}).
Assume
\[
a_1, a_2, a_3 = 0.4, 0.35, 0.3,
\]
so that 3-elimination (Lemma~\ref{lem:nm}) reduces us to prove the following inequality for all Rademacher sums $X'$ with $\var(X')=1$, whose largest weight is $\leq 0.3/(1-0.4^2-0.35^2-0.3^2)^{1/2} < 0.38$:
\begin{equation}\label{eq:demo2}
\begin{gathered}
\hpr{X'}{0.07}{0.69} + \hpr{X'}{0}{0.82} + \hpr{X'}{0}{0.94}
\\ \geq \\
\Pr[X' > 1.57] + \Pr[X' > 1.7] + \Pr[X' > 1.83] + \Pr[X' > 2.58]
\end{gathered}
\end{equation}
We prove~\eqref{eq:demo2} even without the $\hpr{X'}{0.07}{0.69}$ term. Using Chebyshev's inequality~\eqref{eq:cheby_our}:
\begin{equation}\label{eq:demo2-cheby}
\begin{gathered}
\hpr{X'}{0}{0.82} + 0.33\hpr{X'}{0.82}{0.94} + 0.12\hpr{X'}{0.94}{1}
\\ \geq \\
0.7 \Pr[X' > 1.31] + 0.7 \Pr[X' > 1.55] + 0.7 \Pr[X' > 1.77] + 0.7 \Pr[X' > 1.95].
\end{gathered}
\end{equation}
Can we deduce~\eqref{eq:demo2} from~\eqref{eq:demo2-cheby}? Not immediately. If try to deduce~$0.7\cdot$\eqref{eq:demo2} from~\eqref{eq:demo2-cheby} we see that what is `missing', is weaker than:
\[
0.4\hpr{X'}{0}{0.82}-0.12\hpr{X'}{0.94}{1} \geq 0.
\]
Using segment comparison, and specifically Lemma~\ref{lem:seg-compare1}, barely applicable since $(1-0.94) + 2\cdot 0.38 = 0.82$, we conclude $\hpr{X'}{0.94}{1} \leq \hpr{X'}{0}{0.82}$. Also, we are equally satisfied with
\[
\hpr{X'}{0.94}{1} \leq \hpr{X'}{-0.82}{0.82} \leq 2\hpr{X'}{0}{0.82},
\]
which liberally follows from Lemma~\ref{lem:seg-compare1} (and whose analog holds true in the entire range).
\subsection{Subcase $a_1 + a_2 + a_3 \geq 1$ and $a_1 \in (1/3, 0.387)$}
The proof in this case is a bit more complicated than in the other cases, and the demonstration we give does not capture the entire set of arguments we use. However it represents perhaps the most exotic argument, and features a semi-inductive argument along with a 5-elimination.
Assume
\[
a_1 = a_2 = a_3 = a_4 = a_5 = 0.34,
\]
so that 5-elimination (Lemma~\ref{lem:nm}) reduces us to proving the following inequality for all Rademacher sums $X'$ with $\var(X')=1$, whose largest weight is $\leq 0.34/(1-5\cdot 0.34^2)^{1/2} < 0.6$:
\begin{equation}\label{eq:demo3}
\begin{gathered}
4\hpr{X'}{0}{1.01} + 5\hpr{X'}{0.04}{1.01} + \hpr{X'}{1.08}{2.06}
\\ \geq \\
\Pr[X' > 1.01] + 9\Pr[X' > 2.06] + 5\Pr[X' > 3.1] + \Pr[X' > 4.15].
\end{gathered}
\end{equation}
We prove~\eqref{eq:demo3} even with only the first term on the LHS. Using Chebyshev's inequality~\eqref{eq:cheby_our}:
\begin{equation}\label{eq:demo3-cheby}
3\hpr{X'}{0}{1} \geq 9\Pr[X' > 2] + 15\Pr[X' > 3].
\end{equation}
In order to deduce~\eqref{eq:demo3} from~(\ref{eq:demo3-cheby}), we have to show
\begin{equation}\label{eq:demo3-final}
\hpr{X'}{0}{1.01} \geq \Pr[X' > 1.01].
\end{equation}
This inequality is actually not easy to prove. However, it is implied by Tomaszewski's assertion~\eqref{eq:main} for the variable $X'$ which depends on $n-5$ weights (where $n$ is the number of weights in the original Rademacher sum $X$ we discuss)! We comment that in general we should use~\eqref{eq:dzin-intro1} instead of~\eqref{eq:main} to prove~\eqref{eq:demo3-final}, which is possible as~\eqref{eq:dzin-intro1} for $X'$ is implied from Tomaszewski's assertion~\eqref{eq:main} on $n-4$ weights.
\section{Open Problems}\label{sec:open}
As was mentioned in the introduction, we believe the methods developed in this paper can be applied to obtain further results on the distribution of Rademacher sums. We conclude the paper with several related open questions.
\paragraph{Tail bounds for Rademacher sums.} Consider the following general problem.
\newtheorem{problem}[theorem]{Problem}
\begin{problem}
Let $\mathcal{X}$ be the class of all Rademacher sums with variance $1$. Characterize the following function, defined for all $x \in \mathbb{R}$:
\[
F(x) = \sup_{X \in \mathcal{X}} \Pr[X > x].
\]
\end{problem}
\noindent While different sub-cases of this problem were studied in many papers (see, e.g.,~\cite{Dzindzalieta14,Pinelis12} and the numerous references therein),
only a few exact results on it are known (e.g.,~\cite{BD15,Pinelis12}). Theorem~\ref{thm:main} continues the series of exact results, showing that $F(x) = 1/4$ for all $x \in [1, \sqrt{2})$.
A well-known conjecture, due to Hitczenko and Kwapie\'{n}~\cite{HK94}, concerns $F(x)$ for $x = -1$.
\begin{conjecture}[\cite{HK94}]\label{conj:olesz}
Let $X = \sum a_i x_i$ be a Rademacher sum with $\var(X) = 1$. Then
\[
\Pr[X \geq 1] \geq \frac{7}{64}.
\]
\end{conjecture}
Conjecture~\ref{conj:olesz} is an evident counterpart of Tomaszewski's conjecture -- while the latter states that $\Pr[X > 1]$ must be somewhat small, the former declares that $\Pr[X \geq 1]$ must be somewhat large.
The best currently known result toward Conjecture~\ref{conj:olesz} is $\Pr[X > 1] \geq 1/20$ (whenever $X \not \equiv x_1$), proved by Oleszkiewicz~\cite{Oles96} more than 20 years ago. Using our methods (specifically, Proposition~\ref{prop:Prawitz-Rademacher} and Lemma~\ref{lem:nm} with $m \leq 3$) and additional tools, Dvo{\v{r}}{\'{a}}k and the second author~\cite{DK21} proved the stronger bound $\Pr[X \geq 1] \geq 6/64$, along with the sharp bound $\Pr[X > 1] \geq \frac{1}{16}$.
\paragraph{Improved Berry-Esseen type bounds for Rademacher sums.}
Proposition~\ref{prop:ad-hoc2} shows that for a Rademacher sum $X=\sum a_i x_i$ with $\var(X) = 1$ and $0 \leq a_i\leq a_1=0.22$, and for any $x \geq 0$,
\[
\Pr[X \leq x] \geq \Pr[Z \leq x] - \Pr[Z \in \langle 0, a_1]],
\]
where $Z$ is a standard Gaussian. The following conjecture is a natural extension:
\begin{conjecture}
Let $X = \sum a_i x_i$ be a Rademacher sum with $\var(X) = 1$ and $\forall i \colon 0 < a_i \leq a_1$. Let $Z \sim N(0, 1)$ be a standard Gaussian, and $x \in \mathbb{R}$. Then
\begin{equation}\label{eq:be_conj}
|\Pr[X \leq x] - \Pr[Z \leq x]| \leq \Pr[Z \in (0, a_1)] < \frac{1}{\sqrt{2\pi}} a_1.
\end{equation}
\end{conjecture}
Note that first inequality in~\eqref{eq:be_conj} is tight, as is demonstrated by $X = \sum_{i=1}^{n} x_i / \sqrt{n}$ for $n$ odd, and $x=1/\sqrt{n}-\epsilon$
We note that it follows from~\cite[Theorem~1.3 and Remark~1.4(c)]{MS18} that for $X = \sum_{i=1}^{n} x_i / \sqrt{n}$, this inequality holds for any $n,x$.
It appears that our methods can be used to prove the conjecture in part of the range, namely,
$\Pr[Z \leq x] - \Pr[X \leq x] \leq \Pr[Z \in (0, a_1)]$ for all $x \geq 0$, in a way similar to the proof of Proposition~\ref{prop:ad-hoc2}.
\section*{Acknowledgements}
We thank Tom Kalvari, Jiange Li, Lunz Mattner, and Jeremy Schiff, and especially Ron Holzman, for inspiring discussions and useful suggestions.
|
\section{Introduction}
\label{sec:introduction}
We explore mirror symmetry for smoothings of a $3$-dimensional
Gorenstein toric affine variety $V$. Specifically, we try to imagine
what consequences mirror symmetry may have for the classification of
smoothing components of the deformation space $\Defo V$. Conjecture~A
makes the surprising statement that the set of smoothing components of
$\Defo V$ is in bijective correspondence with a set of easily defined
and enumerated 2-variable Laurent polynomials, called $0$-mutable
polynomials. Our Conjecture~B --- in the strong form stated in
Remark~\ref{rem:conjBStrongForm} --- asserts that these smoothing
components are themselves smooth, and computes their tangent spaces
from the corresponding $0$-mutable polynomials.
\medskip
As is customary in toric geometry, $V$ is associated to a strictly
convex 3-dimensional rational polyhedral cone
$\sigma \subseteq N_\mathbb{R}$, where $N$ is a $3$-dimensional lattice; the
Gorenstein condition means that the integral generators of the rays of
$\sigma$ all lie on an integral affine hyperplane $(\usp = 1)$ for
some $\usp \in M := \Hom_\mathbb{Z}(N, \mathbb{Z})$. We denote by $F$ the convex
hull of the integral generators of the rays of $\sigma$, i.e.\
\[
F := \sigma \cap (\usp=1);
\]
this is a lattice polygon (i.e.\ a lattice polytope of dimension $2$)
embedded in the affine $2$-dimensional lattice $(u=1)$. The
isomorphism class of the toric variety $V$ depends only on the
affine equivalence class of the polygon $F$.
If $V$ has an isolated singularity, then $\Defo V$ is finite dimensional and
we know from the work of
Altmann~\cite{altmann_versal_deformation} that there is a $1$-to-$1$
correspondence between the set of irreducible components of $\Defo V$
and integral maximal Minkowski decompositions of $F$.
Altmann also shows that, when taken with their
\emph{reduced} structure, these components are all themselves smooth.
We are interested in the case when $V$ has non-isolated
singularities. Very little is known at this level of generality, but
examples show that the picture for non-isolated singularities is very
different from the one just sketched for isolated singularities. Our
main reason for wanting to work with non-isolated singularities is the
Fanosearch project: we wish to prove a general criterion for smoothing
a toric Fano variety, and Conjecture~A here is just the local case.
Conjecture~A characterizes \emph{smoothing components} of $\Defo V$ in
terms of the combinatorics of the polygon $F$. Specifically, we define
the set $\mathfrak{B}$ of $0$-mutable Laurent polynomials with Newton polygon $F$,
and the conjecture states that there is a canonical bijective
correspondence $\kappa \colon \mathfrak{B}\to \mathfrak{A}$, where
$\mathfrak{A}$ is the set of smoothing components of $\Defo V$.
At first sight the formulation of the conjecture seems strange;
however, the statement makes sense in the context of mirror symmetry,
where (conjecturally) the $0$-mutable polynomials are the mirrors of
the corresponding smoothing components.
In Section~\ref{sec:intr-mirr-symm}, we state a new
Conjecture~B,\footnote{The statement of Conjecture~B comes after
Sec.~\ref{sec:properties-0-mutable} but does not logically depend on
it: if you wish, you can skip directly from
Sec.~\ref{sec:def-of-0-mutable} to Sec.~\ref{sec:conjectures}.}
which implies the existence of a map
$\kappa \colon \mathfrak{B}\to \mathfrak{A}$ --- see
Remark~\ref{rem:conjBimpliesmapdefined}. In that section, we also explain how to
(conjecturally) construct a deformation directly from a $0$-mutable
polynomial in the spirit of the intrinsic mirror symmetry of
Gross--Siebert~\cite{1909.07649, gs_intrinsic_slc} and work of
Gross--Hacking--Keel~\cite{MR3415066}.
The coefficients of the
$0$-mutable Laurent polynomials that appear in our conjecture ought
themselves to enumerate certain holomorphic discs in the corresponding
smoothing, and we would love to see a precise statement along these
lines.
In our view, the conjectures together are nothing other than a
statement of mirror symmetry as a one-to-one correspondence between
two sets of objects, similar to the conjectures made in
\cite{MR3430830} in the context of orbifold del Pezzo surfaces, and
the correspondence between Fano 3-folds and Minkowski polynomials
discovered in \cite{MR3470714}.
In Section~\ref{sec:properties-0-mutable} we give some equivalent
characterisations of $0$-mutable polynomials and begin to sketch some
of their general properties. These properties make it very easy to
enumerate the $0$-mutable polynomials with given Newton polygon.
The material here is rather sketchy --- full details will appear elsewhere;
it serves for context, but it is not logically necessary for the statement
of the conjectures.
The suggestion that there is a simple structure to the set of
smoothing components is surprising in a subject that --- as all serious
practitioners know --- is marred by Murphy's law. In fact, there is a
substantial body of direct and circumstantial evidence for the
conjectures, some of which we present in
Section~\ref{sec:evidence-1}.
In the final Section~\ref{sec:evidence-matej} we compute in detail the
deformation space of the variety $V_F$ associated to the polygon $F$ of
Example~\ref{ex:quadrilateral}, giving evidence for the
conjectures. Some of the reasons for choosing this particular example
are:
\begin{enumerate}
\item The variety $V_F$ is of codimension $5$ and hence it lies outside the
--- still rudimentary but very useful --- structure theory of codimension-$4$
Gorenstein rings \cite{MR3380790}, see also \cite{1812.02594, tom_and_jerry};
\item For this reason, $V_F$ is a good test of the technology of
\cite{fil, fil2} as a tool for possibly proving the conjectures;
\item The polygon $F$ appears as a facet of some of the
$3$-dimensional reflexive polytopes and hence it is immediately
relevant for the Fanosearch project.
\end{enumerate}
As things stand, we are some distance away from being able to prove
the conjectures. We had a tough time even with the example of
Section~\ref{sec:evidence-matej}: while a treatment based on
\cite{fil, fil2} seems possible, the task became so tedious that we
decided instead to rely on Ilten's Macaulay2 \cite{M2} package
\emph{Versal deformations and local Hilbert schemes}
\cite{ilten_package_article}. That package makes it possible to test
the conjectures in many other examples in codimension $\geq 5$.
\subsection*{Notation and conventions}
We work over $\mathbb{C}$, but everything holds over an algebraically closed field of characteristic zero.
We refer the reader to \cite{fulton_book} for an introduction to toric geometry.
All the toric varieties we consider are normal.
We use the following notation.
\begin{longtable}{lp{0.85\textwidth}}
$F$ & a lattice polygon \\
$V$ & the Gorenstein toric affine \mbox{$3$-fold} associated to the cone over $F$ put at height 1\\
$\divV$ & the toric boundary of $V$ \\
$\overline{X}$ & the projective toric surface associated to the normal fan of $F$\\
$\overline{B}$ & the toric boundary of $\overline{X}$ \\
$A$ & the ample line bundle on $\overline{X}$ given by $F$ \\
$X$ & the cluster surface associated to $F$ (the non-toric blowup of
$\overline{X}$ constructed in Section~\ref{sec:properties-0-mutable})\\
$B$ & the strict transform of $\overline{B}$ in $X$ \\
$W^\flat$ & the toric \mbox{$3$-fold} constructed in Section~\ref{sec:intr-mirr-symm}\\
$\overline{W}$ & the toric blowup of $W^\flat$ constructed in Section~\ref{sec:intr-mirr-symm}\\
$W$ & the mirror cluster variety (the non-toric blowup of $\overline{W}$
constructed in Section~\ref{sec:intr-mirr-symm})
\end{longtable}
\subsection*{Thanks to Bill from AC} I was lucky to be a L.E.\ Dickson
Instructor at the University of Chicago in the years 1993--1996, where
I worked in the group led by Bill Fulton. I had studied deformations
of singularities in a seminar held in 1988--89 at the University of
Utah, but in Bill's seminar I learnt about Chow groups and quantum
cohomology. It was a wonderful time in my work life thanks largely to
Bill. This paper features many of the ideas that I learnt in Bill's
seminar and it is very nice to see that they are so relevant in
the study of deformations of singularities.
\subsection*{Acknowledgements} We thank Klaus Altmann, Tom Coates,
Mark Gross, Paul Hacking, Al Kasprzyk, Giuseppe Pitton, and Thomas
Prince for many helpful conversations.
We particularly thank Al Kasprzyk for sharing with us and allowing us
to present some of his unpublished ideas on maximally mutable Laurent
polynomials \cite{KT}.
We discussed with Mark Gross some of our early experiments with
$0$-mutable polynomials.
We owe special thanks to Paul Hacking who read and corrected various
mistakes in earlier versions of the paper --- the responsibility for
the mistakes that are left is of course ours.
Giuseppe Pitton ran computer calculations that provide indirect
evidence for the conjectures.
The idea that mirror Laurent polynomials are characterised by their
mutability goes back to Sergey Galkin \cite{galkin_usnich}.
It will be clear to all those familiar with the issues that this paper
owes a very significant intellectual debt to the work of Gross,
Hacking and Keel \cite{MR3415066} and the intrinsic mirror symmetry of
Gross and Siebert \cite{1909.07649, gs_intrinsic_slc}.
Last but not least, it is a pleasure to thank the anonymous referee
who read our manuscript very carefully, and told us about Ilten's
Macaulay2 package.
\section{Conjecture~A}
\label{sec:main}
\subsection{Gorenstein toric affine varieties}
Consider a rank-$n$ lattice $N\simeq \mathbb{Z}^n$ (usually $n=3$) and, as
usual in toric geometry, its dual lattice $M=\Hom_\mathbb{Z} (N,\mathbb{Z})$. The
$n$-dimensional torus
\[
\mathbb{T}= \Spec \mathbb{C} [M]
\]
is referred to simply as ``the'' torus.\footnote{Sometimes we denote
this torus by $\mathbb{T}_N$, that is, the commutative group scheme
$N\otimes_\mathbb{Z} \Gm$ such that for all rings $R$
$\mathbb{T}_N(R)=N\otimes_{\mathbb{Z}} R^\times$.} Consider a strictly convex
full-dimensional rational polyhedral cone $\sigma \subset N_\mathbb{R}$ and
the corresponding affine toric variety
\[
V = \Spec \mathbb{C}[ \sigma^\vee \cap M].
\]
This is a normal Cohen--Macaulay $n$-dimensional variety.
By definition $V$ is Gorenstein if and only if the pre-dualising sheaf
\[
\omega^0_V = \mathcal{H}^{-n} (\omega_V^\bullet)
\]
is a line bundle. Since our $V$ is Cohen--Macaulay, this is the same as insisting
that all the local rings of $V$ are local Gorenstein rings.
It is known and not difficult to show that $V$ is Gorenstein if and only if there is a
vector $\usp \in M$ such that the integral generators
$\rho_1, \dots, \rho_m$ of the rays of the cone $\sigma$ all lie on the affine
lattice $\mathbb{L}=(\usp = 1)\subset N$. Such vector $\usp$ is called the \emph{Gorenstein degree}.
If $V$ is Gorenstein, then the toric boundary of $V$ is the following effective reduced
Cartier divisor on $V$:
\[
\divV = \Spec \mathbb{C}[\sigma^\vee \cap M]/(x^\usp).
\]
If $V$ is Gorenstein, we set
\[
F := \sigma \cap (\usp=1);
\]
this is an $(n-1)$-dimensional lattice polytope embedded in the affine
lattice $\mathbb{L} = (u=1)$ and it is the convex hull of the integral
generators of the rays of the cone $\sigma$. One can prove that the
isomorphism class of $V$ depends only on the affine equivalence class
of the lattice polytope $F$ in the affine lattice $\mathbb{L}$. Therefore we
will say that $V$ is associated to the polytope $F$.
This establishes a $1$-to-$1$ correspondence between isomorphism
classes of Gorenstein toric affine \mbox{$n$-folds} without torus
factors and $(n-1)$-dimensional lattice polytopes up to affine
equivalence.
\subsection{Statement of Conjecture~A}
\label{sec:statement-A}
In this section we explain everything that is needed to make sense of
the following:
\begin{conA}
Consider a lattice polygon $F$ in a $2$-dimensional affine lattice $\mathbb{L}$.
Let $V$ be the Gorenstein toric affine \mbox{$3$-fold} associated to $F$.
Then there is a canonical bijective function $\kappa \colon \mathfrak{B}\to \mathfrak{A}$ where:
\begin{itemize}
\item $\mathfrak{A}$ is the set of smoothing components of the miniversal deformation space
$\Defo V$,
\item $\mathfrak{B}$ is the set of $0$-mutable polynomials $f\in \mathbb{C}[\mathbb{L}]$ with Newton
polygon $F$.
\end{itemize}
\end{conA}
\begin{rem}
\label{rem:infinite_dimensional}
It is absolutely crucial to appreciate that we are not assuming that
$V$ has an isolated singularity at the toric $0$-stratum. If $V$
does not have isolated singularities, $\Defo V$ is
infinite-dimensional. A few words are in order to clarify what kind
of infinite dimensional space $\Defo V$ is.
In full generality, there is some discussion of this issue in the
literature on the analytic category, see for example~\cite{MR713093,
MR803194}.\footnote{We thank Jan Stevens for pointing out these
references to us. We are not aware of a similar discussion in the
algebraic literature.} In the special situation of interest in
this paper, we take a na\"{\i}ve approach, which we briefly explain,
based on the following two key facts:
\begin{enumerate}[(i)]
\item If $V$ is a Gorenstein toric affine 3-fold, then $T^2_V$ is
finite dimensional. Indeed $V$ has transverse
$A_\star$-singularities in codimension two, hence it is
unobstructed in codimension two, hence $T^2_V$ is a finite length
module supported on the toric $0$-stratum. In fact, there is an
explicit description of $T^2_V$ as a representation of the torus,
see \cite[Section~5]{altmann_andre_quillen_cohomology}, an example
of which is in Lemma~\ref{lem t2} below. This shows that $\Defo V$
is cut out by finitely many equations.
\item On the other hand, the known explicit description of $T^1_V$
as a representation of the torus
\cite[Theorem~4.4]{altmann_one_parameter}, together with the
explicit description of $T^2_V$ just mentioned, easily implies
that each of the equations can only use finitely many
variables.\footnote{The key point is to show that for all fixed
weights $\mathbf{m}\in M$, the equation
$\mathbf{m}=\sum m_i \mathbf{v}_i$ has finitely many solutions
for $m_i\in \mathbb{N}\setminus \{0\}$ and $\mathbf{v}_i\in M$ a weight
that appears non-trivially in $T^1_V$. This type of
consideration is used extensively in the detailed example
discussed in Section~\ref{sec:evidence-matej}.}
\end{enumerate}
Thus we can take $\Defo V$ to be the $\Spec$ of a non-Noetherian
ring, that is, the simplest kind of infinite-dimensional
scheme.\footnote{The situation is not so simple for the universal
family $\mathscr{U}\to \Defo V$. Indeed the equations of $\mathscr{U}$ naturally
involve all the infinitely many coordinate functions on
$T^1_V$. Thus, $\mathscr{U}$ is a bona fide ind-scheme. The language to
deal with this exists but it is not our concern here.}
\end{rem}
\begin{rem}
\label{rem:1}
Conjecture~A does not state what the function $\kappa$ is, nor what
makes it ``canonical.'' The existence of a function $\kappa \colon
\mathfrak{B}\to \mathfrak{A}$ is implied by Conjecture~B stated in
Section~\ref{sec:conjectures} below, see Remark~\ref{rem:conjBimpliesmapdefined}.
\end{rem}
\begin{rem} \label{rem:deformations_variety_or_pair}
Let $V$ be a Gorenstein toric affine \mbox{$3$-fold} and let $\divV$
be the toric boundary of $V$. Let $\Defo (V, \divV)$ be the
deformation functor (or the base space of the miniversal
deformation) of the pair $(V, \divV)$. There is an obvious
forgetful map $\Defo (V, \divV) \to \Defo V$. In this case, since
$\divV$ is an effective Cartier divisor in $V$ and $V$ is affine,
this map is smooth of relative dimension equal to the dimension of
$\coker \left( H^0(\theta_V) \to H^0(N_{\divV / V}) \right)$, where $\theta_V$ is
the sheaf of derivations on $V$ and $N_{\divV / V} = \mathcal{O}_{\partial V}(\partial V)$ is the normal
bundle of $\divV$ inside $V$. In particular, this implies that
$\Defo V$ and $\Defo (V, \divV)$ have exactly the same irreducible
components. One can also see that a smoothing component in
$\Defo V$ is the image of a component of $\Defo (V, \divV)$ where
also $\divV$ is smoothed.
In other words, we can equivalently work with deformations of $V$
or deformations of the pair $(V,\divV)$. The right thing to do in
mirror symmetry is to work with deformations of the pair $(V,\divV)$; however,
the literature on deformations of singularities is all written in
terms of $V$. In most cases it is not difficult to make the translation but this
paper is not the right place for doing that. Thus when possible we
work with deformations of $V$. The
formulation of Conjecture~B in Section~\ref{sec:intr-mirr-symm}
requires that we work with deformations of the pair $(V, \divV)$.
\end{rem}
\subsection{The definition of $0$-mutable Laurent polynomials}
\label{sec:def-of-0-mutable}
This subsection is occupied by the definition of $0$-mutable Laurent
polynomials. The simple key idea --- and the explanation for the name
``$0$-mutable'' --- is that an irreducible Laurent polynomial $f$ is
$0$-mutable if and only if there is a sequence of mutations
\[
f\mapsto f_1\mapsto \cdots \mapsto f_p=\mathbf{1}
\]
starting from $f$ and ending with the constant monomial
$\mathbf{1}$. The precise definition is given below after some
preliminaries on mutations. The definition is appealing and it is
meaningful in all dimensions, but it is not immediately useful if you
want to study $0$-mutable polynomials. Indeed, for example, to prove
that a given polynomial $f$ is $0$-mutable one must produce a chain of
mutations as above and it may not be obvious where to look. It is even
less clear how to prove that $f$ is not $0$-mutable. In
Section~\ref{sec:properties-0-mutable},
Theorem~\ref{thm:equivalent_defns}, we prove two useful
characterizations of $0$-mutable polynomials in two variables. The
first of these states that a polynomial is $0$-mutable if and only if
it is rigid maximally mutable. From this property it is easy to check
that a given polynomial is $0$-mutable, that it is not $0$-mutable,
and to enumerate $0$-mutable polynomials with given Newton
polytope. The second characterization states that a (normalized, see
below) Laurent polynomial in two variables is $0$-mutable if and only
if the irreducible components of its vanishing locus are $-2$-curves
on the cluster surface: see Section~\ref{sec:properties-0-mutable} for
explanations and details.
\medskip
Let $\mathbb{L}$ be an affine lattice and let $\mathbb{L}_0$ be its underlying
lattice.\footnote{In our setup $\mathbb{L}=(\usp =1)\subset N$ does not have
a canonical origin. We try to be pedantic and write $\mathbb{L}_0$ for the
the underlying lattice -- in our example, $\mathbb{L}_0=\Ker \usp$. In
practice this distinction is not super-important and you are free to
choose an origin anywhere you want.}
In other words, $\mathbb{L}_0$ is a free abelian group of finite rank and
$\mathbb{L}$ is a set together with a free and transitive $\mathbb{L}_0$-action. We
denote by $\mathbb{C}[\mathbb{L}]$ the vector space over $\mathbb{C}$ whose basis is made
up of the elements of $\mathbb{L}$. For every $l \in \mathbb{L}$ we denote by $x^l$
the corresponding element in $\mathbb{C}[\mathbb{L}]$. Elements of $\mathbb{C}[\mathbb{L}]$ will
be called (Laurent) polynomials. It is clear that $\mathbb{C}[\mathbb{L}]$ is a
rank-$1$ free module over the $\mathbb{C}$-algebra $\mathbb{C}[\mathbb{L}_0]$. The choice
of an origin in $\mathbb{L}$ specifies an isomorphism
$\mathbb{C}[\mathbb{L}] \simeq \mathbb{C}[\mathbb{L}_0]$.
\begin{dfn}
\label{dfn:normalised} A Laurent polynomial $f\in \mathbb{C}[\mathbb{L}]$ is
\emph{normalized} if for all vertices $v\in \Newt f$ we have $a_v=1$
where $a_v$ is the coefficient of the monomial $x^v$ as it appears
in $f$.
\end{dfn}
In this paper all polynomials are assumed to be normalized unless
explicitly stated otherwise.
If $f\in \mathbb{C}[\mathbb{L}]$ then we say that $f$
\emph{lives} on the smallest saturated affine sub-lattice $\mathbb{L}^\prime\subseteq \mathbb{L}$ such
that $f\in \mathbb{C}[\mathbb{L}^\prime]$. The property of being $0$-mutable only
depends on the lattice where $f$ lives --- here it is crucial that we only allow saturated sub-lattices.
We say that $f$ is an
$r$-variable polynomial if $f$ lives on a rank-$r$ affine lattice.
Let us start by defining $0$-mutable polynomials in $1$ variable.
\begin{dfn} \label{dfn:0-mutable_1_variable} Let $\mathbb{L}$ be an affine
lattice of rank $1$ and let $\mathbb{L}_0$ be its underlying lattice. Let
$v \in \mathbb{L}_0$ be one of the two generators of $\mathbb{L}_0$. A
polynomial $f \in \mathbb{C}[\mathbb{L}]$ is called \emph{$0$-mutable} if
\[
f = (1+x^v)^k x^l
\]
for some $l \in \mathbb{L}$ and $k \in \mathbb{N}$.
\end{dfn}
(It is clear that the definition does not depend on the choice of $v$.)
If $f$ is a Laurent polynomial in 1 variable and its Newton polytope
is a segment of lattice length $k$, then $f$ is $0$-mutable if and only if
the coefficients of $f$ are the $k+1$ binomial coefficients of weight $k$.
\medskip
The definition of $0$-mutable polynomials in more than 1 variable will
be given recursively on the number of variables. Thus from now on we
fix $r \geq 2$ and we assume to know already what it means for a
polynomial of $<r$ variables to be $0$-mutable. Before we can state
what it means for a polynomial $f$ of $r$ variables to be $0$-mutable,
we need to explain how to mutate $f$.
If $\mathbb{L}$ is an affine lattice, we denote by $\Aff (\mathbb{L}, \mathbb{Z})$ the
lattice of affine-linear functions $\varphi \colon \mathbb{L} \to \mathbb{Z}$.
If $\mathbb{L}_0$ denotes the underlying lattice of $\mathbb{L}$, $\varphi$ has a
well-defined \emph{linear part} which we denote by $\varphi_0\colon \mathbb{L}_0\to \mathbb{Z}$.
\begin{dfn}
\label{dfn:mutation}
Let $r\geq 2$ and fix a rank-$r$ affine lattice $\mathbb{L}$.
A \emph{mutation datum} is a pair $(\varphi, h)$ of a non-constant affine-linear
function $\varphi \colon \mathbb{L} \to \mathbb{Z}$ and a $0$-mutable polynomial
$h\in \mathbb{C}[\Ker \varphi_0]$.
Given a mutation datum $(\varphi,h)$ and $f\in \mathbb{C}[\mathbb{L}]$, write (uniquely)
\[
f=\sum_{k\in \mathbb{Z}} f_k
\qquad
\text{where}
\qquad
f_k \in \mathbb{C}[(\varphi =k) \cap \mathbb{L}]
\]
We say that $f$ is \emph{$(\varphi,h)$-mutable} if for all $k<0$
$h^{-k}$ divides $f_k$ (equivalently, if for all $k \in \mathbb{Z}$
$h^kf_k\in \mathbb{C}[\mathbb{L}]$). If $f$ is $(\varphi,h)$-mutable, then the
\emph{mutation} of $f$, with respect to the mutation datum
$(\varphi,h)$, is the polynomial:
\[
\mut_{(\varphi, h)} f =\sum_{k\in \mathbb{Z}} h^k f_k.
\]
\end{dfn}
\begin{rem}
The notion of mutation goes back (at least) to
Fomin--Zelevinsky~\cite{fomin_zelevinsky}. We first learned of
mutations from the work of Galkin--Usnich~\cite{galkin_usnich}
and~Akhtar--Coates--Galkin--Kasprzyk~\cite{sigma}. This paper owes a
significant intellectual debt to the interpretation of mutations
developed in work by Gross, Hacking and Keel, for instance
\cite{MR3314827, MR3350154, MR3415066}.\footnote{Many mathematicians
work on mutations from different perspectives and we apologize for
not even trying to quote all the relevant references here.}
\end{rem}
The following is a recursive definition. The base step is given by
Definition~\ref{dfn:0-mutable_1_variable}.
\begin{dfn}
\label{dfn:0-mutable}
Let $\mathbb{L}$ be an affine lattice of rank $r \geq 2$. We define the set of $0$-mutable polynomials on $\mathbb{L}$
in the following recursive way.
\begin{enumerate}[(i)]
\item If $\mathbb{L}^\prime$ is a saturated affine sub-lattice of $\mathbb{L}$
and $f \in \mathbb{C}[\mathbb{L}^\prime]$ is $0$-mutable, then $f$ is
$0$-mutable in $\mathbb{C}[\mathbb{L}]$.
\item If $f=f_1 f_2$ is reducible, then $f$ is 0-mutable if both
factors $f_1$, $f_2$ are $0$-mutable.\footnote{In order to take the ``product'' $f_1f_2$
one has to choose an origin in $\mathbb{L}$. This choice makes no
difference to the definition.}
\item If $f$ is irreducible, then $f$ is $0$-mutable if a mutation of $f$ is $0$-mutable.
\end{enumerate}
\end{dfn}
Equivalently, the set of $0$-mutable polynomials of $\leq r$
variables is the smallest subset of $\mathbb{C}[\mathbb{L}]$ that contains all $0$-mutable
polynomials of $<r$ variables and that is closed under the operations
of taking products (in particular translation) and mutations of
irreducible polynomials.
\begin{rem}
\label{rem:2}
It follows easily from the definition that $0$-mutable polynomials
are normalized. Indeed:
\begin{enumerate}[(1)]
\item The polynomials in Definition~\ref{dfn:0-mutable_1_variable}
(the base case) are $0$-mutable;
\item The product of two normalized polynomials is normalized;
\item The mutation of a normalized polynomial is normalized.
\end{enumerate}
\end{rem}
\begin{exa}
Let $\mathbb{L}$ be an affine lattice and let $v$ be a primitive vector in
the underlying lattice $\mathbb{L}_0$.
Definition~\ref{dfn:0-mutable_1_variable} and
Definition~\ref{dfn:0-mutable}(i) imply that $(1+x^v)^k x^l$ is
$0$-mutable for all $l \in \mathbb{L}$ and $k \in \mathbb{N}$.
\end{exa}
\begin{exa} \label{exa:conePP123}
Consider the triangle
\begin{equation} \label{eq:triangle_123}
F = \hull{(0,0),(3,0),(3,2)}
\end{equation}
in the lattice $\mathbb{L} = \mathbb{Z}^2$. Let us identify $\mathbb{C}[\mathbb{L}]$ with
$\mathbb{C}[x^\pm, y^\pm]$. One can prove that there are exactly two
$0$-mutable polynomials with Newton polytope $F$, namely:
\begin{align*}
(1+x)^3 + 2(1+x)x^2y + x^3y
\qquad \text{and} \qquad
(1+y)^2 x^3 + 3 (1+y)x^2 + 3x + 1.
\end{align*}
In Figure~\ref{fig:123} we have written the coefficients of these two polynomials
next to the lattice point of $F$ associated to the corresponding
monomial.
It is shown in \cite{1812.02594} that $\Defo V$ has two components, and
that they are both smoothing components, confirming our
conjectures. The calculation there goes back to unpublished work by Jan
Stevens, but see also~\cite{altmann_one_parameter}.
\end{exa}
\begin{figure}
\centering
\begin{tikzpicture}
\foreach \x in {0, ..., 3}
\foreach \y in {0, ..., 0}
{
\node[thickstyle] (\x - \y) at (\x, \y){};
}
\node[thickstyle] (3 - 1) at (3, 1){};
\node[thickstyle] (3 - 2) at (3, 2){};
\node[thickstyle] (2 - 1) at (2, 1){};
\node[thinstyle] (0 - 1) at (0, 1){};
\node[thinstyle] (1 - 1) at (1, 1){};
\node[thinstyle] (0 - 2) at (0, 2){};
\node[thinstyle] (1 - 2) at (1, 2){};
\node[thinstyle] (2 - 2) at (2, 2){};
\draw[thick] (0, 0) -- (3, 0) -- (3, 2) -- cycle;
\node [anchor=east] at (0.25, 0.25) {$1$};
\node [anchor=east] at (1.5, 0.25) {$3$};
\node [anchor=east] at (2.5, 0.25) {$3$};
\node [anchor=east] at (3.5, 0.25) {$1$};
\node [anchor=east] at (2.5, 1.25) {$2$};
\node [anchor=east] at (3.5, 1.25) {$2$};
\node [anchor=east] at (3.5, 2.25) {$1$};
\end{tikzpicture}
\qquad \qquad
\begin{tikzpicture}
\foreach \x in {0, ..., 3}
\foreach \y in {0, ..., 0}
{
\node[thickstyle] (\x - \y) at (\x, \y){};
}
\node[thickstyle] (3 - 1) at (3, 1){};
\node[thickstyle] (3 - 2) at (3, 2){};
\node[thickstyle] (2 - 1) at (2, 1){};
\node[thinstyle] (0 - 1) at (0, 1){};
\node[thinstyle] (1 - 1) at (1, 1){};
\node[thinstyle] (0 - 2) at (0, 2){};
\node[thinstyle] (1 - 2) at (1, 2){};
\node[thinstyle] (2 - 2) at (2, 2){};
\draw[thick] (0, 0) -- (3, 0) -- (3, 2) -- cycle;
\node [anchor=east] at (0.25, 0.25) {$1$};
\node [anchor=east] at (1.5, 0.25) {$3$};
\node [anchor=east] at (2.5, 0.25) {$3$};
\node [anchor=east] at (3.5, 0.25) {$1$};
\node [anchor=east] at (2.5, 1.25) {$3$};
\node [anchor=east] at (3.5, 1.25) {$2$};
\node [anchor=east] at (3.5, 2.25) {$1$};
\end{tikzpicture}
\caption{The two $0$-mutable polynomials whose Newton polytope is the triangle $F$ defined in \eqref{eq:triangle_123}}
\label{fig:123}
\end{figure}
\begin{exa}\label{ex:quadrilateral}
Consider the quadrilateral
\begin{equation} \label{eq:quadrilateral}
F = \hull{(-1,-1),(2,-1),(1,1),(-1,2)}
\end{equation}
in the lattice $\mathbb{L} = \mathbb{Z}^2$. Let us identify $\mathbb{C}[\mathbb{L}]$ with $\mathbb{C}[x^\pm, y^\pm]$.
Consider the polynomial
\[
g = \frac{(1+x)^3 + (1+y)^3 - 1 + x^2 y^2 }{xy},
\]
which is obtained by giving binomial coefficients to the lattice
points of the boundary of $F$ and by giving zero coefficient to the
interior lattice points of $F$. By Lemma~\ref{lem:easy_properties}(2)
every $0$-mutable polynomial with Newton polytope $F$ must coincide
with $g$ on the boundary lattice points of $F$. One can prove that
there are exactly three $0$-mutable polynomials with Newton polytope
$F$, namely:
\begin{gather*}
\alpha = g + 5 + 2x + 2y = \frac{(1+x+2y+y^2)(1+2x+x^2+y)}{xy}, \\
\beta = g + 6 + 3x + 4y = \frac{(1+x)^3 + 3y(1+x)^2 + y^2(1+x)(3+x) + y^3}{xy}, \\
\gamma = g + 6 + 4x + 3y = \frac{(1+y)^3 + 3 x (1+y)^2 + x^2 (1+y)(3+y) + x^3}{xy}.
\end{gather*}
In Figure~\ref{fig:114} we have written down the coefficients of these
three polynomials.
The polynomial $\alpha$ is reducible and it is easy to show that its factors are $0$-mutable.
\begin{figure}
\centering
\begin{tikzpicture}
\foreach \x in {0,1,2}
\foreach \y in {0,1,2}
{
\node[thickstyle](\x - \y) at (\x, \y){};
}
\node[thickstyle](3 - 0) at (3, 0){};
\node[thickstyle](0 - 3) at (0, 3){};
\draw[thick] (0, 0) -- (3, 0) -- (2, 2) -- (0, 3) -- cycle;
\node [anchor=north] at (0, 0) {$1$};
\node [anchor=north] at (1,0) {$3$};
\node [anchor=north] at (2,0) {$3$};
\node [anchor=north] at (3,0) {$1$};
\node[anchor=east] at (0,1) {$3$};
\node[anchor=east] at (0,2) {$3$};
\node[anchor=east] at (0,3) {$1$};
\node[anchor=south] at (2.2,1.9) {$1$};
\node[anchor=north] at (1,1) {$5$};
\node[anchor=north] at (2,1) {$2$};
\node[anchor=north] at (1,2) {$2$};
\node[anchor=north] at (3,3) {$\alpha$};
\end{tikzpicture}
\hfill
\begin{tikzpicture}
\foreach \x in {0,1,2}
\foreach \y in {0,1,2}
{
\node[thickstyle](\x - \y) at (\x, \y){};
}
\node[thickstyle](3 - 0) at (3, 0){};
\node[thickstyle](0 - 3) at (0, 3){};
\draw[thick] (0, 0) -- (3, 0) -- (2, 2) -- (0, 3) -- cycle;
\node [anchor=north] at (0, 0) {$1$};
\node [anchor=north] at (1,0) {$3$};
\node [anchor=north] at (2,0) {$3$};
\node [anchor=north] at (3,0) {$1$};
\node[anchor=east] at (0,1) {$3$};
\node[anchor=east] at (0,2) {$3$};
\node[anchor=east] at (0,3) {$1$};
\node[anchor=south] at (2.2,1.9) {$1$};
\node[anchor=north] at (1,1) {$6$};
\node[anchor=north] at (2,1) {$3$};
\node[anchor=north] at (1,2) {$4$};
\node[anchor=north] at (3,3) {$\beta$};
\end{tikzpicture}
\hfill
\begin{tikzpicture}
\foreach \x in {0,1,2}
\foreach \y in {0,1,2}
{
\node[thickstyle](\x - \y) at (\x, \y){};
}
\node[thickstyle](3 - 0) at (3, 0){};
\node[thickstyle](0 - 3) at (0, 3){};
\draw[thick] (0, 0) -- (3, 0) -- (2, 2) -- (0, 3) -- cycle;
\node [anchor=north] at (0, 0) {$1$};
\node [anchor=north] at (1,0) {$3$};
\node [anchor=north] at (2,0) {$3$};
\node [anchor=north] at (3,0) {$1$};
\node[anchor=east] at (0,1) {$3$};
\node[anchor=east] at (0,2) {$3$};
\node[anchor=east] at (0,3) {$1$};
\node[anchor=south] at (2.2,1.9) {$1$};
\node[anchor=north] at (1,1) {$6$};
\node[anchor=north] at (2,1) {$4$};
\node[anchor=north] at (1,2) {$3$};
\node[anchor=north] at (3,3) {$\gamma$};
\end{tikzpicture}
\caption{The three $0$-mutable polynomials $\alpha$, $\beta$ and $\gamma$ whose Newton polytope is the quadrilateral $F$ defined in \eqref{eq:quadrilateral}}
\label{fig:114}
\end{figure}
Let us consider the following affine-linear functions $\mathbb{L} = \mathbb{Z}^2 \to \mathbb{Z}$:
\begin{align*}
-m^1_{2,1} \colon (a,b) \mapsto b-1, \\
-m^1_{3,1} \colon (a,b) \mapsto b-2, \\
-m^1_{3,2} \colon (a,b) \mapsto 2b-1.
\end{align*}
The level sets of these three affine-linear functions are depicted in Figure~\ref{fig:level_sets}.
\begin{figure}
\centering
\def<desired width>{4cm}
\input{Level_setsx.tex}
\bigskip
\medskip
\caption{The level sets of the three affine-linear functions $\mathbb{L} = \mathbb{Z}^2 \to \mathbb{Z}$ considered in Example~\ref{ex:quadrilateral}}
\label{fig:level_sets}
\end{figure}
We now consider the mutation data $(-m^1_{2,1}, 1+x)$, $(-m^1_{3,2}, 1+x)$ and $(-m^1_{3,1}, 1+x)$.
The polynomial $\alpha$ is mutable with respect to $(-m^1_{2,1}, 1+x)$ and to $(-m^1_{3,2}, 1+x)$ and
\begin{gather*}
\mut_{(-m^1_{2,1}, 1+x)} \alpha = \frac{1+x}{xy} + \frac{3+2x}{x} + \frac{y(3 + 2x + x^2)}{x} + \frac{y^2 (1+x)}{x}, \\
\mut_{(-m^1_{3,2}, 1+x)} \alpha = \frac{1}{xy} + \frac{3+2x}{x} + \frac{y(3+5x + 3x^2 +x^3)}{x} + \frac{y^2 (1+x)^3}{x},
\end{gather*}
but $\alpha$ is not mutable with respect to $(-m^1_{3,1}, 1+x)$.
The polynomial $\beta$ is mutable with respect to all three mutation data $(-m^1_{2,1}, 1+x)$, $(-m^1_{3,2}, 1+x)$ and $(-m^1_{3,1}, 1+x)$, and the mutations are:
\begin{gather*}
\mut_{(-m^1_{2,1}, 1+x)} \beta = \frac{1+x}{xy} + \frac{3(1+x)}{x} + \frac{y(3 +x)(1+x)}{x} + \frac{y^2(1+x)}{x}, \\
\mut_{(-m^1_{3,1}, 1+x)} \beta = \frac{1}{xy} + \frac{3}{x} + \frac{y(3 + x)}{x} + \frac{y^2}{x} = \frac{(1+y)^3 + xy^2}{xy}, \\
\mut_{(-m^1_{3,2}, 1+x)} \beta = \frac{1}{xy} + \frac{3(1+x)}{x} + \frac{y(3 + x)(1+x)^2}{x} + \frac{y^2(1+x)^3}{x}.
\end{gather*}
The polynomial $\gamma$ is not mutable with respect to any of the mutation data $(-m^1_{2,1}, 1+x)$, $(-m^1_{3,2}, 1+x)$, $(-m^1_{3,1}, 1+x)$.
\end{exa}
\section{Properties of $0$-mutable polynomials}
\label{sec:properties-0-mutable}
\subsection{Some easy properties}
If $\mathbb{L}$ is an affine lattice, $f\in \mathbb{C}[\mathbb{L}]$ and $F=\Newt f$, then we write
\[
f=\sum_{l\in F\cap \mathbb{L}} a_l x^l \qquad \text{with} \qquad a_l\in \mathbb{C}.
\]
For every subset $A\subseteq \mathbb{L}_\mathbb{R}$, we write
\[
f \vert_A=\sum_{l \in A\cap \mathbb{L}} a_l x^l.
\]
\begin{lem}
\label{lem:easy_properties}
\begin{enumerate}[(1)]
\item (Non-negativity and integrality) If $f$ is $0$-mutable, then every coefficient of $f$ is a non-negative integer.
\item (Boundary terms) If $f\in \mathbb{C}[\mathbb{L}]$ is $0$-mutable and $F\leq \Newt f$ is a
face, then $f \vert_F$ is $0$-mutable.
\end{enumerate}
\end{lem}
\begin{proof}[Sketch of proof]
(1) is obvious due to the recursive definition of 0-mutable polynomials. Also (2) is easy because it is enough to observe that mutations and products behave well with respect to restriction to faces of the Newton polytope.
\end{proof}
\begin{rem}
A $0$-mutable polynomial may have a zero coefficient at a lattice point of its Newton polytope, e.g.\ $(1+x)(1+xy^2)$.
\end{rem}
\subsection{Rigid maximally mutable polynomials}
\label{sec:rigid-maxim-mutable}
From now on we focus on the two-variable case $r=2$. In what follows,
we give two equivalent characterizations of $0$-mutable polynomials,
one geometric in terms of the associated cluster variety and one
combinatorial in terms of rigid maximally mutable
polynomials.\footnote{The concept of rigid maximally mutable is due to
Al Kasprzyk \cite{KT}. We thank him for allowing us to include his
definition here.}
Let $\mathbb{L}$ be an affine lattice of rank $2$.
For a Laurent polynomial $f\in \mathbb{C}[\mathbb{L}]$, we set
\[
\mathscr{S}(f) = \Bigl\{ \text{mutation data $s=(\varphi, h)$} \ \Bigl\vert \ \text{$f$ is $s$-mutable}\Bigr\}.
\]
Conversely, if $\mathscr{S}$ is a set of mutation data, we denote by
\[
L(\mathscr{S})=\Bigl\{f \in \mathbb{C}[\mathbb{L}]\ \Big\vert\ \text{$\forall s \in \mathscr{S}$, $f$ is $s$-mutable}\Bigr\}
\]
the vector space of Laurent polynomials $f$ that are $s$-mutable for
all the mutation data $s\in \mathscr{S}$. For every polynomial $f \in \mathbb{C}[\mathbb{L}]$, it is clear that $f \in L(\mathscr{S}(f))$.
\begin{dfn}[Kasprzyk]
\label{dfn:rigidMMLP}
Let $\mathbb{L}$ be an affine lattice of rank $2$, and $f\in \mathbb{C}[\mathbb{L}]$.
\begin{enumerate}[(i)]
\item If $f=f_1f_2$ is the product of normalized polynomials $f_1$,
$f_2$, then $f$ is rigid maximally mutable if both factors
$f_1, f_2$ are rigid maximally mutable;
\item If $f$ is normalized and irreducible, then $f$ is
rigid maximally mutable if
\[
L \left( \mathscr{S}(f) \right) = \{\lambda f \mid \lambda \in \mathbb{C} \}.
\]
\end{enumerate}
\end{dfn}
\subsection{Cluster varieties}
\begin{dfn}
\label{dfn:CYpairANDcluster_var}
A \emph{Calabi--Yau (CY) pair} is a pair $(Y, \omega)$ of an
$n$-dimensional quasiprojective normal variety $Y$ and a degree $n$ rational
differential $\omega \in \Omega^n_{k(X)}$, such that
\[
D =-\divi_Y \omega \geq 0
\]
is an effective reduced Cartier divisor on $Y$.\footnote{Like most people,
we mostly work with the pair $(Y, D)$ and omit explicit reference
to $\omega$.}
A \emph{torus chart} on $(Y,\omega)$ is an open embedding
\[
j \colon (\mathbb{C}^\times)^n \hookrightarrow Y\setminus D
\qquad
\text{such that}
\qquad
j^\star (\omega) = \frac{1}{(2\pi \mathrm{i})^n} \frac{\mathrm{d} x_1}{x_1}\wedge \cdots \wedge \frac{\mathrm{d} x_n}{x_n}.
\]
A \emph{cluster variety} is an $n$-dimensional CY pair $(Y, \omega)$ that
has a torus chart.
\end{dfn}
In our situation, the pair $(Y,D)$ will always be log smooth.
\subsection{The cluster surface}
We construct a cluster surface from a
lattice polygon.
Let $\mathbb{L}$ be an affine lattice of rank $2$.
There is a canonical bijection
between the set of lattice polygons $F\subset \mathbb{L}_\mathbb{R}$ up to translation and
the set of pairs $(\overline{X}, A)$ of a projective toric surface $\overline{X}$ and an ample
line bundle $A$ on $\overline{X}$: the torus in question is $\Spec \mathbb{C}[\mathbb{L}_0]$;
the fan of the surface $\overline{X}$ is the normal fan of $F$, and it all works
out such that there is a natural $1$-to-$1$ correspondence between
$F\cap \mathbb{L}$ and a basis of $H^0(\overline{X},A)$.
Fix a lattice polygon $F$ in $\mathbb{L}$ and consider the corresponding
polarised toric surface $(\overline{X}, A)$.
Denote by $\overline{B}$ the toric boundary of $\overline{X}$.
For each edge $E \leq F$,
let $\ell(E)$ be the lattice length of $E$ and let $\overline{B}_E$ be the prime component
of $\overline{B}$ corresponding to $E$;
we have that $\overline{B}_E$ is isomorphic to $\mathbb{P}^1$
and the line bundle $A \vert_{\overline{B}_E}$ has degree $\ell(E)$.
Denote by $x_E \in \overline{B}_E$ the point $[1:-1] \in \mathbb{P}^1$.
For all edges $E \leq F$, blow up $\ell(E)$ times above $x_E$ in the proper
transform of $\overline{B}_E$ and denote by
\[
p\colon (X,B) \longrightarrow (\overline{X}, \overline{B})
\]
the resulting
surface, where $B \subset X$ is the proper transform of the toric
boundary $\overline{B} = \sum_{E \leq F} \overline{B}_E$.
We call the pair $(X,B)$ the \emph{cluster
surface} associated to the lattice polygon $F$.
\begin{thm}
\label{thm:equivalent_defns}
Let $\mathbb{L}$ be a rank-$2$ lattice. The following are equivalent for a
normalized Laurent polynomial $f\in \mathbb{C} [\mathbb{L}]$:
\begin{enumerate}[(1)]
\item $f$ is $0$-mutable;
\item $f$ is rigid maximally mutable;
\item Let
$p \colon (X,B) \to (\overline{X}, \overline{B})$ be the cluster surface
associated to the polygon $F=\Newt f$.
Denote by $Z \subset \overline{X}$ the divisor of zeros of $f$ and by
$Z^\prime \subset X$ the proper transform of $Z$. Every irreducible
component $\Gamma \subset Z^\prime$ is a smooth rational curve
with self-intersection $\Gamma^2=-2$. (Necessarily then $B \cdot \Gamma
=0$ hence $\Gamma$ is disjoint from the boundary $B$.)
\end{enumerate}
\end{thm}
\begin{rem}
\label{rem:4}
The support of $Z^\prime$ is not necessarily a normal crossing
divisor. The irreducible components need not meet transversally, and
$\geq 3$ of them may meet at a point.
\end{rem}
\begin{proof}[Sketch of proof] In proving all equivalences we may and
will assume that the polynomial $f$ is irreducible hence $Z$ is
reduced and irreducible.
The proof uses the following ingredients, which we state without
further discussion or proof:
\begin{enumerate}[(i)]
\item To give a torus chart $j\colon \mathbb{C}^{\times\, 2} \hookrightarrow
X \setminus B$ in $X$ is the same as to give a
\emph{toric model} of $(X,B)$, that is a projective morphism
$q\colon (X, B) \to (X^\prime, B^\prime)$ where $(X^\prime,
B^\prime)$ is a toric pair and $q$ maps
$j(\mathbb{C}^{\times \, 2})$ isomorphically to the torus $X^\prime \setminus
B^\prime$;
\item The work of Blanc~\cite{MR3080816} implies that any two torus charts in $X$ are connected
by a sequence of mutations between torus charts in $X$;
\item A set $\mathscr{S}$ of mutation data specifies a line bundle $\mathcal{L}(\mathscr{S})$ on $X$
such that $H^0\bigl(X,\mathcal{L}(\mathscr{S})\bigr)=L(\mathscr{S})$ and, conversely, every line bundle on $X$
is isomorphic to a line bundle of the form $\mathcal{L}(\mathscr{S})$.
\end{enumerate}
Let us show first that (1) implies (3). To say that an irreducible
polynomial $f$ is $0$-mutable is to say that there exists a sequence
of mutations that mutates $f$ to the constant polynomial $\mathbf{1}$.
This sequence of mutations constructs a new torus chart
$j_1\colon \mathbb{T} = \mathbb{C}^{\times\, 2} \hookrightarrow X\setminus B$ such that the proper
transform $Z^\prime$ -- which is, by assumption, irreducible -- is
disjoint from $j_1 (\mathbb{T})$. This new toric chart gives a new toric model
$p_1 \colon (X,B) \to (X_1, B_1)$ that maps $j_1 (\mathbb{T})$
isomorphically to the torus $X_1 \setminus B_1$ and
hence contracts $Z^\prime$ to a boundary point. $Z^\prime$ is not a $-1$-curve, because those
are all $p$-exceptional, hence $Z^\prime$ is a $-2$-curve.
To show that (3) implies (1), by Lemma~\ref{lem:clustery} below, there
is a new toric model $p_1\colon (X, B) \to (X_1,B_1)$ that contracts
$Z^\prime$ to a point in the boundary. The new toric model then gives a
new torus chart $j_1\colon \mathbb{T} \hookrightarrow X$ such that $Z^\prime $ is
disjoint from $j_1(\mathbb{T})$. By Blanc the induced birational map of tori
\[
j_1^{-1}j \colon \mathbb{T} \dasharrow \mathbb{T}
\]
is a composition of mutations that mutates $f$ to the constant
polynomial.
Let us now show that (3) implies (2). By some tautology, $Z^\prime$ is the
zero divisor of the section $f$ of the line bundle $\mathcal{L}(\mathscr{S})$ on $X$ specified by
the set of mutation data $\mathscr{S}=\mathscr{S}(f)$, and $H^0(X, Z^\prime)=L
(\mathscr{S})$. Since $Z^\prime$ is a $-2$-curve, $L(\mathscr{S})$ is
$1$-dimensional, which is to say that $f$ is rigid maximally mutable.
Finally we show that (2) implies (3). Denote by $\mathcal{L}=\mathcal{L}(\mathscr{S})$
the line bundle on $X$ specified by the set of mutations $\mathscr{S}=\mathscr{S}(f)$, so
that $Z^\prime$ is the zero-locus of a section of $\mathcal{L}$. Note that:
\[
h^2(X, \mathcal{L})=h^0(X, K_X-Z^\prime)=h^0(Y, - \overline{B} -Z^\prime)=0
\]
Riemann--Roch and the fact that $f$ is rigid give:
\[
1=h^0(X, Z^\prime)=h^0(X, \mathcal{L})\geq \chi(X, \mathcal{L})=1+ \frac{1}{2}(Z^{\prime
\,2}+Z^\prime \overline{B})
\]
and hence, because $Z^\prime \overline{B} \geq 0$, we conclude that $Z^{\prime \,
2}\leq 0$ and:
\[
2p_a(Z^\prime)-2=\deg \omega_{Z^\prime} = Z^{\prime \, 2} - Z^\prime \overline{B} \leq 0
\]
so either:
\begin{enumerate}[(i)]
\item $\deg \omega_{Z^\prime}<0$ and then $Z^\prime$ is a smooth
rational curve, and then as above $Z^\prime$ is not a $-1$ curve therefore it is a
$-2$-curve, or
\item $\omega_{Z^\prime}=\mathcal{O}_{Z^\prime}$ and $Z^{\prime\, 2}=Z^\prime
\overline{B}=0$. It follows that $Z^\prime$ is actually disjoint from $\overline{B}$ and
$\mathcal{O}_{Z^\prime}(Z^\prime) = \mathcal{O}_{Z^\prime}$.
The homomorphism
\[
H^0(X, Z^\prime) \to H^0\left(Z^\prime, \mathcal{O}_{Z^\prime}(Z^\prime) \right)=\mathbb{C}
\]
is surjective, hence actually $h^0(X, Z^\prime)=2$, a
contradiction.
\end{enumerate}
This means that we must be in case (i) where $Z^\prime$
is a $-2$-curve.
\end{proof}
\begin{lem}
\label{lem:clustery}
Let $(Y,D)$ be a cluster surface, and
$Z^\prime \subset Y$ an interior $-2$-curve. Then there is a
toric model $q\colon (Y,D) \to (X^\prime, B^\prime)$ that contracts
$Z^\prime$.
\end{lem}
\begin{proof}[Sketch of proof]
First contract $Z^\prime$ to an interior node and then run a
MMP. There is a small number of cases to discuss depending on how
the MMP terminates.
\end{proof}
\section{Conjecture~B}
\label{sec:intr-mirr-symm}
In this section we state a new conjecture --- Conjecture~B --- which
implies, see Remark~\ref{rem:conjBimpliesmapdefined}, the existence of
the map $\kappa\colon \mathfrak{B}\to \mathfrak{A}$ of
Conjecture~A. In the strong form stated in
Remark~\ref{rem:conjBStrongForm}, together with Conjecture~A,
Conjecture~B asserts that the smoothing components of
$\Defo (V, \divV)$ are themselves smooth, and computes their tangent
space explicitly as a representation of the torus. We conclude by
explaining how the two Conjectures~A and~B originate in mirror
symmetry. This last discussion is central to how we arrived at the
formulation of the conjectures, but it is not logically necessary for
making sense of their statement. We work with the version of mirror
symmetry put forward in~\cite{MR3415066} and~\cite{gs_intrinsic_slc,
1909.07649}.\footnote{We thank Paul Hacking for several helpful
discussions on mirror symmetry and for correcting earlier drafts of
this section. We are of course responsible for the mistakes that are
left.}
\medskip
For the remainder of this section fix a lattice polygon $F$ and
denote, as usual, by $V$ the corresponding Gorenstein toric affine
\mbox{$3$-fold} with toric boundary $\partial V$.
In this section we always work with the space $\Defo
(V, \divV)$. Also fix a $0$-mutable polynomial $f \in \mathbb{C}[\mathbb{L}]$ with
$\Newt f = F$.
Conjecture~B associates to $f$ a
$\mathbb{T}$-equivariant family $(\mathscr{U}_f, \mathscr{D}_f) \to \moduli{f}$ of deformations of the pair $(V, \partial V)$.
In the last part of this section we construct from $f$ a $3$-dimensional cluster variety
$(W,D)$, conjecturally the \emph{mirror} of $\moduli{f}$, and hint at an explicit
conjectural construction of the family $\mathscr{U}_f \to \moduli{f}$ from
the degree-$0$ quantum log cohomology of $(W, D)$.
\medskip
Denote by $\sigma \subset N_\mathbb{R}$ the cone over $F$ at height $1$.
As
usual, $\usp \in M = \Hom_\mathbb{Z} (N,\mathbb{Z})$ denotes the Gorenstein degree,
so $F=\sigma \cap \mathbb{L}$ where $\mathbb{L}=(u=1)$.
\subsection{Statement of Conjecture~B}
\label{sec:conjectures}
As in Sec.~\ref{sec:rigid-maxim-mutable}, denote by $\mathscr{S}(f)$ the set
of mutation data of $f$. Recall that an element of $\mathscr{S}(f)$ is a
pair $(\varphi, h)$ consisting of an affine function
$\varphi \in \Aff (\mathbb{L}, \mathbb{Z})$ and a Laurent polynomial
$h\in \mathbb{C}[\Ker \varphi_0]$. Using the restriction isomorphism
$M \simeq \Aff(\mathbb{L}, \mathbb{Z})$, when it
suits us we view a mutation datum $(\varphi, h)$ as a pair of an element $\varphi \in M$ and a
polynomial $h\in \mathbb{C}[\mathbb{L}_0\cap \Ker \varphi] \subseteq \mathbb{C}[N]$.
The most useful mutation data are those where $\varphi$ is strictly
negative somewhere on $F$, and then the minimum of $\varphi$ on $F$ is
achieved on an edge $E\leq F$.\footnote{Recall that in
Definition~\ref{dfn:mutation} we explicitly assume that $\varphi$ is not constant on $F$.} In the
present discussion we want to focus on these mutation data:
\[
\mathscr{S}_{-}(f)= \left\{ (\varphi, h)\in \mathscr{S}(f) \mid \text{for some edge $E\leq
F$, $\varphi|_E$ is constant and $<0$} \right\}.
\]
We want to define a seed $\widetilde{\mathscr{S}}(f)$ on $N$, that is, a set of
pairs $(\varphi,h)$ of a character $\varphi \in M$ and a Laurent
polynomial $h\in \mathbb{C}[\Ker \varphi ]$. The seed we want is
\[
\widetilde{\mathscr{S}}(f)=\mathscr{S}_{-}(f)
\cup \Bigl\{ (-ku,h)\ \Big\vert\ \text{$h$ is a prime factor of $f$ of multiplicity
$k$}\Bigr\}.
\]
\begin{conB}
In this statement, if $U$ is a representation of the torus $\mathbb{T} = \Spec \mathbb{C}[M]$ and $m\in M$
is a character of $\mathbb{T}$, we denote by $U(m)$ the direct summand of $U$
on which $\mathbb{T}$ acts with pure weight $m$.
Let $F\subset \mathbb{L}$ be a lattice polygon, $V$ the
corresponding Gorenstein toric affine $3$-fold, and $f\in \mathbb{C}[\mathbb{L}]$
a $0$-mutable polynomial with $\Newt f = F$.
For every integer $k\geq 1$, denote by $n_k$ the number of prime
factors of $f$ of multiplicity $\geq k$.
Then there is a $\mathbb{T}$-invariant submanifold $\moduli{f} \subset \Defo (V, \divV)$
such that
\[
\dim T_{0} \moduli{f} (m) =
\begin{cases}
1 & \; \text{if $m \not \in \langle-u\rangle_+$ and $\exists$ $(\varphi, h)\in
\widetilde{\mathscr{S}}(f)$ such that $m=\varphi$},\\
n_k & \; \text{if $m=-ku$ for some integer $k \geq 1$},\\
0 & \; \text{otherwise},
\end{cases}
\]
and the general fibre of the
family over $\moduli{f}$ is a pair consisting of a smooth variety and a smooth divisor.
\end{conB}
\begin{rem}
\label{rem:conjBimpliesmapdefined}
If Conjecture~B holds then by openness of versality a general point
of $\moduli{f}$ lies in a unique component of $\Defo (V, \divV)$ and
this gives the map $\kappa\colon \mathfrak{B}\to \mathfrak{A}$ in
the statement of Conjecture~A (see also
Remark~\ref{rem:deformations_variety_or_pair}).
\end{rem}
\begin{rem}[Strong form of Conjecture~B]
\label{rem:conjBStrongForm}
A strong form of Conjecture~B states that the families
$\moduli{f}$ are precisely the smoothing components of
$\Defo (V,\divV)$.
\end{rem}
\begin{exa}[Example~\ref{ex:quadrilateral}
continued] \label{ex:quadrilateral_conjecture} Consider the quadrilateral $F$ in $\mathbb{L} = \mathbb{Z}^2$ defined in \eqref{eq:quadrilateral} and the three $0$-mutable polynomials $\alpha$, $\beta$ and $\gamma$ with Newton polytope $F$.
Set $N = \mathbb{L} \oplus \mathbb{Z} = \mathbb{Z}^3$ and consider the following linear functionals in $M = \Hom_\mathbb{Z}(N,\mathbb{Z}) = \mathbb{Z}^3$:
\begin{align*}
-m^1_{2,1} &= (0,1,-1) & -m^2_{2,1} &= (1,0,-1) \\
-m^1_{3,1} &= (0,1,-2) & -m^2_{3,1} &= (1,0,-2) \\
-m^1_{3,2} &= (0,2,-1) & -m^2_{3,2} &= (2,0,-1)
\end{align*}
and $-u = (0,0,-1)$.
The names $-m^1_{2,1}$, $-m^1_{3,1}$ and $-m^1_{3,2}$ are compatible with the affine-linear functions considered in Example~\ref{ex:quadrilateral} via the restriction isomorphism $M \simeq \Aff(\mathbb{L}, \mathbb{Z})$.
Set $x = x^{(1,0,0)} \in \mathbb{C}[N]$ and $y = x^{(0,1,0)} \in \mathbb{C}[N]$. Then we have:
\begin{gather*}
\widetilde{\mathscr{S}}(\alpha) \supseteq \left\{ (-u, 1+x+2y + y^2) , (-u, 1+y+2x+x^2) \right\}, \\
\widetilde{\mathscr{S}}(\alpha) \supseteq \left\{ (-m^1_{2,1}, 1+x), (-m^1_{3,2}, 1+x), (-m^2_{2,1}, 1+y), (-m^2_{3,2}, 1+y) \right\}, \\
\widetilde{\mathscr{S}}(\beta) \supseteq \left\{(-u, \beta), (-m^1_{2,1}, 1+x), (-m^1_{3,1}, 1+x), (-m^1_{3,2}, 1+x) \right\}, \\
\widetilde{\mathscr{S}}(\gamma) \supseteq \left\{(-u, \gamma), (-m^2_{2,1}, 1+y), (-m^2_{3,1}, 1+y), (-m^2_{3,2}, 1+y) \right\}.
\end{gather*}
Let $V$ be the Gorenstein toric affine $3$-fold associated to $F$.
Conjecture~B states that there are three submanifolds $\moduli{\alpha}$, $\moduli{\beta}$ and $\moduli{\gamma}$ of $\Defo (V, \partial V)$ such that the dimensions of $T_0 \moduli{\alpha}(m)$, $T_0 \moduli{\beta}(m)$ and $T_0 \moduli{\gamma}(m)$ for $m \in \{-u, -m^1_{2,1}, -m^1_{3,1}, -m^1_{3,2}, -m^2_{2,1}, -m^2_{3,1}, -m^2_{3,2} \}$ are written down in the table below.
\[
\begin{array}{cccccccc}
& -u & -m^1_{2,1} & -m^1_{3,1} & -m^1_{3,2} & -m^2_{2,1} & -m^2_{3,1} & -m^2_{3,2} \\
\hline
\dim T_0 \moduli{\alpha}(m) & 2 & 1 & 0 & 1 & 1 & 0 & 1 \\
\dim T_0 \moduli{\beta}(m) & 1 & 1 & 1 & 1 & 0 & 0 & 0 \\
\dim T_0 \moduli{\gamma}(m) & 1 & 0 & 0 & 0 & 1 & 1 & 1
\end{array}
\]
\end{exa}
\subsection{Mirror symmetry interpretation}
\label{sec:mirr-symm-interpr}
Denote by $\mathscr{U}_f \to \moduli{f}$ the deformation family of $V$ induced by the composition
$\moduli{f} \hookrightarrow \Defo(V, \divV) \twoheadrightarrow \Defo V$. We sketch a construction of
$\mathscr{U}_f$ in the spirit of intrinsic mirror symmetry
\cite{gs_intrinsic_slc, 1909.07649}.
\medskip
Let $\sigma^\vee \subset M_\mathbb{R}$ be the dual cone, and let
$s_1, \dots, s_r$ be the primitive generators of the rays of
$\sigma^\vee$. Denote by $W^\flat$ the toric variety --- for the dual
torus $\mathbb{T}_M=\Spec \mathbb{C}[N]$ --- constructed from the fan consisting of
the cones $\{0\}$, $\langle -u\rangle_+$, the $\langle s_j\rangle_+$, and the
two-dimensional cones
\[
\langle -u, s_j\rangle_+
\]
(for $j=1, \dots, r$), and let $D^\flat \subset W^\flat$ be the toric
boundary.
\medskip
Now, the set of edges $E\leq F$ is in $1$-to-$1$ correspondence with
the set $\{ s_1, \dots, s_r \}$, where $E$
corresponds to $s_j$ if $s_j|_E=0$. If
$(\varphi, h) \in \widetilde{\mathscr{S}}(f)$ and
$\varphi \not \in \langle -u\rangle_+$, then there is a unique $j$
such that $\varphi$ is in the cone $\langle -u, s_j\rangle_+$ spanned
by $-u$ and $s_j$.
Let
\[(\overline{W}, \Dbar) \longrightarrow (W^\flat, D^\flat)\]
be the toric variety obtained by
adding the rays $\langle \varphi \rangle_+ \subset M_\mathbb{R}$ whenever
$(\varphi, h)\in \widetilde{\mathscr{S}}(f)$, and (infinitely many)
two-dimensional cones subdividing the cones
$\langle -u, s_j\rangle_+$. Note that $\overline{W}$ is not quasi-compact and
not proper.
\medskip
Finally, we construct a projective morphism \[ (W, D) \longrightarrow (\overline{W},
\Dbar)\] by a sequence of blowups.
In what follows for all $(\varphi, h)\in \widetilde{\mathscr{S}}(f)$ we
denote by $D_{\langle \varphi \rangle_+}\subset \overline{W}$ the
corresponding boundary component, and set
\[
Z_h=(h=0)\subset D_{\langle \varphi \rangle_+}.
\]
The following simple remarks will be helpful in describing the construction.
\begin{enumerate}
\item[(a)] For all positive integers $k$, $(k\varphi, h)\in \mathscr{S}_{-}(f)$
if and only if $(\varphi, h^k)\in \mathscr{S}_{-}(f)$.
\item[(b)] By construction, if $(\varphi, h)\in \mathscr{S}_{-}(f)$, then one has $h=(1+x^e)^k$
(up to translation) for some positive integer $k$,
where $e\in M$ is a primitive lattice vector along the edge $E\leq F$ where
$\varphi$ achieves its minimum.
\end{enumerate}
Let $R\subset M_\mathbb{R}$ be a ray of the fan of $\overline{W}$ other than
$\langle -u\rangle_+$, and let $\varphi
\in M$ be the primitive generator of $R$. It follows from the
remarks just made that there is a largest positive integer $k_R$
such that $(k_R \varphi, 1+x^e)\in \mathscr{S}_{-}(f)$.
\medskip
Our mirror $W$ is obtained from $\overline{W}$ by:
\begin{enumerate}
\item First, as $R$ runs through all the rays of the fan of $\overline{W}$
other than $\langle-u\rangle_+$ in some order, blow up $k_R$ times
above $(1+x^e=0)$ in the proper transform of $D_R$. It can be seen, and it
is a nontrivial fact, that after doing all these blowups the $Z_h$
in the proper transforms of the $D_{\langle -u\rangle_+}$ are
smooth;
\item Subsequently, if $f=\prod h^{k(h)}$ where the $h$ are
irreducible, blow up in any order $k(h)$ times above $Z_h$ in the
proper transform of $D_{\langle -u\rangle_+}$.
\end{enumerate}
The resulting CY pair $(W, D)$ is log smooth, because we have blown up
a sequence of smooth centres. This $(W, D)$ is the \emph{mirror} of
$\moduli{f}$.
One can see that the Mori cone $\NE(W/\overline{W})$ is simplicial, and hence
there is an identification:
\[
\moduli{f} = \Spec \mathbb{C}[\NE(W/\overline{W})]
\]
Mirror symmetry suggests --- modulo issues with
infinite-dimensionality ---
that the ring $QH^0_{\log} (W,D;\mathbb{C}[\NE(W/\overline{W})])$ has a natural filtration and that
one recovers the universal family of pairs $(\mathscr{U}_f, \mathscr{D}_f)\to \mathscr{M}_f$ from this ring out of the
Rees construction.
\begin{rem}
In the context of Conjecture~B, it would be very nice to work out an
interpretation of the coefficients of the $0$-mutable polynomial $f$
as counting certain holomorphic disks on the general fibre of the
family $\mathscr{U}_f\to \moduli{f}$.
\end{rem}
\section{Evidence}
\label{sec:evidence-1}
We have already remarked that the variety $V$ of Example~\ref{exa:conePP123} confirms Conjecture~A. Here we collect some further evidence.
Section~\ref{sec:evidence-matej} is a study of $\Defo V$ where $V$ is the variety of Example~\ref{ex:quadrilateral} and Example~\ref{ex:quadrilateral_conjecture}.
\subsection{Isolated singularities}
Here we fix a lattice polygon $F$ with unit edges, i.e.\ edges with lattice length $1$.
Let $V$ be the Gorenstein toric affine \mbox{$3$-fold} associated to $F$; we have that $V$ has an isolated singularity.
Altmann~\cite{altmann_versal_deformation} proved that there is a 1-to-1 correspondence between the irreducible components of $\Defo V$ and the maximal Minkowski decompositions of $F$.
This restricts to a 1-to-1 correspondence between the smoothing components of $\Defo V$ and the Minkowski decompositions of $F$ with summands that are either unit segments or standard triangles. Here a standard triangle is a lattice triangle that is $\mathbb{Z}^2 \rtimes \mathrm{GL}_2(\mathbb{Z})$-equivalent to $\hull{(0,0),(1,0),(0,1)}$.
On a polygon $F$ with unit edges, the $0$-mutable polynomials are exactly those that are associated to the Minkowski decompositions of $F$ into unit segments and standard triangles. This confirms Conjecture~A.
\subsection{Local complete intersections}
Nakajima \cite{nakajima} has characterised the affine toric varieties that are local complete intersection (lci for short). These are Gorenstein toric affine varieties associated to certain lattice polytopes called Nakajima polytopes. We refer the reader to \cite[Lemma~2.7]{toric_lci_are_resolvable} for an inductive characterisation of Nakajima polytopes. From this characterisation it is very easy to see that every Nakajima polygon is affine equivalent to
\begin{equation} \label{eq:nakajima}
F_{a,b,c} = \hull{(0,0),(a,0),(0,b),(a,b+ac)}
\end{equation}
in the lattice $\mathbb{Z}^2$, for some non-negative integers $a,b,c$ such that $a \geq 1$ and $ b + c \geq 1$.
It is easy to show that the Gorenstein toric affine \mbox{$3$-fold} associated to the polygon $F_{a,b,c}$ is
\[
V_{a,b,c} = \Spec \mathbb{C}[x_1,x_2,x_3,x_4,x_5]/(x_1 x_2 - x_4^c x_5^b, \ x_3 x_4 - x_5^a).
\]
There is a unique $0$-mutable polynomial on $F_{a,b,c}$: this is associated to the unique Minkowski decomposition of $F_{a,b,c}$ into $a$ copies of the triangle $F_{1,0,c} = \hull{(0,0),(1,0),(0,c)}$ and $b$ copies of the segment $\hull{(0,0),(0,1)}$. On the other hand, as $V_{a,b,c}$ is lci, we have that $V_{a,b,c}$ is unobstructed and smoothable, therefore there is a unique smoothing component in the miniversal deformation space of $V_{a,b,c}$. This confirms Conjecture~A.
\section{A worked example}
\addtocontents{toc}{\setcounter{tocdepth}{1}}
\label{sec:evidence-matej}
We explicitly compute the smoothing components of the miniversal
deformation space of the Gorenstein toric affine \mbox{$3$-fold} $V$
associated to the polygon $F$ of Example~\ref{ex:quadrilateral}
(continued in Example~\ref{ex:quadrilateral_conjecture}). We saw that
there exist exactly three $0$-mutable polynomials with Newton polytope
$F$: $\alpha$, $\beta$ and $\gamma$. We explicitly compute the
miniversal deformation space of $V$ and see that it has three
irreducible components, all of which are smoothing components. This
confirms our conjectures.
\subsection{The equations of $V$}
We consider the quadrilateral
\begin{equation} \label{eq:quadrilateral_last}
F = \hull{\vectortwo{-1}{-1}, \vectortwo{2}{-1}, \vectortwo{1}{1}, \vectortwo{-1}{2}}
\end{equation}
in the lattice $\mathbb{L} = \mathbb{Z}^2$.
We consider the cone $\sigma$ obtained by placing $F$ at height $1$,
i.e.\ $\sigma$ is the cone generated by
\begin{equation*}
a_1=\vectorthree{-1}{-1}{1} \qquad a_2=\vectorthree{2}{-1}{1} \qquad a_3=\vectorthree{1}{1}{1} \qquad a_4=\vectorthree{-1}{2}{1}
\end{equation*}
in the lattice $N = \mathbb{L} \oplus \mathbb{Z} = \mathbb{Z}^3$, and the corresponding
Gorenstein toric affine \mbox{$3$-fold}
$V = \Spec \mathbb{C}[\sigma^\vee \cap M]$, where $\sigma^\vee$ is the dual cone of $\sigma$ in the dual lattice
$M = \Hom_\mathbb{Z}(N,\mathbb{Z}) \simeq \mathbb{Z}^3$.
The Gorenstein degree is
\[
\usp = (0,0,1) \in M.
\]
The primitive generators of the rays of the dual cone $\sigma^\vee \subseteq M_\mathbb{R}$ are the vectors
\begin{align*}
s_1 = (0,1,1), \quad s_2 = (1,0,1), \quad s_3 = (-1,2,3), \quad s_4 = (-2,-1,3)
\end{align*}
which are orthogonal to the $4$ edges of $F$.
The Hilbert basis of the monoid $\sigma^\vee \cap M$ is the set of the vectors
\begin{align*}
\usp, \ s_1, \ z_2 = (-1,0,2),\ s_4, \ z_3 = (-1,-1,2), \ s_3, \ z_4 = (0,-1,2), \ s_2.
\end{align*}
Notice that these are the Gorenstein degree $\usp$ and
certain lattice vectors on the boundary
of $\sigma^\vee$.
The elements of the Hilbert basis of $\sigma^\vee \cap M$ are depicted in Figure~\ref{fig:slices}.
\begin{figure}
\centering
\def<desired width>{\textwidth}
\input{Slicesx.tex}
\bigskip
\caption{The intersections of the cone $\sigma^\vee$ with the planes $\mathbb{R}^2 \times \{1\}$, $\mathbb{R}^2 \times \{2\}$ and $\mathbb{R}^2 \times \{3\}$ in $M_\mathbb{R} = \mathbb{R}^3$}
\label{fig:slices}
\end{figure}
\begin{comment}
The set $E \cap \partial \sigma^\vee$ has a cyclic order.
Any three consecutive elements of $E \cap \partial \sigma^\vee$ provide a unique linear relation among them and $\usp$. In particular, if one chooses the three consecutive elements such that $s_j$ lies in the middle, say $\{r, s_j, r'\}$, one will get the following relation
\begin{equation}\label{eq alt eq}
r + r' = \ell(j) \usp + k s_j
\end{equation}
where $\ell(j)$ is the lattice length of the edge of $F$ which is orthogonal to $s_j$ and $k$ is some non-negative integer.
In our example we have
\begin{equation*}
s_2+z_2=3\usp, \quad z_2+z_3=\usp+s_4, \quad z_3+z_4=\usp+s_3, \quad z_4+s_1=3\usp.
\end{equation*}
\end{comment}
The elements of the Hilbert basis of $\sigma^\vee \cap M$ give a closed embedding of $V$ inside $\AA^8$ such that the ideal is generated by binomial equations.
By using rolling factors formats (see \cite{stevens_rolling} and \cite[\S12]{stevens_book}), one can\footnote{We are obliged to the referee for suggesting these equations to us.} see that these equations are:
\begin{gather*}
\mathrm{rank}
\begin{pmatrix}
x_{s_1} & x_{z_2} & x_\usp & x_{s_2} & x_{z_4}\\
x_{z_2} & x_{s_4} & x_{z_3} & x_{z_4} & x_{s_3}
\end{pmatrix}
\leq 1, \\
x_{s_4}x_{s_3}-x^3_{z_3}=0 \qquad \qquad x_{z_2} x_{s_3} - x_{z_3}^2 x_u = 0, \\
x_{z_2}x_{z_4}-x_{z_3} x_\usp^2=0 \qquad \qquad x_{s_1} x_{z_4} - x_\usp^3=0.
\end{gather*}
The singular locus of $V$ has two irreducible components of dimension
$1$: $V$ has generically transverse $A_2$-singularities along each of these.
\subsection{The tangent space}
We consider the tangent space to the deformation functor of $V$, i.e.\ $T^1_V = \Ext^1_{\mathcal{O}_V}(\Omega_V^1, \mathcal{O}_V)$.
This is a $\mathbb{C}$-vector space with an $M$-grading.
For every $m \in M$ we denote by $T^1_V(-m)$ the graded component of $T^1_V$ of degree $-m$.
\begin{lem} \label{lem:t1}
We define $J := \{ (p,q) \in \mathbb{Z}^2 \mid 2 \leq p \leq 3, \ q \geq 1 \}$.
For all $p, q \in \mathbb{Z}$ we set $m^1_{p,q} := p \usp - q s_1$ and
$
m^2_{p,q} := p \usp - q s_2$.
Then
\[
\dim T^1_V(-m) =
\begin{cases}
1 &\text{if } m = \usp, \\
1 &\text{if } m = m^1_{p,q} \text{ with } (p,q) \in J, \\
1 &\text{if } m = m^2_{p,q} \text{ with } (p,q) \in J, \\
0 &\text{otherwise.}
\end{cases}
\]
\end{lem}
\begin{proof}
This is a direct consequence of \cite[Theorem~4.4]{altmann_one_parameter}.
\end{proof}
Some of the degrees of $T^1_V$ are depicted in Figure~\ref{fig:slab}.
\begin{figure}
\centering
\def<desired width>{4cm}
\input{Slabx.tex}
\caption{Some degrees of $T^1_V$ in the plane $\mathbb{R} \times \{0\} \times \mathbb{R} \subseteq M_\mathbb{R} = \mathbb{R}^3$.}
\label{fig:slab}
\end{figure}
\medskip
The base of the miniversal deformation of $V$ is the formal completion (or germ) at the origin
of a closed subscheme of the countable-dimensional affine space $T^1_V$.
We denote by $t_m$
the coordinate on the 1-dimensional $\mathbb{C}$-vector space $T^1_V(-m)$,
when $m = \usp$ or $m \in \{m^1_{p,q}, m^2_{p,q} \}$ with
$(p,q) \in J$.
Since we want to understand the structure of $\Defo V$, we want to analyse the equations of $\Defo V \hookrightarrow T^1_V$
in the variables $t_u$, $t_{m^1_{p,q}}$ and $t_{m^2_{p,q}}$ for $(p,q) \in J$.
The first observation is that each homogeneous first order deformation of $V$ is unobstructed as we see in the following two remarks.
\begin{rem} \label{rem:Minkowski_sum_gives_deform} The 1-dimensional
$\mathbb{C}$-vector space $T^1_V(-\usp)$ gives a first order deformation of
$V$, i.e.\ an infinitesimal deformation of $V$ over
$\mathbb{C}[t_\usp]/(t_\usp^2)$. This deformation can be extended to an
algebraic deformation of $V$ over $\mathbb{C}[t_u]$ as follows
(see \cite{altmann_minkowski_sums}).
Consider the unique non-trivial Minkowski decomposition of $F$ (see Figure~\ref{fig:minkowski_dec_F}):
\begin{equation} \label{eq:Minkowski_deco_of_F}
F = \hull{\vectortwo{0}{0},\vectortwo{1}{0},\vectortwo{0}{2}} + \hull{\vectortwo{-1}{-1},\vectortwo{1}{-1},\vectortwo{-1}{0}}.
\end{equation}
Let $\widetilde{F}$ be the Cayley polytope associated to this Minkowski sum; $\widetilde{F}$ is a $3$-dimensional lattice polytope.
Let $\tilde{\sigma}$ be the cone over $\widetilde{F}$ at height $1$, i.e.\ $\tilde{\sigma}$ is the $4$-dimensional cone generated by
\begin{equation*}
\vectorfour{0}{0}{1}{0}, \vectorfour{1}{0}{1}{0}, \vectorfour{0}{2}{1}{0}, \vectorfour{-1}{-1}{0}{1}, \vectorfour{1}{-1}{0}{1}, \vectorfour{-1}{0}{0}{1}.
\end{equation*}
Let $\widetilde{V}$ be the Gorenstein toric affine \mbox{$4$-fold} $\widetilde{V}$ associated to $\tilde{\sigma}$. Consider the difference of the two regular functions on $\widetilde{V}$ associated to the characters $(0,0,1,0)$ and $(0,0,0,1)$; if we consider this regular function on $\widetilde{V}$ as a morphism $\widetilde{V} \to \AA^1$, we obtain the following cartesian diagram
\begin{equation*}
\xymatrix{
V \ar[r] \ar[d] & \widetilde{V} \ar[d] \\
\{0 \} \ar[r] & \AA^1
}
\end{equation*}
which gives the wanted $1$-parameter deformation of $V$.
\begin{figure}
\begin{tikzpicture}[scale=0.4]
\draw[thick, color=black]
(0,0) -- (3,0) -- (2,2) -- (0,3) -- cycle;
\fill[thick, color=black]
(0,0) circle (3.5pt) (3,0) circle (3.5pt) (2,2) circle (3.5pt)
(2,0) circle (3.5pt) (1,0) circle (3.5pt) (1,1) circle (3.5pt) (2,1) circle (3.5pt) (1,2) circle (3.5pt) (0,2) circle (3.5pt) (0,1) circle (3.5pt) (0,3) circle (3.5pt);
\draw[thick, color=black]
(3, 0.9) node{$\hspace{2.2em}=\hspace{0.5em}$};
\end{tikzpicture}
\begin{tikzpicture}[scale=0.4]
\draw[thick, color=black]
(0,0) -- (1,0) -- (0,2) -- cycle;
\fill[thick, color=black]
(0,0) circle (3.5pt) (1,0) circle (3.5pt) (0,2) circle (3.5pt) (0,1) circle (3.5pt);
\draw[thick, color=black]
(1.5, 0.9) node{$\hspace{2.2em}+\hspace{0.5em}$};
\end{tikzpicture}
\begin{tikzpicture}[scale=0.4]
\draw[thick, color=black]
(0,0) -- (2,0) -- (0,1) -- cycle;
\fill[thick, color=black]
(0,0) circle (3.5pt) (2,0) circle (3.5pt) (0,1) circle (3.5pt) (0,1) circle (3.5pt) (1,0) circle (3.5pt);
\end{tikzpicture}
\caption{The Minkowski decomposition \eqref{eq:Minkowski_deco_of_F} of the quadrilateral $F$ defined in \eqref{eq:quadrilateral_last}}
\label{fig:minkowski_dec_F}
\end{figure}
\end{rem}
\begin{rem}
For every $m \in \{ m^1_{p,q}, m^2_{p,q} \}$ with $(p,q) \in J$,
the first order deformation of $V$ corresponding to
$T^1_V(-m) \simeq \mathbb{C}$ can be extended to an algebraic deformation
of $V$ over $\mathbb{C}[t_m]$ thanks to
\cite[Theorem~3.4]{altmann_one_parameter} (see also \cite{mavlyutov,
petracci_mavlyutov}).
\end{rem}
\subsection{The obstruction space}
We now consider the obstruction space of the deformation functor of $V$:
$T^2_V = \Ext^2_{\mathcal{O}_V}(\Omega^1_V, \mathcal{O}_V)$. This is an $M$-graded
$\mathbb{C}$-vector space. For all $m\in M$ we denote by $T^2_V(-m)$ the
direct summand of degree $-m$.
\begin{lem}\label{lem t2}
If $m \in \{4\usp-s_1, \ 4\usp-s_2, \ 5\usp-s_1-s_2, \ 6\usp-s_1-s_2, \ 9\usp-2s_1-2s_2\}$, then $\dim T^2_V(-m) = 1$. Otherwise $\dim T^2_V(-m) = 0$.
\end{lem}
\begin{proof}
This is a direct computation using formulae in \cite[Section 5]{altmann_andre_quillen_cohomology}.
\end{proof}
\begin{rem}
It immediately follows from the computation in \cite[Section
5]{altmann_andre_quillen_cohomology} that $T^2_V(-m)=0$ if there
exists $a_i$ such that $\langle m,a_i \rangle \leq 0$.
\end{rem}
\subsection{Verifying the conjectures}
Since $\dim T^2_V = 5$, the ideal of the closed embedding $\Defo V \hookrightarrow T^1_V$ has at most $5$ generators. We have the following:
\begin{pro} \label{pro:eq}
The equations of the closed embedding $\Defo V \hookrightarrow T^1_V$ are
\begin{align*}
t_{m^1_{3,1}} t_\usp = 0, \\
t_{m^2_{3,1}} t_\usp = 0, \\
t_{m^1_{3,1}} t_{m^2_{2,1}} + t_{m^1_{2,1}} t_{m^2_{3,1}}= 0, \\
t_{m^1_{3,1}}t_{m^2_{3,1}} = 0, \\
t^2_{m^2_{3,1}} t_{m^1_{3,2}} - t^2_{m^1_{3,1}} t_{m^2_{3,2}} = 0.
\end{align*}
Moreover, $\Defo V$ is non-reduced and has exactly 3 irreducible components; their equations inside $T^1_V$ are:
\begin{enumerate}
\item $t_{m^1_{3,1}}=t_{m^2_{3,1}}=0$,
\item $t_{\usp}=t_{m^1_{3,1}}=t_{m^1_{2,1}}=t_{m^1_{3,2}}=0$,
\item $t_{\usp}=t_{m^2_{3,1}}=t_{m^2_{2,1}}=t_{m^2_{3,2}}=0$.
\end{enumerate}
Every irreducible component of $\Defo V$ is smooth and is a smoothing component.
\end{pro}
\begin{proof}
The proof of the equations of $\Defo V \hookrightarrow T^1_V$ is postponed to
the next section and relies on some computer calculations performed
with Macaulay2. We now assume to know these equations.
The fact that $\Defo V$ is non-reduced and has $3$ irreducible
components $C_1$, $C_2$, $C_3$ with the equations given above can
be checked by taking the primary decomposition of the ideal of
$\Defo V \hookrightarrow T^1_V$. For each $i=1,2,3$, from the equations of
$C_i$ it is obvious that $C_i$ is smooth. We need to prove that
$C_i$ is a smoothing component, i.e.\ the general fibre over $C_i$
is smooth.
The component $C_1$ contains the $1$-parameter deformation
constructed in Remark~\ref{rem:Minkowski_sum_gives_deform}. The
singular locus of the general fibre of this deformation has $2$
connected components with everywhere transverse
$A_2$-singularities; therefore the general fibre of this
deformation is smoothable.
In order to prove that the general fibre over $C_2$ (resp.\ $C_3$)
is smooth, we prove that the general fibre over the $2$-parameter
deformation of $V$ with parameters $t_{m^2_{3,1}}$ and
$t_{m^1_{2,2}}$ (resp.\ $t_{m^1_{3,1}}$ and $t_{m^2_{2,2}}$) is
smooth. This can be done by applying the jacobian criterion to the
output of the computer calculations that we will describe below.
\end{proof}
We now illustrate Conjecture~A and Conjecture~B in our example. Let $C_1$, $C_2$,
$C_3$ be the 3 irreducible components of $\Defo V$, whose equations
are given in Proposition~\ref{pro:eq}.
By Remark~\ref{rem:deformations_variety_or_pair} $\Defo (V, \partial V)$ has 3 irreducible components, $\mathscr{M}_1$, $\mathscr{M}_2$, $\mathscr{M}_3$, each of which lie over exactly one of $C_1$, $C_2$, $C_3$.
For each $i \in \{1,2,3\}$ the smooth morphism $\mathscr{M}_i \to C_i$ induces a surjective linear map $T_0 \mathscr{M}_i \to T_0 C_i$ of linear representations of the torus $\Spec \mathbb{C}[M]$.
Let $\alpha$, $\beta$ and $\gamma$ be the three $0$-mutable polynomials with Newton polytope $F$ (see Example~\ref{ex:quadrilateral}).
By comparing the degrees of $T_0 C_1$, $T_0 C_2$, $T_0 C_3$ with the seeds $\widetilde{\mathscr{S}}(\alpha)$, $\widetilde{\mathscr{S}}(\beta)$, $\widetilde{\mathscr{S}}(\gamma)$ in Example~\ref{ex:quadrilateral_conjecture}, we have that $\alpha$ (resp.\ $\beta$, resp.\ $\gamma$) corresponds to $\mathscr{M}_1$ (resp.\ $\mathscr{M}_2$, resp.\ $\mathscr{M}_3$).
\subsection{Computer computations} \label{sec:computer}
Here we present a proof of Proposition~\ref{pro:eq} which uses the software
Macaulay2 \cite{M2}, in particular the package \verb|VersalDeformations| \cite{ilten_package_article, stevens_computing}.
By observing the degrees of $T^1_V$ (Lemma~\ref{lem:t1}) and the
degrees of $T^2_V$ (Lemma~\ref{lem t2}) it is immediate to see that
each of the 5 equations of $\Defo V \hookrightarrow T^1_V$ can only involve the
following $9$ variables:
\[
t_u \quad
t_{m^1_{3,1}} \quad
t_{m^1_{2,1}} \quad
t_{m^1_{3,2}} \quad
t_{m^1_{2,2}} \quad
t_{m^2_{3,1}} \quad
t_{m^2_{2,1}} \quad
t_{m^2_{3,2}} \quad
t_{m^2_{2,2}}.
\]
We call the corresponding $9$ degrees of $T^1_V$ the `interesting' degrees of $T^1_V$. This implies that there is a smooth morphism $\Defo V \to G$, where $G$ is a finite dimensional germ with embedding dimension $9$.
We now want to use the computer to determine $G$.
We consider the vector
\[
\begin{pmatrix}
3 \\ 4 \\ 5
\end{pmatrix} \in N = \mathbb{Z}^3.
\]
This gives a homomorphism $M \to \mathbb{Z}$ and a $\mathbb{Z}$-grading on on the
algebra $\mathbb{C} [\sigma^\vee \cap M] = H^0(V, \mathcal{O}_V)$, on $T^1_V$, and on
$T^2_V$. We have chosen this particular $\mathbb{Z}$-grading because the
corresponding linear projection is injective on the set
$\{u, m^1_{3,1}, m^1_{2,1}, m^1_{3,2}, m^1_{2,2}, m^2_{3,1},
m^2_{2,1}, m^2_{3,2}, m^2_{2,2} \}$,
which will allow us to identify our $9$ variables above with the
corresponding output of Macaulay2 below. In the following tables we
write down the degrees in $\mathbb{Z}$ of the Hilbert basis of
$\sigma^\vee \cap M$, of the interesting degrees of $T^1_V$, and of
the degrees of $T^2_V$.
\[
\begin{array}{cccccccc}
s_1 & z_2 & s_4 & z_3 & s_3 & z_4 & s_2 & u \\
9 & 7 & 5 & 3 & 4 & 6 & 8 & 5
\end{array}
\]
\[
\begin{array}{ccccccccc}
-u & -m^1_{3,1} & -m^1_{2,1} & -m^1_{3,2} & -m^1_{2,2} & -m^2_{3,1} & -m^2_{2,1} & -m^2_{3,2} & -m^2_{2,2} \\
-5 & -6 & -1 & 3 & 8 & -7 & -2 & 1 & 6
\end{array}
\]
\[
\begin{array}{ccccc}
4u-s_1 & 4u-s_2 & 5u-s_1-s_2 & 6u-s_1 -s_2 & 9u-2s_1-2s_2 \\
-11 & -12 & -8 & -13 & -11
\end{array}
\]
One can see that all non-interesting summands of $T^1_V$ have degree $\geq 9$.
Therefore we are interested in the summands of $T^1_V$ with degree between $-7$ and $8$.
Now we run the following Macaulay2 code, which was suggested to us by the referee.
\begin{verbatim}
S = QQ[s1,z2,s4,z3,s3,z4,s2,u,Degrees=>{9,7,5,3,4,6,8,5}];
M = matrix {{s1,z2,u,s2,z4},{z2,s4,z3,z4,s3}};
I = minors(2,M) + ideal (s4*s3-z3^3,z2*s3-z3^2*u,z2*z4-z3*u^2,s1*z4-u^3);
needsPackage "VersalDeformations"
T1 = cotangentCohomology1(-7,8,I)
T2 = cotangentCohomology2(I)
(F,R,G,C) = versalDeformation(gens(I),T1,T2);
G
\end{verbatim}
The output \verb|T1| describes how the equations of $V \hookrightarrow \AA^8$ are perturbed, at the first order, by the coordinates $t_1, \dots, t_9$ of the interesting part of $T^1_V$. From these perturbations one can compute the degrees of these coordinates and discover the following conversion table between our notation and the output of Macaulay2.
\[
\begin{array}{ccccccccc}
t_1 & t_2 & t_3 & t_4 & t_5 & t_6 & t_7 & t_8 & t_9 \\
t_{m^2_{3,1}} & t_{m^2_{3,2}} & t_{m^2_{2,2}} & t_{m^2_{2,1}} & t_{m^1_{3,1}} & t_{m^1_{3,2}} & t_{m^1_{2,2}} & t_{m^1_{2,1}} & t_u
\end{array}\]
The output \verb|G| describes the miniversal deformation space of $V$ with degrees between $-7$ and $8$, i.e.\ the germ $G$ we wanted to study. This implies that $G$ is the germ at the origin of the closed subscheme of $\AA^9$ defined by the following equations:
\begin{gather*}
t_5 t_9 = 0, \\
t_1 t_9 = 0, \\
t_4t_5+t_1t_8 = 0, \\
t_1 t_5 = 0, \\
t_2 t_5^2 - t_1^2 t_6 = 0.
\end{gather*}
These equations are those in Proposition~\ref{pro:eq}.
The output \verb|F| gives the equations of the deformation of $V$ over the germ $G$.
\begin{rem}
The equations of the germ $G$ are only well defined up to a homogeneous change of coordinates whose jacobian is the identity.
In particular, the quadratic terms of these equations are well defined and can be computed by analysing the cup product $T^1_V \otimes T^1_V \to T^2_V$: this can be done via toric methods \cite{fil, fil2}.
\end{rem}
|
\section{Introduction}\label{sec:introduction}}
\IEEEPARstart{T}{he} interconnection of electronic sensors and actuators, known as the Internet of Things (IoT)~\cite{iot}, is creating enormous opportunities for automating systems around us and improving their efficiency. It is paving the way for the development of smart cities with active monitoring and control of public facilities, smart healthcare, smart transit systems, etc. In recent years, due to the ubiquity of the internet, there has been an increasing trend of offloading computing, control, and storage to the cloud~\cite{comput_offloading}. This is fueling the rapid growth of the IoT as it reduces the physical cost of sensors and actuators. Moreover, connectivity to the cloud opens up endless possibilities for powerful and revolutionary applications, due to the availability of massive computational power and data~\cite{promise_edge_comput}.
Therefore, cloud computing~\cite{Cloud_computing} is now becoming an integral part of the IoT ecosystem particularly for applications involving realtime analytics and Big data~\cite{big_data_iot}.
There is a wide variety of cloud enabled IoT applications that have different data processing needs. For instance, autonomous vehicles on the roads might require information about the shortest available route to the destination. On the other hand, home users might be controlling appliances remotely based on the information obtained from deployed sensors. Some of the IoT applications are highly delay-sensitive, e.g., real time systems such as those involving artificial intelligence (AI), virtual and augmented reality (VR/AR), real time control loops, streaming analytics, etc~\cite{assessment_suitability}. Such applications are referred to as \emph{mission-critical}~\cite{mc_definition} not only due to conventional `life risk' interpretation but also pertaining to the risks of public services interruption, perturbing public order, jeopardizing enterprise operation and causing losses to businesses, etc.
\begin{figure}[t!]
\centering
\includegraphics[width=3.6in]{max_latency.png}
\caption{Tolerable end-to-end delays of some typical mission-critical IoT applications.}\label{max_latency_fig}
\end{figure}
A common characteristic of such applications is the extremely high delay sensitivity. The total delay in the response of a computational request to the cloud server, also referred to as \emph{end-to-end delay}, relies on several factors such as the latency\footnote{We use the term `latency' to refer to the delay caused due to the physical distance that is traveled by the data.}, i.e., round trip time (RTT), in addition to processing time of the computational tasks and the transmission time over the air interface. Fig.~\ref{max_latency_fig} shows the maximum tolerable end-to-end delay of some typical mission-critical IoT (MC-IoT) applications~\cite{latency_report}. It can be observed that most of the applications require an end-to-end delay of less than 10 ms. Some of them are extremely delay sensitive, requiring an end-to-end delay of 1 ms or less, such as VR and real-time industrial control applications. Although the existing cloud servers have the computational power to efficiently compute large amounts of data, the location of the server places a bottleneck on the latency. In other words, a certain time delay is inevitable regardless of the size of the task due to the distance the data has to travel before reaching back to its point of origin. Instead of sending the tremendous amounts of data, generated by the IoT to the cloud, it is more efficient if the data is analyzed at the edge of the network, i.e., close to where it is generated, to reduce the latency~\cite{minimizing_delay_iot}. Hence, a new computing architecture known as fog computing~\cite{fog_iot_cisco}, also referred to as edge computing~\cite{edge_computing_vision}, is now gaining significant attention.
\subsection{Background \& Motivation}
Fog computing is an extension of the cloud such that there are devices located at the edge of the network having computing, storage, and networking capabilities, also referred to as cloudlets or \emph{fog nodes}~\cite{fog_computing_cisco_original}. Due to the reduced distance, the availability of fog nodes can significantly reduce the response time of cloud server to incoming data~\cite{fog_and_iot}.
Hence, fog computing is emerging as one of the key enablers of fog-enabled MC-IoT~\cite{mission_critical} applications.
A cloud service provider (CSP) may have several available fog nodes in addition to the main cloud server for servicing computational requests by MC-IoT applications~\cite{hierarchical_edge_cloud_mobile}. An illustration of the hierarchical fog-cloud architecture~\cite{hierarchical_fog_cloud} for an IoT ecosystem is provided in Fig.~\ref{sys_model_fig}. Each fog node imposes a certain delay in the response to requests coming from a certain geographical region due to its location. Furthermore, the fog nodes may have multiple virtual machine instances (VMIs), similar to the conventional cloud computing, that process the incoming data. The VMIs have different data processing capabilities according to the allocated computing resources, which results in different processing delays. Altogether, the CSP has a set of available VMIs which are characterized by their overall response time. The MC-IoT applications have a varying level of delay tolerance, i.e., the maximum delay in the response of the cloud which does not result in degradation of performance in their operation. Once an application requests the cloud for processing a computational task, the job of the CSP is to instantly make a decision of sending the received data to one of the available VMI. Since the requests by MC-IoT applications arrive at random times and have different levels of delay tolerances, the CSP needs to devise a policy according to which the VMIs are allocated and appropriately price them to maximize its revenue.
\begin{figure}[t]
\centering
\includegraphics[width=3in]{system_model_pro}
\caption{Architecture of a fog-enabled IoT ecosystem in which the MC-IoT devices are connected to the CSP via a gateway. The CSP has several fog nodes, equipped with a range of VMIs, in addition to the main cloud server.}\label{sys_model_fig}
\end{figure}
Since the number available VMIs are limited, it is important to forward the most delay sensitive applications to the best\footnote{The notion of `best' refers to the VMI that offers the lowest end-to-end delay.} available VMI and vice versa to deliver a high quality of experience (QoE) to the users\footnote{The term `users' and `clients' are interchangeably used to refer to the client MC-IoT applications that are using the cloud for data processing.} and consequently generate higher revenue.
Note that improving the QoE of client applications enables the CSP to appropriately charge them for premium services. This implies that a highly delay sensitive applications should be allocated to a low end-to-end delay providing VMI and charged higher prices while delay tolerant applications should be allocated longer end-to-end delay providing VMIs at lower prices. However, the challenge lies in the fact that there is limited information about the nature of upcoming requests in the future.
A highly delay sensitive application may not request for service while the CSP reserves the best available VMIs, which leads to an inefficient utilization of resources resulting in lower QoE of the users. On the other hand, if the delay tolerant applications are allocated the best VMIs, then highly delay sensitive applications may request service in the future and may suffer in performance due to the unavailability of VMIs with low end-to-end delay. Secondly, as the time progresses, if the best available VMIs are not utilized while waiting for highly delay sensitive applications, the CSP may loose the opportunity of generating revenues from them at all and it may have been better to allocate them to less delay sensitive applications. Therefore, there is a need for a dynamically efficient policy that takes these trade-offs into account. We use the following example using the simplest case to elaborate the concept.
\begin{figure}[h]
\vspace{-0.0in}
\centering
\includegraphics[width=3in]{single_revenue_new}
\caption{Expected revenue of the CSP for a single available VMI with varying qualification threshold. \vspace{-0.0in}}\label{single_revenue}
\end{figure}
\begin{example*}
Consider a CSP with only a single VMI available for allocation within a time period of 10 hours. Requests arrive sequentially at the CSP according to a Poisson process with an arrival rate of $\lambda = 20$ requests per hour. Each arriving request is assumed to have a random minimum required response rate (exponentially distributed with mean $\alpha = 1$). It implies that there is a high probability of delay tolerant applications arriving while there is a low probability of delay sensitive applications requesting service. The CSP sets a threshold specifying the barrier on the required response rate beyond which the first arriving request will be allocated to the VMI resulting in an efficiency equal to the product of the barrier and the response rate of the VMI. Consequently, a price equivalent to the qualification threshold is charged. The expected revenue of the CSP is then the product of the barrier and the probability of a qualifying request arriving.
If the barrier is set too low, the probability of a qualifying request within the allocation period will be high but the revenue generated will be low. Similarly, if the bar is set too high, then the probability of an eligible request will be low but the revenue generated if successful allocation takes place will be high. Fig.~\ref{single_revenue} shows the expected revenue against the qualification threshold set by the CSP. It is clear that there exists an optimal qualification threshold which results in maximum revenue of the CSP under uncertainty of arriving requests.
\end{example*}
Note that the example uses a static barrier to filter out computational requests. However, since the allocation period is finite, the threshold needs to be dynamic, i.e., changing with time, in order to maximize the generated revenue. Therefore, there is a need to develop an integrated policy framework that uses the QoE of the users as a basis to maximize the expected revenue generated by the CSP.
\vspace{-0.0in}
\subsection{Related Work}
\textcolor{black}{Efficient resource management in the context of cloud computing for IoT applications is currently an active area of research~\cite{resource_management_iot}. Several different objectives have been targeted for workload offloading to the fog-cloud architecture~\cite{survey_resource_management_cloud}. For instance, in~\cite{service_allocation_fog_cloud}, a framework for allocation of combined fog-cloud resources for IoT services is proposed to minimize the latency experienced by the services. On the other hand, workload allocation in fog-cloud computing for balanced delay and power consumption is provided in~\cite{workload_allocation}. Similarly, the works in~\cite{ansari_fog},~\cite{ansari_fog2}, and~\cite{ansari2} provide useful frameworks for resource provisioning in fog based systems for IoT networks. QoE has recently become an important concern in the context of cloud services~\cite{qoe_in_cloud} as opposed to the quality of service~\cite{qos_iot_fog}. It is directly linked to the latency experienced by the applications. Latency is considered as one of the biggest hurdles in traditional cloud computing for real-time and MC applications~\cite{latency_issue}. The edge-centric computing is by far the most promising approach to reducing latency~\cite{edge_centric_computing}. However, the fog nodes have limited available computational resources which need to be efficiently allocated to arriving tasks.
\textcolor{black}{Resource dispatch and scheduling in cloud systems has been well studied in the literature~\cite{job_scheduling}. The research on allocation and pricing of cloud/fog resources can be broadly categorized into two main directions, i.e., static and dynamic. Both directions have been driven by several different objectives such as QoS, welfare, and revenue maximization, etc. Static allocation refers to the case where jobs are submitted and the CSP decides the allocation of resources based on the current set of jobs running on the cloud.
Often, theoretical tools from queuing theory and optimization are leveraged to develop a scheduling and allocation policy with a specified objective.
Sometimes the decisions are made on a real-time basis, referred to as spot pricing and where the prices are set as the demand changes. However, the revenue maximization is not guaranteed over a certain time horizon. For instance, the allocation of resources for spot markets in cloud computing environments has been explored in~\cite{spot_markets}.
Most such works consider a matching problem between requests and virtual machines, which is not a sequential problem as complete information is available at the time of decision.}
\textcolor{black}{The idea of dynamic pricing and revenue maximization in the presence of stochastic demand has also been investigated in the context of cloud computing~\cite{cloud_pricing}. However, it does not use price discrimination for differentiated services offered by the cloud which is imperative in the case of fog-enabled MC-IoT. Revenue maximization if often accomplished by the use of auctions to extract the maximum revenue under hidden type of the applications~~\cite{mechanism_design_cloud}. However, auctions are not instantaneous, even though they may be implemented in real-time~\cite{static_auction_fogspot}, \cite{bid_the_cloud} \cite{auction_based}. Revenue maximization in a dynamic setting is sparingly investigated in relevant literature due to its increased complexity~\cite{online_job_dispatching}. However, they do not focus on the MC-IoT applications and the revenue maximization aspect. Those that consider real-time decision making, do not consider a finite time horizon to adapt the decision according to the remaining allocation time. This is why we resort to statistical analysis
to dealing with the cloud resource provisioning and pricing decisions for real-time and mission-critical needs of the applications. The pricing and allocation is based on the QoE provided to the users, which in turn is based on the delay tolerance of the applications as well as the response rates of the VMIs. In this work, we consider the case where the dynamics of the allocation process are taken into account. In other words, the sequential arrival, instantaneous allocation, incomplete information, and allocation deadline are the salient features that distinguish this work from the literature.
}
In the following subsection, we highlight the key contributions of this paper in comparison with the existing literature.
\subsection{Contributions}
Most existing work on job dispatching and scheduling is focused on the static assignment of tasks to computational resources in the fog-cloud environment. For instance, a task offloading framework using matching theory has been proposed in~\cite{matching_theory}. However, it is based on static pairing between tasks and fog nodes and the dynamic aspects of task arrival are not considered. Similarly, an index based task assignment and scheduling of tasks is proposed in~\cite{index_task_assignment}. Several works do consider real-time task allocation and dispatch. For instance, an online job dispatching and scheduling algorithm is proposed in~\cite{online_job_dispatching}, where jobs are released in arbitrary order and times by mobile devices and offloaded to unrelated servers with both upload and download delays. However, the approach is based on a greedy algorithm and is not strategic, i.e., does not make use of the statistical information about computational requests to make more effective allocation decisions. Furthermore, pricing and revenue maximization has been looked into sparingly.
In this paper, we present a revenue maximizing perspective towards allocation and pricing in fog based systems designed for mission critical IoT applications. The QoE resulting from the pairing of fog resources with computation requests is used as a basis for pricing. We develop a dynamic policy framework leveraging the literature in economics, mechanism design~\cite{revenue_maximization}, and dynamic revenue maximization~\cite{revenue_book} to provide an implementable mechanism for dynamic allocation and pricing of sequentially arriving IoT requests that maximizes the expected revenue of the CSP.
The developed optimal policy framework assists in both determining which fog node to allocate an incoming task to and the price that should be charged for it for revenue maximization. The proposed policy is statistically optimal, dynamic, i.e., adapts with time, i.e., and is implementable in real-time as opposed to other static matching schemes in the literature. The dynamically optimal solution can be computed offline and implemented in real-time for sequentially arriving computation requests.
The rest of the paper is organized as follows: Section~\ref{Sec:Model} provides a description of the system model, Section~\ref{Sec:Main_Results} provides the details of the dynamically optimal mechanism for QoE based revenue maximization framework along with the main results, and Section~\ref{Sec:Numerical_Experiments} presents the results of numerical experiments with discussions. Finally, Section~\ref{Sec:Conclusion} concludes the paper.
\section{Model Description}\label{Sec:Model}
We consider a CSP in a fog enabled IoT ecosystem having a set of $k$ available fog nodes in addition to the main cloud server and serving a certain geographical region containing MC-IoT devices. The fog nodes have an associated average latency denoted by $l_i \in \mathbb{R}, i \in \{1, \ldots, k\}$ which depends on the distances of the fog nodes from the locations of origin of the processing requests. The number of available VMIs at the fog nodes is denoted by $n_i, i \in \{1, \ldots, k\}$ and each of the available VMI is characterized by its processing delay for a fixed number of computational operations denoted by $\tau_{ij}^{(p)}, i = \{1, \ldots, k\}, j \in \{1, \ldots, n_i\}$. There are a total of $\sum_{i=1}^{k}n_i = N$ VMIs available for allocation by the CSP to sequentially arriving MC-IoT requests. The latency of each individual VMI is denoted by $\tau^{(l)}_{ij} = l_i, \forall j = \{1, \ldots, n_i\}$. The end-to-end delay offered by the VMIs can be evaluated as $\tau_{ij}^{(l)} + \tau_{ij}^{(p)} + \tau^{(o)}$, where $\tau^{(o)}$ represents the other delays including the transmission delay over the air interfaces. Consequently, the average response rate of the $j^{\text{th}}$ VMI at compute node $i$ can be expressed as follows\footnote{Performance metrics and utility of similar forms are used in the literature for fog-enabled IoT systems, e.g.,~\cite{matching_theory,iot_service_delay}.}:
\begin{align}
r_{ij} = \frac{1}{\tau_{ij}^{(l)} + \tau_{ij}^{(p)} + \tau^{(o)}}.
\end{align}
With an abuse of notation, the average response rates of the available VMIs can further be denoted by $\mathbf{r} = [r_1, r_2, \ldots, r_N]$, where $r_1 \geq r_2 \geq \ldots \geq r_N$, and $\mathcal{M}$ defines the mapping of values of the pair $(i,j), i \in \{1, \ldots, k\}, j \in \{1, \ldots, n_i\}$ from the original $r_{ij}$ to the set $n = \{1, \ldots, N\}$ in the new response rates denoted by $r_n$. It implies that the VMI corresponding to a response rate $r_1$ is the best in terms of end-to-end delay for processing IoT data while the one corresponding to $r_N$ is the least favorable in terms of end-to-end delay.
MC-IoT devices in a given geographical region are connected to the CSP via a set of IoT gateways. We assume that requests for processing by MC-IoT applications arrive sequentially at the CSP according to a homogeneous Poisson process with intensity $\lambda$. Each request is characterized by its maximum tolerable delay for successful operation denoted by $d_i, i \geq 1$. In other words, there is a minimum required response rate by the applications. Upon arrival of a request by an MC-IoT application, the CSP has to make a decision to forward the data to one of available VMIs at the fog nodes. An illustration of the system model is provided in Fig.~\ref{sys_model_fig}.
It is pertinent to mention that before the actual task offloading, the task metadata is sent to the cloud server for the allocation decision. Furthermore, we assume that the allocation decision is instantaneous and the delay in transmitting metadata is not relevant in the computation of delay experienced in the transmission and processing of actual application data since it occurs only once before the allocation is made.
The allocation is made for long term access of the resources by the applications. The applications can release the resources anytime based on their requirements. Once the resources are returned, they can be reallocated to other incoming applications. The proposed allocation framework allows this since the solution is dynamically optimal. In other words, it depends on the state, which in this case is the number of VMIs available. Therefore, if the optimal policies have been precomputed for multiple available VMs, then the policy can be switched in real-time and it would still be optimal for the remaining time available.
\subsection{Allocation Efficiency and Quality of Experience}
Upon arrival of a request by an MC-IoT application, the CSP determines the sensitivity of the application to a delay in the response and is subsequently required to allocates it to one of the available VMIs at the fog nodes while charging a particular price. The VMIs are allocated to requesting applications for long term and therefore, once a VMI is allocated to an application, it becomes unavailable for allocation to other applications in the future\footnote{Note that this assumption is only made for obtaining an optimal dynamic policy using an open loop methodology. In practical implementation, the policy can be adapted based on how many VMIs are available.}. Since the VMIs are perishable, i.e., they have no value if not successfully allocated within the allocation time horizon $T$, so it is important for the CSP to optimally allocate the available resources within a certain time frame\footnote{The time horizon refers to the period over which the allocation has to occur which can be related to the demand window.}. It also motivates the idea of dynamic pricing, i.e., to charge higher prices earlier and reduce them as time goes on for maximizing the revenue of the CSP.
Let $x_i = d_i^{-1}$ be the minimum required response rate by the $i^{\text{th}}$ application.
Each application reports this characteristic to the CSP at the time of requesting service. We assume that each $x_i \in \mathbb{R}, i \geq 1,$ is an independent and identically distributed (i.i.d.) random variable with a probability density function (pdf) denoted by $f_X(x)$ and a cumulative distribution function (cdf) denoted by $F_X(x)$\footnote{We assume that the probability distribution of the arriving tasks is known a priori. This is done for analytical tractability and policy development as data-driven approaches are prohibitive in terms of obtaining an implementable dynamic policy.}.
The product $x_ir_j$ can be used as a measure of efficiency when the $i^{\text{th}}$ application is allocated to the $j^{\text{th}}$ available VMI. The QoE as a result of the pairing, denoted by $\Phi(x_i,r_j)$ is quantified as follows:
\begin{align}
\Phi(x_i,r_j) = (x_i r_j)^{\frac{1}{\eta}}, \ \ i,j \in \{1, \ldots, N\},
\end{align}
where $\eta \geq 1$ is a constant controlling the rate of increase in QoE with respect to the efficiency of allocation. The concave nature of the QoE function implies that an increase in allocation efficiency results in diminishing improvements in the QoE of the applications. \textcolor{black}{Note that the choice of utility function is not exclusive and any arbitrary function with similar structural properties, i.e., concave and increasing in both the average response rate and the minimum required rate of the applications , can be used. However, the choice here is made for simplicity and tractability of analysis inspired by the Cobb-Douglas function from econometrics~\cite{cobb_douglas} that models the production output relationship between the amounts of two or more inputs.}
In this section, we provide a framework for dynamically maximizing the expected revenue of the CSP based on the QoE of the users. We focus on a mechanism design approach, which is based on developing optimal implementable policies for allocation and pricing. A direct mechanism is provided whereby each requesting application reports its minimum required response rate and the CSP allocates one of the available VMIs to it. For the optimal allocation to be implementable, an allocation policy and a payment rule is required that is incentive compatible\footnote{Incentive compatibility is a concept from mechanism design theory that ensures that no agent has an incentive to misreport its privately known characteristic.}~\cite{Myerson} in the presence of individually rational users. We first state an allocation rule which satisfies the aforementioned conditions and then provide a pricing strategy that subsequently implements the allocation.
\vspace{-0.0in}
\subsection{Allocation Policy}\label{Sec:Allocation}
Based on the required response rate of each randomly arriving computational request by the IoT devices, the CSP has to allocate a VMI at one of the fog nodes to maximize the expected QoE of the users. It is clear that an application that requires a high response rate should be allocated to the VMI offering a high response rate and vice versa for efficient utilization of resources. However, the problem rests in the fact that an allocation decision has to be made without knowledge of the type of applications that will request in the future. It is shown in the literature~\cite{albright} that a dynamically efficient allocation policy in such situations can be achieved using a partition on the characteristic of the sequentially arriving agent. We provide a deterministic and Markovian allocation policy $\pi_t(x, \mathbf{r}_t)\footnote{Throughout the paper, the subscript $t$ is used to refer to the time dependence.}: \mathbb{R} \rightarrow \mathbf{y}_t$, i.e., at each time $t$, a fixed non-random policy is used that only depends on the current time instant and set of available VMIs at time $t$.
The key result in the allocation policy is provided by the following theorem.
\begin{lemma}[Adapted from~\cite{albright}] \label{theorem_allocation}
A deterministic and Markovian policy at time $t$, i.e., $\pi_t(x, \mathbf{r}_t)$ is implementable iff there exists a set of functions $y_{i}(t)$, $i = 1, \ldots, N_t$, such that $0 < y_{N_t}(t) < y_{N_t-1}(t) < \ldots < y_{1}(t) < y_{0}(t) = \infty$. The allocation policy is such that $\pi_t(x, \mathbf{r}_t) = r_i$ if $x \in [y_{i}(t), y_{i-1}(t)]$ and $\pi_t(x, \mathbf{r}_t) = \emptyset$ if $x < y_{N_t}(t)$.
\end{lemma}
The allocation policy depends on the nature of the requesting applications as well as the available VMIs at time $t$. Lemma~\ref{theorem_allocation} implies that if at time $t$, the requesting application has a required response rate lower than $y_{N_t}(t)$, then the CSP will not allocate any compute node to the application as it aspires to save the VMIs for higher valued application requests in the future. However, if the required response rate is between $y_{i}(t)$ and $y_{i - 1}(t)$, then the $i^{th}$ best available VMI is allocated to the requesting application. It can be observed that the number of partitions or cutoff curves depends on the set of available compute nodes at time $t$.
In the subsequent subsections, we provide a suitable pricing scheme associated with the aforementioned policy and provide the optimal time varying cutoff values for efficient allocation.
\subsection{Pricing Policy}
For the proposed allocation policy, there is a need to appropriately price the applications for their achieved QoE. We assume that the MC-IoT applications are individually rational, i.e., no application will be willing pay more than the QoE it achieves by using the allocated VMI. The partition based allocation policy described in Section~\ref{Sec:Allocation} provides a natural mechanism for pricing the applications for their proposed allocations. Since an application with a required response rate $x \in [y_i(t), y_{i-1}(t)]$ at time $t$ is allocated to a VMI with a response rate $r_i$, it has achieved an improvement in QoE of atleast $(r_i^{\frac{1}{\eta}} - r_{i+1}^{\frac{1}{\eta}})y_i^{\frac{1}{\eta}}(t)$ as compared to the next best allocation. Therefore, it must pay an equivalent price to be allocated to a VMI with response rate $r_i$ as compared to the one with $r_{i+1}$. This process can be continued recursively to obtain implementable prices for each of the available VMIs. It has been shown in literature~\cite{albright} that such incentive compatible pricing structure in which the price is selected on the basis of displaced value is implementable. Therefore, the optimal prices can be completely determined by the implementation conditions. The price charged to an MC-IoT application that is allocated to a $j^{\text{th}}$ best VMI, i.e., $x \in \left[ y_{j}(t), y_{j-1}(t) \right)$, at time $t$ can be expressed as follows:
\begin{align}
P_j(\mathbf{r}_t,t) = \sum_{i = j}^{N_t} \left( r_i^{\frac{1}{\eta}} - r_{i+1}^{\frac{1}{\eta}} \right) y_{i}^{\frac{1}{\eta}}(t),
\end{align}
Note that $r_i > r_{i+1}, \forall i = 1, \ldots, N-1$ due to the initial ordering. The pricing policy is progressive in a relative sense with the lowest QoE achieved as a reference.
We provide a simple example to further elaborate the pricing policy. If there is a single VMI available, the price would be $P_1(\{r_1\},t) = r_1^{\frac{1}{\eta}} y_1^{\frac{1}{\eta}}(t)$, which is equivalent to the QoE of the application to which the VMI was allocated. However, if there are two VMIs avilable, the prices for the lower response rate VMI is set to be $P_2 ( \{ r_1,r_2\}, t ) = r_2^{\frac{1}{\eta}} y_2^{\frac{1}{\eta}}(t)$ and the price for the higher response rate VMI is set to be $P_1(\{ r_1,r_2\}, t ) = r_2^{\frac{1}{\eta}} y_2^{\frac{1}{\eta}}(t) + (r_1^{\frac{1}{\eta}} - r_2^{\frac{1}{\eta}}) y_1^{\frac{1}{\eta}}(t)$ Note that the price for the latter is simply the price of the former plus the improvement in QoE experienced by getting a VMI with response rate $r_1$ instead of $r_2$. In other words, the lowest response rate VMI is priced at the base price equivalent to the QoE achieved by its allocation. The next higher one is priced at the base price plus the improvement in QoE achieved as a result of being allocated a better available VMI.
\section{Main Results}\label{Sec:Main_Results}
In this section, the goal is to maximize the expected revenue generated by the CSP using the pricing strategy developed in Section~\ref{Sec:Model}. We first begin with solving the problem for the case of a single available VMI and then generalize it to multiple VMIs using a recursive approach.
\subsection{Single VMI Case}
If only a single VMI is available to the CSP with a response rate $r_1$, then the expected revenue generated by its allocation within a time period $T$ can be expressed as follows:
\begin{align}
R(\{r_1\},t) &= \int_t^T P_1(\{r_1\},t) h_{1}(s) ds = r_1^{\frac{1}{\eta}} \int_t^T y_{1}^{\frac{1}{\eta}}(s) h_{1}(s) ds, \label{revenue_single_expr}
\end{align}
where $h_1(s)$ is the probability density of waiting time until the first arrival of a request with a required response rate of greater than $y_1(s)$.
The objective of the CSP is to determine the optimal allocation threshold $y_1^{*}(t)$ such that the expected revenue is maximized, i.e.,
\begin{align}
y_1^{*}(t) = \arg \max r_1^{\frac{1}{\eta}} \int_t^T y_{1}^{\frac{1}{\eta}}(s) h_{1}(s) ds.
\end{align}
Note that the revenue expression in~\eqref{revenue_single_expr} involves an integral of a quasi-concave function since for a fixed time $t \geq 0$, the function $y_1^{\frac{1}{\eta}}(t)$ is monotonically increasing while $h_1(t)$ is monotonically decreasing in $y_1(t)$. Hence, a unique maxima can be obtained for every $t$. However, finding the optimal function requires a variational approach.
The optimal dynamic threshold for the case of a single VMI can be determined using the following theorem.
\begin{theorem} \label{single_theorem}
If the CSP has a single available VMI at time $t$, then it is optimal to allocate it to a requesting application if the required delay tolerance $x_i \geq y_1^*(t)$, where the optimal threshold $y_1^*(t)$ solves the following equation:
\begin{align}
y_1^*(t) =& \left( \frac{1 - F_{\hat{X}}((y_1^*)^{\frac{1}{\eta}}(t))}{f_{\hat{X}}((y_1^*)^{\frac{1}{\eta}}(t))} + \right. \notag \\ &\left. \lambda \int_t^T \frac{(1 - F_{\hat{X}}((y_1^*)^{\frac{1}{\eta}}(s)))^2}{f_{\hat{X}}((y_1^*)^{\frac{1}{\eta}}(s))} ds \right)^{\eta}
\end{align}
\begin{proof}
See \textbf{Appendix~\ref{proof_single_theorem}}.
\end{proof}
\end{theorem}
In the subsequent subsection, we extend the approach to the case of multiple available VMIs.
\subsection{Multiple VMI Case} \label{Sec:cutoff}
In this subsection, we first present the case of two available VMIs at the CSP and then generalize it to the case of multiple available VMIs. If there are only two available VMIs with response rates $r_1$ and $r_2$, then the expected revenue of the CSP can be expressed as follows:
\begin{align}
R(\{r_1,r_2\},t) &= \int_0^T \left( P_2(\{r_1,r_2\}t) + R(\{r_1\},t) \right) h_2(t) dt + \notag \\ &\int_0^T \left( P_1(\{r_1,r_2\}t) + R(\{r_2\},t) \right) h_1(t) dt,\notag \\
& = \int_0^T \left( r_2^{\frac{1}{\eta}} y_2^{\frac{1}{\eta}} + R(r_1,t)\right) \times \notag \\ &\lambda (1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(t))) e^{-\int_0^t \lambda (1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(s)))ds } dt + \notag \\ &(r_1^{\frac{1}{\eta}} - r_2^{\frac{1}{\eta}}) \int_0^T \left( y_1^{\frac{1}{\eta}}(t) - R(1,t) \right) \times \notag \\
&\lambda (1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))) e^{ - \int_0^t \lambda (1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(s)))ds} dt.
\end{align}
where $h_1(t)$ represents the density of waiting time till the first arrival of a request with a required response rate of atleast $y_1(t)$ if no request with a required response rate in the interval $[y_2(t), y_1(t))$ has arrived. Similarly, $h_2(t)$ represents the density of waiting time till the first arrival of a request with a required response rate in the interval $[y_2(t), y_1(t))$ if no request with a required response rate in the interval $[y_1(t), \infty )$ has arrived. The optimization problem in this case becomes the following.
\begin{align}
y_2^*(t) &= \arg \max \int_0^T \left( r_2^{\frac{1}{\eta}} (y_2^*)^{\frac{1}{\eta}} + R(r_1,t)\right) \times \notag \\ &\lambda (1 - F_{\hat{X}}((y_2^*)^{\frac{1}{\eta}}(t))) e^{-\int_0^t \lambda (1 - F_{\hat{X}}((y_2^*)^{\frac{1}{\eta}}(s)))ds } dt + \notag \\ &(r_1^{\frac{1}{\eta}} - r_2^{\frac{1}{\eta}}) \int_0^T \left( (y_1^*)^{\frac{1}{\eta}}(t) - R(1,t) \right) \times \notag \\
&\lambda (1 - F_{\hat{X}}((y_1^*)^{\frac{1}{\eta}}(t))) e^{ - \int_0^t \lambda (1 - F_{\hat{X}}((y_2^*)^{\frac{1}{\eta}}(s)))ds} dt.
\end{align}
The optimal dynamic threshold for the case of two available VMIs can be determined using the following theorem.
\begin{theorem}
If the CSP has a two available VMIs at time $t$, then it is optimal to allocate the low response rate VMI to a requesting application if the required delay tolerance $x_i \in [y_2^*(t), y_1^*(t))$, and to allocate the high response rate VMI if the required delay tolerance $x_i \geq y_2^*(t)$ where the optimal threshold $y_2^*(t)$ solves the following equation:
\begin{align}
y_2^*(t) = \Bigg( \frac{1 - F_{\hat{X}}((y_2^*)^{\frac{1}{\eta}}(t))}{F_{\hat{X}}((y_2^*)^{\frac{1}{\eta}}(t))} + &\lambda \int_t^T \frac{(1 - F_{\hat{X}}((y_2^*)^{\frac{1}{\eta}}(s)))^2}{F_{\hat{X}}((y_2^*)^{\frac{1}{\eta}}(s))} ds \notag \\ & - R(1,t) \Bigg)^{\eta}.
\end{align}
\begin{proof}
See~\textbf{Appendix~\ref{proof_allocation_theorem}}.
\end{proof}
\end{theorem}
Note that the optimal threshold $y_2^*(t)$ relies on obtaining the threshold $y_1^*(t)$. In the general case, it can be shown by induction that the optimal thresholds can be obtained recursively using the following theorem:
\begin{theorem} \label{general_theorem}
If there are $N_t$ available VMIs at time $t$, then it is optimal to allocated a VMI with response rate $r_i$ to an incoming request with minimum required response rate $x_i \in [y_i^*(t), y_{i-1}^*(t)]$, where the optimal dynamic thresholds $y_i^*(t)$
satisfy the following recursive equation:
\begin{align}\label{eq_y_general}
y_i^*(t) = & \Bigg( \frac{1 - F_{\hat{X}}(((y_i^*)^{\frac{1}{\eta}})(t))}{f_X(((y_i^*)^{\frac{1}{\eta}})(t))} + \notag \\ & \lambda \int_t^T \frac{(1 - F_{\hat{X}}(((y_{i-1}^*)^{\frac{1}{\eta}})(s)))^2}{f_X(((y_{i-1}^*)^{\frac{1}{\eta}})(s))} ds \ \ - \notag \\ & \lambda \int_t^T \frac{(1 - F_{\hat{X}}(((y_{i}^*)^{\frac{1}{\eta}})(s)))^2}{f_X(((y_{i}^*)^{\frac{1}{\eta}})(s))} ds \Bigg)^{\eta}, i = 2, \ldots, N_t.
\end{align}
\begin{proof}
See~\textbf{Appendix~\ref{proof_general_theorem}}.
\end{proof}
\end{theorem}
Note that the optimal cutoff curves are independent of the response rates of the available VMIs. In fact, they depend only on the number of available VMIs at time $t$ and on the statistical information about the sequentially arriving computational requests by MC-IoT applications. In the following set of corollaries, we provide the results obtained for special cases of the statistical information about arriving requests.
\begin{corollary}\label{corol_exp}
\label{exponential_derivation}
Assuming that the transformed required response rate of sequentially arriving MC-IoT applications, denoted by $\hat{X}$ (See Appendix~\ref{single_theorem}, follows an exponential distribution with a mean of $\alpha^{-1}$, i.e., $f_{\hat{X}}(x) = \alpha e^{-\alpha x}$, and $F_{\hat{X}}(x) = 1 - e^{- \alpha x}$, then the optimal cutoff curves for allocation can be expressed as follows:
\begin{align}
y_1^*(t) = & \frac{1}{\alpha^{\eta}} \left[1 + \log \left( 1 + \frac{\lambda (T - t)}{e} \right) \right]^{\eta},\\
y_2^*(t) = & \frac{1}{\alpha^{\eta}} \left[1 + \log \left( 1 + \frac{\lambda^2 (T - t)^2}{2e \left(\lambda (T - t) + \mathrm{e} \right)} \right) \right]^{\eta}, \\
y_3^*(t) = &\frac{1}{\alpha^{\eta}} \hspace{-0.05cm} \left[1\hspace{-0.05cm} + \hspace{-0.05cm} \log \hspace{-0.1cm}\left( \hspace{-0.05cm}1 \hspace{-0.05cm}+ \hspace{-0.05cm}\frac{\lambda^3 (T - t)^3 }{3e \left( \lambda^2 (T - t)^2 + 2 \mathrm{e}( \lambda (T - t) + e ) \right)}\hspace{-0.1cm} \right) \hspace{-0.05cm} \right]^{\eta}\hspace{-0.1cm},
\end{align}
The remaining lower cutoff curves cannot be easily derived in closed form and can be computed numerically from (10) using the Picard iterative process~\cite{Picard}.
\end{corollary}
\begin{corollary}
Assuming that the transformed required response rate of sequentially arriving MC-IoT applications, denoted by $\hat{X}$ (See Appendix~\ref{single_theorem}), is uniformly distributed in the interval $[0, \beta]$, i.e., $f_{\hat{X}}(x) = \frac{1}{\beta}$ and $F_{\hat{X}}(x) = \frac{x - \beta }{\beta}, x \in [0, \beta]$, then the optimal cutoff curves for the allocation of the best available VMI can be expressed as follows: \vspace{-0.0in}
\begin{align}
y_1^*(t) = \beta^{\eta}\left(1 - \frac{2}{\lambda (T - t) + 4} \right)^{\eta}.
\end{align}
The lower cutoff curves cannot be easily obtained analytically and thus require numerical methods such as the Picard iterative process~\cite{Picard}.
\end{corollary}
\subsection{Expected Revenue}
\begin{figure*}[t]
\centering
\includegraphics[width=4.5in]{flow}
\caption{Resource Allocation and Pricing Flow Diagram.}\label{flow_fig}
\end{figure*}
The expected revenue from the optimal allocation of VMIs to sequentially arriving MC-IoT requests can be expressed by the following theorem:
\begin{theorem}
The expected revenue of the CSP at time $t$ if a total of $N_t$ VMIs with response rate defined by $\mathbf{r}_t = [r_1, r_2, \ldots, r_{N_t}]$ can be expressed as:
\begin{align}
\mathcal{R}(\mathbf{r}_t,t) = \sum_{i=1}^{N_t} r_i^{\frac{1}{\eta}} \left( y_i^{\frac{1}{\eta}}(t) - \frac{1 - F_{\hat{X}}(y_i^{\frac{1}{\eta}}(t))}{f_X(y_i^{\frac{1}{\eta}}(t))} \right).
\end{align}
\begin{proof}
The expected revenue if only a single VMI with unit response rate is available is given by $\mathcal{R}(\{1\},t) = \int_t^s \frac{(1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(s)))^2}{f_X(y_1^{\frac{1}{\eta}}(s))}ds = \left( y_1^{\frac{1}{\eta}}(t) - \frac{1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))}{f_X(y_1^{\frac{1}{\eta}}(t))} \right).$ Similarly, if two VMIs with unit response rates are available, then using Theorem~\ref{single_theorem}, the expected revenue is given by $\mathcal{R}(\{1,1\},t) = \int_t^s \frac{(1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(s)))^2}{f_X(y_2^{\frac{1}{\eta}}(s))}ds = \left( y_1^{\frac{1}{\eta}}(t) - \frac{1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))}{f_X(y_1^{\frac{1}{\eta}}(t))}\right) + \left( y_2^{\frac{1}{\eta}}(t) - \frac{1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(t))}{f_X(y_2^{\frac{1}{\eta}}(t))}\right).$ Using an inductive argument along with the fact that $\mathcal{R}(\{r_j\},t) = r_j^{\frac{1}{\eta}} \mathcal{R}(\{1\},t)$ proves the result.
\end{proof}
\end{theorem}
Notice that the expected revenue is linear in the response rates of the available VMIs and increases if a high response rate or equivalently a low latency is provided by the VMIs at the fog nodes.
\subsection{Implementation of Dynamic VMI Allocation and Pricing}
In this section, we explain the operation of the proposed QoE based revenue maximizing dynamic allocation and pricing framework. Requests for remote computations by MC-IoT applications arrive at the CSP at random times and have a delay sensitivity which is unknown \emph{a priori}. Based on the initial number of available VMIs at the CSP $N_0$, there is a minimum cutoff threshold $y_{N_0}(t)$ which an incoming request has to cross before being allocated a VMI. If the incoming request qualifies for an allocation, the CSP needs to decide which VM to allocate to the user. A lookup table denoted by $\mathcal{T}$ is prepared by the CSP which contains pre-computed optimal dynamic cutoff curves determined in Section~\ref{Sec:cutoff}. Using this lookup table, the request is optimally classified for allocation to one of the available VMIs. The data from the MC-IoT request is forwarded to the selected VMI and a price, also available in the lookup table, is charged to the requesting application. Once the allocation has been completed, the VMI is removed from the set of available VMIs. The lookup tables are updated by removing the least cutoff threshold $y_{N_0}(t)$. The remaining available VMIs are re-arranged in descending order and their prices are updated in the lookup table $\mathcal{T}$. This process is repeated until either all the available VMIs have been successfully allocated or the allocation period has ended.
Note that the lookup tables are useful in the event that an application decides to leave the cloud. Once an application has released the allocated VMI, it becomes available again for re-allocation. Hence, the allocation policy can simply be updated by moving on to the threshold with higher number of available VMIs. This is because the solution is dynamically optimal and can be updated on the go if the state, i.e., the number of available VMIs changes. This procedure has been summarized in \textbf{Algorithm~\ref{alg1}} and the associated flow diagram is provided in Fig.~\ref{flow_fig}.
\begin{algorithm}
\caption{Dynamic VMI Allocation and Pricing}\label{alg1}
\begin{algorithmic}[1]
\Require{Initialize counter = 0, request index $i=1$, starting time $t = 0$, $\mathbf{r}_0 = \{r_i, i = 1, \ldots, N_0: r_1 \geq r_2 \geq \ldots \geq r_{N_0}\}$.}
\While{counter $< N_0$ \textbf{ and } $t < T$}
\State {\begin{tabular}{@{\hspace*{0.5em}}l@{}}Determine the required response rate\\ of the arriving computational request $x_i$.\end{tabular}}
\If{ $x_i \geq y_{N_0 - \text{counter}}(t)$}
\State {\begin{tabular}{@{\hspace*{0.5em}}l@{}}Classify the request using the lookup table, i.e.,\\ determine $j : x_i \in [y_j(t), y_{j-1}(t)]$.\end{tabular}}
\State {\begin{tabular}{@{\hspace*{0.5em}}l@{}}The $i^{\text{th}}$ requesting application is allocated to the\\ $j^{\text{th}}$ highest VMI.\end{tabular}}
\State {\begin{tabular}{@{\hspace*{0.5em}}l@{}}Use the mapping $\mathcal{M}$ to realize the allocation\\ in terms of the fog node and available VMI.\end{tabular}}
\State {\begin{tabular}{@{\hspace*{0.5em}}l@{}}Use $\mathcal{T}$ to charge a price $P_j(\mathbf{r}_t)$ to the requesting\\ application.\end{tabular}}
\State {\begin{tabular}{@{\hspace*{0.5em}}l@{}}Remove the VMI corresponding to $r_j$ from the list\\ of available VMIs.\end{tabular}}
\State {\begin{tabular}{@{\hspace*{0.5em}}l@{}}Re-arrange the set of available VMIs in \\ descending order.\end{tabular}}
\State {\begin{tabular}{@{\hspace*{0.5em}}l@{}}Update lookup table for relevant cutoff curves\\ $\mathcal{T} = \mathcal{T} \backslash y_{N_0 - \text{counter}}(t)$.\end{tabular}}
\State {\begin{tabular}{@{\hspace*{0.5em}}l@{}}Update $\mathcal{T}$ with new prices corresponding to the\\ updated list of available VMIs.
\end{tabular}}
\EndIf
$i \gets i+1$.
\EndWhile
\end{algorithmic}
\end{algorithm}
\begin{figure*}[t!]
\centering
\vspace{-0.0in}
\subfloat[]{\includegraphics[width = 3.1in]{cutoff_exp} \label{cutoff_exp_fig}} \ \
\subfloat[]{\includegraphics[width = 3.1in]{cutoff_unif} \label{cutoff_unif_fig}}
\caption{Optimal cutoff curves for (a) exponentially distributed and (b) uniformly distributed arrival characteristics.}
\label{cutoff_fig}
\end{figure*}
\begin{figure*}[t!]
\centering
\vspace{-0.0in}
\subfloat[]{\includegraphics[width = 3.1in]{exp_eta} \label{exp_eta_fig}} \ \
\subfloat[]{\includegraphics[width = 3.1in]{unif_eta} \label{unif_eta_fig}}
\caption{Optimal first cutoff curves with varying QoE parameter $\eta$ for (a) exponentially distributed and (b) uniformly distributed arrival characteristics.}
\label{cutoff_fig_eta}
\end{figure*}
\begin{figure*}[t!]
\vspace{-0.0in}
\centering
\subfloat[]{\includegraphics[width = 3.1in]{Revenue_lambda} \label{revenue_lambda_exp}} \ \ \ \
\subfloat[]{\includegraphics[width = 3.1in]{Welfare_lambda} \label{welfare_lambda_exp}}\\
\subfloat[]{\includegraphics[width = 3.1in]{Revenue_lambda_unif} \label{revenue_lambda_unif}} \ \ \ \
\subfloat[]{\includegraphics[width = 3.1in]{Welfare_lambda_unif} \label{welfare_lambda_unif}}
\caption{Effect of arrival rate on the expected revenue of the CSP ((a) and (c)) and expected QoE of the users ((b) and (d)) for an exponentially and uniformly distributed arrival characteristic respectively.}
\label{vs_lambda}
\end{figure*}
\vspace{-0.0in}
\section{Numerical Experiments \& Discussion} \label{Sec:Numerical_Experiments}
In this section, we present the results of the numerical experiments performed to evaluate the performance of our proposed allocation and pricing framework.
We first describe the simulation setup and parameter selection. We then provide the results of the simulation and study the behaviour of the framework under varying parameters. Finally, we provide a comparison of the proposed framework with some benchmark allocation strategies and discuss the insights.
\vspace{-0.0in}
\subsection{Experiment Setup}
We assume a CSP with $k=5$ fog data centers in addition to the main cloud servers (Internet data centers) serving a fixed MC-IoT installation area. Each fog data center has $n_i = 20, i = 1, \ldots, 5$, available VMIs for allocation to incoming MC-IoT requests. We have selected a small fog computing architecture in the experiments for the sake of simplicity. However, the experiments can be extended to larger topologies without loss of generality in the results.
The fog data centers are located at increasing distances from the MC-IoT devices which results in increasing latency experienced by the devices while accessing the corresponding data centers. We assume the RTT of the fog nodes experienced by the MC-IoT installation are $l_1 = 0.1$, $l_2 = 0.2$, $l_3 = 0.4$, $l_4 = 0.6$, and $l_5 = 0.8$ ms respectively. The transmission delay over the air interfaces is assumed to be fixed at $\tau^{(o)} = 0.1$ ms. The workload processing times of the VMIs, which depends on the provisioned computing resources, is considered to be uniformly distributed in the interval [0.2,1] ms in the simulations. A demand horizon of $T=12$ hours is assumed during which the allocation takes place. The constant $\eta$ in the QoE function is selected to be $\eta = 1$ for simplicity of results.
\begin{figure*}[t!]
\centering
\vspace{-0.0in}
\subfloat[]{\includegraphics[width = 3.1in]{Revenue_alpha} \label{revenue_alpha_exp}} \ \ \ \
\subfloat[]{\includegraphics[width = 3.1in]{Welfare_alpha} \label{welfare_alpha_exp}}\\
\subfloat[]{\includegraphics[width = 3.1in]{Revenue_alpha_unif} \label{revenue_alpha_unif}} \ \ \ \
\subfloat[]{\includegraphics[width = 3.1in]{Welfare_alpha_unif} \label{welfare_alpha_unif}}
\caption{Effect of the mean characteristic on the expected revenue of the CSP ((a) and (c)) and expected QoE of the users ((b) and (d)) for an exponentially and uniformly distributed arrival characteristic respectively. \vspace{-0.0in}}
\label{vs_alpha}
\end{figure*}
\begin{figure*}[h]
\centering
\vspace{-0.0in}
\subfloat[]{\includegraphics[width = 3.2in]{Number_allocated} \label{num_allocated}} \ \ \ \ \
\subfloat[]{\includegraphics[width = 3.2in]{Aggregate_QoE} \label{qoe_accumulated}}\\ \vspace{-0.0in}
\caption{Time evolution of the allocation and total QoE of the users for exponentially distributed characteristic.} \label{time_evolution}
\end{figure*}
\begin{figure*}[h!]
\centering
\vspace{-0.0in}
\subfloat[]{\includegraphics[width = 3.2in]{comparison_revised2} \label{comp_exp_lambda}} \ \ \ \ \
\subfloat[]{\includegraphics[width = 3.2in]{comparison_mean_revised2} \label{comp_exp_mean}}\\ \vspace{-0.0in}
\subfloat[]{\includegraphics[width = 3.2in]{comparion_unif_lambda_revised2} \label{comp_unif_lambda}} \ \ \ \ \
\subfloat[]{\includegraphics[width = 3.2in]{comparison_unif_mean_revised2} \label{comp_unif_mean}}
\caption{Comparison of expected QoE for exponentially distributed (a \& b) and uniformly distributed ((c) \& (d)) arrival characteristic.}
\label{comparison_fig}
\end{figure*}
\begin{figure*}[h!]
\centering
\vspace{-0.0in}
\subfloat[]{\includegraphics[width = 3.2in]{average_qoe_per_user_exp} \label{qoe_user_exp}} \ \ \ \ \
\subfloat[]{\includegraphics[width = 3.2in]{average_qoe_per_user} \label{qoe_user_unif}}\\ \vspace{-0.0in}
\caption{Comparison of average expected QoE for a typical request for (a) exponentially distributed and (b) uniformly distributed arrival characteristic.}
\label{comparison_user_fig}
\end{figure*}
The MC-IoT applications generate computational requests according to a homogeneous Poisson process that arrive sequentially at the CSP with an arrival rate $\lambda = 10$ requests per hour unless otherwise specified. The reported required response rate of MC-IoT applications is simulated as i.i.d. random variables distributed according to an exponential distribution with mean $\alpha^{-1}$, where $\alpha = 1$, or according to a uniform distribution in the interval $[0, \beta]$, where $\beta = 10$. In the following subsections, we provide the results of the simulations and the obtained insights.
\vspace{-0.0in}
\subsection{Simulation Results}
First we obtain the optimal dynamic cutoff curves obtained for dynamic revenue maximizing VMI allocation. Based on the assumption that the reported required response rate of MC-IoT applications is exponentially distributed and uniformly distributed, the cutoff curves are computed by iteratively solving the differential equations provided in Theorem 2. The first 20 dynamic thresholds for both cases are provided in Fig.~\ref{cutoff_fig}. These are used as a policy for classification of the arriving requests. In general, it can be observed that the thresholds decrease, i.e., the qualification standard is reduced, as the time increases. This implies that the CSP becomes less and less selective as the allocation period is ending to ensure that the maximum number of VMIs are allocated to upcoming requests. However, when there is a lot of time remaining before the allocation period ends, the CSP is more selective ensuring that only the task that reports a sufficiently high required response rate is allocated to a VMI.
It can also be observed from Fig.~\ref{cutoff_exp_fig} that the cutoff curves for the exponentially distributed arrivals have a decreasing allocation margin as the number of VMIs increase. This is due to the possibility of arrival of an highly delay sensitive application for which the barrier is set extremely high. On the other hand, Fig.~\ref{cutoff_unif_fig} shows that under uniformly distributed arrival characteristic, the allocation margin for each of the VMIs is almost constant as the arrival characteristics are equally likely. In both cases, as time approaches the deadline, the cutoff curves approach a constant value equal to the virtual valuation of the agents. It follows that at terminal time, the proposed mechanism is no better than a first-price auction in terms of revenue generation.
In Fig.~\ref{cutoff_fig_eta} illustrates the impact of the QoE parameter $eta$ on the optimal cutoff thresholds. We plot the first cutoff curve, i.e., $y_1(t)$ for varying values of $\eta$ in the case of exponentially distributed and uniformly distributed arrival types in Fig.~\ref{exp_eta_fig} and \ref{unif_eta_fig} respectively. It is clear that the mechanism becomes more selective as $\eta$ increases. However, the thresholds increase much sharply with time for the exponential arrivals as compared to the uniform arrivals.
Next, we investigate the behaviour of the proposed allocation and pricing scheme in response to the changing rate of arrival of MC-IoT requests and the mean arrival characteristic for both exponentially and uniformly distributed arrival types. Fig.~\ref{vs_lambda} shows the expected revenue of the CSP and the expected QoE of the users as the rate of arrival or requests increases. It can be observed from Fig.~\ref{revenue_lambda_exp} and Fig.~\ref{revenue_lambda_unif} that the expected revenue of the CSP increases as the arrival rate increases but saturates at high arrival rates as the available VMIs are exhausted. The exponential arrival characteristic results in a higher expected revenue in general due to the possibility of arrival of highly demanding and highly paying requests, which is not the case in the uniform case. Fig.~\ref{welfare_lambda_exp} and Fig.~\ref{welfare_lambda_unif} depict a similar behaviour in the expected QoE of the users in the two cases. Fig.~\ref{vs_alpha} investigates the behaviour of the expected revenue of the CSP and the expected QoE of the users in response to a change in the mean of the arrival characteristic. It is observed from Fig.~\ref{revenue_alpha_exp} and Fig.~\ref{revenue_alpha_unif} that the expected revenue increases linearly with the mean of the arrival characteristic for both exponential and uniform arrival types. This is because a higher arrival type raises the qualification standard for allocation and consequently the prices. Hence, it does not saturate as the arrival type or rate increases. Finally, Fig.~\ref{welfare_alpha_exp} and Fig.~\ref{welfare_alpha_unif} show a similar increasing behaviour in the expected QoE of the users with an increase in the mean arrival characteristic.
\vspace{-0.0in}
\subsection{Comparison}
To illustrate the performance of the developed optimal allocation mechanism in terms of the QoE of MC-IoT applications, we compare our proposed allocation scheme with the following benchmark strategies:
\begin{enumerate}
\item \textbf{Ideal Allocation:} In this case, we assume there is no uncertainty in the future arrivals and the delay tolerances of all the requests are known a priori. This case is similar to the advance scheduling or reservation scenario where the CSP is aware of the delay tolerances of the applications in advance. It is a useful benchmark since it provides the theoretical upper bound to the achievable QoE under any possible allocation strategy. Note that a loss due to sequentiality and incomplete information will always be there with reference to the ideal allocation.
\item \textbf{Pessimistic Allocation:} In this strategy, the CSP uses the threshold based allocation policy, however, it adopts a pessimistic approach towards future arrivals, i.e., assumes that more delay tolerant requests will arrive in the future as compared to the current one, and allocates the best available VMIs to the requesting applications first. This strategy does not employ any foresight and only makes myopic decisions based on the current arriving task. Hence it attempts to maximize the QoE and the generated revenue by allocating the best available resources to incoming tasks first.
\item \textbf{Optimistic Allocation:} The optimistic allocation policy is the opposite of the pessimistic allocation policy. It always assumes that the arrivals in the future will be more QoE sensitive than the current task at hand. Hence, it saves the best available VMIs while allocating the VMI with the highest end-to-end delay to the current request expecting more QoE sensitive requests to arrive in the future.
\item \textbf{$\epsilon$-Greedy Allocation:} In this strategy, the CSP allocates one of the available VMI in any of the fog nodes with a probability $\epsilon$ to a requesting application with a probability $\epsilon$. In other words, it has an expectation about the nature of the requests that will arrive in the future and hence allocates the available VMIs to an upcoming request with $\epsilon$ probability.
\item \textbf{Periodic Auction:} In this strategy, the available VMIs are periodically auctioned to the requests. The delay tolerances of incoming requests are considered as bids in the auction and the CSP periodically collects the bids and selects the requests to allocate the VMIs. We use a first price auction strategy whereby the highest bidder wins and receives the best available VMI. We assume that the VMIs are allocated one unit at a time. This strategy is adapted from the work in~\cite{most_related_work_comparison}.
\end{enumerate}
Finally, the proposed optimal dynamic strategy makes use of the statistical information about the future arriving tasks to make dynamically optimal decisions for allocating VMIs to arriving requests strategically. For the scenario considered in the simulations, we compare the proposed approach with each of the above mentioned approaches in terms of the average QoE experienced by the MC-IoT applications.
We first provide a time evolution of the allocation procedure in Fig.~\ref{time_evolution}. Fig.~\ref{num_allocated} shows the progression of the number of VMIs allocated as time goes on for each allocation strategy. It can be observed that the auction procedure is able to allocate the maximum number of VMIs while the optimistic approach is able to allocate the least number of VMIs. Fig.~\ref{qoe_accumulated} shoes the corresponding progression of total aggregate QoE with time. Note that while our proposed mechanism does not result in allocation of the most VMIs, but it leads to the maximum aggregate QoE among all the other approaches. Similarly, while auction does allocate more VMIs, it is able to achieve an aggregate QoE similar to the optimistic approach.
The results of the average expected QoE achieved by serviced applications in comparison to the benchmark schemes are provided in Fig.~\ref{comparison_fig}. Fig.~\ref{comp_exp_lambda} and Fig.~\ref{comp_exp_mean} illustrate the average QoE against varying the rate and mean characteristic of the arrival respectively for an exponentially distributed required response rate. Fig.~\ref{comp_unif_lambda} and Fig.~\ref{comp_unif_mean} illustrate the average QoE of serviced applications against varying the rate and mean characteristic of the arrival respectively for a uniformly distributed required response rate. It is clear from the results that the proposed allocation strategy provides the best expected QoE to the users as well as maximizes the revenue of the CSP. However it does suffer from loss of sequentiality as compared with the Ideal allocation scheme. Furthermore, random pairings result in a very low expected QoE while the greedy and optimistic approaches lead to an intermediate average QoE of the users. Another striking result is that the auction strategy may not always be better than other benchmark strategies. Under low arrival rates, the optimistic, greedy, and pessimistic may outperform the auction based strategy. \textcolor{black}{Fig.~\ref{comparison_user_fig} shows the comparison of the average QoE of a typical application request against varying arrival rate of requests for both exponentially (Fig.~\ref{qoe_user_exp}) and uniformly distributed (Fig.~\ref{qoe_user_unif}) arrival characteristic. In general, when the rate of arrival of requests is low, there is a lower expected average QoE for a typical request since the total number of highly demanding applications is lower. This increases sharply as the rate of arrival increases. However, when the arrival rate is very high, the average QoE for a typical application declines since only a few of them can be serviced due to limited resource availability.}
\vspace{-0.0in}
\section{Conclusions} \label{Sec:Conclusion}
In this paper, we provide a QoE based revenue maximizing dynamic resource allocation and pricing framework for fog-enabled MC-IoT applications. We propose an implementable mechanism for allocation of different VMIs available at the fog nodes that result in varying end-to-end delay for user applications.
As opposed to existing works in the literature that focus on static pairing of computational tasks and available fog resources, the proposed resource allocation and pricing strategy is dynamic with instantaneous decision-making as well as takes the hierarchical fog-cloud architecture into account. The developed framework provides an optimal threshold based classification mechanism that uses statistical information about the MC-IoT requests arriving in the future to make dynamically optimal real-time allocation decisions resulting in maximizing revenue of the CSP, based on the QoE of the users. The dynamic thresholds can be pre-computed and used as a lookup table for real-time decision making. Numerical results confirm that the proposed allocation scheme significantly performs better in terms of the QoE achieved by the users in comparison with other benchmark allocation schemes.
\appendices
\section{Proof of Theorem~\ref{single_theorem}} \label{proof_single_theorem}
\textcolor{blue}{}
The proof for the dynamically optimal revenue maximizing curves requires the characterization of the expected revenue of the CSP over the allocation period is a dynamic threshold is used to filter incoming requests. This construction of the proofs is based on the works of Gerkshov and Moldovanu~\cite{revenue_maximization}.
Consider the case where at time $t$ only a single VMI with response rate $r_1$ is available to the CSP for allocation up to time $T$. The expected revenue with a single cutoff curve $y_1(t)$ can be expressed as follows:
\begin{align}
\mathcal{R}(\{r_1\},t) = r_1^{\frac{1}{\eta}} \int_t^T y_{1}^{\frac{1}{\eta}}(s) h_{1}(s) ds,
\end{align}
where $h_{1}(s)$ is the probability density of waiting time until the first arrival of a request with a required response rate of greater than $y_{1}(s)$. The density can be represented by the first arrival in a non-homogeneous Poisson process with intensity $\lambda (1 - F_X(y_{1}(s)))$. For the sake of analytical tractability, we let $\hat{X} = X^{\frac{1}{\eta}}$. Therefore we will use the intensity $\lambda (1 - F_{\hat{X}}(y_{1}^{\frac{1}{\eta}}(s)))$ instead of the former one.
Note that the density of homogeneous Poisson task arrival process is thinned by a factor $(1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(s)))$, which represents the probability that the performance improvement of the arriving task is above the set qualification threshold $y_1^{\frac{1}{\eta}}(s)$. The density can be expressed as follows~\cite{stochastic_book}:
\begin{align}
&h_{1}(s) = \lambda (1 - F_{\hat{X}}(y_{1}^{\frac{1}{\eta}}(s))) \times \notag \\ & \exp \left(- \int_t^s \lambda (1 - F_{\hat{X}}(y_{1}^{\frac{1}{\eta}}(u))) du \right) , \ \ t \leq s \leq T.
\end{align}
Let $H(s) = \int_t^s \lambda \left( 1 - F_{\hat{X}}(y_{1}^{\frac{1}{\eta}}(u)) \right) du$. The expected revenue can then be written as follows:
\begin{align} \label{expected_revenue}
\mathcal{R}(\{r_1\},t) = r_1^{\frac{1}{\eta}} \int_t^T F_{\hat{X}}^{-1} \left( 1 - \frac{H^{\prime}(s)}{\lambda} \right) H^{\prime} (s) e^{-H(s)} ds.
\end{align}
The kernel of integration can be expressed as follows:
\begin{align}
L(s,H(s), H^{\prime}(s)) = F_{\hat{X}}^{-1} \left( 1 - \frac{H^{\prime}(s)}{\lambda} \right) H^{\prime} (s) e^{-H(s)}.
\end{align}
In order to maximize the expected revenue in~\eqref{expected_revenue}, we employ the Euler-Lagrange equation from the calculus of variations~\cite{calculus_variations}. The necessary condition for the revenue maximizing cutoff curves can be expressed as $\frac{\partial L(s,H(s), H^{\prime}(s))}{\partial H(s)} - \frac{d}{dt} \frac{\partial L(s,H(s), H^{\prime}(s))}{\partial H^{\prime}(s)} = 0$. The partial derivatives can be expressed as follows:
\begin{align}
\frac{\partial L}{\partial H(s)} = - F_{\hat{X}}^{-1} \left( 1 - \frac{H^{\prime}(s)}{\lambda} \right) H^{\prime} (s) e^{-H(s)}.
\end{align}
\begin{align}
\frac{\partial L}{\partial H^{\prime}(s)} = e^{-H(s)} \left( F_{\hat{X}}^{-1} \left(1 - \frac{H^{\prime}(s)}{\lambda}\right) \right. \notag \\ \left. - \left( \frac{H^{\prime}(s) }{\lambda f_{\hat{X}} \left(F_{\hat{X}}^{-1} \left(1 - \frac{H^{\prime}(s)}{\lambda}\right)\right)}\right) \right).
\end{align}
\begin{align}
\frac{d}{dt}\frac{\partial L}{\partial H^{\prime}(s)} &= - \left( F_{\hat{X}}^{-1} \left(1 - \frac{H^{\prime}(s)}{\lambda}\right) - \right. \notag \\ & \left. \left( \frac{H^{\prime}(s) }{\lambda f_{\hat{X}} \left(F_{\hat{X}}^{-1} \left(1 - \frac{H^{\prime}(s)}{\lambda}\right)\right)}\right) \right) e^{-H(s)} H^{\prime}(s) + \notag \\ &e^{-H(s)} \left( \frac{- 2 H^{\prime \prime}(s)}{\lambda f_{\hat{X}} \left(F_{\hat{X}}^{-1} \left(1 - \frac{H^{\prime}(s)}{\lambda}\right)\right)}\right. \notag \\
&\left. - \frac{f_{\hat{X}}^{\prime}\left(F_{\hat{X}}^{-1} \left(1 - \frac{H^{\prime}(s)}{\lambda}\right)\right) H^{\prime}(s) H^{\prime \prime}(s)}{ \lambda^2 f_{\hat{X}}^3 \left( F_{\hat{X}}^{-1} \left(1 - \frac{H^{\prime}(s)}{\lambda}\right) \right)} \right),
\end{align}
The Euler-Lagrange equation can be written as follows:
\begin{align}
- (H^{\prime}(s))^2
+& 2 H^{\prime \prime}(s) + \notag \\
&\frac{f_{\hat{X}}^{\prime}\left(F_{\hat{X}}^{-1} \left(1 - \frac{H^{\prime}(s)}{\lambda}\right) \right) H^{\prime}(s) H^{\prime \prime}(s)}{ \lambda f_{\hat{X}}^2 \left( F_{\hat{X}}^{-1} \left(1 - \frac{H^{\prime}(s)}{\lambda}\right) \right)} = 0.
\end{align}
Then, plugging back $H(s) = \int_t^s \lambda (1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(u))) du$, we obtain the following:
\begin{align}
-\lambda (1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(s)))^2 - \frac{2}{\eta} F_{\hat{X}}(y_1^{\frac{1}{\eta}}(s)) y_1^{\frac{1}{\eta}-1}(s) y_1^{\prime}(s) - \notag \\
\frac{ f_{\hat{X}}^{\prime}(y_1^{\frac{1}{\eta}}(s)) (1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(s))) y_1^{\frac{1}{\eta}-1}(s) y_1^{\prime}(s)}{ \eta f_{\hat{X}} (y_1(s))} = 0.
\end{align}
This can be further expressed as follows:
\begin{align}
\frac{d}{ds}\left(y_1^{\frac{1}{\eta}}(s)\right) + \frac{\lambda (1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(s)))^2}{f_{\hat{X}}(y_1^{\frac{1}{\eta}}(s))} = \frac{d}{ds} \left( \frac{ 1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(s))}{ f_{\hat{X}} (y_1^{\frac{1}{\eta}}(s))} \right).
\end{align}
Integrating both sides from $t$ to $T$ results in the following:
\begin{align}
&y_1^{\frac{1}{\eta}}(T) - y_1^{\frac{1}{\eta}}(t) + \lambda \int_t^T \frac{(1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(s)))^2}{f_{\hat{X}}(y_1^{\frac{1}{\eta}}(s))} ds = \notag \\ &\left( \frac{ 1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(T))}{ f_{\hat{X}} (y_1^{\frac{1}{\eta}}(T))} \right) - \left( \frac{ 1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))}{ f_{\hat{X}} (y_1^{\frac{1}{\eta}}(t))} \right)
\end{align}
Using the boundary condition $y_1^{\frac{1}{\eta}}(T) - \frac{1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(T))}{f_{\hat{X}}(y_1^{\frac{1}{\eta}}(T))} = 0$, we reach the following expression:
\begin{align}
- y_1^{\frac{1}{\eta}}(t) + \lambda \int_t^T & \frac{(1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(s)))^2}{f_{\hat{X}}(y_1^{\frac{1}{\eta}}(s))} ds = \notag \\ &- \left( \frac{ 1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))}{ f_{\hat{X}} (y_1^{\frac{1}{\eta}}(t))} \right)
\end{align}
Rearranging the terms results in the following:
\begin{align}
y_1^{\frac{1}{\eta}}(t) = \frac{ 1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))}{ f_{\hat{X}} (y_1^{\frac{1}{\eta}}(t))} + \lambda \int_t^T \frac{(1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(s)))^2}{f_{\hat{X}}(y_1^{\frac{1}{\eta}}(s))} ds. \label{y1}
\end{align}
Equivalently, it can be written as follows:
\begin{align}
y_1(t) = \left( \frac{ 1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))}{ f_{\hat{X}} (y_1^{\frac{1}{\eta}}(t))} + \lambda \int_t^T \frac{(1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(s)))^2}{f_{\hat{X}}(y_1^{\frac{1}{\eta}}(s))} ds \right)^{\eta}. \label{y1}
\end{align}
To complete the proof, we note that the expected revenue is given by $R(r_j,t) = r_j^{\frac{1}{\eta}} R(1,t)$ where
\begin{align}
&R(1,t) = \notag \\ &\int_t^T y_1^{\frac{1}{\eta}}(s) \lambda (1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(s))) e^{- \int_t^s \lambda (1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(s))) dz}ds.
\end{align}
Differentiating the above expression with respect to $t$ gives the following:
\begin{align}
R^{\prime}(1,t) = \lambda (1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t)))(R(1,t) - y_1^{\frac{1}{\eta}}(t)). \label{Revenue_single}
\end{align}
It can be shown that $R(1,t) = \lambda \int_t^T \frac{(1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(s)))^2}{f_{\hat{X}}(y_1^{\frac{1}{\eta}}(s))} ds$ satisfies equation~\eqref{Revenue_single}.
\section{Proof of Theorem~} \label{proof_allocation_theorem}
If two VMIs are available, then the revenue can be expressed as follows:
\begin{align}
\int_0^T &\left( P_2(\{r_1,r_2\},t) + R(\{r_1\},t) \right) h_2(t) dt + \notag \\ &\int_0^T \left( P_1(\{r_1,r_2\},t) + R(\{r_2\},t) \right) h_1(t) dt.
\end{align}
This can be further written as follows:
\begin{align}
&\int_0^T \left( r_2^{\frac{1}{\eta}} y_2^{\frac{1}{\eta}} + R(r_1,t)\right) \times \notag \\ & \lambda (1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(t))) e^{-\int_0^t \lambda (1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(s)))ds } dt + \notag \\ &\int_0^T \left( (r_1^{\frac{1}{\eta}} - r_2^{\frac{1}{\eta}}) y_1^{\frac{1}{\eta}}(t) + R(r_2,t) - R(r_1,t) \right) \times \notag \\
&\lambda (1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))) e^{ - \int_0^t \lambda (1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(s)))ds} dt, \notag \\
&= \int_0^T \left( r_2^{\frac{1}{\eta}} y_2^{\frac{1}{\eta}} + R(r_1,t)\right) \times \notag \\ &\lambda (1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(t))) e^{-\int_0^t \lambda (1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(s)))ds } dt + \notag \\ &(r_1^{\frac{1}{\eta}} - r_2^{\frac{1}{\eta}}) \int_0^T \left( y_1^{\frac{1}{\eta}}(t) - R(1,t) \right) \times \notag \\
&\lambda (1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))) e^{ - \int_0^t \lambda (1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(s)))ds} dt.
\end{align}
Let $G(t) = \int_0^t \lambda (1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(s))) ds $ and $H(t) = \int_0^t \lambda (1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(s))) ds$. Then, the expression can further be written as follows:
\begin{align}
&\int_0^T \left( r_2^{\frac{1}{\eta}} F_{\hat{X}}^{-1} \left( 1 - \frac{H^{\prime}(t)}{\lambda} \right) + r_1^{\frac{1}{\eta}} R(1,t) \right) H^{\prime}(t) e^{- H(t)} dt + \notag \\ &(r_1^{\frac{1}{\eta}} - r_2^{\frac{1}{\eta}}) \int_0^T \left( F_{\hat{X}}^{-1} \left( 1 - \frac{G^{\prime}(t)}{\lambda} \right) \right) G^{\prime \prime}(t) e^{-H(t)} dt.
\end{align}
Therefore, $L_1(t, H(t), H^{\prime}(t)) = \left( r_2^{\frac{1}{\eta}} F_{\hat{X}}^{-1} \left( 1 - \frac{H^{\prime}(t)}{\lambda} \right) + r_1^{\frac{1}{\eta}} R(r_1,t)\right) H^{\prime}(t) e^{-H(t) } + (r_1^{\frac{1}{\eta}} - r_2^{\frac{1}{\eta}})\left( F_{\hat{X}}^{-1} \left( 1 - \frac{G^{\prime}(t)}{\lambda} \right) - R(1,t) \right) G^{\prime}(t) e^{-H(t)}$.
Computing the Euler-Lagrange equation, $\frac{\partial L_1}{\partial H(t)} - \frac{d}{dt} \frac{\partial L_1}{\partial H^{\prime}(t)} = 0$ results in the following:
\begin{align}
&- (r_1^{\frac{1}{\eta}} - r_2^{\frac{1}{\eta}}) G^{\prime}(t) \left( F_{\hat{X}}^{-1} \left(1 - \frac{G^{\prime}(t)}{\lambda}\right) - R(1,t) \right) - \notag \\ &r_1^{\frac{1}{\eta}} R^{\prime}(1,t) -
r_2^{\frac{1}{\eta}}\frac{(H^{\prime}(t))^2 }{ \lambda F_{\hat{X}} \left( F_{\hat{X}}^{-1} \left( 1 - \frac{H^{\prime}(t)}{\lambda} \right) \right)} + \notag\\
&2 r_2^{\frac{1}{\eta}}\frac{ H^{\prime \prime}(t) }{ \lambda F_{\hat{X}} \left( F_{\hat{X}}^{-1} \left( 1 - \frac{H^{\prime}(t)}{\lambda} \right) \right)}
+ \notag \\
& r_2^{\frac{1}{\eta}} \frac{ F_{\hat{X}}^{\prime} \left( F_{\hat{X}}^{-1} \left( 1 - \frac{H^{\prime}(t)}{\lambda} \right) \right)H^{\prime}(t) H^{\prime \prime}(t) }{ \lambda^2 \left( F_{\hat{X}} \left( F_{\hat{X}}^{-1} \left( 1 - \frac{H^{\prime}(t)}{\lambda} \right) \right) \right)^3} = 0,
\end{align}
and $\frac{\partial L_1}{\partial G(t)} - \frac{d}{dt} \frac{\partial L_1}{\partial G^{\prime}(t)} = 0$ results in the following:
\begin{align}
&- H^{\prime}(t) \left( - \frac{G^{\prime} (t)}{ \lambda F_{\hat{X}} \left( F_{\hat{X}}^{-1} \left( 1 - \frac{G^{\prime}(t)}{\lambda} \right) \right) } + F_{\hat{X}}^{-1} \left( 1 - \frac{G^{\prime}(t)}{\lambda} \right) - \right. \notag \\ &\left. R(1,t) \right) - R^{\prime}(1,t) - \frac{ 2 G^{\prime \prime}(t)}{ \lambda F_{\hat{X}} \left( F_{\hat{X}}^{-1} \left( 1 - \frac{G^{\prime}(t)}{\lambda} \right) \right) } + \notag \\
&
\frac{f^{\prime}_X \left( F_{\hat{X}}^{-1} \left( 1 - \frac{G^{\prime}(t)}{\lambda} \right) \right) G^{\prime}(t) G^{\prime \prime} (t)}{\lambda^2 f^3_X \left( F_{\hat{X}}^{-1} \left( 1 - \frac{G^{\prime}(t)}{\lambda} \right) \right) } = 0.
\end{align}
Eventually, it leads to the following differential equations:
\begin{align}
&-(r_1^{\frac{1}{\eta}} - r_2^{\frac{1}{\eta}}) \lambda \left( 1 - F_{\hat{X}} \left( y_1^{\frac{1}{\eta}}(t) \right) \right)
(y_1^{\frac{1}{\eta}}(t) - R(1,t)) - \notag \\& r_1^{\frac{1}{\eta}}(t) R^{\prime}(1,t) - r_2^{\frac{1}{\eta}} \frac{\lambda (1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(t)))^2 }{ F_{\hat{X}}(y_2^{\frac{1}{\eta}}(t))} - 2 r_2^{\frac{1}{\eta}} \frac{ y_2^{\prime}(t) y_2^{\frac{1}{\eta} - 1}(t)}{ \eta} - \notag \\
&
r_2^{\frac{1}{\eta}}\frac{y_2^{\prime}(t) (1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(t))) F_{\hat{X}}^{\prime}(y_2^{\frac{1}{\eta}}(t)) y_2^{\frac{1}{\eta}-1}(t) }{ \eta F_{\hat{X}}^2(y_2^{\frac{1}{\eta}}(t))} = 0, \label{first_eq}
\end{align}
and
\begin{align}
& \frac{2 y_1^{\prime}(t) y_1^{\frac{1}{\eta}-1}(t)}{\eta } - R^{\prime}(1,t) - \notag \\ &\frac{ F_{\hat{X}}^{\prime}(y_1^{\frac{1}{\eta}}(t)) (1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))) y_1^{\prime}(t) y_1^{ \frac{1}{\eta}-1}(t) }{ \eta f_{\hat{X}}^2(y_1^{\frac{1}{\eta}}(t)) } + \lambda (1-F_{\hat{X}}(y_2^{\frac{1}{\eta}}(t))) \notag \\ & \left( \frac{ (1-F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t)))}{ f_{\hat{X}}(y_1^{\frac{1}{\eta}}(t)) } + y_1^{\frac{1}{\eta}}(t) + R(1,t) \right) = 0. \label{second_eq}
\end{align}
Now, we show that a solution to these differential equations is given by the solution to the following system of equations:
\begin{align}
y_1(t) = \left( \frac{1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))}{f_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))} + \lambda \int_t^T \frac{(1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(s)))^2}{f_{\hat{X}}(y_1^{\frac{1}{\eta}}(s))} ds \right)^{\eta}, \label{first}
\end{align}
and
\begin{align}
y_2(t) = \Bigg( \frac{1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(t))}{f_{\hat{X}}(y_2^{\frac{1}{\eta}}(t))} + &\lambda \int_t^T \frac{(1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(s)))^2}{f_{\hat{X}}(y_2^{\frac{1}{\eta}}(s))} ds - \notag \\ & R(1,t) \Bigg)^{\eta}. \label{second}
\end{align}
Differentiating~\eqref{first} with respect to $t$ results in the following:
\begin{align}
2 \frac{y_1^{\frac{1}{\eta}-1} (t) y_1^{\prime}(t) }{\eta} &= - y_1^{\prime}(t) \frac{(1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))) f^{\prime}(y_1^{\frac{1}{\eta}}(t))}{f_{\hat{X}}^2(y_1^{\frac{1}{\eta}}(t))} - \notag \\ &\frac{ \lambda (1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t)))^2}{f_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))}
\end{align}
Substituting this in~\eqref{second_eq} results in the following:
\begin{align}
\left( \lambda(1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))) - \lambda (1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(t))) \right) \times \notag \\ \left( \frac{1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))}{ f_{\hat{X}}(y_1^{\frac{1}{\eta}}(t)) } - y_1^{\frac{1}{\eta}}(t) + R(1,t) \right) = 0,
\end{align}
which is satisfied for all $y_2(t)$ using~\eqref{first}.
Differentiating~\eqref{second} with respect to $t$ results in the following:
\begin{align}
2 \frac{y_2^{\prime}(t) y_2^{\frac{1}{\eta}-1}}{\eta} &= - \frac{(1 - F_{\hat{X}}(y_2(t))F_{\hat{X}}^{\prime}(y_2(t))y_2^{\prime}(t)}{f_{\hat{X}}^2(y_2(t))} - \notag \\ & \frac{\lambda (1 - F_{\hat{X}}(y_2(t)))^2}{F_{\hat{X}}(y_2(t))} - R^{\prime}(1,t).
\end{align}
Substituting this into~\eqref{second_eq} results in the following:
\begin{align}
-(r_1^{\frac{1}{\eta}} - r_2^{\frac{1}{\eta}}) \lambda (1 - F_{\hat{X}}(y_1^{\frac{1}{\eta}}(t))) (y_1^{\frac{1}{\eta}(t)} - \notag \\ R(1,t)) - R^{\prime}(1,t) \left(r_1^{\frac{1}{\eta}} - r_2^{\frac{1}{\eta}} \right) = 0
\end{align}
This holds for $y_1(t)$ from~\eqref{Revenue_single}.
\section{Proof of Theorem~\ref{general_theorem}} \label{proof_general_theorem}
To prove the general case, we show that when two VMIs with unit characteristic are available, the revenue can be computed as follows:
\begin{align}
&R(\{1,1\},t) = \int_t^T \left((y_2^{\frac{1}{\eta}}(t) + \right. \notag \\ & \left. R(1,s)\right) \lambda (1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(s))) e^{- \int_t^s \lambda (1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(z))) dz} ds.
\end{align}
Differentiating with respect to $t$ results in
\begin{align}
&R^{\prime}(\{1,1\},t) = \notag \\ &\lambda (1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(t))) (R(\{1,1\}) - y_2^{\frac{1}{\eta}}(t) - R(1,t)),
\end{align}
It can be shown that $R(\{1,1\},t) = \int_t^T \frac{(1 - F_{\hat{X}}(y_2^{\frac{1}{\eta}}(s))^2)}{f_{\hat{X}}(y_2^{\frac{1}{\eta}}(s))} ds$ satisfies the above differential equation using~\eqref{second}.
Using a similar procedure, it can be be shown that in the general case, the optimal threshold solves the following equation:
\begin{align}
y_i(t) = \left( \frac{1 - F_{\hat{X}}(y_i^{\frac{1}{\eta}}(t))}{f_{\hat{X}}(y_i^{\frac{1}{\eta}}(t))} + R(\boldsymbol{1}_i,t) - R(\boldsymbol{1}_{i-1},t) \right)^{\eta},
\end{align}
where
\begin{align}
R(\boldsymbol{1}_i,t) - \lambda \int_t^T \frac{(1 - F_{\hat{X}}(y_i^{\frac{1}{\eta}}(s)) )^2}{f_{\hat{X}}(y_i^{\frac{1}{\eta}}(s))} ds
\end{align}
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The interest for (geometric) structure-preserving numerical integrators has grown in computing community \cite{hairer06,blanes16,amses18,bochev06,arnold10,amses19,salnikov18}. This is due to the ability of these schemes in reproducing foundamental physical properties (conservation laws, \dots) of the equations and their robustness in long-time integration. Discrete Exterior Calculus (DEC) belongs to this family of integrators. It aims at developing a discrete version of the theory of exterior calculus, and more generally the differential geometry theory, where most equations of physics are formulated. Initially developed by Bossavit for electromagnetism in a series of papers \cite{bossavit98_1,bossavit98_2,bossavit98_3,bossavit98_4,bossavit99_1,bossavit99_2,bossavit99_3,bossavit99_4,bossavit99_5}, DEC was used in fluid mechanics for the resolution of Darcy's equation and the simulation of the dynamics of ideal and viscous fluid flows in basic configurations \cite{elcott07,hirani15,mohamed16a}.
The primary calculus tools of the DEC are discrete differential forms or cochains. An important advantage of DEC is that the Stokes' theorem is naturally verified at discrete level. This is due to the construction of the discrete exterior derivative operator $ｄ$ by duality with the boundary operator. A second advantage is that the DEC framework offers a naturally coherent discretization
of derivative operators (divergence, gradient, curl in 2D and 3D) such that the usual relations
\[\rot \grad=0,\quad\quad\text{and}\quad\quad\dive\rot=0\]
are verified at machine precision. This is due to the fact that these derivative operators are all represented by the exterior derivative $ｄ$ in exterior calculus framework and that in DEC, the discrete $ｄ$ obeys the relation
\begin{equation}
ｄ^2=0, \label{d2}
\end{equation}
just like the continuous $ｄ$.
These properties of DEC permitted for example Elcott et al. \cite{elcott07} to design a circulation-preserving numerical scheme for the simulation of ideal fluid flows. These properties also avoid the apparition of spurious quantities such as an artificial mass, potential or portance in the numerical solution.
When used as discretization method, DEC can be seen as a finite volume method in exterior calculus. Indeed, in this approach, cochains are differential forms integrated over the different elements of the mesh (vertices, edges, triangles, \dots). Disrete exterior calculus has a finite element version, developed mainly by Arnold and his co-workers in \cite{arnold06,arnold10,arnold18}.
A key operator in exterior calculus, needed to express constitutive laws for example, is the Hodge star operator. In a finite element approach, the discrete Hodge operator is built straightforwardly, by applying the continuous Hodge to the polynomial differential forms which constitute the basis. After a suitable inner product, this results in the mass matrix. In DEC, to which the present article is limited, defining a discrete Hodge is less natural because one has to deal with already intregrated forms. In particular, in order to ensure the bijectivity of the discrete Hodge, a dual mesh is necessary in DEC, whereas such a notion is not required in the finite element approach. The circumcentric dual is a popular choice of dual mesh, due to the orthogonal nature of simplicial cells and their circumcentric duals. This orthogonality permits a simple construction of a discrete Hodge, having a diagonal matrix representation \cite{bossavit99_2,hirani03}. The diagonal entries of this discrete Hodge, sometimes called circumcentric or diagonal Hodge, are simply the ratios of the volumes of dual cells and primal simplices.
The diagonal Hodge was initialy designed for completely well-centered simplicial meshes, that are meshes where each simplex contains its circumcenter.
In practice, it is difficult to generate this type of triangulation \cite{Rajan94}. It can be done for some specific simple geometries in $\mathbb{R}^2$ \cite{cassidi81,yuan10,zamfirescu13}. For more complex geometries, VanderZee et al. propose a reprocessing algorithm making a 2D mesh well-centered but it is not very practical and may not work in 3D \cite{vanderzee10}. Hirani and his co-workers extended the use of the diagonal Hodge to Delaunay and to pairwise non-Delaunay meshes, by introducing signed elementary dual volumes \cite{Hirani13,mohamed18}. However, this approach may still lead to zero-volume dual edges and cells when positive contributions equalize negative ones. A reprocessing is then necessary. Mullen et al. \cite{Mullen11hot:hodge-optimized} proposed an optimization of the triangulation to the diagonal Hodge. To this aim, they introduce weighted duals and weighted circumcenter, and construct the mesh by minimising the error of the diagonal Hodge. This technique generates a well-shaped mesh, but not well centered. Moreover, it may become expensive, especially for problems with time-varying domains such as fluid-structure interaction. Note also that all these discrete Hodge operators require that the dual mesh is based on the circumcenter.
Another existing discrete Hodge operator, studied by Bossavit from the beginning of DEC, is the Galerkin Hodge \cite{bossavit98,bossavit99_5}. It corresponds to a mass matrix, with the Whitney forms as shape functions \cite{whitney57,bossavit88}. It is built within a finite element framework. A drawback of using the Galerkin Hodge in (the finite-volume flavoured) DEC is that, not only it introduces an inconsistency but also it is not exact even for constant forms. This can be shown straighforwardly in a right but not isocele triangle. A more DEC- than finite element-flavoured discrete Hodge, but still based on Whitney forms, is proposed in \cite{tarhasaari99}. It is defined through an interpolation with Whitney forms and an integration on dual simplices and is generally called Whitney Hodge. Another discrete Hodge operator which also uses Whitney forms is proposed in \cite{Auchmann06}. Called geometric Hodge, it does not need an interpolation nor integration concept. For symmetry reasons, it requires a dual mesh based on the barycenters. In fact, as noticed in \cite{mohamed16b} in 2D, the geometric Hodge is a symmetric approximation of the Galerkin Hodge where the integration is approximated by a one point quadrature with evaluation only at the barycenter. Like the usual Galerkin and the Whitney Hodge, the geometric Hodge is represented by a sparse but not diagonal matrix. However, numerical comparisons in \cite{mohamed16b} shows that the additional computational time induced by the lake of diagonal structure is not very high.
In the present article, we propose a new construction of a discrete Hodge operator which comply the following requirements. First, it does not need a reprocessing of the primal mesh. This condition is important since, in some situations, modifying the mesh may be harmfull. It may for instance affect the shape of surfaces. Moreover, as mentioned, the induced cost increase may become important in time-varying domains. Second, the construction must be general enough such that the centers on which the dual mesh is based can be changed at convenience, as long as the dual mesh has no degenerate element. For example, the construction of the discrete Hodge must be valid for both a barycentric and an incentric duals. Neither the diagonal nor the geometric Hodge is appropriate for this task since with these operators, the dual mesh is imposed. Lastly, the Hodge must be exact on piecewise constant differential forms, whichever choice is made on the dual mesh.
By proposing such a construction, we intend to develop a discrete Hodge which adapts to the mesh in order to minimize some quantity (the global error, \dots) or to comply to other physical requirements, by choosing algorithmically the most suitable center. These center may even vary from a simplex to another. This optimization process, which will be tackled in a future work, would contrast with that in \cite{Mullen11hot:hodge-optimized} where it is the mesh which adapts to the (diagonal) Hodge operator.
In section 2, a brief reminder on DEC, on the continuous Hodge operator and on the diagonal Hodge is done. A deeper introduction to DEC can be found in the papers of Bossavit \cite{bossavit98_1,bossavit98_2,bossavit98_3,bossavit98_4,bossavit99_1,bossavit99_2,bossavit99_3,bossavit99_4,bossavit99_5} and in \cite{hirani03,desbrun05,crane13} for instance. The development of the new discrete Hodge, followed by illustrative examples, a basic preliminary error analysis and a discussion, is presented in section 3. We end up with numerical experiments, carried out on Poisson equations and on isothermal and non-isothermal fluid dynamics problems, in section 4. Various types of mesh (well-centered mesh, right triangulation and non-Delaunay meshes) will be considered. Convergence analyses will be carried out. Although our goal is not to compare the new discrete Hodge to existing ones, results given by the diagonal Hodge are presented as reference, when available.
\section{Review on DEC}
DEC is a theory of exterior calculus on a discretized domain, which preserves Stokes' theorem and the exactness of relation (\ref{d2}). In what follows, we make a reminder on oriented simplicial discretization of a domain. Next, since the primary tools of exterior calculus are differential forms, we briefly present how to discretize them into cochains. Lastly, the Hodge star operator and its discretization into the diagonal Hodge is reminded.
\subsection{Discretization of the domain\label{sec:discretization_domain}}
Consider a system of partial differential equations defined on an $n$-dimensional spatial domain $M\subset ℝ^m$ for some $m∈ℕ$. The domain $M$ is subdivised into $n$-dimensional simplices. Recall that a $k$-simplex is the convex hull of $k+1$ points. For instance, if $\dim M=3$ then $M$ is subdivised into tetrahedra. We denote $K$ the set composed of these $n$-dimensional simplices, together with their faces, the faces of the faces, and so on untill the vertices. For example, in 3D, $K$ is composed of tetrahedra (3-simplices), triangles (2-simplices), edges (1-simplices) and vertices (0-simplices). In 2D meshes, the top-dimensional simplices are triangles. As an example, Figure \ref{fig:complex} presents a simplicial discretization of a curved surface in $ℝ^3$.
Some regularity conditions are also generally usefull \cite{hirani03,desbrun05,gillette09}.
\begin{figure}[ht]
\centering
\includegraphics[width=4cm]{fig_complex}
\caption{Example of 2D simplicial complex embedded in $ℝ^3$}
\label{fig:complex}
\end{figure}
To each element of $K$ is assigned an orientation, which can simply be defined from an ordering of its nodes. Each simplex has two possible orientations. Two adjacent top-dimensional simplices are required to have the same orientation. By contrast, the orientation of each lower dimensional simplex is arbitrary. A sample 2D oriented complex is presented in Figure \ref{mesh}. For simplicity, and since it is sufficient for our applications, we assume that all the vertices have the same (positive) orientation.
\begin{figure}[ht]\centering
\includegraphics[width=4cm]{fig_orientation}
\caption{Example of a consistently oriented mesh. Arrows represent the orientation of edges and faces}
\label{mesh}
\end{figure}
We denote $K_k$ the set of oriented $k$-dimensional simplices of $K$
\begin{equation*}K_k=\{σ\in K,\ σ\text{ oriented},\ \dim σ=k\}\end{equation*}
and $Ω_k$ the vector space spanned by formal linear combination of elements of $K_k$
\begin{equation*}Ω_k=\operatorname{span}K_k=\left\{c=\sum_{σ_i\in K_k}c_iσ_i,\ c_i\in ℝ\right\}.\end{equation*}
An element $c$ of $Ω_k$ is called a $k-$chain. In theory, the components $c_i$ may take any real value but, in practice, only values in $\{0,1,-1\}$ have physical meaning. A value 0 means that the element $σ_i$ does not belong to the chain, 1 means that it is present in $c$, and $-1$ when it is present but with the opposite orientation. As an example, the chain
\begin{equation}
e_1+e_4+e_5-e_3\in Ω_1
\label{chain}
\end{equation}
in Figure \ref{mesh} constitutes a closed loop.
\subsection{Boundary operator}
The boundary of a $k$-dimensional simplex is the sum of its oriented $(k-1)$-dimensional faces. In this sum, each face is given a sign, depending on wether its orientation is consistent with that of the considered $k$-simplex. For example, the boundary of the face $f_1$ in Figure \ref{mesh} is
\begin{equation*}
∂f_1=e_1+e_2-e_3.
\end{equation*}
The boundary of $e_1$ is
\begin{equation*}
∂e_1=v_2-v_1.
\end{equation*}
The boundary operator $∂$ extends into a linear map from $Ω_k$ to $Ω_{k-1}$ by defining the boundary of a chain $∑_{i∈I}c_iσ_i$, for some set $I$ of indices, as follows:
\begin{equation}
∂⟦∑_{i∈I}c_iσ_i⟧=∑_{i∈I}c_i∂σ_i.
\end{equation}%
The operator $∂$ can be represented by a (sparse) matrix. For instance, the non-zero boundary operators on the mesh in Figure \ref{mesh} are
\begin{equation}
∂_{|Ω_2}=\begin{pmatrix}
1&0\\1&-1\\-1&0\\0&1\\0&1
\end{pmatrix},\quad
∂_{|Ω_1}=\begin{pmatrix}
-1& 0&-1& 0& 0&\\
1&-1& 0&-1& 0&\\
0& 1& 1& 0& 1&\\
0& 0& 0& 1&-1&
\end{pmatrix}.
\label{boundary}
\end{equation}
In equation (\ref{boundary}), $∂_{|Ω_k}$ is the restriction of $∂$ to $Ω_k$, which acts on $k$-simplices.
\subsection{Discrete differential forms and exterior derivative}
A discrete $k$-form, or a $k-$cochain, is an element of the algebraic dual $Ω^k(K)$ of $Ω_k$. Since the elements of $K_k$ form a basis of $Ω_k$, a discrete $k$-form is simply a map which, to each element of $K_k$, assings a real number:
\begin{equation}
ω\in Ω^k(K)\ :\ \begin{array}{ccc}K_k&\longrightarrow& ℝ\\σ&\longmapsto&⟨ω,σ⟩.\end{array}
\label{form}
\end{equation}
Computationally, a $k$-cochain is represented by an array $(⟨ω,σ_i⟩)_{σ_i∈K_k}$.
In DEC, a differential $k$-form $ω\inΩ^k(M)$ on $M$ is discretized into a $k$-cochain $ω_K∈Ω^k(K)$ by taking the pairing $⟨·,·⟩$ as an integration
\[ ⟨ω_K,c⟩=∫_cω,␣␣\text{for all }c∈Ω_k\]
if $k≥1$, and as a simple evaluation
\[ ⟨ω_K,c⟩=ω(c)\]
if $c∈Ω_0$ is a point.
In the sequel, the cochain $ω_K$ will also be denoted $ω$ when there is no confusion.
The discrete exterior derivative operator $ｄ$ is defined as the dual of the boundary operator $∂$:
\begin{equation}
⟨ｄω,c⟩=⟨ω,∂c⟩,\quad \text{ for all } ω\inΩ_k, c\in K_{k+1}.
\label{d}
\end{equation}
It is a linear map from $Ω^k$ to $Ω^{k+1}$. It can be remarked that relation (\ref{d}) would simply be the expression of Stokes' theorem if the pairing $⟨·,·⟩$ was an integration and ω a differential form. Computationally, the matrix $ｄ_{|Ω^k(K)}$ of the restriction of $ｄ$ on the space of $k$-cochains is the transpose of the matrix of $∂_{|Ω_{k+1}}$:
\[ｄ_{|Ω^k(K)}=∂_{|Ω_{k+1}}^\mathsf{T}.\]
Since $∂^2=0$ (the boundary of a boundary is the empty set), it follows by the duality relation (\ref{d}) that
\begin{equation}
ｄ^2=0.
\end{equation}
The following sequence is then, as in the continuous case, exact:
\begin{equation}
0\xrightarrow{\makebox[5mm]{$ｄ$}}Ω^0(K)\xrightarrow{\makebox[5mm]{$ｄ$}}Ω^1(K)\xrightarrow{\makebox[5mm]{$ｄ$}}\cdots\xrightarrow{\makebox[5mm]{$ｄ$}}Ω^{n-1}(K)\xrightarrow{\makebox[5mm]{$ｄ$}}Ω^n(K)\xrightarrow{\makebox[5mm]{$ｄ$}}0.
\label{derham}
\end{equation}
\subsection{The diagonal Hodge}
Having applications in fluid mechanics domain in mind, we assume that $M$ is a flat domain in $ℝ^n$, $n=2$ or 3, from now on, to simplify. It inherits the usual Euclidean metric and the induced right-hand oriented volume form of $ℝ^n$.
To express material laws in exterior calculus framework, the Hodge star operator $\star$ is necessary. It realizes an isomorphism between $\Omega^k(M)$ and $\Omega^{n-k}(M)$. Rather than giving its definition (which can be found for example in \cite{marsden99}), we simply recall that, if $n=2$,
\begin{equation}
\begin{array}{lll}
\star 1=ｄx\wedge ｄy, \\[7pt]
\star ｄx=ｄy,\quad&\star ｄy=-ｄx, \\[7pt]
\star (ｄx\wedge ｄy)=1.
\end{array}
\label{star2d}
\end{equation}
In fact, these relations completely characterize the operator $\star$.
For $n=3$, we have
\begin{equation*}
\begin{array}{lll}
\star 1=ｄx\wedge ｄy \wedge ｄz,\\[7pt]
\star ｄx=ｄy\wedge ｄz,\quad&\star ｄy=ｄz\wedge ｄx,\quad&\star ｄz=ｄx \wedge ｄy,\\[7pt]
\star(ｄy\wedge ｄz)=ｄx,\quad&\star(ｄz\wedge ｄx)=ｄy,\quad&\star(ｄx \wedge ｄy)=ｄz,\\[7pt]
\star (ｄx\wedge ｄy\wedge ｄz)=1.
\end{array}
\end{equation*}
It can be checked that
\begin{equation}
\star\star = -\mathsf{Id}
\label{starstar}
\end{equation}
in 2D and
\[ \star\star = \mathsf{Id} \]
in 3D, where $\mathsf{Id}$ is the identity map. Note also that the Hodge star verifies the relation
\begin{equation}
ω(u_1,\dots,u_k)=\star ω(u_{k+1},\dots,u_{n}) \label{ortho}
\end{equation}
for any positively oriented orthonormal basis $(u_1,\dots,u_n)$ of the tangent space and any differential $k$-form $ω∈Ω^k(M)$.
The discrete Hodge star operator should realize an isomorphism between $Ω^k(K)$ and $Ω^{n-k}(K)$. This is not possible on the same grid $K$ because, generally,
$$\operatorname{dim}Ω^k(K)=\operatorname{card}K_k\not=\operatorname{dim}Ω^{n-k}(K)=\operatorname{card}K_{n-k}.$$
For example, with the mesh on Figure \ref{mesh}, $n$ equals 2 and if $k=0$,
\[\operatorname{dim}Ω^k(K)=4,\quad\quad\operatorname{dim}Ω^{n-k}(K)=2.\]
In order to define a discrete Hodge star operator, a dual mesh $\star K$ is needed. This dual mesh must be constituted by cells such that there is a one-to-one correspondance between $k$-simplices of $K$ and $(n-k)$-dimensional cells of $\star K$. One possible dual mesh is the circumcentric dual. In a 2D mesh such as in Figure \ref{fig:dual}, the circumcentric dual is such that the dual of a triangle is its circumcenter; the dual of a primal edge is the edge connecting the circumcenters of two triangles which share this primal edge; and the dual of a vertex is the 2-cell formed by connecting the circumcenters of the primal triangles which share this vertex. Figure \ref{fig:maillage_dual} presents an example of a mesh on a square and its circumcenter-based dual. Note that the choice of circumcenters as duals of triangles implies the orthogonality between edges and their duals.
\begin{figure}[ht]\centering
\includegraphics[width=9cm]{fig_dual}
\caption{Primal simplices (in blue) of a triangle $f$ (left), an edge $e$ (middle) and a vertex $v$ (right) and their duals $f^*$, $e^*$, $v^*$ (in red) in a 2D mesh}
\label{fig:dual}
\end{figure}
\begin{figure}[ht]\centering
\includegraphics[width=3cm]{fig_maillage_dual}
\caption{Sample 2D mesh (in blue) on a square and its circumcentric dual (in red)}
\label{fig:maillage_dual}
\end{figure}
The boundary and discrete exterior derivative can be transposed to the dual mesh. The orientation on the primal mesh $K$ also induces an orientation on $\star K$ (see \cite{desbrun05})\footnote{To borrow the language of Bossavit \cite{bossavit05}, the (inner-)orientation of each primal simplex, as defined in section \ref{sec:discretization_domain}, constitutes the outer-orientation of its dual cell.}.
The space of $k$-cochains on the dual mesh is denoted $Ω^k(\star K)$.
A discrete Hodge operator, realizing an isomorphism between $Ω^k(K)$ and $Ω^{n-k}(\star K)$, can now defined as:
$$\star :\ Ω^k(K)\longrightarrow Ω^{n-k}(\star K)$$
with
\begin{equation}
\cfrac{⟪\star ω,c^*⟫}{|c^*|}=\cfrac{⟪ω,c⟫}{|c|}, \quad\text{ for any }c\in K_k,
\label{hodge}
\end{equation}
where $c^*\in K_{n-k}^*$ designates the dual of the $k-$simplex $c$. In equation (\ref{hodge}), $|c|$ is a measure of $c$, which is set to $1$ if $c$ is a point and to its Euclidean measure otherwise. Relation (\ref{hodge}) can be understood as follows: the discrete forms $ω$ and $\star ω$, normalized by the measure of the simplex/cell on which they are applied, represent the same density, but act in complementary-dimensional facets. Relation (\ref{hodge}) can also be understood as a discretization of equality (\ref{ortho}), $c$ and $c^*$ forming complementary geometric objects.
Definition (\ref{hodge}) is a low-order approximation of the continuous Hodge star, which turns out to be exact in the particular case where $ω$ is piecewise constant and $c$ and $c^*$ are flat. It induces a representation of the discrete Hodge operator in each dimension as a diagonal matrix, having ratios of cell measures as entries. For example, in a one-triangle mesh as in Figure \ref{triangle_circ}, where $c_t$ and $c_i$ are the circumcenters of the triangle and the edges respectively, the Hodge matrix acting on primal 1-cochains is
\begin{equation}
H_{diagonal}=
\begin{bmatrix}
\displaystyle\dfrac{|e_1^*|}{|e_1|}&0&0\\
0&\displaystyle\dfrac{|e_2^*|}{|e_2|}&0\\
0&0&\displaystyle\dfrac{|e_3^*|}{|e_3|}
\end{bmatrix}. \label{hdiagonal}
\end{equation}
Due to relation (\ref{starstar}), the discrete Hodge matrix acting on dual 1-cochain is the inverse of matrix (\ref{hdiagonal}).
The diagonal structure of $H_{diagonal}$ and the induced computational efficiency makes the diagonal Hodge a popular choice of discrete Hodge operator.
However, to be correctly defined, it requires a well-centered mesh. This means that each $k$-dimensional simplex of the mesh, with $k≥1$, must strictly contain its circumcenter. In 2D, a well-centered mesh can contain only strictly acute triangles. To figure out the problem induced by the non well-centerdness, consider a 2D domain meshed with right triangles. The mesh is not well centered. Circumcenters lie on the boundary of the triangles and some dual edges degenerate into points (see Figure \ref{rectangle}, left, as an example), making matrix (\ref{hdiagonal}) singular. If the mesh contain obtuse triangles then some circumcenters are located outside the triangles. In this case, an extension, introducing a sign convention to the elementary volumes of the dual cells, has been proposed in \cite{mohamed18} to widen the range of applicability of the diagonal Hodge, but bringing no solution for the right-triangularized mesh for example (apart from altering the mesh). %
Moreover, the diagonal Hodge does not allow any choice than the circumcenters as duals of top-simplices. Indeed it has been shown that if $c_t$ and the $c_i$ are not the circumcenters, the diagonal formula (\ref{hdiagonal}) does not lead to a convergent Hodge operator \cite{mohamed16b}.
\begin{figure}
\centering
\includegraphics[width=6cm]{fig_triangle_circ}
\caption{Primal simplices and their circumcentric dual cells in a 2D mesh composed of a single triangle. $c_i$ is the circumcenter of the primal edge $e_i$, for $i=1,2,3$, and $c_t$ is the circumcenter of the triangle. $e_i^*$ is perpendicular to $e_i$.}
\label{triangle_circ}
\end{figure}
In the next section, we propose an alternative discrete Hodge operator.
\section{Analytically constructed discrete Hodge}
In the sequel, we restrict to the case of a bidimensional mesh in $ℝ^2$. We will fix a Cartesian orthonormal frame and denote $v_x$ and $v_y$ the components of a vector $\vec v$ in this frame. In the next subsection, we present the construction of a discrete Hodge on 1-forms. We require that it is exact on locally constant forms.
\subsection{Construction in a one-triangle mesh}
To simplify, consider an oriented bidimensional mesh composed of a single triangle $T$, as in Figure \ref{triangle}. In this figure, the center $c_i$ of the edge $e_i$ is an arbitrary point lying on the edge. The center $c_t$ of the triangle $T$ is also an arbitrary point inside $T$. In the particular case where $c_i$ is the middle point of $e_i$ and $c_t$ is the barycenter (resp. circumcenter, incenter) of $T$, the dual mesh is said barycentric (resp. circumcentric, incentric).
\begin{figure}
\centering
\includegraphics[width=6cm]{fig_triangle_any}
\caption{Primal simplices and their arbitrary-centered dual cells in a 2D mesh composed of a single triangle. The $c_i$'s and $c_t$ are respectively arbitrary interior points of the edges $e_i$'s and of the triangle. The triangle is oriented counterclockwise. Arrows indicate the orientations of the primal edges and the induced orientations of dual edges. The angles $θ_i$ defined in (\ref{angle}) are drawn in red.
}
\label{triangle}
\end{figure}
Let us denote $\vec{e}_i$ and $\vec{e}_i^*$ the vectors representing the oriented edges $e_i$ and $e_i^*$ respectively, and \[\vec u_i=\dfrac{\ve_i}{\|\ve_i\|}\quad\quad\text{and}\quad\quad \vec u_i^*=\dfrac{\ve_i^*}{\|\ve_i^*\|}\] the corresponding unit vectors.
These two vectors are related by a rotation:
\[ \vecteurd{u_{ix}^*}{u_{iy}^*}=\matriced{\cos\theta_i}{-\sin\theta_i}{\sin\theta_i}{\cos\theta_i}\vecteurd{u_{ix}}{u_{iy}} \]
where
\begin{equation}
\theta_i=\widehat{(\vec u_i,\vec u_i^*)}
\label{angle}
\end{equation}
is the angle that they form (see Figure \ref{triangle}).
Consider a differential 1-form
\[ ω=aｄx+bｄy \]
where $a$ and $b$ are scalar functions of the position. Its image by the continuous Hodge star is (see equation (\ref{star2d})):
\[ \star ω=-bｄx+aｄy. \]
We denote $ω_i=⟨ω,e_i⟩$ and $ω^*_i=⟨\star ω,e_i^*⟩$ the values of the corresponding primal and dual cochains.
Assume that $\omega$ is constant inside the triangle. Then
\begin{eqnarray*}
\omega_i^*&=&\int_{e^*_i}\!\star \omega\ =\ \star ω(u_i^*)\ |e_i^*|\ =(-bu_{ix}^*+au_{iy}^*)\ \ |e_i^*| \\
&=& \bigg((au_{ix}+bu_{iy})\ \sin\theta_i+ (-bu_{ix}+au_{iy})\ \cos\theta_i\bigg)|e_i^*|\\
&=& \bigg(ω(\vec u_i)\ \sin\theta_i+ ω(-\tsr J\vec u_i)\ \cos\theta_i\bigg)|e_i^*|,\\
&=& \bigg(÷{ω(\ve_i)}{|e_i|}\ \sin\theta_i+ ÷{ω(-\tsr J\ve_i)}{|\tsr J\ve_i|}\ \cos\theta_i\bigg)|e_i^*|,
\end{eqnarray*}
$\tsr J$ being the skew-symmetric matrix \[\hspace{1 cm}\tsr J=\matriced {0}{-1}{1}{0}.\]
Knowing that $|\tsr J\ve_i|=|\ve_i|$, we have:
\begin{equation}
\omega_i^*=\dfrac{|e_i^*|}{|e_i|}\omega_i\ \sin \theta_i\ \ +\ \ \dfrac{|e_i^*|}{|e_i|}\omega(-\tsr J\ve_i)\ \cos \theta_i.
\label{swi}
\end{equation}
Let's take $i=1$. As soon as the triangle is not degenerate, there exists reals $a^2_1$ and $a_1^3$ such that
\[ -\tsr J\ve_1=a^2_1 \ \ve_2+a^3_1\ \ve_3.\]
So,
\begin{equation}
\omega_1^*=\dfrac{|e_1^*|}{|e_1|}\ \omega_1\ \sin \theta_1\ \ +\ \ \dfrac{|e_1^*|}{|e_1|}\ a_1^2\omega_2\ \cos \theta_1\ \ +\ \ \dfrac{|e_1^*|}{|e_1|}\ a_1^3\omega_3\ \cos \theta_1 .
\label{a}
\end{equation}
Using similar arguments for $i=2$ and 3, we get:
\[\begin{bmatrix}
\omega^*_1\\\\\omega^*_2\\\\\omega^*_3
\end{bmatrix}
=
\begin{bmatrix}
\displaystyle\dfrac{|e_1^*|}{|e_1|}&0&0\\
0&\displaystyle\dfrac{|e_2^*|}{|e_2|}&0\\
0&0&\displaystyle\dfrac{|e_3^*|}{|e_3|}
\end{bmatrix}
\begin{bmatrix}
\sin \theta_1 &a_1^2\cos\theta_1&a_1^3\cos\theta_1 \\\\
a_2^1\cos\theta_2&\sin \theta_2 &a_2^3\cos\theta_2 \\\\
a_3^1\cos\theta_3&a_3^2\cos\theta_3&\sin \theta_3
\end{bmatrix}
\begin{bmatrix}
\omega_1\\\\\omega_2\\\\\omega_3
\end{bmatrix}.
\]
This relation makes the discrete Hodge star operator appear as a product of two matrices. The first one contains ratios of dual and primal volumes. In case of circumcentric dual mesh, the second matrix reduces to the identity matrix and we retrieve the usual diagonal Hodge (\ref{hdiagonal}).
Of course, one can rewrite the angles $θ_i$ and the coefficients $a_i^j$ as functions of the $\ve_i$ and $\ve_i^*$. For example,
\[ \cos\theta_i=\dfrac{\ve_i\cdotp\ve_i^*}{|e_i|\,|e_i^*|},\quad \sin\theta_i=\dfrac{\ve_i\dtimes\ve_i^*}{|e_i|\,|e_i^*|} ,\quad\quad\theta_i\in[0,\pi]\]
where $\vec a·\vec b$ is the inner product of $\vec a$ and $\vec b$. The dotted cross product is defined as the third component of the usual cross product of the extension of $\vec a$ and $\vec b$ to $ℝ^3$:
\[\vec a\dtimes\vec b=a_xb_y-a_yb_x\]
which can also be considered as the real-valued 2D cross product.
Moreover, it is straightforward to show that, for $i≠j$,
\[ a^i_j=\dfrac{\ve_j\cdotp\ve_k}{\ve_i\dtimes\ve_k},\]
where $k≠i,j$. With these relations, the matrix of the Hodge operator acting on discrete primal 1-form is
\[
H =
\begin{bmatrix}
\di\dfrac{\ve_1\dtimes\ve_1^*}{|e_1|^2}&&\di \dfrac{\ve_1\cdotp\ve_1^*}{|e_1|^2}\dfrac{\ve_1\cdotp\ve_3}{\ve_2\dtimes\ve_3}&&\di \dfrac{\ve_1\cdotp\ve_1^*}{|e_1|}\dfrac{\ve_1\cdotp\ve_2}{\ve_3\dtimes\ve_2} \\[15pt]
\di\dfrac{\ve_2\cdotp\ve_2^*}{|e_2|^2}\dfrac{\ve_2\cdotp\ve_3}{\ve_1\dtimes\ve_3}&&\di\dfrac{\ve_2\dtimes\ve_2^*}{|e_2|^2}&&\di \dfrac{\ve_2\cdotp\ve_2^*}{|e_2|^2} \dfrac{\ve_2\cdotp\ve_1}{\ve_3\dtimes\ve_1}\\[15pt]
\di\dfrac{\ve_3\cdotp\ve_3^*}{|e_3|^2}\dfrac{\ve_3\cdotp\ve_2}{\ve_1\dtimes\ve_2}&&\di \dfrac{\ve_3\cdotp\ve_3^*}{|e_3|^2}\dfrac{\ve_3\cdotp\ve_1}{\ve_2\dtimes\ve_1}&&\di\dfrac{\ve_3\dtimes\ve_3^*}{|e_3|^2}
\end{bmatrix}.
\]
\subsection{Examples and remarks\label{sec:examples}}
As illustration, consider the right triangle presented in Figure \ref{rectangle}. The circumcentric dual (left of the figure) is singular because the dual $e_3^*$ of $e_3$ is empty and its measure vanishes. This leads to a singular matrix $H$, as in the diagonal Hodge. With the barycentric dual (Figure \ref{rectangle}, right), $H$ reduces to the following non-singular matrix:
\begin{figure}
\centering
\includegraphics[width=40mm]{fig_rectangle_circ.pdf}
\hspace{1cm}
\includegraphics[width=40mm]{fig_rectangle_bary.pdf}
\caption{Right triangle with side lengths $m$ and $n$. Left: with a circumcentric dual mesh ($\e_3^*$ has a zero length). Right: with a barycentric dual mesh.}
\label{rectangle}
\end{figure}
\[
H_{barycenter}=\begin{bmatrix}
\di÷n{3m}&&\di÷{m}{6n}&&0
\\[10pt]
\di÷n{6m}&&\di÷m{3n}&&0
\\[10pt]
\di÷{n(m^2-n^2)}{6m(m^2+n^2)}&&\di÷{m(m^2-n^2)}{6n(m^2+n^2)}&&\di÷{mn}{3(m^2+n^2)}
\end{bmatrix}.
\]
In an unit right triangle ($m=n=1$), with a barycentric and an incentric dual meshes, $H$ writes:
\[
H_{barycenter}=÷16\begin{bmatrix}
\di2&\di1&0
\\
\di1&\di2&0
\\
0&0&\di1
\end{bmatrix} ,
\quad\quad
H_{incenter}=÷1{4+2√2}\begin{bmatrix}
\di2&\di√2&0
\\
\di√2&\di2&0
\\
0&0&\di2
\end{bmatrix}.
\label{hodge_examples}
\]
The discrete Hodge operator for 0-forms and for 2-forms are composed of measure ratios as in the diagonal Hodge.
For a general mesh, composed of multiple triangles, an usual assembling permits to deduce the global Hodge matrix. This matrix is sparse. It takes a primal 1-cochain into a dual one. Again due to relation (\ref{starstar}), the Hodge operator acting on dual cochains can simply be represented by the inverse of the assembled matrix. However, a full inversion of this matrix is in general expensive. To avoid this cost, an element-wise inversion will be carried out in numerical tests when needed. Moreover, the element-wise inversion maintains the sparsity of the matrix, contrarily to a full inversion. This may significantly reduce the simulation cost for realistic problems. Despite this approximation, we will see in section \ref{sec:poisson} that the newly defined Hodge provides competitive numerical results.
As can be remarked, the new discrete Hodge matrix is not necessarily symmetric. Yet, some authors claim that an Hogde matrix should be symmetric, since
\begin{equation}
\star ω\wedge θ =\star θ\wedge ω \label{starwedge}
\end{equation}
for two arbitrary 1-differential forms ω and θ, $\wedge$ being the wedge product (\cite{Auchmann06,hiptmair01}). Combined to a choice of volume form, this relation represents an inner product. However, relation (\ref{starwedge}) does not imply that the discrete Hodge has to be symmetric. Likewise, a symmetry of the matrix of the Hodge operator does not help in verifying property (\ref{starwedge}). It is the combination of the discrete Hodge matrix with the discrete wedge operator between two discrete 1-forms which should be symmetric. It may also be advocated that a symmetric Hodge matrix would be more advantageous because it would result in a symmetric problem when discretizing a symmetric operator as in (\ref{starwedge}). This argument is however not convincing since the discrete wedge product may destroy the symmetry.
By definition, the symmetry of the Hodge operator matrix alone would mean, in a single-triangle mesh, that
\begin{equation}
\begin{bmatrix}
θ_1\\θ_2\\θ_3
\end{bmatrix}·⦅H
\begin{bmatrix}
ω_1\\ω_2\\ω_3
\end{bmatrix}⦆
=⦅H
\begin{bmatrix}
θ_1\\θ_2\\θ_3
\end{bmatrix}⦆·
\begin{bmatrix}
ω_1\\ω_2\\ω_3
\end{bmatrix}
\label{sym}
\end{equation}
for any pair $ω$ and $θ$ of primal 1-cochains, where the dot symbol $(\ ·\ )$ designates the Euclidean scalar product of $ℝ^3$. Not only the scalar product between cochains has still to be given a sense but also there is no obvious reason to require condition (\ref{sym}).
In fact, the symmetry requirement may be justified in a finite element approach as proposed by Hiptmair in \cite{hiptmair01} but not in DEC. Indeed, in a finite element framework, it is the mass matrix which is called Hodge operator. Yet, this mass matrix represents, not the Hodge operator, but inner products such as (\ref{starwedge}) integrated over elements. Note also that, as recognized in \cite{Auchmann06,Alotto04,cinalli04}, the symmetry restriction seems to be too strong in DEC context.
Which center gives more accurate results for non-constant 1-forms depends on the form. For instance, consider the differential 1-form
\[ω=(x-y)(ｄx-ｄy).\] Table \ref{tab:error_hodge} lists the error of the discrete Hodge operator $H$ on ω, $H$ being either the barycentric or the incentric Hodge. The error is defined as the $ℓ^2$-norm of the dual cochain $H(ω_K)-(\star ω)_{\star K}$. Two computations are carried out. The first one is on a mesh made up of the single unit triangle, having a mean edge length 1.1381. The second computation is on a square meshed with 722 right and isocele triangles (similar to Figure \ref{right_and_acute_mesh}, right) where the mean edge length is $5.6965·10^{-2}$. Table \ref{tab:error_hodge} shows that, with the considered form, the barycenter-based Hodge is more precise than the inceter-based one on the single-triangle mesh. The same conclusion can be drawn in the discretized square. However, in the latter case, the error difference is smaller. For the differential form \[ω=(x+y)(ｄx+ｄy),\] the situation is reversed. Indeed, Table \ref{tab:error_hodge} shows that for this form, the incentric Hodge is about 1.94 times as accurate as the barycentric one, as well on the single-triangle mesh as in the right-triangularized mesh.
\begin{table}[ht]
\centering
$\def1.5{1.5}
\begin{array}{|c|c|c|c|c|}
\hline
&\multicolumn{2}{c|}{\text{Single triangle}}&\multicolumn{2}{c|}{\text{Right mesh}}
\\\cline{2-5}
\text{Differential form}& \text{Barycentric} &\text{Incentric}&\text{Barycentric}&\text{Incentric}
\\\hline
(x-y)(ｄx-ｄy)&0.2946&0.3232&1.5243·10^{-2}&1.5715·10^{-2}
\\\hline
(x+y)(ｄx+ｄy)&0.0589&0.0303&6.6882·10^{-4}&3.4424·10^{-4}
\\\hline
\end{array}
$
\caption{$ℓ^2$-norm of the error of the discrete Hodge operators on the unit right triangle and on a right-triangularized square with 20 points in each direction}
\label{tab:error_hodge}
\end{table}
This very simple test shows that it is not easy to predict which center will give the best results, in terms of precision. Moreover, in more concrete problems, there may be other constraints than the accuracy to account for. The freedom to choose the center makes it conceivable to carry out an optimization of the discrete Hodge operator.
Lastly, let us remark that, in the previous development, there is no restriction on the position of the centers of the simplices, as long as the dual mesh is not degenerate. For example, the center $c_t$ may be outside the triangle, or even outside the domain in the case of multiple-triangle mesh.
In the following section, numerical experiments on (linear) Poisson equation and applications to (non-linear) isothermal and anisothermal fluid flow problems are presented. We will in particular focus on the convergence with the circumcenter-, the barycentrer- and the incenter-based Hodges on different types of mesh.
\section{Numerical results\label{sec:numerical}}
We begin with numerical tests on Poisson problem.
\subsection{Poisson equation\label{sec:poisson}}
In this section, we deal with Poisson equation with Dirichlet boundary conditions in a flat bidimensional domain $M$:
\begin{equation} \label{Poisson_Eq}
\begin{cases}\displaystyle
-\Delta u =f \quad \text{ in } M,\\
u=g \quad\quad\text{\quad on }\partial M,
\end{cases}
\end{equation}
where $f$, $g$ and the unknown $u$ are scalar functions. In exterior calculus formulation, this equation writes
\begin{equation}
\label{eq:poisson_ext}
\begin{cases}\displaystyle
\starｄ\star ｄu =f \quad \text{ in } M,\\
u=g\quad\text{ \quad on }\partial M.
\end{cases}
\end{equation}
In the tests, $M$ is the unit square. Two meshes are considered. The first one is well-centered (see Figure \ref{right_and_acute_mesh}, left) and based on a pattern proposed in \cite{cassidi81}. It will be called ``acute mesh''. On this mesh, the usual diagonal Hodge and the new Hodge operator based on barycentric and on incentric duals will be compared. The second mesh, called ``right mesh'' herein, is more natural and is made of right and isocele triangles (see Figure \ref{right_and_acute_mesh}, right). As the diagonal Hodge cannot be defined on it, the new Hodge operator with barycentric and incentric duals will be used.
\begin{figure}[htb!]
\centering
\includegraphics[scale=0.23]{fig_mesh_acute}\quad\quad
\includegraphics[scale=0.23]{fig_mesh_right}
\caption{Typical acute and right triangulations of the unit square}
\label{right_and_acute_mesh}
\end{figure}
The variable $u$ is placed, as cochain, on dual vertices $σ^*∈K^*_2$. This means that the unknown is the array $(⟨u,σ^*⟩)_{σ^*∈K^*_2}$.
A consequence is that $ｄu$, in equation (\ref{eq:poisson_ext}), is discretized into a dual 1-cochain and the right-most $\star$ operator in that equation is then represeted by the inverse of the assembled Hodge matrix. As mentioned, the inversion is done element-wise numerically.
As precision indicator, we consider the relative error
\begin{equation}
E=÷{‖u-u_{exact}‖_{K^*_2}}{‖u_{exact}‖_{K^*_2}}\label{relative_error_def}
\end{equation}
where $‖v‖_{K^*_2}$ designates a discrete norm of a dual 0-cochain $v$, defined as follows:
\begin{equation}
‖v‖^2_{K^*_2}:=∑_{σ^*\in K^*_2}⟨v,σ^*⟩^2|\sigma^*|.\label{norm_def}
\end{equation}
An object-oriented Fortran code was designed for the simulations. A direct LU solver is selected for the resolution. In a first test, the right-hand side functions $f$ and $g$ are chosen such that the exact solution is the quadratic function
\[u_{exact}=x^2+y^2.\]
Figure~\ref{fig:poisson_d_x2y2} shows the evolution of the relative error $E$ with the mean edge length $Δx_{mean}$. As can be seen in Figure \ref{poisson_sin_acute}, barycentric and incentric dual meshes provide better results than the diagonal Hodge, despite the element-wise inversion needed in the discetized Laplace operator. Note that many triangles are close to right. However, the convergence rates are almost the same with the three types of dual mesh. The left part of Table \ref{tab:poisson_d_x2y2} reveals an almost quadratic convergence.
\begin{figure}
\centering
\begin{subfigure}{\figwidth}
\includegraphics[width=\textwidth]{fig_poisson_d_x2y2_acute_error}
\caption{Acute mesh}\label{poisson_sin_acute}
\end{subfigure}
\begin{subfigure}{\figwidth}
\includegraphics[width=\textwidth]{fig_poisson_d_x2y2_right_error}
\caption{Right mesh}
\end{subfigure}
\caption{Poisson equation with with $u_{exact}=x^2+y^2$: error evolution}
\label{fig:poisson_d_x2y2}
\end{figure}
\begin{table}[ht]
\centering
\begin{tabular}[]{|l|c|c|}\hline
&Acute mesh&Right mesh\\\hline
Circumcentric dual&1.995&--\\\hline
Barycentric dual&1.985&1.923\\\hline
Incentric dual&1.992&1.921\\\hline
\end{tabular}
\caption{Poisson equation with with $u_{exact}=x^2+y^2$: convergence rate}
\label{tab:poisson_d_x2y2}
\end{table}
With the right mesh, the results are less accurate but are still very acceptable as can be observed in Figure \ref{fig:poisson_d_x2y2}b. With the barycentric dual, the relative error is about $7.55·10^{-5}$ when $Δx_{mean}\simeq 2.3·10^{-2}$ (whereas it was $2.28·10^{-5}$ with an acute mesh). The difference between the barycentric and the incentric duals are more pronounced. However, the convergence rates are similar, as can be stated in Table \ref{tab:poisson_d_x2y2}, right.
In a second test, the exact solution is \[u_{exact}=\sin (πx)\,\sinh (πy),\] such that the right-hand side $f$ vanishes. Figure \ref{fig:poisson_d_sin}a shows that, with the acute mesh, the three types of dual present close precisions.
The barycentric solution is a slightly more accurate with the right mesh whereas the incentric solution has clearly lost precision. They have however almost the same convergence rate, as can be seen in Table \ref{tab:poisson_d_sin}.
\begin{figure}
\centering
\begin{subfigure}{\figwidth}
\includegraphics[width=\textwidth]{fig_poisson_d_sin_acute_error}
\caption{Acute mesh}
\end{subfigure}
\begin{subfigure}{\figwidth}
\includegraphics[width=\textwidth]{fig_poisson_d_sin_right_error}
\caption{Right mesh}
\end{subfigure}
\caption{Poisson equation with $u_{exact}=\sin (πx)\,\sinh (πx)$: error evolution}
\label{fig:poisson_d_sin}
\end{figure}
\begin{table}[ht]
\centering
\begin{tabular}[]{|l|c|c|}\hline
&Acute mesh&Right mesh\\\hline
Circumcentric dual&1.975&--\\\hline
Barycentric dual&1.979&1.809\\\hline
Incentric dual&1.982&1.840\\\hline
\end{tabular}
\caption{Poisson equation with $u_{exact}=\sin (πx)\,\sinh (πx)$: convergence rate}
\label{tab:poisson_d_sin}
\end{table}
From these two experiments, it is obvious that when the mesh is well-centered, one cannot predict if the diagonal Hodge is more precise than other discrete Hodge operators. It depends on the solution. By contrast, the barycentric dual gave better accuracy than the incentric dual.
In the previous tests, the equation is linear. In the next subsections, tests on non-linear systems in fluid dynamics domain are presented.
\subsection{Incompressible fluid flow}
Consider an incompressible Newtonian fluid flow of density $\rho$ and kinematic viscosity $\nu$ in a bidimensional domain. The evolution of the velocity $u$ and the pressure $p$ is governed by the Navier-Stokes equations:
\begin{equation}
\begin{cases}\displaystyle
\dfrac{\partial u}{\partial t}+(u·∇)u + \dfrac {1}{\rho}\grad p-\nu\Delta u = 0,\\\\
\dive u =0
\end{cases}
\label{ns}
\end{equation}
To ensure a divergence-free velocity, a stream function formulation is used. Moreover, to get rid of the pressure, a curl is applied on the momentum equation.
The resulting equation is rewritten in exterior calculus formulation and discretized as in \cite{mohamed16a}, (see also equations (\ref{nsaext})). The discretized stream function is placed on primal vertices and, consequently, velocity is discretized into a dual 1-cochain.
\subsubsection{Poiseuille flow}
We consider a Dirichlet problem in the unit square with a Poiseuille flow as exact solution. The accuracy on the stream function ψ and on the velocity $u$ is measured with the relative errors defined similarly as in (\ref{relative_error_def})-(\ref{norm_def}) but with the appropriate simplex/cell set in place of $K_2^*$.
The numerical initial flow is set to zero. The simulation is run until the flow stabilizes. The convergence of the principal variable ψ with the mean edge length is presented in Figure \ref{fig:poiseuille_psi}. It shows a higher precision of the diagonal Hodge in the acute triangulation in the given range of $Δx_{mean}$, but a slightly higher speed of convergence with the new Hodge with barycentric and incentric duals. It is confirmed by the convergence rates in Table \ref{tab:poiseuille_psi_u}, second column. For the velocity however, the circumcentric dual has both a higher precision and a higher convergence speed, as can be observed in Figure \ref{fig:poiseuille_u_acute} and in the third column of Table \ref{tab:poiseuille_psi_u}. Note that a first order interpolation is used for the reconstruction of the approximate velocity field from the discrete 1-form $\star ｄψ$.
\begin{figure}
\centering
\begin{subfigure}{\figwidth}
\includegraphics[width=\textwidth]{fig_poiseuille_acute_error_psi}
\caption{Acute mesh}
\end{subfigure}
\begin{subfigure}{\figwidth}
\includegraphics[width=\textwidth]{fig_poiseuille_right_error_psi}
\caption{Right mesh}
\end{subfigure}
\caption{Poiseuille flow: error on $ψ$}
\label{fig:poiseuille_psi}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}{\figwidth}
\includegraphics[width=\textwidth]{fig_poiseuille_acute_error_u}
\caption{Acute mesh}\label{fig:poiseuille_u_acute}
\end{subfigure}
\begin{subfigure}{\figwidth}
\includegraphics[width=\textwidth]{fig_poiseuille_right_error_u}
\caption{Right mesh}
\end{subfigure}
\caption{Poiseuille flow: error on $u$}
\label{fig:poiseuille_u}
\end{figure}
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|c|c|c|}\hline
&\multicolumn{2}{c|}{Acute mesh}&\multicolumn{2}{c|}{Right mesh}\\\cline{2-5}
&Stream function&Velocity&Stream function&Velocity\\\hline
Circumcentric&2.006 & 1.421&--&--\\
Barycentric&2.184 & 1.106& 1.989 & 1.553\\
Incentric&2.185 & 1.096& 1.989 & 1.539\\
\hline
\end{tabular}
\caption{Poiseuille flow: convergence rates}
\label{tab:poiseuille_psi_u}
\end{table}
In the case of the right triangulation, the barycentric and the incentric duals provide similar precisions, as well for the stream function than for the velocity. For ψ, the errors are visually indistinguishable in Figure \ref{fig:poiseuille_psi}b. For the velocity, the convergence rate is about 1.5, which is higher than the convergence rate of the circumcentric dual in the acute triangulation case (see Table \ref{tab:poiseuille_psi_u}, last column).
\subsubsection{Taylor-Green vortex}
We now simulate a steady Taylor-Green vortex corresponding to the vector field
\begin{equation}u=
\begin{pmatrix}
-\cos x\,\sin y\\\sin x\cos y
\end{pmatrix}
\end{equation}
in the domain $[-π,π]×[-π,π]$. The kinematic viscosity and the density are set to 1.
As previously, the numerical initial flow is set to zero and the simulation is run until the flow stabilizes.
The error on ψ is plotted in Figure \ref{fig:taylor_psi} against the typical edge length. It shows that the decrease of the error is the same for the three types of dual mesh, as well in the acute mesh as in the right mesh. The convergence rate is close to 2.0 in all cases, as reported in Table \ref{tab:taylor_green_convergence_rates}.
\begin{figure}
\centering
\begin{subfigure}{\figwidth}
\includegraphics[width=\textwidth]{fig_taylor_acute_error_psi}
\caption{Acute mesh}
\end{subfigure}
\begin{subfigure}{\figwidth}
\includegraphics[width=\textwidth]{fig_taylor_right_error_psi}
\caption{Right mesh}
\end{subfigure}
\caption{Taylor-Green vortex: error on $ψ$}
\label{fig:taylor_psi}
\end{figure}
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|c|c|c|}\hline
&\multicolumn{2}{c|}{Acute mesh}&\multicolumn{2}{c|}{Right mesh}\\\cline{2-5}
&Stream function&Velocity&Stream function&Velocity\\\hline
Circumcentric&1.996 & 1.187&--&--\\
Barycentric&2.065 & 1.130& 2.018 & 1.735\\
Incentric&2.088 & 1.118& 2.067 & 1.736\\
\hline
\end{tabular}
\caption{Taylor-Green vortex: convergence rates}
\label{tab:taylor_green_convergence_rates}
\end{table}
For the velocity, differences can be observed between the acute and the right meshes. Indeed, it can be seen in Figure \ref{fig:taylor_u} that the decrease is faster in the right mesh. The convergence rate is about 1.74 (for both barycentric and incentric duals) while it is close to 1.14 in the acute mesh (for the three dual types), as can be seen in Table \ref{tab:taylor_green_convergence_rates}.
\begin{figure}
\centering
\begin{subfigure}{\figwidth}
\includegraphics[width=\textwidth]{fig_taylor_acute_error_u}
\caption{Acute mesh}
\end{subfigure}
\begin{subfigure}{\figwidth}
\includegraphics[width=\textwidth]{fig_taylor_right_error_u}
\caption{Right mesh}
\end{subfigure}
\caption{Taylor-Green vortex: error on $u$}
\label{fig:taylor_u}
\end{figure}
In the next test, we deal with an anisothermal fluid flow.
\subsection{Anisothermal fluid flow\label{sec:anisothermal}}
In the case of anisothermal fluid flow with thermal expansion coefficient $\beta$ and thermal diffisivity $\kappa$, and in the limit of the Boussinesq approximation, the evolution of the velocity, pressure and temperature $\theta$ is governed by the equations:
\begin{equation}
\begin{cases}\displaystyle
\dfrac{\partial u}{\partial t}+\dive(u\otimes u) + \dfrac {1}{\rho}\grad p-\nu\Delta u + \beta g \theta\e_y = 0,\\\\
\dive u =0,\\\\
\dfrac{\partial \theta}{\partial t}+\dive(u\theta)-\kappa\Delta\theta =0
\end{cases}
\label{nsa}
\end{equation}
where $-g\e_y$ is the gravity acceleration vector. Let $ω=u^\flat$. In exterior calculus framework, equation (\ref{nsa}) writes
\begin{equation}
\begin{cases}\displaystyle
\dfrac{\partial ω}{\partial t} +\iota_uｄω+ ÷1{ρ}ｄ\overline{p}+ν\starｄ\star ｄω+βgθｄy = 0,
\\\\
\star ｄ\starω=0,
\\\\
\dfrac{\partial \theta}{\partial t}+\starｄ\star (θω)-κ\starｄ\star ｄθ=0.
\end{cases}
\label{nsaext}
\end{equation}
In (\ref{nsaext}), $\iota$ is the interior product and $\overline{p}=p +÷12ρu^2$ is the dynamic pressure. The exterior derivative operator $ｄ$ is applied to the first equation of (\ref{nsaext}) to get rid of the dynamic pressure. A stream-function formulation is used. This means that the equations are solved in the stream function $ψ$ such that $\star ω=ｄψ$ instead of ω, and in the temperature θ. The time discretization scheme is similar to that in \cite{amses18}, page 55.
Assume that $\nu\neq \kappa$. We choose the following traveling wave solution as reference:
\begin{equation}
\begin{cases}\displaystyle
u_x=u_1\e^{\lambda\xi/\nu}+\dfrac{\kappa^2\beta abg}{(c+w)^2(\kappa-\nu)}\theta_1\e^{\lambda\xi/\kappa}
\\\\\displaystyle
u_y=\dfrac{w-au_x}b,
\\\\\displaystyle
p=-\dfrac{\rho\beta bg\kappa\theta_1}{c+w}\e^{\lambda\xi/\kappa}
\\\\\displaystyle
\theta=\theta_1\e^{\lambda\xi/\kappa}
\end{cases}
\label{travel1}
\end{equation}
where the traveling wave variable is
\begin{equation}
\xi=ax+by+ct
\label{var}
\end{equation}
and $a$, $b$, $c$, $w$, $u_1$, $λ$ and $θ_1$ are constants such that $b\neq 0$, $w≠-c$ and $λ=(c+w)/(a^2+b^2)$.
The space domain is $\left[-\frac 12,\frac 12\right] \times\left[-\frac 12,\frac 12\right]$.
The fluid is characterized by $\rho=1$, $\nu=0.2$, $\kappa=0.1$ and $\beta=1$.
The gravity acceleration is set to $g=10$. The values of the different constants in the reference solution are set to $a=b=1$, $c=-1$, $w=2$, $θ_1=\e^{-5}$ and $u_1=2\e^{-5}$. The domain is discretized with a right triangulation (Figure \ref{right_and_acute_mesh}, right). The mean edge length is $1.894\cdot10^{-2}$. The numerical simulation is carried out until time $T$ such that $|\lambda c|T/\kappa=1$.
Figure \ref{travel_psi_error} reproduces the relative errors on the stream function along $x$-axis. Even if this figure presents some weird plateaux due to lack of resolution (and some machine round-off), it clearly shows a good precision with both dual meshes. At worst, the error is 0.025 percent of the norm of the exact stream function along the axis. The mean error over the whole domain is still smaller. Indeed, as listed in Table \ref{tab:travel_error}, the overall (root mean square over the whole domain) error is only 0.008875 percent of the norm of the exact solution with the incentric dual and 3.3 times smaller with the barycentric one.
\begin{figure}
\centering
\includegraphics[width=\figwidth]{fig_traveling_psi_error}
\caption{Traveling wave. Profile of the relative error on the stream function along $y=0$ and at $t=T$}
\label{travel_psi_error}
\end{figure}
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|c|c|c|c|}\hline
Dual mesh&Stream function& Velocity&Temperature \\\hline
Barycentric&$2.651\cdot 10^{-5}$&$7.270\cdot10^{-5}$&$5.529\cdot10^{-3}$\\\hline
Incentric&$8.875\cdot 10^{-5}$&$2.132\cdot10^{-4}$&$5.589\cdot10^{-3}$\\\hline
\end{tabular}
\caption{Traveling wave. Mean relative error}
\label{tab:travel_error}
\end{table}
Since the velocity is computed a posteriori from the stream function, the same conclusion can be drawn on the error. Indeed, the barycentric dual is more precise along the $x$-axis as shown in Figure \ref{travel_u_error}. It is also more precise in mean in the whole domain, as can be checked in Table \ref{tab:travel_error}.
\begin{figure}
\centering
\includegraphics[width=\figwidth]{fig_traveling_u_error}
\caption{Traveling wave. Profile of the relative error on the horizontal velocity along $y=0$ and at $t=T$}
\label{travel_u_error}
\end{figure}
By contrast, for the temperature, the two dual meshes have a similar precision. The error profiles along $x$-axis, presented in Figure \ref{travel_theta_error}, have the same trend. The overall errors are almost equal as can be seen in the last column of Table \ref{tab:travel_error}. In mean, the relative error is about 0.56 percent. These results reveals a good precision of the new discrete Hodge star on a heat transfer problem, with different choices of dual meshes. Convergence rates will be analyzed in the next section.
\begin{figure}
\centering
\includegraphics[width=\figwidth]{fig_traveling_theta_error}
\caption{Traveling wave. Profile of the relative error on the temperature along $y=0$ and at $t=T$}
\label{travel_theta_error}
\end{figure}
\subsection{Non-structured meshes\label{sec:nonstructured}}
To ease the lecture of the article, only two types of primal mesh were considered in the above tests. The first ones were well-centered meshes such that the diagonal Hodge can be used. The second type of mesh was composed of right meshes. These meshes have structured patterns. In this following section, we show that the constructed Hodge gives also satisfactory results on various non-structured and non well-centered meshes. To simplify, only results on barycentric duals are presented.
We set $ν=κ$ and choose the following exact solution:
\begin{equation}
\begin{cases}\displaystyle
u_x=\dfrac{\beta abg\theta_1}{(c+w)(a^2+b^2)}(x_0-\xi)\e^{\lambda\xi/\kappa}
\\\\\displaystyle
u_y=\dfrac{w-au_1}b,
\\\\\displaystyle
p=-\dfrac{\rho\beta bg\kappa\theta_1}{c+w}\e^{\lambda\xi/\kappa}
\\\\\displaystyle
\theta=\theta_1\e^{\lambda\xi/\kappa}
\end{cases}
\label{travel2}
\end{equation}
where, as before, $ξ=ax+by+ct$ is the traveling wave variable.
For the simulation, the fluid characteristics are $\rho=1$, $g=10$, $\nu=\kappa=0.1$, $\beta=1$. We choose $a=b=1$, $c=-1$, $x_0=÷12$ and $w=0$. The spatial domain is $[0,1]\times[-\frac{1}{2},\frac{1}{2}]$.
A first test is carried out on the five meshes shown in Figure \ref{fig:compressed_mesh}. These meshes are increasingly fine, but are all unstructured, non well-centered and non-Delaunay. They have approximately 40 percent of non-Delaunay triangles. Here, a triangle is said non-Delaunay if its circumcircle contains other nodes than its vertices. They are darkened in the figures. Note that a non-Delaunay mesh is always non well-centered \cite{vanderzee10}.
\begin{figure}
\centering
\begin{tabular}{cc}
\includegraphics[width=52mm]{fig_traveling2_compressed_mesh1}
&
\includegraphics[width=52mm]{fig_traveling2_compressed_mesh2}
\\
(a) $Δx_{mean}=0.35620$ & (b) $Δx_{mean}=0.18510$
\\\\
\includegraphics[width=52mm]{fig_traveling2_compressed_mesh3}
&
\includegraphics[width=52mm]{fig_traveling2_compressed_mesh4}
\\
(c) $Δx_{mean}=0.09093$ & (d) $Δx_{mean}=0.04593$
\\\\
\multicolumn{2}{c}{\includegraphics[width=52mm]{fig_traveling2_compressed_mesh5}}
\\
\multicolumn{2}{c}{(e) $Δx_{mean}=0.02285$}
\end{tabular}
\caption{Unstructured meshes with respectively 36.36\%, 40.69\%, 39.41\%, 43.46\% and 44.20\% of non-Delaunay triangles}
\label{fig:compressed_mesh}
\end{figure}
Figure \ref{fig:compressed_convergence} shows the relative overall errors on the stream function and on the temperature with the five meshes. As can be remarked, the error decreases with the mean edge length. The decrease rates, which are reported in Table \ref{tab:compressed_rate}, is about 1.3690 for $ψ$ and 1.1430 for $θ$.
\begin{figure}
\centering
\includegraphics[width=\figwidth]{fig_traveling2_compressed_convergence}
\caption{Unstructured meshes. Convergence of the stream function and the temperature}
\label{fig:compressed_convergence}
\end{figure}
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|c|}\hline
&Stream function&Temperature \\\hline
Convergence rate &1.3690 &1.1430\\\hline
\end{tabular}
\caption{Unstructured meshes. Convergence rates of the stream function and the temperature}
\label{tab:compressed_rate}
\end{table}
The relative error profiles on the horizontal velocity, the stream function and the temperature along the $x$-axis are plotted on Figure \ref{fig:compressed_error}. For simplicity, only the results with the finest mesh, that is mesh (e) of Figure \ref{fig:compressed_mesh}, are shown. The mean edge length in this mesh is $Δx_{mean}=2.285·10^{-2}$. Figure \ref{fig:compressed_error} shows that the errors along the axis remains very small compared to the norm of the exact solutions. The overall errors are recorded in Table \ref{tab:compressed_error}. As can be stated, the error on the velocity is about 2.316 percent of the norm of the exact solution, whereas the error on the stream function and on the temperature are less than 0.4 percent.
\begin{figure}
\centering
\includegraphics[width=85mm]{fig_traveling2_compressed_error_ux_y}
\includegraphics[width=85mm]{fig_traveling2_compressed_error_psi_y}
\includegraphics[width=85mm]{fig_traveling2_compressed_error_theta_y}
\caption{Relative error on mesh (e) of Table \ref{fig:compressed_mesh}}
\label{fig:compressed_error}
\end{figure}
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|c|c|c|c|}\hline
&Stream function &Temperature &Velocity \\\hline
Relative error &$3.581\cdot 10^{-3}$ &$3.682\cdot10^{-3}$&$2.316\cdot10^{-2}$ \\\hline
\end{tabular}
\caption{Relative errors on mesh (e) of Table \ref{fig:compressed_mesh}}
\label{tab:compressed_error}
\end{table}
The meshes used for the previous results are unstructured and highly non-Delaunay. However, they are rather regular in the sense that the triangles have comparable areas and the edges have comparable lengths. In a last test, an analysis on non-Delaunay meshes with some very-badly-shaped triangles are carried out. These meshes are obtained from Delaunay meshes generated by Gmsh \cite{guezaine09}, of which some random vertices are moved, in order to obtain a prescribed ratio of non-Delaunay triangles. Contrarily to the distortion procedure in \cite{mohamed18}, the non-Delaunay triangles are not necessarily pairwise.
Three sets of four increasignly fine meshes are used. In each set, the meshes count respectively 40, 184, 676 and 2658 triangles. The four meshes in the first set comprise approximately 15 percent of non-Delaunay triangles. They are presented in Figure \ref{fig:random15_mesh}. In the second set, the ratio of non-Delaunay triangles is about 25 percent, whereas in the last set, 50 percent of the triangles are non-Delaunay. The meshes in the second set are plotted in Figure \ref{fig:random25_mesh} and the third set can be seen in Figure \ref{fig:random50_mesh}. The mean edge lengths in the three sets range from $2.546·10^{-1}$ to $3.066·10^{-2}$. It can be remarked in Figures \ref{fig:random15_mesh} to \ref{fig:random50_mesh} that the triangles have irregular shapes. Many of them have an angle close to 180 degrees and the area of the triangles varies very significantly. For instance, the minimum triangle area of mesh (d) in Figure \ref{fig:random25_mesh} is $1.962·10^{-7}$ whereas the maximum is $2.228·10^{-3}$, that is more than $10^4$ times as high.
\begin{figure
\centering
\begin{tabular}{cc}
\includegraphics[width=60mm]{fig_traveling2_random15_mesh1}
&
\includegraphics[width=60mm]{fig_traveling2_random15_mesh2}
\\
(a) $Δx_{mean}=0.2546$ & (b) $Δx_{mean}=0.1167$
\\\\
\includegraphics[width=60mm]{fig_traveling2_random15_mesh3}
&
\includegraphics[width=60mm]{fig_traveling2_random15_mesh4}
\\
(c) $Δx_{mean}=0.0609$ & (d) $Δx_{mean}=0.0312$
\end{tabular}
\caption{Set of meshes with 15\% of non-Delaunay triangles}
\label{fig:random15_mesh}
\end{figure}
\begin{figure
\centering
\begin{tabular}{cc}
\includegraphics[width=60mm]{fig_traveling2_random25_mesh1}
&
\includegraphics[width=60mm]{fig_traveling2_random25_mesh2}
\\
(a) $Δx_{mean}=0.2556$ & (b) $Δx_{mean}=0.1186$
\\\\
\includegraphics[width=60mm]{fig_traveling2_random25_mesh3}
&
\includegraphics[width=60mm]{fig_traveling2_random25_mesh4}
\\
(c) $Δx_{mean}=0.0618$ & (d) $Δx_{mean}=0.0307$
\end{tabular}
\caption{Set of meshes with 25\% of non-Delaunay triangles}
\label{fig:random25_mesh}
\end{figure}
\begin{figure
\centering
\begin{tabular}{cc}
\includegraphics[width=60mm]{fig_traveling2_random50_mesh1}
&
\includegraphics[width=60mm]{fig_traveling2_random50_mesh2}
\\
(a) $Δx_{mean}=0.2682$ & (b) $Δx_{mean}=0.1265$
\\\\
\includegraphics[width=60mm]{fig_traveling2_random50_mesh3}
&
\includegraphics[width=60mm]{fig_traveling2_random50_mesh4}
\\
(c) $Δx_{mean}=0.0650$ & (d) $Δx_{mean}=0.0328$
\end{tabular}
\caption{Set of meshes with 50\% of non-Delaunay triangles}
\label{fig:random50_mesh}
\end{figure}
For all of these meshes, the computation runs without any particular complication. The errors decrease with the mean edge length. Figure \ref{fig:random_convergence} reports the evolution of the error on the stream function and on the temperature with the three sets of meshes. They both show a good convergence. For the first set of meshes, the convergence is close to second order for ψ and is at a rate of 1.5159 for θ, as listed on Table~\ref{tab:random_rate}. For the 25\% non-Delaunay meshes, the convergence rates are slightly smaller but remain very interesting. They are 1.6729 for ψ and 1.2154 for θ. For the set of 50\% non-Delaunay meshes, ψ has a similar convergence rate as in the second set. For the temperature, the result is less satisfying since the convergence rate is smaller than 1.
\begin{figure
\centering
\includegraphics[width=84mm]{fig_traveling2_random15_convergence}\\
\includegraphics[width=84mm]{fig_traveling2_random25_convergence}\\
\includegraphics[width=84mm]{fig_traveling2_random50_convergence}
\caption{Convergence of the stream function and of the temperature. From top to bottom: with the first set (Figure \ref{fig:random15_mesh}), the second set (Figure \ref{fig:random25_mesh}) and the third set (Figure \ref{fig:random50_mesh}) of meshes}
\label{fig:random_convergence}
\end{figure}
\begin{table}[ht]
\centering
\begin{tabular}{|c|c|c|}\hline
&Stream function&Temperature \\\hline
15\% non-Delaunay meshes &1.9005 &1.5159\\\hline
25\% non-Delaunay meshes &1.6729 &1.2154\\\hline
50\% non-Delaunay meshes &1.6591 &0.8660\\\hline
\end{tabular}
\caption{Convergence rate}
\label{tab:random_rate}
\end{table}
The values in Table \ref{tab:random_rate} are plotted in Figure \ref{fig:random_rate} which presents visually the evolution of the convergence rate with the mesh quality. This figure shows that the convergence rate decreases less rapidely than the rate of non-Delaunay triangles.
\begin{figure
\centering
\includegraphics[width=\figwidth]{fig_traveling2_random_rate}
\caption{Evolution of the convergence rate with the ratio of non-Delaunay triangles}
\label{fig:random_rate}
\end{figure}
\section{Conclusion}
We proposed a new construction of discrete Hodge operator in the context of DEC. It requires neither a well-centered mesh nor a circumcentric dual as with the diagonal Hodge. It also does not require a barycentric dual as in the geometric Hodge. The new Hodge is exact for piecewise constant forms, whichever interior points are chosen as centers.
As shown by experiments with Poisson's equation, even in a well-centered mesh, the popular diagonal Hodge build on an orthogonal dual does not always give more accurate results than the new discrete Hodge. Anyway, the convergence rates are very close. Next, the new discrete Hodge competes well with the diagonal Hodge, in terms of convergence rate, in the resolution of the non-linear Navier-Stokes equations.
The numerical tests in sections \ref{sec:anisothermal} and \ref{sec:nonstructured} showed the ability of DEC, and of the new discrete Hodge in particular, to solve problems involving thermal transfer.
Tests on various type of meshes were carried out. Structured acute (well-centered) meshes and right meshes were chosen in sections \ref{sec:poisson} to \ref{sec:anisothermal}. Unstructured but rather regular meshes, containing about 40 percent of non-Delaunay triangles were considered in section \ref{sec:nonstructured}. Sets of unstructured meshes with very irregular triangle shape, and comprising 15\%, 25\% or 50\% of non-Delaunay triangles were also used. In all of these configurations, the new Hodge star operator converges well.
Experiments in sections \ref{sec:poisson} to \ref{sec:anisothermal} tend to show that a barycentric dual provides as precise as, or better results than an incentric dual. However, this should not be considered as the general case. Indeed, as shown in section \ref{sec:examples}, the incentric dual may be much more precise than the barycentric one in some cases. Moreover, in other types of problem, there may be other criteria than the error to take into consideration.
Barycenter and incenter were chosen as centers of the dual meshes because of their simplicity. With our general construction, other interior points may also be considered, without breaking the exactness of the discrete Hodge on piecewise constant differential forms. The choice of the interior points may be done algorithmically to minimize some error indicator (for example, the overall error of the discerete Hodge operator on some class of differential forms). The development of such optimization process is curently under study by the authors.
Lastly, extension of the new discrete Hodge operator to the three-dimensional case is under construction.
|
\section{Introduction}
\label{sec:Intro}
The Universe where we reside, to our contemporary knowledge,
is governed by the laws of quantum theory, the information and long-range entanglement, and gravity theory.
Quantum field theory (QFT), especially
gauge field theory, under the name of \emph{Gauge Principle} following Maxwell \cite{Maxwell1865zz}, Hilbert, Weyl \cite{Weyl1929ZPhy},
Pauli \cite{RevModPhys13.203Pauli}, and many pioneers,
forms a foundation of fundamental physics.
Yang-Mills (YM) gauge theory \cite{PhysRev96191YM1954}, generalizing the U(1) abelian gauge group to a non-abelian Lie group,
has been given credits for theoretically and experimentally essential to
describe the Standard Model (SM) physics
\cite{Glashow1961trPartialSymmetriesofWeakInteractions, Salam1964ryElectromagneticWeakInteractions, Weinberg1967tqSMAModelofLeptons}.
The SM of particle physics is a chiral gauge theory in 4d\footnote{{We denote
$d$d for the $d$ spacetime dimensions.
The $d+1$d means the $d$ spatial and 1 time dimensions.
The $n$D means the $n$ space dimensions.}}
encoding three of the four known fundamental forces or interactions
(the electromagnetic, weak, and strong forces, but without gravity).
The SM also classifies all experimentally known elementary particles so far:
Fermions include three generations of quarks and leptons. (See Table \ref{table:SMfermion} for $15 + 1 = 16$ Weyl fermions
in SM for each one of the three generations, where the additional 1 Weyl fermion is the sterile right-handed neutrino).
Bosons include the vector gauge bosons: one electromagnetic force mediator
photon $\gamma$,
the eight strong force mediator
gluon g (denoted the gauge field $A^{\mathrm{a}}$),
the three weak force mediator
$W^\pm$ and $Z^0$
gauge bosons; and the scalar Higgs particle $\phi_H$.
(See Table \ref{table:SMboson} for 12 gauge boson generators for SM.)
While the spin-2 boson, the graviton, has not yet been experimentally verified, and it is not within SM.
Physics experiments had confirmed that at a higher energy of SM,
the electromagnetic and weak forces are unified into an electroweak interaction sector:
Glashow-Salam-Weinberg (GSW) SM
\cite{Glashow1961trPartialSymmetriesofWeakInteractions, Salam1964ryElectromagneticWeakInteractions, Weinberg1967tqSMAModelofLeptons}.
Grand Unifications and
Grand Unified Theories (GUT)
hypothesize that at a further higher energy,
the strong and the electroweak interactions will be unified into an electroweak-nuclear GUT interaction sector.
The GUT interaction is characterized by one larger gauge group
and its force mediator gauge bosons
with a single unified
coupling constant \cite{Georgi1974syUnityofAllElementaryParticleForces, Fritzsch1974nnMinkowskiUnifiedInteractionsofLeptonsandHadrons}.
\footnote{Other than the three known fundamental forces, we also have other GUT forces predicted by GUT.
However, in our present work, we discover a possible new force based on the global anomaly matching,
which we name \text{Topological Force.}
This {Topological Force}
may be a possible new force not included in the known four fundamental forces nor included in GUT.
The {Topological Force} is not fully explored in the prior
particle physics phenomenology
literature, at least not rigorously explored based on the cobordism and nonperturbative global anomaly matching framework.
On the other hand, when we consider the constraints from the anomaly matching,
the gravity mostly plays the role of the gravitational \emph{background probed} fields (instead of \emph{dynamical} gravity),
such as in the gravitational anomaly or the mixed gauge-gravitational anomaly.
We mostly leave out any dynamical gravity outside our model, so the present work is far from a Theory of Everything (TOE).
The only exception that we discuss
the influences of dynamical gravity (or the hypothetical particle: graviton,
such as in Quantum Gravity or Topological Gravity)
will be in \Sec{sec:gravity}.
}
In this article, we revisit the
Glashow-Salam-Weinberg SM
\cite{Glashow1961trPartialSymmetriesofWeakInteractions, Salam1964ryElectromagneticWeakInteractions, Weinberg1967tqSMAModelofLeptons},
with four possible gauge groups:
\begin{eqnarray}
G_{\text{SM}_q} \equiv \frac{{\rm SU}(3)\times {\rm SU}(2)\times {\rm U}(1)}{\mathbb{Z}_q}, \quad q=1,2,3,6.
\end{eqnarray}
We revisit the embedding of this special $q=6$ SM (with a gauge group $\frac{{\rm SU}(3)\times {\rm SU}(2)\times {\rm U}(1)}{\mathbb{Z}_q}$)
into
Georgi-Glashow SU(5) GUT with an SU(5) gauge group \cite{Georgi1974syUnityofAllElementaryParticleForces} (with 24 Lie algebra generators for SU(5) gauge bosons shown in Table \ref{table:SMboson}),
and Fritzsch-Minkowski SO(10) GUT with a Spin(10) gauge group \cite{Fritzsch1974nnMinkowskiUnifiedInteractionsofLeptonsandHadrons}.
Our main motivation to revisit these well-known SM and GUT models, following our prior work \cite{WanWang2018bns1812.11967, WW2019fxh1910.14668},
is that the recent systematic Freed-Hopkins' cobordism classification of topological invariants \cite{Freed2016}
can be applied to classify
all the invertible quantum anomalies,
including
\begin{itemize}
\item all \emph{perturbative local anomalies}, and
\item all \emph{nonperturbative global anomalies},
\end{itemize}
which
can further mathematically and rigorously constrain SM and GUT models.
In fact, many earlier works and recent works suggest the cobordism theory is the underlying math structure of invertible quantum anomalies,
see a list of selective References \cite{Atiyah1975jfAPS, Atiyah1976APS, Atiyah1980APS, Freed0607134, Freed2013bordism, Kapustin2014tfa1403.1467, 1405.7689, Kapustin1406.7329, Kitaev2015, Witten:2015aba, 1711.11587GPW,
2019CMaPh.368.1121Y, Freed2019scoFreedHopkins1908.09916, Witten2019bou1909.08775} and an overview therein.
By the completion of all invertible anomalies, we mean that
it is subject to a given symmetry group
\begin{eqnarray}
G\equiv ({\frac{{G_{\text{spacetime} }} \ltimes {{G}_{\text{internal}} }}{{N_{\text{shared}}}}}),
\end{eqnarray}
where the ${G_{\text{spacetime} }}$ is the spacetime symmetry,\footnote{For example, ${G_{\text{spacetime} }}$ can be the Spin($d$), the double cover of Euclidean rotational symmetry SO($d$) for the QFT of a spacetime dimension $d$.}
the ${{G}_{\text{internal}} }$ is the internal symmetry,\footnote{For example, ${{G}_{\text{internal}} }$ can be the
$G_{\text{SM}_q} \equiv \frac{{\rm SU}(3)\times {\rm SU}(2)\times {\rm U}(1)}{\mathbb{Z}_q}, \quad q=1,2,3,6.$}
the $\ltimes$ is a semi-direct product from a ``twisted''
extension,\footnote{The ``twisted''
extension is due to the symmetry extension from ${{G}_{\text{internal}} }$ by ${G_{\text{spacetime} }}$. For a trivial extension,
the semi-direct product $\ltimes$ becomes a direct product $\times$.}
and the ${N_{\text{shared}}}$ is the shared common normal subgroup symmetry between ${G_{\text{spacetime} }}$
and ${{G}_{\text{internal}} }$.
\Refe{WanWang2018bns1812.11967, WW2019fxh1910.14668, WangWen2018cai1809.11171} applies
Thom-Madsen-Tillmann spectra \cite{thom1954quelques,MadsenTillmann4}, Adams spectral sequence (ASS) \cite{Adams1958},
and Freed-Hopkins theorem \cite{Freed2016} in order
to compute, relevant for SM and GUT, the bordism group
\begin{eqnarray}
\Omega_{d}^{G},
\end{eqnarray}
and a specific version of cobordism group (firstly defined to classify Topological Phases [TP] in \cite{Freed2016})
\begin{eqnarray}\label{eq:TPG}
\Omega^{d}_{G} &\equiv&
\Omega^{d}_{({\frac{{G_{\text{spacetime} }} \ltimes {{G}_{\text{internal}} }}{{N_{\text{shared}}}}})}
\equiv
\mathrm{TP}_d(G).
\end{eqnarray}
For a given $G$,
\Refe{WanWang2018bns1812.11967, WW2019fxh1910.14668, WangWen2018cai1809.11171}
find out corresponding all possible topological terms and all possible anomalies relevant for SM and GUT.
See more mathematical definitions, details, and References therein our prior work \cite{WanWang2018bns1812.11967, WW2019fxh1910.14668}.
Along this research direction, other closely related pioneer and beautiful works, by
{Garcia-Etxebarria-Montero} \cite{GarciaEtxebarriaMontero2018ajm1808.00009} and Davighi-Gripaios-Lohitsiri \cite{2019arXiv191011277D},
use a different mathematical tool, based on Atiyah-Hirzebruch spectral sequence (AHSS) \cite{AtiyahHirzebruch1961} and Dai-Freed theorem \cite{Dai1994kqFreed9405012},
also compute the bordism groups $\Omega_{d}^{G}$ and classify possible global anomalies in SM and GUT.\footnote{A practical comment is that
Adams spectral sequence (ASS) used in
\Refe{Freed2016, WanWang2018bns1812.11967, WW2019fxh1910.14668}
turns out to be more powerful than the
Atiyah-Hirzebruch spectral sequence (AHSS) used in \cite{GarciaEtxebarriaMontero2018ajm1808.00009, 2019arXiv191011277D}.
\begin{itemize}
\item
\Refe{WanWang2018bns1812.11967, WW2019fxh1910.14668, Wan2019sooWWZHAHSII1912.13504, Wan2019oaxWWHAHSIII1912.13514, WanWangv2
} based on Adams spectral sequence (ASS) and
Freed-Hopkins theorem \cite{Freed2016}
includes the more refined data, containing both \emph{module} and \emph{group} structure, thus with the advantages
of having less differentials.
In addition,
we can conveniently read and extract the topological terms and co/bordism invariants from the Adams chart and
ASS \cite{WanWang2018bns1812.11967}.
\item
\Refe{GarciaEtxebarriaMontero2018ajm1808.00009, 2019arXiv191011277D} is based on Atiyah-Hirzebruch spectral sequence (AHSS),
which includes more differentials
and some undetermined extensions.
{It is also not straightforward to extract the topological terms and co/bordism invariants directly from the AHSS.}
\end{itemize}
}
As we will see, many SM and GUT with extra discrete symmetries require the twisted $G$,
whose $\Omega_{d}^{G}$ is very difficult, if not simply impossible,
to be determined via AHSS alone \cite{GarciaEtxebarriaMontero2018ajm1808.00009, 2019arXiv191011277D}.
But the twisted $\Omega_{d}^{G}$ can still be relatively easily computed via ASS \cite{WanWang2018bns1812.11967, WanWangv2}.
Therefore, we will focus on the results obtained in \cite{WanWang2018bns1812.11967, WW2019fxh1910.14668}.
In this work, we have the plans and outline as follows, with topics by sections:
\begin{enumerate}[leftmargin=4.mm, label=Sec.~\textcolor{blue}{[\arabic*]}., ref={[\arabic*]}]
\setcounter{enumi}{1}
\item {\bf Overview on SM, GUT, and Anomalies} in \Sec{sec:2Overview}:
We first overview the ingredients of various SM and GUT to set up the stage in \Sec{sec:2OverviewSMGUT}.
Then we comment on the different types of anomalies,
and we clarify the different uses of these anomalies: including (1) anomaly cancellations of dynamical
gauge fields,\footnote{The perturbative parts of calculations are standard on the QFT textbook \cite{Weinberg1996Vol2, PeskinSchroeder1995book, Zee2003book, Srednicki2007book}}
(2) 't Hooft anomaly matching conditions \cite{tHooft1979ratanomaly} of background fields of global symmetries, and others, in \Sec{sec:2OverviewSMGUTdifferentanomalies} and \Fig{fig:triangle}.
We also describe the
perturbative local anomalies vs
nonperturbative global anomalies;
also bosonic vs fermionic anomalies, etc.
These are presented in \Sec{sec:2Overview}.
\item {\bf Dynamical Gauge Anomaly Cancellation} in
\Sec{sec:DynamicalGaugeAnomalyCancellation}:
We explicitly show the classification of all possible (invertible) anomalies of SM gauge groups by cobordism data in Table \ref{table:SU3SU2U1},
and explicitly check all anomaly cancellations for perturbative local anomalies vs
nonperturbative global anomalies in
\Sec{sec:DynamicalGaugeAnomalyCancellation}.
\item {\bf Anomaly Matching for SM and GUT with Extra Symmetries} in
\Sec{Sec:discretesymmetries}:
It is natural to include
additional extra symmetries (such as the ${ \mathbf{B}- \mathbf{L}}$ baryon minus lepton numbers or $X=5({\bf B}- {\bf L})-4Y$ symmetry with the hypercharge $Y$ \cite{Wilczek1979hcZee},
motivated in \cite{GarciaEtxebarriaMontero2018ajm1808.00009} and \cite{WW2019fxh1910.14668})
into SM and SU(5) GUT.
We show the classification of all possible (invertible) anomalies of some SM and GUT with extra symmetries
by cobordism data in Table \ref{table:SU3SU2U1-discrete}.
We explicitly check their anomaly cancellations for perturbative local anomalies vs
nonperturbative global anomalies in
\Sec{Sec:discretesymmetries}.
\item {\bf Beyond Three ``Fundamental'' Forces: Hidden New Topological Force} in \Sec{sec:HiddenTopologicalSectors}:
We recall that to match the 't Hooft anomalies of global symmetries, there are several ways:
\begin{enumerate}
\item Symmetry-breaking:\\
$\bullet$ (say, discrete or continuous $G$-symmetry breaking. Explicitly breaking or spontaneously breaking [may give Nambu-Goldstone modes]).\\
\item Symmetry-preserving: \\
$\bullet$ Degenerate ground states (like the ``Lieb-Schultz-Mattis theorem \cite{Lieb1961frLiebSchultzAOP, Hastings2003zx},'' may host intrinsic topological orders.\footnote{Topological order is in the sense of Wen's definition \cite{Wen2016ddy1610.03911} and References therein.
The gapped and gauged topological order in the colloquial sense can have fractionalized
excitations such as anyons \cite{Wilczek1990BookFractionalstatisticsanyonsuperconductivity}.
Some of the quantum vacua we look for in 4d and 5d may be regarded as a certain version of
topological quantum computer \cite{Kitaev2003, Kitaev2006}.
}),\\
$\bullet$ Gapless, e.g., conformal field theory (CFT).
There are also novel cases where the anomaly and symmetry together enforces the robustness of gapless ground states \cite{CWang1401.1142, Wan2018djlW2.1812.11955, Cordova2019bsd1910.04962}.\\
$\bullet$
Symmetry-preserving topological quantum field theory (TQFT): anomalous symmetry-enriched topological orders.
\item Symmetry-extension \cite{Wang2017locWWW1705.06728}:
The symmetry-extension in any dimension is a helpful intermediate stepstone,
to construct another earlier scenario: \emph{symmetry-preserving TQFT}, via gauging the extended-symmetry \cite{Wang2017locWWW1705.06728}.
\end{enumerate}
We present the possible ways of matching of global anomalies of ``SM and GUT with extra symmetries,'' in \Sec{sec:HiddenTopologicalSectors}.
The condensed matter realization of 't Hooft matching by an anomalous symmetry-preserving gapped TQFT is especially exotic but truly important to us.
It is also known as the surface topological order in the condensed matter literature. {In a lower dimension, the 2+1d surface topological order
is pointed out firstly by Senthil-Vishwanath \cite{VishwanathSenthil1209.3058}. The surface topological order developments are nicely reviewed in \cite{Senthil1405.4015}.}
The particular symmetric gapped boundary spin-TQFT construction of fermionic SPTs given in \cite{GuoJW1812.11959}
will be especially helpful for our construction.\footnote{Our result (3+1d boundary TQFT of 4+1d bulk) may be regarded
as a \emph{one higher dimensional} fundamental physics realization of the condensed matter systems
in \Refe{Fidkowski1305.5851, Metlitski20141406.3032} (2+1d boundary TQFT of 3+1d bulk).}
It turns out that a certain $\mathbb{Z}_{16}$ global anomaly for SM with an extra discrete $X=5({\bf B}- {\bf L})-4Y$ symmetry
may not be matched by the 15 Weyl fermions per generation.
The resolution, other than introducing a sterile right-handed neutrinos (the 16th Weyl fermions),
can also be that introducing a new gapped topological sector matching some of the anomaly,
but preserves the discrete $\mathbb{Z}_{4,X}$ symmetry.\footnote{More precisely,
it turns out that as suggested by \cite{Hsieh2018ifc1808.02881,Cordova1912.13069},
we can only construct
a $\mathbb{Z}_{4,X}$-symmetric gapped 4d TQFT
to saturate the anomaly of $\mathbb{Z}_{16}$ class from the 5d bordism group $\Omega_5^{{\rm Spin} \times_{\mathbb{Z}_2} \mathbb{Z}_4}=\mathbb{Z}_{16}$,
when the anomaly index is an \emph{even} integer $\upnu \in \mathbb{Z}_{16}$.
For example, the SM with three generations of 15 Weyl fermions has the anomaly index $\upnu = - 3 \in \mathbb{Z}_{16}$,
we can introduce one right-handed neutrino to absorb part of the anomaly $\upnu =1 \in \mathbb{Z}_{16}$.
Then, we also introduce a $\mathbb{Z}_{4,X}$-symmetric gapped 4d TQFT to absorb the remained anomaly $\upnu =2 \in \mathbb{Z}_{16}$.
See more details on the anomaly in \Sec{Sec:discretesymmetries}.
See the construction of $\upnu =2 \in \mathbb{Z}_{16}$ TQFT in \Sec{sec:Z16-nu=2-TQFT}.
}
\item {\bf Ultra Unification: Grand Unification $+$ Topological Force/Matter} in \Sec{sec:UU}:
In certain cases of anomaly matching, we require new hidden gapped topological sectors beyond SM and GUT. We may term the
unification including SM, Grand Unification plus additional topological sectors with Topological Force and Matter
as {Ultra Unification}.
We then suggest possible resolutions to sterile neutrinos, neutrino oscillations,
and Dark Matter, in \Sec{sec:UU}.
\end{enumerate}
\noindent
{\bf Notations:}
Throughout our work,
{we follow the same notations as \cite{WanWang2018bns1812.11967, WW2019fxh1910.14668}.
The imaginary number is $\hspace{1pt}\mathrm{i}\hspace{1pt}\equiv \sqrt{-1}$.
The TQFT stands for a topological quantum field theory,
and the iTQFT stands for an invertible topological quantum field theory.
Follow the Freed-Hopkins notation \cite{Freed2016},
we denote the group $G_1 \times_N G_2 \equiv (G_1 \times G_2)/N$ where $N$ is a common normal subgroup of the groups $G_1$ and $G_2$.
We use the standard notation for characteristic classes \cite{milnor1974characteristic}: $c_i$ for the Chern class, $e_n$ for the Euler class,
$p_i$ for the Pontryagin class, and $w_i$ for the Stiefel-Whitney class.
We abbreviate $c_i(G)$, $e_n(G)$, $p_i(G)$, and $w_i(G)$ for the characteristic classes of the associated vector bundle $V_G$ of the principal $G$ bundle
(normally denoted as $c_i(V_{G})$, $e_n(V_{G})$, $p_i(V_{G})$, and $w_i(V_{G})$). For simplicity, we denote
the Stiefel-Whitney class of the tangent bundle $TM$ of a manifold $M$ as $w_j \equiv w_j(TM)$;
if we do not specify $w_j$ with which bundle, then it is for $TM$.
The PD is defined as taking the Poincar\'e dual.}
{All the product notations between cohomology classes are cup product ``$\smile$'',
such as $w_2 w_3 \equiv w_2(TM)w_3(TM)= w_2(TM)\smile w_3(TM)$.
All the product notations between a cohomology class ${\cal A}$ and any fermionic topological invariant ${\eta}$ (say, the
4d eta invariant $\eta$ of Atiyah-Patodi-Singer \cite{Atiyah1975jfAPS, Atiyah1976APS, Atiyah1980APS}), namely
${\cal A} {\eta}$,
are defined as the value of ${\eta}$ on the submanifold of $M$ which represents the Poincar\'e dual of ${\cal A}$.
Notice that here for ${\cal A} {\eta}$, it is crucial to have ${\cal A}$ as a cohomology class, so we can define its Poincar\'e dual of ${\cal A}$ as $\text{PD}({\cal A})$.
In other words, the ${\cal A} {\eta} \equiv {\eta} (\text{PD}({\cal A}))$.
%
Similarly we define the product notations between a cohomology class and other fermionic invariants also via the Poincar\'e dual PD of cohomology class.
Here are some other examples of fermionic invariants appeared in this work: \\
$\bullet$ The $\tilde{\eta}$ is a mod 2 index of 1d Dirac operator as a cobordism invariant of $\Omega_1^{{\rm Spin}}=\mathbb{Z}_2$.\\
$\bullet$ The ${\eta}'$ is a mod 4 index of 1d Dirac operator as a cobordism invariant of $\Omega_1^{{\rm Spin}\times \mathbb{Z}_4}=\mathbb{Z}_4$.\\
$\bullet$ The Arf invariant \cite{Arf1941} is a mod 2 cobordism invariant of $\Omega_2^{{\rm Spin}}=\mathbb{Z}_2$.
The Arf appears to be the low energy iTQFT of a 1+1d Kitaev fermionic chain \cite{Kitaev2001chain0010440},
whose boundary hosts a single 0+1d real Majorana zero mode on each of open ends.\\
$\bullet$ The Arf-Brown-Kervaire (ABK) invariant \cite{brown1972generalizations, KT1990} is a mod 8 cobordism invariant of $\Omega_2^{\mathrm{Pin}^-}=\mathbb{Z}_8$.
The ABK appears to be the low energy iTQFT
of a 1+1d Fidkowski-Kitaev fermionic chain \cite{FidkowskifSPT1, FidkowskifSPT2}
protected by time reversal and fermion parity symmetry $\mathbb{Z}_2^T \times \mathbb{Z}_2^F$,
whose boundary also hosts some number of 0+1d real Majorana zero modes.
\\
For cobordism or cohomology invariants, we may implicitly make a convention that cohomology classes are pulled back to the manifold $M$.
For example, ${\cal A}_{{\mathbb{Z}_2}} \in \operatorname{H}^1(M,\mathbb{Z}_2)$ is the first cohomology class on $M$ with a $\mathbb{Z}_2$ coefficient,
as a generator from ${\cal A}_{{\mathbb{Z}_2}} \in\operatorname{H}^1(\mathrm{B}(\mathbb{Z}_4/\mathbb{Z}_2^F),\mathbb{Z}_2)$ of ${{\rm Spin} \times_{\mathbb{Z}_2^F} \mathbb{Z}_4}$ pullback to $M$ as the former expression.}
\quad
\newpage
\begin{table}[!t]
$\hspace{-19.8mm}
\begin{tabular}{ccccccc c c c c c c c c}
\hline
$\begin{array}{c}
\textbf{SM}\\
\textbf{fermion}\\
\textbf{spinor}\\
\textbf{field}
\end{array}$
& ${{\rm SU}(3)}$& ${{\rm SU}(2)}$ & ${\rm U}(1)_Y$ & ${\rm U}(1)_{Y_W}$
& ${\rm U}(1)_{\tilde Y}$
& ${\rm U}(1)_{\rm{EM}}$
& ${\rm U}(1)_{{ \mathbf{B}- \mathbf{L}}}$ & ${\rm U}(1)_{X}$ &
$\mathbb{Z}_{4,X}$
& $\mathbb{Z}_{2}^F$ & SU(5) & Spin(10) \\
\hline\\[-4mm]
$\bar{d}_R$& $\bar{\mathbf{3}}$& $\mathbf{1}$ & 1/3 &2/3 & 2 & 1/3 & $-1/3$ & $-3$ &1 & 1 & \multirow{2}{*}{
$\overline{\bf{5}}$} & \multirow{6}{*}{
${\bf{16}}$} \\
\cline{1-11} $l_L$& $\mathbf{1}$& $\mathbf{2}$& $-1/2$& $-1$ & $-3$ & 0 or $-1$ & $-1$ & $-3$ &1 & 1 \\
\cline{1-12} $q_L$& ${{\mathbf{3}}}$& $\mathbf{2}$& 1/6 & 1/3 &1 & 2/3 or $-1/3$ & 1/3 &1 & 1 & 1 & \multirow{3}{*}{
${\bf{10}}$} \\
\cline{1-11} $\bar{u}_R$& $\bar{\mathbf{3}}$& $\mathbf{1}$& $-2/3$ & $-4/3$ & $-4$ & $-2/3$ & $-1/3$ & $1$ &1 & 1 \\
\cline{1-11} $\bar{e}_R= e_L^+$& $\mathbf{1}$& $\mathbf{1}$& 1 & 2 & 6 & 1 & 1 &1 & 1 & 1\\
\cline{1-12} $\bar{\nu}_R= {\nu}_L $& $\mathbf{1}$& $\mathbf{1}$& 0 & 0 & 0 & 0 & 1 & 5 & 1 & 1 & \;\;${\bf{1}}$ &\\
\hline\end{tabular}
$
\caption{We show the quantum numbers of $15 +1 = 16$ left-handed Weyl fermion spinors in each of three generations of matter fields in SM.
The $15$ of 16 Weyl fermion are $ \overline{\bf 5} \oplus {\bf 10}$ of SU(5); namely,
$(\overline{\bf 3},{\bf 1},1/3)_L \oplus ({\bf 1},{\bf 2},-1/2)_L \sim \overline{\bf 5}$
and $({\bf 3},{\bf 2}, 1/6)_L \oplus (\overline{\bf 3},{\bf 1}, -2/3)_L \oplus ({\bf 1},{\bf 1},1)_L \sim {\bf 10}$ of SU(5).
The $1$ of 16 is presented neither in the standard GSW SM nor in the SU(5) GUT, but it is within ${\bf 16}$ of the SO(10) GUT.
The numbers in the Table entries indicate the quantum numbers associated to the representation of the groups given in the top row.
We show the first generation of SM fermion matter fields in Table \ref{table:SMfermion}. There are 3 generations, triplicating Table \ref{table:SMfermion}, in SM.
All fermions have the fermion parity $\mathbb{Z}_{2}^F$ representation 1.}
\label{table:SMfermion}
\end{table}
\begin{comment}
$\hspace{-16mm}
\begin{tabular}{ccccccc c c c c c c}
\hline
$\begin{array}{c}
\textbf{SM}\\
\textbf{fermion}\\
\textbf{spinor}\\
\textbf{field}
\end{array}$
& ${{\rm SU}(3)}$& ${{\rm SU}(2)}$ & ${\rm U}(1)_Y$ & ${\rm U}(1)_{Y_W}$
& ${\rm U}(1)_{{ \mathbf{B}- \mathbf{L}}}$ & ${\rm U}(1)_{X}$ &
$\mathbb{Z}_{4,X}$ & ${\rm U}(1)_{\tilde Y}$ & ${\rm U}(1)_{{\tilde{\mathbf{B}}- \tilde{\mathbf{L}}}}$ & ${\rm U}(1)_{\tilde X}$ & $\mathbb{Z}_{4,\tilde X}$
& $\mathbb{Z}_{2}^F$ \\
\hline\\[-4mm]
$\bar{d}_R$& $\bar{\mathbf{3}}$& $\mathbf{1}$ & 1/3 &2/3 & $-1/3$ & $-3$ &1 & 2& $-1$ & $-9$ & $-1$ & 1 \\
\hline $l_L$& $\mathbf{1}$& $\mathbf{2}$& $-1/2$& $-1$ & $-1$ & $-3$ &1 &$-3$ & $-3$ & $-9$ & $-1$ & 1\\
\hline $q_L$& ${{\mathbf{3}}}$& $\mathbf{2}$& 1/6 & 1/3 & 1/3 &1 & 1&1 & $1$ & $3$ & $-1$ & 1\\
\hline $\bar{u}_R$& $\bar{\mathbf{3}}$& $\mathbf{1}$& $-2/3$ & $-4/3$ & $-1/3$ & $1$ &1 & $-4$& $-1$ & $3$ & $-1$& 1 \\
\hline $\bar{e}_R= e_L^+$& $\mathbf{1}$& $\mathbf{1}$& 1 & 2 & 1 &1 & 1 & 6& $3$ & $3$ & $-1$ & 1\\
\hline $\bar{\nu}_R= {\nu}_L $& $\mathbf{1}$& $\mathbf{1}$& 0 & 0 & 1 & 5 & 1 & 0&$3$ & 15 & $-1$ & 1\\
\hline\end{tabular}$
\end{comment}
\begin{table}[!t]
$ \hspace{14mm}
\begin{tabular}{c c c c c c c c c c c }
\hline
$\begin{array}{c}
\textbf{SM}\\
\textbf{boson}\\
\textbf{scalar or}\\
\textbf{vector }
\textbf{field}
\end{array}$
& ${{\rm SU}(3)}$& ${{\rm SU}(2)}$ & ${\rm U}(1)_Y$
& ${\rm U}(1)_{{ \mathbf{B}- \mathbf{L}}}$ & ${\rm U}(1)_{X}$ &
$\mathbb{Z}_{4,X}$ \\
\hline\\[-4mm]
\hline
$\begin{array}{c}
\textbf{SM Higgs scalar}\\
\phi_H
\end{array}$
& $\mathbf{1}$ & $\mathbf{2}$& 1/2 & 0 & $-2$ & 2 \\
\hline
\hline
\multicolumn{7}{l}{\bf{SM gauge vector bosons} ($1+3+8 =12$ generators)}
\\
\hline
$\begin{array}{c}
\textbf{photon} \;(\gamma)\\
A_Y \text{ or } A
\end{array}$
& ${}$ & \multicolumn{2}{c}{\multirow{3}{*}{
${3\;\;\;\;+\;\;\;1\;}$}} & 0 & 0 & 0 \\
\cline{1-2} \cline{5-7}
$\begin{array}{c}
\textbf{Weak}\\
W^{1,2,3} \text{ or } W^{\pm}, Z^{0}
\end{array}$
& & & & 0 & 0 & 0 \\
\hline
$\begin{array}{c}
\textbf{Strong \rm{(g)}}\\
A^\mathrm{a} \tau^\mathrm{a}
\end{array}$
& ${8}$ & & & 0 & 0 & 0 \\
\hline
\hline
\multicolumn{7}{l}{\bf{BSM Georgi-Glashow gauge vector bosons ($2 \times (3 \times 2 )=12$ generators)}}
\\
\hline
\textbf{X}$^+$ ($+\frac{4}{3}|e|$)
& \multirow{2}{*}{
$\bar{\mathbf{3}}$} & \multirow{2}{*}{
$\mathbf{2}$} & \multirow{2}{*}{
$5/6$} & \multirow{2}{*}{
$2/3$} & 0 & 0
\\
\cline{1-1}
\textbf{Y}$^+$ ($+\frac{1}{3}|e|$)
& & & & & 0 & 0 \\
\cline{1-1}
\textbf{Y}$^-$ ($-\frac{1}{3}|e|$)
& \multirow{2}{*}{
${\mathbf{3}}$} & \multirow{2}{*}{
$\mathbf{2}$} & \multirow{2}{*}{
$-5/6$} & \multirow{2}{*}{
$-2/3$} & 0 & 0 \\
\cline{1-1}
\textbf{X}$^-$ ($-\frac{4}{3}|e|$)
& & & & & 0 & 0 \\
\hline
\end{tabular}
$
\caption{We show quantum numbers of the electroweak Higgs boson $\phi_H$, and
24 gauge bosons corresponding to 24 Lie algebra $su(5)$ generators of Lie group SU(5).
Hereby the representation numbers of gauge fields, we really mean the
gauge fields \emph{transform in the adjoint} of the gauge group.
(But the gauge fields are \emph{not} in the adjoint or other $\mathbf{R}$ representations of the gauge group.)
Also the readers should not be confused with the symmetry charge $X$ (in the Italic form) and its gauge boson $X_g$,
with the Georgi-Glashow (GG) model gauge boson \textbf{X} (in the boldface text form).
The \textbf{X} and \textbf{Y} gauge bosons can carry certain SU(3), SU(2) and U(1) quantum numbers,
because the \textbf{X} and \textbf{Y} sit outside the SM gauge group.
All bosons have the fermion parity $\mathbb{Z}_{2}^F$ representation 0.
See also the caption in Table \ref{table:SMfermion}.}
\label{table:SMboson}
\end{table}
\newpage
\section{Overview on Standard Models, Grand Unifications, and Anomalies}
\label{sec:2Overview}
\subsection{SM and GUT: local Lie algebras to global Lie groups, and representation theory }
\label{sec:2OverviewSMGUT}
We shall first overview the local Lie algebra the representation theory of matter field contents,
and the global Lie groups of of SM and GUT.
Then we will be able to be precise about
the spacetime symmetry group $G_{\text{spacetime}}$ and internal symmetry group $G_{\text{internal}}$ relevant for SM and GUT physics,
\begin{enumerate}[leftmargin=-4.mm, label=\textcolor{blue}{[\Roman*]}., ref={[\Roman*]}]
\item The local gauge structure of Standard Model is the Lie algebra $u(1) \times su(2) \times su(3)$.
This means that the Lie algebra valued 1-form gauge fields take values in the Lie algebra generators of $u(1) \times su(2) \times su(3)$.
There are $1 + 3 + 8 = 12$ Lie algebra generators.
The 1-form gauge fields are the 1-connections of the principals $G_{\text{internal}}$-bundles that we should determine.
\item Fermions are the spinor fields,
as the \emph{sections of the spinor bundles} of the spacetime manifold. For the left-handed Weyl spinor $\Psi_L$,
it is a doublet spin-1/2 representation of spacetime symmetry group $G_{\text{spacetime}}$ (Minkowski ${\rm Spin}(3,1)$ or Euclidean ${\rm Spin}(4)$), denoted as
\begin{eqnarray}
\Psi_L \sim {\bf 2}_L \text{ of } {\rm Spin}(3,1) = \rm SL(2,{\mathbb{C}}), \quad\text{ or }\quad \Psi_L \sim {\bf 2}_L \text{ of } {\rm Spin}(4) = {\rm SU}(2)_L \times {\rm SU}(2)_R.
\end{eqnarray}
These Spin
groups are the double-cover also the universal-cover version of the Lorentz group ${\rm SO}(3,1)_+$ or Euclidean rotation ${\rm SO}(4)$, extended by the fermion parity $\mathbb{Z}_2^F$.
In the first generation of SM, the matter fields as Weyl spinors $\Psi_L$ contain:
\begin{itemize}
\item
The left-handed up and down quarks ($u$ and $d$) form a doublet $\begin{pmatrix}
u\\
d
\end{pmatrix}_L$ in ${\bf 2}$ for the SU(2)$_{\text{weak}}$, and they are in ${\bf 3}$ for the SU(3)$_{\text{strong}}$.
\item
The right-handed up and down quarks, each forms a singlet, $u_R$ and $d_R$, in ${\bf 1}$ for the SU(2)$_{\text{weak}}$. They are in ${\bf 3}$ for the SU(3)$_{\text{strong}}$.
\item
The left-handed electron and neutrino form a doublet $\begin{pmatrix}
\nu_e\\
e
\end{pmatrix}_L$ in ${\bf 2}$ for the SU(2)$_{\text{weak}}$, and they are in ${\bf 1}$ for the SU(3)$_{\text{strong}}$.
\item
The right-handed electron forms a singlet $e_R$ in ${\bf 1}$ for the SU(2)$_{\text{weak}}$,
and it is in ${\bf 1}$ for the SU(3)$_{\text{strong}}$.
\end{itemize}
There are two more generations of quarks:
charm and strange quarks ($c$ and $s$),
and top and bottom quarks ($t$ and $b$).
There are also two more generations of leptons:
muon and its neutrino ($\mu$ and $\nu_\mu$),
and
tauon and its neutrino ($\tau$ and $\nu_\tau$).
So there are three generations (i.e., families) of quarks and leptons:
\begin{eqnarray}
\Bigg(
\begin{pmatrix}
u\\
d
\end{pmatrix}_L \times {\bf 3}_{\text{color}}, \quad\quad \quad u_R \times {\bf 3}_{\text{color}}, \quad\quad\quad\quad d_R \times {\bf 3}_{\text{color}},
\quad\quad\quad
\begin{pmatrix}
\nu_e\\
e
\end{pmatrix}_L,\quad\quad \quad e_R
\quad \Bigg),\\
\Bigg(
\begin{pmatrix}
c\\
s
\end{pmatrix}_L \times {\bf 3}_{\text{color}}, \quad\quad \quad c_R \times {\bf 3}_{\text{color}}, \quad\quad\quad\quad s_R \times {\bf 3}_{\text{color}},
\quad\quad\quad
\begin{pmatrix}
\nu_\mu\\
\mu
\end{pmatrix}_L,\quad\quad \quad \mu_R
\quad \Bigg),\\
\Bigg(
\begin{pmatrix}
t\\
b
\end{pmatrix}_L \times {\bf 3}_{\text{color}}, \quad\quad \quad t_R \times {\bf 3}_{\text{color}}, \quad\quad\quad\quad b_R \times {\bf 3}_{\text{color}},
\quad\quad\quad
\begin{pmatrix}
\nu_\tau\\
\tau
\end{pmatrix}_L,\quad\quad \quad \tau_R
\quad \Bigg).
\end{eqnarray}
In short, for all of them as three generations, we can denote them as:
\begin{eqnarray}
\Bigg(
\begin{pmatrix}
u\\
d
\end{pmatrix}_L \times {\bf 3}_{\text{color}}, \quad\quad \quad u_R \times {\bf 3}_{\text{color}}, \quad\quad\quad\quad d_R \times {\bf 3}_{\text{color}},
\quad\quad\quad
\begin{pmatrix}
\nu_e\\
e
\end{pmatrix}_L,\quad\quad \quad e_R
\quad \Bigg)\times \text{3 generations}. \quad
\end{eqnarray}
In fact, all the following \emph{four} kinds of
\begin{eqnarray}
G_{\text{internal}}=\frac{{\rm SU}(3)\times {\rm SU}(2)\times {\rm U}(1)}{\mathbb{Z}_q}
\end{eqnarray}
with $q=1,2,3,6$ are compatible with the above representations of
fermion fields.\footnote{See an excellent exposition on
$G_{\text{internal}}=\frac{{\rm SU}(3)\times {\rm SU}(2)\times {\rm U}(1)}{\mathbb{Z}_q}$
with $q=1,2,3,6$
in a recent work by Tong \cite{Tong2017oea1705.01853}.}
These $15 \times 3$ Weyl spinors can be written in the following more succinct forms of representations for any of the internal symmetry group $G_{\text{internal}}$ with $q=1,2,3,6$:
\begin{eqnarray} \label{eq:rep-3generations}
\Bigg( ({\bf 3},{\bf 2}, 1/6)_L,({\bf 3},{\bf 1}, 2/3)_R,({\bf 3},{\bf 1},-1/3)_R,({\bf 1},{\bf 2},-1/2)_L,({\bf 1},{\bf 1},-1)_R \Bigg)\times \text{3 generations} \nonumber\\
\Rightarrow
\Bigg( ({\bf 3},{\bf 2}, 1/6)_L,(\overline{\bf 3},{\bf 1}, -2/3)_L,(\overline{\bf 3},{\bf 1},1/3)_L, ({\bf 1},{\bf 2},-1/2)_L,({\bf 1},{\bf 1},1)_L \Bigg)\times \text{3 generations}.
\end{eqnarray}
The triplet given above is listed by their representations:
\begin{eqnarray} \label{eq:rep}
\text{(SU(3) representation, SU(2) representation, hypercharge $Y$)}.
\end{eqnarray}
For example,
$({\bf 3},{\bf 2}, 1/6)$ means that ${\bf 3}$ in SU(3), ${\bf 2}$ in SU(2) and 1/6 for hypercharge.
In the second line of \Eq{eq:rep-3generations}, we transforms the right-handed Weyl spinor
$\Psi_R \sim {\bf 2}_R \text{ of } {\rm Spin}(3,1)$
to its left-handed $\Psi_L \sim {\bf 2}_L \text{ of } {\rm Spin}(3,1)$, while we flip their representation \Eq{eq:rep}
to its complex conjugation representation.\footnote{Note that ${\bf 2}$ and $\overline{\bf 2}$ are the same representation in SU(2), see, e.g., in the context of Yang-Mills gauge theories with discrete symmetries and cobordism \cite{1711.11587GPW}.}
If we include the right-handed neutrinos (say ${\nu_{e}}_R$, ${\nu_{\mu}}_R$, and ${\nu_{\tau}}_R$), they are all in the representation
\begin{eqnarray} \label{eq:nuR-rep}
({\bf 1},{\bf 1},0)_R
\end{eqnarray} with no hypercharge.
We can also represent a right-handed neutrino by the left-handed (complex) conjugation version
\begin{eqnarray} \label{eq:nuL-rep}
({\bf 1},{\bf 1},0)_L.
\end{eqnarray}
Also the complex scalar Higgs field $\phi_H$ is in a representation
\begin{eqnarray}
({\bf 1},{\bf 2}, 1/2).
\end{eqnarray}
In the Higgs condensed phase of SM, the conventional Higgs vacuum expectation value (vev) is chosen to be $\langle \phi_H \rangle =\frac{1}{\sqrt 2}
\begin{pmatrix}
0\\
v
\end{pmatrix}$, which vev has a $Q_{\rm{EM}}=0$.
{Note that
our hypercharge $Y$ is given conventionally by the relation: $Q_{\rm{EM}}=T_3 +Y$ where
$Q_{\rm{EM}}$ is the unbroken (not Higgsed) electromagnetic gauge charge
and $T_3= \frac{1}{2}
\begin{pmatrix}
1 & 0\\
0 & -1
\end{pmatrix}$ is a generator of SU(2)$_{\text{weak}}$. However, some other conventions are common, we list down three conventions
\begin{eqnarray} \label{eq:QEM}
Q_{\rm{EM}}=T_3 +Y=T_3 +\frac{1}{2} Y_W=T_3 +\frac{1}{6} \tilde{Y}.
\end{eqnarray}
In the $Q_{\rm{EM}}=T_3 +\frac{1}{6} \tilde{Y}$ version, we have the integer quantized $\tilde{Y} = 6 Y$. We can rewrite \eq{eq:rep-3generations} as:
\begin{eqnarray} \label{eq:rep-3generations-Sec}
&&\hspace{-18mm}
\Bigg( ({\bf 3},{\bf 2}, Y= 1/6)_L,({\bf 3},{\bf 1}, Y= 2/3)_R,({\bf 3},{\bf 1}, Y= -1/3)_R,({\bf 1},{\bf 2},Y=-1/2)_L,({\bf 1},{\bf 1},Y= -1)_R \Bigg)\times \text{3 generations} \nonumber\\
=
&&\hspace{-6mm}
\Bigg( ({\bf 3},{\bf 2}, Y= 1/6)_L,(\overline{\bf 3},{\bf 1}, Y= -2/3)_L,(\overline{\bf 3},{\bf 1},Y= 1/3)_L, ({\bf 1},{\bf 2},Y= -1/2)_L,({\bf 1},{\bf 1},Y= 1)_L \Bigg)\times
\text{3 generations}\nonumber\\
=
&&\hspace{-6mm}
\Bigg( ({\bf 3},{\bf 2}, Y_W=1/3)_L,({\bf 3},{\bf 1},Y_W=4/3)_R,({\bf 3},{\bf 1},Y_W=-2/3)_R,({\bf 1},{\bf 2},Y_W=-1)_L,({\bf 1},{\bf 1},Y_W=-2)_R \Bigg)
\times \text{3 generations} \nonumber\\
=
&&\hspace{-6mm}
\Bigg( ({\bf 3},{\bf 2}, Y_W= 1/3)_L,(\overline{\bf 3},{\bf 1}, Y_W= -4/3)_L,(\overline{\bf 3},{\bf 1},Y_W= 2/3)_L, ({\bf 1},{\bf 2},Y_W= -1)_L,({\bf 1},{\bf 1},Y_W= 2)_L \Bigg)\times
\text{3 generations}\nonumber\\
\label{eq:rep-3generations-tildeY-Sec}
=
&&\hspace{-6mm}
\Bigg( ({\bf 3},{\bf 2}, \tilde{Y}=1)_L,({\bf 3},{\bf 1}, \tilde{Y}=4)_R,({\bf 3},{\bf 1},\tilde{Y}=-2)_R,({\bf 1},{\bf 2},\tilde{Y}=-3)_L,({\bf 1},{\bf 1},\tilde{Y}=-6)_R \Bigg)\times \text{3 generations} \nonumber\\
=
&&\hspace{-6mm}
\Bigg( ({\bf 3},{\bf 2}, \tilde{Y}=1)_L,(\overline{\bf 3},{\bf 1}, \tilde{Y}=-4)_L,(\overline{\bf 3},{\bf 1},\tilde{Y}=2)_L, ({\bf 1},{\bf 2},\tilde{Y}=-3)_L,({\bf 1},{\bf 1},\tilde{Y}=6)_L \Bigg)\times \text{3 generations}.
\label{eq:Weyl-rep}
\end{eqnarray}
The right hand neutrino $\nu_R$ is in a representation
\begin{eqnarray}
({\bf 1},{\bf 1}, Y=0)
=({\bf 1},{\bf 1}, Y_W=0)
=({\bf 1},{\bf 1}, \tilde{Y} =0).
\end{eqnarray}
times some generation number.
Also the complex scalar Higgs field $\phi_H$ is in a representation
\begin{eqnarray}
({\bf 1},{\bf 2}, Y=1/2)
=({\bf 1},{\bf 2}, Y_W=1)
=({\bf 1},{\bf 2}, \tilde{Y} =3),
\end{eqnarray}
where $\tilde{Y} =3 Y_W= 6 Y$.
}
We organize the above data in Table \ref{table:SMfermion}.
\item If we include the $3 \times 2 + 3 + 3 + 2 +1 =15$ left-handed Weyl spinors from one single generation,
we can combine them as a multiplet of $\overline{\bf 5}$ and {\bf 10} left-handed Weyl spinors of SU(5):
\begin{eqnarray}
(\overline{\bf 3},{\bf 1},1/3)_L \oplus ({\bf 1},{\bf 2},-1/2)_L &\sim& \overline{\bf 5} \text{ of } {\rm SU}(5),\\
({\bf 3},{\bf 2}, 1/6)_L \oplus (\overline{\bf 3},{\bf 1}, -2/3)_L \oplus ({\bf 1},{\bf 1},1)_L &\sim& {\bf 10} \text{ of } {\rm SU}(5).
\end{eqnarray}
Hence these are matter field representations of the SU(5) GUT with a SU(5) gauge group.
Other than the electroweak Higgs $\phi_H$ (which $\phi_H$ can be part of a scalar Higgs in the fundamental
${\bf 5}$ of SU(5) GUT), we may also introduce a different GUT Higgs field $\phi_{GG}$
to break down SU(5) to $\frac{{\rm SU}(3) \times {\rm SU}(2) \times {\rm U}(1)}{\mathbb{Z}_6}$.
The $\phi_{GG}$ is in the adjoint representation of SU(5) as
\begin{eqnarray}
{\bf 24} =({\bf 8},{\bf 1}, {Y}=0) \oplus ({\bf 1},{\bf 3}, {Y}=0) \oplus ({\bf 1},{\bf 1}, {Y}=0) \oplus({\bf 3},{\bf 2}, {Y}=-\frac{5}{6})
\oplus(\bar{\bf 3},{\bf 2}, {Y}=\frac{5}{6}).
\end{eqnarray}
\item If we include an extra right-handed neutrino plus
the $3 \times 2 + 3 + 3 + 2 +1 =15$ left-handed Weyl spinors from one single generation,
we can combine them as a multiplet of 16 left-handed Weyl spinors:
\begin{eqnarray}
\quad \Psi_L \sim {\bf 16}^+ \text{ of } {\rm Spin}(10),
\end{eqnarray}
which sits at the 16-dimensional irreducible spinor representation of Spin(10).
(In fact, ${\bf 16}^+$ and ${\bf 16}^-$-dimensional irreducible spinor representations
together form a ${\bf 32}$-dimensional reducible spinor representation of Spin(10).)
Namely, instead of an SO(10) gauge group,
we should study the SO(10) GUT with a Spin(10) gauge group.
\item {\bf Lie algebra generators and gauge bosons}:
We can count the number of Lie algebra generators to represent the local 1-form gauge field of gauge bosons.
For example, there are $1+3+8 =12$ independent Lie algebra generators thus gauge bosons for the Lie algebra $u(1) \times su(2) \times su(3)$.
There are $24$ independent Lie algebra generators (thus gauge bosons) for the Lie algebra $su(5)$,
and $45$ Lie algebra generators (thus gauge bosons) for the Lie algebra $so(10)$.
In this work, we focus on SM and SU(5) GUT. We list down their quantum numbers in Table \ref{table:SMboson}.
\item {\bf ${\rm U}(1)_{{ \mathbf{B}- \mathbf{L}}}$ symmetry and
${\rm U}(1)_{X}$ symmetry}: \label{RemarkU1}
For SM with ${{G}_{\text{internal}} }=G_{\text{SM}_q}$ of $q=1,2,3,6$,
we have a ${ \mathbf{B}- \mathbf{L}}$ (baryon minus lepton numbers) ${\rm U}(1)_{{ \mathbf{B}- \mathbf{L}}}$ global symmetry.
For SU(5) GUT with ${{G}_{\text{internal}} }={\rm SU}(5)$, we can have a ${\rm U}(1)_{X}$ symmetry \cite{Wilczek1979hcZee},
where the $X$ charge is related to ${ \mathbf{B}- \mathbf{L}}$ via
\begin{eqnarray}
X &\equiv&
5({ \mathbf{B}- \mathbf{L}})-4Y=
5({ \mathbf{B}- \mathbf{L}})-2Y_W
,\\
\tilde X &\equiv& 3 X=
5 \cdot 3 ({ \mathbf{B}- \mathbf{L}})-2 \cdot 3 Y_W
=
5 \cdot 3 ({ \mathbf{B}- \mathbf{L}})-2 \tilde Y =
5 ({ \tilde{\mathbf{B}}- \tilde{\mathbf{L}}})-2 \tilde Y,\\
\tilde Y &=& 3 Y_W = 6 Y.
\end{eqnarray}
Here we have used \Eq{eq:QEM}.
\item {{\bf $\mathbb{Z}_{4, X}$ symmetry, $\mathbb{Z}_2^F$ fermion parity, and discrete symmetries}: \label{RemarkZ4X}\\
We will learn that it is natural to consider a $\mathbb{Z}_{4, X}$ subgroup of ${\rm U}(1)_{X}$, when we attempt to embed the SU(5) GUT to SO(10) GUT. In fact, the
$\mathbb{Z}_{4, X}$ can be regarded as the $\mathbb{Z}_4$ center of the Spin(10) gauge group as $\text{Z}({\rm Spin}(10))=\mathbb{Z}_4
= \mathbb{Z}_{4,X} \supset \mathbb{Z}_{2}^F$ \cite{GarciaEtxebarriaMontero2018ajm1808.00009}.
Since Spin(10) is fully dynamically gauged in SO(10) GUT, the $\mathbb{Z}_{4, X}$ is also dynamically gauged.
In summary, we have
\begin{eqnarray}
&&\text{$\bullet$ ${\rm U}(1)_{{ \mathbf{B}- \mathbf{L}}}$ can be a global symmetry (maybe anomalous) in the SMs for all $q=1,2,3,6$}.\cr
&&\text{$\bullet$ ${\rm U}(1)_{X}$ can be a global symmetry (maybe anomalous) in the SU(5) GUT}.\cr
&&\text{$\bullet$ $\mathbb{Z}_{4, X}$ can be chosen to be \emph{either} a global \emph{or} a gauge symmetry for the SM and SU(5) GUT.}\cr
&&\text{\quad \quad\quad But $\mathbb{Z}_{4, X}$ is a dynamical gauge symmetry in the SO(10) GUT.}\cr
&&\text{$\bullet$ $\mathbb{Z}_{2}^F$
fermion parity has a generator $(-1)^F$ with the fermion number $F$, which
can be a global symmetry} \cr
&&\text{\quad \quad\quad in the SMs (for all $q=1,2,3,6$) and in the SU(5) GUT.} \cr
&&\text{\quad \quad\quad But the $\mathbb{Z}_{2}^F$ is naturally chosen to be a dynamical gauge symmetry in the SO(10) GUT.
}\nonumber
\end{eqnarray}
For the SO(10) GUT, we may call $\mathbb{Z}_{4, X}$ a gauge symmetry colloquially, but we should remind readers that a
gauge symmetry is not really a (global) symmetry, but only a gauge redundancy.
Note that:
\begin{eqnarray}
&&\text{$\bullet$ All fermions from leptons and quarks have a $\mathbb{Z}_{4,X}$ charge 1, see \Table{table:SMfermion}.}\cr
&&\text{$\bullet$ All gauge bosons have a $\mathbb{Z}_{4,X}$ charge 0, but the Higgs have a $\mathbb{Z}_{4,X}$ charge 2, see \Table{table:SMboson}.}\cr
&&\text{$\bullet$ The $\mathbb{Z}_{2}^F$ gives $(-1)$ sign to all fermions (additive notation: charge 1 in $\mathbb{Z}_{2}^F$ in \Table{table:SMfermion}).}\cr
&&\text{$\bullet$ The $\mathbb{Z}_{2}^F$ gives $(+1)$ sign to all bosons (additive notation: charge 0 in $\mathbb{Z}_{2}^F$ in \Table{table:SMboson}).
} \nonumber
\end{eqnarray}
The above facts about $\mathbb{Z}_{4, X}$ charges for bosons or fermions (in \Table{table:SMfermion} and \Table{table:SMboson})
can be understood from
a group extension (as a short exact sequence):
$0 \to \mathbb{Z}_{2}^F \to \mathbb{Z}_{4,X} \to (\mathbb{Z}_{2}^{\cal A} \equiv \frac{\mathbb{Z}_{4,X}}{\mathbb{Z}_{2}^F}) \to 0$.
The fermion parity $\mathbb{Z}_{2}^F$ is a normal subgroup of $\mathbb{Z}_{4,X}$, while the $\mathbb{Z}_{2}^{\cal A}$ is only a quotient group.
So the fermions, with an odd $\mathbb{Z}_{2}^F$ charge, must carry also an odd charge of $\mathbb{Z}_{4,X}$.
The bosons, with an even (i.e., no) $\mathbb{Z}_{2}^F$ charge, must carry also an even charge of $\mathbb{Z}_{4,X}$.\footnote{{Given
the $\mathbb{Z}_{4,X}$ charge state
$| X\rangle$ with $X=0,1,2,3$
as a representation $z^X$.
Note that $z \in {\rm U}(1)$ with $|z|=1$, where we embed the normal subgroup $\mathbb{Z}_2^F \subset \mathbb{Z}_{4,X} \subset {\rm U}(1)$.\\
$\bullet$ The $\mathbb{Z}_{4,X}$ symmetry generator ${\rm U}_{\mathbb{Z}_{4,X}}$ acts on $| X\rangle$,
which becomes ${\rm U}_{\mathbb{Z}_{4,X}} | X\rangle = \hspace{1pt}\mathrm{i}\hspace{1pt}^X | X\rangle$ with $z=\hspace{1pt}\mathrm{i}\hspace{1pt}$. \\
$\bullet$ The subgroup $\mathbb{Z}_{2}^F$ symmetry generator
${\rm U}_{\mathbb{Z}_{2}^F}= ({\rm U}_{\mathbb{Z}_{4,X}})^2$
can also act on $| X\rangle$,
which becomes ${\rm U}_{\mathbb{Z}_{2}^F} | X\rangle = ({\rm U}_{\mathbb{Z}_{4,X}})^2 | X\rangle = \hspace{1pt}\mathrm{i}\hspace{1pt}^{2X} | X\rangle = (-1)^X | X\rangle $.
Thus, we read the fermion parity $(-1)^F$,
the $| 1\rangle$ and $| 3\rangle$ has $-1$ (thus odd 1 in $\mathbb{Z}_{2}^F$), while
the $| 0\rangle$ and $| 2\rangle$ has $+1$ (thus even 0 in $\mathbb{Z}_{2}^F$).
This is also consistent with the fact that the group extension
$0 \to \mathbb{Z}_{2}^F \to \mathbb{Z}_{4,X} \to \mathbb{Z}_{2}^{\cal A} \equiv \frac{\mathbb{Z}_{4,X}}{\mathbb{Z}_{2}^F} \to 0$
requires a nontrivial 2-cocycle $\operatorname{H}^2(\mathrm{B} \mathbb{Z}_{2}^{\cal A}, \mathbb{Z}_{2}^F) = \mathbb{Z}_2$ to specify the extension to $\mathbb{Z}_{4,X}$.}
}
}
These extra symmetries are well-motivated in the earlier pioneer works
\cite{Ibanez1991hvRossPLB, Banks1991xjDine9109045, Csaki1997awMurayama9710105, Dreiner2005rd0512163} and References therein
the recent work \cite{GarciaEtxebarriaMontero2018ajm1808.00009, Hsieh2018ifc1808.02881, GuoJW1812.11959}.
Extra discrete symmetries can be powerful giving rise to new anomaly cancellation constraints.
Part of the new ingredients we will survey are the new global anomalies for discrete symmetries of SMs, in \Sec{Sec:discretesymmetries}.
Let us discuss how these extra groups can be embedded into the total groups in Remarks
\ref{RemarkEmbed1},
\ref{RemarkEmbed2},
\ref{RemarkEmbed3},
and \ref{RemarkEmbed4}.
\item
\label{RemarkEmbed1}
We find the Lie group embedding for the internal symmetry of GUTs and Standard Models
\cite{WW2019fxh1910.14668, WangWen2018cai1809.11171}:
\begin{eqnarray}
{\rm SO}(10)
\supset
{\rm SU}(5)
\supset
\frac{{\rm U}(1) \times {\rm SU}(2)\times{\rm SU}(3)}{\mathbb{Z}_6}.\\
{\rm Spin}(10)
\supset
{\rm SU}(5)
\supset
\frac{{\rm U}(1) \times {\rm SU}(2)\times{\rm SU}(3)}{\mathbb{Z}_6}.
\end{eqnarray}
Only $q=6$, but not other $q=1,2,3$, for $G_{\text{SM}_q}$ can be embedded into Spin(10) nor SU(5).
So from the GUT perspective, it is natural to consider the Standard Model gauge group
$\frac{{\rm U}(1) \times {\rm SU}(2)\times{\rm SU}(3)}{\mathbb{Z}_6}$.\footnote{In fact, the author
believe and declare that $G_{\text{SM}_6} \equiv \frac{{\rm U}(1) \times {\rm SU}(2)\times{\rm SU}(3)}{\mathbb{Z}_6}$
is the correct and natural gauge group of SM.}
\item
\label{RemarkEmbed2}
We also find the following group embedding for the spacetime and internal symmetries of GUTs and Standard Models
(\Refe{WW2019fxh1910.14668, WangWen2018cai1809.11171}, see also \cite{WanWangv2} for the derivations):
\begin{eqnarray} \label{eq:SMembed1}
{\frac{{\rm Spin}(d) \times
{\rm Spin}(10)}{{\mathbb{Z}_2^F}}
\supset
{\rm Spin}(d) \times {\rm SU}(5)
\supset
{\rm Spin}(d) \times \frac{{\rm SU}(3) \times {\rm SU}(2) \times {\rm U}(1)}{\mathbb{Z}_6} }.\\
\label{eq:SMembed2}
{{{\rm Spin}(d) \times
{\rm Spin}(10)}
\supset
{\rm Spin}(d) \times {\rm SU}(5)
\supset
{\rm Spin}(d) \times \frac{{\rm SU}(3) \times {\rm SU}(2) \times {\rm U}(1)}{\mathbb{Z}_6}}.
\end{eqnarray}
\item
\label{RemarkEmbed3}
For an extra ${\rm U}(1)_{{ \mathbf{B}- \mathbf{L}}}$ or ${\rm U}(1)_{X}$ symmetry, we need to consider Spin${}^c \equiv {\rm Spin}(d) \times_{\mathbb{Z}_2^F} {\rm U}(1)$ structure.
We find the embedding:
{
\begin{eqnarray}
{\rm Spin}^c(d)\times {\rm SU}(5)
\supset
{\rm Spin}^c(d) \times \frac{{\rm SU}(3) \times {\rm SU}(2) \times {\rm U}(1)}{\mathbb{Z}_6}.
\end{eqnarray}
}
\item \label{RemarkEmbed4}
For an extra $\mathbb{Z}_{4,X}$ symmetry,
we need to consider ${\rm Spin}(d) \times_{\mathbb{Z}_2} \mathbb{Z}_4 ={\rm Spin}(d) \times_{\mathbb{Z}_2^F} \mathbb{Z}_{4,X} $ structure.
In order to contain these groups embedded in $\frac{{\rm Spin}(d) \times
{\rm Spin}(10)}{{\mathbb{Z}_2^F}}$, it is more naturally to consider:
\begin{eqnarray}
\label{eq:SMembed3}
{\frac{{\rm Spin}(d) \times
{\rm Spin}(10)}{{\mathbb{Z}_2^F}}
\supset
{\rm Spin}(d) \times_{\mathbb{Z}_2^F} \mathbb{Z}_{4,X} \times {\rm SU}(5)
\supset
{\rm Spin}(d) \times_{\mathbb{Z}_2^F} \mathbb{Z}_{4,X} \times \frac{{\rm SU}(3) \times {\rm SU}(2) \times {\rm U}(1)}{\mathbb{Z}_6} }.
\end{eqnarray}
\end{enumerate}
\Refe{WW2019fxh1910.14668, WangWen2018cai1809.11171, GarciaEtxebarriaMontero2018ajm1808.00009, 2019arXiv191011277D}
study the cobordism theory of some of these SM, BSM, and GUT groups.
We shall particular pay attention to $d=5$ for
$\Omega^{d=5}_{G} \equiv
\Omega^{d=d}_{({\frac{{G_{\text{spacetime} }} \ltimes {{G}_{\text{internal}} }}{{N_{\text{shared}}}}})}
\equiv
\mathrm{TP}_{d=5}(G),$ since the 5d cobordism invariants of $\mathrm{TP}_{d=5}(G)$ can classify the 4d invertible anomalies of the total group $G$.
We particularly apply the results in \cite{WW2019fxh1910.14668}, for anomaly constraints and anomaly matchings of BSM physics,
in the following \Sec{sec:DynamicalGaugeAnomalyCancellation}.
\subsection{Classifications of anomalies and their different uses}
\label{sec:2OverviewSMGUTdifferentanomalies}
\begin{figure}[h!]
\centering
({\it 1}) \includegraphics[width=3.in]{anomaly-gggauge-.pdf}
({\it 2}) \includegraphics[width=3.in]{anomaly-G0G0G0-.pdf} \\[2mm]
({\it 3}) \includegraphics[width=3.in]{anomaly-G0gg-.pdf}
({\it 4}) \includegraphics[width=3.in]{anomaly-gG0G0-.pdf}
\caption{Examples of perturbative local anomalies of $\mathbb{Z}$ classes in 4d that can be captured by the free part of cobordism group
$\Omega^{d=5}_{G}
\equiv
\mathrm{TP}_{d=5}(G)$ in \Eq{eq:TPG}, which is in fact descended from the free part of bordism group $\Omega_{d=6}^{G}$.
({\it 1}) Dynamical gauge anomaly.
({\it 2}) 't Hooft anomaly of background (Backgrd.) fields.
({\it 3}) Adler-Bell-Jackiw (ABJ \cite{Adler1969gkABJ, Bell1969tsABJ}) type of anomalies.
({\it 4}) Anomaly that involves two background fields of global symmetries and one dynamical gauge field.
}
\label{fig:triangle}
\end{figure}
To start, let us mention different concepts of anomalies and their different uses. See for example \Fig{fig:triangle} for
perturbative local anomalies captured by Feynman-Dyson graphs.
Generally, we also have nonperturbative global anomalies not captured by Feynman graphs, but their uses are similar as follows:
\begin{enumerate}[leftmargin=-4.mm, label=\textcolor{blue}{({\it\arabic*})}., ref={({\it\arabic*})}]
\item \label{remark:dynamicalanomaly}
Dynamical gauge anomaly of ${{G}_{\text{internal}} }$ (e.g., \Fig{fig:triangle} ({\it 1})):
Anomaly matching cancellation must be zero for a dynamical gauge theory of its group ${{G}_{\text{internal}} }$.
\item \label{remark:tHooftanomaly}
't Hooft anomaly (e.g., \Fig{fig:triangle} ({\it 2})):
Anomaly matching of background symmetry of $G$ including ${{G}_{\text{internal}}}$ for their background fields, surprisingly, need \emph{not} to be
zero for a QFT:\\
\begin{enumerate}[leftmargin=6.mm, label=\textcolor{blue}{($\bullet$ \roman*)}., ref={($\bullet$ \roman*)}]
\item
If the 't Hooft anomaly of a QFT of any $G$ background fields turn out to be exactly zero, this means that these global symmetries of QFT can be realized local and \emph{onsite} (or on an $n$-simplex for a higher generalized global $n$-symmetry \cite{Gaiotto2014kfa1412.5148}) on a lattice.
This also means that we do not need to regularize the QFT living on the boundary of Symmetry-Protected/Enriched Topological states
\cite{Chen2011pg1106.4772, Wen2013oza1303.1803,Wen2013ppa1305.1045, Wang2017locWWW1705.06728, WangWen2018cai1809.11171}.
In fact, the QFT may be well-defined as the low energy theory of a quantum local lattice model in its own dimensions, as a high-energy ultraviolet (UV) completion ---
if it is free from all anomalies, via a cobordism argument \cite{WangWen2018cai1809.11171}.\\
\item
If the 't Hooft anomaly of a QFT of any $G$ background fields turn out to be not zero nor cancelled:
Then this can be regarded as several meanings and implications on QFT dynamics:\\
$\bullet$ {\bf First}, the symmetries of QFT in fact are \emph{non-local} or \emph{non-onsite} (or \emph{non-on-$n$-simplex} for a higher generalized global $n$-symmetry) on a lattice. This means the obstruction of gauging such \emph{non-local} symmetries, thus this
obstruction is equivalent with the definition of 't Hooft anomaly \cite{tHooft1979ratanomaly}.
$\bullet$ {\bf Second}, the QFT can have a hidden sector of the same dimension. We can match the 't Hooft anomaly of QFT with additional sector $S'$ of the same dimension,
but with the $S'$ sector with the opposite 't Hooft anomaly.
So, the combined system, the QFT and $S'$, can have a no 't Hooft anomaly at all.
$\bullet$ {\bf Third}, the QFT can have a hidden sector of one higher dimension. In order to have the symmetries of QFT with 't Hooft anomaly to be \emph{local} or \emph{onsite} instead,
we need to append
and regularize this $d$d QFT living on the boundary of $(d+1)$d Symmetry-Protected/Enriched Topological states (SPTs/SETs)
\cite{Chen2011pg1106.4772, Wen2013oza1303.1803, Wang2017locWWW1705.06728,
WangWen2018cai1809.11171}.\footnote{Symmetry-Protected Topological states (SPTs) are the short-range entangled states as a generalization of the free non-interacting topological insulators and superconductors \cite{2010RMP_HasanKane,2011_RMP_Qi_Zhang, RyuSPT, Kitaevperiod, Wen1111.6341}
with interactions.
Symmetry-Enriched Topological states (SETs) are topological ordered states enriched by global symmetries.
The enthusiastic readers can overview the condensed matter terminology bridging to QFT in \Refe{1711.11587GPW} for QFT theorists,
or the excellent condensed matter reviews \cite{Senthil1405.4015, Wen2016ddy1610.03911}.}
This is related to the Callan-Harvey anomaly inflow \cite{1984saCallanHarvey} of $(d+1)$d bulk and $d$d boundary with their spacetime dimensions differed by one.
$\bullet$ {\bf Fourth},
(some part of) the global symmetry can also be broken spontaneously or explicitly,
so that the 't Hooft anomaly is matched by \emph{symmetry breaking}.
$\bullet$ {\bf Fifth},
Of course, we can also have a certain combination of the First, Second, Third, and Fourth scenarios above, in order to make sense of QFT with 't Hooft anomaly.
\end{enumerate}
\item \label{remark:ABJanomaly}
Adler-Bell-Jackiw (ABJ \cite{Adler1969gkABJ, Bell1969tsABJ}) type of anomalies (e.g., \Fig{fig:triangle} ({\it 3})):
The non-conservations of the global symmetry current ${\cal J}$ is coupled to the background non-dynamical field ${\cal A}$
via the action term $\int {\cal A} \wedge \star {\cal J} \equiv \int \hspace{1pt}\mathrm{d}^d x({\cal A}_{\mu} {\cal J}^{\mu})$ with the Hodge dual star $\star$.
The non-conservation of current is proportional to the anomaly factor in $d$d spacetime
$$
\hspace{1pt}\mathrm{d} (\star {\cal J} ) \propto (F_a)^{d/2} \propto F_a \wedge F_a \wedge \dots.
$$
Here $a$ are dynamical gauge fields; we can also modify the equation appropriate for several different dynamical 1-form or higher-form gauge fields.
For $d=4$, the ABJ anomaly formula is precisely captured by \Fig{fig:triangle} ({\it 3}).
\item \label{remark:4anomaly}
Anomaly that involves two background fields of global symmetries and one dynamical gauge field (e.g., \Fig{fig:triangle} ({\it 4})).
\end{enumerate}
Indeed it is obvious to observe that the anomalies \ref{remark:dynamicalanomaly} are tighten to anomalies \ref{remark:tHooftanomaly}
\begin{enumerate}[label=\textcolor{blue}{(\greek*)}., ref={(\greek*)}]
\item \label{alp}
Anomalies from \ref{remark:dynamicalanomaly} can be related to anomalies from \ref{remark:tHooftanomaly} via the gauging principle.
\item \label{bet}
Anomalies from \ref{remark:tHooftanomaly} can be related to anomalies from \ref{remark:dynamicalanomaly} via the ungauging principle, or via gauging the higher
symmetries \cite{Gaiotto2014kfa1412.5148}
\end{enumerate}
Thus if we learn the gauge group of a gauge theory (e.g., SM, GUT or BSM),
we may identify its ungauged global symmetry group as an internal symmetry group, say ${{G}_{\text{internal}} }$ via
ungauging.\footnote{By gauging or ungauging, also depending on the representation of the matter fields that couple to the gauge theory,
we may gain or lose symmetries or higher symmetries \cite{Gaiotto2014kfa1412.5148}.
It will soon become clear,
for our purpose,
we need to primarily focus on
the ordinary (0-form) internal global symmetries
and their anomalies.
}
Now let us discuss the classifications of anomalies.
By ``all invertible quantum anomalies'' obtain from a cobordism classification, we mean the inclusion of:
\begin{enumerate}[leftmargin=2mm, label=\textcolor{blue}{(\roman*)}., ref={(\roman*)}]
\item
{\bf Perturbative local anomalies} captured by perturbative Feynman graph loop calculations,
classified by the integer group $\mathbb{Z}$ classes, or the free classes in mathematics.
Some selective examples from QFT or gravity are:
\begin{enumerate} [label=\textcolor{blue}{(\arabic*)}:, ref={(\arabic*)}]
\item
Perturbative fermionic anomalies from chiral fermions with U(1) symmetry, originated from Adler-Bell-Jackiw (ABJ) anomalies \cite{Adler1969gkABJ,Bell1969tsABJ}
with $\mathbb{Z}$ classes.
\item Perturbative bosonic anomalies from bosonic systems with U(1) symmetry
with $\mathbb{Z}$ classes \cite{WanWang2018bns1812.11967}.
\item Perturbative gravitational anomalies \cite{AlvarezGaume1983igWitten1984}.
\end{enumerate}
\item {\bf Non-perturbative global anomalies}, classified by a product of finite groups such as $\mathbb{Z}_n$, or the torsion classes in mathematics.
Some selective examples from QFT or gravity are:
\begin{enumerate} [label=\textcolor{blue}{(\arabic*)}:, ref={(\arabic*)}]
\item An SU(2) anomaly of Witten in 4d or in 5d \cite{Witten1982fp} with a $\mathbb{Z}_2$ class, which is a gauge anomaly.
\item A new SU(2) anomaly in 4d or in 5d \cite{WangWenWitten2018qoy1810.00844} with a different $\mathbb{Z}_2$ class, which is a mixed gauge-gravity anomaly.
\item Higher 't Hooft anomalies of $\mathbb{Z}_2$ class for a pure 4d SU(2) YM theory
with a second-Chern-class topological term \cite{Gaiotto2017yupZoharTTT, Wan2018zqlWWZ1812.11968, Wan2019oyr1904.00994} (or the so-called SU(2)$_{\theta =\pi}$ YM):
The higher anomaly involves a discrete 0-form time-reversal symmetry and a 1-form center $\mathbb{Z}_2$-symmetry.
The first anomaly is discovered in \cite{Gaiotto2017yupZoharTTT}; later the anomaly is refined via a mathematical well-defined 5d {bordism invariant as its topological term}, with additional new $\mathbb{Z}_2$ class anomalies found for Lorentz symmetry-enriched
four siblings of Yang-Mills gauge theories \cite{Wan2018zqlWWZ1812.11968, Wan2019oyr1904.00994}.
\item Global gravitational anomalies \cite{Witten1985xe} detected by exotic spheres.
\item Bosonic anomalies: Many types of
bosonic anomalies in diverse dimensions are global anomalies \cite{1403.0617, WangSantosWen1403.5256, Kapustin1404.3230, 1405.7689, JWangthesis}.
These bosonic anomalies only require bosonic degrees of freedom,
but without the requirement of chiral fermions. Many such bosonic anomalies
are related to group cohomology or generalized group cohomology theory, living on the boundary of SPTs \cite{Chen2011pg1106.4772},
closely related to Dijkgraaf-Witten topological gauge theories \cite{Dijkgraaf1989pzCMP}.
\end{enumerate}
\end{enumerate}
Our present work explore global anomalies (generically \emph{not} captured by Feynman graphs).
In order to help the readers to digest the physical meanings of global anomalies,
we should imagine the computation of anomalies on generic curved spacetime manifolds with mixed gauge and gravitational background probes,
in the cobordism theory setting (as in \Fig{fig:Cobordism2006}).
\begin{figure}[h!]
\centering
\includegraphics[width=4.in]{Cobordism-1910-BSM-GUT-figure.pdf}
\caption{General nonperturbative global anomalies \emph{not} captured by Feynman graphs can still be characterized by
generic curved manifolds with mixed gauge and gravitational background probes.
The figure shows a bordism between manifolds. Here $M$ and $M'$ are two closed $d$-manifolds, $N$ is a compact $d+1$-manifold whose boundary is the disjoint union of $M$ and $M'$, {namely $\prt N= M \sqcup M'$.} If there are additional $G$-structures on these manifolds, then the $G$-structure on $N$ is required to be compatible with the $G$-structures on $M$ and $M'$.
If there are additional maps from these manifolds to a fixed topological space, then the maps are also required to be compatible with each other. If these conditions are obeyed, then $M$ and $M'$ are called bordant equivalence
and $N$ is called a bordism between $M$ and $M'$. }\label{fig:Cobordism2006}
\end{figure}
\section{Dynamical Gauge Anomaly Cancellation}
\label{sec:DynamicalGaugeAnomalyCancellation}
In this section, we explicitly check the dynamical gauge anomaly cancellations of various SMs
with four gauge group $G_{\text{SM}_q}\equiv{({\rm SU}(3)\times {\rm SU}(2)\times {\rm U}(1))}/\mathbb{Z}_q$ with $q=1,2,3,6$.
The cobordism classifications of $G_{\text{SM}_q}$'s 4d anomalies are done in \cite{WW2019fxh1910.14668, 2019arXiv191011277D}.
The 4d anomalies can be written as 5d cobordism invariants, which are 5d invertible TQFT (iTQFT).
These 5d cobordism invariants/iTQFT are derived in \cite{WW2019fxh1910.14668},
we summarized the classifications and invariants in Table \ref{table:SU3SU2U1}.
Moreover, as \Refe{2019arXiv191011277D} points out correctly already, we will show that dynamical gauge anomaly
cancellation indeed holds to be true for all $G_{\text{SM}_q}$ with $q=1,2,3,6$.
\Refe{2019arXiv191011277D} has not written down the 5d cobordism invariants nor explicit anomaly polynomials.
However, by simply looking at the group classifications of anomalies, \Refe{2019arXiv191011277D} argues that there is only the famous Witten SU(2) nonperturbative global anomaly
of $\mathbb{Z}_2$ class \cite{Witten1982fp} ,
other anomalies all are perturbative local anomalies captured by Feynman graphs.
So why do we bother to do the calculations again, if \Refe{2019arXiv191011277D} has found all dynamical gauge anomalies cancel for $G_{\text{SM}_q}$?
There are multiple reasons:\\[-10mm]
\begin{itemize}
\item First, we will show that the 5d cobordism invariants obtained in \cite{WW2019fxh1910.14668} indeed match with the anomaly polynomials.
For perturbative local anomalies of $\mathbb{Z}$ classes, see \Fig{fig:triangle-dynamical},
we indeed can show that the 5d cobordism invariants map to some one-loop Feynman graph calculations known in the
standard QFT textbooks \cite{Weinberg1996Vol2, PeskinSchroeder1995book, Zee2003book, Srednicki2007book}.
(In contrast, \Refe{2019arXiv191011277D} focus on global anomalies, and pays less attention on perturbative local anomalies of $\mathbb{Z}$ classes. Follow \Refe{WW2019fxh1910.14668},
we will fill in this
gap by considering all local and all global anomalies.)
We should work through all these correspondences carefully to gain a solid confidence for our understanding of cobordism classifications of anomalies.
\item Second, we can learn how to translate the cobordism data from math into the anomalies in physics.
For example, in Table \ref{table:SU3SU2U1}, we see that the 5d cobordism invariant $c_2({\rm SU}(2))\tilde\eta$ in fact captures the 4d boundary theory has the
Witten SU(2) nonperturbative global anomaly \cite{Witten1982fp}.
\item Third, the most important thing, we will discover new phenomena later when we include the additional discrete symmetries into SM and GUT in \Sec{Sec:discretesymmetries}.
In fact, we will discover entirely new physics that previously have never been figured out in the past.
\end{itemize}
\noindent
{\bf Notations:}
Throughout our work, we write the three SU(2) Lie algebra generator ${\sigma^\mathrm{a}}$ of the rank-2 matrix of fundamental representation
satisfying ${\mathrm{Tr}}[\frac{\sigma^\mathrm{a}}{2} \frac{\sigma^\mathrm{b}}{2}] =\frac{1}{2} \delta_{\mathrm{a} \mathrm{b}}$ with $\mathrm{a}, \mathrm{b} \in \{1,2,3\}$.
We write the eight SU(3) Lie algebra generator ${\tau^\mathrm{a}}$ of the the rank-3 matrix of fundamental representation
satisfying ${\mathrm{Tr}}[\frac{\tau^\mathrm{a}}{2} \frac{\tau^\mathrm{b}}{2}] =\frac{1}{2} \delta_{\mathrm{a} \mathrm{b}}$ with $\mathrm{a}, \mathrm{b} \in \{1,2,\dots,8\}$.
\subsection{Summary of anomalies from a cobordism theory and Feynman diagrams}
\label{sec:localanomaliesFeynman}
\begin{table}[H]
\centering
\hspace*{-7mm}
\begin{tabular}{c c c }
\hline
\multicolumn{3}{c}{Cobordism group $\mathrm{TP}_d(G)$ with $G_{\text{SM}_q}\equiv{({\rm SU}(3)\times {\rm SU}(2)\times {\rm U}(1))}/\mathbb{Z}_q$ with $q=1,2,3,6$}\\
\hline
\hline
& classes & cobordism invariants\\
\hline
\hline\\[-2mm]
\multicolumn{3}{c}{$G={\rm Spin}\times {G_{\text{SM}_1}}$}\\[2mm]
\hline
5d & $\mathbb{Z}^5\times\mathbb{Z}_2$ &
$\begin{matrix}
\mu(\text{PD}(c_1({\rm U}(1)))), \quad \text{CS}_1^{{\rm U}(1)}c_1({\rm U}(1))^2, \quad
\text{CS}_1^{{\rm U}(1)}c_2({\rm SU}(2)) {\sim c_1({\rm U}(1))\text{CS}_3^{{\rm SU}(2)}},\\
\text{CS}_1^{{\rm U}(1)}c_2({\rm SU}(3)) {\sim c_1({\rm U}(1))\text{CS}_3^{{\rm SU}(3)}}, \quad\frac{\text{CS}_5^{{\rm SU}(3)}}{2}, \quad c_2({\rm SU}(2))\tilde\eta
\end{matrix}$\\
\hline
\hline\\[-2mm]
\multicolumn{3}{c}{$G={\rm Spin}\times {G_{\text{SM}_2}}$}\\[2mm]
\hline
5d & $\mathbb{Z}^5$ &
$\begin{matrix}\mu(\text{PD}(c_1({\rm U}(2)))),
\quad
\text{CS}_1^{{\rm U}(2)}c_1({\rm U}(2))^2,
\quad
\frac{\text{CS}_1^{{\rm U}(2)}c_2({\rm U}(2))}{2} {\sim}
\frac{c_1({\rm U}(2)) \text{CS}_3^{{\rm U}(2)}}{2},\\
\text{CS}_1^{{\rm U}(2)}c_2({\rm SU}(3))
{\sim c_1({\rm U}(2)) \text{CS}_3^{{\rm SU}(3)}},
\quad
\frac{\text{CS}_5^{{\rm SU}(3)}}{2}
\end{matrix}$\\
\hline
\hline\\[-2mm]
\multicolumn{3}{c}{$G={\rm Spin}\times {G_{\text{SM}_3}}$}\\[2mm]
\hline
5d & $\mathbb{Z}^5\times\mathbb{Z}_2$ &
$\begin{matrix}\mu(\text{PD}(c_1({\rm U}(3)))),
\quad
\text{CS}_1^{{\rm U}(3)}c_1({\rm U}(3))^2,
\quad
\text{CS}_1^{{\rm U}(3)}c_2({\rm SU}(2)) {\sim c_1({\rm U}(3)) \text{CS}_3^{{\rm SU}(2)}},
\\
\frac{\text{CS}_1^{{\rm U}(3)}c_2({\rm U}(3))+\text{CS}_5^{{\rm U}(3)}}{2}
{\sim
\frac{
c_1({\rm U}(3)) \text{CS}_3^{{\rm U}(3)}
+\text{CS}_5^{{\rm U}(3)}}{2}
},
\quad
\text{CS}_5^{{\rm U}(3)},
\quad
c_2({\rm SU}(2))\tilde\eta
\end{matrix}$ \\
\hline
\hline\\[-2mm]
\multicolumn{3}{c}{$G={\rm Spin}\times {G_{\text{SM}_6}}$}\\[2mm]
\hline
5d & $\mathbb{Z}^5$ &
$\begin{matrix}\mu(\text{PD}(c_1({\rm U}(2)))) \sim \mu(\text{PD}(c_1({\rm U}(3)))),
\quad
\text{CS}_1^{{\rm U}(3)}c_1({\rm U}(3))^2,
\quad
\frac{\text{CS}_1^{{\rm U}(3)}c_2({\rm U}(2))}{2} {\sim \frac{c_1({\rm U}(3)) \text{CS}_3^{{\rm U}(2)}}{2}},
\\
\frac{\text{CS}_1^{{\rm U}(3)}c_2({\rm U}(3))+\text{CS}_5^{{\rm U}(3)}}{2}
{\sim
\frac{ c_1({\rm U}(3)) \text{CS}_3^{{\rm U}(3)}
+\text{CS}_5^{{\rm U}(3)}}{2}
}, \quad
\text{CS}_5^{{\rm U}(3)}
\end{matrix}$ \\
\hline
\hline
\end{tabular}
\caption{The 4d anomalies can be written as 5d cobordism invariants
of
$\Omega^{d=5}_{G} \equiv
\mathrm{TP}_{d=5}(G)$,
which are 5d iTQFTs.
These 5d cobordism invariants/iTQFTs are derived in \cite{WW2019fxh1910.14668}.
We summarized the group classifications of 4d anomalies and their 5d cobordism invariants.
The anomaly classification of $\mathbb{Z}^5$ means that there are 5 perturbative local anomalies (of $\mathbb{Z}$ classes descended from the 6d bordism group
$\Omega_{d=6}^{G}$),
precisely match 5 perturbative one-loop triangle Feynman diagrams in \Fig{fig:triangle-dynamical}.
The anomaly classification of $\mathbb{Z}_2$ means that there is a 1 nonperturbative global anomaly, which
turns out to be Witten SU(2) anomaly \cite{Witten1982fp}.
The $c_j(G)$ is the $j$th Chern class of the associated vector bundle of the principal $G$-bundle.
The $\mu$ is the 3d Rokhlin invariant.
If $\partial M^4=M^3$, then $\mu(M^3)=(\frac{\sigma-{\rm F}\cdot{\rm F}}{8} )(M^4)$, thus
$\mu(\text{PD}(c_1({\rm U}(1))))$ is related to $\frac{c_1({\rm U}(1))(\sigma-{\rm F} \cdot {\rm F} )}{8}$.
Here $\cdot$ is the intersection form of $M^4$.
The ${\rm F} $ is the characteristic 2-surface \cite{Saveliev} in a 4-manifold $M^4$, it obeys the condition ${\rm F}\cdot x=x\cdot x\mod2$ for all $x\in\operatorname{H}_2(M^4,\mathbb{Z})$.
By the Freedman-Kirby theorem, we have $(\frac{\sigma-{\rm F}\cdot{\rm F}}{8} )(M^4)=\text{Arf}(M^4,{\rm F})\mod2$.
The PD is defined as the Poincar\'e dual.
The Arf is a 2d Arf invariant, whose condensed matter realization is the 1+1d Kitaev fermionic chain \cite{Kitaev2001chain0010440}.
The $\tilde\eta$ is a mod 2 index of 1d Dirac operator.
In \Sec{sec:Wittenanomaly} and \Refe{WW2019fxh1910.14668},
we propose that the $\mathbb{Z}_2$ class 5d cobordism invariant $c_2({\rm SU}(2))\tilde\eta$
corresponds to the 4d Witten SU(2) anomaly \cite{Witten1982fp}.
See our notational conventions in Sec.~1 and Sec.~1.2.4 of \Refe{WW2019fxh1910.14668}.
The symbol ``$\sim$'' here means the equivalent rewriting of cobordism invariants on a closed 5-manifold.}
\label{table:SU3SU2U1}
\end{table}
\newpage
\begin{figure}[h!]
\centering
(i) \includegraphics[width=2.0in]{anomaly-U1YU1YU1Y-.pdf}
(ii) \includegraphics[width=2.0in]{anomaly-U1YSU2SU2-.pdf}
(iii) \includegraphics[width=2.0in]{anomaly-U1YSU3SU3-.pdf}
\\[2mm]
(iv) \includegraphics[width=1.8in]{anomaly-U1Ygravgrav-.pdf}
(v) \includegraphics[width=2.2in]{anomaly-SU3SU3SU3-.pdf}
\caption{Examples of dynamical gauge anomaly cancellations in SM. In fact, the 5
perturbative local anomalies from
perturbative one-loop triangle Feynman diagrams
precisely match
anomaly classification of $\mathbb{Z}^5$ obtained from the cobordism group calculations in Table \ref{table:SU3SU2U1} and
\Refe{WW2019fxh1910.14668}.}
\label{fig:triangle-dynamical}
\end{figure}
\subsection{U(1)${}_Y^3$: 4d local anomaly from 5d $\text{CS}_1^{{\mathrm{U}}(1)}c_1({\mathrm{U}}(1))^2$ and 6d $c_1({\mathrm{U}}(1))^3$}
\label{sec:U(1)Y3}
We read from \Refe{WW2019fxh1910.14668} and Table \ref{table:SU3SU2U1} for the $\mathbb{Z}$ class of the 5d cobordism invariants of the following:\\
5d $\text{CS}_1^{{\rm U}(1)}c_1({\rm U}(1))^2$ for $G_{\text{SM}_1}$,
5d $\text{CS}_1^{{\rm U}(2)}c_1({\rm U}(2))^2$ for $G_{\text{SM}_2}$ and $G_{\text{SM}_6}$,
while 5d $\text{CS}_1^{{\rm U}(3)}c_1({\rm U}(3))^2$ for $G_{\text{SM}_3}$ and $G_{\text{SM}_6}$.
These 5d cobordism invariants correspond to the 4d perturbative local anomalies captured by the one-loop Feynman graph:
\begin{eqnarray} \label{eq:anomaly-U1YU1YU1Y}
\includegraphics[width=2.2in]{anomaly-U1YU1YU1Y-.pdf}.
\end{eqnarray}
Without losing generality, we focus on the 4d cubic anomaly (U(1)$_Y$)$^3$ from 5d $\text{CS}_1^{{\rm U}(1)}c_1({\rm U}(1))^2$, which also descends from 6d $c_1({\rm U}(1))^3$
of bordism group $\Omega_6$ in \Refe{WW2019fxh1910.14668}.
Plug in data in \Sec{sec:2OverviewSMGUT}, it is standard to check the anomaly \Eq{eq:anomaly-U1YU1YU1Y} vanishes,\footnote{When
we switch between from the $L$-chiral fermion to the $R$-chiral fermion (anti-chiral fermion), there could be an additional minus
sign.}
\begin{eqnarray}
\sum_\mathrm{q} {\mathrm{Tr}}[(\hat Y_\mathrm{q})^3 ]
&=& \sum_{
{\mathrm{q}_L},
{\mathrm{q}_R}
} (Y_{\mathrm{q}_L})^3 -
(Y_{\mathrm{q}_R})^3\cr
&=&
\frac{1}{2} \delta_{ab} N_{\text{generation}} \cdot \bigg( N_c \cdot \Big( 2 \cdot (1/6)^3 + (-2/3)^3+ (1/3)^3 \Big) + 2 \cdot (-1/2)^3 + (1)^3 + (0)^3 \bigg)
\cr
&=& N_{\text{generation}} \cdot ( - N_c +3) \cdot (1/4),
\end{eqnarray}
which is 0 when $N_c=3$ for 3 colors as we have. The $N_{\text{generation}}$ (or $N_{\text{family}}$) counts the number of generations (same as families).
\subsection{U(1)${}_Y$-SU(2)${}^2$: 4d local anomaly from 5d $\text{CS}_1^{{\rm U}(1)}c_2({\rm SU}(2))$ and 6d $c_1({\rm U}(1))c_2({\rm SU}(2))$}
\label{sec:U(1)YSU(2)2}
We read from \Refe{WW2019fxh1910.14668} and Table \ref{table:SU3SU2U1} for a $\mathbb{Z}$ class of 5d cobordism invariants of the following:\\
5d $\text{CS}_1^{{\rm U}(1)}c_2({\rm SU}(2))$ for $G_{\text{SM}_1}$,
5d $\frac{\text{CS}_1^{{\rm U}(2)}c_2({\rm U}(2))}{2} {\sim}
\frac{c_1({\rm U}(2)) \text{CS}_3^{{\rm U}(2)}}{2}$ for $G_{\text{SM}_2}$,\\
5d $\text{CS}_1^{{\rm U}(3)}c_2({\rm SU}(2)) {\sim c_1({\rm U}(3)) \text{CS}_3^{{\rm SU}(2)}}$ for $G_{\text{SM}_3}$, and
5d $\frac{\text{CS}_1^{{\rm U}(3)}c_2({\rm U}(2))}{2} {\sim \frac{c_1({\rm U}(3)) \text{CS}_3^{{\rm U}(2)}}{2}}$ for $G_{\text{SM}_6}$.
These 5d cobordism invariants correspond to the 4d perturbative local anomalies captured by the one-loop Feynman graph:
\begin{eqnarray} \label{eq:anomaly-U1YSU2SU2}
\includegraphics[width=2.2in]{anomaly-U1YSU2SU2-.pdf}.
\end{eqnarray}
Without losing generality, we focus on the 4d anomaly {U(1)$_Y$-SU(2)$^2$} from
5d $\text{CS}_1^{{\rm U}(1)}c_2({\rm SU}(2))$, which also descends from 6d $c_1({\rm U}(1))c_2({\rm SU}(2))$
of bordism group $\Omega_6$ in \Refe{WW2019fxh1910.14668}.
Plug in data in \Sec{sec:2OverviewSMGUT}, we check the anomaly \Eq{eq:anomaly-U1YSU2SU2} vanishes,
\begin{eqnarray}
\sum_\mathrm{q} {\mathrm{Tr}}[\hat Y_\mathrm{q} \sigma^a \sigma^b ]
&=& \frac{1}{2} \delta_{ab}
(\sum_{
{\mathrm{q}_L},
{\mathrm{q}_R}
} (Y_{\mathrm{q}_L}) -
(Y_{\mathrm{q}_R}))
\cr
&=&
\frac{1}{2} \delta_{ab}N_{\text{generation}} \cdot \bigg( N_c \cdot \Big( 2 \cdot (1/6) \Big) + 2 \cdot (-1/2) \bigg)
\cr
&=& \frac{1}{2} \delta_{ab} N_{\text{generation}} \cdot ( N_c/3 -1),
\end{eqnarray}
which is 0 when $N_c=3$.
\subsection{U(1)${}_Y$-SU(3)${}_c^2$: 4d local anomaly from 5d $\text{CS}_1^{{\rm U}(1)}c_2({\rm SU}(3))$ and 6d $c_1({\rm U}(1))c_2({\rm SU}(3))$}
\label{sec:U(1)YSU(3)2}
We read from \Refe{WW2019fxh1910.14668} and Table \ref{table:SU3SU2U1} for a $\mathbb{Z}$ class of 5d cobordism invariants of the following:\\
5d $\text{CS}_1^{{\rm U}(1)}c_2({\rm SU}(3)) {\sim c_1({\rm U}(1))\text{CS}_3^{{\rm SU}(3)}}$ for $G_{\text{SM}_1}$,
5d $ \text{CS}_1^{{\rm U}(2)}c_2({\rm SU}(3))
{\sim c_1({\rm U}(2)) \text{CS}_3^{{\rm SU}(3)}}$ for $G_{\text{SM}_2}$,
5d $\frac{\text{CS}_1^{{\rm U}(3)}c_2({\rm U}(3))+\text{CS}_5^{{\rm U}(3)}}{2}
{\sim
\frac{
c_1({\rm U}(3)) \text{CS}_3^{{\rm U}(3)}
+\text{CS}_5^{{\rm U}(3)}}{2}
}$ for $G_{\text{SM}_3}$, and
5d $\frac{\text{CS}_1^{{\rm U}(3)}c_2({\rm U}(3))+\text{CS}_5^{{\rm U}(3)}}{2}
{\sim
\frac{ c_1({\rm U}(3)) \text{CS}_3^{{\rm U}(3)}
+\text{CS}_5^{{\rm U}(3)}}{2}
}$ for $G_{\text{SM}_6}$.
(Note that part of the additional contribution from $\text{CS}_5^{{\rm U}(3)}$ or $\text{CS}_5^{{\rm SU}(3)}$ will be separately discussed later in \Sec{sec:SU(3)3}
and \Eq{eq:anomaly-SU3SU3SU3}.)
These 5d cobordism invariants correspond to the 4d perturbative local anomalies captured by the one-loop Feynman graph:
\begin{eqnarray} \label{eq:anomaly-U1YSU3SU3}
\includegraphics[width=2.2in]{anomaly-U1YSU3SU3-.pdf}.
\end{eqnarray}
Without losing generality, we focus on the 4d anomaly {U(1)$_Y$-SU(3)$_c^2$} from
5d $\text{CS}_1^{{\rm U}(1)}c_2({\rm SU}(3)) {\sim c_1({\rm U}(1))\text{CS}_3^{{\rm SU}(3)}}$, which also descends from 6d $c_1({\rm U}(1))c_2({\rm SU}(3))$
of bordism group $\Omega_6$ in \Refe{WW2019fxh1910.14668}.
Plug in data in \Sec{sec:2OverviewSMGUT}, we check the anomaly \Eq{eq:anomaly-U1YSU3SU3} vanishes,
\begin{eqnarray} \label{eq:anomaly-U1YSU3SU3}
\sum_\mathrm{q} {\mathrm{Tr}}[\hat Y_\mathrm{q} \tau^a \tau^b ]
&=& \sum_{
{\mathrm{q}_L},
{\mathrm{q}_R}
} {\mathrm{Tr}}[\hat Y_{\mathrm{q}_L} \tau^a \tau^b ] -
{\mathrm{Tr}}[\hat Y_{\mathrm{q}_R} \tau^a \tau^b ]
= \frac{1}{2} \delta_{ab}
(\sum_{
{\mathrm{q}_L},
{\mathrm{q}_R}
} (Y_{\mathrm{q}_L}) -
(Y_{\mathrm{q}_R}))
\cr
&=& \frac{1}{2} \delta_{ab} N_{\text{generation}} \cdot \bigg( N_c \cdot \Big( 2 \cdot (1/6) + (-2/3)+ (1/3) \Big) \bigg)
\cr
&=& \frac{1}{2} \delta_{ab} N_{\text{generation}} \cdot N_c \cdot 0=0 .
\end{eqnarray}
\subsection{U(1)$_Y$-(gravity)$^2$: 4d local anomaly from 5d $\mu(\text{PD}(c_1({\rm U}(1))))$ and 6d $\frac{c_1({\rm U}(1))(\sigma-{\rm F} \cdot {\rm F} )}{8}$}
\label{sec:U(1)Ygrav2}
We read from \Refe{WW2019fxh1910.14668} and Table \ref{table:SU3SU2U1} for a $\mathbb{Z}$ class of 5d cobordism invariants of the following:\\
5d $\mu(\text{PD}(c_1({\rm U}(1))))$ for $G_{\text{SM}_1}$,
5d $\mu(\text{PD}(c_1({\rm U}(2))))$ for $G_{\text{SM}_2}$ and $G_{\text{SM}_6}$,
while
5d $\mu(\text{PD}(c_1({\rm U}(3))))$ for $G_{\text{SM}_3}$ and $G_{\text{SM}_6}$.
These 5d cobordism invariants correspond to the 4d perturbative local anomalies captured by the one-loop Feynman graph:
\begin{eqnarray} \label{eq:anomaly-U1Ygravgrav}
\includegraphics[width=2.2in]{anomaly-U1Ygravgrav-.pdf}.
\end{eqnarray}
Without losing generality, we focus on the 4d anomaly U(1)$_Y$-gravity$^2$ of U(1)$_Y$-gravitational anomaly
from
5d $\mu(\text{PD}(c_1({\rm U}(1))))$,
which also descends from 6d $\frac{c_1({\rm U}(1))(\sigma-{\rm F} \cdot {\rm F} )}{8}$
of bordism group $\Omega_6$ in \Refe{WW2019fxh1910.14668}.
Plug in data in \Sec{sec:2OverviewSMGUT}, we check the anomaly \Eq{eq:anomaly-U1Ygravgrav} vanishes,
\begin{eqnarray}
\sum_\mathrm{q} {\mathrm{Tr}}[\hat Y_\mathrm{q} ]
&=& \sum_{
{\mathrm{q}_L},
{\mathrm{q}_R}
} (Y_{\mathrm{q}_L})-
(Y_{\mathrm{q}_R})\cr
&=& N_{\text{generation}} \cdot \bigg( N_c \cdot \Big( 2 \cdot (1/6) + (-2/3)+ (1/3) \Big) + 2 \cdot (-1/2) + (1) + (0) \bigg)
\cr
&=& N_{\text{generation}} \cdot ( 0 \cdot N_c +0) =0.
\end{eqnarray}
We remark that if we view the gravity as dynamical fields, then \Eq{eq:anomaly-U1Ygravgrav} checks the dynamical gauge anomaly cancellation of the
\Fig{fig:triangle} ({\it 1}) and Remark \ref{remark:dynamicalanomaly};
if we view the gravity as background probe fields, then \Eq{eq:anomaly-U1Ygravgrav} checks the anomaly cancellation of the type of
\Fig{fig:triangle} ({\it 4}) and Remark \ref{remark:4anomaly}.
\subsection{SU(3)$_c^3$: 4d local anomaly from 5d $\frac{1}{2}{\text{CS}_5^{{\rm SU}(3)}}$ and 6d $\frac{1}{2}{c_3({\rm SU}(3))}$}
\label{sec:SU(3)3}
We read from \Refe{WW2019fxh1910.14668} and Table \ref{table:SU3SU2U1} for a $\mathbb{Z}$ class of 5d cobordism invariants of the following:\\
5d $\frac{1}{2}{\text{CS}_5^{{\rm SU}(3)}}$ for $G_{\text{SM}_1}$ and $G_{\text{SM}_2}$,
and 5d ${\text{CS}_5^{{\rm U}(3)}}$ for $G_{\text{SM}_3}$ and $G_{\text{SM}_6}$.
(Note that part of the contributions from $\text{CS}_5^{{\rm U}(3)}$ also occur in \Sec{sec:U(1)YSU(3)2}.)
These 5d cobordism invariants correspond to the 4d perturbative local anomalies captured by the one-loop Feynman graph:
\begin{eqnarray} \label{eq:anomaly-SU3SU3SU3}
\includegraphics[width=2.2in]{anomaly-SU3SU3SU3-.pdf}.
\end{eqnarray}
Without losing generality, we focus on the 4d anomaly {SU(3)$_c^3$}
from
5d $\frac{1}{2}{\text{CS}_5^{{\rm SU}(3)}}$,
which also descends from 6d $\frac{1}{2}{c_3({\rm SU}(3))}$
of bordism group $\Omega_6$ in \Refe{WW2019fxh1910.14668}.
Plug in data in \Sec{sec:2OverviewSMGUT}, we check the anomaly \Eq{eq:anomaly-SU3SU3SU3} vanishes.
In fact, in the context of SM physics, even without checking explicitly,
it is clear that this {SU(3)$_c^3$} anomaly \Eq{eq:anomaly-SU3SU3SU3} must vanish,
since the color {SU(3)$_c$} is vector gauge theory not chiral gauge theory respect to the color charge.
We recall that only U(1)$_Y$ and SU(2)$_{\text{weak}}$ are chiral gauge theories in SM.
Readers may ask what happen to the {SU(2)$^3$} anomaly by replacing the gauge fields in \Eq{eq:anomaly-SU3SU3SU3} to SU(2), since
SU(2)$_{\text{weak}}$ is chiral? The answer is that {SU(2)$^3$} anomaly does not exist thus must vanish,
because there is no such corresponding 5d cobordism invariant found in \Refe{WW2019fxh1910.14668} and Table \ref{table:SU3SU2U1}.
In fact, for SU(2) and SO($N$) group, all representations have zero 4d perturbative local anomalies, thus
they must have none of $\mathbb{Z}$ classes of 5d cobordism invariants, agreed with \Refe{WW2019fxh1910.14668}.
\subsection{Witten SU(2) anomaly: 4d $\mathbb{Z}_2$ global anomaly from 5d $c_2({\rm SU}(2))\tilde\eta$ and 6d $c_2({\rm SU}(2))\text{Arf}$}
\label{sec:Wittenanomaly}
{The old SU(2) anomaly of Witten in 4d \cite{Witten1982fp}}
is summarized in \cite{WangWenWitten2018qoy1810.00844} for the context we need.
We read from \Refe{WW2019fxh1910.14668} and Table \ref{table:SU3SU2U1} for a $\mathbb{Z}_2$ class of 5d cobordism invariant
and suggest the 4d SU(2) anomaly corresponds to 5d $c_2({\rm SU}(2))\tilde\eta $, and descends from 6d $c_2({\rm SU}(2))\text{Arf}$
of bordism group $\Omega_6$ in \Refe{WW2019fxh1910.14668}.\footnote{As explained in the Notation in the end of \Sec{sec:Intro}:
The $\tilde{\eta}$ is a mod 2 index of 1d Dirac operator as a cobordism invariant of $\Omega_1^{{\rm Spin}}=\mathbb{Z}_2$.
The Arf invariant \cite{Arf1941} is a mod 2 cobordism invariant of $\Omega_2^{{\rm Spin}}=\mathbb{Z}_2$,
whose realization is the 1+1d Kitaev fermionic chain \cite{Kitaev2001chain0010440}.}
\begin{enumerate}
\item
5d $c_2({\rm SU}(2))\tilde\eta$ and 4d Witten SU(2) anomaly: This 4d anomaly is a mod 2 index of $\mathbb{Z}_2$ class
counts the spin-$2r +1/2$ (or ${\bf 4r} + {\bf 2}$ in the dimension of representation)
Weyl spinor as fermion doublet under SU(2) \cite{WangWenWitten2018qoy1810.00844}.
From \Sec{sec:2OverviewSMGUT}, there are four of spin-$2r +1/2$ fermions from
$({\bf 3},{\bf 2}, 1/6)_L$ and $({\bf 1},{\bf 2},-1/2)_L$, multiplied by $N_{\text{generation}}$. So we check overall the Witten anomaly vanishes in SM:
$$
(\text{even number})\vert_{\text{of ${\bf 2}$}} \mod 2 = 0.
$$
\item N\"aively, 5d $c_2({\rm SU}(2))\tilde\eta$ only presents for $G_{\text{SM}_1}$ and $G_{\text{SM}_3}$,
but not for $G_{\text{SM}_2}$ and $G_{\text{SM}_6}$. Readers may wonder how does
Witten anomaly vanish for SM of $q=2,6$? \\
$\bullet$ \Refe{Davighi2020bvi2001.07731} explains nicely and accurately that
the Witten anomaly mutates from a $\mathbb{Z}_2$ class global anomaly
into a subclass of perturbative local $\mathbb{Z}$ class when we changes the SU(2) group to the U(2) group, namely
for $q=2,6$.\\
$\bullet$ \Refe{WW2019fxh1910.14668} gives a formal explanation as follows:
The difference between $q=1$ and $q=2$ case is parallel to the difference between $q=3$ and $q=6$ case.
So without losing generality, we focus on the difference between $q=1$ and $q=2$ cases.
The Witten anomaly from $c_2({\rm SU}(2))\tilde\eta$ of $\mathbb{Z}_2$ for $q=1$ vanishes in $q=2$,
while in contrast the {$\text{CS}_1^{{\rm U}(1)}c_2({\rm SU}(2))
\sim
c_1({\rm U}(1))\text{CS}_3^{{\rm SU}(2)}
$} of $\mathbb{Z}$ for $q=1$ becomes {
$\frac{1}{2}\text{CS}_1^{{\rm U}(2)}c_2({\rm U}(2))
\sim
\frac{1}{2}c_1({\rm U}(2))\text{CS}_3^{{\rm U}(2)}
$}
in $q=2$. The $\text{CS}_1^{{\rm U}(1)}c_2({\rm SU}(2))$ in 5d comes from $c_1({\rm U}(1))c_2({\rm U}(2))$ in 6d.
Here the $c_1({\rm U}(2))c_2({\rm U}(2))\mod2$ is the quotient $\mathbb{Z}_2$, the $c_1({\rm U}(2))c_2({\rm U}(2))$ is the total $\mathbb{Z}$, and the $\frac{1}{2}c_1({\rm U}(2))c_2({\rm U}(2))$ is the normal $\mathbb{Z}$ in the short exact sequence $0 \to \mathbb{Z} \overset{2}{\to} \mathbb{Z} \to \mathbb{Z}_2 \to 0$.
Let us
express the levels of those cobordism invariants as
$k_{q=2}$, $k_{q=1}$, and ${k}_{q=1}'$
respectively. Indeed we can also understand
the short exact sequence as the classes of the levels:
$0 \to
k_{q=1} \in \mathbb{Z} \overset{2}{\to}
k_{q=2} \in \mathbb{Z}
\to k_{q=1}' \in \mathbb{Z}_2 \to 0$.
There is yet another way to explain why $c_1({\rm U}(2))c_2({\rm U}(2))\mod2$ vanishes for $q=2$:
Via Wu formula, the $c_1({\rm U}(2))c_2({\rm U}(2))=\mathrm{Sq}^2c_2({\rm U}(2))=(w_2+w_1^2)c_2({\rm U}(2))=0\mod2$ on the Spin 6-manifolds.
In short, the $k_{q=2} \in \mathbb{Z}$ of local anomalies now also carry information of the
$k_{q=1}' \in \mathbb{Z}_2$ of the Witten SU(2) global anomaly.
\end{enumerate}
In summary of \Sec{sec:DynamicalGaugeAnomalyCancellation}, by checking
five $\mathbb{Z}$ classes of local anomalies and one
$\mathbb{Z}_2$ class of Witten SU(2) global anomaly, we have shown that the for SM with $G_{\text{SM}_q}$ of $q=1,2,3,6$
are indeed free from all {dynamical gauge anomalies, thus dynamical gauge anomaly cancellation holds.}
As we have checked, the $G_{\text{SM}_q}$ is a healthy chiral gauge theory by its own with a dynamical gauge group $G_{\text{SM}_q}$.\footnote{It has been long
sought that $G_{\text{SM}_q}$ is a healthy chiral gauge theory by its own with a dynamical gauge group $G_{\text{SM}_q}$
free from all dynamical gauge anomalies. However, it is only until very recently by the systematic computations of cobordism groups
in \cite{GarciaEtxebarriaMontero2018ajm1808.00009} (which checks $q=1$), \cite{Davighi2020bvi2001.07731} and \cite{WW2019fxh1910.14668}
(which the two papers checks $q=1,2,3,6$)
completing the full checks on
$G_{\text{SM}_q}$.
Without a cobordism classification of anomalies,
previous literature either only check perturbative anomalies,
or may still miss additional global anomaly constraints (as we shall see new anomaly constraints in \Sec{Sec:discretesymmetries}).}
However, what if we include additional global symmetries or gauge sectors?
Such as the ${ \mathbf{B}- \mathbf{L}}$
or
$X \equiv
5({ \mathbf{B}- \mathbf{L}})-4Y$? This motivates us to explore further in the next section \Sec{Sec:discretesymmetries}.
\section{Anomaly Matching for SM and GUT with Extra $({\mathbf{B}- \mathbf{L}})$ Symmetries}
\label{Sec:discretesymmetries}
\begin{figure}[h!]
\centering
(i) \includegraphics[width=2.0in]{anomaly-BLU1YU1Y-.pdf}
(ii) \includegraphics[width=2.0in]{anomaly-BLSU2SU2-.pdf}
(iii) \includegraphics[width=2.0in]{anomaly-BLSU3SU3-.pdf}
\\[2mm]
(iv) \includegraphics[width=1.8in]{anomaly-BLgravgrav-.pdf}
(v) \includegraphics[width=1.8in]{anomaly-BLBLBL-.pdf}
(vi)\includegraphics[width=2.2in]{anomaly-XSU5SU5-.pdf}
\caption{Examples of anomaly constraint for SM (or GUT) with extra symmetries such as ${{ \mathbf{B}- \mathbf{L}}}$ or $X \equiv
5({ \mathbf{B}- \mathbf{L}})-4Y$.
We only show
perturbative local anomalies from
perturbative one-loop triangle Feynman diagrams discussed in \Eq{eq:Spinccobordism}.
We will explore nonperturbative global anomalies (not captured by Feynman diagrams) in later sections.
Assume the gravity contributes as background field: \newline
$\bullet$ If ${{ \mathbf{B}- \mathbf{L}}}$ or $X$ is not gauged, (i), (ii), (iii), and (vi) are ABJ anomalies of \Fig{fig:triangle} ({\it 3}) and Remark \ref{remark:ABJanomaly};\newline
(iv) and (v) are 't Hooft anomalies of \Fig{fig:triangle} ({\it 2}) and Remark \ref{remark:tHooftanomaly}.\newline
$\bullet$ If ${{ \mathbf{B}- \mathbf{L}}}$ or $X$ is gauged, (i)-(iv), (vi) are dynamical gauge anomalies of \Fig{fig:triangle} ({\it 1}) and Remark \ref{remark:dynamicalanomaly};\newline
(v) is an anomaly of \Fig{fig:triangle} ({\it 4}) and Remark \ref{remark:4anomaly}.\newline
If these anomalies are not matched, we can still saturate the anomalies by proposing new sectors appending to the QFT;
we will explore those new physics in \Sec{sec:HiddenTopologicalSectors}.}
\label{fig:triangle-B-L}
\end{figure}
\subsection{SM and GUT with extra continuous symmetries and a cobordism theory}
In \Sec{sec:2OverviewSMGUT} Remark \ref{RemarkEmbed3},
for SM and SU(5) GUT with an extra ${\rm U}(1)_{{ \mathbf{B}- \mathbf{L}}}$ or ${\rm U}(1)_{X}$ symmetry, we need to consider Spin${}^c \equiv {\rm Spin}(d) \times_{\mathbb{Z}_2} {\rm U}(1)$ structure.
We find the embedding:
$$
{\rm Spin}^c(d)\times {\rm SU}(5)
\supset
{\rm Spin}^c(d) \times \frac{{\rm SU}(3) \times {\rm SU}(2) \times {\rm U}(1)}{\mathbb{Z}_6}.
$$
\Refe{2019arXiv191011277D} checks that the 5d bordism group:
\begin{eqnarray}
\Omega_5^{{\rm Spin}^c}( \frac{{{\rm SU}(3)\times {\rm SU}(2)\times {\rm U}(1)}}{\mathbb{Z}_q})=0,
\end{eqnarray}
which means no global anomalies.
\Refe{WanWangv2} computes the following cobordism groups $\mathrm{TP}_5$ and bordism groups $\Omega_6$:
\begin{eqnarray} \label{eq:Spinccobordism}
\begin{array}{ll}
\mathrm{TP}_5({{\rm Spin}^c}\times {\rm SU}(5))=\mathbb{Z}^4, & \Omega_6^{{\rm Spin}^c}({\rm SU}(5))=\mathbb{Z}^4.\\[3mm]
\mathrm{TP}_5({{\rm Spin}^c}\times \frac{{{\rm SU}(3)\times {\rm SU}(2)\times {\rm U}(1)}}{\mathbb{Z}_q})=\mathbb{Z}^{11},& \Omega_6^{{\rm Spin}^c}( \frac{{{\rm SU}(3)\times {\rm SU}(2)\times {\rm U}(1)}}{\mathbb{Z}_q})=\mathbb{Z}^{11}.
\end{array}
\end{eqnarray}
\Refe{WanWangv2} finds that these $\mathrm{TP}_5$ and $\Omega_6$ only contain $\mathbb{Z}$ classes, thus they only correspond to 4d local anomalies
captured by the one-loop Feynman graph shown in \Fig{fig:triangle-B-L}.
We emphasize that if ${\rm U}(1)_{{ \mathbf{B}- \mathbf{L}}}$ or ${\rm U}(1)_{X}$ is free of all anomalies,
then we can dynamically gauge this symmetry. In that case, we should regard the corresponding gauge field as a Spin$^c$ connection
instead of the familiar U(1) gauge fields,
since the original theory requires to be defined on Spin$^c$ manifolds.
(When we mention gauge fields for gauging ${\rm U}(1)_{{ \mathbf{B}- \mathbf{L}}}$ or ${\rm U}(1)_{X}$, what we really have in mind is the
Spin$^c$ connection.)
The ${\rm U}(1)_{{ \mathbf{B}- \mathbf{L}}}$ or ${\rm U}(1)_{X}$ is free of all anomalies if they are free from perturbative local anomalies given by \Eq{eq:Spinccobordism},
since they do not have global anomalies.
Are all the perturbative local anomalies canceled for these {SM and GUT with extra continuous symmetries}?
The baryon and lepton local currents densities are
\begin{eqnarray}
j_{\bfB}^\mu= \frac{1}{3} (\bar q_L \gamma^\mu q_L + \bar u_R \gamma^\mu u_R + \bar d_R \gamma^\mu d_R), \quad
j_{\mathbf{L}}^\mu= (\bar l_L \gamma^\mu l_L +\bar e_R \gamma^\mu e_R +N_{\nu_R} \bar \nu_R \gamma^\mu \nu_R).
\end{eqnarray}
Here $N_{\nu_R}$ is the number of right-handed neutrinos in one generation whose representation is given in \Eq{eq:nuR-rep}:
\\
$\bullet$ In the standard GSW SM, we have $N_{\nu_R}=0$.\\
$\bullet$ In the SU(5) GUT, it is common to have $N_{\nu_R}=0$.\\
$\bullet$ In the SO(10) GUT, we have $N_{\nu_R}=1$.\\[2mm]
We used to believe that there are no perturbative local anomalies for an additional U(1) if this U(1)
is the ${\rm U}(1)_{{ \mathbf{B}- \mathbf{L}}}$ or ${\rm U}(1)_{X}$. Let us check explicitly in the next subsections.
We should pay attention on the anomaly cancellation and its dependence on $N_{\nu_R}=0$ or $1$.
What we will check is the conservation of the ${\rm U}(1)_{{ \mathbf{B}- \mathbf{L}}}$ current,
\begin{eqnarray}
\hspace{1pt}\mathrm{d} \star (j_{\bfB}- j_{\mathbf{L}})= \prt_\mu (j_{\bfB}^\mu - j_{\mathbf{L}}^\mu) \hspace{1pt}\mathrm{d}^d x,
\end{eqnarray}
by taking into account all possible anomaly contributions from cobordism considerations.
\subsubsection{($\bf{B}-\bf{L}$)-U(1)$_Y^2$: 4d local anomaly}
Plug in data from Table \ref{table:SMfermion} to check the 4d local anomaly of $(\bfB-\mathbf{L})$-U(1)$_Y^2$:
\begin{eqnarray} \label{eq:anomaly-BLU1YU1Y}
\includegraphics[width=2.2in]{anomaly-BLU1YU1Y-.pdf},
\end{eqnarray}
\noindent
we find the anomaly factor contributes to \footnote{The minus sign can be interpreted either from the anti-quark of the right-handed $R$-chiral fermion (anti-chiral fermion).}
\begin{eqnarray}
j_{\bfB} &:& N_{\text{generation}} \cdot ({N_c}/{3}) \cdot \bigg(
2 \cdot (1/6)^2 - (-2/3)^2 - (1/3)^2\bigg)
= N_{\text{generation}} \cdot ({N_c}/{3}) \cdot (1/2),\cr
j_{\mathbf{L}}&:& N_{\text{generation}} \cdot \bigg(
2 \cdot (-1/2)^2 + 1^2\bigg)
= N_{\text{generation}} \cdot (1/2),
\end{eqnarray}
such that $\hspace{1pt}\mathrm{d} \star j_{\bfB} \neq 0 $ and $\hspace{1pt}\mathrm{d} \star j_{\mathbf{L}} \neq 0 $,
but $\hspace{1pt}\mathrm{d} \star (j_{\bfB}- j_{\mathbf{L}}) =0$ if $N_c=3$ as the color number it is.
%
\subsubsection{($\bf{B}-\bf{L}$)-SU(2)$^2$: 4d local anomaly}
Plug in data from Table \ref{table:SMfermion} to check the 4d local anomaly of $(\bfB-\mathbf{L})$-SU(2)$^2$:
\begin{eqnarray} \label{eq:anomaly-BLSU2SU2}
\includegraphics[width=2.2in]{anomaly-BLSU2SU2-.pdf}
\end{eqnarray}
\noindent
we find the anomaly factor contributes to
\begin{eqnarray}
j_{\bfB} &:& N_{\text{generation}} \cdot ({N_c}/{3}) \cdot \bigg(
{\mathrm{Tr}}[\frac{\sigma^a}{2} \frac{\sigma^b}{2}] \bigg)
= N_{\text{generation}} \cdot ({N_c}/{3}) \cdot (\delta_{ab} /2), \cr
j_{\mathbf{L}}&:& N_{\text{generation}} \cdot \bigg(
{\mathrm{Tr}}[\frac{\sigma^a}{2} \frac{\sigma^b}{2}] \bigg)
= N_{\text{generation}} \cdot (\delta_{ab} /2) ,
\end{eqnarray}
such that $\hspace{1pt}\mathrm{d} \star j_{\bfB} \neq 0 $ and $\hspace{1pt}\mathrm{d} \star j_{\mathbf{L}} \neq 0 $,
but $\hspace{1pt}\mathrm{d} \star (j_{\bfB}- j_{\mathbf{L}}) =0$ if $N_c=3$ as the color number it is.
\subsubsection{($\bf{B}-\bf{L}$)-SU(3)$^2_c$: 4d local anomaly}
Plug in data from Table \ref{table:SMfermion} to check the 4d local anomaly of $(\bfB-\mathbf{L})$-SU(3)$^2_c$:
\begin{eqnarray} \label{eq:anomaly-BLSU3SU3}
\includegraphics[width=2.2in]{anomaly-BLSU3SU3-.pdf}
\end{eqnarray}
\noindent
we find the anomaly factor contributes to
\begin{eqnarray}
j_{\bfB} &:& N_{\text{generation}} \cdot \bigg(
(2 -1 -1) {\mathrm{Tr}}[\frac{\tau^a}{2} \frac{\tau^b}{2}] \bigg)
= N_{\text{generation}} \cdot 0 \cdot (\delta_{ab} /2) = 0, \cr
j_{\mathbf{L}}&:& N_{\text{generation}} \cdot 0
= 0,
\end{eqnarray}
such that $\hspace{1pt}\mathrm{d} \star j_{\bfB} =\hspace{1pt}\mathrm{d} \star j_{\mathbf{L}}=\hspace{1pt}\mathrm{d} \star (j_{\bfB}- j_{\mathbf{L}}) =0$.
\subsubsection{($\bf{B}-\bf{L}$)-(gravity)$^2$: 4d local anomaly}
\label{sec:BLgravity2}
Plug in data from Table \ref{table:SMfermion} to check the 4d local anomaly of $(\bfB-\mathbf{L})$-(gravity)$^2$:
\begin{eqnarray} \label{eq:anomaly-BLgravgrav}
\includegraphics[width=2.2in]{anomaly-BLgravgrav-.pdf}
\end{eqnarray}
\noindent
we find the anomaly factor contributes to
\begin{eqnarray}
j_{\bfB} &:& N_{\text{generation}} \cdot ({N_c}/{3}) \cdot \bigg(
2 - 1 - 1\bigg)
= 0. \\
j_{\mathbf{L}}&:& N_{\text{generation}} \cdot \bigg(
2 - 1 - N_{\nu_R}\bigg)
= N_{\text{generation}} \cdot (1- N_{\nu_R}) .
\end{eqnarray}
It turns out that $\hspace{1pt}\mathrm{d} \star j_{\bfB} =0$ but $\hspace{1pt}\mathrm{d} \star j_{\mathbf{L}}\neq 0$ unless $N_{\nu_R}=1$. Same for $\hspace{1pt}\mathrm{d} \star (j_{\bfB}- j_{\mathbf{L}}) =0$ only if
$N_{\nu_R}=1$. Perturbative anomaly seems to suggest one right-handed neutrino $N_{\nu_R}=1$ to
saturate the ($\bf{B}-\bf{L}$) current non-conservation.
Are there other ways to saturate this ABJ type anomaly?
When the ($\bf{B}-\bf{L}$) becomes a \emph{discrete} symmetry,
we will be able to resolve the anomaly matching issue with other novel possibilities in \Sec{sec:HiddenTopologicalSectors}.
\subsubsection{($\bf{B}-\bf{L}$)$^3$: 4d local anomaly}
\label{sec:BL3}
Plug in data from Table \ref{table:SMfermion} to check the 4d local anomaly of $(\bfB-\mathbf{L})^3$:
\begin{eqnarray} \label{eq:anomaly-BLBLBL}
\includegraphics[width=2.in]{anomaly-BLBLBL-.pdf}
\end{eqnarray}
\noindent
we find the anomaly factor contributes to
\begin{eqnarray}
j_{\bfB} &:& N_{\text{generation}} \cdot {N_c} \cdot(1/{3})^3 \cdot \bigg(
2 - 1 - 1\bigg)
= 0. \\
j_{\mathbf{L}}&:& N_{\text{generation}} \cdot(1)^3 \cdot \bigg(
2 - 1- N_{\nu_R} \bigg)
= N_{\text{generation}} \cdot (1- N_{\nu_R})
\end{eqnarray}
It turns out that $\hspace{1pt}\mathrm{d} \star j_{\bfB} =0$ but $\hspace{1pt}\mathrm{d} \star j_{\mathbf{L}}\neq 0$ unless $N_{\nu_R}=1$. Same for $\hspace{1pt}\mathrm{d} \star (j_{\bfB}- j_{\mathbf{L}}) =0$ only if
$N_{\nu_R}=1$. Perturbative anomaly seems to suggest one right-handed neutrino $N_{\nu_R}=1$ to
saturate the ($\bf{B}-\bf{L}$) current non-conservation.
Are there other ways to saturate this ABJ type anomaly?
When the ($\bf{B}-\bf{L}$) becomes a \emph{discrete} symmetry,
we will resolve the anomaly matching with other novel possibilities in \Sec{sec:HiddenTopologicalSectors}.
\subsubsection{$X$-SU(5)$^2$: 4d local anomaly}
Recall in \Sec{sec:2OverviewSMGUT}, the ${\rm U}(1)_{{ \mathbf{B}- \mathbf{L}}}$ is not a proper symmetry of SU(5) GUT.
The ``baryon minus lepton number symmetry'' of SU(5) GUT is ${\rm U}(1)_X$.
Plug in data from Table \ref{table:SMfermion} to check 4d local anomaly of $X$-SU(5)$^2$:
\begin{eqnarray} \label{eq:anomaly-XSU5SU5}
\includegraphics[width=2.2in]{anomaly-XSU5SU5-.pdf}
\end{eqnarray}
\noindent
we find the anomaly factor contributed from the representation ${\bf{R}}$ of fermions in SU(5)
as the anti-fundamental ${\bf{R}} = \bar{\bf 5}$
and anti-symmetric ${\bf{R}} = {10}$, from the 15 Weyl fermions
$\bar{\bf 5}\oplus {\bf 10}$ in one generation. Let us check the
$X$ current conservation or violation by ABJ type anomaly:
\begin{eqnarray} \label{eq:XSU(5)2}
\hspace{1pt}\mathrm{d} \star (j_{X})
\propto
\sum_{\bf{R}} X_{\bf{R}} \cdot {\mathrm{Tr}}_{\bf{R}}[ F_{{\rm SU}(5)} \wedge F_{{\rm SU}(5)}]
\propto
\sum_{\bf{R}} X_{\bf{R}} \cdot c_2({{\rm SU}(5)}).
\end{eqnarray}
Here $c_2({{\rm SU}(5)})$ is the second Chern class of SU(5),
which is also related to the 4d instanton number of SU(5) gauge bundle.
For $\bar{\bf 5}\oplus {\bf 10}$ with $N_{\text{generation}}$, from Table \ref{table:SMfermion},
we get the ${\rm U}(1)_X$ charges for
$$X_{\bar{\bf 5}}=-3, \quad X_{\bf 10}=1,$$
so\footnote{To evaluate the $c_2$ or the instanton number
in different representations, ${\bf{R}}_1$ and ${\bf{R}}_2$,
we use the fact that
\begin{eqnarray}
{\mathrm{Tr}}_{{\bf{R}}_1}[F \wedge F] / {\mathrm{Tr}}_{{\bf{R}}_2}[F \wedge F] = (C_2({{\bf{R}}_1})d({{\bf{R}}_1})) / (C_2({{\bf{R}}_2})d({{\bf{R}}_2})),
\end{eqnarray}
here
$C_2({\bf{R}})$ and $d({\bf{R}})$ are respectively the quadratic
Casimir and the dimension of an irreducible representation ${\bf{R}}$.
For the representation ${\bf{R}}$ of SU($N$), we have
\begin{eqnarray}
\begin{array}{ll l}
\hline
{\bf{R}} & d & C_2 \\
\hline
\text{Fundamental} & N & N^2-1\\
\text{Antisymmetric} & N(N-1)/2 & 2(N+1)(N-2)\\
\hline
\end{array}.
\end{eqnarray}
For SU(5) with $N=5$, we get
${\mathrm{Tr}}_{{\bf{10}}}[F \wedge F]=(N-2){\mathrm{Tr}}_{\bar{\bf{5}}}[F \wedge F]=3{\mathrm{Tr}}_{\bar{\bf{5}}}[F \wedge F]$.
\label{ft:SUNrep}
}
\begin{eqnarray}
\hspace{1pt}\mathrm{d} \star (j_{X})
\propto
N_{\text{generation}} \Big(
X_{\bar{\bf 5}}{\mathrm{Tr}}_{\bar{\bf{5}}}[F \wedge F]
+ X_{\bf 10}{\mathrm{Tr}}_{{\bf{10}}}[F \wedge F]
\Big)
=N_{\text{generation}} \cdot 0 =0
\end{eqnarray}
vanishes. We confirm that the ${\rm U}(1)_X$ symmetry is ABJ anomaly free at least perturbatively in SU(5) GUT.
\subsection{SM and GUT with extra discrete symmetries and a cobordism theory}
\label{sec:SMandGUTdiscrete}
In the subsection, we aim to digest better how robust are the anomalies from
\Sec{sec:BLgravity2} and \Sec{sec:BL3} that seems only to be matched with a right-handed neutrino (the 16th Weyl spinor) per generation.
These anomalies are not dynamical gauge anomalies if ($\bf{B}-\bf{L}$)
and $X$
are only global symmetries --- namely the theory is only suffered from 't Hooft anomaly which only results in nonlocal or non-onsite ($\bf{B}-\bf{L}$) symmetry.
However, the ($\bf{B}-\bf{L}$) and $X$ have to be gauged in the SO(10) GUT. If fact, the discrete
$\mathbb{Z}_{4,{X}} \subset {\rm U}(1)_{X}$ acts as the
$\mathbb{Z}_4$ center of Spin(10),
\begin{eqnarray}
\mathbb{Z}_{4,{X}} = Z({\rm Spin}(10)) \subset {\rm Spin}(10)
\end{eqnarray}
needs to be dynamically gauged in the SO(10) GUT.
This fact motivates \Refe{GarciaEtxebarriaMontero2018ajm1808.00009} to use the
$\Omega_5^{{\rm Spin} \times_{\mathbb{Z}_2} \mathbb{Z}_4}=\mathbb{Z}_{16}$ to argue the 16 chiral fermions for the 4d GUT in one generation.
This fact also motivates \Refe{WW2019fxh1910.14668, WanWangv2} to check the following cobordism groups $\mathrm{TP}_5(G)$
with $G \supset {{\rm Spin} \times_{\mathbb{Z}_2^F} \mathbb{Z}_{4,X}}$ summarized in Table \ref{table:SU3SU2U1-discrete}.
\begin{table}[H]
\centering
\hspace*{-18mm}
\begin{tabular}{c c c }
\hline
\multicolumn{3}{c}{
$\begin{array}{c}
\text{ Cobordism group $\mathrm{TP}_d(G)$ with}
\text{
$G_{\text{SM}_q}\equiv{({\rm SU}(3)\times {\rm SU}(2)\times {\rm U}(1))}/\mathbb{Z}_q$
and $q=1,2,3,6$}
\end{array}$
}\\
\hline
\hline
$d$d & classes & cobordism invariants\\
\hline
\hline\\[-2mm]
\multicolumn{3}{c}{$G={\rm Spin}\times_{\mathbb{Z}_2} \mathbb{Z}_4 \times {G_{\text{SM}_1}}$}\\[2mm]
\hline
5d & $\mathbb{Z}^5\times\mathbb{Z}_2\times\mathbb{Z}_4^2\times\mathbb{Z}_{16}$ &
$\begin{matrix}\mu(\text{PD}(c_1({\rm U}(1)))), \quad
\text{CS}_1^{{\rm U}(1)} c_1({\rm U}(1))^2, \quad
{\text{CS}_1^{{\rm U}(1)}c_2({\rm SU}(3)) \sim c_1({\rm U}(1))\text{CS}_3^{{\rm SU}(2)}},\\
{\text{CS}_1^{{\rm U}(1)}c_2({\rm SU}(3)) \sim c_1({\rm U}(1))\text{CS}_3^{{\rm SU}(3)}}, \quad
\frac{({\cal A}_{{\mathbb{Z}_2}})^2\text{CS}_3^{{\rm SU}(3)}+\text{CS}_5^{{\rm SU}(3)}}{2},
\\
({\cal A}_{{\mathbb{Z}_2}}) c_2({\rm SU}(3)),\quad
c_2({\rm SU}(2))\eta', \quad
c_1({\rm U}(1))^2\eta', \quad\eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}}))
\end{matrix}$\\
\hline
\hline\\[-2mm]
\multicolumn{3}{c}{$G={\rm Spin}\times_{\mathbb{Z}_2} \mathbb{Z}_4 \times {G_{\text{SM}_2}}$}\\[2mm]
\hline
5d & $\mathbb{Z}^5\times\mathbb{Z}_2^2\times\mathbb{Z}_4\times\mathbb{Z}_{16}$ &
$\begin{matrix}
\mu(\text{PD}(c_1({\rm U}(2)))),
\quad
{\text{CS}_1^{{\rm U}(2)}c_1({\rm U}(2))^2},
\quad
\frac{({\cal A}_{{\mathbb{Z}_2}})^2\text{CS}_3^{{\rm U}(2)}+\text{CS}_1^{{\rm U}(2)}c_2({\rm U}(2)) }{2}\sim
\frac{({\cal A}_{{\mathbb{Z}_2}})^2\text{CS}_3^{{\rm U}(2)}+c_1({\rm U}(2))\text{CS}_3^{{\rm U}(2)}}{2},\\
\text{CS}_1^{{\rm U}(2)}c_2({\rm SU}(3)) \sim c_1({\rm U}(2))\text{CS}_3^{{\rm SU}(3)},
\quad \frac{({\cal A}_{{\mathbb{Z}_2}})^2\text{CS}_3^{{\rm SU}(3)}+\text{CS}_5^{{\rm SU}(3)}}{2},
\\
({\cal A}_{{\mathbb{Z}_2}})c_2({\rm SU}(3)), \quad ({\cal A}_{{\mathbb{Z}_2}})c_2({\rm U}(2)), \quad c_1({\rm U}(2))^2\eta', \quad \eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}}))
\end{matrix}$ \\
\hline
\hline\\[-2mm]
\multicolumn{3}{c}{$G={\rm Spin}\times_{\mathbb{Z}_2} \mathbb{Z}_4 \times {G_{\text{SM}_3}}$}\\[2mm]
\hline
5d & $\mathbb{Z}^5\times\mathbb{Z}_2\times\mathbb{Z}_4^2\times\mathbb{Z}_{16}$ &
$\begin{matrix}\mu(\text{PD}(c_1({\rm U}(3)))),
\quad
{c_1({\rm U}(3))^2\text{CS}_1^{{\rm U}(3)},
\quad
\text{CS}_1^{{\rm U}(3)}c_2({\rm SU}(2)) \sim c_1({\rm U}(3))\text{CS}_3^{{\rm SU}(2)}},
\\
\frac{({\cal A}_{{\mathbb{Z}_2}})^2\text{CS}_3^{{\rm U}(3)}+\text{CS}_1^{{\rm U}(3)} c_2({\rm U}(3))+\text{CS}_5^{{\rm U}(3)}}{2}
\sim
\frac{({\cal A}_{{\mathbb{Z}_2}})^2\text{CS}_3^{{\rm U}(3)}+c_1({\rm U}(3))\text{CS}_3^{{\rm U}(3)}+\text{CS}_5^{{\rm U}(3)}}{2},
\quad
\text{CS}_5^{{\rm U}(3)},
\\
({\cal A}_{{\mathbb{Z}_2}})c_2({\rm U}(3)),
\quad
c_2({\rm SU}(2))\eta',
\quad
c_1({\rm U}(3))^2\eta',
\quad
\eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}}))
\end{matrix}$\\
\hline
\hline\\[-2mm]
\multicolumn{3}{c}{$G={\rm Spin}\times_{\mathbb{Z}_2} \mathbb{Z}_4 \times {G_{\text{SM}_6}}$}\\[2mm]
\hline
5d & $\mathbb{Z}^5\times\mathbb{Z}_2^2\times\mathbb{Z}_4\times\mathbb{Z}_{16}$ &
$\begin{matrix}\mu(\text{PD}(c_1({\rm U}(3)))),
\quad {c_1({\rm U}(3))^2\text{CS}_1^{{\rm U}(3)},
\quad
\frac{({\cal A}_{{\mathbb{Z}_2}})^2\text{CS}_3^{{\rm U}(2)}+\text{CS}_1^{{\rm U}(3)}c_2({\rm U}(2))}{2}
\sim
\frac{({\cal A}_{{\mathbb{Z}_2}})^2\text{CS}_3^{{\rm U}(2)}+c_1({\rm U}(3))\text{CS}_3^{{\rm U}(2)}}{2}},
\\
\frac{({\cal A}_{{\mathbb{Z}_2}})^2\text{CS}_3^{{\rm U}(3)}+\text{CS}_1^{{\rm U}(3)} c_2({\rm U}(3))+\text{CS}_5^{{\rm U}(3)}}{2}
{\sim
\frac{({\cal A}_{{\mathbb{Z}_2}})^2\text{CS}_3^{{\rm U}(3)}+c_1({\rm U}(3))\text{CS}_3^{{\rm U}(3)}+\text{CS}_5^{{\rm U}(3)}}{2}},
\quad
\text{CS}_5^{{\rm U}(3)},\\
({\cal A}_{{\mathbb{Z}_2}})c_2({\rm U}(3)),
\quad
({\cal A}_{{\mathbb{Z}_2}})c_2({\rm U}(2)),
\quad
c_1({\rm U}(3))^2\eta',
\quad
\eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}}))
\end{matrix}$ \\
\hline
\hline\\[-2mm]
\multicolumn{3}{c}{$G={\rm Spin}\times_{\mathbb{Z}_2} \mathbb{Z}_4 \times {\rm SU}(5)$}\\[2mm]
\hline
5d & $\mathbb{Z}\times\mathbb{Z}_2\times\mathbb{Z}_{16}$ &
$\frac{({\cal A}_{{\mathbb{Z}_2}})^2\text{CS}_3^{{\rm SU}(3)}+\text{CS}_5^{{\rm SU}(3)}}{2}$, \quad
$({\cal A}_{{\mathbb{Z}_2}}) c_2({\rm SU}(5))$,\quad
$\eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}}))$\\
\hline
\hline
\end{tabular}
\caption{Our setup follows Table \ref{table:SU3SU2U1} and \Refe{WW2019fxh1910.14668, WanWangv2}.
The ${\cal A}_{{\mathbb{Z}_2}} \in \operatorname{H}^1(M,\mathbb{Z}_2)$ is the generator from $\operatorname{H}^1(\mathrm{B}(\mathbb{Z}_4/\mathbb{Z}_2^F),\mathbb{Z}_2)$ of ${{\rm Spin} \times_{\mathbb{Z}_2^F} \mathbb{Z}_4}$.
The
$\eta'$ is a $\mathbb{Z}_4$ valued 1d eta invariant which is the extension of a quotient ${\cal A}_{{\mathbb{Z}_2}}$ by the normal 1d $\tilde\eta$.
So $({\cal A}_{{\mathbb{Z}_2}})\equiv ({\cal A}_{{\mathbb{Z}_4}}) \mod 2$ is the quotient, while $\mathbb{Z}_{4,{X}} \subset {\rm U}(1)_{X}$.
The $\eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}}))$ is
the value of ${\eta} \in \mathbb{Z}_{16}$ on the Poincar\'e dual (PD) submanifold of ${\cal A}_{{\mathbb{Z}_2}}$.}
\label{table:SU3SU2U1-discrete}
\end{table}
{We aim to initiate a new approach on matching the nonperturbative global $\mathbb{Z}_{16}$ anomaly
(descended from the perturbative local $\mathbb{Z}$ anomalies of
\Sec{sec:BLgravity2} and \Sec{sec:BL3}) for the missing neutrinos.}
The $\mathbb{Z}^5$ classes perturbative local anomalies are the same results parallel to \Sec{sec:localanomaliesFeynman} and Table \ref{table:SU3SU2U1}.
So in the following subsections, we only focus on checking global anomaly cancellations of $\mathbb{Z}_n$ classes.
Generically Feynman diagrams cannot characterize global anomalies (so we do not present Feynman diagrams below).
But we can characterize global anomalies by generic curved manifolds as in \Fig{fig:Cobordism2006}
with gauge, gravity, or mixed gauge-gravity background fields.
\subsubsection{Witten anomaly $c_2({\rm SU}(2))\tilde\eta$ vs $c_2({\rm SU}(2))\eta'$: 4d $\mathbb{Z}_{2}$ vs $\mathbb{Z}_{4}$ global anomalies}
Follow \Sec{sec:Wittenanomaly}, {the old SU(2) anomaly of Witten in 4d} is a mod 2 class. It is a $\mathbb{Z}_{2}$ global anomaly given
by 5d $c_2({\rm SU}(2))\tilde\eta $ and 6d $c_2({\rm SU}(2))\text{Arf}$ in \cite{WW2019fxh1910.14668} and Table \ref{table:SU3SU2U1-discrete}.
The $c_2({\rm SU}(2))\tilde\eta$ as Witten SU(2) anomaly
can be contributed by the 4d fermion doublet {\bf 2} under the SU(2) representation (or
the (iso-)spin-$2r +1/2$ of SU(2) Weyl spinor for some non-negative integer $r \in \mathbb{Z}_{\geq 0}$).
So to check the anomaly cancellation,
we count the fermion doublet {\bf 2} under SU(2).
There are four of {\bf 2} of SU(2) fermions from
$({\bf 3},{\bf 2}, \tilde{Y}=1)_L$ and $({\bf 1},{\bf 1},\tilde{Y}=6)_L$.
The anomaly from $c_2({\rm SU}(2))\tilde \eta$ counts the number of 4d Weyl spinors
of SU(2) fundamental ${\bf 2}$ mod 2.
From \eq{eq:Weyl-rep}
for $N_{\text{generation}}$, we have:
\begin{eqnarray}
N_{\text{generation}} \cdot (3 + 1) = 0 \mod 2.
\end{eqnarray}
There is also an extended $\mathbb{Z}_{4}$ global anomaly from the 5d cobordism invariant
$c_2({\rm SU}(2))\eta'$ in \cite{WW2019fxh1910.14668} and Table \ref{table:SU3SU2U1-discrete} counting the number of 4d Weyl spinors
of SU(2) fundamental ${\bf 2}$ mod 4.\footnote{As explained in the Notation in the end of \Sec{sec:Intro}:
The ${\eta}'$ is a mod 4 index of 1d Dirac operator as a cobordism invariant of $\Omega_1^{{\rm Spin}\times \mathbb{Z}_4}=\mathbb{Z}_4$.}
From \eq{eq:Weyl-rep}
for $N_{\text{generation}}$, we have:
\begin{eqnarray}
N_{\text{generation}} \cdot (3 + 1) = 0 \mod 4.
\end{eqnarray}
Therefore, we have checked no global SU(2) anomaly of $\mathbb{Z}_{2}$ or $\mathbb{Z}_{4}$ classes for SM with a Spin or ${{\rm Spin} \times_{\mathbb{Z}_2^F} \mathbb{Z}_4}$ structure.
\subsubsection{$({\cal A}_{{\mathbb{Z}_2}}) c_2({\rm SU}(2))$: 4d $\mathbb{Z}_{2}$ global anomaly}
\noindent
The 4d $\mathbb{Z}_{2}$ global anomaly from the 5d cobordism invariant $({\cal A}_{{\mathbb{Z}_2}}) c_2({\rm SU}(2))$,
in \cite{WW2019fxh1910.14668} and Table \ref{table:SU3SU2U1-discrete},
counts the number of 4d left-handed Weyl spinors of SU(2) fundamental ${\bf 2}$ mod 2.
Here $({\cal A}_{{\mathbb{Z}_2}})\equiv ({\cal A}_{{\mathbb{Z}_4}}) \mod 2$, where
${\cal A}_{{\mathbb{Z}_2}} \in \operatorname{H}^1(M,\mathbb{Z}_2)$ is the generator from $\operatorname{H}^1(\mathrm{B}(\mathbb{Z}_4/\mathbb{Z}_2^F),\mathbb{Z}_2)$ of ${{\rm Spin} \times_{\mathbb{Z}_2^F} \mathbb{Z}_4}$.
So $({\cal A}_{{\mathbb{Z}_2}})\equiv ({\cal A}_{{\mathbb{Z}_4}}) \mod 2$ is the quotient, while $\mathbb{Z}_{4,{X}} \subset {\rm U}(1)_{X}$.
The $({\cal A}_{{\mathbb{Z}_2}}) c_2({\rm SU}(2))$ counts the number of 4d Weyl spinors
of SU(2) fundamental ${\bf 2}$ mod 2.
From \eq{eq:Weyl-rep}
for $N_{\text{generation}}$, we have:
\begin{eqnarray}
N_{\text{generation}} \cdot (3 + 1) = 0 \mod 2.
\end{eqnarray}
Thus the anomaly vanishes.
We have no obstruction to gauge the $\mathbb{Z}_4$ by making
${\cal A}_{{\mathbb{Z}_4}}$ dynamical at least from this anomaly cancellation.
\subsubsection{$({\cal A}_{{\mathbb{Z}_2}}) c_2({\rm SU}(3))$: 4d $\mathbb{Z}_{2}$ global anomaly}
\noindent
The 4d $\mathbb{Z}_{2}$ global anomaly from the 5d cobordism invariant $({\cal A}_{{\mathbb{Z}_2}}) c_2({\rm SU}(3))$,
in \cite{WW2019fxh1910.14668} and Table \ref{table:SU3SU2U1-discrete},
counts the number of 4d left-handed Weyl spinors of SU(3) fundamental ${\bf 3}$ mod 2.
From \eq{eq:Weyl-rep},
we count $({\bf 3},{\bf 2}, \tilde{Y}=1)_L$,
$(\overline{\bf 3},{\bf 1}, \tilde{Y}=-4)_L$, and
$(\overline{\bf 3},{\bf 1},\tilde{Y}=2)_L$
with {3 generations}.
For $N_{\text{generation}}$, we have:
\begin{eqnarray}
N_{\text{generation}} \cdot (2 - 1 -1) = 0 \mod 4.
\end{eqnarray}
Thus there is no anomaly. We have no obstruction to gauge the $\mathbb{Z}_4$ by making
${\cal A}_{{\mathbb{Z}_4}}$ dynamical at least from this anomaly cancellation.
\subsubsection{$c_1({\rm U}(1))^2\eta'$: 4d $\mathbb{Z}_{4}$ global anomaly}
The 4d $\mathbb{Z}_{4}$ global anomaly from a 5d cobordism invariant $c_1({\rm U}(1))^2\eta'$ (in \Refe{WW2019fxh1910.14668} and Table \ref{table:SU3SU2U1-discrete})
counts the sum of 4d Weyl spinors' (U(1) charge)$^2$ mod 4.
Let us apply $\tilde{Y}$ for U(1)$_{\tilde{Y}}$ charge
from \eq{eq:Weyl-rep}
with
$N_{\text{generation}}$.
For each generation, we not only have the sum of U(1) charge:
\begin{eqnarray}
3 \cdot 2 \cdot 1+ 3 \cdot 1 \cdot (-4) + 3 \cdot 1 \cdot 2 + 1 \cdot 2 \cdot (-3) + 1 \cdot 1 \cdot 6 =0 \mod 4.
\end{eqnarray}
but also have the sum of
(U(1) charge)$^2$:
\begin{eqnarray}
3 \cdot 2 \cdot 1^2+ 3 \cdot 1 \cdot (-4)^2 + 3 \cdot 1 \cdot 2^2 + 1 \cdot 2 \cdot (-3)^2 + 1 \cdot 2 \cdot (6)^2
= 3 \cdot 4 \cdot 13 =0 \mod 4.
\end{eqnarray}
The $c_1({\rm U}(1))^2$ also counts the U(1) instanton number up to a proportional factor.
Thus there is no anomaly. We have no obstruction to gauge the $\mathbb{Z}_4$ by making the
${\cal A}_{{\mathbb{Z}_4}}$ dynamical at least from this anomaly cancellation.
\subsubsection{$({\cal A}_{{\mathbb{Z}_2}}) c_2({\rm SU}(5))$: 4d $\mathbb{Z}_{2}$ global anomaly}
Similar to \Eq{eq:XSU(5)2}, we consider the discrete 4d $\mathbb{Z}_2$ global anomaly from the
5d cobordism invariant $({\cal A}_{{\mathbb{Z}_2}}) c_2({\rm SU}(5))$ in \cite{WW2019fxh1910.14668, WanWangv2} and Table \ref{table:SU3SU2U1-discrete}. Here $c_2({{\rm SU}(5)})$ is the second Chern class of SU(5),
which is also related to the 4d instanton number of SU(5) gauge bundle.
For $\bar{\bf 5}\oplus {\bf 10}$ with $N_{\text{generation}}$, from Table \ref{table:SMfermion},
we get the $\mathbb{Z}_{4,X}$ charges for
$$X_{\bar{\bf 5}}=-3 = 1 \mod 4, \quad X_{\bf 10}=1 \mod 4.$$
By footnote \ref{ft:SUNrep}, we compute the anomaly factor
\begin{eqnarray}
&&N_{\text{generation}} \Big(
(X_{\bar{\bf 5}} \mod 4) {\mathrm{Tr}}_{\bar{\bf{5}}}[F \wedge F]
+ (X_{ {\bf 10}} \mod 4){\mathrm{Tr}}_{{\bf{10}}}[F \wedge F]
\Big) \cr
&&=N_{\text{generation}} \cdot \Big( 1 \cdot 1 + 1 \cdot 3 \Big) =N_{\text{generation}} \cdot 4 =0 \mod 4.
\end{eqnarray}
This certainly vanishes for the mod 2 anomaly for SU(5) GUT.
There is also another $\mathbb{Z}$ class local anomalies for SU(5) GUT captured by 5d cobordism invariants
$\frac{({\cal A}_{{\mathbb{Z}_2}})^2\text{CS}_3^{{\rm SU}(3)}+\text{CS}_5^{{\rm SU}(3)}}{2}$, we can easily check that SU(5) GUT is free from
any local anomaly given by another 5d cobordism invariant \cite{JW2008.06499}.
\subsubsection{$\eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}}))$: 4d $\mathbb{Z}_{16}$ global anomaly}
\label{sec:etaPDAZ16}
The 4d $\mathbb{Z}_{16}$ global anomaly given by a 5d cobordism invariant $\eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}}))$, in \cite{WW2019fxh1910.14668} and Table \ref{table:SU3SU2U1-discrete},
counts the number mod 16 of 4d left-handed Weyl spinors
($\Psi_L \sim {\bf 2}_L \text{ of } {\rm Spin}(3,1)$ {or} $\Psi_L \sim {\bf 2}_L \text{ of } {\rm Spin}(4) = {\rm SU}(2)_L \times {\rm SU}(2)_R$).
From \eq{eq:Weyl-rep} with $N_{\text{generation}}$ (e.g., {3 generations}),
for each generation, we have:
\begin{eqnarray}
3 \cdot 2 + 3 \cdot 1 + 3 \cdot 1 + 1 \cdot 2 + 1 \cdot 1 = 15 = -1 \mod 16.
\end{eqnarray}
For 1 generation, we need to saturates the anomaly:
\begin{eqnarray} \label{eq:nu-1}
\upnu =-1 \mod 16.
\end{eqnarray}
For 3 generations, we need
\begin{eqnarray}
3 \Bigg( 3 \cdot 2 + 3 \cdot 1 + 3 \cdot 1 + 1 \cdot 2 + 1 \cdot 1 \Bigg)= 45 = -3 \mod 16.
\end{eqnarray}
Therefore we need to saturates the anomaly:
\begin{eqnarray}
\upnu =-3 \mod 16.
\end{eqnarray}
For $N_{\text{generation}}$ generations, we need to saturates the anomaly:
\begin{eqnarray} \label{eq:Ngeneration}
\upnu =-N_{\text{generation}} \mod 16.
\end{eqnarray}
This anomaly can be canceled by adding new degrees of freedom
\begin{eqnarray} \label{eq:NgenerationNnuR}
\upnu =N_{\text{generation}}\cdot (N_{\nu_R}=1) \mod 16.
\end{eqnarray}
The anomaly matching in this \Sec{sec:etaPDAZ16} seems to be matched with a right-handed neutrino (the 16th Weyl spinor) per generation, similar to
\Sec{sec:BLgravity2} and \Sec{sec:BL3}.
This also shows the robustness of \Sec{sec:BLgravity2} and \Sec{sec:BL3} even if we break down ${\rm U}(1)_{\bf{B}-\bf{L}}$ or ${\rm U}(1)_X$
down to $\mathbb{Z}_{4,\bf{B}-\bf{L}}$ or to $\mathbb{Z}_{4,X}$. Again this $\mathbb{Z}_4$ as the center $Z({\rm Spin}(10))$ of Spin(10) is important for the SO(10) GUT.
Are there other ways to match the anomaly other than introducing the right-handed neutrino (the 16th Weyl spinor) per generation?
Let us explore the alternatives in the next subsection.
\subsection{How to match the anomaly? Preserving or breaking the $\mathbb{Z}_{4,X}$ symmetry?}
\label{sec:matchtheanomaly}
Let us summarize what we learn from the anomaly computation and matching in the previous sections.
We have shown that all anomalies presented in Table \ref{table:SU3SU2U1} and Table \ref{table:SU3SU2U1-discrete} can be cancelled,
except the additional anomaly constraint from \Sec{sec:BLgravity2}, \Sec{sec:BL3}, and \Sec{sec:etaPDAZ16} may not be matched unless we obey
\begin{eqnarray}\label{eq:localglobal16}
\begin{array}{lll}
N_{\text{generation}} \cdot (1- N_{\nu_R} + \text{ hidden sector }) &=& 0, \quad\quad\quad\quad\;\text{ from local anomalies of \Sec{sec:BLgravity2} and \ref{sec:BL3}}.\quad\\
N_{\text{generation}} \cdot (1- N_{\nu_R} + \text{ hidden sector }) &=& 0 \mod 16, \text{ from a global anomaly of \Sec{sec:etaPDAZ16}.}
\end{array}
\end{eqnarray}
So the above n\"aively suggests we need the right-handed neutrino (the 16th Weyl spinor) $N_{\nu_R}=1$ per generation.
However, by demanding only a discrete $\mathbb{Z}_{4,X}$ instead of a continuous ${\rm U}(1)_{X}$ symmetry,
the local $\mathbb{Z}$ class anomaly becomes a global $\mathbb{Z}_{16}$ class anomaly. If so,
we do have different ways to match the anomaly, other than introducing the right-handed neutrino (the 16th Weyl spinor) $N_{\nu_R}=1$.
Can we match the anomaly by {\bf additional new
hidden sectors} not yet discovered in SM or in SU(5) Georgi-Glashow GUT?
Let us focus on the $\mathbb{Z}_{4,X}$ symmetry
for the sake of thinking $\mathbb{Z}_{4,X}=Z({\rm Spin}(10))$ is gauged in the SO(10) GUT eventually.
Let us enumerate the possibilities to match the anomaly \Eq{eq:localglobal16}:
\begin{enumerate} [label=\textcolor{blue}{\arabic*}., ref={\arabic*}]
\item {\bf Anomaly matched by a right-handed neutrino (the 16th Weyl spinor)} $N_{\nu_R}=1$:\\
For the right-handed neutrino $\nu_R$ to be massless (or gapless) while preserving the
$\mathbb{Z}_{4,X}$ symmetry, we need to have $\nu_R$ to be a complex Weyl spinor (with a $\mathbb{Z}_{4,X}$ charge $-1 \mod 4$), instead of a real Majorana spinor, in order to have the
$\mathbb{Z}_{4,X}$ symmetry transformation manifest:
\begin{eqnarray} \label{eq:nuRsym}
\nu_R \to \exp(-2 \pi \hspace{1pt}\mathrm{i}\hspace{1pt}/4) \; \nu_R = (-\hspace{1pt}\mathrm{i}\hspace{1pt}) \; \nu_R.
\end{eqnarray}
Since this is a sterile neutrino with a trivial representation of SM gauge group \Eq{eq:nuR-rep}, $({\bf 1},{\bf 1},0)_R$, we can rotate it to left-handed Weyl spinor
$\bar{\nu}_R= {\nu}_L$ with $({\bf 1},{\bf 1},0)_L$ and flips the $\mathbb{Z}_{4,X}$ representation to its complex conjugation:
\begin{eqnarray} \label{eq:nuLsym}
\nu_L \to \exp(2 \pi \hspace{1pt}\mathrm{i}\hspace{1pt}/4)\;\nu_L = (\hspace{1pt}\mathrm{i}\hspace{1pt}) \;\nu_L.
\end{eqnarray}
What can the low energy dynamics of the ${\nu}_R$ be?
\begin{enumerate} [label=\textcolor{blue}{(\arabic*)}., ref={(\arabic*)}]
\item {\bf The $\mathbb{Z}_{4,X}$ preserving massless neutrino}: If the sterile neutrino ${\nu}_R$ remains massless, we can match the anomaly \Eq{eq:localglobal16}
by a symmetric gapless low energy theory with an action on a 4d spacetime $M^4$:
\begin{eqnarray}
\int_{M^4} \bar\nu_R (\hspace{1pt}\mathrm{i}\hspace{1pt} \sigma^\mu \prt_\mu)\nu_R, \quad \text{ or equivalently }
\quad
\int_{M^4} \bar\nu_L (\hspace{1pt}\mathrm{i}\hspace{1pt} \bar \sigma^\mu \prt_\mu)\nu_L,
\end{eqnarray}
with $\bar\sigma \equiv (1, \vec{\sigma})$. Note that $\nu_L$ has a $\mathbb{Z}_{4,X}$ charge 1 and
$\bar\nu_L$ has a $\mathbb{Z}_{4,X}$ charge $-1$, so the action preserves the $\mathbb{Z}_{4,X}$ symmetry.
\item {\bf The $\mathbb{Z}_{4,X}$ preserving Dirac mass term, but the $\mathbb{Z}_{4,X}$ is spontaneously broken by the electroweak Higgs}: The Yukawa-Higgs-Dirac term preserves the $\mathbb{Z}_{4,X}$ symmetry:
\begin{eqnarray}
\int_{M^4} \bar\nu_R \phi_H \nu_L + \bar\nu_L \phi_H \nu_R,
\end{eqnarray}
because $\nu_L$ (or $\bar\nu_R$) has a $\mathbb{Z}_{4,X}$ charge 1,
$\nu_R$ (or $\bar\nu_L$) has a $\mathbb{Z}_{4,X}$ charge $-1$, and $\phi_H$ has a $\mathbb{Z}_{4,X}$ charge 2.
However, when the electroweak symmetry breaking sets in (at a lower energy around 246 G$e$V),
the Higgs condenses, $\langle \phi_H \rangle \neq 0$, and spontaneously breaks the $\mathbb{Z}_{4,X}$ symmetry.
\item {\bf The $\mathbb{Z}_{4,X}$ explicit breaking Majorana mass term}:
If the sterile neutrino ${\nu}_R$ becomes gapped by {Majorana mass} $m_{\text{Maj}}$, the spacetime spinor is in a real Majorana representation.
Then, the $\mathbb{Z}_{4,X}$ symmetry in \Eq{eq:nuRsym} and \Eq{eq:nuLsym} is explicitly broken.
We can still match the anomaly \Eq{eq:localglobal16}
by a $\mathbb{Z}_{4,X}$-symmetry-breaking gapped theory with an action on a 4d spacetime $M^4$:
\begin{eqnarray}
\int_{M^4=\prt M^5} \chi^{\rm T} (\hspace{1pt}\mathrm{i}\hspace{1pt} \bar\sigma^\mu \nabla_\mu)\chi
+\frac{\hspace{1pt}\mathrm{i}\hspace{1pt} m_{\text{Maj}} }{2}( \chi^{\rm T} \sigma^2 \chi+ \chi^\dagger \sigma^2 \chi^* ),
\end{eqnarray}
where we have written the 4-component Majorana spinor as $\Psi_{\text{Maj}}=
\begin{pmatrix}
\chi\\
\hspace{1pt}\mathrm{i}\hspace{1pt} \sigma^2 \chi^*
\end{pmatrix}$ with the transpose T, complex conjugate $*$, and complex conjugate transpose $\dagger$.
\end{enumerate}
See more discussions on Dirac or Majorana masses for three generations of SM fermions in \Sec{sec:MajoranaMassorDiracMass}.
\item {\bf Anomaly matched by new additional or
hidden sectors beyond SM}: Let us hypothesize many scenarios with different low energy dynamics following
\Refe{WW2019fxh1910.14668}'s Sec.~8.2:
\begin{enumerate} [label=\textcolor{blue}{(\arabic*)}., ref={(\arabic*)}]
\item \label{Z4XCFT}
$\mathbb{Z}_{4,X}$-symmetry-preserving anomalous gapless, free or interacting, 4d CFT.
\item \label{Z4XbreakCFT}
$\mathbb{Z}_{4,X}$-symmetry-breaking gapless, free or interacting, 4d CFT.
\item \label{Z4XTQFT}
$\mathbb{Z}_{4,X}$-symmetry-preserving anomalous \emph{long-range entangled} gapped 4d TQFT.
\item \label{Z4XbreakTQFT}
$\mathbb{Z}_{4,X}$-symmetry-breaking gapped 4d TQFT.
\item \label{Z45dSPT}
$\mathbb{Z}_{4,X}$-Symmetry-Protected Topological state (SPTs) in 5d
as a \emph{short-range entangled} gapped phase
captured by a 5d cobordism invariant
\begin{eqnarray}
\exp\bigg(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16} \cdot(-N_{\text{generation}}) \cdot \eta\big(\text{PD}({\cal A}_{{\mathbb{Z}_2}}) \big) \bigg).
\end{eqnarray}
\item \label{Z45dSET}
$\mathbb{Z}_{4,X}$-gauged-(Symmetry)-Enriched Topological state (SETs) in 5d coupled to gravity.
\end{enumerate}
\end{enumerate}
All the above theories are unitary by their own.
We propose that all the above scenarios and their proper \emph{linear combinations},
conventional or exotic, if existing, can saturate the anomaly \Eq{eq:localglobal16}.
Based on the contemporary knowledge of SM physics and experimental hints,\\[2mm]
$\bullet$ Scenario \ref{Z4XCFT} and \ref{Z4XbreakCFT} seem less practical, because
it is less likely to have any new gapless or interacting CFT that we do not observe below the T$e$V energy scale.\footnote{However,
if the hidden new gapless sectors are fully decoupled from SM forces,
they may be weakly interacting massless or gapless matter
account for the Dark Matter sector.}
Also if there is spontaneous symmetry-breaking (SSB),
for ${\rm U}(1)_{X}$ SSB, we expect to observe new Goldstone boson modes; for $\mathbb{Z}_{4,X}$ SSB, we may observe different vacua or domain walls between different vacua.
This shall be falsifiable in the experiments. \\[2mm]
$\bullet$
Scenario \ref{Z4XTQFT} is exotic but very interesting, which we discover new insights into the neutrino physics and Dark Matter.
Since the $\mathbb{Z}_{4,X}$-symmetry is a discrete chiral symmetry,
and the gapped 4d TQFT can be regarded as a gapped phase.
The anomalous $\mathbb{Z}_{4,X}$-symmetry preserving gapped 4d TQFT can be regarded
as \emph{a gapped phase without $\mathbb{Z}_{4,X}$-chiral symmetry breaking}.
However, the 4d TQFT has additional generalized global symmetries \cite{Gaiotto2014kfa1412.5148},
known as the higher-symmetries, whose charged objects are extended objects (1d lines, or 2d surfaces, etc.).
\begin{itemize}[leftmargin=1.8mm]
\item[-] The higher-symmetries of the 4d TQFT can be spontaneously broken. In this sense,
the spontaneously symmetry breaking (SSB) of higher-symmetries imply that the 4d TQFT is \emph{deconfined}.
Below the topological order energy gap, we have the low energy 4d TQFT.
Above the topological order energy gap,
the 4d TQFT hosts \emph{deconfined} excitations of fractionalized particles (from breaking 1d worldline with open 0d ends) or
fractionalized strings (from breaking 2d worldsheet with open 1d boundaries).
\item[-] If the higher-symmetries of the 4d TQFT have no SSB, then this leads to a \emph{gapped confined phase without $\mathbb{Z}_{4,X}$-chiral symmetry breaking}.
However, eventually in a quantum gravity theory,
all the global symmetries must be either \emph{gauged} or \emph{broken}.
Thus we should not have the option of SSB or symmetry-preserving.
Preferably, in a quantum gravity theory,
we can have either of the following options:\\
(1) the higher-symmetries of TQFT are dynamically \emph{gauged} due to a higher energy theory (above GUT),\\
(2) the higher-symmetries of TQFT are \emph{explicitly} broken.
\end{itemize}
Moreover, Scenario \ref{Z4XTQFT} can give rise to Scenario \ref{Z4XbreakTQFT}, if we construct such an anomalous $\mathbb{Z}_{4,X}$-symmetry-preserving TQFT first,
we can break some of the symmetry to obtain the symmetry-breaking gapped TQFT.
We can also construct a mixed symmetry-extension and symmetry-breaking TQFT following \Refe{Wang2017locWWW1705.06728}.
So we should focus on Scenario \ref{Z4XTQFT} explored in \Sec{sec:Hidden4dnon-invertiblenon-abelianTQFT}. \\[2mm]
$\bullet$
Scenario \ref{Z45dSPT} implies that
our 4d SM lives on the boundary of
5d $\mathbb{Z}_{4,X}$-SPTs given by 5d $\eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}}))$.
If Scenario \ref{Z45dSPT} describes our universe, we discover at least an extra dimension from the 5d theory.
We explore this 5d SPT or invertible TQFT theory in \Sec{sec:5dSPT}. \\[2mm]
$\bullet$
Scenario \ref{Z45dSET}
Above certain higher energy scale, the $\mathbb{Z}_{4,X}$ may be dynamically gauged such as in $\mathbb{Z}_{4,X}=Z({\rm Spin}(10))$ of the SO(10) GUT, then
the 4d and 5d bulk are fully gauged and entangled together. The 5d bulk is in fact
a 5d Symmetry-Enriched Topologically ordered state (SETs) that can be coupled to dynamical gravity.
We explore this 5d SETs coupled to dynamical gravity theory in \Sec{sec:gravity}.
\section{Beyond Known ``Fundamental'' Forces: Hidden Topological Force}
\label{sec:HiddenTopologicalSectors}
Follow \Sec{sec:matchtheanomaly}, now we propose new Scenarios, beyond SM and beyond Georgi-Glashow (GG) SU(5) GUT,
to match the $\mathbb{Z}_{16}$ global anomalies
from \Sec{sec:etaPDAZ16}.
We focus on the
Scenario \ref{Z4XTQFT} (thus also Scenario \ref{Z4XbreakTQFT})
for a hidden 4d non-invertible TQFT
in \Sec{sec:Hidden4dnon-invertiblenon-abelianTQFT},
and
for a 5d SPTs or 5d SETs coupled to gravity in Scenario \ref{Z45dSPT} and \ref{Z45dSET} in \Sec{sec:gravity}.
\subsection{Hidden 5d invertible TQFT or 5d Symmetry-Protected Topological state (SPTs)}
\label{sec:5dSPT}
We can saturate the missing anomaly of $\upnu =- N_{\text{generation}}\mod 16$
in
\Eq{eq:Ngeneration} by a 5d invertible TQFT (or a 5d SPTs protected by
$G={\rm Spin}\times_{\mathbb{Z}_2} \mathbb{Z}_4 \times {G_{\text{internal}}}$-symmetry
in a condensed matter language) with a 5d partition function:
\begin{eqnarray}
{\bf Z}_{\text{5d-iTQFT}}=\exp\bigg(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16} \cdot(-N_{\text{generation}}) \cdot \eta\big(\text{PD}({\cal A}_{{\mathbb{Z}_2}}) \big) \bigg\rvert_{M^5}\bigg).
\end{eqnarray}
Here we define the eta invariant
$$\eta \in \mathbb{Z}_{16}, \quad \text{ and }\quad \eta\big(\text{PD}({\cal A}_{{\mathbb{Z}_2}}) \big) \in \mathbb{Z}_{16}$$
slightly different by a proportional factor in the math literature.
Let us overview quickly what we need about the
{APS eta invariant $\upeta$} \cite{Atiyah1975jfAPS, Atiyah1976APS, Atiyah1980APS, gilkey1985eta, Witten:2015aba}.
Let us consider the 4d $\upeta$ invariant and then relate to our 4d $\eta$\footnote{The
4d $\eta$ invariant has condensed matter realizations. It is known as the
class DIII topological superconductor [TSC] in the free fermion limit with a $\mathbb{Z}$ classification \cite{RyuSPT, Kitaevperiod, Wen1111.6341}.
The $\upnu = 1$ of this TSC is the Balian-Werthamer (BW) state of the B phase of
${}^3$He liquid (${}^3$He B) \cite{Volovik2003Book}. A 2+1d single Majorana fermion can live on the
surface protected by time reversal symmetry in the non-interacting free system.
Including the time-reversal $\mathbb{Z}_4^T$ symmetry preserving interactions (preserving the $\mathrm{Pin}^+$ structure),
the 4d $\eta$ invariant is reduced from a free $\mathbb{Z}$ class to a torsion $\mathbb{Z}_{16}$ class.}
invariant and 5d $\eta\big(\text{PD}({\cal A}_{{\mathbb{Z}_2}}) \big)$ invariant.
On an Euclidean signature curved spacetime manifold $M^d$, the path integral of a massive
5d Dirac fermion spinor $\psi$ with a mass $m$
coupled to a background gauge field $A$ or a background gravity of a metric $g_{\mu \nu}$, the Euclidean path integral is
\begin{equation}
\int [{\cal D} {\psi}] [{\cal D}\bar{\psi}] \exp\left(-S_E \right)=\int [{\cal D} {\psi}] [{\cal D}\bar{\psi}] \exp\bigg(
-\int_{M^d} \hspace{1pt}\mathrm{d}^4x_E\, \sqrt{ \det g}
\big( \bar\psi (\slashed{D}_A+ m)\psi \big)
\bigg)=\det(\slashed{D}_A+ m),
\end{equation}
where \textit{locally} $\slashed{D}_A \equiv e^\mu_{\mu'} \gamma^{\mu'} (\partial_\mu+\hspace{1pt}\mathrm{i}\hspace{1pt} \omega_\mu-\hspace{1pt}\mathrm{i}\hspace{1pt} A_\mu)$, $e^\mu_{\mu'}$ is a vielbein,
with a spin connection $\omega_\mu$, and a gauge connection $A_\mu$ for a gauge bundle of a group $\tilde{G}$. More explicitly, the components are
$\omega_\mu=\hspace{1pt}\mathrm{i}\hspace{1pt} \omega_\mu^{\lambda\nu}[\gamma^\lambda,\gamma^\nu]/8$ and $A_\mu=A^{\mathrm{a}}_\mu \rm{T}^{\mathrm{a}}$ with $\rm{T}^{\mathrm{a}}$ generators of the Lie algebra $\text{Lie}(\tilde{G})$. We also specify the transition functions that relate fields
$\psi$ and $\bar{\psi}$ (two independent Grassmann fields)
and $A$ on different patches in order to \textit{globally} define the Dirac operator $\slashed{D}_A$ on $M^d$.
The transition function also leaves the local expression of the partition function invariant.
If we define $m \to \infty$ to be a trivial gapped vacua without any topological feature, then the
$m \to -\infty$ gapped vacua can host a nontrivial iTQFT or SPTs in $d$d.
Follow the notations in \cite{1711.11587GPW}, the
$d$d partition function of nontrivial iTQFT or SPTs
can be defined via a ratio\footnote{The $\det(-\hspace{1pt}\mathrm{i}\hspace{1pt}\slashed{D}_A)=\prod_\lambda \lambda$.
One can regularize it by $\det(-\hspace{1pt}\mathrm{i}\hspace{1pt}\slashed{D}_A)= \prod_\lambda \frac{\lambda}{\lambda + \hspace{1pt}\mathrm{i}\hspace{1pt} {\rm M}}$ via a Pauli-Villars regulator of mass ${\rm M}$.
The APS $\eta$-invariant \cite{Atiyah1975jfAPS, Atiyah1976APS, Atiyah1980APS}
is a regularization of
$\sum_{\lambda} (\text{sign}\,\lambda)$.
}
\begin{equation}
{\bf Z}_{{\text{$d$d-iTQFT}}}[A]= \lim_{|m|\rightarrow \infty}\frac{\det(\slashed{D}_A-|m|)}{\det(\slashed{D}_A+|m|)}\equiv
\lim_{|m|\rightarrow \infty}\prod_\lambda \frac{\hspace{1pt}\mathrm{i}\hspace{1pt} \lambda -|m|}{\hspace{1pt}\mathrm{i}\hspace{1pt} \lambda +|m|}
=\frac{1}{2}(N_0+\lim_{s\rightarrow {0}+}\sum_{\lambda\neq 0} (\text{sign}\,\lambda)\cdot |\lambda|^{-s})
\end{equation}
as a regularized sum of eigenvalues of Dirac operator.
Here $\slashed{D}_A$ is anti-Hermitian, so $-\hspace{1pt}\mathrm{i}\hspace{1pt}\slashed{D}_A$ is Hermitian.
The $\lambda$ are eigenvalues of $-\hspace{1pt}\mathrm{i}\hspace{1pt}\slashed{D}_A$ so $\lambda \in {\mathbb{R}}$ are real. \\
$N_0$ are the number of the operator $-\hspace{1pt}\mathrm{i}\hspace{1pt}\slashed{D}_A$'s zero modes.
Depending on the underlying $G$-structure of manifolds and $-\hspace{1pt}\mathrm{i}\hspace{1pt}\slashed{D}_A$
(such as $G={\mathrm{Pin}^+}$ \cite{Witten:2015aba}, $G={\mathrm{Pin}^c}$ \cite{gilkey1985eta, Metlitski1510.05663}, or $G={\mathrm{Pin}^+\times_{\mathbb{Z}_2}{\rm SU}(2)}$ \cite{1711.11587GPW}),
the 4d iTQFT/SPT partition functions are respectively:
\begin{eqnarray}
\Big({\bf Z}_{{\text{4d-iTQFT}}} \Big)^\upnu
\xrightarrow{|m|\rightarrow \infty}
\left\{\begin{array}{lll}
\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{2} \cdot\upnu \cdot \upeta_{\mathrm{Pin}^+}), & \text{ with } \upeta_{\text{Pin}^+} \in \frac{1}{8}\mathbb{Z}, &\quad \upnu\in \mathbb{Z}_{16}.\\
\exp({2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}} \cdot\upnu \cdot \upeta_{\mathrm{Pin}^c}), & \text{ with }\upeta_{\text{Pin}^c} \in \frac{1}{8}\mathbb{Z}, &\quad \upnu\in \mathbb{Z}_{8}.\\
\exp({2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}} \cdot\upnu \cdot \upeta_{\mathrm{Pin}^+\times_{\mathbb{Z}_2}{\rm SU}(2)}), & \text{ with }\upeta_{\mathrm{Pin}^+\times_{\mathbb{Z}_2}{\rm SU}(2)} \in \frac{1}{4}\mathbb{Z}, &\quad \upnu\in \mathbb{Z}_{4}.
\end{array}
\right.
\end{eqnarray}
However, in our work, we adjust above definitions to make $\eta_{\text{Pin}^+} \in \mathbb{Z}_{16}$ via $\eta_{\text{Pin}^+} = 8 \upeta_{\text{Pin}^+}$ mod 16, so:
\begin{eqnarray} \label{eq:4dZ16iTQFT}
\Big({\bf Z}_{{\text{$4$d-iTQFT}}} \Big)^\upnu
\xrightarrow{|m|\rightarrow \infty}
\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16} \cdot\upnu \cdot \eta_{\mathrm{Pin}^+} \bigg\rvert_{M^4}), & \text{ with } \eta_{\text{Pin}^+} \in \mathbb{Z}_{16}, &\quad \upnu\in \mathbb{Z}_{16}.
\end{eqnarray}
Similarly, based on the Smith homomorphism \cite{2018arXiv180502772T, GarciaEtxebarriaMontero2018ajm1808.00009, Hsieh2018ifc1808.02881, GuoJW1812.11959
\footnote{In fact, there are more Smith homomorphisms in any dimension. Related discussions along these maps are abundant, see
\cite{2018arXiv180502772T} and also \cite{Kapustin1406.7329, GuoJW1812.11959, WanWangv2}:
\begin{eqnarray} \label{eq:Smithmap}
\begin{array}{rrrl}
\Omega_8^{\mathrm{Pin}^+} =\mathbb{Z}_{32} \times \mathbb{Z}_2&\to& \Omega_7^{{\rm Spin} \times {\mathbb{Z}_2}}=\mathbb{Z}_{16}&\quad \text{generated by $\mathbb{RP}^7$},\\
\Omega_7^{{\rm Spin} \times {\mathbb{Z}_2}}=\mathbb{Z}_{16}&\to&
\Omega_6^{\mathrm{Pin}^-} =\mathbb{Z}_{16}& \quad \text{generated by $\mathbb{RP}^6$},\\
\Omega_6^{\mathrm{Pin}^-} =\mathbb{Z}_{16} &\to&
\Omega_5^{{\rm Spin} \times_{\mathbb{Z}_2} \mathbb{Z}_4}=\mathbb{Z}_{16}&\quad \text{generated by $\mathbb{RP}^5$},\\
\Omega_5^{{\rm Spin} \times_{\mathbb{Z}_2} \mathbb{Z}_4}=\mathbb{Z}_{16} &\to&
\Omega_4^{\mathrm{Pin}^+} =\mathbb{Z}_{16}& \quad \text{generated by $\mathbb{RP}^4$},\\
\Omega_4^{\mathrm{Pin}^+} =\mathbb{Z}_{16}
&\to& \Omega_3^{{\rm Spin} \times {\mathbb{Z}_2}}=\mathbb{Z}_{8}&\quad \text{generated by $\mathbb{RP}^3$},\\
\Omega_3^{{\rm Spin} \times {\mathbb{Z}_2}}=\mathbb{Z}_{8}&\to&
\Omega_2^{\mathrm{Pin}^-} =\mathbb{Z}_{8}&\quad \text{generated by $\mathbb{RP}^2$},\\
\Omega_2^{\mathrm{Pin}^-} =\mathbb{Z}_{8}&\to&
\Omega_1^{{\rm Spin} \times_{\mathbb{Z}_2} \mathbb{Z}_4}=\mathbb{Z}_{4}&\quad \text{generated by $\mathbb{RP}^1=S^1/\mathbb{Z}_2=S^1$},\\
\Omega_1^{{\rm Spin} \times_{\mathbb{Z}_2} \mathbb{Z}_4}=\mathbb{Z}_{4}
&\to&
\Omega_0^{\mathrm{Pin}^+} =\mathbb{Z}_{2} &\quad \text{generated by $\mathbb{RP}^0 = $ point}.
\end{array}
\end{eqnarray}
Notice that $\operatorname{H}^1(\mathbb{RP}^n,\mathbb{Z}_2)=\mathbb{Z}_2$, for example by using $n=5$, if we choose
${{\cal A}_{{\mathbb{Z}_2}}}$ for $\operatorname{H}^1(\mathbb{RP}^n,\mathbb{Z}_2)=\mathbb{Z}_2$, then $\mathbb{RP}^4$ and $\mathbb{RP}^5$ detect all $\upnu \in \mathbb{Z}_{16}$:
\begin{eqnarray}
\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16} \upnu \left. \eta\right\vert_{{M^4=\mathbb{RP}^4}}) &=&\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16} \upnu ).\cr
\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16}\upnu \left.\eta\big(\text{PD}({\cal A}_{{\mathbb{Z}_2}})\big) \right\vert_{M^5=\mathbb{RP}^5}) &=&\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16} \upnu).
\end{eqnarray}
}
\begin{eqnarray} \label{eq:5to4Smith}
\Omega_5^{{\rm Spin} \times_{\mathbb{Z}_2} \mathbb{Z}_4}=\mathbb{Z}_{16} \quad \text{(generated by $\mathbb{RP}^5$)}
\stackrel{ \cap {\cal A}_{{\mathbb{Z}_2}}}{\xrightarrow{\hspace*{1cm}}}
\Omega_4^{\mathrm{Pin}^+} =\mathbb{Z}_{16} \quad \text{(generated by $\mathbb{RP}^4$)},
\end{eqnarray}
we characterize the 5d iTQFT whose manifold generator for bordism group is $\mathbb{RP}^5$ at $\upnu=1$ below.
The cobordism invariant for any $\upnu$ corresponds to a 5d iTQFT partition function
\begin{eqnarray} \label{eq:5dSPT}
\Big({\bf Z}_{{\text{$5$d-iTQFT}}} \Big)^\upnu
\xrightarrow{|m|\rightarrow \infty}
\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16} \cdot\upnu \cdot \eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}})) \bigg\rvert_{M^5}), & \text{ with } \eta = \eta_{\text{Pin}^+} \in \mathbb{Z}_{16}, &\quad \upnu\in \mathbb{Z}_{16}.
\end{eqnarray}
Given a $G \supseteq {{\rm Spin} \times_{\mathbb{Z}_2^F} \mathbb{Z}_4}$ structure, the cohomology class
${\cal A}_{{\mathbb{Z}_2}} \in \operatorname{H}^1(M,\mathbb{Z}_2)$ is the generator from $\operatorname{H}^1(\mathrm{B}(\mathbb{Z}_4/\mathbb{Z}_2^F),\mathbb{Z}_2)$.
So $({\cal A}_{{\mathbb{Z}_2}})\equiv ({\cal A}_{{\mathbb{Z}_4}}) \mod 2$ is the quotient for the $\mathbb{Z}_4$ gauge field ${\cal A}_{{\mathbb{Z}_4}}$, which is the background field for the symmetry $\mathbb{Z}_{4,{X}} \subset {\rm U}(1)_{X}$.
The $\eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}}))$ is
the value of ${\eta} \in \mathbb{Z}_{16}$ on the Poincar\'e dual (PD) submanifold of the cohomology class
${\cal A}_{{\mathbb{Z}_2}}$. This PD is the precise meaning of \Eq{eq:5to4Smith} taking
${ \cap {\cal A}_{{\mathbb{Z}_2}}}$ from 5d to 4d.
Thus, to summarize, the full 5d iTQFT and the 4d SM or SU(5) GG GUT model action $S_{\text{SM or GUT}}$
with $N_{\text{generation}}$
together can make the anomaly \Eq{eq:Ngeneration} matched via
the full 5d-4d coupled partition function
\begin{eqnarray} \label{eq:UU-GUT}
{\bf Z}_{\text{5d-4d}}
=
\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16} \cdot\upnu \cdot \eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}})) \rvert_{M^5}) \bigg\rvert_{\upnu=-N_{\text{generation}}}\cdot
\int [{\cal D} {\psi}] [{\cal D}\bar{\psi}][{\cal D} A]\cdots \exp\left( \hspace{1pt}\mathrm{i}\hspace{1pt} S_{\text{SM or GUT}} \right \rvert_{M^4}). \quad\quad
\end{eqnarray}
with a bulk-boundary correspondence $\prt {M^5}={M^4}$.
The full ${\bf Z}_{\text{5d-4d}}$ is gauge invariant, in particular also invariant under the background
$\mathbb{Z}_{4,X}$ transformation (at least at the higher-energy of GUT scale).
This concludes the
5d SPTs coupled to 4d SM or GUT in the Scenario \ref{Z45dSPT}.
\subsection{Hidden 4d non-invertible TQFT or 4d symmetry-enriched topological order}
\label{sec:Hidden4dnon-invertiblenon-abelianTQFT}
Now we explore the Scenario \ref{Z4XTQFT} (thus also Scenario \ref{Z4XbreakTQFT})
for a hidden 4d non-invertible TQFT in \Sec{sec:Hidden4dnon-invertiblenon-abelianTQFT} to match the missing anomaly.
In fact to construct such an {\bf anomalous symmetry-preserving 4d TQFT} (here we preserve $\mathbb{Z}_{4,X}$ in ${{\rm Spin} \times_{\mathbb{Z}_2} \mathbb{Z}_4}$ structure),
we take the inspirations from the quantum condensed matter phenomenon in one lower dimension,
known as the {\bf anomalous symmetry-enriched boundary topological order} in 3d (2+1d spacetime)
living on the boundary of 4d SPTs (3+1d spacetime).
In condensed matter thinking, it can be understood as saturating the 't Hooft anomaly by a symmetric gapped sector,
without breaking any symmetry breaking and without gapless modes,
via smearing the 't Hooft anomaly and anomalous symmetry into the long-range entanglement.
In the context of 3d boundary and 4d bulk,
a novel surface topological order was firstly pointed out by
Vishwanath-Senthil in an insightful work \cite{VishwanathSenthil1209.3058}. Later on many people follow up on developing the
surface topological order constructions (see overviews in \cite{Senthil1405.4015, Wen2016ddy1610.03911} and \cite{Wang2017locWWW1705.06728}).
We need to generalize this condensed matter idea to find an anomalous symmetric 4d TQFT living on the boundary of 5d fermionic SPTs \Eq{eq:5dSPT}.
In particular, we require the symmetry-extension method \cite{Wang2017locWWW1705.06728} and
the spin-TQFT construction from a previous work \cite{GuoJW1812.11959}.
\subsubsection{Symmetry extension $[\mathbb{Z}_2] \to {\rm Spin} \times {\mathbb{Z}_{4,X}} \to {\rm Spin} \times_{\mathbb{Z}_2^F} {\mathbb{Z}_{4,X}}$ and a 4d $[\mathbb{Z}_2]$ gauge theory}
\label{sec:sym-ext}
First, we can rewrite the 5d iTQFT partition function \eq{eq:5dSPT} on a 5d manifold $M^5$ in terms of
\begin{multline} \label{eq:5dSPT-4dTQFT}
\Big({\bf Z}_{{\text{$5$d-iTQFT}}} \Big)^\upnu
=
\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16} \cdot\upnu \cdot \eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}})) \bigg\rvert_{M^5})\\
=
\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16} \cdot\upnu \cdot \Big( 8 \cdot\frac{p_1(TM)}{48}(\text{PD}({\cal A}_{{\mathbb{Z}_2}})) + 4 \cdot\text{Arf} (\text{PD}(({\cal A}_{{\mathbb{Z}_2}})^3) )
+ 2 \cdot{\tilde{\eta}} (\text{PD}(({\cal A}_{{\mathbb{Z}_2}})^4) ) + ({\cal A}_{{\mathbb{Z}_2}})^5 \Big) \bigg\rvert_{M^5})\\
=
\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16} \cdot\upnu \cdot \Big( 8 \cdot\frac{\sigma}{16}(\text{PD}({\cal A}_{{\mathbb{Z}_2}})) + 4 \cdot\text{Arf} (\text{PD}(({\cal A}_{{\mathbb{Z}_2}})^3) )
+ 2 \cdot{\tilde{\eta}} (\text{PD}(({\cal A}_{{\mathbb{Z}_2}})^4) ) + ({\cal A}_{{\mathbb{Z}_2}})^5 \Big) \bigg\rvert_{M^5}),
\end{multline}
for a generic ${\upnu=-N_{\text{generation}}} \in \mathbb{Z}_{16}$.
Here are some explanations on the mathematical inputs of \eq{eq:5dSPT-4dTQFT}
(also partly explained in the Notation in the end of \Sec{sec:Intro}):\\
$\bullet$ The $p_1(TM)$ is the first Pontryagin class of spacetime tangent bundle $TM$ of the manifold $M$.
Via the Hirzebruch signature theorem, we have
$$\frac{1}{3}\int_{\Sigma^4} p_1(TM) =\sigma(\Sigma^4) =\sigma$$
on a 4-manifold $\Sigma^4$, where $\sigma$ is the signature of $\Sigma^4$.
So in \eq{eq:5dSPT-4dTQFT}, we evaluate the $\frac{1}{3}\int_{\Sigma^4} p_1(TM) =\sigma$ on the Poincar\'e dual $\Sigma^4$ manifold of
the $({\cal A}_{{\mathbb{Z}_2}})$ cohomology class within the $M^5$.\\
{$\bullet$ The $\tilde{\eta}$ is a mod 2 index of 1d Dirac operator as a cobordism invariant of $\Omega_1^{{\rm Spin}}=\mathbb{Z}_2$.\\
$\bullet$ The Arf invariant \cite{Arf1941} is a mod 2 cobordism invariant of $\Omega_2^{{\rm Spin}}=\mathbb{Z}_2$,
whose realization is the 1+1d Kitaev fermionic chain \cite{Kitaev2001chain0010440}.\\
$\bullet$ The $({\cal A}_{{\mathbb{Z}_2}})^5$ is a mod 2 class purely bosonic topological invariant, which corresponds to a 5d bosonic SPT phase given by
the group cohomology class data $\operatorname{H}^5(\mathrm{B}\mathbb{Z}_2,{\rm U}(1))=\mathbb{Z}_2$, which is also one of the $\mathbb{Z}_2$ generators in $\Omega_5^{{\rm SO}}(\mathrm{B}\mathbb{Z}_2)$.
}
Below we ask whether we can construct a fully gauge-invariant 5d-4d coupled partition function preserving the
${{\rm Spin} \times_{\mathbb{Z}_2} \mathbb{Z}_4}$ structure:
\begin{eqnarray} \label{eq:5dSPT-4dTQFT-0}
\Big({\bf Z}_{{\text{$5$d-iTQFT}}} \Big)^\upnu
\Big({\bf Z}_{{\text{$4$d-TQFT}}} \Big).
\end{eqnarray}
Preserving the
${{\rm Spin} \times_{\mathbb{Z}_2} \mathbb{Z}_4}$ structure means
that under the spacetime background transformation and the ${\cal A}_{{\mathbb{Z}_2}}$ background gauge transformation,
the 5d-4d coupled partition function is still fully gauge invariant.
\begin{enumerate}[leftmargin=-1.mm]
\item
When ${\upnu}$ is odd, such as ${\upnu}=1,3,5,7,\dots \in \mathbb{Z}_{16}$,
\Refe{Hsieh2018ifc1808.02881} suggested that the symmetry-extension method \cite{Wang2017locWWW1705.06728} cannot help to construct a symmetry-gapped TQFT.
Furthermore, Cordova-Ohmori \cite{Cordova1912.13069} proves that a symmetry-preserving gapped TQFT phase is impossible for
this odd ${\upnu} \in \mathbb{Z}_{16}$ anomaly from $\Omega_5^{{\rm Spin} \times_{\mathbb{Z}_2} \mathbb{Z}_4}=\mathbb{Z}_{16}$.
The general statement in \cite{Cordova1912.13069} is that
given an anomaly index $\upnu \in \Omega_5^{{\rm Spin} \times_{\mathbb{Z}_2} \mathbb{Z}_{4}}=\mathbb{Z}_{16}$,
we can at most construct a fully symmetric gapped TQFT
\emph{if and only if}
$$ 4 \mid 2 \upnu.$$
Namely, 4 has to be a divisor of $2 \upnu$.
Apparently, the $4 \mid 2 \upnu$ is true only when $\upnu$ is even,
while the $4 \mid 2 \upnu$ is false when $\upnu$ is odd.
Since ${\upnu=-N_{\text{generation}}}$, the case of ${\upnu}=1$ (for a single generation) and ${\upnu}=3$ (for three generations) are particularly important
for the high energy physics phenomenology.
{This means that we are not able to
directly construct any 4d symmetric gapped TQFT that explicitly matches the same $\mathbb{Z}_{16}$ anomaly
for one right-handed neutrino (${\upnu}=1$) or three right-handed neutrinos (${\upnu}=3$).}
\item When ${\upnu}$ is even, such as ${\upnu}=2,4,6,8,\dots \in \mathbb{Z}_{16}$,
\Refe{Hsieh2018ifc1808.02881} suggested that the symmetry-extension method \cite{Wang2017locWWW1705.06728} can trivialize the 't Hooft anomaly.
Furthermore, Cordova-Ohmori \cite{Cordova1912.13069} shows that there is no obstruction to construct
a symmetry-preserving gapped TQFT phase for any even ${\upnu_{\text{even}} } \in \mathbb{Z}_{16}$.
We can verify the claim by rewriting \eq{eq:5dSPT-4dTQFT} in terms of the $(\frac{\upnu_{\text{even}}}{2}) \in \mathbb{Z}_{8}$ index
with a 2d Arf-Brown-Kervaire (ABK) invariant:
\begin{multline} \label{eq:5dSPT-4dTQFT-Z8}
\Big({\bf Z}_{{\text{$5$d-iTQFT}}} \Big)^{\upnu_{\text{even}}}
=
\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16} \cdot{\upnu_{\text{even}}} \cdot \eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}})) \bigg\rvert_{M^5})
=
\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{8} \cdot (\frac{{\upnu_{\text{even}}}}{2}) \cdot \Big(
\text{ABK} (\text{PD}(({\cal A}_{{\mathbb{Z}_2}})^3) )
\Big) \bigg\rvert_{M^5}) \\
=
\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{8} \cdot (\frac{{\upnu_{\text{even}}}}{2}) \cdot \Big( 4 \cdot \text{Arf} (\text{PD}(({\cal A}_{{\mathbb{Z}_2}})^3) )
+ 2 \cdot{\tilde{\eta}} (\text{PD}(({\cal A}_{{\mathbb{Z}_2}})^4) ) + ({\cal A}_{{\mathbb{Z}_2}})^5 \Big) \bigg\rvert_{M^5}) .
\end{multline}
Notice that \eq{eq:5dSPT-4dTQFT-Z8} can become trivialized if we can trivialize the $({\cal A}_{{\mathbb{Z}_2}})^2$ factor.
In fact, the topological term $({\cal A}_{{\mathbb{Z}_2}})^2$ can be trivialized by the symmetry extension \cite{Wang2017locWWW1705.06728}
$$
0 \to \mathbb{Z}_2 \to {\mathbb{Z}_{4,X}} \to \frac{\mathbb{Z}_{4,X}}{\mathbb{Z}_2^F} \to 0.
$$
Namely, the cocycle topological term $({\cal A}_{{\mathbb{Z}_2}})^2$ in $\operatorname{H}^2(\mathrm{B}(\frac{\mathbb{Z}_{4,X}}{\mathbb{Z}_2^F}),{\rm U}(1))$ becomes a coboundary once we
lifting the $\frac{\mathbb{Z}_{4,X}}{\mathbb{Z}_2^F}$-gauge field ${\cal A}_{{\mathbb{Z}_2}}$
to a ${\mathbb{Z}_{4,X}}$ gauge field in
$\operatorname{H}^2(\mathrm{B}{\mathbb{Z}_{4,X}},{\rm U}(1))$. So this suggests that the following symmetry extension for the spacetime-internal symmetry\footnote{In fact,
the recent works \cite{Wan2019sooWWZHAHSII1912.13504, PrakashJW2011.13921, PTWtoappear} suggest that many cobordism invariants given in \eq{eq:Smithmap} can be trivialized via a $\mathbb{Z}_2$ extension. For example, the following group extension to $\mathrm{EPin}(d)$ via
\begin{eqnarray}
1 \to \mathbb{Z}_2 \to \mathrm{EPin}(d) \to \mathrm{Pin}^{-}(d) \to 1
\end{eqnarray}
can trivialize the cobordism invariants when $\upnu = 4 \in \mathbb{Z}_{8} = \Omega_2^{\mathrm{Pin}^-}$ at $d=2$,
and for any $\upnu \in\mathbb{Z}_{16} = \Omega_6^{\mathrm{Pin}^-}$ at $d=6$.
The $\mathrm{EPin}(d)$ is firstly introduced in \cite{Wan2019sooWWZHAHSII1912.13504}.
The following extension to ${\rm Spin}(d) \times\mathbb{Z}_4$ via
\begin{eqnarray}
1 \to \mathbb{Z}_2 \to {\rm Spin}(d) \times\mathbb{Z}_4 \to {\rm Spin}(d) \times\mathbb{Z}_2 \to 1
\end{eqnarray}
can trivialize the cobordism invariants for any even $\upnu_{\text{even}} $ $\in \mathbb{Z}_{8} = \Omega_3^{{\rm Spin} \times\mathbb{Z}_2}$ at $d=3$
and for any $\upnu \in\mathbb{Z}_{16} = \Omega_7^{{\rm Spin} \times\mathbb{Z}_2}$ at $d=7$.
The following group extension to $\mathrm{EPin}(d)$ via
\begin{eqnarray}
1 \to \mathbb{Z}_2 \to \mathrm{EPin}(d) \to \mathrm{Pin}^{+}(d) \to 1
\end{eqnarray}
can trivialize the cobordism invariants for any even $\upnu_{\text{even}} $ $\in \mathbb{Z}_{16} = \Omega_4^{\mathrm{Pin}^+}$ at $d=4$.
The following extension to ${\rm Spin}(d) \times\mathbb{Z}_4$ via
\begin{eqnarray}
1 \to \mathbb{Z}_2 \to {\rm Spin}(d) \times\mathbb{Z}_4 \to {\rm Spin}(d) \times_{\mathbb{Z}_2} \mathbb{Z}_4 \to 1
\end{eqnarray}
can trivialize the cobordism invariants for any even $\upnu_{\text{even}} $ $\in \mathbb{Z}_{16} = \Omega_5^{{\rm Spin} \times_{\mathbb{Z}_2} {\mathbb{Z}_4}}$ at $d=5$.
This is similar to what \Refe{RyanThorngren2001.11938} suggested that if the anomaly can be matched by a symmetric gapped boundary TQFT, then a
$\mathbb{Z}_2$ gauge theory almost always work.
}
\begin{eqnarray} \label{eq:Z2-SpinxZ4}
1 \to [\mathbb{Z}_2] \to {\rm Spin} \times {\mathbb{Z}_{4,X}} \to {\rm Spin} \times_{\mathbb{Z}_2^F} {\mathbb{Z}_{4,X}} \to 1.
\end{eqnarray}
can fully trivialize any even ${\upnu_{\text{even}}} \in \mathbb{Z}_{16}$ cobordism invariant given in \eq{eq:5dSPT-4dTQFT-Z8}.
The $[\mathbb{Z}_2]$ means that we can gauge the normal subgroup $[\mathbb{Z}_2]$ in the total group ${\rm Spin} \times {\mathbb{Z}_{4,X}}$.
This symmetry extension \eq{eq:Z2-SpinxZ4} also means that a 4d $[\mathbb{Z}_2]$ gauge theory can preserve the
${\rm Spin} \times_{\mathbb{Z}_2^F} {\mathbb{Z}_{4,X}}$ symmetry while
can also saturate the even ${\upnu_{\text{even}}} \in \mathbb{Z}_{16}$ anomaly.
This 4d $[\mathbb{Z}_2]$ gauge theory is the anomalous symmetric gapped non-invertible TQFT
that we aim for in the Scenario \ref{Z4XTQFT}.
\end{enumerate}
\subsubsection{4d anomalous symmetric $\mathbb{Z}_2$ gauge theory with $\upnu = 2 \in \mathbb{Z}_{16}$ preserving ${\rm Spin} \times_{\mathbb{Z}_2^F} {\mathbb{Z}_{4,X}}$}
\label{sec:Z16-nu=2-TQFT}
For example, when
${\upnu_{\text{even}}} =2 \in \mathbb{Z}_{16}$,
we have \eq{eq:5dSPT-4dTQFT-0} with the input of 5d bulk iTQFT \eq{eq:5dSPT-4dTQFT-Z8},
then we can explicitly construct the partition function on a 5d manifold $M^5$ with a 4d boundary
$M^4 \equiv \partial M^5$ as,\footnote{Follow the Notation in the end of \Sec{sec:Intro},
we use the $\smile$ notation for the cup product between cohomology classes, or between a cohomology class and a fermionic topological invariant.
We use the $\cup$ notation for the surgery gluing the boundaries of two manifolds within relative homology classes.
So the $M_1 \cup M_2$ means gluing the boundary $\partial M_1 = \overline{\partial M_2}$ such that the common orientation of
$\overline{\partial M_2}$ is the reverse of ${\partial M_2}$.
}
\begin{multline} \label{eq:5dSPT-4dTQFT-explicit}
\Big({\bf Z}_{{\text{$5$d-iTQFT}}} \Big)^{{\upnu_{\text{even}}} =2 } \cdot
\Big({\bf Z}_{{\text{$4$d-$\mathbb{Z}_2$-gauge theory}}} \Big) \\
=
\sum_{c\in {\partial'}^{-1}(\partial [\text{PD}({\cal A}^3)])}\hspace{1pt}\mathrm{e}^{\frac{2\pi\hspace{1pt}\mathrm{i}\hspace{1pt}}{8}\text{ABK}(c\cup \text{PD}({\cal A}^3))} \quad\quad\quad\quad\quad\quad\\
\cdot \frac{1}{2^{|\pi_0(M^4)|}} \sum_{\substack{a\in C^1(M^4,\mathbb{Z}_2) ,\\b\in C^2(M^4,\mathbb{Z}_2)} }(-1)^{\int_{ M^4} b (\delta a+{\cal A}^2)} \cdot \hspace{1pt}\mathrm{e}^{\frac{2\pi\hspace{1pt}\mathrm{i}\hspace{1pt}}{8}\text{ABK}(c\cup \text{PD}'({\cal A} a))}
\cdot
\end{multline}
Here we write the mod 2 cohomology class $\mathbb{Z}_2$ gauge field as ${\cal A} \equiv {\cal A}_{{\mathbb{Z}_2}} \in \operatorname{H}^1(M^5, \mathbb{Z}_2)$.
We generalize the boundary TQFT construction in the Section 8 of \cite{GuoJW1812.11959}.
Here are some remarks about this 5d bulk iTQFT-4d boundary TQFT partition function \eq{eq:5dSPT-4dTQFT-explicit}:
\begin{enumerate} [leftmargin=.mm, label=\textcolor{blue}{\arabic*)}., ref={\arabic*)}]
\item
The $(\text{PD}({\cal A}^3))$ is a 2d manifold taking the Poincar\'e dual (PD) of 3-cocycle ${\cal A}^3$ in the $M^5$;
but the 2d manifold $(\text{PD}({\cal A}^3))$ may touch the
the 4d boundary $\partial M^5 =M^4$.
The 1d boundary $\partial (\text{PD}({\cal A}^3))$ can be regarded as the 1d intersection
between the 2d $(\text{PD}({\cal A}^3))$ and the $M^4$.
\item \noindent
More precisely, for a ${\rm Spin} \times_{\mathbb{Z}_2} \mathbb{Z}_4$ manifold $M^5$ with a boundary,
we have used the Poincar\'e-Lefschetz duality for a manifold with boundaries:
\begin{eqnarray}
{\cal A}^2\in \operatorname{H}^3(M^5,\mathbb{Z}_2)\xrightarrow{\cong} \operatorname{H}_2(M^5, M^4,\mathbb{Z}_2)\ni \text{PD}({\cal A}^3).
\end{eqnarray}
\item \noindent
For any pair $({\rm S},{\rm S}')$, where ${\rm S}'$ is a subspace of $S$, the short exact sequence of chain complexes
\begin{eqnarray}
0\to C_*({\rm S}')\to C_*({\rm S})\to C_*({\rm S},{\rm S}')\to 0
\end{eqnarray}
induces a long exact sequence of homology groups
\begin{eqnarray}
\cdots\to \operatorname{H}_n({\rm S}')\to \operatorname{H}_n({\rm S})\to \operatorname{H}_n({\rm S},{\rm S}')\xrightarrow{\partial} \operatorname{H}_{n-1}({\rm S}')\to\cdots.
\end{eqnarray}
Here $C_n({\rm S},{\rm S}'):=C_n({\rm S})/C_n({\rm S}')$,
the $\operatorname{H}_n({\rm S},{\rm S}')$ is the relative homology group, and $\partial$ is the boundary map.\\
$\bullet$ Take $({\rm S},{\rm S}')=(\text{PD}({\cal A}^3),\partial \text{PD}({\cal A}^3))$, we denote the boundary map by $\partial$:
\begin{eqnarray} \label{eq:map-partial}
\operatorname{H}_2( \text{PD}({\cal A}^3), \partial \text{PD}({\cal A}^3) ) \xrightarrow{\partial} \operatorname{H}_{1}( \partial \text{PD}({\cal A}^3)).
\end{eqnarray}
Here $\text{PD}({\cal A}^3)$ is not a closed 2-manifold, but which has a boundary closed 1-manifold $\partial \text{PD}({\cal A}^3)$. \\
$\bullet$ Take $({\rm S},{\rm S}')=(M^4= \partial M^5, \partial \text{PD}({\cal A}^3))$, we denote another boundary map by $\partial_1$:
\begin{eqnarray} \label{eq:map-partial1}
\operatorname{H}_2(M^4, \partial \text{PD}({\cal A}^3) ) \xrightarrow{\partial_1} \operatorname{H}_{1}( \partial \text{PD}({\cal A}^3)).
\end{eqnarray}
Both $M^4= \partial M^5$ and $\partial \text{PD}({\cal A}^3)$ are closed manifolds, of 4d and 1d, respectively.
\item \noindent
Now, the $c$ is defined as a 2d surface living on the boundary $M^4$.
The ${\partial_1}c$ uses the boundary map \eq{eq:map-partial1}'s $\partial_1$ of $c$ on the $M^4$.
We can compensate the 2-surface $\text{PD}({\cal A}^3)$ potentially with a 1-boundary,
by gluing it with $c$ to make a closed 2-surface.
To do so, we require both $\text{PD}({\cal A}^3)$ and $c$ share the same 1d boundary.
\item
\noindent
The ${c\in {\partial_1}^{-1}(\partial [\text{PD}({\cal A}^3)])}$ also means
\begin{eqnarray} \label{eq:partial'c=partialPDA3}
{{\partial_1} c = \partial [\text{PD}({\cal A}^3)] = [\partial \text{PD}({\cal A}^3)]}.
\end{eqnarray}
This is exactly the requirement that the bulk 2-surface $\text{PD}({\cal A}^3)$ (living in $M^5$) and
the boundary 2-surface $c$ (living on $M^4$) share the same 1d boundary. \\
\noindent
$\bullet$ The $[\text{PD}({\cal A}^3)]$ means the fundamental class and the relative homology class of the 2d manifold $\text{PD}({\cal A}^3)$.\\
\noindent
$\bullet$ The {$\partial [\text{PD}({\cal A}^3)] = [\partial \text{PD}({\cal A}^3)]$} means the boundary of the fundamental class
(via the boundary $\partial$ map in \eq{eq:map-partial})
is equivalent to the fundamental class of the boundary $\partial$ of $\text{PD}({\cal A}^3)$.
Beware that the two $\partial$ operations in {$\partial [\text{PD}({\cal A}^3)] = [\partial \text{PD}({\cal A}^3)]$} have different meanings.
\item
\noindent Note that when $M^5$ is a closed manifold with no boundary $M^4= \partial M^5= \emptyset$ thus $c = \emptyset$,
then the term
$$\sum_{c\in {\partial_1}^{-1}(\partial [\text{PD}({\cal A}^3)])}\hspace{1pt}\mathrm{e}^{\frac{2\pi\hspace{1pt}\mathrm{i}\hspace{1pt}}{8}\text{ABK}(c\cup \text{PD}({\cal A}^3))}
\text{ is equivalently reduced to }
\hspace{1pt}\mathrm{e}^{\frac{2\pi\hspace{1pt}\mathrm{i}\hspace{1pt}}{8}\text{ABK}( \text{PD}({\cal A}^3))}
=\Big({\bf Z}_{{\text{$5$d-iTQFT}}} \Big)^{\upnu_{\text{even}}=2}.$$
This is a satisfactory consistent check, consistent with the 5d bulk-only iTQFT at ${\upnu_{\text{even}}}=2$ in \eq{eq:5dSPT-4dTQFT-Z8}.
\item \noindent
The
$a\in C^1(M^4,\mathbb{Z}_2)$ means that $a$ is a 1-cochain,
and the $b\in C^2(M^4,\mathbb{Z}_2)$ means that {$b$ is a 2-cochain}.
The factor $(-1)^{\int_{ M^4} b (\delta a+{\cal A}^2)}=
\exp(\hspace{1pt}\mathrm{i}\hspace{1pt} \pi {\int_{ M^4} b (\delta a+{\cal A}^2)})$ gives the weight of the
4d $\mathbb{Z}_2$ gauge theory.
The $b \delta a$ term is the BF theory written in the mod 2 class.\footnote{The continuum QFT version of this $\mathbb{Z}_2$ gauge theory is
$\exp(\hspace{1pt}\mathrm{i}\hspace{1pt} {\int_{ M^4}\frac{2}{2 \pi} b \hspace{1pt}\mathrm{d} a+ \frac{1}{\pi^2} b {\cal A}^2)})$, where
the $b$ integration over a closed 2-cycle, $\oiint b$, can be $n \pi$ with some integer $n \in \mathbb{Z}$.
The $a$ and ${\cal A}$ integration over a closed 1-cycle, $\oint a$ and $\oint {\cal A}$, can be $n \pi$ with some integer $n \in \mathbb{Z}$.}
The path integral sums over these distinct cochain classes.
The variation of $b$ gives the equation of motion $(\delta a+{\cal A}^2) = 0 \mod 2.$
In the path integral, we can integrate out $b$ to give the same constraint $(\delta a+{\cal A}^2) = 0 \mod 2.$
This is precisely the trivialization of the second cohomology class,
\begin{eqnarray} \label{eq:trivializationA2}
{\cal A}^2= ({\cal A}_{{\mathbb{Z}_2}})^2 =\delta a \mod 2
\end{eqnarray}
so the 2-cocycle becomes a 2-coboundary which splits to a 1-cochain $a$. This is exactly the condition imposed by the symmetry extension \eq{eq:Z2-SpinxZ4}:
$1 \to [\mathbb{Z}_2] \to {\rm Spin} \times {\mathbb{Z}_{4,X}} \to {\rm Spin} \times_{\mathbb{Z}_2^F} {\mathbb{Z}_{4,X}} \to 1$,
where the ${\cal A}^2$ term in $ {\rm Spin} \times_{\mathbb{Z}_2^F} {\mathbb{Z}_{4,X}}$ becomes trivialized as a coboundary
${\cal A}^2 =\delta a$
(so ${\cal A}^2= 0$ in terms of a cohomology or cocycle class)
in ${\rm Spin} \times {\mathbb{Z}_{4,X}}$.
\item \noindent
The $\frac{1}{2^{|\pi_0(M^4)|}}$ factor mod out the gauge redundancy for the boundary 4d $\mathbb{Z}_2$ gauge theory.
The $\pi_0(M^4)$ is the zeroth homotopy group of $M^4$, namely
the set of all path components of $M^4$.
Thus, we only sum over the gauge inequivalent classes in the path integral summation.
\item \noindent
The ${\text{ABK}}(c\cup \text{PD}({\cal A}^3))$ is defined on a 2d manifold with $\mathrm{Pin}^-$ structure.
Recall the $M^5$ has the ${\rm Spin} \times_{\mathbb{Z}_2} {\mathbb{Z}_{4}}$ structure.
If $M^5$ is closed, then there is a natural Smith map \eq{eq:Smithmap} to induce the 2d $\mathrm{Pin}^-$ structure on the closed surface via $\text{PD}({\cal A}^3)$.
However, the $M^5$ has a boundary $M^4$, so $ \text{PD}({\cal A}^3)$ may not be closed --- the previously constructed closed 2-surface $(c\cup \text{PD}({\cal A}^3))$
is meant to induce a 2d $\mathrm{Pin}^-$ structure.\footnote{We do not yet know whether it is always possible to induce a unique 2d $\mathrm{Pin}^-$ structure on
$(c\cup \text{PD}({\cal A}^3))$ for any possible pair of data $(M^5,M^4= \partial M^5)$ given any $\mathrm{Pin}^-$ $M^5$. However, we claim that it is possible
to find some suitable $M^4$ so that the 2d $(c\cup \text{PD}({\cal A}^3))$ has $\mathrm{Pin}^-$ induced, thus in this sense the ${\text{ABK}}(c\cup \text{PD}({\cal A}^3))$ is defined.
\label{ft:inducePin-1}}
Then we compute the ABK on this closed 2-surface $(c\cup \text{PD}({\cal A}^3))$.
\item \noindent
Let us explain
the other term ${\text{ABK}(c\cup \text{PD}'({\cal A} a))}$ of the 4d boundary TQFT in \eq{eq:5dSPT-4dTQFT-explicit}.
The $\text{PD}'$ is the Poincar\'e dual on $M^4= \partial M^5$.
Since
the fundamental classes of $M^5$ and $M^4= \partial M^5$ are related by
\begin{eqnarray}
[M^5]\in \operatorname{H}_5(M^5, M^4,\mathbb{Z}_2)\xrightarrow{\partial}\operatorname{H}_4( M^4,\mathbb{Z}_2)\ni[ M^4] = [\partial M^5].
\end{eqnarray}
Here we have the following relations:\footnote{{Let us clarify
the notations: $\partial$, $\partial'$, and $\partial_1$.
The boundary notation $\partial$ may mean as (1) taking the boundary, or (2) in the boundary map of relative homology class in \eq{eq:map-partial}.
It should be also clear to the readers
that\\
$\bullet$ the $\partial$ is associated with the operations on objects living in the bulk $M^5$ or ending on the boundary $M^4$,\\
$\bullet$ while the $\partial'$ is associated with the operations on objects living on the boundary $M^4$ alone.\\
The $\partial_1$ is defined as another boundary map in \eq{eq:map-partial1}.}}
\begin{eqnarray} \label{eq:PD-cap-relation}
\text{PD}&=&[M^5]\cap, \cr
\text{PD}' &=&[ M^4]\cap=[\partial M^5]\cap =\partial[ M^5]\cap,\\
{\partial'} \text{PD}'({\cal A} a) &=&
{\text{PD}'(\delta ({\cal A} a))=\text{PD}'({\cal A}^3)=\partial [\text{PD}({\cal A}^3)]=\partial_1 c.}
\end{eqnarray}
$\bullet$ The cap product $\cap$ here is to define PD homology class, so $\text{PD}({\cal A}) = [M^5] \cap {\cal A}$.\\
$\bullet$ Here we use $[M^4]=[\partial M^5]=\partial [M^5] $:
the fundamental class of boundary of $M^5$ gives
the boundary of fundamental class.\\
$\bullet$ {The \eq{eq:PD-cap-relation}'s first equality ${\partial'} \text{PD}'({\cal A} a) =\text{PD}'(\delta ({\cal A} a))$ uses
the coboundary operator $\delta$ on the cohomology class ${\cal A} a$.}\\
$\bullet$ The \eq{eq:PD-cap-relation}'s second equality $\text{PD}'(\delta ({\cal A} a))=\text{PD}'({\cal A}^3)$
uses the condition $\delta {\cal A} =0$ and the trivialization condition \eq{eq:trivializationA2}: $\delta a = \cA^2$. \\
$\bullet$ The \eq{eq:PD-cap-relation}'s third equality $\text{PD}'({\cal A}^3)=\partial [\text{PD}({\cal A}^3)]$,
we use ``\emph{the naturality of the cap product}''.
There are natural pushforward
and pullback maps on homology and cohomology,
related by the projection formula, also known as ``the naturality of the cap product.''
\\
$\bullet$ The \eq{eq:PD-cap-relation}'s last equality $\partial [\text{PD}({\cal A}^3)]=\partial_1 c$ is based on \eq{eq:partial'c=partialPDA3}.
Importantly, as a satisfactory consistency check, this also shows that the 2-surface $c$ obeys:
\begin{eqnarray}
\text{the $c$
in ${\text{ABK}}(c\cup \text{PD}({\cal A}^3))$ is the same $c$ in ${\text{ABK}(c\cup \text{PD}'({\cal A} a))}$.}
\end{eqnarray}
$\bullet$ Similarly as before, the union $(c\cup\text{PD}'({\cal A} a))$ is a closed 2-surface and we induce a $\mathrm{Pin}^-$ structure on this 2-surface.
So we can compute the ABK on this closed 2-surface $(c\cup\text{PD}'({\cal A} a))$ on the $M^4$.\footnote{Similar to Footnote \ref{ft:inducePin-1},
we do not yet know whether it is always possible to induce a unique 2d $\mathrm{Pin}^-$ structure on
$(c\cup\text{PD}'(Aa))$ on any $M^4$ given any $\mathrm{Pin}^-$ $M^5$. However, we claim that it is possible
to find suitable some $M^4$ so that the 2d $\mathrm{Pin}^-$ is induced, thus in this sense the ${\text{ABK}}(c\cup\text{PD}'(Aa))$ is defined.
\label{ft:inducePin-2} }
\end{enumerate}
In summary, we have constructed the 4d $\mathbb{Z}_2$-gauge theory in \eq{eq:5dSPT-4dTQFT-explicit}
preserving the $({\rm Spin} \times_{\mathbb{Z}_2^F} {\mathbb{Z}_{4,X}})$-structure, namely it is a $({\rm Spin} \times_{\mathbb{Z}_2^F} {\mathbb{Z}_{4,X}})$-{symmetric} TQFT
but with ${\upnu_{\text{even}}}=2 \in \mathbb{Z}_{16}$ anomaly.
This can be used to compensate the anomaly
$\upnu =- N_{\text{generation}}\mod 16$
with $N_{\text{generation}}=2$, two generations of missing right-handed neutrinos.
We could not however directly construct the symmetric gapped TQFT for
$\upnu$ is odd (thus symmetric TQFTs not possible for $N_{\text{generation}}=1$ or $3$), due to the obstruction found in
\cite{Hsieh2018ifc1808.02881, Cordova1912.13069}.
However, we can still compensate the
$\upnu =- N_{\text{generation}} = -3\mod 16$ anomaly by:
\begin{enumerate}
\item
Appending the
symmetric gapped 4d TQFT with the anomaly ${\upnu}=2 \in \mathbb{Z}_{16}$,
plus an additional one generation of right-handed neutrino $\nu_R$
with the anomaly ${\upnu}=1 \in \mathbb{Z}_{16}$.
This single right-handed neutrino $\nu_R$ can pair with any of the three generation of
left-handed neutrino $\nu_L$ to give a Dirac mass via the Yukawa-Higgs term.
The quantum interference between the three generation of neutrinos (three $\nu_L$ and one $\nu_R$)
in the TQFT background (of ${\upnu}=2$) may give rise to
the neutrino oscillations observed in the phenomenology (see \Sec{sec:neutrino-oscillations}).
\item Appending the
symmetric gapped 4d TQFT with the anomaly ${\upnu}=2 \in \mathbb{Z}_{16}$,
plus an additional 5d iTQFT in an extra dimension
with the anomaly ${\upnu}=1 \in \mathbb{Z}_{16}$.
\end{enumerate}
Either of the above combinations still complete a scenario for the full anomaly matching.
By the same logic, we can also propose many more other possible full anomaly matching scenarios,
such as the linear combinations of scenarios given in \Sec{sec:matchtheanomaly}.
\subsection{Gapping Neutrinos: Dirac mass vs Majorana mass vs Topological mass}
\label{sec:HiddenTopologicalSectorsTopologicalmass}
It should be clear that the 4d TQFT in \Sec{sec:Hidden4dnon-invertiblenon-abelianTQFT}
with an index $\upnu_{\text{even}} \mod 16$ in \Eq{eq:Ngeneration}
\emph{cannot} be directly access from gapping the free fermion theory ($\upnu_{\text{even}} \mod 16$ number of
Weyl fermions or Majorana fermions).
In order to access the 4d TQFT, we have to:
\begin{enumerate}
\item Go to a higher-energy mother effective field theory (EFT) at a deeper UV (ultraviolet), such that the mother EFT can
land at IR (infrared) with two different types of low energy physics: 4d free fermions on one side, and the 4d TQFT on another side.
\item Access via the dual fermionic vortex zero mode bound state condensation
and go through a 4d topological quantum phase transition \cite{subirsachdev2011book, XGWen2004Book}
or ``4d mirror symmetry \cite{Hori2003VafaMirrorsymmetry}'' (i.e., a duality of QFTs in the version of 4d spacetime). We comment more about the 4d duality
in \Sec{sec:4dduality}.
\end{enumerate}
Before then, let us clarify the meanings of Mass or Energy Gap.
The masses for {\bf left-handed neutrinos} (thus also right-handed antineutrinos) are experimentally estimated to be very small, nearly million times smaller than the electron mass \cite{PDG2008}
\begin{eqnarray}
m_{\nu_e}, m_{\nu_\mu}, m_{\nu_\tau} < 1 e{\rm V} \ll m_{e} \simeq 0.51 {\rm M}e{\rm V}.
\end{eqnarray}
More precisely,
the flavor states (${\nu_e}$, ${\nu_\mu}$, ${\nu_\tau}$)
are superpositions of the different mass eigenstates (${\nu_1}$, ${\nu_2}$, ${\nu_3}$). So a flavor state should be weighed and averaged over different masses of the different mass eigenstates.
The current experiments show that the sum of the masses of the three neutrinos should also be below about $1e$V.
These mass bounds hold for neutrinos regardless being Dirac fermion or Majorana fermion particles.
These mass bounds only apply to left-handed neutrinos (thus also right-handed antineutrinos, or its own anti-particle if a neutrino is a Majorana fermion).
But we do not yet know about the mass, or mass bound, or the existence of
{\bf right-handed neutrinos} (called {\bf sterile neutrinos} because they are in the trivial representation $({\bf 1},{\bf 1},0)_R$
in \Eq{eq:nuR-rep} and do not interact with any of three SM forces).
They also can be regarded as the left-handed $({\bf 1},{\bf 1},0)_L$ with complex conjugation on the representation.
Let us overview the two known mass generation mechanism and propose a new third mechanism (a topological mass or energy gap mechanism to gap the
neutrinos):\footnote{By a {\bf gap},
we mean giving a {\bf mass gap} or an {\bf energy gap} to the system's energy spectrum,
which is the eigenenergy values of the quantum Hamiltonian of the system. (E.g. Solving a big matrix eigenvalues in the linear algebra.)
The mass gap usually already assumes the free particle descriptions exist. However, in the interacting systems such as
CFT or strongly-correlated many-body quantum matter, we may not always be able to find a suitable free particle description.
In the later case, we can still have an energy gap from the interacting or many-body physics.
The energy gap can be a generalization of the particle mass gap for the interacting or many-body systems.
After all, Einstein had told us long ago the energy is the mass: $\Delta E \sim (\Delta m) c^2$.}
\begin{enumerate} [leftmargin=.mm, label=\textcolor{blue}{\arabic*}., ref={\arabic*}]
\item \label{Diracmass}
{\bf Dirac mass} mechanism \cite{Dirac19281, Dirac19282}: Dirac mass and Anderson-Higgs mechanism are believed to give other Standard Model particles their masses.
\item \label{Majoranamass}
{\bf Majorana mass} mechanism \cite{Majorana1937}: This requires that the neutrino and antineutrino to be the same particle.
If a neutrino is indeed a Majorana fermion, then lepton-number violating processes such as neutrino-less double beta decay would be allowed.
The neutrino-less double beta decay is not allowed if neutrinos are Dirac fermions.
\item \label{Topologicalmass}
{\bf Topological mass} mechanism (or {\bf Topological Energy Gap} mechanism):
This describes the gapped excitations of interacting systems
(typical many-body quantum matter). The underlying low energy theory
can be either invertible TQFT (invertible or SPT phases), or non-invertible TQFT (topological order).
The energy gap is induced by the interaction mechanism
\cite{FidkowskifSPT1, FidkowskifSPT2, Wen2013ppa1305.1045, Wang2013ytaJW1307.7480,
You2014oaaYouBenTovXu1402.4151, YX14124784, BenTov2014eea1412.0154,
BenTov2015graZee1505.04312, Wang2018ugfJW1807.05998, WangWen2018cai1809.11171}.
The typical quadratic mass term breaks the chiral symmetry;
but the interaction-induced energy gap can be fully symmetric without any chiral symmetry breaking \cite{RazamatTong2009.05037}.
\end{enumerate}
Is the mass or energy gap necessarily associated with a single particle picture (only defined if there is a free limit)?
Or can the mass gap associated with the whole quantum system in a strongly correlated interacting limit?
It turns out that a free particle mass may be suitable for {\bf Dirac mass} \ref{Diracmass} and {\bf Majorana mass} \ref{Majoranamass} mechanisms;
but a free particle mass may not be appropriate for the
{\bf Topological mass} \ref{Topologicalmass} mechanism.
\subsubsection{Dirac Mass or Majorana Mass}
\label{sec:MajoranaMassorDiracMass}
In more details, we write
Dirac mass (\ref{Diracmass}) and Majorana mass (\ref{Majoranamass}) as follows.
%
By the 2-component Weyl spinor notations in 4d,
the undotted indices are for left-handed spinor fields, the
dotted indices are for right-handed spinor fields.
We write the 4-component spinors (such as Dirac or Majorana) in terms of two
2-component Weyl spinors:\footnote{Here we use the 2-component Weyl spinor notation.
Not to confuse the ${``\mu''}$ in muon neutrino index ${\nu_{\mu}}$ with the spacetime index $\mu$.
For two general left-handed Weyl spinors, say $\chi$ and $\chi'$ (where each component is a Grassman number with anti-commutation properties
$\chi_{\al } \chi'_{\bt} = - \chi'_{\bt} \chi_{\al }$), we have
$$
\chi \chi' \equiv \chi^{\al } \chi'_{\al} \equiv \epsilon^{\al \bt} \chi_{\bt} \chi'_{\al} = -\epsilon^{ \bt \al} \chi_{\bt} \chi'_{\al} =
- (\chi^{\rm T})_{\bt}(\hspace{1pt}\mathrm{i}\hspace{1pt} \sigma^2)_{\bt \al }\chi'_{\al }
=- \epsilon^{\al \bt} \chi'_{\al } \chi_{\bt}= \epsilon^{ \bt \al} \chi'_{\al} \chi_{\bt} \equiv \chi' \chi
, \quad\quad
$$
$$\hspace{-14mm}
( \chi \chi' )^\dagger =(\chi^{\al } \chi'_{\al})^\dagger=(\chi'_{\al})^\dagger(\chi^{\al } )^\dagger=(\chi'^\dagger_{\dot \al})(\chi^{\dagger\dot \al })
\equiv\chi'^\dagger \chi^\dagger
= (\chi'^{\dagger}_{\dot\al})(\hspace{1pt}\mathrm{i}\hspace{1pt} \sigma^2)_{\dot \al \dot \bt}\chi^\dagger_{\dot \bt }
= (\chi'^{\dagger}_{\dot\al})(\hspace{1pt}\mathrm{i}\hspace{1pt} \sigma^2)_{ \dot \al \dot\bt}\chi^*_{\dot \bt }
=(\chi^\dagger_{\dot \al})(\chi'^{\dagger\dot \al }) \equiv\chi^\dagger \chi'^\dagger =( \chi' \chi)^\dagger.
$$
We use the convention $ \epsilon^{\al \bt} = \epsilon^{\bt \al} = -\epsilon_{\al \bt}= \epsilon_{\bt \al} =
\epsilon^{\dal \dot{\beta}} = \epsilon^{\dot{\beta} \dal} = -\epsilon_{\dal \dot{\beta}}= \epsilon_{\dot{\beta} \dal} =(\hspace{1pt}\mathrm{i}\hspace{1pt} \sigma^2)_{\al \bt} =
\begin{pmatrix}
0 & 1 \\
- 1& 0
\end{pmatrix}_{\al \bt}$,
also
$\chi^{\al } \equiv \epsilon^{\al \bt} \chi_{\bt}$ and
$(\chi^{\al } )^\dagger=(\chi^{\dagger\dot \al })=(\epsilon^{\dot \al \bt} \chi_{\bt})^\dagger=(\chi_{\bt})^\dagger \epsilon^{\dot \al\bt}
=(\hspace{1pt}\mathrm{i}\hspace{1pt} \sigma^2)_{\dot \al \dot\bt}\chi^\dagger_{\dot \bt }
=(\hspace{1pt}\mathrm{i}\hspace{1pt} \sigma^2)_{\dot \al \dot \bt}\chi^*_{\dot \bt }$.
The
Hermitian conjugate switches the two SU(2) Lie algebras in the Lie algebra of the Lorentz group
${\rm Spin}(3,1)=\rm{SL}(2,{\mathbb{C}})$.
${\rm Spin}(4)={\rm Spin}(3) \times {\rm Spin}(3)={\rm SU}(2)_L \times {\rm SU}(2)_R$.
The $h.c.$ is Hermitian complex conjugate.
}
\begin{eqnarray}\label{eq:Diracmass}
&\hspace{-10mm}
\text{Dirac spinor }
\Psi_{\rm D}\equiv \begin{pmatrix}
\chi_{\al}\\
{\zeta}_{}^{\dagger \dot\al}
\end{pmatrix}.
\quad
\overline{\Psi}_{\rm D}\equiv({\zeta}_{}^{ \al}, \chi_{\dot \al}^\dagger ).
&
\text{Dirac mass: }
\overline{\Psi}_{\rm D} {\Psi}_{\rm D}= {\zeta}_{}^{ \al} \chi_{\al} +\chi_{\dot \al}^\dagger {\zeta}_{}^{\dagger \dot\al}
=\zeta \chi + ( \zeta \chi )^\dagger.
\quad\quad\\
&\label{eq:Majoranamass}
\hspace{-10mm}
\text{Majorana spinor }
\Psi_{\rm M}\equiv\begin{pmatrix}
\chi_{\al}\\
\chi_{}^{\dagger \dot\al}
\end{pmatrix}.
\quad
\overline{\Psi}_{\rm M}\equiv({ \chi}_{}^{ \al}, \chi_{\dot \al}^\dagger ).
&\text{Majorana mass: }
\overline{\Psi}_{\rm M} {\Psi}_{\rm M}=
\chi \chi + h.c.=\chi \chi + ( \chi \chi )^\dagger.
\quad\quad\quad
\end{eqnarray}
If the neutrino mass is generated by Dirac mass (\ref{Diracmass}) like other Standard Model fermions,
then the mass term needs to be an SU(2) singlet. The left-handed neutrino from the SU(2) doublet ${l_L}_{\nu}$
would have the Yukawa interactions with
the SU(2) doublet Higgs $\phi_H$ and an SU(2) singlet
right-handed neutrino $\chi_\nu$.
Another mechanism is that neutrino mass can be generated by a Majorana mass (\ref{Majoranamass}), which would require the neutrino and antineutrino to be the same particle (particle as its anti-particle).
The anomaly $\upnu =- N_{\text{generation}}\mod 16$ in \Eq{eq:Ngeneration} dictates that in the free fermion limit,
we need to add $\upnu =N_{\text{generation}} = 3\mod 16$ right-handed spinors to SM to match the anomaly.\footnote{This
anomaly $\upnu =- N_{\text{generation}}\mod 16$ in \Eq{eq:Ngeneration}
also rules out many BSM introducing more than $N_{\text{generation}} = 3 \mod 16$ of sterile neutrinos or right-handed Weyl spinors.}
Denote the right-handed spinor of sterile neutrinos of 3 generations $(\nu_e, \nu_{\mu}, \nu_{\tau})$ as
$(\chi_{\nu_e},
\chi_{\nu_\mu},
\chi_{\nu_\tau})$,
we can write down the free
quadratic non-interacting action where $\bar\sigma \equiv (1, \vec{\sigma})$,
\begin{eqnarray}
\chi_{\nu_e}^{\dagger}\hspace{1pt}\mathrm{i}\hspace{1pt} \bar\sigma^{\mu} {\partial}_{\mu}
\chi_{\nu_e}
+
\chi_{\nu_\mu}^{\dagger}\hspace{1pt}\mathrm{i}\hspace{1pt} \bar\sigma^{\mu} {\partial}_{\mu}
\chi_{\nu_\mu}
+
\chi_{\nu_\tau}^{\dagger}\hspace{1pt}\mathrm{i}\hspace{1pt} \bar\sigma^{\mu} {\partial}_{\mu}
\chi_{\nu_\tau}+\frac{1}{2}
\bigg((\chi_{\nu_e},
\chi_{\nu_\mu},
\chi_{\nu_\tau})
\begin{pmatrix}
&&\\
&M_{} &\\
&&
\end{pmatrix}
\begin{pmatrix}
\chi_{\nu_e}\\
\chi_{\nu_\mu}\\
\chi_{\nu_\tau}
\end{pmatrix}
+ h.c.
\bigg).
\end{eqnarray}
Along the rank-3 mass matrix $M$, its diagonalized elements represent the Majorana mass \Eq{eq:Majoranamass}.
More generally, we can pair
$\upnu =N_{\text{generation}} = 3$ right-handed Weyl spinors (sterile neutrinos) with each other and with the
3 left-handed Weyl spinors via another
free non-interacting matrix term action:
\begin{eqnarray}
\frac{1}{2}
\bigg(
(
\Big({l_L}_{\nu_e},
{l_L}_{\nu_\mu},
{l_L}_{\nu_\tau}\Big)\frac{\<\phi_H\>}{|{\<\phi_H\>}|},
\chi_{\nu_e}^\dagger,
\chi_{\nu_\mu}^\dagger,
\chi_{\nu_\tau}^\dagger)\quad
\Biggl(\mkern-5mu
\begin{tikzpicture}[baseline=-.65ex]
\matrix[
matrix of math nodes,
column sep=1ex,
] (m)
{
\;\;\;\;0_{\text{\;\;\;}} & M_{\text{Dirac}} \\
M_{\text{Dirac}} & M_{\text{S}}\\
};
\draw[dashed]
([xshift=0.55ex] m-1-1.north east) -- ([xshift=0.ex]m-2-1.south east);
\draw[dashed]
([yshift=-0.2ex]m-1-1.south west) -- ([yshift=-0.19ex] m-1-2.south east);
\node[above,text depth=1pt] at (m-1-1.north) {$\scriptstyle 3$};
\node[above,text depth=1pt] at (m-1-2.north) {$\scriptstyle 3$};
\node[left,overlay] at ([xshift=-1.8ex]m-1-1.west) {$\scriptstyle 3$};
\node[left,overlay] at ([xshift=-1.2ex]m-2-1.west) {$\scriptstyle 3$};
\end{tikzpicture}\mkern-5mu
\Biggr)
\begin{pmatrix}
{l_L}_{\nu_e} \frac{\<\phi_H\>}{|{\<\phi_H\>}|} \\
{l_L}_{\nu_\mu}\frac{\<\phi_H\>}{|{\<\phi_H\>}|}\\
{l_L}_{\nu_\tau}\frac{\<\phi_H\>}{|{\<\phi_H\>}|}\\
\chi_{\nu_e}^\dagger\\
\chi_{\nu_\mu}^\dagger\\
\chi_{\nu_\tau}^\dagger
\end{pmatrix} + h.c.
\bigg).
\end{eqnarray}
There is a rank-6 mass matrix.
Here the Dirac mass scale, for example, is given by Higgs vev $M_{\text{Dirac}} \sim |{\<\phi_H\>}|$.
We remind that the flavor states can be superpositions of the different mass eigenstates.
In above, we rewrite the three generations of SU(2) doublet ${\bf 2}$ of left-handed neutrino
$\Big({l_L}_{\nu_e},
{l_L}_{\nu_\mu},
{l_L}_{\nu_\tau}\Big)$ pair
with the SU(2) doublet ${\bf 2}$ Higgs
as $\Big({l_L}_{\nu_e},
{l_L}_{\nu_\mu},
{l_L}_{\nu_\tau}\Big)\frac{\<\phi_H\>}{|{\<\phi_H\>}|}$.
The usual seesaw mechanism \cite{Minkowski1977seesaw}
sets the scale of $|M_{\text{S}}| \gg |M_{\text{Dirac}}|$.
So the three mass eigenstates have mass $\simeq \frac{|M_{\text{Dirac}}|^2}{|M_{\text{S}}|} \ll |M_{\text{Dirac}}|$ (for the observed 3 neutrinos much smaller other
Dirac mass of M$e$V or G$e$V scales),
while the other three mass eigenstates have mass $\simeq {|M_{\text{S}}|}$ which can set to be the GUT scale (thus too heavy yet to be detected).
We should emphasize that by having {Dirac mass or Majorana mass} to any of the Weyl fermion spinors,
this quadratic mass would break the $\mathbb{Z}_{4,X}$ symmetry (\Eq{eq:nuRsym} and \Eq{eq:nuLsym} assigned to the complex Weyl fermions).
More precisely, as explained in \Sec{sec:matchtheanomaly},
the Dirac mass term preserves the $\mathbb{Z}_{4,X}$ symmetry, but the $\mathbb{Z}_{4,X}$ symmetry is \emph{spontaneously} broken by the electroweak Higgs condensate ${\<\phi_H\>} \neq 0$.
In contrast, the Majorana mass term \emph{explicitly} breaks the $\mathbb{Z}_{4,X}$ symmetry.
Is it necessary to break $\mathbb{Z}_{4,X}$ symmetry in order to saturate the anomaly by a gapped theory?
No, we do not have to break the $\mathbb{Z}_{4,X}$ symmetry if we introduce
a Topological Mass/Energy Gap for Weyl fermions, see
\Sec{sec:TopologicalMassandTopologicalEnergyGap}.
\subsubsection{Topological Mass and Topological Energy Gap}
\label{sec:TopologicalMassandTopologicalEnergyGap}
Now we introduce {\bf Topological mass} and {\bf Topological energy gap} mechanism.
The concept of free particle mass may not be appropriate here,
so we should digest the mechanism from the interacting theory viewpoint.
We may colloquially call any of the following mass gaps as the topological energy gap:
\begin{enumerate}[leftmargin=2.5mm, label=\textcolor{blue}{(\arabic*)}., ref={(\arabic*)}]
\item \label{appr:Interaction}
{\bf Interaction-induced mass or interaction-induced energy gap}:
This idea has been used to classify the topological phases of interacting quantum matter.
Given the symmetry $G$ (including
the ${G_{\text{spacetime} }}$ and ${{G}_{\text{internal}} }$) of the system,
are the two ground states (i.e., two vevs) deformable to each other by preserving the symmetry $G$?
This is the key question for the community studying the classification of Symmetry-Protected Topological state (SPTs).\\[2mm]
$\bullet$ Fidkowski-Kitaev \cite{FidkowskifSPT1, FidkowskifSPT2} had shown
that 1+1d $\mathbb{Z}$ classification \cite{RyuSPT, Kitaevperiod, Wen1111.6341}
of $T^2=+1$ topological superconductor with $\mathbb{Z}_2^T \times \mathbb{Z}_2^F$ symmetry can be reduced to
$\mathbb{Z}_8$ class. Fidkowski-Kitaev may be the first example of showing the interaction can produce the energy gap
between 8 Majorana fermions in 1+1d without breaking the original symmetry. (See recent discussions along QFT reviewed in \cite{BenTov2015graZee1505.04312, Tong20198Majorana1906.07199}.)\\[2mm]
$\bullet$ Kitaev \cite{Kitaev2015} and Fidkowski-Chen-Vishwanath (FCV) \cite{Fidkowski1305.5851} suggested that
3+1d $\mathbb{Z}$ classification of $T^2=(-1)^F$ topological superconductor (TSC) with $\mathbb{Z}_4^T \supset \mathbb{Z}_2^F$ symmetry can be reduced to
$\mathbb{Z}_{16}$ class. This implies that the 16 number of 2+1d Majorana fermions on the boundary of 3+1d TSC can open up an energy gap
without introducing any free quadratic mass: neither Majorana nor Dirac masses.\\[2mm]
$\bullet$ Wen \cite{Wen2013ppa1305.1045}, and the author and Wen \cite{Wang2013ytaJW1307.7480, Wang2018ugfJW1807.05998, WangWen2018cai1809.11171},
suggest that all the $G$-anomaly-free gapless theory can be fully gapped
without breaking $G$-symmetry.
This idea includes introducing a random disorder new Higgs field \cite{Wen2013ppa1305.1045};
or introducing the proper-designed direct \emph{non-perturbative interactions} \cite{Wang2013ytaJW1307.7480, WangWen2018cai1809.11171},
(which are usually irrelevant or marginal operator deformations viewed from IR QFTs)
or introducing a \emph{symmetric gapped boundary} \cite{Wang2018ugfJW1807.05998}.
Many of such examples are applied to construct chiral fermion or chiral gauge theories on the lattice.
This approach is pursued also by You-BenTov-Xu \cite{You2014oaaYouBenTovXu1402.4151, YX14124784}.
BenTov and Zee \cite{BenTov2014eea1412.0154, BenTov2015graZee1505.04312} names this mechanism as Kitaev-Wen mechanism,
or \emph{the mass without mass}.
In the lattice gauge theory, gapping the mirror fermions dated back to Eichten-Preskill \cite{Eichten1985ftPreskill1986};
a recent work by Kikukawa also suggested the same conclusion that the SO(10) GUT mirror fermion can be gapped \cite{Kikukawa2017ngf1710.11618}.
This deformation of $G$-anomaly-free theory is consistent with Seiberg's conjecture that the
deformation classes of QFTs is constrained by symmetry and anomaly \cite{NSeiberg-Strings-2019-talk}.
\item \label{appr:VortexCondensation}
{\bf Vortex condensation}: This is an approach commonly used in condensed matter literature for 2+1d strongly-correlated
systems. The idea is that the symmetry-breaking defects (such as vortices) may trap the zero modes and which carry nontrivial quantum number.
The question is to find which number of vortex zero modes with what kind of symmetry assignment,
can the vortices be proliferated to restore the broken global symmetry ---
this would drive the quantum phase transition between the symmetry-breaking phase and the symmetry-restoring phase.
This approach, called the vortex condensation, has been used to construct 2+1d surface topological orders, see the condensed matter review \cite{Senthil1405.4015}.
\item \label{appr:SymmetryExtension}
{\bf Symmetry-extension and symmetry-preserving gapped topological order/TQFT}:
As mentioned briefly in \Sec{sec:sym-ext},
the symmetry-extension mechanism \cite{Wang2017locWWW1705.06728}
and
higher-symmetry extension generalization \cite{Wan2018djlW2.1812.11955}
are a rather exotic mechanism.
We trivialize the anomaly and introduce a gapped phase,
not by breaking $G$ to its subgroup $G_{\text{sub}} \subset G$, but by extending it to a larger group
$G_{\text{total}}$ which can be regarded as a fibration of the original group $G$ as a quotient group.
%
(See the down-to-earth lattice constructions in any dimension \cite{Wang2017locWWW1705.06728} and in 1+1d bulk \cite{Prakash2018ugo1804.11236, PrakashJW2011.13921})
%
This mechanism is a useful intermediate step stone,
to construct a \emph{symmetry-preserving TQFT}, via gauging the extended-symmetry \cite{Wang2017locWWW1705.06728}.
This approach is applicable to bosonic systems \cite{Wang2017locWWW1705.06728, Wang1801.05416} and fermionic systems \cite{GuoJW1812.11959, Kobayashi2019lep1905.05391, PrakashJW2011.13921} in any dimension.
In contrast to the well-known gapped phase saturate the anomaly via {\bf symmetry breaking} (either global symmetry breaking or Anderson-Higgs mechanism for gauge theory), this approach is based on {\bf symmetry extension} (thus beyond symmetry breaking and Anderson-Higgs mechanism).
\end{enumerate}
The {\bf Topological mass} and {\bf Topological energy gap} mechanism (including
\ref{appr:Interaction}, \ref{appr:VortexCondensation} and \ref{appr:SymmetryExtension}), in fact,
is obviously beyond the familiar Dirac, Majorana mass, and seesaw mechanism \cite{Minkowski1977seesaw}.
In a colloquial sense, we do not have a Higgs field $\phi_H$ breaking the symmetry and gives vev $ \<\phi_H\> \neq 0$.
In certain case, we can consider a new randomly disordered Higgs field $\phi_h$ such that
\cite{Wen2013ppa1305.1045, WangWen2018cai1809.11171}
\begin{eqnarray}
\<\phi_h\> = 0, \quad\quad \<|\phi_h|^2\> \neq 0.
\end{eqnarray}
So Topological mass/energy gap is a quantum behavior beyond the mean field quadratic semiclassical theory, beyond Anderson-Higgs mechanism, and beyond Landau-Ginzburg symmetry-breaking paradigm.
\subsubsection{4d duality for Weyl fermions and ``mirror symmetry''}
\label{sec:4dduality}
Before apply the Topological Mass from \Sec{sec:TopologicalMassandTopologicalEnergyGap} to neutrino physics furthur,
we like to introduce a potential helpful supersymmetry (SUSY) duality in 4d known as Seiberg duality \cite{Seiberg19949411149} studied in ${\cal N}=1$ theory and supersymmetric quantum chromodynamics (SQCD). Seiberg duality is an ${\cal N}=1$
electric magnetic duality in SUSY non-abelian gauge theories with the weak-strong duality.
On the left-hand side of the duality may have
quarks and gluons;
on the right-hand side dual theory,
they become the solitons (such as nonabelian magnetic monopoles) of the elementary fields.
When the left-hand side theory is Higgsed by an expectation value of a squark, the right-hand side dual theory's is confined.
Massless glueballs, baryons, and magnetic monopoles in the confined strongly coupled description in the left-hand side theory
becomes some weakly coupled elementary quarks in the right-hand side dual Higgs description.
Schematically, there is an IR duality between left-hand side (LHS) and right-hand side (RHS) under a renormalization group (RG) flow
for ${\cal N}=1$:
\begin{multline}
\label{eq:4dSeibergdual}
{\rm SU}(N_c) \text{ gauge theory with $N_f$ chiral and antichiral multiplets $Q$, $\tilde{Q}$ in color fundamental $N_c$, $\bar{N}_c$} \\
\xleftrightarrow{\text{IR duality}}
{\rm SU}(N_f-N_c) \text{ gauge theory with $q$ and $\tilde{q}$ in color fundamental $N_f-N_c$, $\overline{N_f-N_c}$ and meson $M$}.
\end{multline}
Include the representations of chiral multiplet/superfields, we have the relations:
\begin{eqnarray}
\begin{tabular}{ccc }
\hline
4d ${\cal N}=1$ Seiberg duality & LHS: {SQCD} & RHS: dual theory \\
\hline
\hline
{color gauge group} & SU($N_c$) & SU($N_f-N_c$) \\
\hline\\[-4mm]
Same global internal symmetries &
\multicolumn{2}{c}{${\rm SU}(N_f)_L \times {\rm SU}(N_f)_R \times {\rm U}(1)_B \times {\rm U}(1)_R$}\\
\hline
$\begin{array}{c}
\text{Chiral multiplet/superfields:}\\
\text{Representation {\bf R}}
\end{array}$
& $Q: (N_f, 1, 1, \frac{N_f-N_c}{N_f})$ & $q: ( 1, N_f, \frac{-1}{N_f-N_c}, \frac{N_c}{N_f})$ \\
& $\tilde{Q}: ( 1, \bar{N}_f, -1, \frac{N_f-N_c}{N_f})$ & $\tilde{q}: ( \bar{N}_f, 1, \frac{1}{N_f-N_c}, \frac{N_c}{N_f})$ \\
& & $M: ( {N}_f, \bar{N}_f, 0, \frac{2(N_f-N_c)}{N_f})$ \\
\hline
\end{tabular}
\end{eqnarray}
We are particularly interested in the case when the RHS flows to free Weyl spinors, which means that
we can choose as simple as $N_c=2$ and $N_f=3$, and ${\cal N}=1$:
\begin{eqnarray} \label{eq:15}
{\rm SU}(N_c=2) \text{ with $N_f=3$ chiral and anti-chiral multiplets $Q$ and $\tilde{Q}$}
\xleftrightarrow{\text{IR duality}}
15 \text{ Weyl spinors}.
\end{eqnarray}
For Weyl spinor counting
we have $N_f=3$ chiral multiplets and $N_f=3$ anti-chiral multiplets,
each is the fundamental or anti fundamental ${\bf 2}$ or $\bar{\bf 2}$ of SU(2),
thus they contribute $2 \cdot 2 \cdot 3 = 12$ Weyl spinors.
There is also a vector multiplet which sits at the adjoint representation ${\bf 3}$ of SU(2),
this contributes another $3$ Weyl spinors.
So in total we have $2 \cdot 2 \cdot 3 +3 = 15$ Weyl spinors.
The $N_f=N_c +1=3$ is interesting because it sits at the lower boundary ($3 N_c/2 = N_f$)
of $3 N_c/2 < N_f < 3 N_c$, where the origin of the moduli space is an interacting CFT and
non-abelian Coulomb phase.
Also this case we have $N_f=N_c +1$ thus two moduli spaces are identical but
the interpretations of the singularity at the origin are different --- massless particles can be regarded as, either
strongly coupled mesons and baryons on LHS, or weakly interacting or free quarks on RHS.
This duality helps as $ 15 \mod 16 = -1 \mod 16$ so to cancel the anomaly
\Eq{eq:nu-1} as $\upnu = -N_{\nu_R} \mod 16= -1 \mod 16 $ in one generation of SM.
One way to simplify 15 Weyl spinors to
1 Weyl spinor on RHS would be that adding 1 Weyl spinor on both sides in the trivial representation,
and adding \emph{nonperturbative deformations} to gap the RHS completely:
\begin{multline} \label{eq:4ddual}
{\rm SU}(N_c=2) \text{ with $N_f=3$ chiral and anti-chiral multiplets $Q$ and $\tilde{Q}$ + deformations} \\
\xleftrightarrow{\text{IR duality}}
(\text{gapping 16 Weyl spinors via nonperturbative interacting deformations}) \\
+ (- 1) \text{ Weyl spinors (in the complex conjugate representation)}.
\end{multline}
We leave details of this construction in an upcoming work \cite{toappear}.
In the following subsections, we can argue several phenomenon of gapping Weyl spinors based on this proposed duality \Eq{eq:4ddual}.
The hope is that
we can access the 4d TQFT
from the dual fermionic vortex zero mode bound state condensation via the topological quantum phase transition \cite{subirsachdev2011book, XGWen2004Book}
or ``4d mirror symmetry \cite{Hori2003VafaMirrorsymmetry} description'' (i.e., a duality of QFTs in the version of 4d spacetime).
In a general colloquial sense, this duality is also related to the particle-vortex duality \cite{Peskin19771978, DasguptaHalperin1981}.
The renown supersymmetric version of duality in 3+1d includes, for example, the ${\cal N}=2$ Seiberg-Witten theory \cite{SeibergWitten1994rs9407087},
and the ${\cal N}=1$ Seiberg duality \cite{Seiberg19949411149}, and many other theories (see a review \cite{Intriligator1995auSeiberg9509066}).
A recent ongoing pursuit, along
a fermionic non-supersymmetric version of particle-vortex duality in 3+1d,
includes
\cite{Anber2018iof1805.12290, Cordova2018acb1806.09592DumitrescuClay, BiSenthil1808.07465,
Wan2018djlW2.1812.11955, BiLakeSenthil1910.12856, WangYouZheng1910.14664}.\footnote{There were also a fermionic version in 2+1d of particle-vortex duality proposed in \cite{Son20151502.03446, Wang2015Senthil1505.05141, Metlitski20151505.05142} and formalized in \cite{SeibergWitten20161606.01989, KarchTong20161606.01893, MuruganNastase20161606.01912}. They have achieved great success on understanding condensed matter phenomena in 2+1d.
They may provide insightful guidelines to the 3+1d, namely 4d duality construction.}
\subsubsection{Gapping 3, 2, or 1 Weyl spinors / sterile right-handed neutrinos + extra sectors}
\label{sec:Gapping3Weylspinors}
In \Sec{sec:MajoranaMassorDiracMass}, the 4d anomaly $\upnu =- N_{\text{generation}}\mod 16$ in \Eq{eq:Ngeneration} dictates that in the free particle limit,
we need to add $\upnu =N_{\text{generation}} = 3\mod 16$ right-handed spinors to match the anomaly. But as mentioned, in \Sec{sec:Hidden4dnon-invertiblenon-abelianTQFT},
there is an obstruction
to obtain a Topological Mass: there does \emph{not} exist any 4d noninvertible TQFT to saturate the $\upnu =N_{\text{generation}} = 3\mod 16$ anomaly,
shown by \cite{Hsieh2018ifc1808.02881,Cordova1912.13069}.
{In order to gap 3 Weyl spinors in a topological fashion, we can obtain a 4d $\upnu_{\text{even}}=2$-TQFT,
with a left-over right-handed neutrino (which introduces Dirac/Majorana masses):
\begin{multline} \label{eq:TQCP-1}
\text{4d $(\upnu = 3)$ massless right-handed neutrinos}\\
\xrightarrow{\text{4d topological quantum phase transition}}\\
\text{4d $(\upnu_{\text{even}}=2)$-TQFT + } \\
\text{1 right-handed neutrino gives Dirac/Majorana masses} \\
\text{to the other 3 of SM's left-handed neutrinos.}
\end{multline}
We can go through a quantum phase transition
by using 3 times of \Eq{eq:4ddual}.
We may access the 4d noninvertible TQFT from the LHS.
On the other hand, we do not gain the expectation values for the quadratic mass to two generations of fermions:
\begin{eqnarray} \label{eq:interaction-mass}
\newcommand*{\multicolumn{1}{c|}{0}}{\multicolumn{1}{c|}{0}}
\begin{pmatrix}
M_{11} & M_{12}& M_{13}\\
M_{21} &M_{22} & M_{23}\\
M_{31} & M_{32}& M_{33}
\end{pmatrix}
\xrightarrow{\text{mean field vev}}
\text{quadratic mass term vev} \propto
\begin{pmatrix}
0 &\multicolumn{1}{c|}{0} &0\\
0 & \multicolumn{1}{c|}{0} & 0 \\\cline{1-3}
0 & \multicolumn{1}{c|}{0} & \# \<\phi_H^{\nu}\> \text{ or } M_{\text{Majorana}}
\end{pmatrix}.
\end{eqnarray}
Schematically, on the left-hand side of \eq{eq:interaction-mass},
we may include the quadratic fermion pairing term
with the $M_{ij}$ stands for possible highly-order interaction (or momentum-dependent pairing) terms.
On the right-hand side, we ask what are the quadratic mass at the mean-field level.
Two generations have the zero mean-field mass. But one generation has left with a Dirac mass ($\# \<\phi_H^{\nu}\>$) or Majorana mass ($M_{\text{Majorana}}$).}
If we aim to preserve the $\mathbb{Z}_{4,X}$ symmetry at the Lagrangian level, then we should choose the
Dirac mass term instead of Majorana mass term.
However, as mentioned in \Sec{sec:matchtheanomaly}, Yukawa-Higgs-Dirac mass term spontaneously breaks the $\mathbb{Z}_{4,X}$,
while the Majorana mass explicitly breaks the $\mathbb{Z}_{4,X}$. Neither scenario involving Dirac or Majorana mass preserves the $\mathbb{Z}_{4,X}$ at the dynamical level.
Nonetheless, we can preserve the $\mathbb{Z}_{4,X}$ at the dynamical level if we trade a single generation of 4d right-handed neutrino
to a 5d $(\upnu = 1)$-iTQFT. So we can consider the $\mathbb{Z}_{4,X}$-symmetric topological quantum phase transition:
\begin{multline} \label{eq:TQCP-2}
\text{4d $(\upnu = 2)$ massless right-handed neutrinos + 5d $(\upnu = 1)$-iTQFT}\\
\xrightarrow{\text{4d topological quantum phase transition}}\\
\text{4d $(\upnu_{\text{even}}=2)$-TQFT + }
\text{5d $(\upnu = 1)$-iTQFT.}
\end{multline}
To complete a story for high-energy phenomenology,
we can add
``(4d SM with $(N_{\text{generation}} = 3) \times$ 15 Weyl fermions)''
on both sides of topological quantum phase transitions in
\eq{eq:TQCP-1} and
\eq{eq:TQCP-2}, in order to apply to SM and BSM physics.
\section{Ultra Unification: Grand Unification $+$ Topological Force and Matter}
\label{sec:UU}
In the previous sections, we had shown that in order to match a nonperturbative $\mathbb{Z}_{16}$ global anomaly (\Sec{sec:SMandGUTdiscrete}) but still preserve\footnote{By
preserving $\mathbb{Z}_{4,{X}}$, we mean the $\mathbb{Z}_{4,{X}}$-symmetry is preserved at some higher energy scale above the electroweak scale. Of course,
the energy scale lower such as the Higgs scale 125 G$e$V, the usual Yukawa-Higgs-Dirac mass terms would spontaneously break the $\mathbb{Z}_{4,{X}}$.}
$\mathbb{Z}_{4,{X}} $ of SM and SU(5) GUT Georgi-Glashow (GG) model with only 15 Weyl fermions per generation,
we can introduce a new hidden gapped sector appending to the SM and GUT: \\[-10mm]
\begin{itemize}
\item
Topological Mass/Energy Gap (\Sec{sec:Hidden4dnon-invertiblenon-abelianTQFT} and \Sec{sec:HiddenTopologicalSectorsTopologicalmass}) to gap the 16th Weyl fermions (right-handed neutrinos). The outcome low energy gives rise to a 4d non-invertible TQFT.
\item 5d invertible TQFT (iTQFT) with one extra dimension (\Sec{sec:5dSPT}, known as 5d SPTs in a quantum condensed matter analogy).
\end{itemize}
Overall, we consider the combinations of the above two solutions plus additional scenarios enlisted in \Sec{sec:matchtheanomaly}.
In either cases,
we require new hidden gapped topological sectors beyond SM and GUT. We name the
unification including SM, Grand Unification, plus additional unitary gapped topological sectors with Topological Force and Matter
as {\bf Ultra Unification}.
\subsection{4d-5d Theory: Quantum Gravity and Topological Gravity coupled to TQF
}
\label{sec:gravity}
\begin{enumerate}[leftmargin=0mm, label=\textcolor{blue}{\Roman*]}:, ref={\Roman*]}]
\item
The combinations of solutions from
\Sec{sec:matchtheanomaly}, including adding 4d non-invertible TQFT and 5d invertible TQFT,
mean that we can propose a new schematic partition function / path integral, generalizing \Eq{eq:UU-GUT} to
\begin{multline} \label{eq:UU-GUT-1}
{\bf Z}_{\overset{\text{5d-iTQFT/}}{\text{4d-QFT}}}[{\cal A}_{{\mathbb{Z}_4}}]
=
\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16} \cdot\upnu_{\rm{5d}} \cdot \eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}})) \rvert_{M^5}) \cdot
\int [{\cal D} {\psi}] [{\cal D}\bar{\psi}][{\cal D} A][{\cal D} \phi_H][{\cal D} {a}] [{\cal D} b]\cdots \\
\exp( \hspace{1pt}\mathrm{i}\hspace{1pt} \left. S_{\text{4d-SM/GUT}}[\psi, \bar{\psi}, A, \phi_H, \dots, {\cal A}_{\mathbb{Z}_4}] \right \rvert_{M^4}
+ \hspace{1pt}\mathrm{i}\hspace{1pt} \left. S_{\text{4d-TQFT}}^{(\upnu_{\rm{4d}})} [a, b, \dots , {\cal A}_{{\mathbb{Z}_2}}] \right\rvert_{M^4}
)
\bigg\rvert_{\upnu_{\rm{5d}} - \upnu_{\rm{4d}}=-N_{\text{generation}}}.
\quad\quad
\end{multline}
Partition function depends on the (${\cal A}_{{\mathbb{Z}_2}} = {\cal A}_{{\mathbb{Z}_4}}$ mod 2) background fields.
The anomaly \Eq{eq:Ngeneration} is now matched by:
\begin{eqnarray} \label{eq:Ngeneration-nu45}
\upnu = \upnu_{\rm{5d}} - \upnu_{\rm{4d}}=-N_{\text{generation}} \mod 16.
\end{eqnarray}
The $S_{\text{4d-SM/GUT}}$ is the 4d SM or GUT action.
The $\psi, \bar{\psi}, A, \phi_H$ are SM and GUT quantum fields,
where $\psi, \bar{\psi}$ are the 15 or 16 Weyl spinor fermion fields,
the $A$ are 12 or 24 gauge bosons given by gauge group Lie algebra generators,
and $\phi_H$ is the electroweak Higgs (we can also add GUT Higgs).
The $S_{\text{4d-TQFT}}^{(\upnu_{\rm{4d}})}$ is a 4d noninvertible TQFT outlined in \Sec{sec:Hidden4dnon-invertiblenon-abelianTQFT}
(in particular, $\upnu_{\rm{4d}}= \upnu_{\rm{even}}= \pm 2, \dots$).
The ${a}$ and ${b}$ (and possibly others fields from fermionic invariants such as the ABK invariant)
are TQFT gauge fields (locally differential 1-form $a$ and 2-form $b$, as anti-symmetric tensor gauge connections).
\item Many different perspectives guide to an understanding that at high enough energy scale, every global symmetry should be either
\emph{dynamically gauged} or
\emph{explicitly broken} \cite{Misner1957mtWheeler, Polchinski2003bq0304042, Banks2010zn1011.5120, Harlow2018tngOoguri1810.05338}.
By every global symmetry, we include the internal symmetry ${{G}_{\text{internal}} }$,
the spacetime ${{G}_{\text{spacetime}} }$, the fermion parity $\mathbb{Z}_2^F$, the time reversal symmetry $\mathbb{Z}_2^T$ or $\mathbb{Z}_4^{T} \supset \mathbb{Z}_2^F$, and so on.
Below we discuss the consequences on gauging some of the symmetries on the boundary or in the bulk.
Our notations on the bulk and boundary symmetries associated with the group extension
follow the conventions of \cite{Wang2017locWWW1705.06728, Wang1801.05416}:
\begin{enumerate}[leftmargin=1.5mm]
\item {\bf Gauge the $\mathbb{Z}_2$ on the 4d}:
We can gauge $a$ $\in$ ${\operatorname{H}^1(M, \mathbb{Z}_{2})}$ on the boundary
which sits at the normal subgroup $[\mathbb{Z}_2]$
of the symmetry extension
\eq{eq:Z2-SpinxZ4}:
\begin{eqnarray} \label{eq:Z2-SpinxZ4-Z2}
1 \to [\mathbb{Z}_2]_{\text{bdry gauge}} \to {\rm Spin} \times {\mathbb{Z}_{4,X}} \to ({\rm Spin} \times_{\mathbb{Z}_2^F} {\mathbb{Z}_{4,X}})_{\text{Bulk/bdry sym}} \to 1,
\end{eqnarray}
while the 5d bulk and 4d boundary both keep the global symmetry $ {\rm Spin} \times_{\mathbb{Z}_2^F} {\mathbb{Z}_{4,X}}$.
In this case, there is a 4d $\mathbb{Z}_2$ gauge theory living on the boundary of a 5d iTQFT.
\item
{\bf Gauge the $\frac{\mathbb{Z}_{4,X}}{\mathbb{Z}_2^F}$ in 5d} and {\bf gauge the $\mathbb{Z}_{4,{X}}$ on the 4d}:
We can also dynamically gauge ${\cal A}_{{\mathbb{Z}_2}}$ $\in$ ${\operatorname{H}^1(M, \mathbb{Z}_{4,X}/\mathbb{Z}_2^F)}$ which sits at the normal subgroup $\mathbb{Z}_2$
of the following short exact sequence respect to the bulk:
\begin{eqnarray} \label{eq:Z2-Spin}
1 \to [\frac{\mathbb{Z}_{4,X}}{\mathbb{Z}_2^F}]_{\text{Bulk gauge}} \to {\rm Spin} \times_{\mathbb{Z}_2^F} {\mathbb{Z}_{4,X}}\to ({\rm Spin})_{\text{Bulk sym}} \to 1.
\end{eqnarray}
{{If we also gauge the $[\frac{\mathbb{Z}_{4,X}}{\mathbb{Z}_2^F}]$ out of the 5d bulk symmetry ${\rm Spin} \times_{\mathbb{Z}_2^F} {\mathbb{Z}_{4,X}}$,
then the 5d bulk short-range entangled gapped iTQFT can become a
5d bulk long-range entangled gapped $\mathbb{Z}_2$ gauge TQFT with a
fermionic rotational ${\rm Spin}(d)$ symmetry definable on manifolds with a ${\rm Spin}$ structure.}}
This is due to gauging the normal subgroup of the above group extension \eq{eq:Z2-Spin}.
Once the $[\frac{\mathbb{Z}_{4,X}}{\mathbb{Z}_2^F}]$ is dynamically gauged in 5d,
it also implies that the $[\mathbb{Z}_{4,{X}}]$ should be dynamically gauged on the 4d boundary theory.
(It is natural that the $\mathbb{Z}_{4,{X}}$ is gauged above some GUT scale,
especially given that the $\mathbb{Z}_{4,{X}}$ is a gauge subgroup $\mathbb{Z}_{4,{X}} = Z({\rm Spin}(10)) \subset {\rm Spin}(10)$ of the SO(10) GUT, see more discussions in \cite{JW2008.06499}.)
By doing so, we need to sum over
${\cal A}_{{\mathbb{Z}_4}}$ $\in$ ${\operatorname{H}^1(M, \mathbb{Z}_{4,X})}$ for a given spacetime topology and geometry.
We can gauge the normal subgroup $\mathbb{Z}_{4,X}$ out of the 4d boundary symmetry ${\rm Spin} \times {\mathbb{Z}_{4,X}}$ in
the group extension,
\begin{eqnarray} \label{eq:Z4-Spin}
1 \to [{\mathbb{Z}_{4,X}}]_{\text{bdry gauge}} \to {\rm Spin} \times_{} {\mathbb{Z}_{4,X}} \to ({\rm Spin})_{\text{Bulk/bdry sym}} \to 1.
\end{eqnarray}
\item
{\bf Gauge the ${\mathbb{Z}_2^F} \times \frac{\mathbb{Z}_{4,X}}{\mathbb{Z}_2^F}$ in 5d} and {\bf gauge the ${\mathbb{Z}_2^F} \times \mathbb{Z}_{4,{X}}$ on the 4d}:
Other than gauging ${\cal A}_{{\mathbb{Z}_2}}$ $\in$ ${\operatorname{H}^1(M, \mathbb{Z}_{4,X}/\mathbb{Z}_2^F)}$, we can also gauge the fermion parity $\mathbb{Z}_2^F$.
So we gauge the normal subgroup out of the following short exact sequence respect to the bulk:
\begin{eqnarray} \label{eq:Z2-SO}
1 \to [{\mathbb{Z}_2^F} \times \frac{\mathbb{Z}_{4,X}}{\mathbb{Z}_2^F}]_{\text{Bulk gauge}} \to {\rm Spin} \times_{\mathbb{Z}_2^F} {\mathbb{Z}_{4,X}}\to ({\rm SO})_{\text{Bulk sym}} \to 1.
\end{eqnarray}
Then the 5d bulk becomes a
long-range entangled gapped TQFT with a bosonic
${\rm SO}(d=5)$ symmetry defined on 5d manifolds with a special orthogonal group ${\rm SO}$ structure.
For the boundary theory,
we gauge the normal subgroup out of the following short exact sequence respect to the boundary:
\begin{eqnarray} \label{eq:Z4-SO}
1 \to [{\mathbb{Z}_2^F} \times {\mathbb{Z}_{4,X}}]_{\text{bdry gauge}} \to {\rm Spin} \times_{} {\mathbb{Z}_{4,X}} \to ({\rm SO})_{\text{Bulk/bdry sym}} \to 1.
\end{eqnarray}
By gauging the fermion parity $\mathbb{Z}_2^F$, this means the UV completion at a higher energy scale should be a bosonic system,
presumably emerged from a local tensor product Hilbert space ---
this is analogous to the ``It from Qubit'' and the quantum condensed matter view
\cite{WangWen2018cai1809.11171}.
\item {\bf Gauge the $\mathbb{Z}_{4,{X}}$ with the spacetime Spin and SO symmetry}:
We can also sum over ${\cal A}_{{\mathbb{Z}_4}}$ $\in$ ${\operatorname{H}^1(M, \mathbb{Z}_{4,X})}$ together with
all spacetime ${{M \in \forall\; \text{topology},{\forall \;\text{geometry}}}}$.
Let us consider the case the spacetime is a closed
$M=M^5$ without boundary.
But summing over all spacetime $M$ certainly diverges, one needs to make sense of the partition function
by regularization. To deal with such a path integral is a challenging problem similar to topological quantum gravity.
We are not be able to solve it for now. In summary, a schematic path integral says:
\begin{multline} \label{eq:5dQG}
\sum_{{M \in \forall\; \text{topology},{\forall \;\text{geometry}}}}
\sum_{{\cal A}_{{\mathbb{Z}_2}} \in \operatorname{H}^1(M, \mathbb{Z}_{4,X}/\mathbb{Z}_2^F)}
\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16} \cdot\upnu_{\rm{5d}} \cdot \eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}})) \rvert_{M^5}) \\
\xrightarrow{\;\text{regularize}\;}
\sum_{\underset{{\cal A}_{{\mathbb{Z}_2}} \in \operatorname{H}^1(M, \mathbb{Z}_{4,X}/\mathbb{Z}_2^F)}{{M \in \; \{\text{topo/geo}\},}}}
\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16} \cdot\upnu_{\rm{5d}} \cdot \eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}})) \rvert_{M^5}).
\end{multline}
A few thoughts can help to attack this challenging problem on regularization of summing over spacetimes.
\begin{itemize}
\item
In 2d topological gravity, we can sum over conformal structures.
This can be a finite dimensional integral for a fixed topology.
\item One can simplify the problem to sum over different topologies given by the 5d $\eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}}))$ invariant
with $M^5$ of ${{\rm Spin} \times_{\mathbb{Z}_2} \mathbb{Z}_4}= {{\rm Spin} \times_{\mathbb{Z}_2^F} \mathbb{Z}_{4,X}}$ structure
\item
Recent work by Dijkgraaf and Witten on 2d topological gravity provides a guide to the analogous 2d partition function \cite{Dijkgraaf2018Witten1804.03275} related to the 2d Arf invariant. The theory \Eq{eq:5dQG} may be regarded as a 5d gravity (dynamical, quantum, and topological gravity)
related to the 5d $\eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}}))$ invariant.
\end{itemize}
\end{enumerate}
\item Suppose we find a way to regularize the path integral,
then we can consider the theory \Eq{eq:5dQG} with boundary, where we can place the 4d SM theory. We thus rewrite
\Eq{eq:UU-GUT-1} as the fully gauged version of a schematic path integral:
\begin{multline} \label{eq:UU-GUT-2-grav}
{\bf Z}_{\overset{\text{5d-TQFT.QG/}}{\text{4d-QFT}}}
=
\sum_{\underset{{\cal A}_{{\mathbb{Z}_2}} \in \operatorname{H}^1(M, \mathbb{Z}_4/\mathbb{Z}_2^F)}{{M \in \{\text{topo/geo}\},}}}
\exp(\frac{2\pi \hspace{1pt}\mathrm{i}\hspace{1pt}}{16} \cdot\upnu_{\rm{5d}} \cdot \eta(\text{PD}({\cal A}_{{\mathbb{Z}_2}})) \rvert_{M^5}) \cdot
\int [\cD{\cal A}_{{\mathbb{Z}_4}}]
[{\cal D} {\psi}] [{\cal D}\bar{\psi}][{\cal D} A][{\cal D} \phi_H][{\cal D} a ] [{\cal D} b]\cdots \\
\exp( \hspace{1pt}\mathrm{i}\hspace{1pt} \left. S_{\text{4d-SM/GUT}}[\psi, \bar{\psi}, A, \phi_H, \dots, {\cal A}_{\mathbb{Z}_4}] \right \rvert_{M^4}
+ \hspace{1pt}\mathrm{i}\hspace{1pt} \left. S_{\text{4d-TQFT}}^{(\upnu_{\rm{4d}})}[ a , b, \dots , {\cal A}_{\mathbb{Z}_2}] \right\rvert_{M^4}
)
\bigg\rvert_{\upnu_{\rm{5d}} - \upnu_{\rm{4d}}=-N_{\text{generation}}}.
\quad\quad
\end{multline}
Recall we gauge ${\cal A}_{{\mathbb{Z}_4}}$ as it is the ${\mathbb{Z}_{4,X}}$ gauge field and ${\cal A}_{{\mathbb{Z}_4}} \mod 2 = {\cal A}_{{\mathbb{Z}_2}}$.
This is a 4d and 5d coupled fully gauged quantum system ---
where the 4d has SM, GUT and noninvertible TQFT,
and the 5d can allow a certain gravity (dynamical, quantum and topological) coupled to the 5d TQFT.
The 1-form gauge field ${\cal A}_{{\mathbb{Z}_4}}$ (obtained by gauging the 0-form ${\mathbb{Z}_{4,X}}$ symmetry) can mediate between
the 5d bulk, and 4d boundary (the 5d bulk and the 4d SM and BSM gapped topological sectors).
\end{enumerate}
\subsection{Braiding statistics and link invariants in 4d or 5d: \newline
Quantum communication with gapped hidden sectors from our Standard Model} \label{subsec:braiding}
Follow \Sec{sec:gravity}, above a higher energy scale (way above SM and above SU(5) GUT but around SO(10) GUT),
the discrete $X=5({\bf B}- {\bf L})-4Y$ symmetry (the ${\mathbb{Z}_{4,X}}$ symmetry)
becomes dynamically gauged, with a dynamical gauge vector boson
mediator $X_g$, mathematically associated with a 1-cochain gauge field ${\cal A}_{{\mathbb{Z}_4}}$.\footnote{Again
this is a discrete $\mathbb{Z}_{4,X}$ or ${\rm U}(1)_X$ gauge theory's gauge boson $X_g$, different from the SU(5) GUT's continuous nonabelian Lie group's gauge boson X and Y.
We have been denoted $X_g$'s 1-cochain gauge field as
${\cal A}_{{\mathbb{Z}_2}}$ for the case of $\frac{\mathbb{Z}_{4,X}}{\mathbb{Z}_2^F}$ symmetry.
We have been denoted $X_g$'s 1-cochain gauge field as
${\cal A}_{{\mathbb{Z}_4}}$ for the case of $\mathbb{Z}_{4,X}$ symmetry.}
So the entangled Universe in 4d and 5d may be described by \Eq{eq:UU-GUT-2-grav}.
Once the $\mathbb{Z}_{4,X}$ is dynamically gauged,
the 4d and 5d gauged TQFT sectors become noninvertible TQFTs.
In fact such a medium can be regarded as some version of {\bf topological quantum computer} \cite{Kitaev2003, Kitaev2006} with intrinsic topological orders \cite{Wen2016ddy1610.03911}.
There is an immediate question: Do we have any way to communicate or interact with
objects living in 4d or 5d TQFTs? Could we communicate with
those objects within SM particles that we human beings and creatures are made of?
Another way to phrase this question is: Can we detect the Topological Force?
For example in the TQFTs of \eq{eq:5dSPT-4dTQFT-explicit}, we have some gauge invariant extended line and surface operators:
\begin{eqnarray}
\oint a+\frac{1}{2} {\cal A}_{{\mathbb{Z}_2}},\quad\quad
\oiint b+\frac{1}{2}{\cal A}_{{\mathbb{Z}_2}}^2,
\end{eqnarray}
invariant under the gauge transformation:
${\cal A}_{{\mathbb{Z}_2}}\to {\cal A}_{{\mathbb{Z}_2}}+\delta\lambda$, \;
$a\to a+\lambda\delta\lambda$,
and $b\to b+\lambda\delta\lambda\delta\lambda$, where $\lambda$ is a 0-cochain valued in $\operatorname{H}^1(M,\mathbb{Z}_2)$.
Since the extended ${a}$ line and ${b}$ surface operators do not end on the SM particles, nor do those $a$ and $b$ extended operators
interact with SM gauge forces,
it seems that we SM particles cannot communicate with the TQFT
at least n\"aively.
{At least, we SM particles cannot directly interact with the 4d TQFT sector via the SM gauge forces}.
But in fact, the answer is {\bf YES}, {we SM particles can indeed communicate with the 4d TQFT sector, only via the ${\cal A}_{{\mathbb{Z}_2}}$ gauge field.}
Moreover, {we SM particles can also communicate with the 5d TQFT sector via the ${\cal A}_{{\mathbb{Z}_4}}$ gauge field.}
The idea is that the TQFT path integral \eq{eq:5dSPT-4dTQFT-explicit} schematically
\begin{eqnarray} \label{eq:Z4dnabTQFTgaugeZ4}
\int [\cD{\cal A}_{{\mathbb{Z}_2}} ] \int [\cD b] [\cD a] \;
\exp{\Big(\hspace{1pt}\mathrm{i}\hspace{1pt} \pi \int_{ M^4} b (\delta a+{\cal A}_{{\mathbb{Z}_2}}^2)
+\frac{2\pi\hspace{1pt}\mathrm{i}\hspace{1pt}}{8}\text{ABK}(c\cup \text{PD}'({\cal A}_{{\mathbb{Z}_2}} a)) \Big)}.
\end{eqnarray}
contains a term with a schematic coupling:
\begin{eqnarray} \label{eq:ABKAa}
\pi \int (b \smile \delta a + \delta {\cal B} \smile {\cal A}_{{\mathbb{Z}_2}} + b \smile {\cal A}_{{\mathbb{Z}_2}} \smile {\cal A}_{{\mathbb{Z}_2}}) + \frac{1}{4} (\text{ABK} \smile {\cal A}_{{\mathbb{Z}_2}} \smile a).
\end{eqnarray}
For the purpose which will become clear soon,
we introduce a dynamical 2-cochain ${\cal B}$ field which is natural if the ${\cal A}_{{\mathbb{Z}_2}}$ is dynamically gauged, with
${\cal B} \smile \delta {\cal A}_{{\mathbb{Z}_2}} = \delta {\cal B} \smile {\cal A}_{{\mathbb{Z}_2}}$ on a closed 4-manifold.
Again the schematic cup product term
$(\text{ABK} \smile {\cal A}_{{\mathbb{Z}_2}} \smile a)$,
coupling between a fermionic invariant ABK and the cohomology class $Aa$, is formally $\text{ABK}(c\cup \text{PD}'({\cal A}_{{\mathbb{Z}_2}} a))$ defined in \Sec{sec:Z16-nu=2-TQFT}.
We denote $\text{PD}'$ as the Poincar\'e dual on the 4-manifold $M^4= \partial M^5$.
So what are the experimental designs for us to
communicate with ``TQFT sectors''?\footnote{In fact,
previous works on braiding statistics and link invariants,
such as the multi-string 3-loop braiding process
and other exotic braiding process (\cite{WangLevin1403.7437, Jiang1404.1062, Wang1404.7854} and \cite{CWangMLevin1412.1781,
Putrov2016qdo1612.09298PWY, 1602.05951WWY, Wang2019diz1901.11537}),
provide helpful guidelines to these topological link designs, shown in \Fig{fig:BSMlink1}.
}
The action term $(\text{ABK} \smile {\cal A}_{{\mathbb{Z}_2}} \smile a)$ in \Eq{eq:ABKAa}
prompts us to design a configuration in \Fig{fig:BSMlink1}.
To motivate the discussion,
the fermionic topological term
$(\text{ABK} \smile {\cal A}_{{\mathbb{Z}_2}} \smile a)
\sim
\text{ABK}(\text{PD}'({\cal A}_{{\mathbb{Z}_2}} a))$
is similar to
the fermionic topological invariant
$(\text{Arf} \smile a_1 \smile a_2)=
\text{ABK}(\text{PD}'(a_1 \smile a_2) )
=
\text{ABK}(\text{PD}'(a_1) \cap \text{PD}'(a_2) )
$
studied in
\cite{GuoJW1812.11959}.
The $(\text{Arf} \smile a_1 \smile a_2)$
corresponds to a $\mathbb{Z}_2$ torsion class cobordism invariant in the subgroup of
$\Omega^{{\rm Spin} \times (\mathbb{Z}_4)^2}_{4} = (\mathbb{Z}_4)^2 \times \mathbb{Z}_2$.
The $a_1$ and $a_2$ are the mod 2 classes of 1-cochain of two distinct $\mathbb{Z}_4$ gauge fields.
Only this $\mathbb{Z}_2$ generator is intrinsically fermionic.
\Refe{GuoJW1812.11959} shows that the Sato-Levine invariant \cite{sato1984cobordisms},
$\text{Arf}(V_{(1)} \cup V_{(2)})$, can detect a certain link configuration.
The link configuration $L$ is a disjoint union of two surface links $\Sigma_{(1)}^2$ and $\Sigma_{(2)}^2$,
so $L = \Sigma_{(1)}^2 \sqcup \Sigma_{(2)}^2$. The ``$\sqcup$'' means the disjoint union.
The $L$ is a semi-boundary link \cite{sato1984cobordisms},
so that, by definition, there exists Seifert volumes $V_{(1)}^3$ and $V_{(2)}^3$ for each link component $\partial V_{(j)}^3= \Sigma_{(j)}^2$,
such that the intersections $V_{(1)}^3 \cap \Sigma_{(2)}^2 = \emptyset$ and
$\Sigma_{(1)}^2 \cap V_{(2)}^3 = \emptyset$ are empty.
To design a detectable link configuration for
the fermionic topological term
$(\text{ABK} \smile {\cal A}_{{\mathbb{Z}_2}} \smile a)
\sim
\text{ABK}(\text{PD}'({\cal A}_{{\mathbb{Z}_2}} a))$,
we start from finding appropriate two 2-surfaces $ \Sigma_{(1)}^2$ and $ \Sigma_{(2)}^2$ to be linked,
see \Fig{fig:BSMlink1}.
We have a 2-cochain ${\cal B}$ field dual to the 1-cochain ${\cal A}_{{\mathbb{Z}_2}}$ which is dynamically gauged.
In this case, the 2-surface $\Sigma_{(1)}^2$ fundamental class can be paired with a 2-surface operator from the cohomology class ${\cal B}$ field,
$
\oiint_{L_1} {\cal B} +\dots.
$
And another 2-surface $\Sigma_{(2)}^2$ fundamental class can be paired with a 2-surface operator from the cohomology class $b$ field,
$
\oiint_{L_2} b +\dots
$
with extra $\dots$ terms to make the 2-surface operator gauge invariant \cite{toappear}.
\begin{figure}[t!]
\centering
(a) \includegraphics[width=1.24in]{whitehead.pdf}
(b) \includegraphics[width=4.in]{BSM-link-Sato-Levine.pdf}
\caption{(a) The Whitehead link formed by a disjoint union of two-component 1-loops, $\ell = \ell_1 \sqcup \ell_2$, is detectable by the Sato-Levine invariant. (See related QFTs explored in \cite{Putrov2016qdo1612.09298PWY,GuoJW1812.11959}.)
(b)
A schematic nontrivial link configuration in a 4d spacetime (inside the large black circle)
that can carry an odd $\mathbb{Z}_{4,X}$ charge (also an odd $\frac{\mathbb{Z}_{4,X}}{\mathbb{Z}_2^F}$ charge)
measured by a codimension 1 operator $\star {\cal J}_{{\mathbb{Z}_2}}$ (from the 5d bulk perspective) or $\star {\cal J}_{{\mathbb{Z}_4}}$ (from the 4d boundary perspective).
The $\mathbb{Z}_{4,X}$ charge is trapped non-locally within the surface link.
The surface link $L = \Sigma_{(1)}^2 \sqcup \Sigma_{(2)}^2$ contains two surfaces obtained from a spun version of Whitehead link.
Let a $D^3$ ball contain the Whitehead link $\ell$.
The $\Sigma_{(1)}^2$ rotates the $\ell_1$ along an ${S^1}'$ of $D^3 \times {S^1}'$ outside the $D^3$.
The $\Sigma_{(2)}^2$ rotates the $\ell_2$ along the blue dashed arrow around the $\ell_1$ circle within the $D^3$.
From a 4d boundary theory perspective, we have the ${\cal A}_{{\mathbb{Z}_4}} $ along a 1-line $I^1$ (drawn in the red dashed)
Poincar\'e dual (PD) to the $\star {\cal J}_{{\mathbb{Z}_4}}$.
From a 5d bulk theory perspective, we have the ${\cal A}_{{\mathbb{Z}_2}} $ along a 1-line $I^1$ PD to the $\star {\cal J}_{{\mathbb{Z}_2}}$.
}
\label{fig:BSMlink1}
\end{figure}
Let us explain how to design a semi-boundary two-component link $L = \Sigma_{(1)}^2 \sqcup \Sigma_{(2)}^2$ (in \Fig{fig:BSMlink1} (b),
this link configuration is only an example, there could be other configurations trapping also a $\mathbb{Z}_{4,X}$ charge):
\begin{enumerate}[leftmargin=2.5mm]
\item
First, we can start from thinking of a Whitehead link $\ell = \ell_1 \sqcup \ell_2$ in a 3-ball $D^3$
shown in \Fig{fig:BSMlink1} (a).
In \Fig{fig:BSMlink1} (a), the $\ell_1$ and $\ell_2$ are both $S^1$ circles linked in a $D^3$.
\item Then, we make the $\ell_1$ circle spun around another ${S^1}'$ circle along the ${S^1}'$ of $D^3 \times {S^1}'$ outside of the $D^3$ sphere.
(So the readers can imagine this ${S^1}'$ rotation of $D^3$ going outside of the paper.)
If $M^4$ is a sphere $S^4$, it can be chosen as $S^4= (D^3 \times {S^1}') \cup ( S^2 \times D^2 )$, gluing the $(D^3 \times {S^1}')$ with another $( S^2 \times D^2 )$
along their common boundary $( S^2 \times {S^1}')$.
So this spun $\ell_1$ gives us a 2-torus that we call this 2-surface $ \Sigma_{(1)}^2 \cong \ell_1 \times {S^1}'$.
We place a cohomology class ${\cal B}$ field on the $\ell_1$.
\item We also make the $\ell_2$ rotating along the blue dashed arrow, circling around the $\ell_1$ component.
We call this 2-surface $ \Sigma_{(2)}^2 \cong \ell_2 \times S^1$.
We place a cohomology class $b$ field on the $\ell_2$.
\item This semi-boundary two-component 2-surface link $L = \Sigma_{(1)}^2 \sqcup \Sigma_{(2)}^2$ in \Fig{fig:BSMlink1} (b), with ${\cal B}$ and $b$ field inserted,
is a nontrivial link configuration which detects the fermionic topological term
$(\text{ABK} \smile {\cal A}_{{\mathbb{Z}_2}} \smile a)$.
\item We can also replace the 2-cochain ${\cal B}$ field to its dual 1-cochain ${\cal A}_{{\mathbb{Z}_2}}$ field.
In the sense, we can interpret the coupling
$\pi \int {\cal A}_{{\mathbb{Z}_2}} \smile \Big( \delta {\cal B} + b \smile {\cal A}_{{\mathbb{Z}_2}} + \frac{1}{4} (\text{ABK} \smile a) + \dots \Big)$
as the ${\cal A}_{{\mathbb{Z}_2}}$ gauge field coupled to a current ${\cal J}_{{\mathbb{Z}_2}}$ (with a Hodge dual $\star$), say in a schematic differential form:
\begin{eqnarray} \label{eq:AZ2JZ2}
\int {\cal A}_{{\mathbb{Z}_2}} \wedge \star {\cal J}_{{\mathbb{Z}_2}} +\dots.
\end{eqnarray}
The ${\cal B}$ sits at the 2-surface $\Sigma_{(1)}^2$,
so the Hodge dual of the current
$ \star {\cal J}_{{\mathbb{Z}_2}} \propto \delta {\cal B} +\dots $ sits at the Seifert 3-volume $V_{(1)}^3$.
If so, we learn that the 1-cochain gauge field ${\cal A}_{{\mathbb{Z}_2}}$ can sit along a 1-line $I^1$,
Poincar\'e dual (PD) to the Seifert volume $V^3_{(1)}$,
drawn along the red dashed line $I^1$ in \Fig{fig:BSMlink1} (b).
\item The ${\cal A}_{{\mathbb{Z}_2}}$ gauge field in \eq{eq:AZ2JZ2} is actually the dynamical gauge field living in the 5d bulk TQFT perspective; in fact, on the 4d boundary,
the ${\cal A}_{{\mathbb{Z}_2}}$ gauge field and the $a$ gauge field can be combined together
to become a $\mathbb{Z}_4$ gauge field: ${\cal A}_{{\mathbb{Z}_4}}$.
The relation between
the boundary $a$ $\in$ $ {\operatorname{H}^1(M, \mathbb{Z}_{2})}$,
the boundary ${\cal A}_{{\mathbb{Z}_4}}$ $\in$ ${\operatorname{H}^1(M, \mathbb{Z}_{4,X})}$,
and
the bulk ${\cal A}_{{\mathbb{Z}_2}}$ $\in$ ${\operatorname{H}^1(M, \mathbb{Z}_{4,X}/\mathbb{Z}_2^F)}$ gauge fields
can be understood
by a group extension (following \eq{eq:Z2-Spin}
and \eq{eq:Z4-Spin}):
\begin{eqnarray}
0 \to [{\mathbb{Z}_2}]_{\text{bdry gauge $a$}} \to [{\mathbb{Z}_{4,X}}]_{\text{bdry gauge ${\cal A}_{{\mathbb{Z}_4}}$}} \to [\frac{\mathbb{Z}_{4,X}}{\mathbb{Z}_2^F}]_{\text{Bulk gauge ${\cal A}_{{\mathbb{Z}_2}}$}} \to 0.
\end{eqnarray}
Recall that ${\cal A}_{{\mathbb{Z}_2}} = {\cal A}_{{\mathbb{Z}_4}} \mod 2$, similarly ${\cal J}_{{\mathbb{Z}_2}} = {\cal J}_{{\mathbb{Z}_4}} \mod 2$.
Thus we have the coupling term \eqn{eq:AZ2JZ2} in 5d, but we have the following term on the 4d boundary:
\begin{eqnarray}
\int {\cal A}_{{\mathbb{Z}_4}} \wedge \star {\cal J}_{{\mathbb{Z}_4}} +\dots.
\end{eqnarray}
\item Due to \eq{eq:ABKAa},
we can view the spun version of Whitehead link as a source of
an odd $\mathbb{Z}_{4,X}$ charge (also an odd $\frac{\mathbb{Z}_{4,X}}{\mathbb{Z}_2^F}$ charge), wrapped inside a
current $\star {\cal J}_{{\mathbb{Z}_4}}$ in a 3-volume (or $\star {\cal J}_{{\mathbb{Z}_2}}$ in a 4-volume) \cite{GuoJW1812.11959, toappear}.
\end{enumerate}
In fact, the odd $\mathbb{Z}_{4,X}$
charge $q_X$ of the topological link means that the dynamically gauged Wilson line with a line integral
$\exp(\hspace{1pt}\mathrm{i}\hspace{1pt} q_X \int {\cal A}_{{\mathbb{Z}_4}})$
can mediate between a topological link \emph{on one end} (${\text{p}_1}$)
and any other object carrying a nontrivial $\mathbb{Z}_{4,X}$ charges \emph{on the other end} (${\text{p}_2}$).
So we only need to look for all SM and GUT particles carrying the $\mathbb{Z}_{4,X}$ charges (especially the odd $\mathbb{Z}_{4,X}$ charge).
We now look at \Table{table:SMfermion} and \ref{table:SMboson}, all the left-handed Weyl spinors carry $\mathbb{Z}_{4,X}$ charge $+1$ (and the
right-handed Weyl spinors carry a complex conjugate of $\mathbb{Z}_{4,X}$ charge $-1 = 3 \mod 4$).
The electroweak Higgs $\phi_H$ carries an even $\mathbb{Z}_{4,X}$ charge $2 \mod 4$.
So, yes, in fact, all SM Weyl spinors carry odd $\mathbb{Z}_{4,X}$ charge, thus, all SM Weyl spinors, say $\psi_{\text{SM}}$, can be the other end of the
line integral $\exp(\hspace{1pt}\mathrm{i}\hspace{1pt} q_X \int {\cal A}_{{\mathbb{Z}_4}}$) of
$\mathbb{Z}_{4,X}$. All the odd number of bound states (like proton and neutrons in our body, also electrons) presumably can carry
the odd $\mathbb{Z}_{4,X}$.
In summary, we may have a schematic communication between a SM particle $\psi_{\text{SM}}$ and a topological link via a line operator
\begin{eqnarray}
\exp(\hspace{1pt}\mathrm{i}\hspace{1pt} \pi \oiint b + \dots ) \bigg\vert_{{\text{a link $L$ formed around }\text{p}_1}}\cdot \exp(\hspace{1pt}\mathrm{i}\hspace{1pt} q_X \int_{\text{p}_1}^{\text{p}_2} {\cal A}_{{\mathbb{Z}_4}}) \cdot \psi_{\text{SM}}({\text{p}_2}),
\end{eqnarray}
where ${\cal A}_{{\mathbb{Z}_4}}$ is the discrete $\mathbb{Z}_{4,X}$ gauge field.
A candidate link $L$ is shown in \Fig{fig:BSMlink1} (b).
The $\mathbb{Z}_{4,X}$ discrete gauge boson ${\cal A}_{{\mathbb{Z}_4}}$
is the mediator of Topological Force.
Topological Force has effects of long-range remote interactions but only through braiding and fusion statistics, thus
Topological Force is presumably weaker than the GUT forces, strong, electroweak forces.
In summary, if either the 4d or 5d gapped topological sectors exist,
our SM world and the gapped topological sector
can be mediated and communicated by Topological Force.
\subsection{Topological Force as a new force?}
We had mentioned that under some assumptions
about the discrete $\mathbb{Z}_{4,X}$ symmetry unbroken at high energy and only 15 Weyl fermions per generation,
it is natural to include a nonperturbative topological sector of 4d TQFT, 5d iTQFT, or 5d TQFT, appending to the SM or the GUT.
The Topological Force derived here is not within three known Fundamental Forces or GUT forces.
So we may view the Topological Force as a new force.
%
Our model in \Sec{sec:gravity} suggests that perhaps the Topological Force can be a new force in the model \Eq{eq:UU-GUT-1}.
But the Topological Force may mix with the gravity (the Fourth force) in the model \Eq{eq:UU-GUT-2-grav}, when the $\mathbb{Z}_{4,X}$ is gauged and the
geometry/topology are summed but regularized.
It is not clear how the gravity in \Sec{sec:gravity} has anything to do with Einstein gravity.
It may be interesting to explore this direction in the future.
\subsection{Neutrinos}
\label{sec:neutrino-oscillations}
{\bf Neutrino oscillations} may also be explained if we consider the Majorana zero modes of the vortices in the 4d TQFT defects in \Sec{sec:HiddenTopologicalSectors}.
The left-handed neutrinos (confirmed in the experiments) are nearly gapless/ massless. When neutrinos traveling through the 4d TQFT defects,
we may observe nearly gapless left-handed neutrino flavor oscillations interfering with the Majorana zero modes trapped by the vortices in the 4d TQFT defects.
A possible scenario is to arrange an even $\upnu_{\rm{4d}}$ anomaly index for the 4d TQFT
$$
\upnu_{\rm{4d},\rm{even}}= \pm 2, \pm 4, \dots \in \mathbb{Z}_{16},
$$
so there is an additional internal rotational symmetry within the 4d TQFT.
For example, given a $\upnu_{\rm{4d},\rm{even}}= 2$ of 4d TQFT,
there is an additional U(1) internal rotational symmetry.\footnote{We emphasize this additional U(1) is an extra symmetry,
distinct from the $({\bf B}- {\bf L})$, the $X=5({\bf B}- {\bf L})-4Y$, the SM, or the GUT gauge groups.
In one lower dimension, for a 3d boundary of a 4d bulk iTQFT given by the 4d APS $\eta$ invariant of $\Omega_4^{\mathrm{Pin}^+}=\mathbb{Z}_{16}$,
\Refe{Metlitski20141406.3032} introduces a similar extra U(1) symmetry when $\upnu =\pm 2 \in \mathbb{Z}_{16}$ of $\Omega_4^{\mathrm{Pin}^+}$.}
We can imagine a mother EFT with a $\upnu_{\rm{4d},\rm{even}}= 2 \in \mathbb{Z}_{16}$ anomaly
as a deeper UV theory which lands to two possible IR theories:\\
$\bullet$ One side has two right-handed neutrinos, a free fermion theory with a $\upnu_{\rm{4d},\rm{even}}= 2 \in \mathbb{Z}_{16}$ anomaly.\\
$\bullet$ Another side has a 4d TQFT also with a $\upnu_{\rm{4d},\rm{even}}= 2 \in \mathbb{Z}_{16}$ anomaly.\\
As a dual description of TQFT on the free fermion side, we have
the internal rotational U(1) or O(2) symmetry realized as an internal flavor rotation between two right-handed neutrinos (as 3+1d Weyl or Majorana fermions).
The TQFT defects are U(1) symmetry breaking vortices (or vortex strings).
The vortices can trap Majorana zero modes below the vortex energy subgap.
\noindent
{\bf SM and GUT Models with right-handed neutrino, 4d TQFT, and 5d iTQFT}:
We may still need to introduce at least a right-handed neutrino in order to give a conventional Dirac mass to
the three left-handed neutrinos.
To match the $\upnu= -N_{\text{generation}}= -3 \in \mathbb{Z}_{16}$ anomaly,
we had considered the scenario in \eq{eq:TQCP-1}:
\begin{eqnarray} \label{eq:system-1}
\left\{
\begin{array}{l}
\text{$\bullet$ 4d SM or GUT with $\upnu = 3 \times 15$ Weyl fermions.}\\
\text{$\bullet$ 4d $(\upnu_{\text{even}}=2)$-TQFT.} \\
\text{$\bullet$ 4d $\upnu=1$ right-handed neutrino gives Dirac/Majorana masses} \\
\quad
\text{to the other 3 of SM's left-handed neutrinos.}
\end{array}
\right.
\end{eqnarray}
If we want to introduce also a 5d iTQFT but still keep a 4d TQFT and a right-handed neutrino (in order to give a conventional Dirac mass to
left-handed neutrinos), we can also propose another scenario:
\begin{eqnarray} \label{eq:system-2}
\left\{
\begin{array}{l}
\text{$\bullet$ 4d SM or GUT with $\upnu = 3 \times 15$ Weyl fermions.}\\
\text{$\bullet$ 4d $(\upnu_{\text{even}}=4)$-TQFT or $(\upnu_{\text{even}}=2+k)$-TQFT.} \\
\text{$\bullet$ 5d $(\upnu = -2)$-iTQFT or $(\upnu = -k)$-iTQFT.}\\
\text{$\bullet$ 4d $\upnu=1$ right-handed neutrino gives Dirac/Majorana masses} \\
\quad \text{to the other 3 of SM's left-handed neutrinos.}
\end{array}
\right.
\end{eqnarray}
We have a generic $k \in \mathbb{Z}_{\text{even}}$. The above models, \eq{eq:system-1} and \eq{eq:system-2}, all have a 0 mod 16 index, thus free from the $\mathbb{Z}_{16}$ anomaly.
\noindent
{\bf Right-handed sterile neutrinos are not sterile to $\mathbb{Z}_{4,X}$ gauge field}:
We should emphasize again that although the
right-handed neutrinos are sterile to SM forces,
they are \emph{not} sterile to the $\mathbb{Z}_{4,X}$ gauge field ${\cal A}_{{\mathbb{Z}_4}}$, because they carry the odd $\mathbb{Z}_{4,X}$ charge.
\subsection{Dark Matter}
{\bf Dark Matter}: In the heavy sector of TQFTs that we described in \Sec{sec:HiddenTopologicalSectors},
the surface $b$ is the worldsheet of heavy string excitations.
Those are new heavy objects whose mass is in the scale of a topological order energy gap, $\Delta_{\text{TQFT}}$,
possibly around the GUT scales (for example, suggested in \cite{JW2008.06499}, the $\Delta_{\text{TQFT}}$ can be in between the SU(5) GUT and the SO(10) GUT scales).
So these new heavy higher-dimensional extended objects may be a suitable candidate account for the abundant Dark Matter in the Unvierse.
\noindent
{\bf Dark Energy}: We have very little to say about the Dark Energy. Except that in the gravity version (sum over geometry and topology; dynamical, quantum, and topological gravity)
of partition function in \Eq{eq:UU-GUT-2-grav}, it may be worthwhile to investigate the underlying energy stored in this partition function.
It seems that 4d and 5d topological sectors can be very heavy and store a much huge amount of energy than what we knew of from our Standard Model world.
\section{Acknowledgements}
JW is grateful to his previous collaborators for fruitful past researches
as helpful precursors for the present work.
JW appreciates Email correspondences with Robert Gompf, Miguel Montero, Kantaro Ohmori, Pavel Putrov, Ryan Thorngren, Zheyan Wan, and Yunqin Zheng;
and the mental support
from Shing-Tung Yau.\footnote{Instead of writing or drawing an image of the author's mental feelings, a piece of
Johannes Brahms's music ``The Variations and Fugue on a Theme by Handel, Op. 24 (1861)'' may illuminate this well.}
JW thanks the participants of Quantum Matter in Mathematics and Physics program at
Harvard University CMSA for the enlightening atmosphere.
Part of this work is presented by JW in the workshop Lattice for Beyond the Standard Model physics 2019 at Syracuse University (May 2-3, 2019)
\cite{JW-BSM-2019-talk}, also at Higher Structures and Field Theory at Erwin Schr\"odinger Institute in Wien (August 4, 2020) \cite{ESIJWUltraUnification},
and at Harvard University Particle Physics Lunch (November 30, 2020) and String Lunch (December 4, 2020).
JW was supported by
NSF Grant PHY-1606531.
This work is also supported by
NSF Grant DMS-1607871 ``Analysis, Geometry and Mathematical Physics''
and Center for Mathematical Sciences and Applications at Harvard University.
\section{Bibliography}
\bibliographystyle{Yang-Mills}
|
\section{Introduction}
\label{sec:1}
Haze is a typical atmospheric phenomenon which occurs when the dust, smoke and other particles accumulate in relatively dry air. These particles absorb and scatter the light greatly, thus attenuating the scene radiance reflected from scene point and confuse it with the scattering light. Haze will lead to a decrease in visibility of the scene point and the images captured under this weather condition will become poor in contrast and lose the visual details. Many vision tasks such as object detection would suffer from performance degradation due to these terrible hazy images. Therefore, image dehazing, as a preprocessing step and visual enhancement technology, has been extensively researched and achieves remarkable performance~\citep{Tan, DehazeGAN, DehazeNet, DCPDN, AOD-Net, SFSU, CMAda, NHR}.
In recent, many researchers have shifted their focus to remove haze from a single image which is more promising but more challenging in practice since without any extra information beyond the observed image itself. A variety of methods have been proposed~\citep{DCP,CAP,GRM,FVR} by employing a widely recognized atmospheric scattering physical model~\citep{Physical-Model}, which could be roughly divided into two categorizes: prior- and learning-based methods.
\begin{figure}[!t]
\label{Figure:Introduction}
\begin{center}
\subfigure[]{
\label{Figure:First:Haze}
\includegraphics[scale=0.1]{./Fig/Fig.1-2/fig1_haze}
}
\subfigure[]{
\label{Figure:First:DehazeNet}
\includegraphics[scale=0.1]{./Fig/Fig.1-2/fig1_dehazenet}
}
\subfigure[]{
\label{Figure:First:YOLY}
\includegraphics[scale=0.1]{./Fig/Fig.1-2/fig1_dehaze}
}
\subfigure[]{
\label{Figure:First:T}
\includegraphics[scale=0.1]{./Fig/Fig.1-2/fig1_T}
}
\end{center}
\caption{A visual illustration of the proposed YOLY. a) a real world hazy image; b) the clean image recovered by DehazeNet; c) the clean image recovered by YOLY; (d) the transmission map estimated by YOLY. Note that DehazeNet is pretrained using a collection of hazy-clean image pairs as did in~\citep{DehazeGAN}, whereas our method obtains the result only using the observed hazy image itself. From the results, one could find that our method qualitatively performs better than DehazeNet. Zoom-in is recommended to see the detailed comparisons.}
\end{figure}
To be specific, prior-based methods employ some handcrafted priors derived from the intrinsic properties of image, such as texture, contrast, and chromatic aberration. For instance, \cite{DCP} observe there exists the dark channel in the local patches of the outdoor haze-free images and accordingly propose using such a dark channel prior (DCP) to estimate the transmission map and atmospheric light for reconstructing the clean image. With the assumption of the image depth is positively correlated to the difference between brightness and saturation, \cite{CAP} propose color attenuation prior (CAP) to estimate the transmission map. Although these methods have achieved remarkable performance, the dehazing quality heavily depends on the consistency between the adopted prior and the actual image properties.
To alleviate the dependence of priors, learning-based methods especially deep learning based models have recently attracted much attention and a lot of efforts have been devoted in recent~\citep{DehazeNet, AOD-Net, GridDehazeNet, DCPDN, EPDN, SUFSPSD, JTME, MSPPDN}. Different from prior-based methods, the parameters of atmospheric scattering model are learned from a large scale training dataset. For example, DehazeNet~\citep{DehazeNet} estimates the transmission map by utilizing a deep neural network trained on a large scale dataset with the supervision of the ground-truth transmission map. Although these learning methods have achieved state-of-the-art performance in image dehazing, almost all of them work in a supervised and trained manner. Namely, they usually require a large scale training dataset which is with some kind of ground truths (\textit{e.g.}, hazy-clean image pair). Once the conditions are unsatisfied, these learning-based methods would be failed.
In practice, however, it is daunting even impossible to collect a large scale dataset with the desirable ground truth due to the variation in scene and the other factors such as illumination. Therefore, most of methods resort to collect some clean images first and then synthesize the corresponding hazy images via the atmospheric scattering model with the handcrafted parameters. However, the synthesized hazy images are probably less informative and inconsistent with the real hazy images, which would lead to the domain shift issue when the model trained on the synthetic dataset is applied to the real-world hazy images. To address the above issues, it is highly expected to develop a novel deep neural network which could work in an unsupervised and untrained manner simultaneously, while achieving the promising performance. However, to the best of our knowledge, such a challenging task is less touched heretofore.
In this paper, we propose a novel neural network called you only look yourself (\textbf{YOLY}) which employs three joint subnetworks (\textit{i.e.}, J-Net, T-Net, and A-Net) to disentangle a given hazy image into three component layers, \textit{i.e.}, scene radiance layer, transmission map layer, and atmospheric light layer. After that, these three layers are further used to reconstruct the observed input. Thanks to the proposed novel objective function and network structure, YOLY performs image dehazing by only using the information contained in the observed single hazy image. In other words, the proposed network performs like a ``real'' single image dehazing method, which does not follow the conventional paradigm of training the neural network on an image set with some ground truths. The major advantages of our YOLY is that it could avoid the labor-intensive data collection and domain shift issue like existing deep learning based methods, while achieving promising results. Besides, benefiting from the disentanglement of scene radiance, transmission map, and atmospheric light, our method provides an effective way to synthesize new hazy images in a learning-based rather than handcrafted fashion, which could also be the first study afak.
To summarize, the contributions are given as follows:
\begin{itemize}
\item To the best of our knowledge, this work could be one of the first unsupervised and untrained neural networks for single image dehazing, which removes haze without training on a image collection and the ground-truth clean image pairs. To the best of knowledge, the closest work is DDIP~\citep{Double-DIP} which is with significantly difference with the proposed YOLY and the discussion on their distinction is presented in Section~\ref{sec:2}.
\item A new neural network (\textit{i.e.}, YOLY) is proposed, which consists of three joint disentanglement subnetworks. In brief, two non-degenerate convolutional neural networks (J-Net and T-Net) are used to obtain the clean image and the transmission map, and a variational auto-encoder (A-Net) is used to obtain the atmospheric light. The whole network is learned in a self-supervised manner, which only explore and exploit the information rooted in the observed hazy image as the supervision.
\item Our method could also be used to synthesize new hazy images by transferring the haze from one image to another clean image, thus avoiding handcrafting the parameters of the physical model.
\end{itemize}
\section{Related Work}
\label{sec:2}
In this section, we briefly introduce some existing image dehazing methods from the perspective of prior- and learning-based category. Besides, we will also introduce recent development in unsupervised neural networks for image enhancement.
\subsection{Prior-based methods}
Almost all existing prior-based methods are shallow models, which aim to explore various of handcrafted priors from the hazy-free images, such as texture, contrast and chromatic aberration. With the prior, the transmission map and atmospheric light could be estimated, and thus the under-constrained dehazing problem could be well posed. Following this paradigm, dark channel prior~\citep{DCP} was proposed based on the following observation. Namely, most local patches in haze-free outdoor images contain some pixels with very low intensities in at least one color channel. By using such a prior, DCP successfully estimate the transmission map and atmospheric light of hazy image. Besides, a variety of priors are employed based on different observations/assumptions, such as color attenuation prior (CAP)~\citep{CAP}, non-local color prior (NCP)~\citep{NLD} and haze-line prior (HLP)~\citep{HLP}.
Although promising performance has been achieved by these prior-based methods, the dehazing performance is not always desirable due to the inconsistency between the adopted prior and changeable environment. Moreover, these methods are shallow models which might be with limited capacity of handling complex data.
\subsection{Learning-based methods}
Different from the prior-based methods, learning-based methods adopt a data-driven manner to learn the transmission map and/or atmospheric light. In recent, motivated by the success of neural networks, some studies~\citep{DehazeNet, MSCNN, DehazeGAN, AOD-Net, GridDehazeNet, EPDN} have been conducted to apply neural networks to image dehazing, which have achieved the state-of-the-art performance. For instance, \cite{DehazeNet} proposed DehazeNet which utilizes a trainable convolutional neural network to estimate the transmission map under the supervision of the ground-truth transmission maps. \cite{MSCNN} proposed a multi-scale convolutional neural network wherein a coarse- and fine-scale network are combined to estimate the transmission map for dehazing. \cite{DehazeGAN} proposed simultaneously estimating the transmission map and the atmospheric light by using a generative adversarial network. In summary, like the neural network in other tasks, the success of these deep deep image dehazing methods also rely on a large scale training dataset which is with some truthful supervisors such as the hazy-clean image pairs.
The differences of our method with these existing deep learning based methods~\citep{DehazeNet, MSCNN, DehazeGAN, AOD-Net, GridDehazeNet, EPDN} are given in the following aspects: 1) the proposed YOLY works in an unsupervised rather than supervised manner. In other words, our method does not need the hazy-clean pair images; 2) YOLY is an ``untrained'' instead of ``trained'' model. In other words, YOLY does not require training on a dataset, which could directly handle each single hazy image when it is observed. These two advantages make our method avoid the labor-intensive data collection and the domain-shift issues of using the synthetic hazy images to address the real-world images; 3) our method could disentangle the clean image, the transmission map, and the atmospheric light from the hazy image. This makes transferring haze from the real scenes to another image possible. Namely, our method provides an effective solution to synthesize new hazy images in a data-driven rather than human-specific way.
\subsection{Unsupervised Deep Image Enhancement Methods}
Although there are only few efforts~\citep{Double-DIP} in developing unsupervised approach for single image dehazing, some methods have been proposed for other image enhancement tasks in recent~\citep{N2N,N2V,DIP,Double-DIP,DD}. For example, Noise2Noise (N2N)~\citep{N2N} shows that simple statistical arguments lead to new capabilities in data denoising using neural networks. However, it has to use a corrupted image set with the same noise distribution to train the neural network. In other words, it cannot handle the single image case like our method does despite the difference in tasks. Deep image prior (DIP)~\citep{DIP} is another recently proposed unsupervised method which fits a corrupted image using a random noise vector and early-stoping strategy to recover the clean image. However, it is a daunting task to determine the training epoch at which the desirable result is obtained.
Although both the aforementioned methods and our YOLY are unsupervised methods, there are largely different. First, most of the aforementioned methods are not specifically designed for single image dehazing. Will be shown in our experiments, they cannot achieve encouraging performance in such a challenging task (haze is a kind of signal-dependent noises). Second, the methods such as N2N and its variants require using a data collection for training, whereas our method will only use the observed hazy image itself. Third, our method is based on the layer disentanglement idea which is different from these methods in the methodology. It should be pointed out that our method is also remarkably different from the recently proposed DDIP~\citep{Double-DIP} in the following two aspects. On one hand, the loss and the network structure are totally different. To be specific, our method employs variational inference to model the atmospheric light, whereas DDIP adopts a U-Net-like structure to fit the image same with DIP. Moreover, our YOLY utilizes the color attenuation as a supervisor to estimate the clean image, whereas DDIP mainly employs the early-stopping fitting strategy~\citep{DIP}. On the other hand, the input and the working mechanism are different. In brief, DDIP takes three random noises as inputs and feeds them into three generator networks to fit the hazy image, which utilizes the properties of DIP. In contrast, our method directly feeds the hazy image as the conditional input into three subnetworks so that different layers are disentangled. In other words, DDIP performs layer composition in a bottom-to-top fashion, whereas our YOLY performs layer disentanglement in a top-to-bottom fashion.
\begin{figure*}[!th]
\begin{center}
\includegraphics[scale=0.6]{./Fig/Fig.1-2/fig2_net}
\end{center}
\caption{\label{Figure:Net}The architecture of YOLY. In brief, YOLY consists of three joint learning subnetworks which are the clean image estimation network (J-Net), the transmission map estimation network (T-Net), and the atmospheric light estimation network (A-Net). Taking a single hazy image as the input, these three subnetworks disentangle the input into three different layers which are then utilized to reconstruct the input hazy image in a decomposition-composition fashion. Note that, the atmospheric light $f_{A}(x)$ would be approximately even exact homogeneous if the transmission map $f_{T}(x)$ could be well recovered as shown in the figure. HSV denotes the operation of transforming the recovered image to its HSV version.
\end{figure*}
\section{Proposed Method}
\label{sec:3}
Given a single hazy image $x$ as the input, we aim to recover the clean image $J(x)$ without using the information beyond the image content itself. The basic idea of our methods is to disentangle $x$ into three layer components using three joint subnetworks as shown in Fig~\ref{Figure:Net}. More specifically, YOLY simultaneously feeds $x$ into a clean image estimation network (J-Net), a transmission map estimation network (T-Net), and an atmospheric light estimation network (A-Net). After that, the outputs of them are further combined to reconstruct $x$ at the top of YOLY through the atmospheric scattering physical model. In such a way, the whole model is learned in an unsupervised fashion and these subnetworks are optimized in an end-to-end manner. Formally, at the top layer of YOLY, we aim to minimize the following loss:
\begin{equation}
\label{eq:L_Rec}
\mathcal{L}_{Rec} = \|I(x)-x\|_{p},
\end{equation}
where $\|\cdot\|_{p}$ denotes $p$-norm of a given matrix. In this paper, we simply adopt Frobenius norm. $I(x)$ is computed by composing the outputs of the three subnetworks via
\begin{equation}
\label{physical model}
I(x)=J(x)T(x)+A(1-T(x))
\end{equation}
where $J(x)$ denotes the clean image predicted by J-Net, $T(x)$ is the medium transmission map predicted by T-Net, and $A$ is the global atmospheric light on each pixel coordinates predicted by A-Net. It should be pointed out that some algorithms~\citep{DehazeGAN} have been proposed by learning these factors based on the above physical model, however, to the best of our knowledge, there is few efforts have been devoted to developing unsupervised methods so far.
The loss $\mathcal{L}_{Rec}$ is designed to constrain the entire network including the subnetworks to well reconstruct the hazy image $x$ after layer disentanglement. In other words, it guides the layer disentanglement and composition through incorporating the haze generation process. In the following, we will illustrate how these three networks could obtained desired layer components, besides directly utilizing the above self-supervision.
\subsection{J-Net}
J-Net aims to predict the clean image $J(x)$ from the hazy image $x$. As demonstrated in~Fig.(\ref{Figure:Net}), J-Net takes a non-degenerate architecture by following~\citep{DCGAN}. Namely, our J-Net does not implement the down-sampling operation, thus preventing the loss of detail in $J(x)$. More specifically, J-Net only consists of the convolutional layer, batch normalization layer~\citep{BN}, and LeakyReLU activation. In the last layer, we choose sigmoid function to normalize the output into $[0, 1]$. More details about the network implementation could refer to the experimental setting and supplementary materials.
To supervise J-Net, we propose the following loss function
\begin{equation}
\label{eq:LJ}
\mathcal{L}_{J} = \|V(f_{J}(x)) - S(f_{J}(x))\|_{p},
\end{equation}
where $V(f_{J}(x))$ denotes the brightness of $f_{J}(x)$, and $S(f_{J}(x))$ denotes the saturation of $f_{J}(x)$.
The loss term $\mathcal{L}_{J}$ is designed based on the observation in~\citep{CAP}. Namely, the depth of clean image is positively correlated to the difference between brightness and saturation of a clean image. To utilize this prior in an unsupervised manner, we recast the prior as the above formulation, \textit{i.e.}, the difference between the value and saturation should be as small as possible in the predicted $J(x)$. Eq.~(\ref{eq:LJ}) has two advantages. On one hand, the formulation has sub-gradients and pluggable into our model to enjoy the joint optimization through back-propagation. On the other hand, it makes possibility to recover the clean image without using the ground-truth clean image.
\subsection{T-Net}
As the clean background and the transmission map are dependent of the input $x$, we adopt the similar network structure for J-Net and T-Net. There are only two differences between them. To be specific, the output layer of J-Net is with three channels, whereas the output layer of T-Net is with only one channel for computational efficiency. On the other hand, T-Net does not employ an explicit loss, which only utilizes the self-supervision back-propagated from the top layer of YOLY to guide the optimization.
\subsection{A-Net}
A-Net aims to estimate the global atmospheric light from the observed images. As the global atmospheric light $A$ is independent of the image content and owns the global property, it is reasonable to assume that $A$ is sampled from a latent Gaussian distribution. Accordingly, we recast the learning of $A$ as a variational inference problem~\citep{VAE}. To be specific, A-Net consists of an encoder, a symmetric decoder, and an intermediate block. Both the encoder and the decoder consist of four blocks. In the encoder, the blocks consist of a convolutional layer, a ReLU activation function, and a max pooling layer in sequence. In the decoder, the blocks sequentially perform upsampling, convolution, batch normalization, and ReLU activation. To learn the latent Gaussian model, the intermediate block is used to transform the output (\textit{i.e.}, $z$) of the encoder to the latent Gaussian distribution $\mathcal{N}(\mu_{z}, \sigma_{z}^{2})$, \textit{i.e.}, $z\rightarrow \{\mu_{z}, \sigma_{z}^{2}\}$, where $\mu_{z}$ and $\sigma_{z}^{2}$ are mean and variance of the learned Gaussian model. Through resampling from the Gaussian model, the reconstruction of the latent code could be generated, \textit{i.e.}, $\mathcal{N}(\mu_{z}, \sigma_{z}^{2})$ $\rightarrow\hat{z}$. After that, $\hat{z}$ is fed into the decoder to obtain the reconstruction of the disentangled atmospheric light $f_{A}(x)$.
The loss function for A-Net is formulated as below:
\begin{equation}
\label{eq:LA}
\mathcal{L}_{A}=\mathcal{L}_{H}+\mathcal{L}_{KL}+\lambda\mathcal{L}_{Reg},
\end{equation}
where $\mathcal{L}_{H}$ is the loss between the disentangled atmospheric light $f_{A}(x)$ and the initial hint $A(x)$, where $A(x)$ is estimated from $x$. $\mathcal{L}_{KL}$ and $\mathcal{L}_{Reg}$ denote the loss for variational inference and regularization term, respectively. $\lambda$ is a nonnegative parameter to balance the regularization.
To be exact, $\mathcal{L}_{H}$ is defined as
\begin{equation}
\label{eq:LH}
\mathcal{L}_{H}=\|f_{A}(x)-A(x)\|_{F}.
\end{equation}
$\mathcal{L}_{KL}$ aims to minimize the difference between the latent code $z$ and the corresponding reconstruction $\hat{z}$ re-sampled from the Gaussian model. To enjoy the end-to-end optimization using the standard stochastic gradient methods, the reparameterization trick could be used to yield a lower bound estimator~\citep{VAE}. Mathematically,
\begin{align}
\label{eq:KL}
\mathcal{L}_{KL}&=KL(\mathcal{N}(\mu_z, \sigma^2_z) \arrowvert\arrowvert \mathcal{N}(0, I))\notag\\
&=\frac{1}{2}\sum_{i}\left((\mu_{z_{i}})^{2}+(\sigma_{z_{i}})^{2}-1-\log(\sigma_{z_{i}})^{2}\right)
\end{align}
where $KL(\cdot)$ denotes the Kullback-Leibler divergence between two distributions, $z_{i}$ denotes the $i$-th dimension of $z$.
To avoid overfitting, we enforce $\mathcal{L}_{Reg}$ on the outputs of A-Net, \textit{i.e.}, $f_{A}(x)$. Formally,
\begin{equation}
\label{eq:LReg}
\mathcal{L}_{Reg}(x)=\frac{1}{2m}\sum_{i=1}^{m}(x_{i}-\frac{1}{|N(x_{i})|}\sum_{y_{i}\in N(x_{i})}y_{i})^{2},
\end{equation}
where $\mathcal{N}(x_{i})$ is the second order neighborhood of $x_{i}$, $|\mathcal{N}(x_{i})|$ is the neighborhood size, and $m$ denotes the pixel number of $x$. Clearly, the regularizations play a role of mean filtering, which enforce $A(x)$ to be smooth. Note that, the high-frequency details of the recovered haze-free image might lose if the above regularization is enforced on the output of J-Net.
In summary, the total loss of our YOLY is as below:
\begin{align}
\label{eq:finalloss}
\mathcal{L}
=&\mathcal{L}_{Rec}+\mathcal{L}_{J}+\mathcal{L}_{H}+\mathcal{L}_{KL}+\lambda\mathcal{L}_{Reg}\notag\\
=&\|I(x)-x\|_{F}^{2}+\|V(f_{J}(x))-S(f_{J}(x))\|_{F}^{2}\notag\\
&+\|f_{A}(x)-A(x)\|_{F}^{2}\notag\\
&+\sum_{i}\left((\mu_{z_{i}})^{2}+(\sigma_{z_{i}})^{2}-1-\log(\sigma_{z_{i}})^{2}\right)\notag\\
&+\frac{\lambda}{m}\sum_{i=1}^{m}(f_{A}(x)_{i}-\frac{1}{|N(f_{A}(x)_{i})|}\sum_{y_{i}\in N(f_{A}(x)_{i})}y_{i})^{2}
\end{align}
\section{Experiments}
In this section, we evaluate our method on two synthetic datasets and two real-world datasets, comparing with 14 baseline methods in terms of PSNR and SSIM. In the following, we will first demonstrate the experimental setting, and then show the qualitative and quantitative results on the datasets, as well as the time cost of YOLY. After that, we will show the effectiveness of our method for haze transfer. Finally, the ablation study is presented to verify the effectiveness of the proposed method.
\subsection{Experimental Settings}
In this part, we elaborate on the used datasets, baselines, the evaluation metrics, and the implementation details.
\textbf{Datasets:}
RESIDE~\citep{RESIDE} is a new large scale haze image dataset, of which the testing subsets consist of Synthetic Objective Testing Set (SOTS) and Hybrid Subjective Testing Set (HSTS). To be specific, SOTS consists of 500 indoor hazy images, which are synthesized by the physical model with handcrafted parameters. HSTS contains 10 synthetic haze images and 10 real-world hazy images. In our experiments, we take SOTS and HSTS for evaluations. Besides, we also manually collect 10 hazy real-world images from Internet for a more comprehensive investigation.
\textbf{Baselines:}
For comprehensive comparisons, we compare the proposed YOLY with 14 methods which are divided into three groups, namely, four learning-based dehazing methods, five prior-based dehazing methods and five unsupervised deep image enhancement methods. It should be pointed out that, both the prior-based and unsupervised deep methods remove haze from image without using the ground-truth clean image, and their major difference is that the former is the shallow model whereas the latter is based on deep neural networks.
To be specific, the learning-based dehazing methods are DehazeNet~\citep{DehazeNet}, MSCNN~\citep{MSCNN}, AOD-Net~\citep{AOD-Net} and CAP~\citep{CAP}. Here, although CAP use prior information, we still classify it into the learning-based category because it employs the ground-truth transmission as the supervisor. The prior-based dehazing approaches are DCP~\citep{DCP}, FVR~\citep{FVR}, BCCR~\citep{BCCR}, GRM~\citep{GRM}, NLD~\citep{NLD}. The unsupervised deep image enhancement methods contain N2N~\citep{N2N}, N2V~\citep{N2V}, DIP~\citep{DIP}, DD~\citep{DD} and DDIP~\citep{Double-DIP}. Noticed that, N2N, N2V, DD, and DIP are specifically designed for other image enhancement tasks rather than image dehazing, which are compared in our experiments for two reasons. On one hand, there is no unsupervised deep model for singe image dehazing excepted DDIP so far. Thus, the comparisons with them could provide a more extensive study. On the other hand, the inferior performance achieved by them shows that it would achieve undesirable result if these methods are simply applied to the dehazing task. Furthermore, among all the tested methods, only N2V, DD, DIP, DDIP, and our YOLY are untrained neural networks, which means that they do not require training data and only use the given hazy image.
\textbf{Evaluation Metrics:} Like~\citep{MSCNN,AOD-Net,Double-DIP,EPDN, DCPDN}, two popular metrics are used in the quantitative comparisons, \textit{i.e.}, PSNR and SSIM. Higher value of these metrics indicates better dehazing performance.
\textbf{Experimental Configurations:} We conduct experiments on an NVIDIA Titan RTX GPU in PyTorch. To optimize YOLY, we employ the ADAM optimizer~\citep{Adam} with the default learning rate and the maximal epoch of 500. For better reproducibility, we do not exhaustively tune parameters for our method and instead fix $\lambda = 0.1$ for all the evaluations. To initialize the hint, we use the method similar to \citep{DCP}. Regarding some of baselines, we directly refer to the best results reported in the original papers. For the baselines without the corresponding results, we carry out them by using the source codes provided by the authors and adopting their parameter settings. The source code of YOLY will be released on Github.
\begin{table*}
\centering
\caption{Results on the synthetic indoor database (SOTS). The bold number indicates the best method of each category of methods.}
\label{Table:SOTS}
\begin{tabular}{c| cccc ccccc}
\toprule
\multicolumn{1}{c|}{\multirow{2}{*}{Metrics}} & \multicolumn{4}{c|}{Learning-based Dehazing Methods} & \multicolumn{5}{c}{Prior-based Dehazing Methods} \\
\cline{2-10}
& DehazeNet & MSCNN & AOD-Net & \multicolumn{1}{c|}{CAP} & DCP & FVR & BCCR & GRM & NLD \\
\hline
PSNR & \textbf{21.14} & 17.57 & 19.06 & \multicolumn{1}{c|}{19.05} & 16.62 & 15.72 & 16.88 & \textbf{18.86} & 17.29\\
SSIM & 0.8472 & 0.8102 & \textbf{0.8504} & \multicolumn{1}{c|}{0.8364} & 0.8179 & 0.7483 & 0.7913 & \textbf{0.8553} & 0.7489\\
\midrule
\multicolumn{1}{c|}{\multirow{2}{*}{Metrics}} & \multicolumn{6}{c|}{Unsupervised Neural Networks} \\
\cline{2-7}
& N2N & N2V & DIP & DD & \multicolumn{1}{c|}{DDIP} & \multicolumn{1}{c|}{Ours} \\
\cline{1-7}
PSNR & 14.49 & 10.67 & 12.28 & 11.92 & \multicolumn{1}{c|}{16.97} & \multicolumn{1}{c|}{\textbf{19.41}}
\\
SSIM & 0.7078 & 0.5397 & 0.5782 & 0.6404 & \multicolumn{1}{c|}{0.7147} & \multicolumn{1}{c|}{\textbf{0.8327}} \\
\bottomrule
\end{tabular}
\end{table*}
\begin{figure*}[!t]
\def \m_width{0.075}
\def \a_height{1cm}
\def \b_height{1cm}
\begin{center}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/Input/1400_3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DehazeNet/1400_3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/MSCNN/1400_3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/AOD-Net/1400_3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DCP/1400_3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/GRM/1400_3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/N2V/1400_3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DIP/1400_3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DD/1400_3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DDIP/1400_3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/VALID/1400_3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth,height=\a_height]{./Fig/SOTS/GT/1400_3.pdf}}
\end{center}
\vspace{-0.7cm}
\begin{center}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/Input/1401_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DehazeNet/1401_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/MSCNN/1401_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/AOD-Net/1401_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DCP/1401_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/GRM/1401_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/N2V/1401_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DIP/1401_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DD/1401_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DDIP/1401_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/VALID/1401_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth,height=\a_height]{./Fig/SOTS/GT/1401_6.pdf}}
\end{center}
\vspace{-0.7cm}
\begin{center}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/Input/1410_4.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DehazeNet/1410_4.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/MSCNN/1410_4.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/AOD-Net/1410_4.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DCP/1410_4.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/GRM/1410_4.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/N2V/1410_4.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DIP/1410_4.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DD/1410_4.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DDIP/1410_4.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/VALID/1410_4.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth,height=\a_height]{./Fig/SOTS/GT/1410_4.pdf}}
\end{center}
\vspace{-0.7cm}
\begin{center}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/Input/1421_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DehazeNet/1421_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/MSCNN/1421_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/AOD-Net/1421_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DCP/1421_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/GRM/1421_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/N2V/1421_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DIP/1421_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DD/1421_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DDIP/1421_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/VALID/1421_6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth,height=\a_height]{./Fig/SOTS/GT/1421_6.pdf}}
\end{center}
\vspace{-0.7cm}
\begin{center}
\setcounter{subfigure}{0}
\subfigure[]{\label{Figure:SOTS:Input}\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/Input/1436_2.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DehazeNet/1436_2.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/MSCNN/1436_2.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/AOD-Net/1436_2.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DCP/1436_2.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/GRM/1436_2.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/N2V/1436_2.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DIP/1436_2.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DD/1436_2.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/DDIP/1436_2.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/SOTS/VALID/1436_2.pdf}}
\subfigure[]{\label{Figure:SOTS:GT}\includegraphics[width=\m_width\textwidth,height=\a_height]{./Fig/SOTS/GT/1436_2.pdf}}
\end{center}
\caption{\label{Figure:SOTS} Visual Results on SOTS. From the left to the right column (\textit{i.e.}, Figs.~(\ref{Figure:SOTS:Input}--\ref{Figure:SOTS:GT})), the input hazy image, DehazeNet~\citep{DehazeNet}, MSCNN~\citep{MSCNN}, AOD-Net~\citep{AOD-Net}, DCP~\citep{DCP}, GRM~\citep{GRM}, N2V~\citep{N2V}, DIP~\citep{DIP}, DD~\citep{DD}, DDIP~\citep{Double-DIP}, our YOLY and the ground truth are presented in column-wise. Some areas are highlighted by red rectangles and zooming-in is recommended for a better visualization and comparison. }
\end{figure*}
\begin{table*}[!t]
\centering
\caption{Results on the synthetic outdoor database (HSTS). The bold number indicates the best method of each category of methods.}
\label{Table:HSTS}
\begin{tabular}{c| cccc ccccc}
\toprule
\multicolumn{1}{c|}{\multirow{2}{*}{Metrics}} & \multicolumn{4}{c|}{Learning-based Dehazing Methods} & \multicolumn{5}{c}{Prior-based Dehazing Methods} \\
\cline{2-10}
& DehazeNet & MSCNN & AOD-Net & \multicolumn{1}{c|}{CAP} & DCP & FVR & BCCR & GRM & NLD\\
\hline
PSNR & \textbf{24.48} & 18.64 & 20.55 & \multicolumn{1}{c|}{21.53} & 14.84 & 14.48 & 15.08 & 18.54 & \textbf{18.92}\\
SSIM & \textbf{0.9153} & 0.8168 & 0.8973 & \multicolumn{1}{c|}{0.8726} & 0.7609 & 0.7624 & 0.7382 & \textbf{0.8184} & 0.7411 \\
\midrule
\multicolumn{1}{c|}{\multirow{2}{*}{Metrics}} & \multicolumn{6}{c|}{Unsupervised Neural Networks} \\
\cline{2-7}
& N2N & N2V & DIP & DD & \multicolumn{1}{c|}{DDIP} & \multicolumn{1}{c|}{Ours} \\
\cline{1-7}
PSNR & - & 11.79 & 14.55 & 14.66 & \multicolumn{1}{c|}{20.91} & \multicolumn{1}{c|}{\textbf{23.82}} \\
SSIM & - & 0.5450 & 0.5573 & 0.6409 & \multicolumn{1}{c|}{0.8842} & \multicolumn{1}{c|}{\textbf{0.9125}} \\
\bottomrule
\end{tabular}
\end{table*}
\begin{figure*}[!t]
\def \m_width{0.075}
\def \a_height{2cm}
\def \b_height{1cm}
\begin{center}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/HSTS/Input/2.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/HSTS/DehazeNet/2.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/HSTS/MSCNN/2.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/HSTS/AOD-Net/2.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/HSTS/DCP/2.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/HSTS/GRM/2.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/HSTS/N2V/2.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/HSTS/DIP/2.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/HSTS/DD/2.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/HSTS/DDIP/2.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/HSTS/Ours/2.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth,height=\a_height]{./Fig/HSTS/GT/2.pdf}}
\end{center}
\vspace{-0.7cm}
\begin{center}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\b_height]{./Fig/HSTS/Input/6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\b_height]{./Fig/HSTS/DehazeNet/6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\b_height]{./Fig/HSTS/MSCNN/6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\b_height]{./Fig/HSTS/AOD-Net/6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\b_height]{./Fig/HSTS/DCP/6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\b_height]{./Fig/HSTS/GRM/6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\b_height]{./Fig/HSTS/N2V/6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\b_height]{./Fig/HSTS/DIP/6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\b_height]{./Fig/HSTS/DD/6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\b_height]{./Fig/HSTS/DDIP/6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\b_height]{./Fig/HSTS/Ours/6.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/HSTS/GT/6.pdf}}
\end{center}
\vspace{-0.7cm}
\setcounter{subfigure}{0}
\begin{center}
\subfigure[]{\label{Figure:HSTS:Input}\includegraphics[width=\m_width\textwidth, height=\b_height]{./Fig/HSTS/Input/8.pdf}}
\subfigure[]{\label{Figure:HSTS:DehazeNet}\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/HSTS/DehazeNet/8.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/HSTS/MSCNN/8.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/HSTS/AOD-Net/8.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/HSTS/DCP/8.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/HSTS/GRM/8.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/HSTS/N2V/8.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/HSTS/DIP/8.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/HSTS/DD/8.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/HSTS/DDIP/8.pdf}}
\subfigure[our]{\label{Figure:HSTS:SL}\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/HSTS/Ours/8.pdf}}
\subfigure[]{\label{Figure:HSTS:GT}\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/HSTS/GT/8.pdf}}
\end{center}
\caption{\label{Figure:HSTS} Visual Results on HSTS. From the left to the right column (\textit{i.e.}, Figs.(\ref{Figure:HSTS:Input}--\ref{Figure:HSTS:GT})), the input hazy image, DehazeNet~\citep{DehazeNet}, MSCNN~\citep{MSCNN}, AOD-Net~\citep{AOD-Net}, DCP~\citep{DCP}, GRM~\citep{GRM}, N2V~\citep{N2V}, DIP~\citep{DIP}, DD~\citep{DD}, DDIP~\citep{Double-DIP}, our YOLY and the ground truth are presented in column-wise. Zooming-in is recommended for a better visualization.}
\end{figure*}
\subsection{Comparisons on Synthetic Datasets}
Table~\ref{Table:SOTS} and Fig.~\ref{Figure:SOTS} report the quantitative and qualitative results on the indoor testing dataset (\textit{i.e.}, SOTS). It is worthy to note that, we do not illustrate the visualization results of FVR, BCCR, and NLD due to the space limitation and their relatively inferior performance. From Table~\ref{Table:SOTS} and Fig.~\ref{Figure:SOTS}, one could observe that:
\begin{itemize}
\item Compared with the unsupervised deep methods, YOLY is 2.44 and 0.1180 higher than the best of them in terms of PSNR and SSIM, respectively. Compared with the unsupervised shallow method (\textit{i.e.}, prior-based), YOLY achieves a gain of 0.55 over the best method in SSIM.
\item YOLY surpasses all supervised methods excepted DehazeNet in the quantitative comparisons. In the qualitative comparisons, YOLY shows the best visual results as shown in Fig.~\ref{Figure:SOTS}, especially in the border areas of the figures. For example, DehazeNet fails to obtain a desirable result from the first image and remove the haze from the area close to the chair in the second image.
\item On the dataset, YOLY takes about 34.08s to obtain the final result for each image, \textit{i.e.}, it costs 68.15ms to run each epoch. Note that, this is the whole time cost of YOLY and no more training is required.
\end{itemize}
Different from the indoor scenes, the outdoor scenes are usually much more complex in depth, illustration, and background. As shown in Table~\ref{Table:HSTS} and Fig.~\ref{Figure:HSTS}, our method achieves promising dehazing performance in the outdoor scenes (HSTS). More specifically,
\begin{itemize}
\item Table~\ref{Table:HSTS} shows that our method is remarkably superior to all the unsupervised baselines including deep and prior-based shallow methods. More specifically, it outperforms the best unsupervised deep method by 2.91 and 0.0283 in terms of PSNR and SSIM, respectively. It is also 4.9 and 0.0941 higher than the best prior-based dehazing methods. Note that, N2N cannot achieve result on the HSTS dataset because it requires multiple hazy samples from the same scene whereas this dataset only includes a single sample for each scene.
\item Although YOLY is quantitively worse than DehazeNet, it shows better recovery performance in the visual comparison (see Fig.~\ref{Figure:HSTS:DehazeNet} of DehazeNet and Fig.~\ref{Figure:HSTS:SL} of YOLY for example). In fact, the haze-free images recovered by our YOLY seem more favorite than the ground truth clean image (Fig.~\ref{Figure:HSTS:GT}) which might be corrupted during data collection.
\item The visualization results partially show the limitations of the traditional prior-based methods in handling the complex scenes which may violate the adopted prior. More specifically, DCP shows some distortions in the sky area. The main reason is that the sky and the bright areas do not satisfy the assumption of DCP, thus resulting in inaccurate estimation of the transmission map.
\item It is worth noting that all unsupervised deep models excepted DDIP and YOLY cannot perform well in the qualitative and quantitative comparisons. The reason is that they are not specifically designed for dehazing. This shows the difficulty and necessity in developing new deep unsupervised single image dehazing methods.
\item On the dataset, YOLY takes about 28.58s for handling each image averagely and each epoch costs 57.17ms.
\end{itemize}
\begin{figure*}[!t]
\def \m_width{0.095}
\def \a_height{1.2cm}
\def \b_height{1.2cm}
\begin{center}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/RealWorld/Input/3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/RealWorld/DCP/3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/RealWorld/GRM/3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/RealWorld/DehazeNet/3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/RealWorld/MSCNN/3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/RealWorld/AOD-Net/3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/RealWorld/DIP/3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/RealWorld/DD/3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/RealWorld/DDIP/3.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/RealWorld/YOLY/3.pdf}}
\end{center}
\vspace{-0.7cm}
\begin{center}
\setcounter{subfigure}{0}
\subfigure[]{\label{Figure:RW:Input}\includegraphics[width=\m_width\textwidth, height=\b_height]{./Fig/RealWorld/Input/4.pdf}}
\subfigure[]{\label{Figure:RW:DehazeNet}\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/RealWorld/DehazeNet/4.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/RealWorld/MSCNN/4.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/RealWorld/AOD-Net/4.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/RealWorld/DCP/4.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/RealWorld/GRM/4.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/RealWorld/DIP/4.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/RealWorld/DD/4.pdf}}
\subfigure[]{\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/RealWorld/DDIP/4.pdf}}
\subfigure[]{\label{Figure:RW:SL}\includegraphics[width=\m_width\textwidth,height=\b_height]{./Fig/RealWorld/YOLY/4.pdf}}
\end{center}
\caption{\label{Figure:RW} Visual Results on the Real-World Dataset. From the left to the right column (\textit{i.e.}, Figs.(\ref{Figure:RW:Input}--\ref{Figure:RW:SL})), the input hazy image, DehazeNet~\citep{DehazeNet}, MSCNN~\citep{MSCNN}, AOD-Net~\citep{AOD-Net}, DCP~\citep{DCP}, GRM~\citep{GRM}, DIP~\citep{DCP}, DD~\citep{DD}, DDIP~\citep{Double-DIP} and our method are presented in column-wise. Some areas are highlighted by red rectangles and zooming-in is recommended for a better visualization and comparison. }
\end{figure*}
\begin{figure*}[!t]
\begin{center}
\def \m_wid{0.0743}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/INPUT/example-04-haze}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DehazeNet/example-04-haze}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/MSCNN_Dehazing/example-04-haze}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/AOD-Net/example-04-haze}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DCP/example-04-haze}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/GRM/example-04-haze}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/N2V/example-04-haze}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DIP/example-04-haze}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DD/example-04-haze}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DDIP/example-04-haze}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY/example-04-haze}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY_T/example-04-haze}}
\end{center}
\vspace{-0.8cm}
\begin{center}
\def \m_wid{0.0254}
\def \two_wid{0.0508}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/INPUT/chenwu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DehazeNet/chenwu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/MSCNN_Dehazing/chenwu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/AOD-Net/chenwu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DCP/chenwu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/GRM/chenwu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/N2V/chenwu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DIP/chenwu}}
\subfigure{\includegraphics[scale=\two_wid]{./SupplementaryMaterial/Fig/RealWorld/DD/chenwu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DDIP/chenwu}}
\subfigure{\includegraphics[scale=\two_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY/chenwu}}
\subfigure{\includegraphics[scale=\two_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY_T/chenwu}}
\end{center}
\vspace{-0.8cm}
\begin{center}
\def \m_wid{0.0558}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/INPUT/chengde_skyless}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DehazeNet/chengde_skyless}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/MSCNN_Dehazing/chengde_skyless}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/AOD-Net/chengde_skyless}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DCP/chengde_skyless}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/GRM/chengde_skyless}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/N2V/chengde_skyless}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DIP/chengde_skyless}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DD/chengde_skyless}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DDIP/chengde_skyless}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY/chengde_skyless}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY_T/chengde_skyless}}
\end{center}
\vspace{-0.8cm}
\begin{center}
\def \m_wid{0.0466}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/INPUT/HazyDay}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DehazeNet/HazyDay}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/MSCNN_Dehazing/HazyDay}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/AOD-Net/HazyDay}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DCP/HazyDay}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/GRM/HazyDay}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/N2V/HazyDay}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DIP/HazyDay}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DD/HazyDay}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DDIP/HazyDay}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY/HazyDay}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY_T/HazyDay}}
\end{center}
\vspace{-0.8cm}
\begin{center}
\def \m_wid{0.0488}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/INPUT/hewu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DehazeNet/hewu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/MSCNN_Dehazing/hewu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/AOD-Net/hewu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DCP/hewu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/GRM/hewu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/N2V/hewu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DIP/hewu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DD/hewu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DDIP/hewu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY/hewu}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY_T/hewu}}
\end{center}
\vspace{-0.8cm}
\begin{center}
\def \m_wid{0.0591}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/INPUT/menglong}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DehazeNet/menglong}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/MSCNN_Dehazing/menglong}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/AOD-Net/menglong}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DCP/menglong}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/GRM/menglong}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/N2V/menglong}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DIP/menglong}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DD/menglong}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DDIP/menglong}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY/menglong}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY_T/menglong}}
\end{center}
\vspace{-0.8cm}
\begin{center}
\def \m_wid{0.0245}
\def \two_wid{0.0489}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/INPUT/shanxiang}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DehazeNet/shanxiang}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/MSCNN_Dehazing/shanxiang}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/AOD-Net/shanxiang}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DCP/shanxiang}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/GRM/shanxiang}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/N2V/shanxiang}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DIP/shanxiang}}
\subfigure{\includegraphics[scale=\two_wid]{./SupplementaryMaterial/Fig/RealWorld/DD/shanxiang}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DDIP/shanxiang}}
\subfigure{\includegraphics[scale=\two_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY/shanxiang}}
\subfigure{\includegraphics[scale=\two_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY_T/shanxiang}}
\end{center}
\vspace{-0.8cm}
\begin{center}
\def \m_wid{0.0470}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/INPUT/IMG_0752}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DehazeNet/IMG_0752}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/MSCNN_Dehazing/IMG_0752}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/AOD-Net/IMG_0752}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DCP/IMG_0752}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/GRM/IMG_0752}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/N2V/IMG_0752}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DIP/IMG_0752}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DD/IMG_0752}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DDIP/IMG_0752}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY/IMG_0752}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY_T/IMG_0752}}
\end{center}
\vspace{-0.8cm}
\begin{center}
\def \m_wid{0.047}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/INPUT/indonesia-fires}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DehazeNet/indonesia-fires}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/MSCNN_Dehazing/indonesia-fires}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/AOD-Net/indonesia-fires}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DCP/indonesia-fires}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/GRM/indonesia-fires}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/N2V/indonesia-fires}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DIP/indonesia-fires}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DD/indonesia-fires}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DDIP/indonesia-fires}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY/indonesia-fires}}
\subfigure{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY_T/indonesia-fires}}
\end{center}
\vspace{-0.8cm}
\begin{center}
\def \m_wid{0.0564}
\def \two_wid{0.1126}
\setcounter{subfigure}{0}
\subfigure[]{\label{Figure:RW:Input}\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/INPUT/zhuhai}}
\subfigure[]{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DehazeNet/zhuhai}}
\subfigure[]{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/MSCNN_Dehazing/zhuhai}}
\subfigure[]{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/AOD-Net/zhuhai}}
\subfigure[]{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DCP/zhuhai}}
\subfigure[]{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/GRM/zhuhai}}
\subfigure[]{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/N2V/zhuhai}}
\subfigure[]{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DIP/zhuhai}}
\subfigure[]{\includegraphics[scale=\two_wid]{./SupplementaryMaterial/Fig/RealWorld/DD/zhuhai}}
\subfigure[]{\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/DDIP/zhuhai}}
\subfigure[]{\label{Figure:RW:YOLY}\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY/zhuhai}}
\subfigure[]{\label{Figure:RW:YOLY}\includegraphics[scale=\m_wid]{./SupplementaryMaterial/Fig/RealWorld/YOLY_T/zhuhai}}
\end{center}
\caption{\label{Figure:RW} Visual Results on the Real-World scenes. From the left to the right column (\textit{i.e.}, Figs.(\ref{Figure:RW:Input}--\ref{Figure:RW:YOLY})), the input hazy image, DehazeNet~\citep{DehazeNet}, MSCNN~\citep{MSCNN}, AOD-Net~\citep{AOD-Net}, DCP~\citep{DCP}, GRM~\citep{GRM}, N2V~\citep{N2V}, DIP~\citep{DCP}, DD~\citep{DD}, DDIP~\citep{Double-DIP}, our method and the estimated transmission map are presented. Some areas are highlighted by red rectangles and zooming-in is recommended for a better visualization and comparison. }
\end{figure*}
\subsection{Comparisons on Real World Dataset}
To demonstrate the effectiveness of the proposed method in real-world hazy scenes, we carry out qualitative experiments on the HSTS real-world image set which is without the ground-truth clean image.
From Fig.~\ref{Figure:RW}, one could observe that our YOLY demonstrates the best visualization result in almost all scenes. For example, although DehazeNet, MSCNN, AOD-Net, DCP, and DDIP successfully remove most of haze in the pictures, they fail to handle the areas with much more details. Besides, the methods such as DCP and DDIP also suffer from the color distortions in the background. DehazeNet and AOD-Net lose some details after dehazing in the low-light areas. In contrast, our method could be immune from these issues and get a much more favorite recovery. On this dataset, YOLY takes about 59.27s for dehazing each image and each epoch costs about 118.54ms.
Besides the results on the HSTS real-world image set, we also conduct comparisons on 10 hazy images collected from Internet by us. As shown in Fig.~\ref{Figure:RW}, the proposed YOLY is remarkably superior to the baselines. Moreover, we also illustrate the transmission map learned by YOLY for a more comprehensive investigation in the last column.
\subsection{Haze Transfer}
Recent works have witnessed the effectiveness of supervised neural networks in image dehazing. To well train the neural networks, a large scale hazy-clean image pairs are required, and most of works resort to the synthetic dataset by manually specifying the parameters of the physical model. Such a handcrafted haze creation solution has suffered from a variety of limitations, \textit{e.g.}, the domain shift when using the synthetic hazy images to address the real-world images. To solve this problem, it is highly expected to develop new haze creation methods which work in a learning manner. However, to the best of our knowledge, there are few efforts have devoted to this problem so far.
As a by-product of our layer disentanglement idea, our method could transfer haze from a given hazy image to another clean one. In other words, we provide a novel solution to haze creation. To transfer the haze, our method first disentangles the atmospheric light and the transmission map from the hazy image. After that, the model is further used to generate the new hazy images w.r.t. a given clean image.
To demonstrate the effectiveness of our method in haze transfer, we conduct the following experiments, \textit{i.e.}, transferring haze 1) from an indoor image to another indoor clean image, see Fig.~\ref{Figure:Transfer_i2i}; 2) from an indoor image to another outdoor clean image, see Fig.~\ref{Figure:Transfer_i2o}; 3) from an outdoor image to another outdoor clean image. From the results, one could find that our method could successfully transfer the haze as expected. The performance dominance of YOLY is more distinct in Fig.~\ref{Figure:Transfer_r2o}, where the haze source comes from the real world. More specifically, one could observe that the haze nearby the camera is much heavier than other areas in the second image of Fig.~\ref{Transfer:b_r2o}. The major reason might attribute to the difficulty in estimating the depth information of outdoor image. In contrast, the hazy images generated by our method less suffers from this issue, which shows the promising haze creation performance of our method.
\begin{figure}[!t]
\def \m_width{0.11}
\def \a_height{1.5cm}
\begin{center}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2i/1411/1411_gt.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2i/1411/1411_hazy.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2i/1411/1411_transfer.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2i/1411/hazy_ori.pdf}}
\end{center}
\vspace{-0.7cm}
\begin{center}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2i/1413/1413_gt.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2i/1413/1413_hazy.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2i/1413/1413_transfer.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2i/1413/hazy_ori.pdf}}
\end{center}
\vspace{-0.7cm}
\begin{center}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2i/1423/1423_gt.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2i/1423/1423_hazy.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2i/1423/1423_transfer.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2i/1423/hazy_ori.pdf}}
\end{center}
\vspace{-0.7cm}
\begin{center}
\setcounter{subfigure}{0}
\subfigure[]{\label{Transfer:a_i2i}\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2i/1423/1423_gt.pdf}}
\subfigure[]{\label{Transfer:b_i2i}\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2i/1423/1423_hazy.pdf}}
\subfigure[]{\label{Transfer:c_i2i}\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2i/1423/transfer_replace.pdf}}
\subfigure[]{\label{Transfer:d_i2i}\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2i/1423/ori_replace.pdf}}
\end{center}
\caption{\label{Figure:Transfer_i2i} Haze transferring from \textbf{synthetic indoor hazy} images to \textbf{indoor clean} images. Fig.~\ref{Transfer:a_i2i} and \ref{Transfer:b_i2i} are the ground truth clean image and the corresponding handcrafted hazy image. Fig.~\ref{Transfer:c_i2i} is our hazy image synthesized by transferring the haze from Fig.~\ref{Transfer:d_i2i} to Fig.~\ref{Transfer:a_i2i}. From the result, one could find that the handcrafted hazy images show some distortions due to the wrong depth information, \textit{i.e.}, the area near to the border of the piano in the first image and the chair near to the camera in the third and fourth images. There are also some unreal haze distribution in the handcrafted hazy image, \textit{i.e.}, the area near to the camera in the second image. Some areas are highlighted by red rectangles and zooming-in is recommended for a better visualization and comparison. }
\end{figure}
\begin{figure}[!t]
\def \m_width{0.11}
\def \a_height{1.5cm}
\begin{center}
\setcounter{subfigure}{0}
\subfigure[]{\label{Transfer:a_i2o}\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2o/1/gt.pdf}}
\subfigure[]{\label{Transfer:b_i2o}\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2o/1/hazy.pdf}}
\subfigure[]{\label{Transfer:c_i2o}\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2o/1/transfer.pdf}}
\subfigure[]{\label{Transfer:d_i2o}\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/i2o/1/ori.pdf}}
\end{center}
\caption{\label{Figure:Transfer_i2o} Haze transferring from \textbf{synthetic indoor hazy} images to \textbf{outdoor clean} images. Fig.~\ref{Transfer:a_i2o} and \ref{Transfer:b_i2o} are the ground truth clean image and the corresponding handcrafted hazy image. Fig.~\ref{Transfer:c_i2o} is our hazy image synthesized by transferring the haze from Fig.~\ref{Transfer:d_i2o} to Fig.~\ref{Transfer:a_i2o}. One could find out that the haze is average distributed in the handcrafted hazy image. In contrast, the transferred hazy images show that the haze is light near the camera and heavy in other areas. Namely, our transferred hazy images is more real than the handcrafted ones.
}
\end{figure}
\begin{figure}[!t]
\def \m_width{0.11}
\def \a_height{3cm}
\begin{center}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/r2o/2/clear.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/r2o/2/hazy.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/r2o/2/transfer.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/r2o/2/ori.pdf}}
\end{center}
\vspace{-0.7cm}
\def \m_width{0.11}
\def \a_height{1.5cm}
\begin{center}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/r2o/7/clear.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/r2o/7/hazy.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/r2o/7/transfer.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/r2o/7/ori.pdf}}
\end{center}
\vspace{-0.7cm}
\begin{center}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/r2o/9/clear.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/r2o/9/hazy.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/r2o/9/transfer.pdf}}
\subfigure{\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/r2o/9/ori.pdf}}
\end{center}
\vspace{-0.7cm}
\begin{center}
\setcounter{subfigure}{0}
\subfigure[]{\label{Transfer:a_r2o}\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/r2o/10/clear.pdf}}
\subfigure[]{\label{Transfer:b_r2o}\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/r2o/10/hazy.pdf}}
\subfigure[]{\label{Transfer:c_r2o}\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/r2o/10/transfer.pdf}}
\subfigure[]{\label{Transfer:d_r2o}\includegraphics[width=\m_width\textwidth, height=\a_height]{./Fig/Transfer/r2o/10/ori.pdf}}
\end{center}
\caption{\label{Figure:Transfer_r2o} Haze transferring from \textbf{real-world hazy} images to \textbf{outdoor clean} images. Fig.~\ref{Transfer:a_r2o} and \ref{Transfer:b_r2o} are the ground truth clean image and the corresponding handcrafted hazy image. Fig.~\ref{Transfer:c_r2o} is our hazy image synthesized by transferring the haze from Fig.~\ref{Transfer:d_r2o} to Fig.~\ref{Transfer:a_r2o}. From the result, one could find out that there are a lot unreal haze distributions in the handcrafted hazy images. For example, the second image shows a heavy haze near the camera and light haze in the sky area, which is inconsistent with the real-world cases. As a contrast, our hazy image is faithful to such common knowledge.
}
\end{figure}
\subsection{Ablation Study}
To verify the effectiveness of our loss function, we conduct an ablation study on the HSTS dataset by removing one of $\mathcal{L}_{H}$, $\mathcal{L}_{KL}$, $\mathcal{L}_{J}$, and $\mathcal{L}_{Reg}$. From Table~\ref{Table:Ablation}, one could see that: 1) our method leverages the advantages of variational inference in haze removal with the formulation of $\mathcal{L}_{H}$ and $\mathcal{L}_{KL}$; 2) the statistical information used on J-Net could improves the performance of the model; 3) the performance of YOLY is slightly improved with the regularization on the estimation of atmospheric light.
\begin{table}
\centering
\caption{Ablation Study on the HSTS database.}
\label{Table:Ablation}
\begin{footnotesize}
\begin{tabular}{*{6}{c}}
\toprule
Metrics & $\mathcal{L}_{H}$ & $\mathcal{L}_{KL}$ & $\mathcal{L}_{J}$ & $\mathcal{L}_{Reg}$ & Ours \\
\midrule
PSNR & 21.16 & 21.68 & 22.53 & 22.28 & 23.82 \\
SSIM & 0.8733 & 0.8690 & 0.9039 & 0.8836 & 0.9125 \\
\bottomrule
\end{tabular}
\end{footnotesize}
\end{table}
\section{Conclusion}
In this paper, we propose an unsupervised and untrained image dehazing neural network which separates the observed hazy image into scene radiance layer, transmission map layer and atmospheric light layer. Three advantages of the proposed YOLY are: 1) unsupervised characteristic means that the method does not use the information beyond the image content; 2) untrained characteristic means that the method does not require using an image collection for training like most existing neural networks do; 3) transferable hazy capacity which could synthesize new hazy images by extracting the haze from a given image in a learning- and unsupervised-fashion. Extensive experiments on two synthesis datasets and one real-world dataset demonstrate the promising performance of our method in the quantitative and qualitative comparisons.
Although our method remarkably outperforms most existing unsupervised shallow and deep methods, it is only comparable to the state-of-the-art supervised image dehazing approaches. Thus, it is promising to continually improve its performance in future. Moreover, it is also valuable to explore how to extend our method to other image/video enhancement tasks such as denoising, inpainting, and so on.
\bibliographystyle{spbasic}
\bibliographystyle{spmpsci}
\bibliographystyle{spphys}
|
\section{Introduction to layered semiconductors}
\label{optical properties}
\vspace{-0.2cm}
The physical properties of atomic monolayers often change dramatically from those of their parent bulk materials. Prime examples are monolayers of graphite (graphene) and MoS$_2$, as their ultimate thinness makes them extremely promising for applications in electronics and optics. At the same time they give access to new degrees of freedom of the electronic system such as the valley index or interactions between quasi-particles such as excitons (Coulomb bound electron-hole pairs). Additional functionalities emerge as these materials are stacked in van der Waals heterostructures \cite{ubrig2020design}. In addition to the currently investigated materials, recently about 1800 materials were predicted to be exfoliable, stable in monolayer form \cite{mounet2018two} and therefore tools for investigating the properties of these emerging layered materials are of prime importance.
Here we provide a guide to optical emission and absorption spectroscopy for atomically thin layered materials, commonly carried out in optical microscopes for increased spatial resolution. Optical spectroscopy gives access to key information such as the bandgap, exciton binding energy and absorption strength of a material. Combining spatial and polarisation resolution gives access to the spin and valley physics in monolayers and also in heterostructures. In the latter optical, transitions are tunable over a wide wavelength range and electron-hole pairs can experience nanoscale moir\'e confinement potentials for quantum optics experiments and investigating collective effects of electronic excitations \cite{seyler2019signatures, shimazaki2020strongly}. Moreover, optical spectroscopy techniques can be applied to semiconducting, magnetic layered materials such as chromium trihalides for probing their magnetisation \cite{sun2019giant,ubrig2019low,mak2019probing,paradisanos2020prominent}. Optical spectroscopy also reveals magnetic proximity effects and charge transfer as non-magnetic and magnetic layers are placed in direct contact to form heterostructures \cite{zhong2017van,PhysRevLett.124.197401,lyons2020interplay}. For applications in photonics optical spectroscopy reveals how light matter coupling is enhanced when layered materials are placed in optical cavities or on resonators \cite{sortino2019enhanced,paik2019interlayer}. Optical spectroscopy can be used as a non-invasive technique for studying lattice structure, interlayer coupling and stacking, that provides complementary information to direct atomic-resolution imaging from electron microscopy \cite{zhang2017interlayer,sushko2019high,andersen2019moir,shree2019high,holler2020lowfrequency}. \\
\indent The target of this review is to give an overview on what kind of information on novel layered materials we can access in practical optical spectroscopy, how the optical spectra are impacted by several distinct parameters such as the set-up and equipment used, the experimental conditions (temperature, external fields), the sample structure and very importantly the active layer environment. The electronic excitations in an atomically thin layer are strongly impacted by the substrate and encapsulating layers. This leads to two directions for experiments, (i) access intrinsic properties of the layers placed in a well controlled environment (i.e. in-between two inert buffer layers) or (ii) the layered material acts as a probe as we make use of the interaction of the optical excitations with the direct environment to investigate, for example, the magnetisation of adjacent layers or detecting molecules in the proximity \cite{zhao2019functionalization}. \\
\indent Optical properties of layered semiconductors, using the model system of transition metal dichalcogenide (TMD) semiconductors are introduced in the remainder of Section \ref{optical properties}, together with fabrication methods for typical sample structures. The equipment used for optical spectroscopy set-ups, commercially available systems or components assembled in a laboratory, are discussed in Section \ref{equipment}. Optical spectroscopy techniques used to uncover the main optical transitions in TMD materials and how they are impacted by the sample structure and the surrounding layers are detailed in the remainder of Section \ref{spectroscopy}. How to access spin and valley polarisation effects using optical methods is introduced in Section \ref{polarization}. Finally, opportunities for Raman scattering and second harmomic generation (SHG) are outlined in Section \ref{orientation}.
\begin{figure*}[t]
\centering
\includegraphics[width=0.95\linewidth]{Fig1}
\caption{\textbf{Excitons in TMD monolayers and heterobilayers for optical spectroscopy}. \textbf{a,} Top view of MoS$_2$ monolayer hexagonal crystal structure, here transition metal atoms are shown in blue, the chalcogen atoms in black. The armchair and zigzag directions are indicated by purple and grey arrows. \textbf{b,} Schematic illustrations of valley selective optical transition which obey chiral selection rules addressed optically using $\sigma^+$ ($\sigma^-$) polarised light \cite{cao2012valley,xiao2012coupled,mak2012control,sallen2012robust,zeng2012valley}, see the details in Section \ref{polarization}. The direct bandgap is at the $K$-points of the Brillouin zone for TMD monolayers. \textbf{c,} Schematic of van der Waals heterobilayers. The individual monolayers show intralayer excitons (orange oval) a bound pair of an electron and hole which resides in a single layer. Interlayer excitons can be found in heterobilayers where the two charge carriers are located in different layers (green oval). Inset shows the hydrogen-like, exciton states which correspond to different energy levels, designated by their principal quantum number $n$. \textbf{d,} Band alignment between 2D semiconductor monolayers calculated in \cite{liu2016van}. Arrow shows the wavelength of the bandgap for different materials varying from the visible to near-infrared. }
\label{fig1}
\end{figure*}
\subsection{Optical properties of layered semiconductors : transition metal dichalcogenides}
\indent Interlayer van der Waals forces are considerably weaker than intralayer covalent bonding which makes a very large number of crystals exfoliable \cite{frindt1966single,mounet2018two}. TMD monolayers consist of a hexagonally oriented transition metal layer covalently bonded with top and bottom chalcogen layers \cite{dickinson1923crystal,wilson1969transition}, see FIG.~\ref{fig1}a. Multilayers are coupled by weak van der Waals forces and their symmetry is given by the stacking order \cite{van2019stacking,frondel1970molybdenite}, $i.e.$ the relative twist angle being 0$^{o}$ or 60$^{o}$. TMDs such as MoS$_2$ occur in different phases with semiconducting or metallic properties, here we focus on the semiconducting phase (2$H$).
The electronic and optical properties of TMDs change drastically as their thickness is reduced from bulk to atomic bilayer and monolayers. Bulk and few-layered TMDs exhibit an indirect bandgap, however, TMD monolayers show a direct bandgap. This gives rise to strong photoluminescence (PL) emission \cite{mak2010atomically,splendiani2010emerging} and also to a large, tunable absorption in monolayers \cite{tonndorf2013photoluminescence,moody2015intrinsic}. The light-matter interaction is dominated by excitons, Coulomb bound electron hole pairs, and this interaction is strongly enhanced when the incoming electromagnetic wave is resonant with the energy of the excitonic states \cite{wang2015giant}.\\
\indent The binding energies of excitons in TMD monolayers are of the order of several hundreds of meV. This is a consequence of the large electron and hole effective masses \cite{kormanyos2015k}, the reduced dielectric screening and the spatial confinement of the carriers. As a result, excitonic effects dominate the optical properties even at room temperature and beyond \cite{he2014tightly,chernikov2014exciton,ugeda2014giant,wang2018colloquium}.
A strong spin-orbit splitting of about 200~meV (for Mo-based) and 400~meV (W-based) appears in the valence bands at the $K$-point \cite{ramasubramaniam2012large,song2013transport,kormanyos2015k}. As a consequence, two separate interband optical transitions are observed in absorption, named A (transition from the upper valence band) and B (transition from the lower valence band) \cite{wilson1969transition}.\\
\indent Interestingly, the strong Coulomb interaction is extremely sensitive to the dielectric screening associated with a spatially inhomogeneous environment \cite{rytova2018screened,Keldysh} as discussed in Section \ref{spectroscopy}.
This implies that the exciton transition energy, and to a larger extent, both the exciton binding energy and the free carrier bandgap can be tuned by engineering the local dielectric environment \citep{raja2017coulomb,waldecker2019rigid}. However, a possible undesirable consequence is that local dielectric fluctuations from disorder and impurities can result in strongly broadened optical transitions \cite{rhodes2019disorder,cadiz2017excitonic}.\\
\indent Optical transitions that are dipole-, momentum- and spin-allowed are referred to as 'bright' and the excitons recombine by emitting a photon. In layered materials excitons with different spatial orientations of the optical dipole, either in-plane or out-of the layer plane (see scheme in FIG.~\ref{fig:box}), participate in optical transitions in WSe$_2$ \cite{zhou2017probing} and InSe$_2$ \cite{brotons2019out}, for example. Taking into account also optical transitions that rely on phonon absorption or emission or spin-mixing of the different electronic states results in a large number of possible optical transitions ('bright' and 'dark') for a specific material.
The emission and detection efficiency can be optimised in applied magnetic fields and by selecting a particular light polarisation and propagation direction \cite{wang2017plane,robert2017fine,robert:2020measurement}. These different types of optical transitions can be addressed selectively in the optical microscope set-up as detailed below. Mainly the \textit{interband} transitions between the valence band and conduction band are probed, but also transitions between excitonic states \cite{pollmann2015resonant} or \textit{intersubband} transitions in multilayers can be observed \cite{schmidt2018nano}.\\
\indent When the electron and hole reside within the same layer, we speak of intralayer excitons, as sketched in FIG.~\ref{fig1}c. Interlayer excitons can form in TMD heterobilayers due to the type-II (or staggered) band alignment \cite{liu2016van} with the photoexcited electrons and holes residing in different layers, see FIG.~\ref{fig1}d. These excitons are referred to as indirect in real space. The difference in the lattice constants between the monolayers forming the heterobilayer affects (i) the alignment of $K$-points which determines if the interlayer transition is direct or indirect in reciprocal space, i.e. a phonon is needed in addition to a photon in the emission process \cite{kang2013band,rivera2015observation} and, (ii) formation of moir\'e effects/reconstructions \cite{van2014tailoring,sushko2019high,weston2019atomic,sung2020broken}. The latter leads to a periodic modulation of the electron and hole bandstructure depending on the difference between the lattice constants and/or twist angle, see FIG.~\ref{fig3}a. The depth and periodicity of the moir\'e potential can generate localised emitters (individual excitons) or collective excitations (trapping of excitons) \cite{yu2017moire}, with encouraging first reports all using optical spectroscopy \cite{shimazaki2020strongly,regan2020mott,wang2019evidence}. This approach for generating periodic, nanoscale potentials can certainly be extended to new types of van der Waals heterostructures.
\subsection{Layered semiconductor samples : monolayers and heterostuctures}
\label{fabrication}
Widespread fabrication methods of monolayer samples on common SiO$_2$/Si substrates include mechanical exfoliation from bulk crystals (top-down) and bottom-up chemical vapor deposition (CVD). Promising results of high quality monolayers are also obtained by molecular beam epitaxy (MBE) \cite{dau2018beyond,pacuski2020narrow}.\\
\indent Exfoliated TMD crystals from high quality bulk show commonly defect densities of around $10^{12}$ $\text{cm}^{-2}$ \cite{rhodes2019disorder}, which is still considerably larger than in III-V semiconductor nanostructures. Currently the impact of defect concentration and type on the luminescence efficiency is widely investigated, with improvements reported for samples treated with super-acids \cite{amani2015near}. Exfoliation is widely used because of simple handling even outside clean-room facilities and cost efficiency. However, there are several limitations: (a) the location of a single monolayer on the stamp/substrate is random and searching for a monolayer among flakes of different thickness is time consuming (b) the monolayers are relatively small with an average lateral dimension of tens of micrometer, and (c) the yield for finding monolayer per unit of surface area is low. \\
\indent As an alternative to exfoliation, CVD allows direct growth of monolayer material on a large surface area
\cite{lee2012synthesis,kobayashi2015growth,rhyee2016high,george2019controlled}. Here, the monolayer dimensions and the number of monolayers per unit of surface area are considerably larger than for exfoliation, see FIG.~\ref{fig:box}, with the limiting case of monolayer material covering the entire substrate surface. Recently, it has been demonstrated that lateral, as well as vertical TMD heterostructures can directly grow on flat or patterned SiO$_2$/Si substrates \cite{sahoo2018one,li2020general}.
Detaching CVD-grown samples from the growth substrate to deterministically fabricate heterostructures is possible using, for instance, water assisted pick-up techniques \cite{jia2016large,paradisanos2020controlling}. \\
\indent For CVD-grown and exfoliated samples, the intrinsic quality can be evaluated by low-temperature optical spectroscopy. Emission from defects and inhomogeneous broadening in the optical spectra can be used as a diagnostic tool. In addition to the intrinsic layer quality, the impact of the underlying substrate can also be a limiting factor of the optical quality. This has been demonstrated by comparing low-temperature optical spectra between TMD monolayers (both CVD and exfoliated) on SiO$_2$ and encapsulated with hexagonal boron nitride (hBN)
\cite{cadiz2017excitonic,raja2019dielectric,shree2019high}. It is possible to further improve the optical response by optimising the thickness of the top and bottom hBN in encapsulated TMD samples \cite{lien2015engineering,robert2018optical,fang:2019control}, see FIG.~\ref{fig2}b.\\
\indent Individual layers from high quality bulk crystals of different layered materials can be assembled into van der Waals heterostructures by using deterministic dry stamping \cite{Gomez:2014a} or direct pick-up \cite{purdie2018cleaning}.
This allows controlled transfer of layers at precise locations on the substrate. An optical micrograph of a van der Waals heterostructure is shown in FIG.~\ref{fig2}a. During the transfer micro-bubbles (blisters) can appear due to trapped air, water or hydrocarbons \cite{purdie2018cleaning}. Agglomeration of the bubbles can be achieved by thermal annealing. This leaves clean, smooth areas with sharp interfaces \cite{doi:10.1021/acs.nanolett.7b01248}. In layered TMDs, the presence of bubbles or wrinkles can introduce defect emission due to strain and/or carrier localisation. This can lower the optical transition energy and lead to the appearance of localised emission, see FIG.~\ref{fig2}b and section \ref{spectroscopy}.\\
\indent Besides the intrinsic carrier density of a sample, charge impurities from a disordered substrate or adsorbates can introduce charge potential fluctuations, which strongly impact the optical properties. This is commonly observed in TMD monolayers on SiO$_2$/Si. As a result, transitions corresponding to charged exciton states (trions) can be detected in low temperature PL, shown in FIG.~\ref{fig2}b,c. Independent control of the carrier density in gated field-effect devices
\cite{wang2017probing,PhysRevLett.124.027401} is thus crucial to study the optical properties in the charge neutral, p- or n-doped regimes \cite{courtade2017charged,wang2017probing}.
\begin{figure*}[t]
\centering
\includegraphics[width=0.99\linewidth]{Fig2}
\caption{\textbf{Variation in photoluminescence (PL) response for different experimental conditions}. \textbf{a,} Optical micrograph of a typical van der Waals heterostructure sample containing a WSe$_2$ monolayer between hBN top and bottom encapsulation layer, sample homogeneity and sample imperfections (bubbles and wrinkles) can be seen. The homogeneous and non-homogeneous regions of the sample are labelled 1 and 2 in white, respectively. \textbf{b,} Typical PL spectra at T = 4~K recorded at two different locations 1 and 2 of the sample shown in \textbf{a}. The main exciton transitions (neutral excitons (A:1s), charged excitons (T) and spin- forbidden dark exciton (D) are quenched and localised emitters appear when PL is recorded on the bubbles or on wrinkles. However, strong PL emission corresponding to the main excitons from WSe$_2$ monolayers is recorded on the flat region of the sample. \textbf{c,} Sketch of MoSe$_2$ monolayer sample structures in different dielectric environment. The left panel is the sample structure of a non-encapsulated TMD monolayer on SiO$_2$ and the TMD monolayer encapsulated in hBN is on the right panel. Typical PL spectra recorded at room temperature and at cryogenic temperature on the encapsulated and the non-encapsulated sample are shown in the bottom panel. The linewidth reduces significantly for the encapsulated sample at T = 4~K compared to non-encapsulated sample.
\textbf{d,} hBN encapsulated MoSe$_2$ layer suspended over a trench \cite{PhysRevLett.124.027401} and corresponding spectra are shown in \textbf{e}.
The PL emission energy remains constant on suspended samples on different locations on the monolayer showing sample homogeneity \cite{PhysRevLett.124.027401}.
\textbf{f,} WS$_2$ monolayer encapsulated in hBN on a SiO$_2$/Si substrate, WS$_2$ layer plasma treated to generate optically active defects. Typical PL emission spectra for CW laser (532~nm) excitation at 0.01 $\mu$W and 5 $\mu$W at T=4~K. The PL emission intensity of the main excitons are clearly visible at high power density, whereas these features are almost not detectable at low laser power density. PL spectra at low laser power density reveal that carriers can be trapped efficiently by defect sites and recombine by emitting photons at lower energy. Therefore at low laser power the PL emission of defects is considerably stronger than the free exciton emission.}
\label{fig2}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.80\linewidth]{Fig3}
\caption{\textbf{Description of a typical microscope for optical spectroscopy}: Micro-photoluminescence and reflectivity set-ups have to fulfill several key criteria: High spatial resolution for sample mapping and accessing specific sample areas, high detection efficiency for example for quantum emitters in layered materials, combination of two detectors (one CCD/CMOS camera for sample imaging to guide the experiment and another camera to accumulate the emission signal as a function of wavelength). All this can be combined in a microscope, where excitation and detection pass through the same objective (epifluorescence geometry). The spatial resolution is given by the diameter $\Delta$ of the detection spot. This depends on the numerical aperture ($NA$) of the objective and the wavelength $\lambda$ through the Rayleigh criterion as $\Delta=\frac{0.61\lambda}{NA}$ \cite{hecht2002optics}, where $NA>0.8$ is common in commercial objectives. Excitons in layered materials can have an optical dipole either in-plane or out-of-plane, see inset representing emitters with in-plane and out-of-plane optical dipoles.
Objectives with high $NA$ also collect part of the emission from an out-of-plane optical dipole.
A typical microscope set-up can be used to perform with similar equipment photoluminescence, reflectance, photoluminescence excitation (PLE), Raman scattering and second harmonic generation (SHG) spectroscopy experiments. Van der Waals heterostructures of different materials can be studied for temperatures T = 4 - 300 K. The samples are placed on nonmagnetic nanopositioners (travel distance several millimetres). An objective located inside a cryostat is more stable for long signal integration times. The other microscope components remain at room temperature. The microscope design has several modules: The excitation light propagates through an optical fibre via the lower horizontal arm. The upper horizontal arm consists of a camera and broad-band light source for sample imaging. The vertical arm is used for detection. The signal is collected by a spectrometer (here we show a Czerny-Turner geometry) with a diffraction grating coupled to a highly sensitive Si-CCD camera. Optionally an optical fibre can be used in the detection path. The fibre transports the signal from the microscope to the spectrometer entrance and also acts as a spatial pin-hole, making this a confocal arrangement for increased lateral (xy) and depth (z) resolution \cite{pawley2006handbook}. Polarisation analysis is a key motivation for studying layered materials such as TMDs for spintronics and valleytronics. Both linear polarisation and circular polarisation analysis can be achieved. For example, in the excitation part a Glan-Taylor (GT) polariser and a halfwave plate ($\lambda/2$) can be inserted. It is important to take into account the impact of the beamsplitters (BS) on the polarisation state. The quarter wave plate ($\lambda/4$) transforms the linear into circular polarisation and vice versa. Instead of a simple waveplate, a liquid crystal retarder (where a precise retardance of $\lambda / 2$, $\lambda / 4$ etc. can be adjusted by the applied bias) can optionally be used in the detection path. This avoids a macroscopic beam movement which can occur during rotation of a typical waveplate as the retardance/polarisation is changed. The performance of all optical set-up components is wavelength dependent, such as limited bandwidth of reflection coatings, chromatic aberrations of the objective or optical retardance for specific wavelengthd. Examples of optical components and polarisation analysis/rejection used in a microscope can be found in \cite{pawley2006handbook,kuhlmann2013dark,benelajla2020extreme}. }
\label{fig:box}
\end{figure*}
\section{Measuring optical absorption and luminescence in layered semiconductors}
\label{spectroscopy}
The fundamental optical transitions in TMDs lie in the energy range from $\approx$ 1.1~eV (monolayer MoTe$_2$) up to $\approx$ 2.1 eV (monolayer WS$_2$), see FIG.~\ref{fig1}d. Ferromagnetic semiconductors, such as CrBr$_3$ and CrI$_3$ cover similar transition energies \cite{bermudez1979spectroscopic,molinasnchez2019magnetooptical}. Interlayer excitons in heterostructures such as MoS$_2$/WSe$_2$ can reach emission wavelengths above 1100~nm ($<$1.1~eV), approaching the telecommunication bands \cite{karni2019infrared}.
Black phosphorous is a layered semiconducting material with a direct bandgap that strongly varies with the number of layers and covers the visible (monolayer) to mid-infrared (bulk) spectral region \cite{ling2015renaissance}. The same evolution of band gap change versus thickness occurs in PtSe$_2$ \cite{ansari2019quantum}, but here the bandgap is indirect as for Si, so applications for detectors are possible.
At the opposite end of the spectrum layered hexagonal BN has a bandgap in the deep ultraviolet at 6~eV (200~nm) \cite{cassabois2016hexagonal}.
Below we describe the working principles of experiments to study absorption and emission of the optical transitions in layered semiconductors.
\subsection{Optical spectroscopy equipment}
\label{equipment}
A typical spectroscopy set-up contains a light source and several optical components to guide the excitation light to the sample and the signal to the spectrometer (monochromator) and detector. A charged-coupled device (CCD) or a complementary metal–oxide–semiconductor (CMOS) camera is also essential for the sample imaging. In this section we outline the characteristics of the main components and explain in FIG.~\ref{fig:box} the working principle of a versatile micro-spectroscopy set-up, widely used in commercial and also home-built systems. \\
\indent \textbf{Light sources.}- The main parameters for the laser excitation depend on the application and the sample's band structure: (i) the wavelength is selected in accordance to the investigated optical transitions (ii) the laser can be continuous-wave (CW) or pulsed. Pulsed lasers are more suitable for time-resolved experiments. For femtosecond (fs) or picosecond (ps) pulse duration the spectral width should be taken into account when investigating transitions close in energy. At the same time, the laser peak power should be calculated to avoid sample damage. High beam quality (aiming for perfect collimation, i.e. low $M^{2}$ factor) lasers are preferable for focusing the beam tightly to a diffraction limited spot, considering the small lateral dimensions of many exfoliated samples. For CW excitation, laser diodes can be typically selected between 375~nm and 2000~nm. For experiments requiring tunable wavelength excitation, convenient solutions for the 700-1000~nm range include Titanium Sapphire (Ti:sapphire) lasers \cite{spence199160}, which can be either pulsed (ps/fs) or continuous.
To cover the wavelength range between 500-700~nm and 1000-1600~nm, an optical parametric oscillator can be coupled to the Ti:sapphire laser combined with a doubling crystal. Dye lasers can also be used, where the choice of dye and its solvent allows for the selection of the emission range. Absorption or reflection measurements are performed using a broadband white-light source to cover the full visible wavelength range. Often a simple halogen lamp suffices. When only one specific optical transition is investigated, a monochromatic LED or SLED with 10-20~nm spectral bandwidth can be used. This enables a good compromise between a small spot size and sufficient excitation power. Other solutions include laser driven light-sources or super continuum white lasers \cite{alfano2016supercontinuum} that allow to select a broad or narrow wavelength range for excitation, which makes them versatile for photoluminescence but also for broad band absorption experiments. \\
\indent \textbf{Optical components.}- The optical components used in the set-up are selected for a specific wavelength range depending on both the excitation source and the emission wavelength. These include the polarisation components, microscope objectives, and lenses (ideally achromatic doublets) shown in FIG.~\ref{fig:box}. Homogeneous areas (flat surface, no charge fluctuations) in typical exfoliated sample have lateral dimensions down to a few micrometers. Therefore small excitation/detection spots close to the diffraction limit are crucial to record spectra with transition linewidths limited by the homogeneous, not inhomogeneous, broadening. A diffraction limited spot diameter can be achieved by using high numerical aperture ($NA$) objectives. \\
\indent \textbf{Detection.}- The final target is to detect the intensity as a function of wavelength of the light emitted from/scattered by the sample. The signal is focused onto the entrance slit of a spectrometer. The collected signal is then dispersed by a monochromator, which can host different diffraction gratings, where a small (large) number of lines/mm allows studying a broad (narrow) spectral range. The signal can then be detected by a CCD or high quality CMOS chips. Alternatively, the monochromator can be left out and a simple combination of filters can be used in front of the detector. \\
\indent \textbf{Experimental conditions.}- Control of the ambient conditions is crucial. Many experiments are carried out at room temperature but low temperature experiments are necessary to access particular optical transitions. In simple bath cryostats the sample is kept in thermal contact with a liquid helium bath at T = 4~K, either via helium exchange gas or a cold finger. The main drawback is that the bath needs to be periodically refilled with liquid helium. Alternatively, closed-cycle cryostats liquify the helium gas using external compressors and allow continuous operation. The external compressor needs to be mechanically decoupled from the sample space to minimise vibrations. A piezo-based 3-axis stage with nanometer step-size is used to place a specific area of the layer of interest in the focal point of the objective. For high mechanical stability of the set-up, a low-temperature compatible microscope objective can be placed inside the cryostat, see FIG.~\ref{fig:box}. \\
\indent The sample holder can also be placed inside the bore of a superconducting coil to apply magnetic fields. In this case the nanopositioners, as well as the objective lens must be made of non-magnetic materials such as titanium and beryllium copper. The sample needs to be placed at the centre of the coil which limits in practice the sample size and also the optical access (beam diameter). The motivation for magneto-optics is manifold such as extracting the valley Zeeman splitting, and hence identifying the origin of new excitonic transitions including interlayer excitons in TMD homo- and heterobilayers, trilayers and bulk \cite{arora2017interlayer,leisgang2020giant}. Furthermore, it is possible to investigate valley polarisation dynamics and 'brightening' of otherwise spin forbidden-transitions \cite{zhang2017magnetic}. Attention should be given to undesired Faraday effects in certain optical components. Magnetic ions in the glass result in an undesired rotation of the linear polarisation in the presence of magnetic fields. This needs to be compensated by other polarisation control elements \cite{wang2016control}.
\subsection{Absorption spectroscopy}
\label{absorption}
Strictly speaking, measuring absorption ($A$) requires to measure both transmittance ($T$) and reflectance ($R$) where $A = 1-R-T$. Transmission measurements require a transparent substrate and a detection path different from the excitation path, for example a separate microscope objective on each side of the sample or alternatively one objective combined with a fibre on the other side of the sample \cite{goryca2019revealing}. In practice often reflectivity is measured as it is the simplest experiment for samples on substrates like SiO$_2$/Si that are not transparent. To get a quantity independent of the optical response of the set-up, one generally measures the reflectivity contrast defined as $(R_\text{sam}-R_\text{sub})/R_\text{sub}$, where $R_\text{sam}$ is the intensity reflection coefficient of the sample with the TMD layer and $R_{sub}$ comes from the same structure without the TMD layer. The spectra obtained this way are for brevity commonly referred to as absorption in the literature.
\\
\indent The optical properties of a material can be seen in a simple classical picture as the interaction between light (electromagnetic radiation), and various types of oscillators \cite{klingshirn2012semiconductor}. In TMD monolayers, the dominating oscillators are exciton resonances. Therefore, in reflectivity different exciton resonances are accessible up to room temperature as they possess strong oscillator strength and high density of states (DOS). This allows to observe the Rydberg series of the A-exciton: 1$s$, 2$s$, 3$s$...(see FIG.~ \ref{fig2}b bottom panel), thus giving a measure of the exciton binding energy and the single particle bandgap \cite{chernikov2014exciton,goryca2019revealing}. Other optical transitions related to defect states in the gap or other exciton complexes which posses weaker oscillator strength and/or comparatively lower density of states are difficult to trace in absorption, although they might appear in photoluminescence emission, as discussed below. \\
\indent \textbf{Monolayers.}- The energy of the A-exciton transition in TMD monolayers is given by the difference of the single particle bandgap (of unbound electrons and holes) and the exciton binding energy. Engineering the dielectric environment (and hence all energy scales linked to the Coulomb interaction) results in significant changes in the single particle bandgap \cite{waldecker2019rigid} and the exciton binding energy of TMD monolayers. But the shift in the global A-exciton transition energy is rather small, as changes in single particle bandgap and the exciton binding energies partially compensate each other. Compare FIG.~\ref{fig2}c for monolayer MoSe$_2$ transition energies with and without hBN encapsulation, which are very close in value. However, the linewidth in absorption is significantly impacted by dielectric disorder \cite{raja2019dielectric}. For instance, bubbles, wrinkles, polymer residues and hBN have different dielectric constants. Therefore, a non-uniform dielectric environment affects the energy of the exciton transitions and the overall shape of reflectivity spectra. Uniform dielectric slabs such as thick (tens to hundreds of nm) hBN layers can be exploited to steer the absorption. The visibility of exciton resonances in absorption is mainly influenced by the thickness of hBN and SiO$_2$, see FIG.~\ref{fig2}b. This is due to thin-film interference effects, the bottom hBN thickness determines how far the monolayer is from the Si/SiO$_2$ interface, which acts as a mirror. The choice of hBN thickness of the heterostructure can be optimised using a transfer matrix approach to increase the visibility of the targeted transitions \cite{robert2018optical}. In this process, the particular energy of the excitonic resonance of the layered semiconductor should be taken into account \cite{robert2018optical}. Recent results in TMD materials placed in front of mirrors show a modulation in the absorption strength of up to 100\% due to interference/cavity effects \cite{horng2019perfect,scuri2018large,back2018realization}. The strong influence of the dielectric environment on the light-matter interaction of atomically thin semiconducting membranes motivates a great potential for sensing applications including novel device architectures with precisely tunable optical properties. \\
\indent \textbf{Multilayers.}-
In general the nature (direct or indirect) and the energy of the bandgap evolves as a function of the layer number for a given material.
In addition, the absorption of layered semiconductors such as black phosphorous \cite{ ling2015renaissance} and ReSe$_2$ \cite{ho1998absorption, zhang2016tunable} reveals information on the crystal structure of these particular materials as it is highly anisotropic in the layer plane, as a direct result of a highly anisotropic lattice structure.
In TMD multi-layers and even bulk strong excitonic features are reported even at room-temperature in early studies \cite{wilson1969transition}. In addition to these features typically attributed to the intralayer A- and B-exciton, more recently the observation of interlayer excitons (formed by carriers in 2 adjacent layers) has been reported in absorption of bulk samples \cite{arora2017interlayer,horng2018observation,arora2018valley}. Absorption of interlayer excitons is also reported in homobilayers and homotrilayers of MoS$_2$ \cite{Gerber:2019,slobodeniuk2019fine}. In these systems the transition energy of the absorption can be tuned through the application of an electric field perpendicular to the layers (Stark shift) over 120~meV and the interaction between interlayer and intralayer excitons can be investigated \cite{leisgang2020giant,sung2020broken,lorchat2020dipolar}.
\begin{figure*}
\centering
\includegraphics[width=169mm]{Fig4}
\caption{\textbf{Moiré interlayer excitons in heterobilayers.} \textbf{a,} Two different monolayer materials stacked vertically that display a moir\'e pattern due to slight lattice mismatch and twist angle. Different local atomic alignments in the heterostructure show different optical properties \cite{yu2017moire,shimazaki2019moir,ciarrocchi2019polarization,shimazaki2020strongly}. Three spots labelled 1, 2 and 3 are chosen for PL measurements on MoSe$_2$ monolayer, WSe$_2$ monolayer and MoSe$_2$/WSe$_2$ heterostructure, respectively, and corresponding spectra are shown in \textbf{c}. \textbf{b,} SEM imaging of hBN-encapsulated twisted WSe$_2$ bilayers with a spatially varying reconstruction pattern due to the interaction between the respective layers after stacking \cite{sushko2019high}. \textbf{c,} Example of photoluminescence spectrum from an hBN-encapsulated MoSe$_2$ monolayer (solid red curve), MoSe$_2$/WSe$_2$ heterostructure (solid black curve) and WSe$_2$ monolayer (solid blue curve) \cite{ciarrocchi2019polarization}. Intralayer exciton emission is observed from the MoSe$_2$ and WSe$_2$ monolayers. The interlayer exciton emission (IX$_1$ and IX$_2$) appear in energy below the intralayer resonances from the heterostructure. \textbf{d,} Electrical control of interlayer excitons in MoSe$_2$/WSe$_2$ heterobilayer. Colormap of PL emission spectra as a function of applied gate voltages $V_{TG}$ and $V_{BG}$ when sweeping at constant doping. Stark shift of the interlayer excitons in applied electric fields perpendicular to the layers showing out-of-plane electric dipole \cite{ciarrocchi2019polarization}. \textbf{e,} Polarisation-resolved second harmonic generation (SHG) intensity of the individual monolayers from a different publication compared to panels \textbf{c,d}, indicating the armchair directions of the monolayers which determine the twist angle ($\theta$) between the WSe$_2$ and MoSe$_2$ layers \cite{nagler2017giant}.}
\label{fig3}
\end{figure*}
\subsection{Photoluminescence spectroscopy}
Luminescence experiments are widely used for studying the macroscopic optical properties of materials as well as their microscopic electronic excitation, for the evaluation of crystalline quality (presence of defects) and for testing novel optoelectronic devices \cite{pelant2012luminescence, haunschild2010quality}. Luminescence is defined as a surplus of the electromagnetic radiation (light) emitted by a solid, in addition to its equilibrium radiation described by Planck’s law. This surplus energy is transformed into detectable luminescence radiation. During the process of luminescence the electrons are excited to higher energy states (by a \textit{light}-source in the case of \textit{photo}-luminescence). Subsequently the carriers relax in energy for example through phonon emission, followed by photon emission. The succession of all involved relaxation and recombination events lasts a relatively long time, which is a main difference compared to other types of so-called secondary radiation: reflected light and scattered light (for example Raman). Once the material is excited with a light pulse, the luminescence continues to decay for some time and can be recorded in time-resolved photoluminescence \cite{pelant2012luminescence}, see a review on time-resolved spectroscopy \cite{balocchi2012time} for technical details. These experiments give information on carrier relaxation and recombination times. In addition to time-resolved photoluminescence (an incoherent technique), also important information can be gleaned from coherent spectroscopy such as four-wave mixing and two-color pump-probe experiments \cite{jakubczyk2016radiatively,moody2015intrinsic,hao2016coherent}. \\
\indent In absorption spectra optical transitions with large oscillator strength and high density of states dominate. In contrast, the emission spectrum given in photoluminescence experiments can be dominated by other transitions as these experiments probe the population of a state. Typically, optical transitions at lower energies are detected in PL as carriers relax towards these lower energy states before radiative recombination. It is therefore instructive to compare emission with absorption spectra (measured in reflection geometry) as in FIG.~\ref{fig2}b. For the material investigated, monolayer WSe$_2$, the main bright exciton transition (A:1s) that dominates in absorption is not generally the strongest feature in PL emission, as electron-hole pairs can relax towards states lying at lower energy to recombine \cite{wang2017plane}. \\
\indent \textbf{Sample temperature.}- At room temperature, the main transitions such as the A- and B-excitons in TMDs will be broadened due to scattering with phonons, compare spectra at T=4~K and 300~K in FIG.~\ref{fig2}c. Carriers or excitons are mobile at high temperature and defect potentials in the lattice with only shallow confinement energy will not act as efficient trapping sites. At low temperature phonon absorption is reduced and the linewidth reveals the sample quality (inhomogeneous broadening versus homogeneous broadening \cite{fang:2019control}). This is illustrated in FIG.~\ref{fig2}c. Spectral lineshape and main emission energy can change compared to high temperature. Carriers can get trapped at sufficiently deep defect potentials. Also excitonic complexes with lower binding energy, such as trions, are now stable and contribute to the PL signal, see FIG.~\ref{fig2}b,c,f for three different materials \cite{moody2015intrinsic,cadiz2017excitonic,Robert:2017,shree2018observation}.\\
\indent \textbf{Excitation power.}- Laser power plays an important role as it controls the number of photoexcited carriers. Strictly speaking, the power density (defined as the average power per unit area) is directly related to the photogenerated exciton density in layered semiconductors. It is inversely proportional to the square of the focused beam radius, thus it is directly related to the numerical aperture of the objective lens and laser wavelength. Let us take an example of a sample with a finite value of defect density: for a tightly focused beam (spot size $\approx$ 1~$\mu$m), at low laser power (typically hundreds of $n$W) all carriers can be trapped at defect sites and the PL signal of free excitons is not visible, as in FIG.~\ref{fig2}f, lower panel. Increasing laser power (few $\mu$W) fills all defect sites and free exciton PL can be measured in addition to defect emission, see FIG.~\ref{fig2}f upper panel. Further increase in power (tens to hundreds $\mu$W) will lead to such a high concentration of excitons that exciton-exciton interactions start to play a role. At high excitation density very different phenomena such as biexciton formation and exciton-exciton annihilation can be studied depending on the material \cite{nagler2018zeeman,sun2014observation,barbone2018charge,paradisanos2017room}. Understanding emission for high exciton (carrier) concentrations is crucial for applications such as lasing and for investigating collective states such as condensates \cite{wang2019evidence,sigl2020condensation}.\\
\indent \textbf{Sample quality and dielectric environment.}- Two monolayer samples exfoliated from the same quality bulk crystal but placed on different substrates can exhibit very different PL characteristics, see FIG.~\ref{fig2}c for a comparison of MoSe$_2$ on SiO$_2$ and on hBN, respectively.
The total line broadening comes both from homogeneous contributions given by $\approx$ 1~ps lifetime (below 1 meV linewidth) and inhomogenous contribution from sample imperfections (impurities, defects, interface, substrate, etc) \cite{jakubczyk2016radiatively}.
The key role of high quality hBN buffers with low defect density is to provide atomic flatness \cite{Taniguchi:2007a} for monolayer deposition and a very clean, homogeneous dielectric environment. The hBN bulk bandgap of 6~eV \cite{cassabois2016hexagonal} is high enough to use it as an essentially transparent encapsulation layer for many materials, providing an ideal environment to address intrinsic properties of 2D-TMDs and preserving good optical quality of air-sensitive materials such as CrI$_3$ or black phosphorous. As the inhomogenous broadening is largely suppressed in high quality samples of TMD monolayers, the linewidth starts to be an indication of the exciton lifetime (homogeneous broadening) and can be tuned by carefully choosing the encapsulating hBN thickness - placing the monolayer on a node or anti-node of the electromagnetic field in the multilayer structure \citep{fang:2019control}. \\
\indent PL emission is sensitive to the presence of wrinkles or bubbles which induce strain and localisation potentials in encapsulated monolayers. Small strain ($1\%$) in monolayer induces a rather significant (about 50 meV) shift in bandgap energy \cite{zhu2013strain} which can explain shifts in absolute emission energy from sample to sample and for different areas of the same sample. Therefore, in practice PL emission intensity as well as linewidth varies on different location of TMD monolayers, see FIG.~\ref{fig2}b when the detection spot (1 $\mu$m diameter) is scanned across a WSe$_2$ monolayer. Remarkably, exciton emission energy that does \textit{not} vary as a function of the detection spot position has been observed in hBN-encapsulated MoSe$_2$ monolayers, suspended over a trench, see FIG.~\ref{fig2}d,e \cite{PhysRevLett.124.027401}, indicating a stable, inert environment.\\
\indent The general study of the dielectric environment, surface quality, flatness, charging events and their impact on optical properties is very important also for other nanostructures such as carbon nano-tubes \cite{berger2009optical,hirana2010strong,ai2011suppression,noe2018environmental,raynaud2019superlocalization} and layered perovskites \cite{blancon2018scaling} which show strong excitonic effects. \\
\indent\textbf{Interference effects.}- Layered materials are usually placed on a substrate such as Si with an SiO$_2$ layer of typically 85~nm. As discussed earlier for white light absorption experiments, optical interference will also be important for the laser excitation beam and the PL emission as a function of the SiO$_2$ thickness and possibly the encapsulation layer thickness. Reversely, for constant SiO$_2$ thickness absorption and emission intensities and directivity will depend on the wavelength, as detailed in \cite{lien2015engineering,zhang2015interference}. In practice the thickness of the SiO$_2$ layer on top of Si is chosen to maximise optical contrast of monolayers already during sample fabrication, as discussed in detail for graphene on SiO$_2$ \cite{roddaro2007optical}. \\
\indent\textbf{Emission dynamics.}- The detected luminescence signal from a layered semiconductor is the result of an intricate interplay between radiative and non-radiative energy relaxation. Time-resolved PL can be performed using a pulsed laser excitation and measuring the recombination (emission) time. In clean TMD samples the strong exciton oscillator strength leads to an intrinsic radiative lifetime of the order of 1~ps at low temperature \cite{lagarde2014carrier}. Localised emitters recombine considerably slower and also dark excitons have a lifetime up to 3 orders of magnitude longer. In the time domain, low temperature measurements on high quality samples allow spectrally isolating each transition (either with bandpass filters or with a spectrometer) and then studying the emission dynamics of each optical transition separately \cite{Robert:2016a,Robert:2017,fang:2019control}.\\
\indent \textbf{Optical dipole orientation.}- In TMD monolayers the main optical transitions have an in-plane optical dipole, they therefore emit light normal to the monolayer plane. However, in addition to these bright (spin-allowed) transitions also excitons that have an out-of-plane optical dipole emit light \cite{wang2017plane}. Out-of-plane dipole transitions are also prominent in InSe \cite{bandurin2017high,brotons2019out}. Due to the small sample dimensions most experiments are carried out in a microscope using an objective with high ($>$0.8) $NA$, see inset on microscope objective in FIG.~\ref{fig:box}. As a result, PL emission containing out-of-plane and also in-plane components in the monolayer are detected. In WSe$_2$ and WS$_2$ monolayers, dark excitons are prominent, see FIG.~\ref{fig2}b and lead to exotic, higher order complexes (such as biexcitons made up of a dark and a bright exciton and so-called dark trions \cite{barbone2018charge,PhysRevLett.124.196802}), that can be identified by monitoring the orientation of the exciton dipole. The role of out-of-plane dipole emission is also investigated for quantum emitters in WSe$_2$ \cite{luo2020exciton}. Brightening of (spin-) dark states due to an increased mixing of the spin-states in WSe$_2$, WS$_2$, MoSe$_2$ and MoS$_2$ monolayers can be observed in low-temperature magneto-PL experiments by applying strong (ideally several tens of T) in-plane magnetic fields \cite{robert2017fine,lu2019magnetic,robert:2020measurement}.\\
\indent \textbf{Multilayers.}- PL spectroscopy is useful also in TMD heterobilayers with type II alignement for the examination of spatially indirect interlayer exciton (see FIGs.~\ref{fig1}c and \ref{fig3}c) with large binding energies ($\approx$ 150 meV) \cite{rivera2015observation,rivera2018interlayer,jauregui2019electrical}. A long period moir\'e pattern (see sketch in FIG.~\ref{fig3}a) offers new directions to explore and control exciton arrays in twisted TMD heterobilayers from potentials that trap individual excitons to the formation of minibands. This allows physics related to the Mott-insulator \cite{shimazaki2020strongly}, for potential applications in quantum optoelectronic devices \cite{andersen2019moir,sung2020broken}. Some key characteristics of interlayer excitons include a long-lifetime (ns), a wide transition energy tunability that ranges over several hundreds of meV via applied electric fields, see FIG.~\ref{fig3}d and a characteristic Zeeman splitting when compared to intralayer excitons \cite{arora2018valley}.
\subsection{Photoluminescence excitation spectroscopy}
In photoluminescence excitation spectroscopy (PLE) the PL emission intensity for a chosen energy is recorded for different photon excitation energies. Tunable lasers or powerful white light sources are used as an excitation source. The linewidth and tuning step of the source will determine the spectral resolution of the PLE experiment. The measured PL intensity will depend on two factors (i) the absorption strength at the excitation energy and (ii) the efficiency of energy relaxation followed by radiative recombination (in competition with non-radiative channels). This combined dependence on both absorption and energy relaxation (often through phonon emission) make PLE spectroscopy a very interesting tool for several investigations : \\
\indent \textbf{Interlayer excitons.}- A PL signal enhancement of interlayer excitons is observed when the laser excitation energy is resonant with intralayer states in one of the layers, confirming that interlayer excitons form via charge transfer processes between the layers \cite{rivera2015observation,hong2014ultrafast,ciarrocchi2019polarization}, see FIGs.~\ref{fig1}d and \ref{fig3}c. In general electronic coupling or charge transfer between layers can be investigated by tuning a laser in resonance with an electronic transition in one layer and monitoring PL emission at an energy corresponding to the adjacent layer or to the heterostructure.\\
\indent \textbf{Measuring excited exciton states.}-
PLE can be also used to establish a link between optical transitions with similar microscopic origin within the same monolayer. In MoS$_2$ monolayers, B-exciton states energetically overlap with the excited A-exciton states (A:2$s$, A:3$s$..,). PLE spectroscopy allows to distinguish the excited states by collecting the emission intensity of the ground state, A:1$s$, as a function of the excitation laser energy, scanned over the energy of A:2$s$, A:3$s$, etc. Besides states with $s$-symmetry, also $p$-states can be examined. To access $p$-states, two-photon absorption processes are necessary and therefore the laser energy needs to be tuned to half of the transition energy \cite{cassabois2016hexagonal}. The identification of high-excited exciton states in one and two-photon-PLE is a powerful method to evaluate the impact of different dielectric environments on the energy evolution of the exciton states. Furthermore, it is possible to extract the exciton binding energies \cite{hill2015observation} and investigate predictions of splittings of the $p$-exciton states \cite{srivastava2015signatures}. However, one should note that the crystal symmetry or disorder effects can mix $s$ and $p$ exciton states \cite{glazov2017intrinsic,berghauser2016optical}. \\
\indent \textbf{Identification of dominant phonon modes for energy relaxation.}- In addition to key information on absorption, PLE is used to identify efficient relaxation channels. In PLE experiments on MoSe$_2$ monolayers a periodic oscillation in energy is observed over an energy range without any expected exciton resonance (roughly constant absorption \cite{wang2015exciton}). These maxima are all equally spaced in energy by longitudinal acoustic phonons at the M point of the Brillouin zone, LA(M), revealing the efficient energy relaxation of excitons through emission of LA(M) phonons \cite{chow2017phonon,shree2018observation}. This experimental observation was possible due to the spectrally narrow excitation source, resolving fine separations between different peaks related to phonon emission \cite{shree2018observation,PhysRevB.93.155407}.
\section{Accessing spin-valley polarisation in optical spectroscopy}
\label{polarization}
The symmetry of the electronic states in monolayers and multilayer crystals governs the optical selection rules for light polarisation in emission and absorption \cite{xiao2012coupled,yu2017moire}, see FIG.~\ref{fig1}b, as studied since several decades for semiconductor nanostructures \cite{dyakonov2017spin}. For polarisation analysis, linear polarisers and waveplates can be inserted in the detection and excitation path of the set-up, see FIG.~\ref{fig:box} for practical details. Exciting a system with polarised light can address a specific spin or valley state, see FIG.~\ref{fig1}b. The emitted light gives information on spin and valley dynamics in time-integrated PL experiments. The circular polarisation in time-integrated experiments depends on the exact ratio of PL emission time $\tau_{PL}$ versus depolarisation time $(\tau_{depol})$ as
$P_c = P_0/(1 + \frac{\tau_{PL}}{\tau_{depol}})$, where $P_0$ is the initially generated polarisation which could depend on the excitation energy \cite{kioseoglou2012valley,tornatzky2018resonance}. \\
\indent \textbf{TMD monolayers.}- The interband transitions in monolayer materials such as MoS$_2$ are governed by chiral selection rules as optical transitions in the $K^+$ ($K^-$) valley are $\sigma^+$ ($\sigma^-$) polarised. For neutral, bright excitons the intrinsic lifetime is of the order of 1~ps, so from time-integrated PL experiments that report $P_c$ values in the order of 50\% one can infer that $\tau_{depol}$ is at least of this order of magnitude. In practice more sophisticated pump-probe measurements reveal very short valley lifetimes for neutral excitons \cite{jakubczyk2016radiatively}. The ratio $\frac{\tau_{PL}}{\tau_{depol}}$ can be tuned by placing monolayers in optical microcavities \cite{dufferwiel2017valley, dufferwiel2018valley}. Longer valley lifetimes in monolayers are reported for resident carriers \cite{dey2017gate} not excitons, measured with pump-probe techniques such as Kerr rotation employed for probing polarisation in semiconducting or metallic nano-structures \cite{mak2019probing}.\\
\indent \textbf{TMD heterobilayers.}- Optical spectroscopy can be used to probe the local atomic registry, i.e. how metal and chalcogen atoms are aligned in the top with respect to the bottom layer \cite{yu2017moire,seyler2019signatures,ciarrocchi2019polarization}. Here information can be gleaned on the formation of nano-scale, periodic moir\'e potentials, see sample sketch in FIG.~\ref{fig3}a. But similar to bilayer graphene, reconstruction can occur when two TMD layers are brought in contact, which can be visualised using imaging techniques such as transmission electron microscopy (TEM) or scanning electron microscopy (SEM). Recently Sushko et al. \cite{sushko2019high} reported SEM imaging of hBN-encapsulated twisted WSe$_2$ bilayers showing that a spatially varying reconstruction pattern develops due to the interaction between the respective layers after stacking, see FIG.~\ref{fig3}b. Polarisation selection rules probed in PL also carry information on different stackings (H-type or R-type for 60$^{\circ}$ or 0$^{\circ}$ twist angle, respectively) \cite{yu2018brightened}. Therefore, polarisation-resolved optical spectroscopy together with direct atomic-resolution imaging of the lattice is a very powerful combination for analysing the formation of moir\'e potentials \cite{andersen2019moir}. The PL experiment samples over a spot diameter of 1 $\mu$m, whereas moir\'e potentials can occur with a periodicity of nanometers (see FIG.~\ref{fig3}a), which leads to averaging effects. The intrinsic lifetime of interlayer excitons is of the order of ns at low temperature and not ps as in monolayers, which allows for imaging exciton and polarisation spatial diffusion in PL maps \cite{rivera2016valley, unuchek2019valley}. The physics of both intralayer and interlayer excitons can be accessed in the monolayer and bilayer regions of the same sample, as in FIG.~\ref{fig3}a and c. \\
\indent \textbf{Experiments in applied magnetic fields.}- The circular polarisation can be manipulated by applying external magnetic fields \cite{molas2019probing,srivastava2015valley}. Interesting examples are heterobilayers, where a giant Zeeman splitting of 26~meV at B = 30~T for interlayer excitons induces near-unity valley polarisation measured in PL emission \cite{nagler2017giant}. In monolayer MoSe$_2$ a field of 7~T results in near unity polarisation of electrons probed in absorption and emission \cite{PhysRevLett.118.237404}.
\section{Optical techniques for accessing crystal quality and orientation}
\label{orientation}
\indent \textbf{Raman spectroscopy} is based on the analysis of laser light scattered by a material. During this process the crystal typically absorbs (or emits) energy in the form of lattice vibrations - phonons. The analysis of the scattered light's energy and polarisation reveals information on the crystal symmetry and quality, doping and where applicable alloying and stacking in multilayers. A typical PL set-up (shown in FIG.~\ref{fig:box}) can be conveniently adapted to collect the Raman spectra by selecting a suitable set of filters according to the wavelength of the excitation laser. Typically, filters for Raman spectroscopy reject the excitation laser with a cut-off frequency of few tens of $cm^{-1}$ ($\approx$ 10~meV from the single mode energy).
The selection of laser wavelength $\lambda$ has an important impact on the spectral sensitivity as the intensity of the Raman signal is proportional to $\lambda^{-4}$.
Key parameters of the excitation laser include spectral linewidth ($\leq$1~MHz), frequency and power stability, spectral purity ($\geq$65~dB side-mode suppression ratio), beam quality (close to Gaussian) and output power.
The excitation wavelength in the Raman scattering of TMDs is also important because of the presence of excitonic states. When the photon energy matches the transition energy of a real state it gives rise to a strong signal enhancement and appearance of new features, associated to symmetry dependent electron-phonon interactions (resonant Raman scattering) \cite{PhysRevLett.114.136403,scheuschner2015interlayer}.\\
\indent Information on the structural phase and composition of materials can be obtained by means of Raman spectroscopy \cite{oliver2017structural}.
This allows for example investigating currently debated link between ferromagnetic ordering and structural phase transitions in CrI$_3$ \cite{ubrig2019low} as a function of temperature. Raman spectroscopy can reveal anisotropy in the crystal structure of, for example, ReSe$_2$, which can be directly linked to measurements of anisotropic optical absorption in the same material \cite{wolverson2014raman}.
Electron-phonon interactions can have significant effects on the Raman frequencies. As a result, doping effects can be effectively monitored in TMD monolayers for example with the out-of-plane phonon, $A'_{1}$, due to its strong electron-phonon coupling \cite{chakraborty2012symmetry,bertolazzi2017engineering}. It is also possible to extract quantitative information about the presence of uniaxial strain since the in-plane phonon energy, $E'$, decreases with applying tensile strain and a splitting occurs (degeneracy is lifted) \cite{conley2013bandgap}. An estimation of the monolayer crystal quality and presence of defects can be realised due to the activation of defect-induced zone-edge phonon modes, such as the LA(M) \cite{mignuzzi2015effect}. Furthermore, disorder and interference effects originating from the substrate impact the intensity and spectral shape of the optical phonons in the monolayer \cite{buscema2014effect}. A powerful and reliable means to determine the number of TMD layers with atomic-level precision is to measure the energy difference between the two main vibrational modes ($E$ and $A$ phonons), affected by interlayer interactions \cite{lee2010anomalous}. Apart from the high-frequency ($\geq$ 80~$cm^{-1}$) spectral range, the number of layers can be identified by collecting optical signatures of the rigid layer vibrations (breathing and shear modes) in the ultra-low frequency range \cite{zhang2013raman}.
In this case, important information on the interlayer interaction and determination of the stacking order in multilayers can be obtained \cite{van2019stacking}. In TMD multilayers resonant Raman spectroscopy can also provide a fingerprint of the extension of excitons over several layers, as otherwise symmetry forbidden modes are activated for the so-called C-exciton region \cite{scheuschner2015interlayer} in energy above the A- and B-exciton. Raman spectroscopy can also be used to investigate the competition between formation of periodic moir\'e potentials and local reconstruction (compare FIGs.~\ref{fig3}a and b) in artificially stacked WSe$_2$/MoSe$_2$ and bilayer MoS$_2$ as a function of twist angle \cite{holler2020lowfrequency,debnath2020evolution}.\\
\indent \textbf{Second-harmonic generation} (SHG) is a nonlinear optical process that converts two photons of the same frequency into one photon of twice the orignal frequency. It is a powerful technique to analyse the orientation and symmetry properties of 2D materials. For SHG experiments, the optical set-up is typically coupled to a pulsed laser that is capable of generating sufficient peak power for this nonlinear optical process. The SHG signal depends on the elements of the second-order susceptibility tensor $\chi^{(2)}$ \cite{klingshirn2012semiconductor}, which are non-vanishing for non-centrosymmetric media (i.e. odd number of TMD layers) along the armchair direction of TMDs, see FIG.~\ref{fig1}a. This crystalographic direction can thus be directly determined by rotating the linear polarisation in the experiment. The resulting polar plot of the SHG intensity reveals the crystallographic orientation of the material, useful to precisely measure the relative twist angles (stacking) of homo- and heterobilayers \cite{shinde2018stacking}, see FIG.~\ref{fig3}e. Novel techniques exploit this effect to map with high spatial resolution ($\approx$ 400~nm) the armchair orientation in twisted bilayers \cite{shg2019twist}, as well as in large ($>10^{4}$~$\mu$m$^{2}$) monolayer areas and evaluate their crystal quality since dislocations and grain boundaries can affect the armchair orientation i.e. change the lattice vector \cite{shg2018ultrahigh}. Also the presence of uniaxial strain can be quantified by measuring the SHG intensity along different polarisation directions \cite{mennel2019second}. \\
\indent The SHG response of a material does not only reflect the crystal structure but also depends on the electronic excitations \cite{farenbruch2020magneto}. The efficiency of the SHG signal can be enhanced by several orders of magnitude by selecting the excitation energy to be in resonance with excitonic states of the investigated materials \cite{wang2015giant}. This opens the way for investigating in general the role of electronic excitations (exciton resonances) on the SHG response of a material \cite{seyler2015electrical,PhysRevLett.103.057203}. Using the sensitive SHG response to detect the energy position of electronic transitions is termed 'SHG spectroscopy'. In bilayer CrI$_3$ SHG has been shown to originate from the layered antiferromagnetic order, which breaks both the spatial inversion symmetry and the time-reversal symmetry \cite{sun2019giant} of this centro-symmetric crystal. This makes SHG a highly sensitive probe also for magnetic ordering in layered materials.
|
\section{Introduction}
This paper focuses on the problem of creating schedules for Time-Triggered traffic on the TTEthernet network.
The objective is to minimize the makespan of the Time-Triggered traffic and, thus, maximize the continuous gap for the Event-Triggered traffic.
The aim of the paper is to develop the scheduling algorithm and, consequently, to compare the concept of the makespan minimization with the concept of the porosity optimization proposed by Steiner in~\cite{Steiner2011}.
The TTEthernet network is a network protocol which offers Time-Triggered communication and preserves the backward compatibility with Ethernet.
Ethernet (IEEE 802.3) has been the prevalent technology for home and office networks over the last few decades.
Thanks to its widespread adoption, it has developed into a mature technology offering high bandwidth with cheap and readily available hardware.
\begin{figure}[t]
\resizebox{\columnwidth}{!}
{
\includegraphics{Images/Infrastructure.pdf}
}
\caption{Example of the TTEthernet network infrastructure with routing and scheduling of message $m_1$ from node l to node q}
\label{Fig:Infrastructure}
\end{figure}
Conventional computer networks are mainly used for on-demand data transfer without an immediate physical impact on the real world.
In case of failure, the transmission can usually be repeated without causing major difficulties.
Therefore, the focus is on the bandwidth and efficiency with only moderate demands on reliability.
In contrast, in industrial applications various control loops of physical devices are realized by the network, e.g. data from sensors are transferred to a processing unit which then sends commands to actuators.
Any disturbance can, thus, have immediate effects on the real world with possible severe consequences.
As jitter (variance in transmission times) is detrimental to the function of the control loops, determinism is often required.
In addition, individual devices in the network are often limited in hardware and need to operate in demanding environments.
Due to these differences, different technologies and protocols were traditionally used for conventional computer networks and industrial networks.
Ethernet has been used for home and office networks while various Fieldbus networks (such as CAN~\cite{Davis2013} or Profinet~\cite{Hanzalek2010}) have been used for industrial applications.
However, with the increasing integration of industrial systems, increasing demands on the volume of data transferred and also the maturing and development of the Ethernet, there has been a trend of using Ethernet-based networks for industrial applications as well.
This trend has become even more pronounced with the ever increasing amounts of data transfers necessary to facilitate features like real-time image processing and recognition or communication among individual units in a smart system.
Therefore, extensions of Ethernet are being developed to meet the demands of industrial applications.
An overview of the development of industrial networks is given by \cite{Galloway:Intro}.
TTEthernet is a promising extension of Ethernet, which provides determinism and fault-tolerance while being compatible with standard Ethernet.
Besides the TTEthernet standard, there has been ongoing effort on the standardization of extensions to Ethernet for scheduled traffic by the IEEE 802.1 Time-Sensitive Networking Task Group.
Determinism with the strictest guarantees is achieved through a fixed schedule for the traffic.
Therefore, synthesizing a good (exactly what this means will be discussed later) schedule which meets all the requirements and deadlines is essential for the performance of the network.
Because TTEthernet allows for complex topologies, the scheduling involves additional complexity compared to bus or passive star topologies of networks like FlexRay~\cite{ISOFlexRay} or CAN.
\subsection{TTEthernet Overview}
TTEthernet (TT stands for Time-Triggered) is an extension of Ethernet for
deterministic communication developed as joint project among the Vienna University of Technology~\cite{kopetz2005}, TTTech and Honeywell, and
standardized as SAE~AS~6802~\cite{AS6802} in 2011.
TTEthernet operates at Level 2 of the ISO/OSI model, above the physical layer of Ethernet.
It requires a switched network with fully duplex physical links, such as Fast Ethernet physical link 100BASE-TX or Automotive Ethernet standard 100BASE-T1, so that unpredictable conflicts, while accessing a shared medium, are avoided.
An example of the TTEthernet infrastructure is depicted in Fig.~\ref{Fig:Infrastructure}.
TTEthernet specifies a protocol for clock synchronization and the rules for managing the traffic on the network.
After an initial startup phase when the clocks of the devices in the network are synchronized for the first time, the operation of TTEthernet is periodic.
The clocks are being periodically synchronized to counter any possible clock drift when in steady operation.
This period is called the \emph{integration cycle}.
The messages which follow a deterministic schedule are also periodic.
The least common multiple of their period is called the \emph{cluster cycle}.
TTEthernet integrates traffic of different time-criticality levels into one physical network.
There are three traffic classes in TTEthernet.
These classes, ordered by decreasing priority, are Time-Triggered (TT), Rate-Constrained (RC) and Best-Effort (BE) traffic.
\begin{figure}[t]
\resizebox{\columnwidth}{!}
{
\includegraphics{Images/LinkSchedule.pdf}
}
\caption{Example of communication on a link in one cluster cycle}
\label{Fig:Link_schedules}
\end{figure}
The TT traffic class has the highest priority, and sub-$\mu$s jitter can be achieved (depending on the network devices).
The TT messages are periodic.
We assume that they are strictly periodic (i.e. no jitter is allowed) in agreement with~\cite{TS2015}.
Their schedule is calculated offline and then loaded into the individual devices.
The schedule is repetitive with a hyperperiod of the cluster cycle.
A simple example of the TT traffic together with the RC traffic on one direction of a physical link is presented in Fig.~\ref{Fig:Link_schedules}.
In the figure, the integration cycles are situated in rows and x-axis represents the time instants in the particular integration cycle.
The figure shows that messages $TT_1$, $TT_3$ and $TT_4$ have the same period (twice the integration cycle) and $TT_2$ has another period (four times the integration cycle).
The length of the cluster cycle is equal to four times the length of the integration cycle.
The dark message at the beginning of each integration cycle is the synchronization message.
The synchronization message is used for periodical clock synchronization among nodes.
This message represents the time allocated for the synchronization mechanism of the TTEthernet and it does not participate in the optimization process.
The schedule also provides temporal isolation and enables fault tolerance.
This is due to the fact that not only the sending of a frame is scheduled, but its reception is scheduled as well.
If a TT frame arrives outside the \emph{acceptance window} (i.e. the time the frame is supposed to arrive considering the synchronization inaccuracy), it is discarded by the receiver.
This mechanism is called the \emph{temporal firewall}.
For traffic with less strict precision requirements, the RC traffic class can be used.
This traffic class conforms to the ARINC~664p7 specification \cite{ARINC664p7} (also called AFDX).
It offers greater flexibility because only the frame routing needs to be determined offline.
The messages themselves are event-driven within some limitations.
The RC traffic represents the event-triggered communication.
Thus, it does not follow any schedule known in advance.
It is organized in so-called \emph{virtual links}.
As stated in \cite{AFDXdetailed}, a virtual link is an analogy to the ARINC~429~\cite{ARINC429} single-source multi-drop bus.
The virtual link determines the routing of the messages associated with it.
Furthermore, there are two parameters: the maximum allowed frame size and the bandwidth allocation gap, associated with each virtual link.
The bandwidth allocation gap represents the minimum allowed length of an interval between consecutive frames on the virtual link.
This effectively limits the bandwidth of the virtual link.
In return for this limitation, the maximum possible delay of any RC message can be calculated offline.
Standard Ethernet traffic can also be transmitted through the network.
Even standard Ethernet devices, unaware of TTEthernet, can communicate through the network.
Such traffic is called the Best-Effort (BE) traffic and has the lowest priority.
When the TT traffic is used together with other traffic classes, a TT frame could be delayed by another RC or BE frame.
This happens when a TT frame arrives while an RC or BE frame is in transmission.
Then, unless the transmission of the RC or BE frame can be interrupted, the TT frame has to wait until the transmission is finished.
Therefore, a method of handling such situations, called the \emph{traffic integration policy}, is needed.
There are three policies for the integration of different traffic classes as mentioned by \cite{TS2012}: Pre-emption, Shuffling and Timely block.
The Timely block integration policy, which causes no extra delay of the TT traffic, is used in this paper.
In this case, an RC frame can only be transmitted if there is enough time for the transmission of the entire frame before the next TT frame is scheduled.
If there is insufficient time, the transmission of the RC frame is postponed until after the TT frame is transmitted.
It additionally means, that the TT traffic follows the schedule without any delays.
\subsection{Related works}
Many recent publications focus on the scheduling of Time-Triggered communication on the Ethernet network.
The most significant research in this area was done by Steiner~et~al.
Their first paper~\cite{Steiner2010} described the basic constraints for scheduling the TT traffic.
The scheduling problem was solved by the reformulation of the constraints to the SMT (Satisfiability modulo theories) model.
The native SMT formulation was used just for small cases with up to 100 message instances.
The second approach presented allows one to find a schedule for bigger instances.
It uses the SMT solver as a backend for scheduling smaller groups of frames, into which the frames are divided.
After a group of frames is scheduled, the positions of the corresponding frame instances are fixed, and the next group is scheduled iteratively.
The paper does not consider other traffic classes, and the only criterion is to find any feasible schedule.
The SMT formulation of the Time-Trigged message scheduling problem has been adapted to 802.1Qbv~\cite{Craciunas2016} in 2016.
In~\cite{Steiner2011}, which builds on and extends~\cite{Steiner2010}, traffic from other classes than just the TT class is considered.
To reserve capacity for rate-constrained traffic, the concept of schedule porosity, i.e. inserting blank slots reserved for the RC traffic into the schedule, is introduced.
The pessimistic time analysis for the RC traffic is proposed to evaluate the concept.
Tamas-Selicean~et~al. published a new method for calculating the worst-case delay~\cite{TS2015} which promises much tighter estimates compared to the previous work.
The improved precision comes, however, at the cost of much more expensive computation.
As noted by~\cite{Steiner:NewHorizons}, porosity scheduling has the disadvantage that gaps introduced at the beginning of the scheduling process do not consider the profile of the RC traffic.
The idea of using the metaheuristic approach for scheduling the TTEthernet communication by TabuSearch was proposed in~\cite{TS2012} and further extended and more intensely evaluated in~\cite{TS2015designopt}.
The operators of the TabuSearch~\cite{Baghel2012} algorithm take the RC worst-case delay calculation into account and try to change the current schedule in such a way that all the real-time constraints imposed on the RC traffic are satisfied.
Craciunas~et~al. took the given TTEthernet communication schedule into account while scheduling the tasks on the communication endpoints in~\cite{Craciunas2014}.
They extended the work and introduced the combined network-level and task-level scheduling in~\cite{Craciunas2015}.
The CPU tasks are modeled in a very similar way to the network transmission tasks.
A CPU is modeled as another physical link in the network.
A task running on this CPU is modeled as a frame which needs to be transmitted over this physical link in one direction, and its transmission time is equal to the required CPU processing time.
We have formulated our problem as the Resource Constrained Project Scheduling Problem (RCPSP~\cite{Neumann2012}), as will be described later, and its Multi-Mode version (MMRCPSP)~\cite{Heller2016}.
RCPSP is already a well-studied optimization problem.
The survey for the problem and its related variants can be found in~\cite{Hartmann2009}.
The Multi-Mode version, where activities have several alternative modes with different parameters, was studied by Schnell~et~al. in~\cite{SchnellHartl:MMwithPrec}.
The authors developed the exact algorithm by extending the SCIP~Solver~\cite{SCIP}.
Added constraint handlers (i.e. functions for constraints propagation) allowed to directly cooperate between a low-level constraint integer program solver and high-level MMRCPSP constraints and objective.
Another exact approach is described in~\cite{Sitek2016} where the constraint-based modeling tool CP Optimizer was used to solve the RCPSP problem.
The expressive power and universality of the CP Optimizer allows one to easily extend or modify the RCPSP model, which is used for justification of the makespan and porosity optimization in the paper.
\subsection{Paper outline and contribution}
The paper is organized as follows: Section~\ref{Sec:ProblemStatement} describes the studied problem of the TT message scheduling in the TTEthernet network.
In Section~\ref{Sec:Algorithm}, the proposed method is described consisting of a message routing algorithm, a load-balancing heuristic and an RCPSP based formulation of the scheduling problem.
The method is evaluated from the resulting schedule length point of view and also the RC traffic worst-case delay point of view in Section~\ref{Sec:ExperimentalResults}. Section~\ref{Sec:Conclusion} concludes the paper.
The main contribution of the paper is the investigation of a new concept for creating schedules of the TT traffic in the TTEthernet network so that, in opposite to the preceding studies, it focuses on shortening the makespan (the latest completion of any message transmission among all integration cycles and links - see Fig.~\ref{Fig:Link_schedules}) of the TT traffic instead of introducing porosity (blank slots in the TT traffic schedules that are reserved for RC and BE traffic).
The makespan serves as a good measurement for the schedule quality evaluation.
The proposed idea has been inspired by the FlexRay communication scheme, where the Time-Triggered and Event-Triggered segments are separated.
A further contribution is the novel formulation of the TTEthernet scheduling problem as an RCPSP problem.
Both contributions are evaluated and discussed.
\section{Problem statement}
\label{Sec:ProblemStatement}
This paper aims to design a method for finding feasible periodic schedules for Time-Triggered communication on the TTEthernet network such that the maximal part of the remaining bandwidth can be preserved for RC and BE messages.
Each message $m_i$ from a set of the TT messages $M$ has assigned:
\begin{itemize}
\item $t_i$ - period
\item $c_i$ - message length in the number of bits considering headers and interframe gap
\item $d_i$ - deadline
\item $r_i$ - release date
\item $q_i$ - identifier of the transmitting node
\item $Q_i$ - set of the receiving nodes
\end{itemize}
The length of the resulting schedule is determined by the cluster cycle $cc$ (40\,ms in Fig.~\ref{Fig:Link_schedules}).
The length of the integration cycle $ic$ (10\,ms in Fig.~\ref{Fig:Link_schedules}) is assumed to be the greatest common divisor of message periods (i.e. $ic = gcd_{i | m_i \in M}(t_i)$).
In other words, all the periods $t_i$ has to be an integer multiple of the length of integration cycle $ic$.
The schedule is so-called strictly periodic, which means that the next occurrence of message $m_i$ in a particular link (further called \emph{message occurrence} - see Fig.~\ref{Fig:Messages_terminilogy}) appears in the schedule exactly $t_i$ time units after the current one.
The positions of all message occurrences of message $m_i$ in the strictly periodic schedule can be deduced from the position of the first message occurrence.
The transmission time of message $m_i$ has to be smaller than or equal to the integration cycle and its length $c_i$ does not exceed the maximal Ethernet frame length of 1530\,bytes.
It would not be possible to send a synchronization message otherwise.
Deadline $d_i$ and release date $r_i$ are assumed to have the value in the range $0 \leq r_i \leq d_i \leq t_i$.
Each node $e_i$ in the network has its identifier assigned.
The nodes are divided into two classes: \emph{redistribution nodes} and \emph{communication endpoints}.
The communication endpoints are nodes that generate or process the data (e.g. sensors, actuators, control units and other ECUs).
Thus, the identifier of any communication endpoint can be assigned to message $m_i$ as transmitter $q_i$ or one of the receivers from set $Q_i$.
The redistribution nodes, on the other side, are switches without any own data to transmit and serve as intermediary nodes for communication.
The TTEthernet infrastructure consists of nodes and links which interconnect them.
Each link $k_{i, j}$ from a set of links $K$ connects two nodes $e_i$ and $e_j$.
This connection covers just one direction of the full-duplex communication.
Therefore, two links $k_{i,j}$ and $k_{j,i}$ model one physical link between nodes $e_i$ and $e_j$.
These two links are independent from the scheduling point of view.
\begin{figure}[t]
\resizebox{\columnwidth}{!}
{
\includegraphics{Images/Messages_terminilogy.pdf}
}
\caption{Visualization of the difference between message occurrence and message instance}
\label{Fig:Messages_terminilogy}
\end{figure}
A feasible schedule has to fulfill the following hard constraints:\\
\textbf{Completeness constraint:} Each message $m_i \in M$ has to be scheduled.\\
\textbf{Contention-free constraint:} Any link is capable of transferring at most one message at a time.\\
\textbf{Precedence constraint:}
A sequence of the links $S_i^q = (k_{l,m}, k_{m,o}, ..., k_{p, q})$ represents the path that message $m_i$ has to go from transmitter $q_i = e_l$ to receiver $e_q \in Q_i$ through redistribution nodes $e_m, ..., e_p$.
An example of such a message transmission is presented in Fig.~\ref{Fig:Infrastructure} where communication endpoints are titled by "ECU" and the redistribution nodes has arrows painted on the top side. The front side of all nodes is labeled by its name.
Only one direction of each physical link is labeled in Fig.~\ref{Fig:Infrastructure} for the sake of simplicity.
The instance of message $m_i$ in link $k_{l,m}$ is denoted as \emph{message instance} $m_i^{l,m}$.
Thus, all the transmissions of some message $m_i$ in one particular link represents the same message instance.
The message occurrence, on the other hand, represents all the transmissions of some message $m_i$ in one particular integration cycle.
The difference between message instance and message occurrence is graphically explained in Fig.~\ref{Fig:Messages_terminilogy}.
The figure shows the detail view on the sub-segment of the network infrastructure from Fig.~\ref{Fig:Infrastructure} with node $e_i$, $e_m$ and $e_o$ only.
The both links of any physical link are labeled here already.
The path $S_i^q$ is not known in advance if the infrastructure does not have a tree topology.
Therefore, part of the optimization process is to find an appropriate path.
The message transition from transmitting node $q_i$ to all the receivers from $Q_i$ has to be accomplished in one integration cycle.
In the real multi-hop networks, each hop introduces a technical delay caused by queuing in the ingress and egress port, etc.
Such a delay in a switch is represented by parameter $\tau$ for TT messages.
The value of $\tau$ can be in range from \SI{1}{\micro\second} to \SI{2,4}{\micro\second} according to the network configuration~\cite{Steinbach2010}.
Sequence $S_i^q$ entails the generalized precedence constraints because message $m_i$ has to be scheduled in link $k_{m,o}$ $\tau$ time units after it is scheduled in $k_{l,m}$ if $k_{l,m}$ precedes $k_{m,o}$ in $S_i^q$.
The coherent TT traffic segment should be compressed as much as possible to preserve the maximum part of the remaining bandwidth for the RC and BE traffic.
This idea follows the practice from the FlexRay bus or Profinet where the TT traffic has its communication segment.
The TT traffic can be scheduled at the beginning of the integration cycle, and the remaining \emph{TT-free gap} (coherent gap in integration cycle without the TT traffic) is preserved for the RC and BE traffic.
The shortest TT-free gap among all links $k_{i,j}$ is denoted as a \emph{critical gap} (see Fig.~\ref{Fig:Link_schedules}).
Considering the constraints above, the goal of the scheduling is to find a schedule for TT messages which maximizes the critical gap and, thus, minimizes its makespan.
\section{Algorithm}
\label{Sec:Algorithm}
The algorithm proposed in the paper is divided into three stages.
In the first stage (Sec.~\ref{ssec:Routing}), the routing of the messages is established.
In the second stage (Sec.~\ref{SSec:ICAP}), the algorithm finds the assignment of the messages to the particular integration cycles.
The transmission times for each message occurrence in each link are scheduled in the last stage (Sec.~\ref{ssec:scheduling}).
\subsection{Determination of Time-Triggered messages routing}
\label{ssec:Routing}
The network topology is often a tree in industrial networks.
It means that there are no cycles and, therefore, there exists only one possible path from a communication endpoint to any other endpoint.
Thus, the routing determination is trivial in such a case.
However, the TTEthernet does not restrict the network topology, according to the specification, to the tree.
The cycles introduce new redundant paths for messages.
The redundant paths can relieve busy links and serve as a backup during a partial network malfunction.
However, the TT~messages have to know which path they are routed through in advance.
Therefore, the first stage of the algorithm finds the routing.
Each hop the message has to pass implies the increase in overall traffic of the network.
The long path also causes the prolongation of the end-to-end delay because the message is delayed in the ingress and egress buffer of each redistribution node in the path.
This means that utilizing the shortest path algorithm provides an efficient routing with the minimal number of hops.
The network topology is transformed to graph where edges represent the physical links and its weight is set to one.
The Floyd-Warshall algorithm~\cite{Floyd1962} is used, consequently, to find a routing among all nodes in the network.
All the messages follow the routing consequently.
Considering that the aim of this routing is to minimize delay caused by the switch hoping, the routing over the shortest path can cause unbalanced load among links.
Thus, there are cases where it is better to choose another routing strategy.
However, this default routing strategy can be substituted by any other routing that follows demands of the particular application.
Other routing strategies can be found for example in \cite{Andrade2016} or \cite{Jouy2014}.
The routing defines the links in which the messages are to be scheduled and specify the precedence relations among message instances of each message.
\subsection{Integration cycle assignment problem}
\label{SSec:ICAP}
The used idea how to distribute messages among the integration cycles comes from the multiprocessor scheduling area.
In the multiprocessor scheduling area, if all the workload of tasks is distributed among the processors evenly, then the schedule makespan has a good chance to be minimal.
Following that, the algorithm tries to distribute the messages among integration cycles evenly.
All the precedence constraints, time lags imposed by the switch delay $\tau$ and real-time constraints are relaxed here.
The integration cycle assignment problem is formulated as the following ILP model:
\begin{equation}
\resizebox{\columnwidth}{!}
{$\displaystyle
\nonumber
\begin{aligned}
& \min_{x_{i,j}}
& & z
& & & \\
& \text{s.t.}
& & \sum_{j} x_{i,j} = 1
& & & \forall i\\
&
& & \sum_{m_i \in k_{l,m}}{c_i^{l,m} \cdot x_{i,j\bmod t_i}} \leq z
& & & \forall j,l,m\;|\;j \in \{0\,...\,\frac{cc}{ic}\}\\
&
& & x_{i,j} = 0
& & & \forall i,j\;|\;d_i < j\cdot ic \\
&
& & x_{i,j} = 0
& & & \forall i,j\;|\;r_i > j\cdot ic + ic \\
& & x_{i,j} \in \{0,1\}; \; z \in R
& & & \forall i,j\\
\end{aligned}
\label{Equ:ILPCLR}
\vspace{0.2cm}
$}
\end{equation}
\addtocounter{equation}{1}
The binary variable $x_{i,j} = 1$ if message $m_i$ is scheduled to the integration cycle $j \in \{0\, ...\, t_i\}$ and zero otherwise.
The parameter $c_i^{l,m}$ represents the transmission time of message $m_i$ on line $l_{l,m}$.
Note, that if the links are configured to have a different bandwidth then the transmission time of the same message varies among the links.
The first constraint assures that the first message occurrence appears in exactly one of the possible integration cycles.
Thus, it satisfies the completeness constraint.
The second constraint makes the variable $z$ to have the value equal to or greater than the time needed to exchange all messages in any integration cycle of any link in the network.
The constraint is evaluated for each resource (each link and each integration cycle in the cluster cycle) such that the transmission time of all message occurrences assigned to the particular resource is summed up.
The resulting total must be less than or equal to variable $z$.
The aim of the ILP model is to find such an assignment that minimizes $z$.
Thus, the maximal time needed for message exchange among all resources is minimized.
The last two constraints force messages to be assigned to the integration cycle which can satisfy the release date and deadline constraints.
The resulting assignment balances the load among the resources, follows the routing of the messages and satisfies the release date and deadline constraints.
\subsection{Link schedule creation problem}
\label{ssec:scheduling}
The RCPSP formulation, inspired by the RCPSP formulation of the traffic scheduling on ProfinetIRT~\cite{Hanzalek2010}, is used for the message scheduling problem.
Thus, let us provide a brief overview for RCPSP first.
\subsubsection{RCPSP Overview}
The RCPSP problem used in the paper is classified as PSm,1,1\textbar temp\textbar $\text{C}_\text{max}$ by Graham's notation.
Translated to common language, the problem is a Project Scheduling problem with \textit{m} resources, one unit of each resource available, and each activity demands at most one unit of the resource.
The problem is constrained by temporal constraints (time lags among activities), and the objective is to minimize the makespan $\text{C}_\text{max}$.
The problem is defined as a sextuplet $(V, p, E, R, B, b)$.
$V = \{A_0, A_1, ..., A_n, A_{n+1}\}$ is a set of non-preemptive \emph{activities} that should be scheduled.
$A_0$ and $A_{n+1}$ are special dummy activities where $A_0$ represents the start of the schedule and $A_{n+1}$ the end of the schedule, respectively.
$p = \{0, p_1, ..., p_n, 0\}$ is the vector of the activities duration.
The dummy activities have zero duration.
The resulting schedule would be shifted otherwise.
$E$ is a set of pairs representing time constraints.
If activity $A_i$ has a temporal relation to activity $A_j$ then $(A_i, A_j) \in E$.
Each pair in $E$ is valued by the start-start time lag $l_{i,j}$.
Note that the time lag between any activity $A_i$ and dummy activity $A_{n+1}$ is equal to the activity duration $p_i$.
The temporal constraints are often visualized by the so-called activity-on-node graph $G(V,E)$.
The set $R = \{R_1, ..., R_q\}$ is a set of $q$ resources considered in the problem.
$B$, consequently, represents the number of units available in the resources.
All the resources are unary ($B_k = 1 \;\;\forall k \in R$) in our case.
Parameter $b$ is a set of activity demands where $b_{i,k}$ represents the amount of resource $R_k$ demanded by the execution of activity $A_i$.
The value of $b_{i,k}$ is binary ($b_{i,k} \in \{0, 1\} \;\; \forall i \in V, k \in R$) in the PSm,1,1\textbar temp\textbar $\text{C}_\text{max}$ problem.
The goal is to assign the start time to the activities such that resource demands and time constraints are satisfied and the start time of activity $A_{n+1}$ is minimal.
An interested reader is referred to \cite{Brucker1999} for more detailed information on the RCPSP problem.
\subsubsection{Formulation of the link schedule creation problem to RCPSP}
The stated TTEthernet scheduling problem can be formulated as the PSm,1,1\textbar temp\textbar Cmax problem, considering the following conditions are met:
\begin{enumerate}
\item Routing: It is known in which links each message appears
\item Integration cycle assignment: It is known in which integration cycles each message appears
\end{enumerate}
The first condition is satisfied by the first stage of the algorithm where the messages are assigned to links according to the routing.
The second condition is also met and the assignment of each message to the integration cycles is obtained from the second stage of the algorithm.
In the RCPSP model~\cite{Neumann2012}, the message instances are represented by the activities.
Thus, the set of message instances is translated to the set of activities $V$ where activity $A_i^{l,m}$ corresponds to message instance $m_i^{l,m}$.
The dummy activities $A_0$ and $A_{n+1}$ are artificially added to $V$.
The duration $p_i^{l,m} = c_i^{l,m}$ of activity $A_i^{l,m}$ expresses the transmission time of message $m_i$ on link $k_{l,m}$.
However, we assume that the bandwidth of all physical links is the same for the sake of simplicity in this paper.
The resulting start time of activity $A_i^{l,m}$ will, consequently, represent the offset of message instance $m_{i}^{l,m}$ in the particular integration cycles in link $k_{l,m}$.
Recall that the integration cycles, in which the message appears, are determined by the first message occurrence.
The set of resources $R$ is obtained from $I$ - the set of integration cycles in the cluster cycle and $K$ - the set of links in the network.
The number of resources used in the model is then $|R| = |I| \cdot |K|$.
Each resource $R_i^{l,m}$ represents the usage of link $k_{l,m}$ in the integration cycle $ic_i$.
The resources have unary availability $B_i^{l,m} = 1$.
Correspondingly, the message demands are unary too, which ensures that the contention-free constraint is satisfied.
Considering that the given conditions are met, the activities demand can be directly derived from the routing and the message to the integration cycles assignment.
The notation $b_{i,j}^{l,m}$, which represents the amount of resource $R_j^{l,m}$ demanded by activity $A_i^{l,m}$, is used from this point further.
If message $m_i$ is to be routed through link $k_{l,m}$, among others, and it is known that the message appears, for example, in $ic_x$, $ic_y$ and $ic_z$ then \mbox{$b_{i,x}^{l,m}=1$}, \mbox{$b_{i,y}^{l,m}=1$}, \mbox{$b_{i,z}^{l,m}=1$}.
All other demands for the activity representing message instance $m_i^{l,m}$ will be equal to zero.
\begin{figure}[t]
\centering
\resizebox{0.65\columnwidth}{!}
{
\includegraphics{Images/Graph.pdf}
}
\caption{Example of activity-on-node graph for outlined message $m_1$ from~Fig.~\ref{Fig:Infrastructure}}
\label{Fig:Graph}
\end{figure}
All the real-time constraints are modeled by set $E$ and time lags.
In order to model release date $r_i$ and deadline $d_i$, it is necessary to know their relative values $\tilde{r}_i$ and $\tilde{d}_i$ with respect to the integration cycle in which the message is transmitted.
The values are obtained by subtracting the integration cycle start time from $r_i$ or $d_i$ respectively.
The release date of the message $m_i$ is transformed to the positive time lag led from the dummy activity $A_0$ to activity $A_i^{l,m}$ where $k_{l,m}$ is the first link in the path the message traverses.
The time lag is equal to the relative release date $\tilde{r}_i$.
The deadline, on the other hand, is transformed to a negative time lag led from activity $A_i^{p,q}$, where $k_{p,q}$ is the last link before receiving the communication endpoint, to the dummy activity $A_0$.
The value of the deadline time lag is equal to the negative value of the relative deadline $\tilde{d}_i$ extended by duration $p_i^{p,q}$ (i.e. $-(\tilde{d}_i + p_i^{p,q})$).
Note that if there is more receivers in set $Q_i$ then it is necessary to add the deadline time lag for each receiver separately.
Technical delays and precedence constraints are represented by the positive precedence time lags between two consecutive message instances $m_i^{l,m}$ and $m_i^{m,n}$.
The precedence time lag consists of duration $p_i^{l,m}$ and redistribution nodes delay parameter $\tau$ in store-and-forward networks.
In Fig.~\ref{Fig:Graph}, the example of the activity-on-node graph for message $m_1$ is presented.
Message $m_1$ can be also observed in Fig.~\ref{Fig:Infrastructure} or in its crop in Fig.~\ref{Fig:Messages_terminilogy} - the one with a thick outline.
The links assumed here have a bandwidth of 10 Mbit/s and the message consists of 672 bits (64 octets of Ethernet frame + 12 octets interframe gap + 8 octets preamble and start of frame delimiter).
The message is transmitted by node $e_l$ and received by nodes $e_q$ and $e_u$.
It traverses the redistribution nodes $e_m$, $e_o$ and $e_p$ and the redistribution node delay $\tau$ is \SI{1}{\micro\second}.
The relative release date $\tilde{r}_1$, expressed by the edge from $A_0$ to $A_1^{l,m}$, is \SI{10}{\micro\second}.
The relative deadline $\tilde{d}_1$, represented by the edge from $A_1^{p,u}$ and $A_1^{p,q}$ to $A_0$, is \SI{4.5}{\milli\second}.
The time lags allow one to limit the latency between the message transmission and reception or the precedence constraints between the messages (e.g. derived from the application level constraints).
However, these constraints are out of the scope of our problem statement and are not used in the paper.
With a given configuration of the RCPSP model, the objective of the RCPSP (i.e. to minimize the maximal makespan) corresponds to the objective of our problem statement.
In other words, the minimal RCPSP makespan assures the maximal critical gap.
The scheduling problem can be formulated as MMRCPSP as well.
In this case, the integration cycle assignment is not necessary to be solved in advance, because it can be formulated as part of the MMRCPSP~\cite{Heller2016}.
However, the complexity of the resulting MMRCPSP model is significantly higher and, thus, the solver needs more time to find a good or even any solution (see Sec.~\ref{ssec:Quality}).
\section{Experimental results}
\label{Sec:ExperimentalResults}
The proposed scheduling method was tested on a PC with Intel\textsuperscript{ \textregistered} Core\texttrademark\;i7-4610M CPU (two cores with 3~GHz and hyper-threading) and 8~GB RAM.
The algorithm uses the CPLEX ILP Solver for solving the Integration cycle assignment problem.
The RCPSP problem was solved by both the CP Optimizer and SCIP Solver.
However, the results obtained by the CP Optimizer were significantly better in all the cases.
Thus, the results of the SCIP Solver are omitted here.
\subsection{Benchmark instance sets}
Eight different benchmark sets were used for testing purposes.
Seven of them were generated by our instance generator.
They represent the artificial instances.
The last instance is the real problem instance obtained from our industrial partner.
The synthetic instances are divided into seven groups according to the number of the used TT messages.
The sets are called according to that - from \emph{Set\_20TT} (set with twenty TT messages) to \emph{Set\_2000TT}.
The length of the integration cycle equals thousand times the number of the TT messages, which approximately caused the utilization of half of the bandwidth.
Each such set contains thirty independently generated benchmark instances.
During the instance creation, the algorithm for generating the network topology is selected randomly.
The generator provides four possible algorithms - the star topology generator, the snowflake topology generator, the Barab\'{a}si-Albert algorithm for random tree generation and a random topology generator (the Barab\'{a}si-Albert algorithm with additional redundant links).
The Barab\'{a}si-Albert algorithm is slightly modified.
The nodes of the network with degree 2 are eliminated to remove long linear network segments which are uninteresting from the scheduling point of view.
The network consists of twenty communication endpoints for all the topologies.
Each link has the bandwidth of 1\;Gbit/s.
The messages have a randomly chosen transmitter and set of receivers from the set of communication endpoints.
Thus, any communication endpoint can send a message to an arbitrary subset of other communication endpoints.
The payload of the messages is taken randomly from interval of 46 to 256 bytes.
The message periods should be harmonic or close to harmonic to keep the length of the cluster cycle reasonable.
Thus, the periods are from the domain $t_i = 2^n3^mic$ where $n \in \mathbb{N}$ and $m \in \{0,1\}$.
The release times and deadlines are also generated randomly such that $r_i < d_i \leq t_i$.
The last benchmark set \emph{Set\_Indust} consists of one benchmark instance given to us by our industrial partner.
The instance evaluated the behavior of the proposed method on instances from a natural environment which often does not directly correspond to the artificial ones.
The network consists of 6 redistribution nodes and 59 communication endpoints and contains alternative paths.
The instance has 1018 messages to exchange with periods from domain $t_i \in \{12.5\text{ms}, 25\text{ms}, 50\text{ms}, 100\text{ms}, 200\text{ms}, 1000\text{ms}\}$ and efficient payloads (payload that contain useful information) from domain $c_i \in [12, 9224]$ bits.
There are no hard real-time constraints imposed on the instance.
Each message has from one up to forty-two receivers.
\subsection{Quality and Performance evaluation of the solver}
\label{ssec:Quality}
The provided method for the TT message scheduling has been evaluated on the described benchmark set and compared with the Multi-mode method.
The Multi-mode method developed by us is similar to the described algorithm, but it uses the Multi-mode RCPSP instead of the general RCPSP model.
In the Multi-mode RCPSP, the alternative modes can be assigned to the activities, and the resulting schedule uses just one of the provided activities.
The TTEthernet problem can be formulated as a Multi-mode RCPSP without the necessity to decide in which integration cycles each message appears in advance.
Thus, it is not necessary to solve the integration cycle assignment problem heuristically by ILP.
The Multi-mode RCPSP can return better solution than our described algorithm denoted as ICAP in Table~\ref{Tab:QualityResults} that uses heuristic integration cycle assignment, but the computational complexity increases significantly.
\begin{table}[t]
\caption{Quality comparison of different approaches}
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{l|r|r|r|r}
\multicolumn{3}{r}{}&\multicolumn{1}{r}{ICAP}&Multi-mode\\
&N [-]&LB [ns]&$\text{C}_{\text{max}}$\;[ns]&$\text{C}_{\text{max}}$\;[ns]\\
\hline
Set\_20TT&124&9626&13430&\textbf{13255}\\
Set\_50TT&312&19468&\textbf{22908}&23678\\
Set\_100TT&624&35603&\textbf{40450}&42291\\
Set\_200TT&1243&66594&\textbf{74663}&79733\\
Set\_500TT&3107&157317&182185&\textbf{181913}\\
Set\_1000TT&6244&318017&\textbf{354577}&354950\\
Set\_2000TT&12460&616446&\textbf{663532}&-\\
Set\_Industrial&5844&59888&\textbf{63568}&-\\
\hline \hline
Average&&177429&194708&-\\
\end{tabular}}
\label{Tab:QualityResults}
\end{table}
Table~\ref{Tab:QualityResults} represents the results obtained by the methods.
The captions of the benchmark sets are situated in the first column.
The second column presents the average number of message instances/activities in the set~(N).
Recall that each message has one message instance for each link the message passes.
Therefore, the number of message instances is often much higher than the number of messages.
All the remaining columns shows the results for the particular algorithm. The values are averaged over thirty independent instances in each set.
The third column shows the lower bound~(LB) for the problem.
The lower bound is obtained from the ILP solution of the integration cycle assignment problem.
The makespan of our problem can never be shorter than the sum of the messages transmission time exchanged in any link if the message to the integration cycle assignment is balanced.
That is why the ILP solution can be used as a lower bound for the whole problem too.
However, it is often not possible to obtain a makespan equal to the lower bound because time lags are not considered in the ILP model.
Therefore, it is not-tight lower bound.
The fourth column contains the average makespan value of the instances in the given set for our provided algorithm~(ICAP) in nanoseconds.
In comparison, the fifth column contains the makespan for the Multi-mode method.
The time limit for the computation of each instance was set to 300 seconds for both methods.
The small theoretical instances with 20 TT messages were solved to an optimum by RCPSP solver by both methods in time.
The final results for these small instances are better for the Multi-mode version because it is not heuristically guided by the ILP solution and, therefore, no states of the search space are pruned.
One can observe the gap between the optimal solution of the Multi-mode method and the non-tight lower bound.
The Multi-mode RCPSP solver was not able to solve bigger instances to an optimum in time.
This leads one to the fact that the guided ICAP method obtains better solutions (except the set with 500 TT messages where relaxing the time-lags in the ICAP formulation led to the weaker integration cycle assignment) from this point further.
Furthermore, no feasible solution was found by the Multi-mode RCPSP solver for instances with two thousand messages in 300 seconds.
Note that the solutions obtained by the ICAP method are not more than 15\% distant from the lower bound for instances with more than 50 TT messages.
\subsection{Worst-case RC traffic delay evaluation}
To justify the used scheduling method, the solver based on the porosity idea from~\cite{Steiner2011} was implemented to compare the makespan optimization method with the porosity optimization method with respect to the RC traffic.
The porosity is represented by the set of gaps introduced into the schedule.
Consequently, the optimization objective is to maximize the length of these gaps.
Compared to the makespan optimization, which aims to introduce just one big gap at the end of the schedule, the porosity optimization distributes the free bandwidth through the schedule.
Note that unlike in other protocols that has strict separations of time-triggered and event-triggered segments, the introduced porosity gaps as well as one critical gap introduced by makespan optimization does not separate the TT traffic and RC traffic so strictly.
It means that even if the makespan is e.g. 6\,ms (just like in the Fig.~\ref{Fig:Link_schedules}) the RC message can be transmitted sooner if there is no TT traffic scheduled in the particular integration cycle (e.g. message $RC_2$ in Fig.~\ref{Fig:Link_schedules}).
Moreover, the RC traffic can be transmitted even in the gaps of the schedule where they can fit (e.g. message $RC_1$ in Fig.~\ref{Fig:Link_schedules}).
\begin{figure}[h]
\resizebox{\columnwidth}{!}
{
\includegraphics{Images/RC_results4.pdf}
}
\caption{Worst-Case delays for the RC traffic}
\label{Fig:RC_Results}
\end{figure}
To evaluate the difference between the makespan optimization and the porosity optimization in an objective way, the Worst-Case delay calculation for the RC messages taken from~\cite{TS2015} was employed.
In Fig.~\ref{Fig:RC_Results}, the average delays among the instances in the particular sets and its standard deviations are presented.
The evaluation was made for all the artificial sets.
However, Set\_2000TT has been omitted from the graph because it caused graph scaling which worsens its clarity.
The worst-case RC traffic delay of Set\_2000TT was $797 \pm 18$\SI{}{\micro\second} for makespan optimization and it was $799 \pm 18$\SI{}{\micro\second} for porosity optimization.
According to the results, the makespan optimization behaves comparable and even slightly better than the porosity optimization with respect to the RC traffic.
The standard deviation is higher in the case of Set\_500TT because there were a few significant outliers.
Overall, the obtained Worst-Case delays are more stable in the case of the makespan optimization.
\section{Conclusion}
\label{Sec:Conclusion}
The paper in hand focuses on the problem of the Time-Triggered communication scheduling on the TTEthernet network.
To incorporate the Event-Triggered Rate-Constrained traffic with TT traffic is very important for mixed-criticality applications \cite{Novak2016}.
The majority of communication is Event-Triggered in industrial applications nowadays.
However, the necessity to incorporate safety-related communication, e.g. in automotive or avionics industry, pushes system developers to use the TT traffic because it can be more easily verified and certified.
We have investigated the idea of separating the Time-Triggered traffic and Event-Triggered traffic, which was inspired by the scheme of the FlexRay bus communication cycle.
The aim has been to minimize the length of the Time-Triggered schedule and leave the rest of the communication bandwidth free for the Rate-Constrained traffic.
For such a problem statement, we have designed the algorithm based on the ILP formulation of the integration cycle assignment problem (load balancing among integration cycles) and the RCPSP formulation of the message scheduling problem.
The experiments performed on several artificial benchmark sets and one real-case instance show that the designed method is able to obtain near-optimal results (for bigger instances in 15\% of the non-tight lower bound) concerning the proposed criterion.
Moreover, the RC traffic worst-case delay calculations suggest that the proposed concept (minimization of the schedule makespan) can provide slightly better delays in average than the porosity imposed to the schedule by gaps.
For the future work, we would like to study the behavior of the provided method in the environment of the 802.1Qbv~\cite{QbvEthernet} network.
The Ethernet technology is used in the automotive, avionics, etc. industry nowadays, where the development cycle of the products is specific - the new product models are created taking a backward compatibility with the previous ones into account.
Thus, we are also going to study the incremental scheduling case where backward compatibility needs to be preserved.
\section*{Acknowledgment}
This work was supported by the Grant Agency of the Czech Republic under the Project FOREST GACR P103-16-23509S.
\bibliographystyle{ieeetr}
|
\section{Introduction}
Motivated by pre-trained models like BERT \cite{devlin2018bert} and GPT \cite{radford2018improving} which have significantly improved the performance of many NLP tasks, researchers \cite{lu2019vilbert,li2019unicoder,su2019vl,li2019visualbert,chen2019uniter} have noticed the importance of pre-training for vision-language tasks, e.g., Visual Question Answering(VQA) \cite{antol2015vqa} and
Visual Commonsense Reasoning (VCR) \cite{zellers2019recognition}.
Existing vision-language pre-training methods attempt to learn joint representations through visual grounding tasks on large image-text datasets, including Masked Language Modelling based on randomly-masked sub-words, Masked Region Prediction and Image-Text Matching at the image/text-level. However, based on randomly-masking and predicting the sub-words, current models did not distinguish common words and words describing the detailed semantics \cite{johnson2015image}, e.g., objects(``man", ``boat"), attributes of objects(``boat is white"), relationships between objects(``man standing on boat"). These methods neglect the importance of constructing detailed semantic alignments across vision and language, therefore the trained models can not well represent fine-grained semantics required by some real scenes. As shown in Figure \ref{tab_nodecase}, the detailed semantics are essential to distinguish the listed scenes which mainly differ in objects, attributes and relationships. Hence, better joint vision-language representations should characterize detailed semantic alignments across the modalities.
\begin{figure*}[t]
\centering
\begin{tabular}{ccc}
\multicolumn{1}{c}{(a) Objects} & \multicolumn{1}{c}{(b) Attributes} & \multicolumn{1}{c}{(c) Relationships} \\
\hline
\begin{minipage}{0.3\textwidth}
\centering
\includegraphics[width=0.72\linewidth]{kiss_dog.jpg}
\end{minipage}
&
\begin{minipage}{0.3\textwidth}
\centering
\includegraphics[width=0.64\linewidth]{purple_toy.jpg}
\end{minipage}
&
\begin{minipage}{0.3\textwidth}
\centering
\includegraphics[width=0.76\linewidth]{riding_bike.jpg}
\end{minipage}
\\
\multicolumn{1}{c}{\small{A tan \textcolor{red}{\textbf{dog}} and a little girl kiss.}}
&
\multicolumn{1}{c}{\small{A black dog playing with a \textcolor{red}{\textbf{purple}} toy.}}
&
\multicolumn{1}{c}{\small{
A man in red plaid \textcolor{red}{\textbf{rides}} his bike in a park.}}
\\
\begin{minipage}{0.3\textwidth}
\centering
\includegraphics[width=0.72\linewidth]{kiss_cat.jpg}
\end{minipage}
&
\begin{minipage}{0.3\textwidth}
\centering
\includegraphics[width=0.64\linewidth]{green_toy.jpg}
\end{minipage}
&
\begin{minipage}{0.30\textwidth}
\centering
\includegraphics[width=0.80\linewidth]{fixing_bike.jpg}
\end{minipage}
\\
\multicolumn{1}{c}{\small{The little girl is kissing the brown \textcolor{red}{\textbf{cat}}.}}
&
\multicolumn{1}{c}{\small{A black dog playing with a \textcolor{red}{\textbf{green}} toy.}}
&
\multicolumn{1}{c}{\small{
An older man \textcolor{red}{\textbf{repairing}} a bike tire in a park.}}
\\
\end{tabular}
\caption{Similar scene pairs from the Flick30K datasets \cite{young2014image}. It is the detailed semantics that determine the interpretation of the scenes, objects (dog, cat) in scene pair (a), attributes(purple, green) in scene pair (b) and relationships(rides, repairing) in scene pair (c). }
\label{tab_nodecase}
\end{figure*}
Inspired by the knowledge masking strategy of ERNIE \cite{sun2019ernie2}, which aims at learning more structured knowledge by masking phrases and named entities rather than individual sub-words, we propose ERNIE-ViL, that incorporates knowledge obtained from scene graphs \cite{johnson2015image} to construct better representations for vision-language joint modelling. Through constructing Scene Graph Prediction tasks, ERNIE-ViL puts more emphasis on detailed semantic alignments across vision and language. Concretely, we implement these pre-training tasks by masking and predicting different types of nodes in the scene graph parsed from the sentence. By concentrating on understanding detailed semantic words rather than common words, these Scene Graph Prediction tasks force the model to extract object/attribute/relationship information from the visual modality, thus establish semantic connections between vision and language.
Pre-training with the Scene Graph Prediction tasks, ERNIE-ViL learns the vision-language detailed semantic alignments.
We pre-train ERNIE-ViL on two large commonly-used image-text out-of-domain datasets, namely Conceptual Captions \cite{sharma2018conceptual} and SBU Captions \cite{ordonez2011im2text}. To evaluate the performance of ERNIE-ViL, we conduct experiments on various vision-language tasks, (1) Visual Question Answering (VQA 2.0) \cite{antol2015vqa}, (2) Visual Commonsense Reasoning (VCR) \cite{zellers2019recognition}, (3) Region-to-Phrase Grounding (RefCOCO+) \cite{kazemzadeh2014referitgame}, (4, 5) Image-text Retrieval / Text-image Retrieval (Flickr30K) \cite{young2014image}. On all these tasks, ERNIE-ViL obtains significant improvements compared to those models pretrained on the same datasets. Especially, on the Region-to-Phrase grounding task that
relies more heavily on detailed semantic alignments, we achieve an improvement of 2.4\% on both testsets. To compare with the models pretrained on both out-of-domain and in-domain datasets, we continually pre-train ERNIE-ViL on MS-COCO \cite{lin2014microsoft} and Visual-Genome \cite{krishna2017visual} (in-domain datasets for downstream tasks). ERNIE-ViL achieves the state-of-the-art performances on all downstream tasks. Also ERNIE-ViL obtains the best single model performance and ranks the first place on the leaderboard with an absolute improvement of 3.7\% on the Q$\to$AR task compared to the state-of-the-art performance. And our code and pre-trained models are scheduled to be public.
Overall, the contributions of our method are three-folds:
\begin{itemize}
\item To the best of our knowledge, ERNIE-ViL is the first work that has introduced structured knowledge to enhance vision-language pre-training.
\item ERNIE-ViL constructs Scene Graph Prediction tasks during the pre-training of vision-language joint representations, putting more emphasis on the cross-modal detailed semantics alignments.
\item ERNIE-ViL achieves state-of-the-art performances on 5 downstream cross-modal tasks and ranks the first place on the VCR leaderboard.
\end{itemize}
\begin{figure*}[t]
\centering
\includegraphics[width=1.95\columnwidth]{scene_graph_prediction.pdf}
\caption{Illustration of Scene Graph Prediction tasks for ERNIE-ViL. Given detected regions of the image and token sequence of the text, ERNIE-ViL uses a two-stream cross-modal Transformers network to model the joint vision-language representations. Based on the scene graph parsed from the text using Scene Graph Parser, we construct Object Prediction, Attribute Prediction and Relationship Prediction tasks to learn cross-modal detailed semantics alignments.}
\label{fig_framework}
\end{figure*}
\section{Related Works}
\subsection{Cross-modal Pre-training}
Inspired by text pre-training models \cite{devlin2018bert}, many cross-modal pre-training models for vision-language have been proposed.
These researchers put their efforts mainly on three aspects, which are model architecture, pre-training tasks and pre-training data.
\begin{itemize}
\item \textbf{Model Architecture} Current works are based on different variables of Transformers \cite{vaswani2017attention}. Most of them \cite{li2019unicoder,su2019vl,zhou2019unified,li2019visualbert,huang2020pixel} used a uniform cross-modal Transformer modelling both image and text representations, while the others like ViLBERT \cite{lu2019vilbert} and LXMERT \cite{tan2019lxmert} were based on two-stream cross-modal Transformers, which brings more specific representations for images and texts.
\item \textbf{Pre-training Tasks} Inspired by the pre-training tasks in text models, Masked Language Model and similar Masked Region Prediction tasks \cite{lu2019vilbert} are utilized in cross-modal pre-training. And similar to Next-Sentence Prediction, Image-Text Matching \cite{lu2019vilbert,su2019vl,chen2019uniter} task is also widely used. However, based on randomly masking and predicting sub-words, these methods did not distinguish the common words and words describing the detailed semantics. Hence, the cross-modal fine-grained semantic alignments cannot be well characterized in those learned joint representations.
\item \textbf{Pre-training Data} Unlike text pre-training models that can leverage tremendous natural language data, vision-language tasks require high-quality aligned image-text data that are hard to obtain. Conceptual Captions\cite{sharma2018conceptual} and SBU Captions\cite{ordonez2011im2text} are two widely-used datasets for image-text pre-training, with 3.0M and 1.0M image-description pairs respectively. These two datasets are out-of-domain for vision-language downstream tasks, while some existing works \cite{chen2019uniter,huang2020pixel} incorpate in-domain datasets, such as MS-COCO and Visual-Genome, that are highly correlated with downstream tasks.
\end{itemize}
\subsection{Scene Graph}
Scene graphs contain structured knowledge of visual scenes, including the present objects, attributes of objects, and relationships between objects. As a beneficial prior knowledge describing the detailed semantics of images and captions, scene graphs have led to many state-of-the-art models in image captioning \cite{yang2019auto}, image retrieval \cite{wu2019unified}, VQA \cite{zhang2019empirical} and image generation \cite{johnson2018image}.
\section{Approach}
In this section, we first introduce the architecture of ERNIE-ViL. Then we illustrate our newly-proposed Scene Graph Prediction tasks. Finally, pre-training with Scene Graph Prediction tasks in ERINE-ViL is introduced.
\subsection{Model Architecture}
The vision-language model aims at learning the joint representations that integrates information of both modalities and the alignments across the modalities. The inputs of ERNIE-ViL are a sentence and an image. Given a sequence of words and an image, we introduce the methods to embed the inputs to the feature space and the vision-language encoder.
\paragraph{Sentence Embedding} We adopt the similar word pre-prossessing method as BERT. The input sentence is tokenized into sub-word tokens using WordPiece approach. Special tokens such as $\textit{[CLS]}$ and $\textit{[SEP]}$ are also added to the tokenized text sequence to form the text sequence as $\{\textit{[CLS]}, w1,\dots w_{T} ,\textit{[SEP]}\}$. The final embedding for each sub-word token is generated by combining its original word embedding, segment embedding and sequence position embedding.
\paragraph{Image Embedding} For the image, we first use a pre-trained object detector to detect the salient image regions from the image. The pooling features before multi-class classification layer are utilized as the region features. We also encode the location features for each region via a 5-dimensional vector $(\frac{x_1}{W},\frac{y_1}{H},\frac{x_2}{W},\frac{y_2}{H},\frac{(y_2-y_1)(x_2-x_1)}{WH})$ for the region position and the fraction of image area covered, where $(x_1,y_1)$ and $(x_2,y_2)$ denote
the coordinates of top-left and bottom-right corner while $W$ and $H$ are the width and height of the input image. The location vectors are projected to form the location features, which are then summed with the region visual features. We also add a special feature $\textit{[IMG]}$ that denotes the representation of the entire image (i.e. mean-pooled visual features with a spatial encoding corresponding to the entire image) to form the final region sequence $\{\textit{[IMG]}, v_{1},\dots, v_{I}\}$.
\paragraph{Vision-Language Encoder} Given the embedding of image regions and the words for the sentence $\{\textit{[IMG]}, v_{1},\dots, v_{I} , \textit{[CLS]}, w1,\dots w_{T} ; \textit{[SEP]}\}$, we use two-stream cross-modal Transformers to joint model the intra-modal and inter-modal representations. Similar to ViLBERT \cite{lu2019vilbert}, ERNIE-ViL consists of two parallel Transformer encoders for image and text segments, which are cross-attended with cross-modal Transformer blocks. The model outputs embeddings for each input of both the image and text. We take $h_{\textit{[IMG]}}$ and $h_{\textit{[CLS]}}$ as the holistic image and text representations.
\subsection{Scene Graph Prediction}\label{sec_task}
Detailed semantics, includes objects, attributes of objects, and relationships between objects, are essential to the understanding of visual scenes \cite{johnson2015image}. As the scene shown in Figure \ref{fig_framework}, detailed semantics describes the visual scene from different aspects. The objects, such as ``cat", ``car", ``woman" are the fundamental elements in the scene. And associated attributes, such as ``little", ``brown", ``blue" characterize shape and color of objects. Relationships such as ``on top of", ``putting" represent the spatial connections and actions between objects. Therefore detailed semantics are crucial in accurately understanding visual scenes.
Since the goal of vision-language joint representations is to engrave the semantic connections across modalities, detailed semantic alignments are significantly important in cross-modal learning.
Scene graphs encode various fine-grained semantic information. Utilizing structured knowledge obtained from scene graphs, ERNIE-ViL learns the cross-modal detailed semantic alignments. As shown in Figure \ref{fig_framework}, according to the scene graph parsed from the text, we construct the corresponding Scene Graph Prediction tasks, including Object Prediction task, Attribute Prediction task, and Relationship Prediction task. These tasks force ERNIE-ViL to model the correlations of detailed semantics across modalities. For example, as the relationship words ``on top of" is masked, based on the language context, the model may predict that the missing word is ``under" or ``into". These words are grammatically fluent in the sentence, but are inconsistent with the scene ``the cat is on top of the car". Through training the Relationship Prediction task, the model obtains the spatial relation of the corresponding objects(``car", ``cat") from the image, thus can accurately predict that the missing word is ``on top of". Through constructing Scene Graph Prediction tasks, ERNIE-ViL learns cross-modal detailed semantic alignments.
\paragraph{Scene graph parsing}
Given the text sentence $\mathbf{w}$, we parse it into a scene graph\cite{johnson2015image}, which denotes as $G(\mathbf{w})=<O(\mathbf{w}),E(\mathbf{w}),K(\mathbf{w})>$, where $O(\mathbf{w})$ is the set of objects mentioned in $\mathbf{w}$, $E(\mathbf{w})\subseteq O(\mathbf{w})\times R(\mathbf{w})\times O(\mathbf{w})$ is the set of hyper-edges representing relationship triplets, and $R(\mathbf{w})$ is the set of relationship nodes between object nodes. $K(\mathbf{w})\subseteq O(\mathbf{w})\times A(\mathbf{w})$ is the set of attribute pairs, where $A(\mathbf{w})$ is the set of attribute nodes associated with object nodes. Scene graphs describe the objects in more details with various associated attributes and relationships between objects. Thus integrating the knowledge of scene graphs can benefit learning more fine-grained joint representations for the vision-language. In this paper, the Scene Graph Parser provided by Anderson \cite{anderson2016spice} is adopted to parse texts to scene graphs. For a more intuitive understanding, we illustrate a specific case for the parsed scene graph from the text in
Table \ref{tab_sg}.
\begin{table}
\begin{center}
\begin{tabular}{c|c}
\hline
\multicolumn{1}{c|}{\multirow{3}{*}{sentence: $\mathbf{w}$}} & A woman in blue dress is putting \\
\multicolumn{1}{c|}{} & her little white cat on top of a \\
\multicolumn{1}{c|}{} & brown car in front of her house. \\ \hline
\multicolumn{1}{c|}{\multirow{1}{*}{objects:$O(\mathbf{w})$}} & dress, woman, cat, car, house \\ \hline
\multicolumn{1}{c|}{\multirow{1}{*}{relationships:$R(\mathbf{w})$}} & in, putting, on-top-of, in-front-of \\ \hline
\multicolumn{1}{c|}{\multirow{1}{*}{ attributes: $A(\mathbf{w})$}} & blue, white, little, brown \\ \hline
\end{tabular}
\end{center}
\caption{The scene graph parsed from the caption of the visual scene. For simplicity, we only list all the nodes leaving out the connections between them. }
\label{tab_sg}
\end{table}
\subsubsection{Object Prediction}
Objects are the dominant elements of visual scenes, thus playing an important role in constructing the representations of semantic information. Predicting the objects forces the model to build the vision-language connections at object level.
Firstly, for all the object nodes in the scene graph, we randomly select 30\% of them to mask. And for each selected object node $O(\mathbf{w})$, we replace it with the special token $\textit{[MASK]}$ in probability of 80\%, another random token in probability of 10\%, and keep it in probability of 10\%. Note that the objects actually correspond to the sub-sequences of text in the sentence, therefore the object masking are implemented by masking the corresponding sub-sequences in the text.
For Object Prediction, ERNIE-ViL recover these masked object tokens, denoted as $\mathbf{w}_{o_{i}}$, based on their surrounding words $\mathbf{w}$ and all image regions $\mathbf{v}$, by minimizing the negative log-likelihood:
\begin{equation}
\mathcal{L}_{obj}(\theta)=-E_{(\mathbf{w},\mathbf{v})\sim D}\log(P(\mathbf{w}_{o_{i}}| \mathbf{w}_{\setminus \mathbf{w}_{o_{i}}},\mathbf{v}))
\end{equation}
\subsubsection{Attribute Prediction}
Attributes characterize the specific information of the visual objects, such as color or shape of the objects, therefore representing the detailed information in the visual scenes in more fine-grained level.
Similarly, we randomly select 30\% of the attribute pairs in the scene graph, and the mask strategy here is the same as that in Object Prediction. Since the attribute nodes in the scene graph are attached to objects, we keep the associated object while masking out the attribute node $A(\mathbf{w})$ in each selected $K(\mathbf{w})\subseteq O(\mathbf{w})\times A(\mathbf{w})$.
Given object words $w_{o_{i}}$ in attribute pair $ \langle w_{o_i},w_{a_{i}} \rangle$, Attribute Prediction is to recover the masked tokens $w_{a_{i}}$ of attribute pairs. Based on the object tokens $w_{o_{i}}$ , other surrounding words $\mathbf{w}$ and all image regions v, Attribute Prediction minimizes the negative log-likelihood:
\begin{equation}
\mathcal{L}_{attr}(\theta)=-E_{(\mathbf{w},\mathbf{v})\sim D}\log(P(\mathbf{w}_{a_{i}}|\mathbf{w}_{o_{i}}, \mathbf{w}_{\setminus\mathbf{w}_{a_{i}}},\mathbf{v}))
\end{equation}
\subsubsection{Relationship Prediction}
Relationships describe the actions (semantic) or relative position (geometry) between the objects of the visual scenes, which contributes to distinguish scenes with same objects but different relationships.
Thus, ERNIE-ViL constructs the Relationship Prediction task to learn cross-modal relationships connections. When performing the mask strategy of selected relationship triplets $E(\mathbf{w})\subseteq O(\mathbf{w})\times R(\mathbf{w}) \times O(\mathbf{w})$, we keep the objects and mask out the relationship node $R(\mathbf{w})$. Specifically,
given object tokens $\mathbf{w}_{o_{i1}},\mathbf{w}_{o_{i2}}$ in relationship triplet $ \langle \mathbf{w}_{o_{i1}},\mathbf{w}_{r_i},\mathbf{w}_{o_{i2}} \rangle$, this task recovers the masked relationship tokens, predicting the probability for each masked relation tokens $\mathbf{w}_{r_{i}}$. Thus the context for the prediction is the given object tokens $\mathbf{w}_{o_{i1}},\mathbf{w}_{o_{i2}}$, other surrounding words from the text and all image regions $\mathbf{v}$. The loss for this task is:
\begin{equation}
\mathcal{L}_{rel}(\theta)=-E_{(\mathbf{w},\mathbf{v})\sim D}\log(P(\mathbf{w}_{r_{i}}|\mathbf{w}_{o_{i1}},\mathbf{w}_{o_{i2}}, \mathbf{w}_{\setminus \mathbf{w}_{r_i}},\mathbf{v}))
\end{equation}
\subsection{Pre-training with Scene Graph Prediction}
Simliar to ViLBERT\cite{lu2019vilbert}, ERNIE-ViL also adopts Masked Language Modelling(MLM) to capture the syntactic and lexical information in the text. Moreover, Masked Region Prediction and Image-text Matching are utilized for visual modality and cross-modality respectively. The losses for all these pre-training tasks are summed.
\section{Experiments}
\subsection{Training ERNIE-ViL}
\paragraph{Pre-training Data} We use the Conceptual Captions (CC) dataset \cite{sharma2018conceptual} and SBU Captions (SBU) dataset \cite{ordonez2011im2text} as pre-training data. CC is a collection of 3.3 million image-caption pairs automatically scraped from alt-text enabled web images and SBU is a similar vision-language dataset which has 1.0 million image-caption pairs. Since some links have become broken, only about 3.0 million pairs for CC dataset and 0.8 million pairs for SBU dataset are available and utilized in our experiments. Note that CC and SBU are image-caption pairs automatically collected from the web and have no intersections with the down-stream task datasets, thus act as out-of-domain datasets for training vision-language models.
\paragraph{Implementation Details}
For each image-text pair in the training, the pre-processing is performed as follows. For the image, we adopt Faster R-CNN \cite{anderson2018bottom} to select salient image regions and extract region features. Specifically, regions with class detection probability exceeds a confidence threshold of 0.2 are selected and 10 to 36 boxes are kept. And for each kept region, the mean-pooled convolutional representation is used as the region feature. For the text, we parse the scene graph from the sentence using the Scene Graph Parser and adopt WordPieces to tokenize the sentence similar to BERT.
For the masking strategies, we randomly mask 15\% of tokens, 30\% of scene graph nodes, and 15\% of image regions. For the Image-text Matching task, we randomly select a image for each text to form the negative image-text pair. Note that only items in the positive pairs will be considered for token and region prediction tasks.
We train ERNIE-ViL on two scale settings: ERNIE-ViL-base and ERNIE-ViL-large, which mainly differ in model depth of the text stream. The detailed settings of text and visual streams are shown in Table \ref{tab_setting}. And similar to VilBERT\cite{lu2019vilbert}, cross-transformers are used to co-at tent the two streams. We initialize the text stream parameters from ERNIE 2.0 \cite{sun2019ernie2}, and implement ERNIE-ViL via PaddlePaddle. After then, ERINE-ViL is pre-trained on a total batch size of 512 for 700k steps on 8 V100 GPUs, using adam optimizer with initial learning rates of 1e-4 and Noam \cite{vaswani2017attention} as learning rate decay schedule.
\begin{table}[t]
\centering
\setlength{\tabcolsep}{1.8mm}{
\small
\begin{tabular}{c|cccc|cccc}
\hline
& \multicolumn{4}{c|}{Base} & \multicolumn{4}{c}{Large} \\
&$L$ & $H$ & $A$ & $F$ & $L$ & $H$ & $A$ & $F$ \\ \hline
Text &12 &768 &12 &3072 &24 &1024 &16 & 4096 \\
Visual & 6 &1024& 8 & 1024 &6 &1024 &16 &4096 \\
\hline
\end{tabular}
}
\caption{Settings for ERNIE-ViL model. $L$: number of layers, $H$ : hidden size, $A$ : number of self-attention heads, $F$ : feed-forward/filter size.}
\label{tab_setting}
\end{table}
\subsection{Downstream Tasks}
\subsubsection{Visual Commonsense Reasoning (VCR)}
The Visual Commonsense Reasoning (VCR) \cite{zellers2019recognition} task contains two sub-tasks: visual question answering (Q$\to$A) and answer justification (QA$\to$R), which are both multiple choice problems. The holistic setting (Q$\to$AR) requires both the chosen answer and chosen rationale to be correct. In visual question answering (Q$\to$A) task, we concatenate the question and each candidate answer for the language modality. We take dot product of final hidden state $h_{[\textit{CLS}]}$ and $h_{[\textit{IMG}]}$ to predict matching score with an additional FC layer. For the answer justification (QA$\to$R) task, we concatenate the question, the answer and each candidate rationale as the input of the text stream. Similar with UNITER \cite{chen2019uniter}, a second-stage pre-training is adopted on VCR dataset. And then we fine-tune the model over 6 epochs with a batch size of 64 and adopt Adam optimizer with initial learning rate of 1e-4.
\subsubsection{Visual Question Answering (VQA)}
The VQA task requires answering natural language questions according to images. VQA 2.0 dataset \cite{antol2015vqa} contains 204k images and 1.1M questions about these images. Also additional question-answer pairs from Visual Genome are used for data augmentation as in UNITER \cite{chen2019uniter}.
We treat VQA as a multi-label classification task – assigning a soft target score to each answer based on its relevancy to the 10 human answer responses. We take dot product of final hidden state $h_{[\textit{CLS}]}$ and $h_{[\textit{IMG}]}$ to map this representation into 3,129 possible answers with an additional two-layer MLP.
Fine-tuning of VQA model is performed over 12 epochs on batch size of 256 and using Adam optimizer with initial learning rate of 1e-4.
\begin{table*}[t]
\centering
\renewcommand{\arraystretch}{1.2}
\tabcolsep 6.4pt
\small
\begin{tabular}{llllllllllll}
\toprule[0.7pt]
\multicolumn{1}{c}{\multirow{2}{*}{Domains}} & \multicolumn{1}{c}{\multirow{2}{*}{Models}} &
&
\multicolumn{3}{c}{VCR} &
&
\multicolumn{3}{c}{RefCOCO+} \\
\cline{8-10}
\cline{4-6}
\multicolumn{1}{c}{} & \multicolumn{1}{c}{} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{\texttt{Q$\to$A}} & \multicolumn{1}{c}{\texttt{QA$\to$R}} & \multicolumn{1}{c}{\texttt{Q$\to$AR}} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{\texttt{val}} &
\multicolumn{1}{c}{\texttt{testA}} & \multicolumn{1}{c}{\texttt{testB}} \\
\hline
\multirow{7}{*}{Out-of-domain} & \multicolumn{1}{c}{UNITER-base} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{72.78} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} \\
& \multicolumn{1}{c}{Unicoder-VL-base} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{72.6(73.4)} & \multicolumn{1}{c}{74.5(74.4)} & \multicolumn{1}{c}{54.4(54.9)} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} \\
& \multicolumn{1}{c}{ViLBERT-base} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{72.42(73.3)} & \multicolumn{1}{c}{74.47(74.6)} & \multicolumn{1}{c}{54.04(54.8)} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{72.34} & \multicolumn{1}{c}{78.52} & \multicolumn{1}{c}{62.61} \\
& \multicolumn{1}{c}{VLBERT-base} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{73.8(-)} & \multicolumn{1}{c}{74.4(-)} & \multicolumn{1}{c}{55.2(-)} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{71.60} & \multicolumn{1}{c}{77.72} & \multicolumn{1}{c}{60.99} \\
& \multicolumn{1}{c}{ERNIE-ViL-base} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{76.37(77.0)} & \multicolumn{1}{c}{79.65(80.3)} & \multicolumn{1}{c}{61.24(62.1)} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{74.02} & \multicolumn{1}{c}{80.33} & \multicolumn{1}{c}{\textbf{64.74}} \\
\cline{2-10}
& \multicolumn{1}{c}{VLBERT-Large} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{75.5(75.8)} & \multicolumn{1}{c}{77.9(78.4)} & \multicolumn{1}{c}{58.9(59.7)} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{72.59} & \multicolumn{1}{c}{78.57} & \multicolumn{1}{c}{62.30} \\
& \multicolumn{1}{c}{ERNIE-ViL-Large} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{\textbf{78.52(79.2)}} & \multicolumn{1}{c}{\textbf{83.37(83.5)}} & \multicolumn{1}{c}{\textbf{65.81(66.3)}} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{\textbf{74.24}} & \multicolumn{1}{c}{\textbf{80.97}} & \multicolumn{1}{c}{64.70} \\
\hline
\multirow{3}{*}{\shortstack{Out-of-domain\\+ in-domain}} & \multicolumn{1}{c}{UNITER-large} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{77.22(77.3)} & \multicolumn{1}{c}{80.49(80.8)} & \multicolumn{1}{c}{62.59(62.8)} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{75.90} & \multicolumn{1}{c}{81.45} & \multicolumn{1}{c}{66.70}\\
&\multicolumn{1}{c}{VILLA-large} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{78.45(78.9)} & \multicolumn{1}{c}{82.57(82.8)} & \multicolumn{1}{c}{65.18(65.7)} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{\textbf{76.17}} & \multicolumn{1}{c}{81.54} & \multicolumn{1}{c}{66.84}\\
&\multicolumn{1}{c}{ERNIE-ViL-large} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{\textbf{78.98(-)}} & \multicolumn{1}{c}{\textbf{83.70(-)}} & \multicolumn{1}{c}{\textbf{66.44(-)}} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{75.89} & \multicolumn{1}{c}{\textbf{82.37}} & \multicolumn{1}{c}{\textbf{66.91}}\\
\bottomrule[0.7pt]
\end{tabular}
\begin{tabular}{llllllllllllll}
\toprule[0.7pt]
\multicolumn{1}{c}{\multirow{2}{*}{Domains}} & \multicolumn{1}{c}{\multirow{2}{*}{Models}} & &
\multicolumn{2}{c}{VQA} &
&
\multicolumn{3}{c}{IR-Flickr30K}
& &
\multicolumn{3}{c}{TR-Flickr30K}
\\
\cline{7-9}
\cline{4-5}
\cline{11-13}
\multicolumn{1}{c}{} & \multicolumn{1}{c}{} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{\texttt{test-dev}} &
\multicolumn{1}{c}{\texttt{test-std}} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{\texttt{R@1}} &
\multicolumn{1}{c}{\texttt{R@5}} & \multicolumn{1}{c}{\texttt{R@10}} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{\texttt{R@1}} &
\multicolumn{1}{c}{\texttt{R@5}} & \multicolumn{1}{c}{\texttt{R@10}} \\
\hline
\multirow{7}{*}{Out-of-domain} & \multicolumn{1}{c}{UNITER-base} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{71.56} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} &
\multicolumn{1}{c}{-} \\
& \multicolumn{1}{c}{Unicoder-VL-base} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{71.50} & \multicolumn{1}{c}{90.90} & \multicolumn{1}{c}{94.90} & \multicolumn{1}{c}{} &
\multicolumn{1}{c}{86.20} & \multicolumn{1}{c}{96.30} & \multicolumn{1}{c}{99.00} \\
& \multicolumn{1}{c}{VLBERT-base} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{71.16} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{} &
\multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} \\
& \multicolumn{1}{c}{ViLBERT-base} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{70.55} & \multicolumn{1}{c}{70.92} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{58.20} & \multicolumn{1}{c}{84.90} & \multicolumn{1}{c}{91.52} & \multicolumn{1}{c}{} &
\multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} \\
& \multicolumn{1}{c}{ERNIE-ViL-base} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{73.18} & \multicolumn{1}{c}{73.36} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{74.44} & \multicolumn{1}{c}{92.72} & \multicolumn{1}{c}{95.94} & \multicolumn{1}{c}{} &
\multicolumn{1}{c}{86.70} & \multicolumn{1}{c}{97.80} & \multicolumn{1}{c}{99.00} \\
\cline{2-13}
& \multicolumn{1}{c}{VLBERT-large} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{71.79} & \multicolumn{1}{c}{72.22} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{} &
\multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} \\
& \multicolumn{1}{c}{ERNIE-ViL-large} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{\textbf{73.78}} & \multicolumn{1}{c}{\textbf{73.96}} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{\textbf{75.10}} & \multicolumn{1}{c}{\textbf{93.42}} & \multicolumn{1}{c}{\textbf{96.26}} & \multicolumn{1}{c}{} &
\multicolumn{1}{c}{\textbf{88.70}} & \multicolumn{1}{c}{\textbf{97.30}} & \multicolumn{1}{c}{\textbf{99.10}} \\
\hline
\multirow{4}{*}{\shortstack{Out-of-domain\\+ in-domain}} & \multicolumn{1}{c}{UNITER-large} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{73.82} & \multicolumn{1}{c}{74.02} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{75.56} & \multicolumn{1}{c}{94.08} & \multicolumn{1}{c}{96.76} & \multicolumn{1}{c}{} &
\multicolumn{1}{c}{87.30} & \multicolumn{1}{c}{98.00} & \multicolumn{1}{c}{\textbf{99.20}} \\
& \multicolumn{1}{c}{OSCAR-large} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{73.61} & \multicolumn{1}{c}{73.82} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{} &
\multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} & \multicolumn{1}{c}{-} \\
& \multicolumn{1}{c}{VILLA-large} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{74.69} & \multicolumn{1}{c}{74.87} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{76.26} & \multicolumn{1}{c}{\textbf{94.24}} & \multicolumn{1}{c}{\textbf{96.84}} & \multicolumn{1}{c}{} &
\multicolumn{1}{c}{87.90} & \multicolumn{1}{c}{97.50} & \multicolumn{1}{c}{98.80} \\
& \multicolumn{1}{c}{ERNIE-ViL-large} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{\textbf{74.95}} & \multicolumn{1}{c}{\textbf{75.10}} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{\textbf{76.66}} & \multicolumn{1}{c}{94.16} & \multicolumn{1}{c}{96.76} & \multicolumn{1}{c}{} &
\multicolumn{1}{c}{\textbf{89.20}} & \multicolumn{1}{c}{\textbf{98.50}} & \multicolumn{1}{c}{\textbf{99.20}} \\
\bottomrule[0.7pt]
\hline
\end{tabular}
\caption{Results of downstream vision-language tasks for ERNIE-ViL model, compared with previous state-of-the-art pre-trained models.
IR: Image Retrieval. TR: Text Retrieval. For VCR task which has private test set, we only report the test results (in parentheses) for ERNIE-ViL models pre-trained on out-of-domain datasets.}
\label{tab_result}
\end{table*}
\subsubsection{Grounding Referring Expressions}
The referring expression task is to localize an image region given a natural language reference. We evaluate the task on RefCOCO+ dataset \cite{kazemzadeh2014referitgame}. Bounding box proposals provided by Mattnet \cite{yu2018mattnet} are utilized. The representation for each region is denoted by its final hidden state $h_{v_i}$ with an additional FC layer. Each region $i$ is labelled as positive only when the IoU between it and the ground truth box is over 0.5. We fine-tune the model over 20 epochs with a batch size of 256 and adopt Adam optimizer with initial learning rate of 1e-4.
\subsubsection{Image Retrieval \& Text Retrieval}
Caption-based image retrieval is a task of identifying an image from a pool based on a caption describing its content. Flickr30K \cite{young2014image} contains 31,000 images and 5 captions for each image. Adopting the same split in ViLBERT \cite{lu2019vilbert}, we use each of 1,000 images for validation and for testing and the rest for training. We take dot product of final hidden state of $h_{[\textit{CLS}]}$ and $h_{[\textit{IMG}]}$ to predict matching score $s(\mathbf{w},\mathbf{v})$ for each image-text pair with an additional FC layer. We utilize circle loss \cite{sun2020circle} with 20 random negative samples for each image-text pair.
We trained 40 epochs using Adam optimizer with a initial learning rate 1e-5.
\subsection{Results}
\begin{table*}[t]
\centering
\renewcommand{\arraystretch}{1.2}
\tabcolsep 6.4pt
\small
\begin{tabular}{llllllllllllllll}
\toprule[0.7pt]
\multirow{2}{*}{\shortstack{initialized text \\ stream parameters }}&
\multicolumn{1}{c}{\multirow{2}{*}{pre-training tasks}}&
&
\multicolumn{1}{c}{VCR} &
&
\multicolumn{1}{c}{VQA} &
&
\multicolumn{1}{c}{RefCOCO+} &
&
\multicolumn{1}{c}{IR} &
&
\multicolumn{1}{c}{TR} &
\\
\cline{4-4}
\cline{6-6}
\cline{8-8}
\cline{10-10}
\cline{12-12}
& \multicolumn{1}{c}{} &
&\multicolumn{1}{c}{\texttt{Q$\to$AR(dev)}} &
&
\multicolumn{1}{c}{\texttt{dev}} &
&
\multicolumn{1}{c}{\texttt{val}} &
&
\multicolumn{1}{c}{\texttt{R@1(dev)}} &
&
\multicolumn{1}{c}{\texttt{R@1(dev)}} &
\\
\hline
\multicolumn{1}{c}{BERT} &
\multicolumn{1}{c}{w/o \textbf{SGP}} & \multicolumn{1}{c}{} &
\multicolumn{1}{c}{59.06} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{72.38}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{72.81} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{70.74} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{85.00}\\
\multicolumn{1}{c}{BERT} &
\multicolumn{1}{c}{w/ \textbf{SGP}} & \multicolumn{1}{c}{} &
\multicolumn{1}{c}{59.92} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{73.04}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{73.50} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{72.96} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{87.40}\\
\multicolumn{1}{c}{ERNIE-2.0} &
\multicolumn{1}{c}{w/ \textbf{SGP}} & \multicolumn{1}{c}{} &
\multicolumn{1}{c}{61.24} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{73.18}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{74.02} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{73.58} &
\multicolumn{1}{c}{} &
\multicolumn{1}{c}{87.80}\\
\bottomrule[0.7pt]
\end{tabular}
\caption{Results of downstream vision-language tasks for ERNIE-ViL pre-trainging with/without Scene Graph Prediction (\textbf{SGP}) tasks, and using different text stream parameters initialization.
IR \& TR: image retrieval \& text retrieval on Flickr30K. }
\label{tab_abl}
\end{table*}
\iffalse
\begin{table*}[tb]
\centering
\begin{tabular}{c|c|c|c|c|c|c}
\hline
\multirow{2}{*}{initialized text-stream parameters } & \multirow{2}{*}{pre-training tasks} &\multicolumn{1}{c|}{VCR} & \multicolumn{1}{c|}{VQA} & \multicolumn{1}{c|}{RefCOCO+} & \multicolumn{1}{c|}{IR-Flickr30K} &
\multicolumn{1}{c}{TR-Flickr30K} \\
& & Q$\to$AR(dev) & test-dev & dev & R@1(dev) & R@1(dev) \\ \hline
BERT & w/o \textbf{SGP} & 59.06 & 72.38 & 72.81 & 70.74 & 85.00 \\
BERT & w/ \textbf{SGP} & 59.92 & 73.04 & 73.50 & 72.96 & 87.40 \\
ERNIE-2.0 & w/ \textbf{SGP} & 61.24 & 73.18 & 74.02 & 73.58 & 87.80 \\ \hline
\end{tabular}
\caption{Effectiveness of proposed Scene Graph Prediction (\textbf{SGP}) tasks. We illustrate the comparisons between different initialized text parameters and different pre-training tasks.
IR: image retrieval, TR: text retrieval.}
\label{tab_abl}
\end{table*}
\fi
\begin{table*}[tb]
\centering
\begin{tabular}{ccccc}
\toprule[0.7pt]
&Image & Text & with \textbf{SGP} task & without \textbf{SGP} task \\
\hline
1&\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.70\linewidth]{disc.jpg}
\end{minipage}
&
\begin{minipage}{0.20\textwidth}
\small{ a black dog about to catch a flying \textcolor{red}{\textbf{disc}} .}
\end{minipage}
&
\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.7\linewidth]{disc_sg.png}
\end{minipage}
&
\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.7\linewidth]{disc_base.png}
\end{minipage}
\\
2&\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.70\linewidth]{red.jpg}
\end{minipage}
&
\begin{minipage}{0.20\textwidth}
\small{two men wearing \textcolor{red}{\textbf{red}} jackets are looking out over some water and one man has yellow earphones on his ears .}
\end{minipage}
&
\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.7\linewidth]{red_sg.png}
\end{minipage}
&
\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.7\linewidth]{red_base.png}
\end{minipage}
\\
3&\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.70\linewidth]{kicking.jpg}
\end{minipage}
&
\begin{minipage}{0.20\textwidth}
\small{ a little boy in a green shirt \textcolor{red}{\textbf{kicks}} a ball}
\end{minipage}
&
\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.7\linewidth]{kick_sg.png}
\end{minipage}
&
\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.7\linewidth]{kick_base.png}
\end{minipage}
\\
\bottomrule[0.8pt]
\end{tabular}
\caption{Examples of cloze test predictions for ERNIE-ViL pre-training with and without \textbf{SGP} tasks. Masked token are colored in bold and red. The probabilities of the top 5 predictions, denoted as the light purple bars, are listed in the right columns.}
\label{tab_cloze}
\end{table*}
We compare ERNIE-ViL against other cross-modal pre-training models and the results are illustrated in Table \ref{tab_result}.
\begin{table}[t]
\centering
\renewcommand{\arraystretch}{1.2}
\tabcolsep 6.4pt
\small
\begin{tabular}{llllllllllll}
\toprule[0.7pt]
\multirow{2}{*}{Nodes} & &
\multicolumn{2}{c}{without \textbf{SGP} tasks}&
&
\multicolumn{2}{c}{with \textbf{SGP} tasks} &
\\
\cline{3-4}
\cline{6-7}
& \multicolumn{1}{c}{} & \multicolumn{1}{c}{\texttt{ACC@1}} & \multicolumn{1}{c}{\texttt{ACC@5}} &
&
\multicolumn{1}{c}{\texttt{ACC@1}} &
\multicolumn{1}{c}{\texttt{ACC@5}} &
\\
\hline
objects & \multicolumn{1}{c}{} & \multicolumn{1}{c}{57.14} & \multicolumn{1}{c}{79.22} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{\textbf{58.34}} & \multicolumn{1}{c}{\textbf{80.80}} \\
attributes & \multicolumn{1}{c}{} & \multicolumn{1}{c}{44.32} & \multicolumn{1}{c}{67.58} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{\textbf{46.16}} & \multicolumn{1}{c}{\textbf{70.30}} \\
relationships & \multicolumn{1}{c}{} & \multicolumn{1}{c}{47.57} & \multicolumn{1}{c}{68.10} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{\textbf{50.65}} & \multicolumn{1}{c}{\textbf{71.54}} \\
overall & \multicolumn{1}{c}{} & \multicolumn{1}{c}{49.75} & \multicolumn{1}{c}{71.75} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{\textbf{51.75}} & \multicolumn{1}{c}{\textbf{74.31}} \\
\bottomrule[0.7pt]
\end{tabular}
\caption{Cloze test results for ERNIE-ViL. An improvement of 2.0\% on overall ACC@1 between models with/without \textbf{SGP} tasks. }
\label{tab_clone_acc}
\end{table}
\iffalse
\begin{table}[tb]
\centering
\begin{tabular}{c|cc|cc}
\hline
& \multicolumn{2}{c|}{without \textbf{SGP tasks }} & \multicolumn{2}{c}{with \textbf{SGP} tasks} \\
& ACC@1 & ACC@5 & ACC@1 & ACC@5 \\ \hline
objects & 57.14 & 79.22 & \textbf{58.34} & \textbf{80.80} \\
attributes & 44.32 & 67.58 & \textbf{46.16} & \textbf{70.30} \\
relationships & 47.57 & 68.10 & \textbf{50.65} & \textbf{71.54} \\
overall & 49.75 & 71.75 & \textbf{51.75} & \textbf{74.31} \\
\hline
\end{tabular}
\end{table}
\fi
\iffalse
\begin{table}[tb]
\centering
\begin{tabular}{c|cc}
\hline
1&Image & Text \\
\hline
&\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.70\linewidth]{disc.jpg}
\end{minipage}
&
\begin{minipage}{0.20\textwidth}
\small{ a black dog about \\ to catch a flying \textcolor{red}{\textbf{disc}} .}
\end{minipage} \\ \hline
& \small{BERT, w SG} & \small{BERT, w/o SG} \\ \hline
&\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.7\linewidth]{disc_sg.pdf}
\end{minipage}
&
\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.7\linewidth]{disc_base.pdf}
\end{minipage} \\ \hline
\end{tabular}
\end{table}
\begin{table}
\centering
\begin{tabular}{c|cc}
\hline
2 &Image & Text \\
\hline
&\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.70\linewidth]{red.jpg}
\end{minipage}
&
\begin{minipage}{0.20\textwidth}
\small{two men wearing \textcolor{red}{\textbf{red}} jackets are looking out over some water and one man has yellow earphones on his ears.}
\end{minipage} \\ \hline
& \small{BERT, w SG} & \small{BERT, w/o SG} \\ \hline
&\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.7\linewidth]{red_sg.pdf}
\end{minipage}
&
\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.7\linewidth]{red_base.pdf}
\end{minipage} \\ \hline
\end{tabular}
\end{table}
\begin{table}
\begin{tabular}{c|cc}
\hline
3 &Image & Text \\
\hline
&\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.70\linewidth]{kicking.jpg}
\end{minipage}
&
\begin{minipage}{0.20\textwidth}
\small{a little boy in a green shirt \textcolor{red}{\textbf{kicks}} a ball.}
\end{minipage} \\ \hline
& \small{BERT, w SG} & \small{BERT, w/o SG} \\ \hline
&\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.7\linewidth]{kicks_sg.pdf}
\end{minipage}
&
\begin{minipage}{0.20\textwidth}
\includegraphics[width=0.7\linewidth]{kicks_base.pdf}
\end{minipage} \\ \hline
\end{tabular}
\caption{Effectiveness of Scene Graph Predictions Strategies for ERINE-ViL. Each sample represents an image-text pair with the masked token colored in red.}
\label{tab_cloze}
\end{table}
\fi
Among the methods pre-trained on the same out-of-domain datasets (CC and SBU),
ERNIE-ViL obtains the best performances on all 5 downstream tasks. For the visual reasoning tasks, ERNIE-ViL-large achieves a significant improvement of 6.60\% on VCR (Q$\to$AR) task and 1.74\% on VQA (test-std) task compared with VLBERT-large. On visual grounding task, ERNIE-ViL-large obtains an improvement of 2.40\% for both testA split and testB split on RefCOCO+ task compared to VLBERT-large. On the cross-modal retrieval tasks, where no large models pre-trained on out-of-domain datasets has released results, ERNIE-ViL-base achieves an imporvement of 2.94\% on R@1 for image retrieval and 0.50\% on R@1 for text retrieval compared with Unicoder-VL-base.
For further comparison with those models pretrained with both out-of-domain and in-domain datasets, we pre-train ERINE-ViL with all these datasets. As illustrated in Table \ref{tab_result}, ERINE-ViL-large acheives state-of-the-art performances on these tasks compared to existing works, e.g., UNITER, OSCAR\cite{li2020oscar} and VILLA\cite{gan2020large}.
\subsection{Analysis}
\paragraph{Effectiveness of Scene Graph Prediction tasks}
To verify the effectiveness of Scene Graph Prediction (\textbf{SGP}) tasks, we first conduct experiments with ERNIE-ViL-base settings based on the text parameters initialized from BERT. As illustrated in Table \ref{tab_abl}, pre-training with \textbf{SGP} tasks
in ERNIE-ViL brings significant improvements across all downstream tasks. Especially on Grounding Referring Expressions and Retrieval tasks, those require understanding detailed semantics alignments, \textbf{SGP} tasks make an improvement of 0.69\% accuracy on RefCOCO+ and 2.22\% of R@1 for image retrieval on Flickr30K.
Note that text parameter initialized from ERNIE 2.0 can lead to further improvements on all tasks and a relatively large improvement on VCR task. We considere that through continually learning on various pre-training tasks, ERNIE 2.0 learned more common sense knowledge which benefits the VCR task.
Overall, the \textbf{SGP} tasks significantly contribute to the state-of-the-art results of ERNIE-ViL.
\paragraph{Cloze Test} To get a more intuitively understanding of the improvements brought by SGP tasks, we conduct the language cloze test conditioned on the visual modality. In the cloze test, language tokens represent detailed semantics (objects, attributes and relationships) are masked from the text and the model is required to infer them with the context from both the text and the image. To construct the dataset, we sampled 15,000 image-text pairs from Flickr30K dataset and 5,000 objects, attributes and relationships tokens
each are selected. For the prediction, the top one accuracy (ACC@1) and top five accuracy (ACC@5) are adopted as the evaluation metric. The comparison of prediction results between two models, which are pre-trained models with \textbf{SGP} task and without \textbf{SGP} task, are illustrated in Table \ref{tab_clone_acc}.
The text-stream parameters of both models are initialized from BERT. An absolute improvement of 1.20\% for objects, 3.08\% for relationships and 1.84\% for attributes on ACC@1 demonstrates that ERNIE-ViL pre-trained with \textbf{SGP} tasks learns better cross-modal detailed semantics alignments.
Moreover, we illustrate some cases in Table \ref{tab_cloze}, and the top $5$ possible predictions are shown in the right columns. As in case 1-2, model pre-trained without \textbf{SGP} tasks cannot make the right predictions as it didn't learn accurate alignments of detailed semantics, without distinguishing common words and detailed semantics words while pre-training.
While in case 3, the model can predict the reasonable tokens but with lower confidence compared with model pre-trained with \textbf{SGP} tasks.
\section{Conclusion}
We proposed ERNIE-ViL to learn the joint representations of vision and language. In addition to conventional MLM for cross-modal pre-training, we introduce Scene graph Prediction tasks to characterize the cross-modal detailed semantic alignments. Experiment results on various downstream tasks demonstrate the improvements of incorporating structured knowledge obtained from scene graphs during cross-modal pre-training. For future work, scene graphs extracted from images could also be incorporated into cross-modal pre-training. Moreover, Graph Neural Networks that integrate more structured knowledge could be considered as well.
\iffalse
\section{NOTES}
\noindent Congratulations on having a paper selected for inclusion in an AAAI Press proceedings or technical report! This document details the requirements necessary to get your accepted paper published using PDF\LaTeX{}. If you are using Microsoft Word, instructions are provided in a different document. AAAI Press does not support any other formatting software.
The instructions herein are provided as a general guide for experienced \LaTeX{} users. If you do not know how to use \LaTeX{}, please obtain assistance locally. AAAI cannot provide you with support and the accompanying style files are \textbf{not} guaranteed to work. If the results you obtain are not in accordance with the specifications you received, you must correct your source file to achieve the correct result.
These instructions are generic. Consequently, they do not include specific dates, page charges, and so forth. Please consult your specific written conference instructions for details regarding your submission. Please review the entire document for specific instructions that might apply to your particular situation. All authors must comply with the following:
\begin{itemize}
\item You must use the 2021 AAAI Press \LaTeX{} style file and the aaai21.bst bibliography style files, which are located in the 2021 AAAI Author Kit (aaai21.sty, aaai21.bst).
\item You must complete, sign, and return by the deadline the AAAI copyright form (unless directed by AAAI Press to use the AAAI Distribution License instead).
\item You must read and format your paper source and PDF according to the formatting instructions for authors.
\item You must submit your electronic files and abstract using our electronic submission form \textbf{on time.}
\item You must pay any required page or formatting charges to AAAI Press so that they are received by the deadline.
\item You must check your paper before submitting it, ensuring that it compiles without error, and complies with the guidelines found in the AAAI Author Kit.
\end{itemize}
\section{Copyright}
All papers submitted for publication by AAAI Press must be accompanied by a valid signed copyright form. They must also contain the AAAI copyright notice at the bottom of the first page of the paper. There are no exceptions to these requirements. If you fail to provide us with a signed copyright form or disable the copyright notice, we will be unable to publish your paper. There are \textbf{no exceptions} to this policy. You will find a PDF version of the AAAI copyright form in the AAAI AuthorKit. Please see the specific instructions for your conference for submission details.
\section{Formatting Requirements in Brief}
We need source and PDF files that can be used in a variety of ways and can be output on a variety of devices. The design and appearance of the paper is strictly governed by the aaai style file (aaai21.sty).
\textbf{You must not make any changes to the aaai style file, nor use any commands, packages, style files, or macros within your own paper that alter that design, including, but not limited to spacing, floats, margins, fonts, font size, and appearance.} AAAI imposes requirements on your source and PDF files that must be followed. Most of these requirements are based on our efforts to standardize conference manuscript properties and layout. All papers submitted to AAAI for publication will be recompiled for standardization purposes. Consequently, every paper submission must comply with the following requirements:
\begin{quote}
\begin{itemize}
\item Your .tex file must compile in PDF\LaTeX{} --- (you may not include .ps or .eps figure files.)
\item All fonts must be embedded in the PDF file --- including includes your figures.
\item Modifications to the style file, whether directly or via commands in your document may not ever be made, most especially when made in an effort to avoid extra page charges or make your paper fit in a specific number of pages.
\item No type 3 fonts may be used (even in illustrations).
\item You may not alter the spacing above and below captions, figures, headings, and subheadings.
\item You may not alter the font sizes of text elements, footnotes, heading elements, captions, or title information (for references and mathematics, please see the limited exceptions provided herein).
\item You may not alter the line spacing of text.
\item Your title must follow Title Case capitalization rules (not sentence case).
\item Your .tex file must include completed metadata to pass-through to the PDF (see PDFINFO below).
\item \LaTeX{} documents must use the Times or Nimbus font package (you may not use Computer Modern for the text of your paper).
\item No \LaTeX{} 209 documents may be used or submitted.
\item Your source must not require use of fonts for non-Roman alphabets within the text itself. If your paper includes symbols in other languages (such as, but not limited to, Arabic, Chinese, Hebrew, Japanese, Thai, Russian and other Cyrillic languages), you must restrict their use to bit-mapped figures. Fonts that require non-English language support (CID and Identity-H) must be converted to outlines or 300 dpi bitmap or removed from the document (even if they are in a graphics file embedded in the document).
\item Two-column format in AAAI style is required for all papers.
\item The paper size for final submission must be US letter without exception.
\item The source file must exactly match the PDF.
\item The document margins may not be exceeded (no overfull boxes).
\item The number of pages and the file size must be as specified for your event.
\item No document may be password protected.
\item Neither the PDFs nor the source may contain any embedded links or bookmarks (no hyperref or navigator packages).
\item Your source and PDF must not have any page numbers, footers, or headers (no pagestyle commands).
\item Your PDF must be compatible with Acrobat 5 or higher.
\item Your \LaTeX{} source file (excluding references) must consist of a \textbf{single} file (use of the ``input" command is not allowed.
\item Your graphics must be sized appropriately outside of \LaTeX{} (do not use the ``clip" or ``trim'' command) .
\end{itemize}
\end{quote}
If you do not follow these requirements, your paper will be returned to you to correct the deficiencies.
\section{What Files to Submit}
You must submit the following items to ensure that your paper is published:
\begin{itemize}
\item A fully-compliant PDF file that includes PDF metadata.
\item Your \LaTeX{} source file submitted as a \textbf{single} .tex file (do not use the ``input" command to include sections of your paper --- every section must be in the single source file). (The only allowable exception is .bib file, which should be included separately).
\item The bibliography (.bib) file(s).
\item Your source must compile on our system, which includes only standard \LaTeX{} 2020 TeXLive support files.
\item Only the graphics files used in compiling paper.
\item The \LaTeX{}-generated files (e.g. .aux, .bbl file, PDF, etc.).
\end{itemize}
Your \LaTeX{} source will be reviewed and recompiled on our system (if it does not compile, your paper will be returned to you. \textbf{Do not submit your source in multiple text files.} Your single \LaTeX{} source file must include all your text, your bibliography (formatted using aaai21.bst), and any custom macros.
Your files should work without any supporting files (other than the program itself) on any computer with a standard \LaTeX{} distribution.
\textbf{Do not send files that are not actually used in the paper.} We don't want you to send us any files not needed for compiling your paper, including, for example, this instructions file, unused graphics files, style files, additional material sent for the purpose of the paper review, and so forth.
\textbf{Do not send supporting files that are not actually used in the paper.} We don't want you to send us any files not needed for compiling your paper, including, for example, this instructions file, unused graphics files, style files, additional material sent for the purpose of the paper review, and so forth.
\textbf{Obsolete style files.} The commands for some common packages (such as some used for algorithms), may have changed. Please be certain that you are not compiling your paper using old or obsolete style files.
\textbf{Final Archive.} Place your PDF and source files in a single archive which should be compressed using .zip. The final file size may not exceed 10 MB.
Name your source file with the last (family) name of the first author, even if that is not you.
\section{Using \LaTeX{} to Format Your Paper}
The latest version of the AAAI style file is available on AAAI's website. Download this file and place it in the \TeX\ search path. Placing it in the same directory as the paper should also work. You must download the latest version of the complete AAAI Author Kit so that you will have the latest instruction set and style file.
\subsection{Document Preamble}
In the \LaTeX{} source for your paper, you \textbf{must} place the following lines as shown in the example in this subsection. This command set-up is for three authors. Add or subtract author and address lines as necessary, and uncomment the portions that apply to you. In most instances, this is all you need to do to format your paper in the Times font. The helvet package will cause Helvetica to be used for sans serif. These files are part of the PSNFSS2e package, which is freely available from many Internet sites (and is often part of a standard installation).
Leave the setcounter for section number depth commented out and set at 0 unless you want to add section numbers to your paper. If you do add section numbers, you must uncomment this line and change the number to 1 (for section numbers), or 2 (for section and subsection numbers). The style file will not work properly with numbering of subsubsections, so do not use a number higher than 2.
\subsubsection{The Following Must Appear in Your Preamble}
\begin{quote}
\begin{scriptsize}\begin{verbatim}
\def\year{2021}\relax
\documentclass[letterpaper]{article}
\usepackage{aaai21}
\usepackage{times}
\usepackage{helvet}
\usepackage{courier}
\usepackage[hyphens]{url}
\usepackage{graphicx}
\urlstyle{rm}
\def\UrlFont{\rm}
\usepackage{graphicx}
\usepackage{natbib}
\usepackage{caption}
\frenchspacing
\setlength{\pdfpagewidth}{8.5in}
\setlength{\pdfpageheight}{11in}
\pdfinfo{
/Title (AAAI Press Formatting Instructions for Authors
Using LaTeX -- A Guide)
/Author (AAAI Press Staff, Pater Patel Schneider,
Sunil Issar, J. Scott Penberthy, George Ferguson,
Hans Guesgen, Francisco Cruz, Marc Pujol-Gonzalez)
/TemplateVersion (2021.1)
}
\end{verbatim}\end{scriptsize}
\end{quote}
\subsection{Preparing Your Paper}
After the preamble above, you should prepare your paper as follows:
\begin{quote}
\begin{scriptsize}\begin{verbatim}
\begin{document}
\maketitle
\begin{abstract}
\end{abstract}\end{verbatim}\end{scriptsize}
\end{quote}
\noindent You should then continue with the body of your paper. Your paper must conclude with the references, which should be inserted as follows:
\begin{quote}
\begin{scriptsize}\begin{verbatim}
|
\section{Other Normalization Methods}
\label{app:normalization}
A simple normalization technique is to map scores on an environment $j$ that are in the range $[a_j,b_j]$ to $[0,1]$, i.e., $g(x,j) \coloneqq (x-a_j) / (b_j - a_j)$ \citep{bellemare13arcade}. However, this can result in normalized performance measures that cluster in different regions of $[0,1]$ for each environment. For example, consider one environment where a the minimum is $-100$, the maximum is $10$ and a uniform random policy can score around $0$. Similarly consider a second environment where the minimum score is $10$, maximum is $1,\!000$, and random gets around $20$. On the first environment, algorithms will tend to have a normalized performance near $1$ and in the second case most algorithms will have a normalized performance near $0$. So in the second environment algorithms will likely appear worse than algorithms in the first regardless of how close to optimal they are. This means the normalized performances are not really comparable.
A different version of this approach uses the minimum and maximum mean performance of each algorithm \citep{bellemare13arcade,balduzzi2018eval}. Let $\hat \mu_{i,j}$ be the sample mean of $X_{i,j}$. Then this normalization method uses the following function, $\bar g(i,j) \coloneqq (\hat \mu_{i,j} - \min_{i'} \hat \mu_{i',j}) / (\max_{i'} \hat \mu_{i',j} - \min_{i'} \hat \mu_{i',j})$. This sets the best algorithm's performance on each environment to $1$ and the worst to $0$, spreading the range of values out over the whole interval $[0,1]$. This normalization technique does not correct for nonlinear scaling of performance. As a result algorithms could be near $0$ or $1$ if there is an outlier algorithm that does very well or poorly. For example, one could introduce a terrible algorithm that just chooses one action the whole time. This makes the environment seem easier as all scores would be near $1$ except for this bad algorithm. We would like the evaluation procedure to be robust to the addition of poor algorithms.
An alternative normalization technique proposed by \citet{whiteson2011eval} uses the probability that one algorithm outperforms another on an environment, $j$, i.e., $\Pr(X_{i,j} \ge X_{k,j})$. This technique is intuitive and straight forward to estimate but neglects the difference in score magnitudes. For example, consider that algorithm $i$ always scores a $1.0$ and algorithm $k$ always scores $0.99$, the probability that $i$ is better than $k$ is $1$, but the difference between them is small, and the normalized score of $1.0$ neglects this difference.
\section{$\alpha$-Rank and our Implementation}
\label{app:alpharank}
The $\alpha$-Rank procedure finds a solution to a game by computing the stationary distribution of strategy profiles when each player is allowed to change their strategy. This is done by constructing a directed graph where nodes are pure strategies and edges have weights corresponding to the probability that one of the players switches strategies. This graph can be represented by a Markov matrix, $C \in [0,1]^{|\mathcal{S}| \times |\mathcal{S}|}$. The entry $C_{s,s'}$ corresponds to a probability of switching from a strategy $s$ to $s'$. Only one player is allowed to change strategies at a time, so the possible transitions for a strategy $s=(i,(j,k))$, are any strategies $s'=(i',(j,k))$ or $s'=(i,(j',k'))$ for all $i',k' \in \mathcal{A}$ and $j' \in \mathcal{M}$.
The typical $\alpha$-Rank procedure uses transition probabilities, $C_{s,s'}$, that are based on a logistic transformation of the payoff difference $u_l(s') - u_l(s)$. These differences are scaled by a parameter $\alpha$ and as $\alpha$ approaches $\infty$, the transition matrix approximates the Markov Conley chain (MCC), which is the motivation for using $\alpha$-Rank as a solution concept for games. See the work of \citet{omidshafiei2019alpharank} for more detailed information.
The entries of the matrix for valid transitions are:
\begin{equation}
C_{s,s'} =
\begin{cases}
\eta \frac{1-\exp{\left (-\alpha (u_l(s')-u_l(s)) \right )}}{1-\exp{\left (-\alpha n(u_l(s')-u_l(s)) \right )}} & \text{if $u_l(s') \!= u_l(s)$}\\
\frac{\eta}{n} & \text{if $u_l(s') = u_l(s)$}\\
\end{cases}
\text{ and } C_{s,s} = 1 - \sum_{s' \ne s}C_{s,s'},
\end{equation}
where $\eta = (\sum_{k} |\mathcal{S}^k|-1)^{-1}$, $l$ represents the player who switched from strategy $s$ to $s'$, $n$ is the population constant (we set it to $50$ following the prior work).
The equilibrium over strategies is then given by the stationary distribution $d$ of the Markov chain induced by $C$, i.e., $d$ is a distribution such that $d = d \tilde C$. The equilibrium solution $p^*, q^*$ are then the sum of probabilities in $d$ for each strategy, i.e., $p^*_{s_1} = \sum_{s_2 \in \mathcal{S}_2} d_{s_1,s_2}$ and $q^*_{s_2} = \sum_{s_1 \in \mathcal{S}_1} d_{s_1,s_2}$.
The aggregate performance can then be computed using $q^*$ as in \eqref{eq:aggforalgi}.
Theoretically, the hyperparameter $\alpha$ could be chosen arbitrarily high and the matrix $C$ would still be irreducible, i.e., for all $s \in \mathcal{S}$, $d_{s} > 0$ and $d$ is unique. However, due to numerical precision issues, a high value of $\alpha$ sets transition probabilities to zero for some dominated strategies, i.e., $u_l(s') < u_l(s)$, which can result in a matrix that is reducible. The suggested method to chose $\alpha$ is to tune it on a logarithmic scaled to find the highest value such that the transition matrix, $C$, is still irreducible \citep{omidshafiei2019alpharank}.
This strategy works when the payoffs are known, but when they represent empirical samples of performance, then the value of $\alpha$ chosen will depend on the empirical payoff functions. Setting $\alpha$ based solely on the empirical payoffs could introduce bias to the matrix based on that sample. So we need a different solution without a data dependent hyperparameter.
In the MCC graph construction, all edges leading to strategies with strictly greater payoffs have the same positive weight. All edges that lead to strategies with the same payoff have he same weight but less than that of the strictly greater payoff. There are no transitions to strategies with worse payoffs.
As $\alpha \to \infty$ the transitions probabilities quickly saturate to $\eta$ if $u_l(s') > u_l(s)$ and $0$ if $u_l(s') < u_l(s)$. So we construct the transition matrix, $C$, differently using the saturation values to set the transition probabilities $C$ is close to the MCC construction.
The entries for $C$ that represent valid transitions in the graph are:
\begin{equation}
\label{eq:cmat}
C_{s,s'} \coloneqq
\begin{cases}
\eta & \text{if $u_l(s') > u_l(s)$}\\
\frac{\eta}{m} & \text{if $u_l(s') = u_l(s)$}\\
0 & \text{otherwise}
\end{cases}
\text{ } C_{s,s} \coloneqq 1 - \sum_{s' \ne s}C_{s,s'}.
\end{equation}
However, this often makes the transition matrix reducible, i.e., the stationary distribution might have mass on only one strategy.
To ensure $C$ is irreducible we follow the damping approach used in PageRank \citep{page1999pagerank}, i.e., $\tilde C = \gamma C + (1-\gamma) (1/|\mathcal{S}|)$, where $\gamma \in (0,1)$ is a hyperparameter and $1-\gamma$ represents the probability of randomly switching to any strategy in $\mathcal{S}$. During a Monte-Carlo simulation of transitions through $\tilde C$, states will transition from $s$ to $s'$ according to $C$, but with probability $1-\gamma$ the transition ignores $C$ and switches to some strategy $s' \in \mathcal{S}$ chosen uniformly at random.
For $\gamma=1$ the matrix is unchanged and represents the MCC solution, but is reducible. For $\gamma$ near one, the stationary distribution will be similar to the solution given by the MCC solution with high weight placed on dominate strategies and small weight given to weak ones. As $\gamma \to 0$ the stationary distribution becomes more uniform as it is only considering shorter sequences of transitions before a random switch occurs.
This method differs from the infinite-$\alpha$ approach presented by \citet{rowland2019alphabounds}, but in the limit as $\gamma \to 1$ and $\alpha \to \infty$, the solutions have small differences. The approach using $\gamma$ has a benefit in that there is no data dependent hyperparameter and it has a simple interpretation.
We chose to set $\gamma = (|\mathcal{S}|-1) / (|\mathcal{S}|)$ so that the expected number of transitions to occur before a random jump is $|\mathcal{S}|$. This allows for propagation of transition probabilities to cover every strategy combination. We could have chosen to set $\gamma$ near one, e.g., $\gamma = 1-10^{-8}$, but this would make the computation of the confidence intervals take longer. This is because optimizing the $C$ within confidence intervals $[C^-, C^+]$ (defined in the next section) is equivalent finding the optimal value function of a Markov decision process (MDP) with a discount parameter of $\gamma$. See the work of \citet{kerchove2007maxpage,fercoq2013pagerankopt,csaji2014pagerankopt} for more information on this connection. Solving and MDP with a discount $\gamma$ near $1.0$ causes the optimization process of value iteration and policy iteration to converge slower than if $\gamma$ is small. So we chose $\gamma$ such that it could still find solutions near the MCC solution, but remain computationally efficient.
Using this new definition of $C$ we use the following alternative but equivalent method to compute the aggregate performance more efficiently \citep{fercoq2013pagerankopt}:
\begin{align}
\label{eq:aggalgivalue}
y_i = \frac{1 - \gamma}{|\mathcal{S}|} \sum_{s \in \mathcal{S}} v(s) \quad
v = (I - \gamma C)^{-1} R_i,
\end{align}
where $R_i \in \mathbb{R}^{|\mathcal{S}|}$ is a vector with entries $R_i(s) \coloneqq \mathbf{E}[F_{X_{k,j}}(X_{i,j})]$ with $s=(s_1,s_2)$ and $s_2 = (j,k)$. Notice that $s_1$ is ignored because $i$ is already specified by $R_i$.
\section{Confidence Intervals on the Aggregate Performance}
\label{app:confagg}
\begin{table}
\centering
\begin{tabular}{|r|l|}
\hline
\multicolumn{2}{|c|}{\textbf{Symbol List}} \\
\hline
\multicolumn{1}{|c|}{\textbf{Symbol}} & \multicolumn{1}{c|}{\textbf{Description}} \\
\hline
\rowcolor[HTML]{EFEFEF}
$\mathcal{A}$ & set of algorithms in the evaluation \\
$\mathcal{M}$ & set of environments in the evaluation \\
\rowcolor[HTML]{EFEFEF}
$X_{i,j}$ & random variable representing performance of algorithm $i$ on environment $j$ \\
$T_{i,j}$ & number of samples of performance for algorithm $i$ on environment $j$ \\
\rowcolor[HTML]{EFEFEF}
$x_{i,j,t}$ & the~$t^\text{th}$sample of performance of algorithm~ $i$ on environment $j$ and sorted such that $x_{i,j,t-1} \le x_{i,j,t}$ \\
$D$ & data set containing all samples of performance for each algorithm on each environment \\
\rowcolor[HTML]{EFEFEF}
$y \in \mathbb{R}^{|\mathcal{A}|}$ & $y_i$ is the aggregate performance for each algorithm $i$ \\
$Y^-,Y^+ \in \mathbb{R}^{|\mathcal{A}|}$ & lower and upper confidence intervals on $y$ computed using $D$ \\
\rowcolor[HTML]{EFEFEF}
$\delta \in (0,0.5]$ & confidence level for the aggregate performance \\
$F_{X_{i,j}}$ & cumulative distribution function (CDF) of$X_{i,j}$ and is also used for normalization \\
\rowcolor[HTML]{EFEFEF}
$\hat F_{X_{i,j}}$ & empirical cumulative distribution function constructed using samples $x_{i,j,\cdot}$ \\
$F^-_{X_{i,j}}$, $F^+_{X_{i,j}}$ & lower and upper confidence intervals on $F_{X_{i,j}}$ computed using $D$ \\
\rowcolor[HTML]{EFEFEF}
$z_{i,j,k}$ & performance of algorithm $i$, i.e., $z_{i,j,k} = \mathbf{E}[F_{X_{k,j}}(X_{i,j})]$ \\
$Z^-_{i,j,k},Z^+_{i,j,k}$ & lower and upper confidence intervals on $z_{i,j,k}$ computed using $D$. \\
\rowcolor[HTML]{EFEFEF}
$s_1 \in \mathcal{S}_1$ & strategy for player $p$ where $\mathcal{S}_1 = \mathcal{A}$ and $s_1$ is often denoted using $i $ \\
$s_2 \in \mathcal{S}_2$ & strategy for player $q$ where $\mathcal{S}_2 = \mathcal{M} \times \mathcal{A}$ and $s_2$ is often denoted using $(j,k)$ \\
\rowcolor[HTML]{EFEFEF}
$s \in \mathcal{S}$ & joint strategy where $\mathcal{S} = \mathcal{S}_1 \times \mathcal{S}_2,$ and $s=(s_1,s_2)$ is often denoted as $(i,(j,k))$ \\
$p \in \Delta(\mathcal{S}_1)$ & strategy for player $p$ represented as a distribution over $\mathcal{S}_1$ \\
\rowcolor[HTML]{EFEFEF}
$q \in \Delta(\mathcal{S}_2)$ & strategy for player $q$ represented as a distribution over $\mathcal{S}_2$ \\
$u_p(s)$ & payoff for player $p$ when $s$ is played, i.e.,$ u_p(s) = \mathbf{E}[F_{X_{k,j}}(X_{i,j})]$ \\
\rowcolor[HTML]{EFEFEF}
$u_q(s)$ & payoff for player $q$ when $s$ is played, i.e., $u_q(s) = - u_p(s)$ \\
$u_l^-(s), u_l^+(s)$ & confidence intervals on $u_l(s)$ for player $l \in \{p,q\}$ computed using D \\
\hline
\end{tabular}
\caption{List of symbols used to create confidence intervals on the aggregate performance.}
\label{tbl:symbols}
\end{table}
In this section, we detail the PBP procedure for computing confidence intervals $Y^-$ and $Y^+$ on the aggregate performance $y$ and prove that they hold with high probability.
That is, we show that for any confidence level $\delta \in (0, 0.5]$;
\begin{equation}
\Pr(\forall i \in \mathcal{A}, \ y_i \in [Y^-_i, Y^+_i]) \ge 1 - \delta.
\end{equation}
We will first describe the PBP procedure to compute confidence intervals and then prove that they valid.
A list of the symbols used in the construction of confidence intervals and their description are provided in Table \ref{tbl:symbols} to refresh the reader.
The steps to compute the confidence intervals are outlined in Algorithm \ref{alg:pbp}
Recall that the aggregate performance for an algorithm $i$ is
\begin{equation}
y_i \coloneqq \sum_{j=1}^{|\mathcal{M}|}\sum_{k=1}^{|\mathcal{A}|} q^*_{j,k} \mathbf{E}[F_{X_{k,j}}(X_{i,j})],
\end{equation}
where $q^*$ is the equilibrium solution to the game specified in Section \ref{sec:summarization}.
To compute valid confidence intervals $Y^-,Y^+$ on $y$ using a dataset $D$, the uncertainty of $q^*$ and mean normalized performance $z_{i,j,k} = \mathbf{E}[F_{X_{k,j}}(X_{i,j})]$. PBP accomplishes this by three primary steps. The first step is to compute confidence intervals $Z^-_{i,j,k},Z^+_{i,j,k}$ on $z_{i,j,k}$ such that
\begin{equation}
\Pr \left (\forall (i,j,k) \in \mathcal{A} \times \mathcal{M} \times \mathcal{A},\ z_{i,j,k} \in [Z^-_{i,j,k}, Z^+_{i,j,k}] \right ) \ge 1-\delta.
\end{equation}
The second step is to compute the uncertainty set $\mathcal{Q}$ containing all possible $q^*$ that are compatible with $Z^-$ and $Z^+$.
The last step is to compute the smallest and largest possible aggregate performances for each algorithm over these sets, i.e.,
\begin{align}
Y^-_i = \min_{q \in \mathcal{Q}} \sum_{j=1}^{|\mathcal{M}|}\sum_{k=1}^{|\mathcal{A}|} q_{j,k} Z^-_{i,j,k} \text{ and }\quad
Y^+_i = \max_{q \in \mathcal{Q}} \sum_{j=1}^{|\mathcal{M}|}\sum_{k=1}^{|\mathcal{A}|} q_{j,k} Z^+_{i,j,k}. \\
\end{align}
PBP follows this process, except in the last two steps $\mathcal{Q}$ is never explicitly constructed to improve computational efficiency.
Intuitively, the procedure provides valid confidence intervals because all values to compute the aggregate performance depend on the normalized performance. So by guaranteeing with probability at least $1-\delta$ that the true mean normalized performances will be between $Z^-$ and $Z^+$, then so long as the the upper (lower) confidence interval computed is at least as large (small) as the maximum (minimum) of the aggregate score for any setting of $z \in [Z^-, Z^+]$, the confidence intervals will be valid.
We break the rest of this section into two subsections. The first subsection discusses constructing the confidence intervals on the mean normalized performance and proving their validity. The second subsection describes how to construct the confidence intervals on the aggregate performance proves their validity.
\begin{algorithm}[tb]
\caption{Performance Bound Propagation (PBP)}
\label{alg:pbp}
\begin{algorithmic}[1]
\STATE {\bfseries Input:} dataset $D$ containing samples of performance and a confidence level $\delta \in (0,0.5])$
\STATE {\bfseries Output:} $Y^-$, $Y^+$ confidence intervals on the aggregate performance
\\
\hrulefill
\STATE $\delta^\prime \gets \delta / (|\mathcal{A}||\mathcal{M}|)$;
\STATE $\texttt{sort\_ascending}(\{x_{i,j,t}\}_{t=1}^{T_{i,j}})$;
\STATE // Compute confidence intervals for the CDFs \label{alg:lines:dkwstart}
\FOR{$i,j \in \mathcal{A} \times \mathcal{M}$}
\STATE $F_{X_{i,j}}^-, F_{X_{i,j}}^+ \gets \texttt{dkw\_bound}(\{x_{i,j,t}\}_{t=1}^{T_{i,j}}, \delta^\prime)$; // computation shown in \eqref{eq:dkw} \label{alg:lines:dkwstop}
\ENDFOR
\STATE // Compute confidence intervals on the mean normalized performance
\FOR{$i,j,k \in \mathcal{A} \times \mathcal{M} \times \mathcal{A}$}
\STATE $Z^-_{i,j,k} \gets F^-_{X_{k,j}}(x_{i,j,T}) - \sum_{t=0}^{T-1} \left [ F^-_{X_{k,j}}(x_{i,j,t+1})-F^-_{X_{k,j}}(x_{i,j,t}) \right ] F^+_{X_{i,j}}(x_{i,j,t})$; \\
\STATE $Z^+_{i,j,k} \gets F^+_{X_{k,j}}(x_{i,j,T+1}) - \sum_{t=1}^{T} \left [ F^+_{X_{k,j}}(x_{i,j,t+1})-F^+_{X_{k,j}}(x_{i,j,t}) \right ] F^-_{X_{i,j}}(x_{i,j,t})$;
\ENDFOR
\STATE // Construct game quantities
\STATE $\mathcal{S} = \mathcal{A} \times (\mathcal{M} \times \mathcal{A})$; \ strategy profile set
\STATE $\gamma \gets \frac{|\mathcal{S}|-1}{|\mathcal{S}|}$
\STATE $C^-, C^+ \gets \texttt{bound\_markov\_matrix}(Z^-, Z^+)$ as defined in \eqref{eq:cbounds}.
\STATE // Optimize aggregate performance over all possible $C \in [C^-,C^+]$
\FOR{$i \in \mathcal{A}$}
\STATE $v \gets \texttt{find\_optimal\_valuefunction}(C^-, C^+, R_i = -Z^-_{i,\cdot,\cdot})$; // solve \eqref{eq:lpbound}
\STATE $Y^-_i \gets \frac{(1-\gamma)}{|\mathcal{S}|} \sum_{s \in \mathcal{S}}|v(s)|$
\STATE $v \gets \texttt{find\_optimal\_valuefunction}(C^-, C^+, R_i = Z^+_{i,\cdot,\cdot})$; // solve \eqref{eq:lpbound}
\STATE $Y^+_i \gets \frac{(1-\gamma)}{|\mathcal{S}|} \sum_{s \in \mathcal{S}}|v(s)|$
\ENDFOR
\end{algorithmic}
\end{algorithm}
\subsection{Confidence intervals on the normalized performance}
The normalized performance $z_{i,j,k} = \mathbf{E}[F_{X_{k,j}}(X_{i,j})]$ has two unknowns, $F_{X_{k,j}}$ and the distribution of $X_{i,j}$.
To compute confidence intervals on $z_{i,j,k}$ for all $i,j,k$, confidence intervals are needed on the output on all distribution functions $F_{X_{i,j}}$.
The confidence intervals on the distributions can then be combined to get confidence intervals on $z_{i,j,k}$.
To compute confidence intervals on $F_{X_{i,j}}$ we assume that $X_{i,j}$ is bounded on the interval $[a_j, b_j]$ for all $i \in \mathcal{A}$ and $j \in \mathcal{M}$. Let $\hat F_{X_{i,j}}$ be the empirical CDF with
\begin{equation}
\hat F_{X_{i,j}}(x) \coloneqq \frac{1}{T_{i,j}} \sum_{t=1}^{T_{i,j}} \mathbf{1}_{x_{i,j,t} \le x},
\end{equation}
where $T_{i,j}$ is the number of samples of $X_{i,j}$, $x_{i,j,t}$ is the $t^\text{th}$ sample of $X_{i,j}$, and $\mathbf{1}_{A} = 1$ if event $A$ is true and $0$ otherwise.
Using the Dvoretzky--Kiefer--Wolfowitz (DKW) inequality \citep{dvoretzky1956bound} with tight constants \citep{massart1990bound}, we define $F^-_{X_{i,j}}$ and $F^+_{X_{i,j}}$ to be the lower and upper confidence intervals on $F_{X_{i,j}}$, i.e.,
\begin{equation}
\label{eq:dkw}
\begin{aligned}
F^+_{X_{i,j}}(x) \coloneqq &\
\begin{cases}
1 & \text{if $x \ge b$}\\
\min(1.0, \hat F_{X_{i,j}}(x) + \epsilon) & \text{if $a \le x < b$}\\
0 & \text{if $x < a$} \\
\end{cases} \\
F^-_{X_{i,j}}(x) \coloneqq &\
\begin{cases}
1 & \text{if $x \ge b$}\\
\max(0.0, \hat F_{X_{i,j}}(x) - \epsilon) & \text{if $a \le x < b$}\\
0 & \text{if $x < a$} \\
\end{cases}
\end{aligned}
\quad \text{ and } \epsilon = \sqrt{\frac{\ln \frac{2}{\delta'}}{2T_{i,j}}},
\end{equation}
where $\delta' \in (0,0.5]$ is a confidence level and we use $\delta' = \delta / (|\mathcal{A}||\mathcal{M}|)$.
By the DKW inequality with tight constants the following property is known:
\begin{prop}[DKW with tight constants confidence intervals]
\label{prop:dkw}
\begin{equation}
\Pr \left ( \forall x \in \mathbb{R}, \ \ F_{X_{i,j}}(x) \in [F_{X_{i,j}}^-(x), F_{X_{i,j}}^+(x)] \right ) \ge 1-\delta'.
\end{equation}
\end{prop}
\begin{proof}
See the works of \citet{dvoretzky1956bound} and \citet{massart1990bound}.
\end{proof}
Further, by the union bound we have that
\begin{align}
\label{eq:unioncdf}
\Pr \left (\forall i \in \mathcal{A}, \forall j \in \mathcal{M}, \forall x \in \mathbb{R}, \ \ F_{X_{i,j}}(x) \in [F_{X_{i,j}}^-(x), F_{X_{i,j}}^+(x)] \right ) \ge 1-\delta.
\end{align}
To construct confidence intervals on the mean normalized performance, we will use Anderson's inequality \citep{anderson1969mean}. Let $X$ be a bounded random variable on $[a,b]$, with sorted samples $x_1 \le x_2 \le \dots \le x_T$, $x_0=a$, and $x_{T+1}=b$. Let $g \colon \mathbb{R} \to \mathbb{R}$ be a monotonically increasing function. Anderson's inequality specifies for a confidence level $\delta \in (0, 0.5]$ the following high confidence bounds on $\mathbf{E}[g(X)]$:
\begin{equation}
\begin{aligned}
\mathbf{E}[g(X)] \ge & \ g(x_{T}) &- \sum_{t=0}^{T-1} \left [ g(x_{t+1})-g(x_{t}) \right ] F^+_{X}(x_{t}) \\
\mathbf{E}[g(X)] \le & \ g(x_{T+1}) &- \sum_{t=1}^{T} \left [ g(x_{t+1})-g(x_{t}) \right ] F^-_{X}(x_{t}), \\
\end{aligned}
\end{equation}
where $F^{+/-}_X$ uses the DKW inequality with tight constants and as defined in \eqref{eq:dkw}.
Anderson's inequality can be used to bound the mean normalized performance since $F_{X_{k,j}}$ is a monotonically increasing function and $\delta \in (0,0.5]$. Since $F_{X_{k,j}}$ is unknown, we replace $g$ in Anderson's inequality with $F^-_{X_{k,j}}$ for the lower bound and $F^+_{X_{k,j}}$ for the upper bound. This gives the following confidence intervals for $z_{i,j,k}$:
\begin{equation}
\begin{aligned}
Z^-_{i,j,k} = & \ F^-_{X_{k,j}}(x_{i,j,T}) &- \sum_{t=0}^{T-1} \left [ F^-_{X_{k,j}}(x_{i,j,t+1})-F^-_{X_{k,j}}(x_{i,j,t}) \right ] F^+_{X_{i,j}}(x_{i,j,t}) \\
Z^+_{i,j,k} = & \ F^+_{X_{k,j}}(x_{i,j,T+1}) &- \sum_{t=1}^{T} \left [ F^+_{X_{k,j}}(x_{i,j,t+1})-F^+_{X_{k,j}}(x_{i,j,t}) \right ] F^-_{X_{i,j}}(x_{i,j,t}), \label{eq:normbounds} \\
\end{aligned}
\end{equation}
where $T = T_{i,j}$, $x_{i,j,0} = a_j$, and $x_{i,j,T+1} = b_j$.
We now prove the following lemma:
\begin{lemma}
\label{lem:meannormalized}
If $Z^-$ and $Z^+$ are computed by \eqref{eq:normbounds}, then:
\begin{equation}
\Pr \left (\forall i,k \in \mathcal{A}, \forall j \in \mathcal{M}, \ \ z_{i,j,k} \in [Z^-_{i,j,k}, Z^+_{i,j,k}] \right ) \ge 1-\delta.
\end{equation}
\end{lemma}
\begin{proof}
By Anderson's inequality we know that $Z^+_{i,j,k}$ is an high confidence upper bound on $\mathbf{E}[F^+_{X_{k,j}}(X_{i,j})]$ and similarly $Z^-_{i,j,k}$ is a high confidence lower bound on $\mathbf{E}[F^+_{X_{k,j}}(X_{i,j})]$, i.e.,
\begin{align}
&\ \Pr \left (\mathbf{E}[F^-_{X_{k,j}}(X_{i,j})] \ge Z^-_{i,j,k} \right ) \ge 1-\delta'/2 \\
&\ \Pr \left (\mathbf{E}[F^+_{X_{k,j}}(X_{i,j})] \le Z^+_{i,j,k} \right ) \ge 1-\delta'/2.
\end{align}
By Property \ref{prop:dkw} we know that $\Pr \left ( \forall x \in \mathbb{R}, \ \ F_{X_{k,j}}(x) \in [F_{X_{k,j}}^-(x), F_{X_{k,j}}^+(x)] \right ) \ge 1-\delta'$, thus
\begin{equation}
\Pr \left (\forall i,k \in \mathcal{A}, \forall j \in \mathcal{M}, \ \ z_{i,j,k} \in [Z^-_{i,j,k}, Z^+_{i,j,k}] \right ) \ge 1-2\delta',
\end{equation}
where $2\delta'$ comes from combining the failure rates of confidence intervals on the CDFs $F_{X_{i,j}}$ and $F_{X_{k,j}}$. The confidence intervals on the mean normalized performances can only fail if the confidence intervals on CDFs fail. As stated in \eqref{eq:unioncdf}, all confidence intervals on the CDFs contain the true CDFs with probability at least $1-\delta$. Thus, all mean normalized performances hold with probability at least $1 - \delta$.
\end{proof}
The confidence intervals given by $Z^-$ and $Z^+$ are guaranteed to hold for $T \ge 1$, and $\delta \in (0,0.5]$, but are often conservative requiring a large number samples to identify a statistically meaningful result.
So we empirically test alternatives that have either stricter assumptions or weaker theoretical justification.
\subsection{Confidence intervals on the aggregate performance}
In this section we will provide details on how to compute the confidence intervals on the aggregate performance using the confidence intervals on the mean normalized performance and then prove that they hold with high confidence. To construct confidence intervals and prove their validity we will make the following steps. First, we show that for a fixed weighting $q$ that valid confidence intervals can be computed directly by using interval arithmetic. Second, we describe how to characterize the uncertainty of the game. Third, we make a connection between aggregate performance using the equilibrium solution $q^*$ and the optimal average reward for a Markov decision process. Lastly we describe an optimization procedure for computing the optimal average reward, which corresponds to finding the lower and upper confidence intervals.
Before discussing how to bound the aggregate performance using the game theoretic solution, consider the case when weights $q$ can be any probability distribution over algorithms and environments chosen before the experiment begins.
Let weights $q \in [0,1]^{|\mathcal{A}| \times |\mathcal{M}|}$, such that $\sum_{j \in \mathcal{M}} \sum_{k \in \mathcal{A}} q_{j,k}=1$.
Let $\tilde y_i = \sum_{j \in \mathcal{M}} \sum_{k \in \mathcal{A}} q_{j,k} z_{i,j,k}$ be the aggregate performance for an algorithm $i \in \mathcal{A}$.
The corresponding confidence intervals for $\tilde y_i$ are $\tilde Y^-_i = \sum_{j \in \mathcal{M}} \sum_{k \in \mathcal{A}} q_{j,k} Z^-_{i,j,k}$ and $\tilde Y^+_i = \sum_{j \in \mathcal{M}} \sum_{k \in \mathcal{A}} q_{j,k} Z^+_{i,j,k}$.
\begin{lemma}
\label{lem:aggsum}
If weights $q$ are independent of the data $D$, then:
\begin{equation}
\Pr \left (\forall i \in \mathcal{A}, \ \tilde y_i \in [\tilde Y_i^-, \tilde Y_i^+] \right ) \ge 1 - \delta
\end{equation}
\end{lemma}
\begin{proof}
Applying the result of Lemma \ref{lem:meannormalized}, all confidence intervals produced by $Z^-_{i,j,k}$ and $Z^+_{i,j,k}$ contain $z_{i,j,k}$ with probability $1-\delta$. So interval arithmetic can be used to compute confidence intervals on the aggregate performance without changing the probability of failure, i.e., $\tilde Y^-_i = \sum_{j \in \mathcal{M}} \sum_{k \in \mathcal{A}} q_{j,k} Z^-_{i,j,k}$ and $\tilde Y^+_i = \sum_{j \in \mathcal{M}} \sum_{k \in \mathcal{A}} q_{j,k} Z^+_{i,j,k}$ for each algorithm $i \in \mathcal{A}$.
With these intervals then
\begin{align}
\Pr \left (\forall i \in \mathcal{A}, \ \tilde y_i \in [\tilde Y_i^-, \tilde Y_i^+] \right ) \ge &\ 1 - \Pr \left (\bigcup_{i,k \in \mathcal{A},j \in \mathcal{M}} z_{i,j,k} \notin [Z^-_{i,j,k}, Z^+_{i,j,k}] \right ) \\
\ge &\ 1 - \delta.
\end{align}
\end{proof}
This method of aggregating performance highlights how the uncertainty of normalized scores propagates to the confidence intervals of the aggregate performance for a fixed weighting. Next we will show how to compute confidence intervals on $y$ when using the dynamic weighting produce by the equilibrium solution to the two player game.
Instead of considering all possible equilibrium solutions $q^*$, recall that $y_i$ depends on the Markov matrix $C$ as shown in \eqref{eq:aggalgivalue}. To construct confidence intervals on $y_i$ the uncertainty in creating the matrix $C$ as defined in \eqref{eq:cmat} needs to be considered \citep{rowland2019alphabounds}.
The definition of $C$ assumes certainty of payoffs of each strategy, but the empirical payoffs have uncertainty corresponding to $Z^-$ and $Z^+$, i.e., for $s=(i,(j,k))$, $u^-_p(s)=Z^-_{i,j,k}$, $u^+_p(s)=Z^+_{i,j,k}$, $u^-_q(s)=-Z^+_{i,j,k}$, $u^+_q(s)=-Z^-{i,j,k}$. As a result, when the payoff confidence intervals overlap for two strategies $s$ and $s'$ this creates uncertainty in $C$. We define $C^-,C^+ \in [0,1]^{|\mathcal{S}| \times |\mathcal{S}|}$ as the lower and upper confidence intervals on $C$ with entries
\begin{equation}
\label{eq:cbounds}
\begin{aligned}
C^-_{s,s'}, C^+_{s,s'} \coloneqq &
\begin{cases}
(\eta,\eta) & \text{if $u^-_l(s') > u^+_l(s)$}\\
(0,0) & \text{if $u^-_l(s) > u^+_l(s')$}\\
(\frac{\eta}{n},\frac{\eta}{n}) & \text{if $u^{+/-}_l(s) = u^{+/-}_l(s')$}\\
(0,\eta) & \text{otherwise}
\end{cases} \quad \forall s' \in \mathcal{S} \setminus \{s\} \\
C^-_{s,s}, C^+_{s,s} \coloneqq &\ \left (1 - \sum_{s' \ne s}C^+_{s,s'},\ 1 - \sum_{s' \ne s}C^-_{s,s'} \right )
\end{aligned}
\end{equation}
To get the bounds on the aggregate performance over all possible $C$ in these intervals the uncertainty of the stationary distribution on $\tilde C = \gamma C + (1-\gamma)(1/|\mathcal{S}|)$ has to be consider.
This can be accomplished by first computing the minimum and maximum values of the stationary distribution for each strategy \citep{rowland2019alphabounds} and then finding the minimum and maximum aggregate performances for all possible stationary distributions in this limits.
However, individually computing these two quantities leads to looser bounds than directly estimating the minimum and maximum aggregate performance over all possible $C$ because it ignores the correlations in the confidence intervals of the stationary distribution.
To compute the minimum and maximum aggregate performance over all possible $C$, we need to make a connection between the average performance using the stationary distribution of $\tilde C$ and the average performance before termination on $C$.
Let $d$ be a stationary distribution over strategy profiles $\mathcal{S}$ induce by the Markov matrix $\tilde C = \gamma C + (1-\gamma)(1/|\mathcal{S}|)$.
Let $q$ be the distribution of strategies for player $q$ contained in $d$.
\begin{lemma}
\label{lem:disttovalue}
The following are equivalent
\begin{align}
y_i = &\ \sum_{(j,k) \in \mathcal{S}_2} q_{j,k}z_{i,j,k} \\
y_i = &\ \frac{1-\gamma}{|\mathcal{S}|}\sum_{s \in \mathcal{S}} v(s),
\end{align}
where $v = (I - \gamma C)^{-1}R_i$, $R_i \in \mathbb{R}^{|\mathcal{S}|}$ such that $R_i(s) = z_{i,j,k}$ for $s=(\cdot, (j,k))$.
\end{lemma}
\begin{proof}
We know that $q_{j,k} = \sum_{i \in \mathcal{S}_1} d(i,(j,k))$. This implies $y_i = \sum_{(j,k) \in \mathcal{S}_2} q_{j,k}z_{i,j,k} = \sum_{s \in \mathcal{S}} d(s) R_i(s)$. Applying $8.2.12$ of \citet{puterman1994mdp}, we have the relation
\begin{align}
v + y_i = &\ R_i + \tilde C v \\
v + y_i = &\ R_i + \gamma C v + \frac{(1-\gamma)}{|S|}\mathbf{1}^\top v,
\end{align}
where here $\mathbf{1} \in \mathbb{R}^{|\mathcal{S}|}$ is a vector of all ones. Then for $v = (I-\gamma C)^{-1}R_i$, $y_i=\frac{(1-\gamma)}{|S|}\mathbf{1}^\top v$.
\end{proof}
Finally, using techniques developed by \citet{fercoq2013pagerankopt}, the lower and upper bounds of the aggregate performance $y_i$ can be bound by solving the following optimization problem:
\begin{equation}
\label{eq:lpbound}
\begin{aligned}
\min_{v} \quad & \sum_{s \in \mathcal{S}} v(s) \\
\textrm{s.t.} \quad & v(s) \ge R_i(s) + \gamma \sum_{s'\in \mathcal{S}} C_{s,s'} v(s')\\
\quad & C^- \le C \le C^+ \\
& \sum_{s' \in \mathcal{S}} C_{s,s'} = 1 \quad \forall s \in \mathcal{S},
\end{aligned}
\end{equation}
where $C$ is a free variable, $R_i(s) = -Z^-_{i,j,k}$, and $R_i(s) = Z^+_{i,j,k}$ are used to obtain the lower and upper bounds on $y_i$, respectively. Both bounds are computed as $Y_i^{+/-}=(1-\gamma)(1/|\mathcal{S}|) \sum_{s \in \mathcal{S}}|v(s)|$ using the respective solutions $v$. The absolution value of $v$ is taken to account for the negativity of $R_i = -Z^-$. We compute the solution $v$ in polynomial time using policy iteration based approach similar to \citet{fercoq2013pagerankopt}, but modify their algorithm to fit our matrix $C$. We detail our exact method in Appendix \ref{app:valueiteration}.
Now to prove the main result we will use the previous lemmas in connection with PBP.
\pbpthm*
\begin{proof}
From Lemma \ref{lem:meannormalized} we know that $Z^-$ and $Z^+$ are valid $1-\delta$ confidence intervals on $z$. Thus, applying Lemma \ref{lem:aggsum} we know that a valid $1-\delta$ confidence intervals can be computed by a weighted sum of lower and upper bounds $Z^-$ and $Z^+$, for any joint probability distribution $q$ over environments and algorithms. Through Lemma \ref{lem:disttovalue} this is equivalent to assuming fixed Markov matrix $C$.
The true matrix $C$ is unknown, so the minimum and maximum intervals need to be found over a set $\mathcal{C}$ that contains all transition matrices that are compatible with $Z^-$ and $Z^+$.
Let $\mathcal{C} = \prod_{s \in \mathcal{S}} \mathcal{C}_s$, where $\mathcal{C}_s$ is the polytope of transition probabilities starting from strategy profile $s \in \mathcal{S}$, i.e.,
$$\mathcal{C}_s = \{C_{s,\cdot} \colon \forall {s' \in \mathcal{S}}, \ C_{s,s'} \in [C^-_{s,s'}, C^+_{s,s'}], \sum_{s' \in \mathcal{S}} C_{s,s'}=1 \}.$$
The minimum and maximum aggregate values $Y^-$ and $Y^+$ computed over $\mathcal{C}$ can be found in polynomial time using linear programming, value iteration, or policy iteration \citep{fercoq2013pagerankopt}.
PBP finds the minimum and maximum confidence intervals over all $C \in \mathcal{C}$, thus, they represent valid $1-\delta$ confidence intervals for each algorithm.
\end{proof}
\section{Policy Iteration for Bounding the Aggregate Performance}
\label{app:valueiteration}
This section we detail our approach for optimizing the aggregate performance over the uncertainty of the Markov matrix $C$ to compute confidence intervals.
Recall that the upper and lower high-confidence bounds on the aggregate performance for algorithm $i$ can be found by solving the optimization problem in \ref{eq:lpbound}.
Alternatively one can use a Dynamic Programming approach either using value iteration or policy iteration for more efficient optimization \citep{fercoq2013pagerankopt}.
Using value iteration has better scaling to the size of the state space in the optimization problem than policy iteration.
However, we found that in small and moderate sized problems policy iteration was sufficient and used it in our experiments.
Our method is similar to that of \citet[Algorithm 1]{fercoq2013pagerankopt}, except that we use policy iteration instead of value iteration and a modification to the dynamic programming operator since the transition probabilities in our problem do not depend on the number of out going edges.
Algorithm \ref{alg:policyiteration} shows pseudocode of the policy iteration to find the the optimal value function $v^*$ through the modification of the matrix $C \in \mathcal{C}$.
The algorithm takes as input lower and upper bounds $C^-, C^+ \in \mathbb{R}^{|\mathcal{S}| \times | \mathcal{S}|}$ on $C$, a reward function $R \in \mathbb{R}^{|\mathcal{S}|}$, a discount factor $\gamma \in (0,1)$, a tolerance on the distance to optimal, $tol \in \mathbb{R}$, and the maximum number of iterations $max\_itrs \in \mathbb{N}^+$. We set $\gamma = (1-|\mathcal{S}|)/|\mathcal{S}|$ and $tol=1 \times 10^{-7}$, $max\_itrs = 400$ as default parameters. When finding the upper confidence interval, $y^+_i$, for algorithm $i$, $R$ is a vector such that $R(s) = Z^+_{i,j,k}$ for all $s=(\cdot,(j,k)) \in \mathcal{S}$. Similarly, when finding $y_i^-$, $R$ is such that $R(s) = -Z^-_{i,j,k}$.
\begin{algorithm}[tb]
\caption{Policy Iteration for aggregate performance optimization}
\label{alg:policyiteration}
\begin{algorithmic}
\STATE {\bfseries Input:} $C^-, C^+ \in \mathbb{R}^{|\mathcal{S}|\times |\mathcal{S}|}$, $R \in \mathbb{R}^{|\mathcal{S}|}$, $\gamma \in (0,1) = (1-|\mathcal{S}|)/|\mathcal{S}|$, $tol \in \mathbb{R} = 1 \times 10^{-7}$, $max\_itrs \in \mathbb{N}^+ = 400$
\STATE Initialize: $C \gets C^-$, $v \gets \mathbf{0} \in \mathbb{R}^{|\mathcal{S}|}$
\STATE $changed \gets \texttt{True}$
\STATE $iteration \gets 0$
\WHILE{changed}
\STATE $changed \gets \texttt{False}$
\STATE $iteration \gets iteration + 1$
\IF{$iteration \ge max\_itrs$}
\STATE \texttt{break}
\ENDIF
\FOR{$s \in \mathcal{S}$}\STATE $C_s' \gets \texttt{update\_transition\_row}(C^-_s,C^+_s,R(s),v,\gamma)$
\IF{$||C_s - C_s'||_\infty \ge 1 \times 10^{-8}$}
\STATE $changed \gets \texttt{True}$
\STATE $C_s \gets C_s'$
\ENDIF
\ENDFOR
\STATE $v' \gets (I - \gamma C)^{-1}R$
\STATE $\epsilon \gets ||v-v'||_\infty$ // max change in value function
\STATE $\epsilon_{v^*} \gets (2 \epsilon \gamma) / (1-\gamma)$ // error bound on distance to $v^*$
\STATE $\epsilon_{aggregate} \gets (1-\gamma) \epsilon_{v^*}$ // bound on the maximum error to the aggregate performance
\IF{$\epsilon_{aggregate} < tol$}
\STATE $changed \gets \texttt{False}$
\ENDIF
\STATE $v \gets v'$
\ENDWHILE
\STATE {\bfseries Return:} $(1-\gamma) \texttt{mean}(|v|)$
\end{algorithmic}
\end{algorithm}
Algorithm \ref{alg:policyiteration} has two main steps. First, greedily optimize $C$ with respect to the current value function $v$, i.e., for all $s \in \mathcal{S}$
\begin{equation}
C_s = \argmax_{C_s \in \mathcal{C}_s} R(s) + \gamma C_s v.
\end{equation}
We provide pseudocode for this step in Algorithm \ref{alg:policyupdate}.
The second step is a value function update, which we compute exactly by solving the system of linear equations
$ v = R + \gamma C v$,
by setting $v = (I-\gamma C)^{-1}R$.
These steps repeat until $C$ stops changing or a bound on the maximum absolute error in aggregate performance is below some threshold.
The confidence interval on the aggregate performance is then returned as $(1-\gamma)/|\mathcal{S}| \sum_{s \in \mathcal{S}} |v(s)|$.
Since $C$ is sparse optimization to the code can be made to drastically speed up computation when $\mathcal{S}$ is large. We make some of these modifications in our implementation.
\begin{algorithm}[tb]
\caption{\texttt{update\_transition\_row} procedure for updating $C_s$}
\label{alg:policyupdate}
\begin{algorithmic}
\STATE {\bfseries Input:} $C^-_s, C^+_s \in \mathbb{R}^{|\mathcal{S}|}$, $r \in \mathbb{R}$, $v \in \mathbb{R}^{|\mathcal{S}|}$, $\gamma \in (0,1)$
\STATE Initialize: $C_s \gets C^-_s$, $c \gets \texttt{sum}(C^-_s)$
\STATE $w \gets r + \gamma v$
\STATE $idxs \gets \texttt{argsort}(w, \texttt{direction}=decreasing)$
\FOR{$i=1$ {\bfseries to} $|\mathcal{S}|$}
\STATE $idx \gets idxs[i]$
\STATE $\Delta c \gets \min(C^+_s-C^-_s, 1-c)$
\STATE $C_s[idx] \gets C_s[idx] + \Delta c$
\STATE $c \gets c + \Delta c$
\ENDFOR
\STATE {\bfseries Return:} $C_s$
\end{algorithmic}
\end{algorithm}
Due to small numerical errors this version of policy iteration may not keep the same policy, $C$, between successive iterations when at the optimal solution. To ensure that the procedure stops in a reasonable time and closely approximates the true solution we employ three techniques: an iteration limit, halting computation when $C$ is $\epsilon$-close to between iterations, and stopping when a bound on the distance to true solution is below a tolerance, $tol$.
The first two approaches are straight forward an in the pseudocode.
To bound the distance to the true solution we leverage prior work on bounding the distance of the current value function to the optimal value function.
Consider the value function $v$ and the subsequent value function $v'$ obtained after one application of the Bellman operator.
In our problem the Bellman operator is
\begin{equation}
v(s) = \max_{C_s \in \mathcal{C}_s} R(s) + \gamma \sum_{s' \in \mathcal{S}}C_{s,s'} v(s').
\end{equation}
Let $\epsilon = \max_{s \in \mathcal{S}} |v(s)-v'(s)|$.
Then the distance $\epsilon_{v^*} = \max_{s \in \mathcal{S}}|v'(s)-v^*(s)|$ of $v'$ to the optimal value function $v^*$, is bounded above by $\epsilon$ \citep{williams1993bounds}, i.e,
$$\epsilon_{v^*} \le (2 \epsilon \gamma) / (1-\gamma).$$
We translate this this bound into a bound on the error to the confidence interval of $y_i$.
Let $y_i$ be the confidence interval computed using $v'$ and $y_i^*$ be the confidence interval computed using $v^*$.
Then an upper bound on the error $\epsilon_{aggregate} = |y_i^*-y_i|$ is
\begin{align}
\epsilon_{aggregate} = &\ |y_i^* - y_i| \\
= &\ \left | \frac{1-\gamma}{|\mathcal{S}|} \sum_{s \in \mathcal{S}} |v^*(s)| - \frac{1-\gamma}{|\mathcal{S}|} \sum_{s \in \mathcal{S}} |v'(s)| \right | \\
= &\ \left | \frac{1-\gamma}{|\mathcal{S}|} (||v^*||_1 - ||v'||_1) \right | \\
\le &\ \frac{1-\gamma}{|\mathcal{S}|} ||v^*-v'||_1 \\
= &\ \frac{1-\gamma}{|\mathcal{S}|} |\mathcal{S}| ||v^*-v'||_\infty \\
= &\ (1-\gamma) ||v^*-v'||_\infty \\
\le &\ (1-\gamma) \max_{s \in \mathcal{S}}|v^*(s)-v'(s)| \\
= &\ (1-\gamma) \epsilon_{v^*} = 2 \epsilon \gamma.
\end{align}
\section{Algorithm Definitions}
\label{app:definitions}
This section provides the complete definition for each algorithm used in the experiments.
Each algorithm is made complete by defining a distribution from, which hyperparameters are sampled. Table \ref{tab:hyperparametertable} shows distributions or values for any hyperparameter used in this work. For the continue state space environments, all of the algorithms use Fourier basis and linear function approximation \citep{konidaris2011fourier}. Note that $U(a,b)$ indicates a uniform random variable on $[a,b)$, and $U(\{\ldots \})$ indicates that a variable is sampled uniformly at random from a set of finite values.
Note that these ranges should not be considered optimal and could easily be improved for the environments in this experiment. The ranges were chosen to reflect a lack of knowledge about what optimal settings on an environment are and to be reflective of ranges one might expect optimal hyperparameters to fall in. The manual setting of these ranges leaks information based on our own experience with the algorithms and environments. However, since the definition is completely specified on these environments any favor to one algorithm could be easily identified and test for. Furthermore, any adaptive algorithm that can adjust these parameters during learning is likely to be superior than specifying better ranges through experience in this exact setup. Still, one should not tune ranges to fit any given set of environments.
\begin{table}
\centering
\begin{tabular}{rrcc}
\hline
\multicolumn{1}{c}{\textbf{Algorithm}} & \multicolumn{1}{c}{\textbf{Hyperparameter}} & \textbf{Discrete} & \textbf{Continuous} \\ \hline
\rowcolor[HTML]{EFEFEF}
All & $\lambda$ & \multicolumn{2}{c}{\cellcolor[HTML]{EFEFEF}$U(0,1)$} \\
All & $\gamma$ & \multicolumn{2}{c}{$\Gamma - e^{U(\ln 10^{-4},\ln 0.05)}$} \\
\rowcolor[HTML]{EFEFEF}
All & Value function & Tabular & Linear with Fourier basis \\
All Sarsa$(\lambda)$ and Q$(\lambda)$ & $\epsilon$ & \multicolumn{2}{c}{$U(0,1)$} \\
\rowcolor[HTML]{EFEFEF}
Sarsa$(\lambda)$ and Q$(\lambda)$ & $\alpha_q$ & $e^{U(\ln 10^{-3},\ln 10^{-1})}$ & $e^{U(\ln 10^{-6},\ln 10^{-3})}$ \\
Sarsa$(\lambda)$-s and Q$(\lambda)$-s & $\alpha_q$ & $e^{U(\ln 10^{-3},\ln 10^{-1})}$ & $e^{U(\ln 10^{-3},\ln 10^{0})/|\phi|}$ \\
\rowcolor[HTML]{EFEFEF}
All AC, PPO, and NACTD & Policy & Tabular Softmax & Linear Softmax with Fourier basis \\
AC, NACTD & $\alpha_v$ & $e^{U(\ln 10^{-3},\ln 10^{-1})}$ & $e^{U(\ln 10^{-6},\ln 10^{-3})}$ \\
\rowcolor[HTML]{EFEFEF}
AC, NACTD & $\alpha_p$ & $e^{U(\ln 10^{-3},\ln 10^{-1})}$ & $e^{U(\ln 10^{-6},\ln 10^{-3})}$ \\
AC-S & $\alpha_v$ & $e^{U(\ln 10^{-3},\ln 10^{-1})}$ & $e^{U(\ln 10^{-3},\ln 10^{0})/|\phi|}$ \\
\rowcolor[HTML]{EFEFEF}
AC-S, AC-Parl2 & $\alpha_p$ & $e^{U(\ln 10^{-3},\ln 10^{-1})}$ & $e^{U(\ln 10^{-3},\ln 10^{0})/(|\phi| \times \text{num\_actions)}}$ \\
NAC-TD & $\alpha_w$ & $e^{U(\ln 10^{-3},\ln 10^{-1})}$ & $e^{U(\ln 10^{-6},\ln 10^{-3})}$ \\
\rowcolor[HTML]{EFEFEF}
NAC-TD & normalize\_gradient & \multicolumn{2}{c}{\cellcolor[HTML]{EFEFEF}True} \\
PPO & clip & $U(0.1,0.3)$ & \\
\rowcolor[HTML]{EFEFEF}
PPO & entropy\_coef & \multicolumn{2}{c}{\cellcolor[HTML]{EFEFEF}$e^{U(\ln 10^{-8},\ln 10^{-2})}$} \\
PPO & steps\_per\_batch & \multicolumn{2}{c}{$2^{U(\log_2 64, \log_2 256)}$} \\
\rowcolor[HTML]{EFEFEF}
PPO & epochs & \multicolumn{2}{c}{\cellcolor[HTML]{EFEFEF}$U(\{1,\dotsc,10\})$} \\
PPO & batch\_size & \multicolumn{2}{c}{$2^{U(\log_2 16, \log_2 \min(64, \text{steps\_per\_batch}))}$} \\
\rowcolor[HTML]{EFEFEF}
PPO & Adam-$\epsilon$ & \multicolumn{2}{c}{\cellcolor[HTML]{EFEFEF}$10^{-5}$} \\
PPO & Adam-$\alpha$ & $e^{U(\ln 10^{-3},\ln 10^{-1})}$ & $e^{U(\ln 10^{-6},\ln 10^{-3})}$ \\
\rowcolor[HTML]{EFEFEF}
PPO & Adam-$\beta_1$ & \multicolumn{2}{c}{\cellcolor[HTML]{EFEFEF}$0.9$} \\
PPO & Adam-$\beta_2$ & \multicolumn{2}{c}{$0.999$} \\
\rowcolor[HTML]{EFEFEF}
Fourier basis & dorder & N/A & $U(\{0,\dotsc,9\})$ \\
Fourier basis & iorder & N/A & $U(\{1,\dotsc,9\})$ \\
\rowcolor[HTML]{EFEFEF}
Fourier basis & trig & N/A & $\texttt{cos}$
\end{tabular}
\caption{This table show the distributions from which each hyperparameter is sampled. The All algorithm means the hyperparameter and distribution were used for all algorithms. Steps sizes are labeled with various $\alpha$s. The discount factor $\gamma$ an algorithm uses is scaled down from $\Gamma$ that is specified by the environment. For all environments used in this work $\Gamma = 1.0$. PPO uses the same learning rate for both the policy and value function. The max dependent order on the Fourier basis is limited such that no more than $10,\!000$ features are generated as a result of dorder. }
\label{tab:hyperparametertable}
\end{table}
\section{Environments}
\label{app:environments}
This section describes the environments used in the experiments.
All environments are listed in Table \ref{tab:envtable}.
Environments were recreated in the Julia language and many implementations follow closely to the that in the RLPy repository \citep{geramifard2015rlpy}.
Most environments are best described by either the paper publishing the environment or by examining the source code we provide. We also describe the discrete environments below.
There are eight discrete domains used in the work four Gridworld environments and four chain environments.
The Gridworld environments are an $N \times N$ grid with the agent starting every episode in the top left corner and goal state in the bottom right. The reward is $-1$ at every step until the goal state is reached, then the episode is then terminated. In every state there are four actions: up, down, left, and right. The transition dynamics are either deterministic, meaning an up action sends the agent up one state unless it is outside the map, or the dynamics are stochastic, meaning the agent might randomly move to one of the states perpendicular to the intended direction or stays in the current state.
The chain environments are $N$ chains where there are $N$ states each with a connection only to the state directly to the left or right of it and end points only connect to the one state they are next to. The agent starts in state one (far left of the chain) and the goal state is state $N$ (far right of the chain). The reward is $-1$ every step until the goal state is reached and then the episode terminates. Both gridworld and chain MDPs terminate episodes in a finite time based on the size of the problem. Gridworld problems terminate after $20N^2$ steps have been taken and chain environments terminate after $20N$ steps are taken.
\begin{table}
\centering
\begin{tabular}{lcc}
\hline
\multicolumn{1}{c}{\textbf{Environment}} & \textbf{Num Episodes} & \textbf{State Space} \\ \hline
Gridworld 5 Deterministic & 100 & Discrete \\
Gridworld 5 Stochastic & 100 & Discrete \\
Gridworld 10 Deterministic & 100 & Discrete \\
Gridworld 10 Stochastic & 100 & Discrete \\
Chain 10 Deterministic & 100 & Discrete \\
Chain 10 Stochastic & 100 & Discrete \\
Chain 50 Deterministic & 100 & Discrete \\
Chain 50 Stochastic & 100 & Discrete \\
Acrobot & 500 & Continuous \\
Cart-Pole & 100 & Continuous \\
MountainCar & 100 & Continuous \\
PinBall Empty & 100 & Continuous \\
PinBall Box & 100 & Continuous \\
Pinball Medium & 100 & Continuous \\
PinBall Single & 200 & Continuous
\end{tabular}
\caption{This table list every used in this paper along with the number of episodes each algorithm was allowed to interact with the environment and its type of state space. }
\label{tab:envtable}
\end{table}
\section{Alternative Bounding Techniques}
\label{app:altbounds}
As pointed out in our description of PBP, we use the nonparametric concentration inequalities DKW and Anderson's inequalities. These inequalities are often conservative and lead to conservative confidence intervals. So we investigate two alternatives PBP-t a method that replaces DKW and Anderson's inequality with the one based on Students t-distribution and the percentile bootstrap.
In PBP-t everything is the same as PBP except we compute confidence intervals on the mean normalized performance as follows
\begin{equation}
\begin{aligned}
Z^-_{i,j,k} = & \ \mu_{i,j,k} - \frac{\hat \sigma}{\sqrt{T_{i,j}}} t_{1-\delta',T_{i,j}-1}\\
Z^+_{i,j,k} = & \ \mu_{i,j,k} + \frac{\hat \sigma}{\sqrt{T_{i,j}}} t_{1-\delta',T_{i,j}-1}\\
\end{aligned}
\end{equation}
where $z_{i,j,k,t} = \hat F_{X_{k,j}}(x_{i,j,t})$, $\mu_{i,j,k} = \frac{1}{T_{i,j}} \sum_{t=1}^{T_{i,j}} z_{i,j,k,t}$, $\hat \sigma = \sqrt{\sum_{t=1}^{T_{i,j}} (\mu_{i,j,k} - z_{i,j,k,t})^2 / (T_{i,j}-1)}$, $t_{1-\delta', \nu}$ is the $100(1-\delta')$ percentile of Student's t-distribution with $\nu$ degrees of freedom, and we set $\delta' = \delta / (|\mathcal{A}||\mathcal{M}|)$. Notice that there are $|\mathcal{A}|^2|\mathcal{M}|$ comparisons being made, so $\delta' = \delta / (|\mathcal{A}|^2|\mathcal{M}|)$ should be used to account for more the multiple comparisons. However, it is likely that if one comparison with an algorithm $i$ fails then that there will be failures with the other $|\mathcal{A}|-1$ algorithms so we use the smaller $\delta'$ as a heuristic for tighter confidence intervals.
In the bootstrap procedure, we use the percentile bootstrap with a confidence level of $\delta' = \delta/(|\mathcal{A}||\mathcal{M}|)$ and $10,\!000$ bootstrap samples of the aggregate performance. Each bootstrap is formed by re-sampling the performance of each algorithm on each environment for the collected data. Then the aggregate performance for each bootstrap is computed. The lower and upper confidence intervals are given by the $100(\delta'/2)$ and $100(1-\delta'/2)$ percentile from the bootstrap aggregate performance.
Since $\delta'/2$ is often really small, $10,\!000$ bootstrap samples are needed to get confidence intervals that more accurately reflect the true confidence intervals. This requires a substantial amount of compute time and can take over four hours for an original sample size of $10,\!000$.
\section{Confidence interval test experiment}
\label{app:boundsexperiment}
To test the different bounding techniques, we estimate the failure rate of each confidence interval technique at different sample sizes. For this experiment we execute $1,\!000$ trials of the evaluation procedure using samples sizes (trials per algorithm per environment) of $10$, $30$, $100$, $1,\!000$, and $10,\!000$.
There are a total of $11.14$ million samples per algorithm per environment. To reduce computation costs, we limit this experiment to only include the Sarsa-Parl2, Q-Parl2, AC-Parl2, and Sarsa($\lambda$)-s. Additionally, we reduce the environment set to be the discrete environments and mountain car. Then we compute the proportion of violations for any confidence interval. All methods use the same data, so the results are not independent. We choose not to run independent samples of the bounds to limit our environmental impact.
The method to compute of the proportion of violations and number of significant pairwise comparison can be found in the source code.
It it important to note that when this experiment was run, there was a bug in the code that made the step-size for the Parl2 methods smaller by a factor of $0.1$. This does not invalidate the results, only that the algorithms run are not equivalent those used in the other experiments in this paper. The main impact of this difference was that Sarsa-Parl2 and Q-Parl2 did were not as effective on the discrete MDPs (though they diverged even less) and their scores were nearly the same. Since both of these algorithms had near identical scores on most of the environments, it became almost impossible to differentiate them, so detecting five out of six ($83\%$) statistically significant comparisons is considered optimal for this experiment.
\section{Performances}
\label{app:indperf}
This section illustrates the distribution of performance of each algorithm. Tables of the average performance (rounded to the tenths place) along with the algorithm rank on that environment. In the Figures showing the performance distributions the shaded regions represent $100(1-0.05/|\mathcal{A}||\mathcal{M}|)\% \approx 99.9697\%$ confidence intervals computed via DKW.
In the performance tables $100(1-0.05/|\mathcal{A}||\mathcal{M}|)\%$ confidence intervals are shown in parentheses.
\newpage
\begin{figure}[ht!]
\centering
\includegraphics[width=0.6\textwidth]{images/group/Classic_Control_Environments.png}
\label{fig:classic_dist}
\end{figure}
\begin{figure}[hb!]
\centering
\includegraphics[width=0.6\textwidth]{images/group/PinBall_Environments.png}
\label{fig:pinball_dist}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[width=0.6\textwidth]{images/group/Gridworld_Environments.png}
\label{fig:gridworld_dist}
\end{figure}
\begin{figure}[hb!]
\centering
\includegraphics[width=0.6\textwidth]{images/group/Chain_Environments.png}
\label{fig:chain_dist}
\end{figure}
\newpage
\begin{table}[h!]
\centering
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{\textbf{Acrobot}} \\
\midrule
Algorithm & Mean & Rank \\
\midrule
\rowcolor[HTML]{EFEFEF}
Sarsa-Parl2 & -20.6 (-24.6, -18.1) & 1 (2, 1) \\
Q-Parl2 & -25.7 (-29.8, -22.8) & 2 (6, 1) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$)-s & -28.3 (-32.3, -26.3) & 3 (6, 2) \\
Q($\lambda$)-s & -30.3 (-34.3, -27.9) & 4 (9, 2) \\
\rowcolor[HTML]{EFEFEF}
PPO & -30.5 (-34.5, -26.7) & 5 (9, 2) \\
AC-parl2 & -30.6 (-34.6, -28.8) & 6 (9, 2) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$) & -34.7 (-38.7, -32.4) & 7 (10, 4) \\
Q($\lambda$) & -35.7 (-39.7, -33.3) & 8 (10, 4) \\
\rowcolor[HTML]{EFEFEF}
AC-s & -36.0 (-40.0, -33.5) & 9 (10, 4) \\
AC & -41.4 (-45.4, -37.3) & 10 (11, 7) \\
\rowcolor[HTML]{EFEFEF}
NAC-TD & -47.1 (-51.1, -43.0) & 11 (11, 10) \\
\bottomrule
\end{tabular}
\label{tab:perf_acrobot}
\centering
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{\textbf{Cart-Pole}} \\
\midrule
Algorithm & Mean & Rank \\
\midrule
\rowcolor[HTML]{EFEFEF}
Sarsa-Parl2 & 469.2 (448.3, 490.0) & 1 (2, 1) \\
Q-Parl2 & 450.3 (429.6, 471.1) & 2 (2, 1) \\
\rowcolor[HTML]{EFEFEF}
AC-parl2 & 390.4 (369.7, 411.2) & 3 (3, 3) \\
Sarsa($\lambda$)-s & 347.5 (326.5, 368.4) & 4 (7, 4) \\
\rowcolor[HTML]{EFEFEF}
Q($\lambda$)-s & 345.5 (324.5, 366.3) & 5 (7, 4) \\
AC & 338.7 (317.8, 359.6) & 6 (7, 4) \\
\rowcolor[HTML]{EFEFEF}
AC-s & 320.6 (300.0, 341.5) & 7 (9, 4) \\
Q($\lambda$) & 291.0 (270.1, 311.9) & 8 (10, 7) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$) & 287.2 (265.9, 308.6) & 9 (10, 7) \\
PPO & 276.3 (256.1, 297.1) & 10 (11, 8) \\
\rowcolor[HTML]{EFEFEF}
NAC-TD & 244.7 (224.0, 265.5) & 11 (11, 10) \\
\bottomrule
\end{tabular}
\label{tab:perf_cartpole}
\end{table}
\begin{table}[h!]
\centering
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{\textbf{Mountain Car}} \\
\midrule
Algorithm & Mean & Rank \\
\midrule
\rowcolor[HTML]{EFEFEF}
AC-parl2 & -791.6 (-894.0, -688.7) & 1 (3, 1) \\
Sarsa-Parl2 & -843.2 (-945.0, -740.5) & 2 (4, 1) \\
\rowcolor[HTML]{EFEFEF}
AC-s & -966.0 (-1068.0, -863.1) & 3 (4, 1) \\
Sarsa($\lambda$)-s & -1024.2 (-1125.8, -923.5) & 4 (5, 2) \\
\rowcolor[HTML]{EFEFEF}
Q($\lambda$)-s & -1197.3 (-1298.9, -1094.9) & 5 (6, 4) \\
Q-Parl2 & -1269.3 (-1371.2, -1166.4) & 6 (6, 5) \\
\rowcolor[HTML]{EFEFEF}
AC & -2477.1 (-2576.0, -2374.3) & 7 (8, 7) \\
NAC-TD & -2489.3 (-2589.1, -2386.4) & 8 (8, 7) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$) & -2769.1 (-2867.8, -2666.2) & 9 (10, 9) \\
Q($\lambda$) & -2779.3 (-2877.9, -2676.4) & 10 (10, 9) \\
\rowcolor[HTML]{EFEFEF}
PPO & -3026.3 (-3124.9, -2923.5) & 11 (11, 11) \\
\bottomrule
\end{tabular}
\label{tab:perf_mntcar}
\end{table}
\begin{table}[h!]
\centering
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{\textbf{Chain 10 Deterministic}} \\
\midrule
Algorithm & Mean & Rank \\
\midrule
\rowcolor[HTML]{EFEFEF}
Q-Parl2 & -9.2 (-13.2, -9.1) & 1 (2, 1) \\
Sarsa-Parl2 & -9.3 (-13.3, -9.2) & 2 (2, 1) \\
\rowcolor[HTML]{EFEFEF}
AC-parl2 & -19.0 (-23.0, -17.7) & 3 (3, 3) \\
Sarsa($\lambda$) & -27.9 (-31.8, -26.9) & 4 (9, 4) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$)-s & -27.9 (-31.8, -26.9) & 4 (9, 4) \\
Q($\lambda$) & -27.9 (-31.9, -26.9) & 6 (9, 4) \\
\rowcolor[HTML]{EFEFEF}
Q($\lambda$)-s & -27.9 (-31.9, -26.9) & 6 (9, 4) \\
AC & -32.2 (-36.2, -29.9) & 8 (9, 4) \\
\rowcolor[HTML]{EFEFEF}
AC-s & -32.2 (-36.2, -29.9) & 8 (9, 4) \\
PPO & -38.0 (-41.9, -36.5) & 10 (10, 10) \\
\rowcolor[HTML]{EFEFEF}
NAC-TD & -89.7 (-93.7, -85.7) & 11 (11, 11) \\
\bottomrule
\end{tabular}
\centering
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{\textbf{Chain 10 Stochastic}} \\
\midrule
Algorithm & Mean & Rank \\
\midrule
\rowcolor[HTML]{EFEFEF}
Q-Parl2 & -12.7 (-16.7, -12.2) & 1 (2, 1) \\
Sarsa-Parl2 & -12.8 (-16.8, -12.4) & 2 (2, 1) \\
\rowcolor[HTML]{EFEFEF}
AC-parl2 & -22.6 (-26.6, -21.2) & 3 (3, 3) \\
Q($\lambda$) & -34.9 (-38.9, -33.6) & 4 (9, 4) \\
\rowcolor[HTML]{EFEFEF}
Q($\lambda$)-s & -34.9 (-38.9, -33.6) & 4 (9, 4) \\
Sarsa($\lambda$) & -35.1 (-39.0, -33.7) & 6 (9, 4) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$)-s & -35.1 (-39.0, -33.7) & 6 (9, 4) \\
AC & -36.9 (-40.8, -34.7) & 8 (9, 4) \\
\rowcolor[HTML]{EFEFEF}
AC-s & -36.9 (-40.8, -34.7) & 8 (9, 4) \\
PPO & -43.8 (-47.6, -42.1) & 10 (10, 10) \\
\rowcolor[HTML]{EFEFEF}
NAC-TD & -75.2 (-79.1, -71.2) & 11 (11, 11) \\
\bottomrule
\end{tabular}
\label{tab:perf_chain10s}
\end{table}
\begin{table}[h!]
\centering
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{\textbf{Chain 50 Deterministic}} \\
\midrule
Algorithm & Mean & Rank \\
\midrule
\rowcolor[HTML]{EFEFEF}
Sarsa-Parl2 & -78.3 (-97.9, -71.5) & 1 (2, 1) \\
Q-Parl2 & -92.7 (-112.3, -74.6) & 2 (2, 1) \\
\rowcolor[HTML]{EFEFEF}
AC-parl2 & -231.0 (-250.5, -211.3) & 3 (3, 3) \\
AC & -425.5 (-444.1, -405.8) & 4 (6, 4) \\
\rowcolor[HTML]{EFEFEF}
AC-s & -425.5 (-444.1, -405.8) & 4 (6, 4) \\
PPO & -462.2 (-480.4, -442.6) & 6 (10, 4) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$) & -479.5 (-498.3, -462.4) & 7 (10, 6) \\
Sarsa($\lambda$)-s & -479.5 (-498.3, -462.4) & 7 (10, 6) \\
\rowcolor[HTML]{EFEFEF}
Q($\lambda$) & -481.6 (-500.4, -464.0) & 9 (10, 6) \\
Q($\lambda$)-s & -481.6 (-500.4, -464.0) & 9 (10, 6) \\
\rowcolor[HTML]{EFEFEF}
NAC-TD & -928.4 (-946.8, -908.7) & 11 (11, 11) \\
\bottomrule
\end{tabular}
\centering
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{\textbf{Chain 50 Stochastic}} \\
\midrule
Algorithm & Mean & Rank \\
\midrule
\rowcolor[HTML]{EFEFEF}
Sarsa-Parl2 & -101.6 (-121.1, -96.4) & 1 (2, 1) \\
Q-Parl2 & -111.1 (-130.6, -97.5) & 2 (2, 1) \\
\rowcolor[HTML]{EFEFEF}
AC-parl2 & -239.3 (-258.7, -219.7) & 3 (3, 3) \\
AC & -463.2 (-481.5, -443.6) & 4 (6, 4) \\
\rowcolor[HTML]{EFEFEF}
AC-s & -463.2 (-481.5, -443.6) & 4 (6, 4) \\
PPO & -474.4 (-492.2, -455.3) & 6 (6, 4) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$) & -546.9 (-565.2, -529.2) & 7 (10, 7) \\
Sarsa($\lambda$)-s & -546.9 (-565.2, -529.2) & 7 (10, 7) \\
\rowcolor[HTML]{EFEFEF}
Q($\lambda$) & -550.3 (-568.7, -532.5) & 9 (10, 7) \\
Q($\lambda$)-s & -550.3 (-568.7, -532.5) & 9 (10, 7) \\
\rowcolor[HTML]{EFEFEF}
NAC-TD & -897.4 (-915.3, -877.8) & 11 (11, 11) \\
\bottomrule
\end{tabular}
\label{tab:perf_chain50s}
\end{table}
\begin{table}[h!]
\centering
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{\textbf{Gridworld 10 Deterministic}} \\
\midrule
Algorithm & Mean & Rank \\
\midrule
\rowcolor[HTML]{EFEFEF}
Sarsa-Parl2 & -49.5 (-90.7, -48.9) & 1 (2, 1) \\
Q-Parl2 & -60.8 (-102.0, -48.5) & 2 (2, 1) \\
\rowcolor[HTML]{EFEFEF}
AC-parl2 & -130.6 (-171.6, -115.1) & 3 (3, 3) \\
PPO & -199.6 (-240.4, -188.8) & 4 (4, 4) \\
\rowcolor[HTML]{EFEFEF}
Q($\lambda$) & -290.6 (-331.0, -278.9) & 5 (10, 5) \\
Q($\lambda$)-s & -290.6 (-331.0, -278.9) & 5 (10, 5) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$) & -291.8 (-332.3, -281.3) & 7 (10, 5) \\
Sarsa($\lambda$)-s & -291.8 (-332.3, -281.3) & 7 (10, 5) \\
\rowcolor[HTML]{EFEFEF}
AC & -324.7 (-365.2, -295.3) & 9 (10, 5) \\
AC-s & -324.7 (-365.2, -295.3) & 9 (10, 5) \\
\rowcolor[HTML]{EFEFEF}
NAC-TD & -1141.3 (-1181.6, -1099.9) & 11 (11, 11) \\
\bottomrule
\end{tabular}
\centering
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{\textbf{Gridworld 10 Stochastic}} \\
\midrule
Algorithm & Mean & Rank \\
\midrule
\rowcolor[HTML]{EFEFEF}
Sarsa-Parl2 & -67.5 (-108.6, -66.1) & 1 (2, 1) \\
Q-Parl2 & -75.4 (-116.4, -64.8) & 2 (2, 1) \\
\rowcolor[HTML]{EFEFEF}
AC-parl2 & -146.4 (-187.3, -135.8) & 3 (3, 3) \\
PPO & -229.5 (-270.1, -217.6) & 4 (4, 4) \\
\rowcolor[HTML]{EFEFEF}
Q($\lambda$) & -339.2 (-379.2, -326.9) & 5 (10, 5) \\
Q($\lambda$)-s & -339.2 (-379.2, -326.9) & 5 (10, 5) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$) & -342.3 (-382.4, -330.0) & 7 (10, 5) \\
Sarsa($\lambda$)-s & -342.3 (-382.4, -330.0) & 7 (10, 5) \\
\rowcolor[HTML]{EFEFEF}
AC & -347.2 (-387.4, -329.7) & 9 (10, 5) \\
AC-s & -347.2 (-387.4, -329.7) & 9 (10, 5) \\
\rowcolor[HTML]{EFEFEF}
NAC-TD & -864.1 (-904.2, -823.3) & 11 (11, 11) \\
\bottomrule
\end{tabular}
\label{tab:perf_gw10s}
\end{table}
\begin{table}[h!]
\centering
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{\textbf{Gridworld 5 Deterministic}} \\
\midrule
Algorithm & Mean & Rank \\
\midrule
\rowcolor[HTML]{EFEFEF}
Sarsa-Parl2 & -12.2 (-22.5, -12.1) & 1 (2, 1) \\
Q-Parl2 & -12.5 (-22.7, -12.0) & 2 (2, 1) \\
\rowcolor[HTML]{EFEFEF}
AC-parl2 & -32.2 (-42.4, -30.1) & 3 (3, 3) \\
PPO & -51.1 (-61.2, -49.0) & 4 (10, 4) \\
\rowcolor[HTML]{EFEFEF}
Q($\lambda$) & -51.4 (-61.5, -49.6) & 5 (10, 4) \\
Q($\lambda$)-s & -51.4 (-61.5, -49.6) & 5 (10, 4) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$) & -51.9 (-62.0, -50.1) & 7 (10, 4) \\
Sarsa($\lambda$)-s & -51.9 (-62.0, -50.1) & 7 (10, 4) \\
\rowcolor[HTML]{EFEFEF}
AC & -62.0 (-72.1, -58.2) & 9 (10, 4) \\
AC-s & -62.0 (-72.1, -58.2) & 9 (10, 4) \\
\rowcolor[HTML]{EFEFEF}
NAC-TD & -168.1 (-178.3, -157.9) & 11 (11, 11) \\
\bottomrule
\end{tabular}
\centering
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{\textbf{Gridworld 5 Stochastic}} \\
\midrule
Algorithm & Mean & Rank \\
\midrule
\rowcolor[HTML]{EFEFEF}
Sarsa-Parl2 & -17.0 (-27.2, -16.6) & 1 (2, 1) \\
Q-Parl2 & -17.1 (-27.4, -16.3) & 2 (2, 1) \\
\rowcolor[HTML]{EFEFEF}
AC-parl2 & -37.2 (-47.3, -35.4) & 3 (3, 3) \\
PPO & -57.2 (-67.3, -54.9) & 4 (10, 4) \\
\rowcolor[HTML]{EFEFEF}
Q($\lambda$) & -61.5 (-71.6, -59.4) & 5 (10, 4) \\
Q($\lambda$)-s & -61.5 (-71.6, -59.4) & 5 (10, 4) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$) & -61.5 (-71.6, -59.4) & 7 (10, 4) \\
Sarsa($\lambda$)-s & -61.5 (-71.6, -59.4) & 7 (10, 4) \\
\rowcolor[HTML]{EFEFEF}
AC & -67.6 (-77.6, -64.8) & 9 (10, 4) \\
AC-s & -67.6 (-77.6, -64.8) & 9 (10, 4) \\
\rowcolor[HTML]{EFEFEF}
NAC-TD & -125.4 (-135.5, -115.4) & 11 (11, 11) \\
\bottomrule
\end{tabular}
\label{tab:perf_gw5s}
\end{table}
\begin{table}[h!]
\centering
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{\textbf{Pinball Box}} \\
\midrule
Algorithm & Mean & Rank \\
\midrule
\rowcolor[HTML]{EFEFEF}
Sarsa-Parl2 & 8823.2 (8513.4, 8998.4) & 1 (1, 1) \\
AC-parl2 & 7875.4 (7565.9, 8170.2) & 2 (2, 2) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$)-s & 6288.3 (5980.0, 6569.4) & 3 (4, 3) \\
AC-s & 5961.2 (5653.1, 6251.1) & 4 (7, 3) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$) & 5603.0 (5295.1, 5889.7) & 5 (8, 4) \\
AC & 5602.5 (5295.1, 5886.8) & 6 (8, 4) \\
\rowcolor[HTML]{EFEFEF}
NAC-TD & 5546.8 (5243.2, 5838.0) & 7 (8, 4) \\
PPO & 5184.6 (4882.6, 5447.4) & 8 (9, 5) \\
\rowcolor[HTML]{EFEFEF}
Q($\lambda$)-s & 4728.6 (4421.8, 5020.1) & 9 (11, 8) \\
Q($\lambda$) & 4449.9 (4143.3, 4740.8) & 10 (11, 9) \\
\rowcolor[HTML]{EFEFEF}
Q-Parl2 & 4246.5 (3937.2, 4539.9) & 11 (11, 9) \\
\bottomrule
\end{tabular}
\centering
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{\textbf{Pinball Empty}} \\
\midrule
Algorithm & Mean & Rank \\
\midrule
\rowcolor[HTML]{EFEFEF}
Sarsa-Parl2 & 8942.1 (8631.8, 9115.3) & 1 (1, 1) \\
AC-parl2 & 8041.0 (7730.8, 8333.4) & 2 (2, 2) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$)-s & 6382.4 (6073.4, 6664.1) & 3 (5, 3) \\
AC-s & 6155.8 (5846.5, 6444.9) & 4 (7, 3) \\
\rowcolor[HTML]{EFEFEF}
AC & 5814.2 (5505.4, 6097.3) & 5 (8, 3) \\
Sarsa($\lambda$) & 5778.6 (5469.8, 6063.7) & 6 (8, 4) \\
\rowcolor[HTML]{EFEFEF}
NAC-TD & 5710.1 (5405.2, 5998.0) & 7 (8, 4) \\
PPO & 5401.3 (5097.3, 5666.6) & 8 (9, 5) \\
\rowcolor[HTML]{EFEFEF}
Q($\lambda$)-s & 4891.8 (4584.1, 5182.5) & 9 (11, 8) \\
Q($\lambda$) & 4602.3 (4294.8, 4892.1) & 10 (11, 9) \\
\rowcolor[HTML]{EFEFEF}
Q-Parl2 & 4487.9 (4178.3, 4781.5) & 11 (11, 9) \\
\bottomrule
\end{tabular}
\label{tab:perf_pbempt}
\end{table}
\begin{table}[h!]
\centering
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{\textbf{Pinball Medium}} \\
\midrule
Algorithm & Mean & Rank \\
\midrule
\rowcolor[HTML]{EFEFEF}
Sarsa-Parl2 & 8402.1 (8094.2, 8625.5) & 1 (1, 1) \\
AC-parl2 & 7128.1 (6820.6, 7429.2) & 2 (2, 2) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$)-s & 5667.4 (5361.4, 5949.4) & 3 (4, 3) \\
AC-s & 5195.5 (4890.3, 5487.2) & 4 (7, 3) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$) & 5050.5 (4745.1, 5336.5) & 5 (7, 4) \\
AC & 4835.7 (4531.8, 5123.8) & 6 (7, 4) \\
\rowcolor[HTML]{EFEFEF}
NAC-TD & 4652.5 (4353.9, 4940.2) & 7 (9, 4) \\
PPO & 4227.1 (3932.0, 4493.3) & 8 (10, 7) \\
\rowcolor[HTML]{EFEFEF}
Q($\lambda$)-s & 4084.7 (3780.6, 4375.1) & 9 (10, 7) \\
Q($\lambda$) & 3689.6 (3386.2, 3981.9) & 10 (11, 8) \\
\rowcolor[HTML]{EFEFEF}
Q-Parl2 & 3404.5 (3097.9, 3699.2) & 11 (11, 10) \\
\bottomrule
\end{tabular}
\centering
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{\textbf{Pinball Single}} \\
\midrule
Algorithm & Mean & Rank \\
\midrule
\rowcolor[HTML]{EFEFEF}
Sarsa-Parl2 & 3754.6 (3470.3, 4023.3) & 1 (1, 1) \\
AC-parl2 & 1696.7 (1418.4, 1989.3) & 2 (2, 2) \\
\rowcolor[HTML]{EFEFEF}
Sarsa($\lambda$)-s & 514.1 (240.8, 793.2) & 3 (5, 3) \\
Sarsa($\lambda$) & 119.2 (-151.7, 408.1) & 4 (7, 3) \\
\rowcolor[HTML]{EFEFEF}
AC-s & 103.8 (-158.5, 389.6) & 5 (7, 3) \\
Q-Parl2 & -73.4 (-352.7, 214.7) & 6 (7, 4) \\
\rowcolor[HTML]{EFEFEF}
AC & -197.1 (-451.4, 95.1) & 7 (7, 4) \\
Q($\lambda$)-s & -746.6 (-1005.9, -457.2) & 8 (10, 8) \\
\rowcolor[HTML]{EFEFEF}
Q($\lambda$) & -986.1 (-1244.4, -693.8) & 9 (10, 8) \\
NAC-TD & -1229.4 (-1463.7, -934.7) & 10 (11, 8) \\
\rowcolor[HTML]{EFEFEF}
PPO & -1602.2 (-1789.1, -1327.7) & 11 (11, 10) \\
\bottomrule
\end{tabular}
\label{tab:perf_pbsingle}
\end{table}
\section{Introduction}
\label{sec:intro}
When applying reinforcement learning (RL), particularly to real-world applications, it is desirable to have algorithms that reliably achieve high levels of performance without requiring expert knowledge or significant human intervention.
For researchers, having algorithms of this type would mean spending less time tuning algorithms to solve benchmark tasks and more time developing solutions to harder problems.
Current evaluation practices do not properly account for the uncertainty in the results \citep{henderson2018deep} and neglect the difficulty of applying RL algorithms to a given problem.
Consequently, existing RL algorithms are difficult to apply to real-world applications \citep{arnold2019realworld}.
To both make and track progress towards developing reliable and easy-to-use algorithms, we propose a principled evaluation procedure that quantifies the difficulty of using an algorithm.
For an evaluation procedure to be useful for measuring the usability of RL algorithms, we suggest that it should have four properties.
First, to ensure accuracy and reliability, an evaluation procedure should be \textit{scientific}, such that it provides information to answer a research question, tests a specific hypothesis, and quantifies any uncertainty in the results.
Second, the performance metric captures the \textit{usability} of the algorithm over a wide variety of environments. For a performance metric to capture the usability of an algorithm, it should include the time and effort spent tuning the algorithm's hyperparameters (e.g., step-size and policy structure).
Third, the evaluation procedure should be \textit{nonexploitative} \citep{balduzzi2018eval},
meaning no algorithm should be favored by performing well on an over-represented subset of environments or by abusing a particular score normalization method.
Fourth, an evaluation procedure should be \textit{computationally tractable}, meaning that a typical researcher should be able to run the procedure and repeat experiments found in the literature.
As an evaluation procedure requires a question to answer, we pose the following to use throughout the paper:
\textit{which algorithm(s) perform well across a wide variety of environments with little or no environment-specific tuning}? Throughout this work, we refer to this question as the \textit{general evaluation question}.
This question is different from the one commonly asked in articles proposing a new algorithm, e.g., the \textit{common question} is, can algorithm X outperform other algorithms on tasks A, B, and C?
In contrast to the common question, the expected outcome for the general evaluation question is not to find methods that maximize performance with optimal hyperparameters but to identify algorithms that do not require extensive hyperparameter tuning and thus are easy to apply to new problems.
In this paper, we contend that the standard evaluation approaches do not satisfy the above properties, and are not able to answer the general evaluation question.
Thus, we develop a new procedure for evaluating RL algorithms that overcomes these limitations and can accurately quantify the uncertainty of performance.
The main ideas in our approach are as follows.
We present an alternative view of an algorithm such that sampling its performance can be used to answer the general evaluation question.
We define a new normalized performance measure, \textit{performance percentiles}, which uses a relative measure of performance to compare algorithms across environments.
We show how to use a game-theoretic approach to construct an aggregate measure of performance that permits quantifying uncertainty.
Lastly, we develop a technique, \textit{performance bound propagation} (PBP), to quantify and account for uncertainty throughout the entire evaluation procedure.
We provide source code so others may easily apply the methods we develop here.\footnote{Source code for this paper can be found at \url{https://github.com/ScottJordan/EvaluationOfRLAlgs}.}
\section{Notation and Preliminaries}
In this section, we give notation used in this paper along with an overview of an evaluation procedure.
In addition to this section, a list of symbols used in this paper is presented in Appendix \ref{app:confagg}.
We represent a performance metric of an algorithm, $i \in \mathcal{A}$, on an environment, $j \in \mathcal{M}$, as a random variable $X_{i,j}$. This representation captures the variability of results due to the choice of the random seed controlling the stochastic processes in the algorithm and the environment.
The choice of the metric depends on the property being studied and is up to the experiment designer. The performance metric used in this paper is the average of the observed returns from one execution of the entire training procedure, which we refer to as the average return.
The cumulative distribution function (CDF), $F_{X_{i,j}} \colon \mathbb{R} \to [0,1]$, describes the performance distribution of algorithm $i$ on environment $j$ such that $F_{X_{i,j}}(x) \coloneqq \Pr(X_{i,j} \le x)$.
The quantile function, $Q_{X_{i,j}}(\alpha) \coloneqq \inf_{x} \{ x \in \mathbb{R} | F_{X_{i,j}}(x) \ge \alpha \}$, maps a cumulative probability, $\alpha \in (0,1)$, to a score such that $\alpha$ proportion of samples of $X_{i,j}$ are less than or equal to $Q_{X_{i,j}}(\alpha)$.
A normalization function, $g \colon \mathbb{R} \times \mathcal{M} \to \mathbb{R}$, maps a score, $x$, an algorithm receives on an environment, $j$, to a normalized score, $g(x,j)$, which has a common scale for all environments.
In this work, we seek an aggregate performance measure, $y_i \in \mathbb{R}$, for an algorithm, $i$, such that $y_i \coloneqq \sum_{j=1}^{|\mathcal{M}|} q_j \mathbf{E}[g(X_{i,j},j)]$, where $q_j \ge 0$ for all $j \in \{1,\dotsc,|\mathcal{M}|\}$ and $\sum_{j=1}^{|\mathcal{M}|}q_j = 1$. In Section \ref{sec:aggregation}, we discuss choices for the normalizing function $g$ and weightings $q$ that satisfy the properties specified in the introduction.
The primary quantities of interest in this paper are the aggregate performance measures for each algorithm and confidence intervals on that measure. Let $y \in \mathbb{R}^{|\mathcal{A}|}$ be a vector representing the aggregate performance for each algorithm. We desire confidence intervals, $Y^-,Y^+ \in \mathbb{R}^{|\mathcal{A}|} \times \mathbb{R}^{|\mathcal{A}|}$, such that, for a confidence level $\delta \in (0,0.5]$,
\begin{equation}
\Pr \left (\forall i \in \{1,2,\dotsc,|\mathcal A|\}, y_i \in [Y^-_i,Y^+_i] \right) \ge 1 - \delta.
\end{equation}
To compute an aggregate performance measure and its confidence intervals that meet the criteria laid out in the introduction, one must consider the entire evaluation procedure. We view an evaluation procedure to have three main components: data collection, data aggregation, and reporting of the results. During the data collection phase, samples are collected of the performance metric $X_{i,j}$ for each combination $(i,j)$ of an algorithm $i \in \mathcal{A}$ and environment $j \in \mathcal{M}$.
In the data aggregation phase, all samples of performance are normalized so the metric on each environment is on a similar scale, then they are aggregated to provide a summary of each algorithm's performance across all environments. Lastly, the uncertainty of the results is quantified and reported.
\section{Data Collection}
\label{sec:datacollection}
In this section, we discuss how common data collection methods are unable to answer the general evaluation question and then present a new method that can.
We first highlight the core difference in our approach to previous methods.
The main difference between data collection methods is in how the samples of performance are collected for each algorithm on each environment.
Standard approaches rely on first tuning an algorithm's hyperparameters, i.e., any input to an algorithm that is not the environment, and then generating samples of performance.
Our method instead relies on having a definition of an algorithm that can automatically select, sample, or adapt hyperparameters. This method can be used to answer the general evaluation question because its performance measure represents the knowledge required to use the algorithm. We discuss these approaches below.
\subsection{Current Approaches}
\label{sec:datacollectioncurrent}
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{images/EvaluationTuneAndReport2.pdf}
\caption{Data collection process of the tune-and-report method. The yellow box indicates trials using different random seeds.}
\label{fig:persampling}
\end{figure}
A typical evaluation procedure used in RL research is the \textit{tune-and-report} method. As depicted in Figure \ref{fig:persampling}, the tune-and-report method has two phases: a tuning phase and a testing phase. In the tuning phase, hyperparameters are optimized either manually or via a hyperparameter optimization algorithm. Then after tuning, only the best hyperparameters are selected and executed for $T$ trials using different random seeds to provide an estimate of performance.
The tune-and-report data collection method does not satisfy the usability requirement or the scientific requirement.
Recall that our objective is to capture the difficulty of using a particular algorithm.
Because the tune-and-report method ignores the amount of data used to tune the hyperparameter, an algorithm that only works well after significant tuning could be favored over one that works well without environment-specific tuning, thus, violating the requirements.
Consider an extreme example of an RL algorithm that includes all policy parameters as hyperparameters. This algorithm would then likely be optimal after any iteration of hyperparameter tuning that finds the optimal policy.
This effect is more subtle in standard algorithms, where hyperparameter tuning infers problem-specific information about how to search for the optimal policy, (e.g., how much exploration is needed, or how aggressive policy updates can be).
Furthermore, this demotivates the creation of algorithms that are easier to use but do not improve performance after finding optimal hyperparameters.
The tune-and-report method violates the scientific property by not accurately capturing the uncertainty of performance. Multiple i.i.d.~samples of performance are taken after hyperparameter tuning and used to compute a bound on the mean performance. However, these samples of performance do not account for the randomness due to hyperparameter tuning. As a result, any statistical claim would be inconsistent with repeated evaluations of this method.
This has been observed in several studies where further hyperparameter tuning has shown no difference in performance relative to baseline methods \citep{lucic2018gans,melis2018eval}.
The evaluation procedure proposed by \citet{dabney2014adaptive} addresses issues with uncertainty due to hyperparameter tuning and performance not capturing the usability of algorithms. Dabney's evaluation procedure computes performance as a weighted average over all $N$ iterations of hyperparameter tuning, and the entire tuning process repeats for $T$ trials. Even though this evaluation procedure fixes the problems with the tune-and-report approach, it violates our computationally tractable property by requiring $TN$ executions of the algorithm to produce just $T$ samples of performance.
In the case where $N=1$ it is not clear how hyperparameters should be set.
Furthermore, this style of evaluation does not cover the case where it is prohibitive to perform hyperparameter tuning, e.g., slow simulations, long agent lifetimes, lack of a simulator, and situations where it is dangerous or costly to deploy a bad policy.
In these situations, it is desirable for algorithms to be insensitive to the choice of hyperparameters or able to adapt them during a single execution. It is in this setting that the general evaluation question can be answered.
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{images/EvaluationCompleteAlgorithm2.pdf}
\caption{Data collection process using complete algorithm definitions. The yellow box indicates using different random seeds.}
\label{fig:completedata}
\end{figure}
\subsection{Our Approach}
\label{sec:datacollectionnew}
In this section, we outline our method, \textit{complete data collection}, that does not rely on hyperparameter tuning.
If there were no hyperparameters to tune, evaluating algorithms would be simpler. Unfortunately, how to automatically set hyperparameters has been an understudied area. Thus, we introduce the notion of a complete algorithm definition.
\begin{defn}[Algorithm Completeness]
An algorithm is complete on an environment $j$, when defined such that the only required input to the algorithm is meta-information about environment $j$, e.g., the number of state features and actions.
\end{defn}
Algorithms with a complete definition can be used on an environment and without specifying any hyperparameters.
Note that this does not say that an algorithm cannot receive forms of problem specific knowledge, only that it is not required.
A well-defined algorithm will be able to infer effective combinations of hyperparameters or adapt them during learning.
There are many ways to make an existing algorithm complete. In this work, algorithms are made complete by defining a distribution from which to randomly sample hyperparameters.
Random sampling may produce poor or divergent behavior in the algorithm, but this only indicates that it is not yet known how to set the hyperparameters of the algorithm automatically. Thus, when faced with a new problem, finding decent hyperparameters will be challenging.
One way to make an adaptive complete algorithm is to include a hyperparameter optimization method in the algorithm. However, all tuning must be done within the same fixed amount of time and cannot propagate information over trials used to obtain statistical significance.
Figure \ref{fig:completedata} shows the complete data collection method. For this method we limit the scope of algorithms to only include ones with complete definitions; thus, it does not violate any of the properties specified. This method satisfies the scientific requirement since it is designed to answer the general evaluation question, and the uncertainty of performance can be estimated using all of the trials. Again, this data collection method captures the difficulty of using an algorithm since the complete definition encodes the knowledge necessary for the algorithm to work effectively.
The compute time of this method is tractable, since $T$ executions of the algorithm produces $T$ independent samples of performance.
The practical effects of using the complete data collection method are as follows.
Researchers do not have to spend time tuning each algorithm to try and maximize performance.
Fewer algorithm executions are required to obtain a statistically meaningful result.
With this data collection method, improving upon algorithm definitions will become significant research contributions and lead to algorithms that are easy to apply to many problems.
\section{Data Aggregation}
\label{sec:aggregation}
Answering the general evaluation question requires a ranking of algorithms according to their performance on all environments $\mathcal{M}$. The aggregation step accomplishes this task by combining the performance data generated in the collection phase and summarizing it across all environments.
However, data aggregation introduces several challenges.
First, each environment has a different range of scores that need to be normalized to a common scale.
Second, a uniform weighting of environments can introduce bias.
For example, the set of environments might include many slight variants of one domain, giving that domain a larger weight than a single environment coming from a different domain.
\subsection{Normalization}
\label{sec:normalization}
The goal in score normalization is to project scores from each environment onto the same scale while not being exploitable by the environment weighting.
In this section, we first show how existing normalization techniques are exploitable or do not capture the properties of interest. Then we present our normalization technique: performance percentiles.
\subsubsection{Current Approaches}
We examine two normalization techniques: performance ratios and policy percentiles. We discuss other normalization methods in Appendix \ref{app:normalization}.
The performance ratio is commonly used with the Arcade Learning Environment to compare the performance of algorithms relative to human performance \cite{mnih2015dqn,machado2018ale}.
The performance ratio of two algorithms $i$ and $k$ on an environment $j$ is $\mathbf{E}[X_{i,j}]/\mathbf{E}[X_{k,j}]$.
This ratio is sensitive to the location and scale of the performance metric on each environment, such that an environment with scores in the range $[0,1]$ will produce larger differences than those on the range $[1000, 1001]$. Furthermore, all changes in performance are assumed to be equally challenging, i.e., going from a score of $0.8$ to $0.89$ is the same difficulty as $0.9$ to $0.99$. This assumption of linearity of difficulty is not reflected on environments with nonlinear changes in the score as an agent improves, e.g., completing levels in Super Mario.
A critical flaw in the performance ratio is that it can produce an arbitrary ordering of algorithms when combined with the arithmetic mean, $\sum_j q_j \mathbf{E}[X_{i,j}] / \mathbf{E}[X_{k,j}]$ \citep{fleming1986geom},
meaning a different algorithm in the denominator could change the relative rankings.
Using the geometric mean can address this weakness of performance ratios, but does not resolve the other issues.
Another normalization technique is \textit{policy percentiles}, a method that projects the score of an algorithm through the performance CDF of random policy search \citep{dabney2014adaptive}.
The normalized score for an algorithm, $i$, is $F_{X_{\Pi,j}}(X_{i,j})$, where $F_{X_{\Pi,j}}$ is the performance CDF when a policy is sampled uniformly from a set of policies, $\Pi$, on an environment $j$, i.e, $\pi \sim U(\Pi)$.
Policy percentiles have a unique advantage in that performance is scaled according to how difficult it is to achieve that level of performance relative to random policy search.
Unfortunately, policy percentiles rely on specifying $\Pi$, which often has a large search space. As a result, most policies will perform poorly, making all scores approach $1.0$. It is also infeasible to use when random policy search is unlikely to achieve high levels of performance.
Despite these drawbacks, the scaling of scores according to a notion of difficulty is desirable, so we adapt this idea to use any algorithm's performance as a reference distribution.
\subsubsection{Our Approach}
\label{sec:normalizationours}
An algorithm's performance distribution can have an interesting shape with large changes in performance that are due to divergence, lucky runs, or simply that small changes to a policy can result in large changes in performance \citep{jordan2018cdf}.
These effects can be seen in Figure \ref{fig:cdf_cartpole}, where there is a quick rise in cumulative probability for a small increase in performance.
Inspired by \citet{dabney2014adaptive}'s policy percentiles, we propose \textit{performance percentiles}, a score normalization technique that can represent these intricacies.
The probability integral transform shows that projecting a random variable through its CDF transforms the variable to be uniform on $[0,1]$ \citep{dodge2006oxford}. Thus, normalizing an algorithm's performance by its CDF will equally distribute and represent a linear scaling of difficulty across $[0,1]$. When normalizing performance against another algorithm's performance distribution, the normalized score distribution will shift towards zero when the algorithm is worse than the normalizing distribution and shift towards one when it is superior.
As seen in Figure \ref{fig:cdf_cartpole}, the CDF can be seen as encoding the relative difficulty of achieving a given level of performance, where large changes in an algorithm's CDF output indicate a high degree of difficulty for that algorithm to make an improvement and similarly small changes in output correspond to low change in difficulty.
In this context difficulty refers to the amount of random chance (luck) needed to achieve a given level of performance.
To leverage these properties of the CDF, we define performance percentiles, that use a weighted average of each algorithm's CDF to normalize scores for each environment.
\begin{defn}[Performance Percentile]
\label{def:performancepercentile}
In an evaluation of algorithms, $\mathcal{A}$, the performance percentile for a score $x$ on an environment, $j$, is $F_{\bar X_j}(x, w_j)$, where $F_{\bar X_j}$ is the mixture of CDFs $F_{\bar X_j}(x,w_j) \coloneqq \sum_{i=1}^{|\mathcal{A}|} w_{j,i} F_{X_{i,j}}(x)$, with weights $w_j \in \mathbb{R}^{|\mathcal{A}|}$,
$\sum_{i=1}^{|\mathcal{A}|} w_{j,i} = 1$, and $\forall i \ w_{j,i} \ge 0$.
\end{defn}
So we can say that performance percentiles capture the performance characteristic of an environment relative to some averaged algorithm. We discuss how to set the weights $w_j$ in the next section.
\begin{figure
\centering
\includegraphics[width=0.48\textwidth]{images/cdf_cartpole_normalizedpts.png}
\caption{This plot shows the CDF of average returns for the Sarsa-Parl2, Sarsa($\lambda$), and Actor-Critic algorithms on the Cart-Pole environment. Each line represents the empirical CDF using $10,\!000$ trials and the shaded regions represent the $95\%$ confidence intervals. To illustrate how the performance percentiles work, this plot shows how samples of performance (black dots) are normalized by each CDF, producing the normalized scores (colored dots). The correspondence between a single sample and its normalized score is shown by the dotted line.}
\label{fig:cdf_cartpole}
\end{figure}
Performance percentiles are closely related to the concept of (probabilistic) performance profiles \cite{Dolan2002profiles,barreto2010probeval}. The difference being that performance profiles report the cumulative distribution of normalized performance metrics over a set of tasks (environments), whereas performance percentiles are a technique for normalizing scores on each task (environment).
\subsection{Summarization}
\label{sec:summarization}
A weighting over environments is needed to form an aggregate measure.
We desire a weighting over environments such that no algorithm can exploit the weightings to increase its ranking. Additionally, for the performance percentiles, we need to determine the weighting of algorithms to use as the reference distribution. Inspired by the work of \citet{balduzzi2018eval}, we propose a weighting of algorithms and environments, using the equilibrium of a two-player game.
In this game, one player, $p$, will try to select an algorithm to maximize the aggregate performance, while a second player, $q$, chooses the environment and reference algorithm to minimize $p$'s score. Player $p$'s pure strategy space, $\mathcal{S}_1$, is the set of algorithms $\mathcal{A}$, i.e., $p$ plays a strategy $s_1=i$ corresponding to an algorithm $i$. Player $q$'s pure strategy space, $\mathcal{S}_2$, is the cross product of a set of environments, $\mathcal{M}$, and algorithms, $\mathcal{A}$, i.e., player $q$ plays a strategy $s_2=(j,k)$ corresponding to a choice of environment $j$ and normalization algorithm $k$.
We denote the pure strategy space of the game by $\mathcal{S} \coloneqq \mathcal{S}_1 \times \mathcal{S}_2$. A strategy, $s \in \mathcal{S}$, can be represented by a tuple $s=(s_1,s_2)=(i,(j,k))$.
The utility of strategy $s$ is measured by a payoff function $u_p \colon \mathcal{S} \to \mathbb{R}$ and $u_q \colon \mathcal{S} \to \mathbb{R}$ for players $p$ and $q$ respectively. The game is defined to be zero sum, i.e., $u_q(s) = -u_p(s)$. We define the payoff function to be
$u_p(s) \coloneqq \mathbf{E}[F_{X_{k,j}}(X_{i,j})]$.
Both players $p$ and $q$ sample strategies from probability distributions $p \in \Delta(\mathcal{S}_1)$ and $q \in \Delta(\mathcal{S}_2)$, where $\Delta(\mathcal{X})$ is the set of all probability distributions over $\mathcal{X}$.
The equilibrium solution of this game naturally balances the normalization and environment weightings to counter each algorithm's strengths without conferring an advantage to a particular algorithm.
Thus, the aggregate measure will be useful in answering the general evaluation question.
After finding a solution $(p^*,q^*)$, the aggregate performance measure $y_i$ for an algorithm $i$ defined as
\begin{equation}
\label{eq:aggforalgi}
y_i \coloneqq \sum_{j=1}^{|\mathcal{M}|}\sum_{k=1}^{|\mathcal{A}|} q^*_{j,k} \mathbf{E}[F_{X_{k,j}}(X_{i,j})].
\end{equation}
To find a solution $(p^*,q^*)$, we employ the $\alpha$-Rank technique \citep{omidshafiei2019alpharank}, which returns a stationary distribution over the pure strategy space $\mathcal{S}$. $\alpha$-Rank allows for efficient computation of both the equilibrium and confidence intervals on the aggregate performance \citep{rowland2019alphabounds}. We detail this method and details of our implementation in Appendix \ref{app:alpharank}.
\section{Reporting Results}
\label{sec:reportingresults}
As it is crucial to quantify the uncertainty of all claimed performance measures, we first discuss how to compute confidence intervals for both single environment and aggregate measures, then give details on displaying the results.
\subsection{Quantifying Uncertainty}
\label{sec:aggbound}
In keeping with our objective to have a scientific evaluation, we require our evaluation procedure to quantify any uncertainty in the results. When concerned with only a single environment, standard concentration inequalities can compute confidence intervals on the mean performance. Similarly, when displaying the distribution of performance, one can apply standard techniques for bounding the empirical distribution of performance.
However, computing confidence intervals on the aggregate has additional challenges.
Notice that in \eqref{eq:aggforalgi} computing the aggregate performance requires two unknown values: $q^*$ and the mean normalized performance, $\mathbf{E}[F_{X_{k,j}}(X_{i,j})]$.
Since $q^*$ depends on mean normalized performance, any uncertainty in the mean normalized performance results in uncertainty in $q^*$.
To compute valid confidence intervals on the aggregate performance, the uncertainty through the entire process must be considered.
We introduce a process to compute the confidence intervals, which we refer to as \textit{performance bound propagation} (PBP).
We represent PBP as a function $\text{PBP} \colon \mathcal{D} \times \mathbb{R} \to \mathbb{R}^{|\mathcal{A}|} \times \mathbb{R}^{|\mathcal{A}|}$, which maps a dataset $D \in \mathcal{D}$ containing all samples of performance and a confidence level $\delta \in (0,0.5]$, to vectors $Y^-$ and $Y^+$ representing the lower and upper confidence intervals, i.e., $(Y^-, Y^+) = \text{PBP}(D,\delta)$.
The overall procedure for PBP is as follows, first compute confidence intervals for each $F_{X_{i,j}}$, then using these intervals compute confidence intervals on each mean normalized performance, next determine an uncertainty set $\mathcal{Q}$ for $q^*$ that results from uncertainty in the mean normalized performance, finally for each algorithm find the minimum and maximum aggregate performance over the uncertainty in the mean normalized performances and $\mathcal{Q}$.
We provide pseudocode in Appendix \ref{app:confagg} and source code in the repository.
We prove that PBP produces valid confidence intervals for a confidence level $\delta \in (0, 0.5]$ and a dataset $D$ containing $T_{i,j} >1$ samples of performance for all algorithms $i \in \mathcal{A}$ and environments $j \in \mathcal{M}$.
\begin{restatable}{thm}{pbpthm}
If $(Y^-,Y^+) = \text{PBP}(D,\delta)$, then
\begin{equation}
\Pr \left ( \forall i \in {1,2,\dotsc,|\mathcal{A}|},\ y_i \in [Y^-_i, Y^+_i] \right ) \ge 1-\delta.
\end{equation}
\end{restatable}
\begin{proof}
Although the creation of valid confidence intervals is critical to this contribution, due to space restrictions it is presented in Appendix \ref{app:confagg}.
\end{proof}
\subsection{Displaying Results}
\begin{figure
\centering
\includegraphics[width=0.48\textwidth]{images/cdf_bounds_acab.png}
\caption{This plot shows the distribution of average returns for the Actor-Critic algorithm on the Acrobot environment. The $x$-axis represents a probability and the $y$-axis represents the average return such that the proportion of trials that have a value less than or equal to $y$ is $x$, e.g., at $x=0.5$, $y$ is the median. Each line represents the empirical quantile function using a different number of trials and the shaded regions represent the $95\%$ confidence intervals computed using concentration inequalities.
%
In this plot, the larger the area under the curve, the better the performance.
%
This plot highlights the large amount of uncertainty when using small sample sizes and how much it decreases with more samples. }
\label{fig:return_ab}
\end{figure}
In this section, we describe our method for reporting the results. There are three parts to our method: answering the stated hypothesis, providing tables and plots showing the performance and ranking of algorithms for all environments, and the aggregate score, then for each performance measure, provide confidence intervals to convey uncertainty.
The learning curve plot is a standard in RL and displays a performance metric (often the return) over regular intervals during learning.
While this type of plot might be informative for describing some aspects of the algorithm's performance, it does not directly show the performance metric used to compare algorithms, making visual comparisons less obvious.
Therefore, to provide the most information to the reader, we suggest plotting the distribution of performance for each algorithm on each environment. Plotting the distribution of performance has been suggested in many fields as a means to convey more information, \citep{Dolan2002profiles,farahmand2010cdfevo,reimers2017scoredist,cohen2018distributed}.
Often in RL, the object is to maximize a metric, so we suggest showing the quantile function over the CDF as it allows for a more natural interpretation of the performance, i.e., the higher the curve, the better the performance \citep{bellemare13arcade}. Figure \ref{fig:return_ab} show the performance distribution with $95\%$ confidence intervals for different sample sizes.
It is worth noting that when tuning hyperparameters the data needed to compute these distributions is already being collected, but only the results from the tuned runs are being reported.
By only reporting only the tuned performance it shows what an algorithm can achieve not what it is likely to achieve.
\section{Experimental Results}
\label{sec:experiments}
In this section, we describe and report the results of experiments to illustrate how this evaluation procedure can answer the general evaluation question and identify when a modification to an algorithm or its definition improves performance. We also investigate the reliability of different bounding techniques on the aggregate performance measure.
\subsection{Experiment Description}
To demonstrate the evaluation procedure we compare the algorithms: Actor-Critic with eligibility traces (AC) \cite{sutton2018reinforcement}, Q($\lambda$), Sarsa($\lambda$), \cite{suttonbaro1998book}, NAC-TD \citep{morimura2005ntd,degris2012inac,thomas2014bias}, and proximal policy optimization (PPO) \citep{schulman2017ppo}.
The learning rate is often the most sensitive hyperparameter in RL algorithms.
So, we include three versions of Sarsa$(\lambda)$, Q$(\lambda)$, and AC: a base version, a version that scales the step-size with the number of parameters (e.g., Sarsa$(\lambda)$-s), and an adaptive step-size method, Parl2 \citep{dabney2014adaptive}, that does not require specifying the step size.
Since none of these algorithms have an existing complete definition, we create one by randomly sampling hyperparameters from fixed ranges. We consider all parameters necessary to construct each algorithm, e.g., step-size, function approximator, discount factor, eligibility trace decay. For the continuous state environments, each algorithm employs linear function approximation using the Fourier basis \cite{konidaris2011fourier} with a randomly sampled order. See Appendix \ref{app:definitions} for full details of each algorithm.
These algorithms are evaluated on $15$ environments, eight discrete MDPs, half with stochastic transition dynamics, and seven continuous state environments: Cart-Pole \citep{florian2007cartpole}, Mountain Car \citep{suttonbaro1998book}, Acrobot \citep{sutton1995acrobot}, and four variations of the pinball environment \citep{konidaris2009pinball,geramifard2015rlpy}. For each independent trial, the environments have their dynamics randomly perturbed to help mitigate environment overfitting \citep{whiteson2011eval}; see code for details. For further details about the experiment see Appendix \ref{app:environments}.
While these environments have simple features compared to the Arcade Learning Environment \citep{bellemare13arcade}, they remain useful in evaluating RL algorithms for three reasons. First is that experiments finish quickly. Second, the environments provide interesting insights into an algorithm's behavior. Third, as our results will show, there is not yet a complete algorithm that can reliably solve each one.
We execute each algorithm on each environment for $10,\!000$ trials. While this number of trials may seem excessive, our goal is to detect a statistically meaningful result. Detecting such a result is challenging because the variance of RL algorithms performance is high; we are comparing $|\mathcal{A}| \times |\mathcal{M}| = 165$ random variables, and we do not assume the performances are normally distributed. Computationally, executing ten thousand trials is not burdensome if one uses an efficient programming language such as Julia \citep{bezanson2017julia} or C++, where we have noticed approximately two orders of magnitude faster execution than similar Python implementations. We investigate using smaller sample sizes at the end of this section.
\begin{table}
\begin{tabular}{lcr}
\toprule
\multicolumn{3}{c}{\textbf{Aggregate Performance}} \\
\midrule
\textbf{Algorithms} & \multicolumn{1}{c}{\textbf{Score}} & \textbf{Rank} \\ \midrule
\rowcolor[HTML]{EFEFEF}
Sarsa-Parl2 & $0.4623 (\textbf{0.3904}, 0.5537)$ & 1 (2,1) \\
Q-Parl2 & $0.4366 (\textbf{0.3782}, 0.5632)$ & 2 (2,1) \\
\rowcolor[HTML]{EFEFEF}
AC-Parl2 & $0.1578 (0.0765, \textbf{0.3129})$ & 3 (11,3) \\
Sarsa$(\lambda)$-s & $0.0930 (0.0337, 0.2276)$ & 4 (11,3) \\
\rowcolor[HTML]{EFEFEF}
AC-s & $0.0851 (0.0305, 0.2146)$ & 5 (11,3) \\
Sarsa$(\lambda)$ & $0.0831 (0.0290, 0.2019)$ & 6 (11,3) \\
\rowcolor[HTML]{EFEFEF}
AC & $0.0785 (0.0275, 0.2033)$ & 7 (11,3) \\
Q$(\lambda)$-s & $0.0689 (0.0237, 0.1973)$ & 8 (11,3) \\
\rowcolor[HTML]{EFEFEF}
Q$(\lambda)$ & $0.0640 (0.0214, 0.1780)$ & 9 (11,3) \\
\rowcolor[HTML]{EFEFEF}
NAC-TD & $0.0516 (0.0180, 0.1636)$ & 10 (11,3) \\
PPO & $0.0508 (0.0169, 0.1749)$ & 11 (11,3) \\
\bottomrule
\end{tabular}
\caption{Aggregate performance measures for each algorithm and their rank. The parentheses contain the intervals computed using PBP and together all hold with $95\%$ confidence. The bolded numbers identify the best ranked statistically significant differences.}
\label{tab:agg_res}
\end{table}
\subsection{Algorithm Comparison}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{images/aggregate_perf_bounds.png}
\caption{The aggregate performance for each algorithm with confidence intervals using PBP, PBP-t, and bootstrap. The width of each interval is scaled so all intervals hold with $95\%$ confidence.}
\label{fig:aggperf}
\end{figure}
The aggregate performance measures and confidence intervals are illustrated in Figure \ref{fig:aggperf} and given in Table \ref{tab:agg_res}. Appendix \ref{app:indperf} lists the performance tables and distribution plots for each environment. Examining the empirical performances in these figures, we notice two trends. The first is that our evaluation procedure can identify differences that are not noticeable in standard evaluations. For example, all algorithms perform near optimally when tuned properly (indicated by the high end of the performance distribution). The primary differences between algorithms are in the frequency of high performance and divergence (indicated by low end of the performance distribution). Parl2 methods rarely diverge, giving a large boost in performance relative to the standard methods.
The second trend is that our evaluation procedure can identify when theoretical properties do or do not make an algorithm more usable. For example, Sarsa($\lambda$) algorithms outperform their Q($\lambda$) counterparts. This result \emph{might} stem from the fact that Sarsa($\lambda$) is known to converge with linear function approximation \citep{perkins2002linearsarsa} while Q($\lambda$) is known to diverge \citep{baird1995residual,wiering2004divql}. Additionally, NAC-TD performs worse than AC despite that natural gradients are a superior ascent direction. This result is due in part because it is unknown how to set the three step-sizes in NAC-TD, making it more difficult to use than AC. Together these observations point out the deficiency in the way new algorithms have been evaluated. That is, tuning hyperparameters hides the lack of knowledge required to use the algorithm, introducing bias that favors the new algorithm. In contrast, our method forces this knowledge to be encoded into the algorithm, leading to a more fair and reliable comparison.
\subsection{Experiment Uncertainty}
While the trends discussed above might hold true in general, we must quantify our uncertainty. Based on the confidence intervals given using PBP, we claim with $95\%$ confidence that on these environments and according to our algorithm definitions, Sarsa-Parl2 and Q-Parl2 have a higher aggregate performance of average returns than all other algorithms in the experiment.
It is clear that $10,\!000$ trials per algorithm per environment is not enough to detect a unique ranking of algorithms using the nonparametric confidence intervals in PBP.
We now consider alternative methods, PBP-t, and the percentile bootstrap.
PBP-t replaces the nonparameteric intervals in PBP with ones based on the Student's t-distribution.
We detail these methods in Appendix \ref{app:altbounds}.
From Figure \ref{fig:aggperf}, it is clear that both alternative bounds are tighter and thus useful in detecting differences.
Since assumptions of these bounds are different and not typically satisfied, it is unclear if they are valid.
\begin{table}
\begin{tabular}{|l|lr|lr|lr|}
\hline
\multicolumn{7}{|c|}{\textbf{Confidence Interval Performance}} \\ \hline
& \multicolumn{2}{c|}{PBP} & \multicolumn{2}{c|}{PBP-t} & \multicolumn{2}{c|}{Bootstrap} \\ \hline
Samples & FR & SIG & FR & SIG & FR & SIG \\ \hline
\rowcolor[HTML]{EFEFEF}
$10$ & 0.0 & 0.0 & 1.000 & 0.00 & 0.112 & 0.11 \\
$30$ & 0.0 & 0.0 & 0.000 & 0.00 & 0.092 & 0.37 \\
\rowcolor[HTML]{EFEFEF}
$100$ & 0.0 & 0.0 & 0.000 & 0.02 & 0.084 & 0.74 \\
$1,\!000$ & 0.0 & 0.0 & 0.000 & 0.34 & 0.057 & 0.83 \\
\rowcolor[HTML]{EFEFEF}
$10,\!000$ & 0.0 & 0.33 & 0.003 & 0.83 & 0.069 & 0.83 \\ \hline
\end{tabular}
\caption{Table showing the failure rate (FR) and proportion of significant pairwise comparison (SIG) identified for $\delta=0.05$ using different bounding techniques and sample sizes. The first column represents the sample size. The second, third, and fourth columns represent the results for PBP, PBP-t, and bootstrap bound methods respectively. For each sample size, $1,\!000$ experiments were conducted.}
\label{tab:boundresults}
\end{table}
To test the different bounding techniques, we estimate the failure rate of each confidence interval technique at different sample sizes. For this experiment we execute $1,\!000$ trials of the evaluation procedure using sample sizes (trials per algorithm per environment) of $10$, $30$, $100$, $1,\!000$, and $10,\!000$.
There are a total of $11.14$ million samples per algorithm per environment. To reduce computation costs, we limit this experiment to only include Sarsa($\lambda$)-Parl2, Q($\lambda$)-Parl2, AC-Parl2, and Sarsa($\lambda$)-s. Additionally, we reduce the environment set to be the discrete environments and Mountain Car. We compute the failure rate of the confidence intervals, where a valid confidence interval will have a failure rate less than or equal to $\delta$, e.g., for $\delta=0.05$ failure rate should be less than $\le 5\%$. We report the failure rate and the proportion of statistically significant pairwise comparisons in Table \ref{tab:boundresults}.
All methods use the same data, so the results are not independent.
The PBP method has zero failures indicating it is overly conservative.
The failure rate of PBP-t is expected to converge to zero as the number of samples increase due to the central limit theorem.
PBP-t begins to identify significant results at a sample size of $1,\!000$, but it is only at $10,\!000$ that it can identify all pairwise differences.\footnote{Sarsa-Parl2 and Q-Parl2 have similar performance on discrete environments so we consider detecting $83\%$ of results optimal.}
The bootstrap technique has the tightest intervals, but has a high failure rate.
These results are stochastic and will not necessarily hold with different numbers of algorithms and environments. So, one should use caution in making claims that rely on either PBP-t or bootstrap. Nevertheless, to detect statistically significant results, we recommend running between $1,\!000$, and $10,\!000$ samples, and using the PBP-t over bootstrap.
While this number of trials seems, high it is a necessity as comparison of multiple algorithms over many environments is a challenging statistical problem with many sources of uncertainty.
Thus, one should be skeptical of results that use substantially fewer trials.
Additionally, researchers are already conducting many trials that go unreported when tuning hyperparameters.
Since our method requires no hyperparameter tuning, researchers can instead spend the same amount of time collecting more trials that can be used to quantify uncertainty.
There are a few ways that the number of trials needed can be reduced.
The first is to think carefully about what question one should answer so that only a few algorithms and environments are required.
The second is to use active sampling techniques to determine when to stop generating samples of performance for each algorithm environment pair \citep{rowland2019alphabounds}.
It is important to caution the reader that this process can bias the results if the sequential tests are not accounted for \citep{howard2018confseq}.
Summarizing our experiments, we make the following observations. Our experiments with complete algorithms show that there is still more work required to make standard RL algorithms work reliably on even extremely simple benchmark problems. As a result of our evaluation procedure, we were able to identify performance differences in algorithms that are not noticeable under standard evaluation procedures. The tests of the confidence intervals suggest that both PBP and PBP-t provide reliable estimates of uncertainty. These outcomes suggest that this evaluation procedure will be useful in comparing the performance of RL algorithms.
\section{Related Work}
This paper is not the first to investigate and address issues in empirically evaluating algorithms. The evaluation of algorithms has become a signficant enough topic to spawn its own field of study, known as \textit{experimental algorithmics} \citep{fleischer2002algorithmics,mcgeoch2012guide}.
In RL, there have been significant efforts to discuss and improve the evaluation of algorithms \citep{whiteson2011special}.
One common theme has been to produce shared benchmark environments, such as those in the annual reinforcement learning competitions \citep{whiteson2008competition,dimitrakakis2014competition}, the Arcade Learning Environment \citep{bellemare13arcade}, and numerous others which are to long to list here.
Recently, there has been a trend of explicit investigations into the reproducibility of reported results \citep{henderson2018deep,islam2017reproducibility,khetarpal2018reevaluate,colas2018seeds}. These efforts are in part due to the inadequate experimental practices and reporting in RL and general machine learning \citep{pineau2020reproducibility,lipton2018trends}.
Similar to these studies, this work has been motivated by the need for a more reliable evaluation procedure to compare algorithms. The primary difference in our work to these is that the knowledge required to use an algorithm gets included in the performance metric.
An important aspect of evaluation not discussed so far in this paper is competitive versus scientific testing \citep{hooker1995testing}. Competitive testing is the practice of having algorithms compete for top performance on benchmark tasks. Scientific testing is the careful experimentation of algorithms to gain insight into how an algorithm works.
The main difference in these two approaches is that competitive testing only says \textit{which} algorithms worked well but not \textit{why}, whereas scientific testing directly investigates the what, when, how, or why better performance can be achieved.
There are several examples of recent works using scientific testing to expand our understanding of commonly used methods.
\citet{lyle2019distribuional} compares distributional RL approaches using different function approximation schemes showing that distributional approaches are only effective when nonlinear function approximation is used.
\citet{tucker2018mirage} explore the sources of variance reduction in action dependent control variates showing that improvement was small or due to additional bias. \citet{witty2018measuring} and \citet{atrey2020saliency} investigate learned behaviors of an agent playing Atari 2600 games using ToyBox \cite{foley2018toybox}, a tool designed explicitly to enable carefully controlled experimentation of RL agents.
While, at first glance the techniques developed here seems to be only compatible with competitive testing, this is only because we specified question with a competitive answer.
The techniques developed here, particularly complete algorithm definitions, can be used to accurately evaluate the impact of various algorithmic choices. This allows for the careful experimentation to determine what components are essential to an algorithm.
\section{Conclusion}
\label{sec:conclusion}
The evaluation framework that we propose provides a principled method for evaluating RL algorithms.
This approach facilitates fair comparisons of algorithms by removing unintentional biases common in the research setting.
By developing a method to establish high-confidence bounds over this approach, we provide the framework necessary for reliable comparisons.
We hope that our provided implementations will allow other researchers to easily leverage this approach to report the performances of the algorithms they create.
\section*{Acknowledgements}
The authors would like to thank Kaleigh Clary, Emma Tosch, and members of the Autonomous Learning Laboratory: Blossom Metevier, James Kostas, and Chris Nota, for discussion and feedback on various versions of this manuscript.
Additionally, we would like to thank the reviewers and meta-reviewers for their comments, which helped improved this paper.
This work was performed in part using high performance computing equipment obtained under a grant from the Collaborative R\&D Fund managed by the Massachusetts Technology Collaborative.
This work was supported in part by a gift from Adobe.
This work was supported in part by the Center for Intelligent Information Retrieval. Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect those of the sponsor.
Research reported in this paper was sponsored in part by the CCDC Army Research Laboratory under Cooperative Agreement W911NF-17-2-0196 (ARL IoBT CRA). The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Army Research Laboratory or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein.
|
\section{Introduction}
\label{sec: introduction}
Multi-component fluids are very common in nature, where degree of polydispersity is often attributed to a distribution of particle sizes. The outcome of the particle size variation has an important role, e.g., in the glass forming ability of the colloidal fluids. In experiments, polydispersity offers a means to avoid crystallization and thus allows one to study very high viscous fluids.\cite{pusey1987,auer2001} Application of these materials demands understanding their physics and our ability to manipulate them for industrial processing. Various equilibrium aspects of size polydisperse systems (e.g., packing, thermodynamics, etc.) have been discussed in earlier works.\cite {Stell, Sollich, Warren,Salacuse1982_JCP77,Salacuse1984_JCP81,Sollich1998_PRL80,Warren1998_PRL80,Evans2001_JCP114,Sollich2002_JPCM14, Wilding2006_JCP125,Buzzacchi2006_PRE73,Wilding2008_PRE77,Jacobs2013_JCP139} On the other hand, polydispersity is very common in biological systems as well, where a cell contains large amount of different types of macromolecules that differ in their shape, size, and interactions leading to interesting dynamical behavior and geometry controlled kinetics.\cite{IMSokolov2012_SoftMatter,RMetzler2000_PhysicsReports,eli2012_PhysToday,IIzeddin2014_eLife,Benichou,jeon2016} In our recent work,\cite{Shagolsem2015_JCP,Shagolsem2016_JCP} a model complex fluid characteristic of biological systems was introduced, where we focused on interaction multiplicity (i.e., energy polydispersed) among the constituents -- a minimal physical model in which {\it all particles are different} (APD) in the sense that their interaction parameters are chosen at random from a given distribution. The results of our earlier work on statics and dynamics of APD fluid close to solidification transition are briefly discussed below.
In an energy polydisperse fluid, although the particles have random interactions, the local self-organization show that the particles relax into a non-random state characterized by clustering of particles according to the values of their pair interaction parameters which become more pronounced as temperature is decreased throughout the fluid region. This particle identity ordering has consequences in the particles' dynamical behavior. For example, when the particle identity ordering sets in, different types of particles find themselves in different local environments and thus show distinct diffusion coefficients. It is important to note that size polydispersity leads to entropic effects associated with packing a large number of dissimilar objects in an efficient way thus increasing the number of available configurations, whereas interaction polydispersity affects the weights of the different configurations through neighborhood identity ordering. Furthermore, several aspects of the system, viz., identity ordering and metastable cluster, and effect on tensile strength (where a multicomponent model solids shows limiting tensile strength that far exceed the corresponding one-component solids) are reported in references~\cite{azizi2019,kulveer2019}.
In the present paper, non-equilibrium aspects of the energy polydispersed fluid in thin-film geometry is addressed, where we study energy polydisperse Lennard-Jones fluid subjected to linear flow field. There is large volume of work on the rheological studies of colloidal dispersion and soft-matter in general.\cite{brown2014,chen2010} Recent advances in the field allow access to the microscopic structural changes during the process and thus gives better understanding of the mechanical response of these complex fluids. In particular, we study the effect of flow field on particle distribution in the film, spatial ordering of particles, and local identity ordering. Also, macroscopic response is studied by measuring stress components and hence shear viscosity. Finally, the role of substrate quality on the aforesaid quantities of the fluid is also investigated by considering two types of substrates (non-selective and selective substrates) described in the following section. \\
Our paper is organized as follows: In section~\ref{sec: model-simulation-details}, we present the model and simulation details. Particle distribution in the film through density profile is reported in section~\ref{sec: density-profile}, particle ordering and flow-field induced melting in section~\ref{sec: identity-ordering}, and macroscopic response in section~\ref{sec: macro-response}. Finally, in section~\ref{sec: conclusion} we summarize and discuss the results of this work.
\section{Model and simulation details}
\label{sec: model-simulation-details}
We perform coarse-grained MD simulations (NVT-ensemble) in two dimensions considering $N=2500$ particles in a simulation box of dimension $L_x=80\sigma$ and $L_y=49\sigma$, with $\sigma$ particle diameter. All the particles have same mass $m$ and size $\sigma$ which are set to unity.
The particles interact via Lennard-Jones (LJ) potential
\begin{equation}
U_{ij}(r) = 4\epsilon_{ij}\left[(\sigma/r)^{12}-(\sigma/r)^{6}\right]~,
\label{eqn: LJ-potential}
\end{equation}
which is cut-off and shifted to zero at $r_{\rm c}=2.5\sigma$. Here, $\epsilon_{ij}$ and $r$ are the interaction strength and separation between a pair of particles $i$ and $j$ respectively.
The equation of motion of $i^{\rm th}$ particle is given by the Langevin equation
\begin{equation}
m \frac{d^2 \bf r_i}{dt^2} + \zeta \frac{d \bf r_i}{dt} = -\frac{\partial U}{\partial \bf r_i} + {\bf f_i}~,
\label{eqn: langevin}
\end{equation}
with $\bf r_i$ the position of particle $i$, $\zeta$ the friction coefficient (assumed to be the same for all particles), $U$ is the sum of all the pair potentials $U_{ij}$ ($i\ne j$), and $\bf f_i$ a random external force with zero mean and second moment proportional to the product of temperature $T$ and $\zeta$. All the physical quantities are expressed in LJ reduced units, with LJ time $\tau_{_{\rm LJ}}=1$.\cite{allen} We use velocity-Verlet scheme to integrate the equations of motion with time step $\delta t = 0.005$. \\
In order to introduce energy polydispersity, we assign interaction parameter $\epsilon_i$ to each particle $i$ which is drawn randomly from a uniform distribution in the range 1-4, where the value of interaction strength $\epsilon_i$ defines the particle identity. Between particles $i$ and $j$ the interaction strength is given by $\epsilon_{ij}=\sqrt{\epsilon_i\epsilon_j}$ (following Lorentz-Berthelot anzatz \cite {Berthelot}). The distribution $P(\epsilon_{ij})$ is shown in fig.~\ref{fig: fig1_schematic}(a), where the most probable value $\epsilon_{ij}^{\rm max}=2.0$ and mean $\overline{\epsilon}_{ij}=2.42$. As a reference system, we simulate a one-component (1C) LJ system at the same density with interaction parameter $\epsilon_{ij}=2.5$ (mid-point of the interval $1-4$). As reported in earlier studies, \cite{Shagolsem2015_JCP,Shagolsem2016_JCP} at this number density $\rho^\ast=0.69$ the system behaves as a liquid at high temperatures. In general, the phase diagram of APD system differs from that of a pure one component (1C) fluid, \cite{Wilding2008_PRE77} where in 1C system (at the same density) it passes through the triple point and thus goes directly from fluid to solid-gas coexistence region upon cooling.\cite{Abraham1981_PhysRep,patashinsky}
The liquid-solid transition takes place within a narrow temperature range in which the mean interaction energy decreases sharply upon cooling from the liquid to solid-fluid coexistence region. The transition temeperature is estimated to be $T^\ast \approx 1.0$ for both APD and 1C systems (for bulk systems), and interesting dynamical differences are observed close to $T^*$ as detailed in references \cite{Shagolsem2015_JCP,Shagolsem2016_JCP}. \
\begin{table}[h]
\caption{Pair-wise interaction strength among the species}
\centering
\begin{tabular}{|l|l|l|}
\hline
\bf{Between $i-j$} & \bf{Strength $\epsilon_{ij}$} & \bf{Nature} \\
\hline
fluid - fluid & polydisperse & attractive (different strength) \\
\hline
fluid - wall & (a) $2.5$ & attractive (same strength) \\
& & -- {\it non-selective wall case} \\
& (b) polydisperse & attractive (different strength) \\
& & -- {\it selective wall case} \\
\hline
wall - wall & $1.0$ & repulsive \\
\hline
\end{tabular}
\label{table: params}
\end{table}
In the current study, fluid particles are confined by two all-atom walls in slit geometry of width $D=46\sigma$, see fig.~\ref{fig: fig1_schematic}, and thus the system is periodic along X-axis only. Two types of confining walls are considered: (i) {\it non-selective walls} (ns-Wall), where the wall particles have no preferential interaction with the fluid particles, and we set wall-fluid interaction strength $\epsilon_{\rm wall - fluid} = 2.5$ for all particle types, and (ii) {\it selective walls} (s-Wall), where the wall-fluid interaction strength varies depending on the particle type. Among the wall particles we set $\epsilon_{\rm wall}=1.0$. In both the wall types, interaction cut-off distance is set to $r_c=2.5\sigma$. However, in this study, we do not consider the case of ns-W made by setting $r_c=1.12\sigma$ (i.e., purely repulsive case). Various interaction strengths prensent in the system are briefly summarized in table~\ref{table: params}.
\begin{figure}[ht]
\includegraphics*[width=0.45\textwidth]{figs1/fig1_schematic-1.pdf}
\begin{center}
\caption{Equilibrium configurations of the APD system relaxed at $\delta=0$ for neutral substrate. Fluid flows along X-direction and the velocity gradient is along the Y-direction. Wall moves along with the fluid. Particles are colored according to their interaction parameter $\epsilon_i$ in the range 1-4, see color scale bar.}
\label{fig: fig1_schematic}
\end{center}
\end{figure}
It is known that the presence of substrate can supress or enhance onset of solidification and thus shift the transition temperature to a lower or higher values.
For example, simulations study of freezing and melting of Lennard-Jones methane confined within slit pores by Miyahara and Gubbins show that when the adsorbate-wall interactions were weaker compared to the adsorbate-adsorbate interactions, the freezing temperature is significantly lower than in the bulk, as reported in most of the experiments for silica pores. In contrast, a large increase in the freezing temperature is observed for strongly attractive materials, such as porous-carbons. No appreciable change in the freezing temperature is observed when the adsorbate-wall and adsorbate-adsorbate interactions have similar strength.
A detailed investigation is necessary to understand the effect of confinement on the solidification transition of APD fluids under variation of parameters such as density, film thickness, etc., and we leave this task for a future study. However, for the purpose of current work, approximate values of $T^*$ is located for the confined systems by performing heating/cooling cycles and using the procedure detailed in reference \cite{Shagolsem2015_JCP}.
For the considered density, it is observed that $T^*$ deviates from that of the bulk (albeit very small) and follow the trend: $T^*_{\rm bulk}(\approx 0.98) < T^*_{\rm s-w}(\approx 0.997) < T^*_{\rm ns-w}(\approx 1.01)$. Since the walls are attractive elevation of $T^*$ compared to the bulk is expected. However, as defined above, ns-Wall attracts all the particles with same strength and s-Wall, depending on particle type, attracts with different strengths. Consequently, particle enrichment near the walls is more for ns-Wall type and thus expect to solidify at a higher temperature, and for stronger polymer-wall interaction strength $T^*$ is expected to increase further. Since our interest is in the fluid phase we focus our attention to $T\ge T^*$ (otherwise stated clearly). \\
In order to study the effect of flow field, we introduce a linear velocity profile defined as
\begin{equation}
v(y) = \frac{2 v_0}{D}\left(y-\frac{D}{2}\right)~,
\end{equation}
where $y$ varies from $0$ to $D$, as shown in figure~\ref{fig: fig1_schematic}. Furthermore, the upper(lower) walls are moving with velocity $+v_0(-v_0)$, and the shear rate is defined as
\begin{equation}
\dot{\gamma}=\frac{2v_0}{D}~.
\end{equation}
It is important to note that by imposing linear flow profile and moving the walls simultaneously we neglect the slipping near walls.
Typically, we relax the sample for $2.5\times 10^4\tau_{_{\rm LJ}}$ followed by shear for $5\times 10^5\tau_{_{\rm LJ}}$ corresponding to 100 million MD time steps. The results reported here are obtained in the steady state. The simulations are carried out using open source program LAMMPS.\cite{lammps}
\section{Density profile}
\label{sec: density-profile}
Typical steady state configurations of the system for $\dot{\gamma}=0.003$ at temperatures below, at, and above the transition, i.e., $\delta\equiv \frac{T-T^*}{T^*}=-0.1,0,0.1$, respectively are shown in figure~\ref{fig: simu-snapshot}. For $T>T^*$ the particles are homogeneously disributed in the film and with decrease in temperature the system becomes inhomogeneous (i.e., development of voids) and the particles with large $\epsilon_i$ values clumped together forming a core and smaller $\epsilon_i$ particles surrounding it. As observed in the equilibrium studies,\cite{Shagolsem2015_JCP} this particle identity ordering is expected at small and moderate shear rates. However, it is interesting to note that the difference in substrate quality has effect on the particle distribution near the interface.
\begin{figure}[ht]
\includegraphics*[width=0.4\textwidth]{figs1/fig2_schematic.png}
\begin{center}
\caption{Typical configurations of the 1C and APD systems in the steady state for shear rate $\dot{\gamma}=0.003$ shown for (A) $\delta=-0.1$, (B) $\delta=0$, and (C) $\delta=0.1$. Particles are colored according to their $\epsilon_i$ values, with $\epsilon_i=1$(blue) and $\epsilon_i=4$(red), see color scale bar.}
\label{fig: simu-snapshot}
\end{center}
\end{figure}
To understand the particle distribution in the film we calculate the averaged density profile at different positions of the film and compare with 1C system. In order to obtain the density $\rho$ as a function of distance $y$ from the walls, we divide the film into several layers and calculate the density in each layer. In figure~\ref{fig: apd-rho-compare}, we compare the density profiles of 1C and APD systems obtained at different values of shear rates and three different values of temperatures $\delta =-0.1,0,0.1$.
\begin{figure}[ht]
\includegraphics*[width=0.45\textwidth]{figs1/fig-rho-compare-1.pdf}
\begin{center}
\caption{Laterally averaged particle density $\rho$ at different positions in the film indicated by the normalize distance $y/D$. The walls are located at $y=0$ and $y=1$.}
\label{fig: apd-rho-compare}
\end{center}
\end{figure}
As expected, at high temperature the particle distribution is homogeneous and hence $\rho(y)$ is flat in the bulk, see figure~\ref{fig: apd-rho-compare} for $\delta=0.1$. However, in the wall-particle interface there is particle enrichment for both 1C and APD with non-selective walls, as shonw by the slight increase in $\rho(y)$ as $y$ approaches walls. Whereas decrease of $\rho(y)$ near walls is seen for APD with selective substrate indicating particle depletion. No shear rate dependence on the density profile is observed at this temperature. \\
However, at lower temperatures ($\delta=0,-0.1$) the difference in the profile is enhanced both in bulk and interface regions. For 1C and APD (ns-Wall) the density profile develops a minimum and at further lower temperature for APD (ns-Wall) and relatively high shear rates particles there is strong depletion near interface. It is interesting to note that for APD (s-Wall) the density profile remains qualitatively the same, i.e. maximum in the bulk and minimum at the interfaces, and weak dependence on the shear rate. In the following, we discuss the spatial ordering of particles and local identity ordering.
\section{Particle ordering and shear induced melting}
\label{sec: identity-ordering}
To characterize the spatial organization of particles we calculate the radial distribution function defined as
\begin{equation}
g(r)=\frac{1}{\rho_{_0} N}\left<\sum_{i}\sum_{j\ne i} \delta(r-r_{ij}) \right>~,
\end{equation}
where $\rho_{_0}$ is the particle number density in the system, and $r_{ij}=r_i-r_j$ distance between $i^{th}$ and $j^{th}$ particles.
\begin{figure}[ht]
\includegraphics*[width=0.45\textwidth]{figs1/fig-gr2d-compare-1.pdf}
\begin{center}
\caption{Radial distribution function $g(r)$ for (a) 1C system, (b) APD with ns-Wall, and (c) APD with s-Wall obtained at $\delta=-0.1$ for different values of shear rates $\dot\gamma$ indicated in the figure. In all the systems, $g(r)$ becomes liquid-like at high shear rate indicating shear-induced melting.}
\label{fig: gr2d}
\end{center}
\end{figure}
Since for $T>T^*$ or $\delta > 0$ the system is in liquid state we focus on $T<T^*$ (where the system is in solid-gas coexistence) in order to understand the effect of shear on spatial organization. In figure~\ref{fig: gr2d}, $g(r)$ is displayed for 1C and APD systems obtained at $\delta=-0.1$ for different values of shear rates. Higher order peaks corresponding to hexagonal packing are observed for small $\dot\gamma$ values which is gradually transformed to liquid-like ordering upon increasing $\dot\gamma$. Such shear-induced melting and distortion of $g(r)$ when subjected to shear-flow in simple liquids were observed in earlier works.\cite{hess1980,evans1984,hess1985} The associated effects in the transport coefficients due to structural reorganization is discussed in section~\ref{sec: macro-response}. \\
Another aspect of particle ordering is the identity of neighboring particles. Neighborhood identity ordering is quantified by means of effective interaction parameter of particle $i$, defined as
\begin{equation}
\epsilon_i^{\rm eff}=\frac {1}{n_b}\sum_{j=1}^{n_b}\epsilon_{ij}~,
\label{eqn: epsilon}
\end{equation}
where the sum over $j$ goes over all the $n_b$ neighboring particles within a cut-off radius $r_c=1.7$ which roughly corresponds to the minimum between the first and the second peaks of the radial distribution function. The value of $\epsilon_i^{\rm eff}$ ranges from 0 (for no neighbors within $r_c$) to 4. \\
\begin{figure}[ht]
\includegraphics*[width=0.5\textwidth]{figs1/fig_configs_compare-1.png}
\begin{center}
\caption{Typical steady state configurations of APD systems shown for three different values of shear rates $\dot{\gamma}\approx 0.0004,~0.0043,~0.0261$ at (a) $\delta=0$ and (b) $\delta=0.1$. Particles are colored according to their effective interaction parameter $\epsilon_i^{\rm eff}$ values. See color scale bar for $\epsilon_i^{\rm eff}$ values.}
\label{fig: configs-under-shear}
\end{center}
\end{figure}
In figure~\ref{fig: configs-under-shear}, we display the systems under shear, where particles are colored according to their $\epsilon_i^{\rm eff}$, shown for different shear rates at $\delta=0,~0.1$. One can see in the figure, for both selective and non-selective walls cases, there are regions of high $\epsilon_i^{\rm eff}$ surrounded by regions with low $\epsilon_i^{\rm eff}$ values. For $\delta>0$ particles are homogeneously distributed in the film (as discussed above) and observe smaller regions of relatively high $\epsilon_i^{\rm eff}$ values, and as $T\rightarrow T^*$ high $\epsilon_i^{\rm eff}$ regions gets bigger.
\begin{figure}[ht]
\includegraphics*[width=0.45\textwidth]{figs1/fig-eeff-vs-d-compare-1.pdf} \\ \vspace{0.2cm}
\includegraphics*[width=0.45\textwidth]{figs1/fig-apd-eeff-histo-compare-2.pdf}
\begin{center}
\caption{(Upper panel) Average effective interaction parameter $\epsilon_i^{\rm eff}$ as a function of normalized film thickness $y/D$ shown for non-selective wall (nsW) and selective wall (sW) types. (Lower panel) Distribution of $\epsilon_i^{\rm eff}$ at three different values of $\dot\gamma$ indicated in the figure at $\delta=-0.1,0$ for both nsW and sW.}
\label{fig: apd-eeff-compare1}
\end{center}
\end{figure}
To gain further insight we proceed to understand the spatial ordering of particles' identity at different locations of the film characterized by laterally averaged $\epsilon_i^{\rm eff}$ profile and $\epsilon_i^{\rm eff}$ distributions as shown in figure~\ref{fig: apd-eeff-compare1}. We observe that the laterally averaged $\epsilon_i^{\rm eff}$ profile closely follows that of the corresponding density profile shown in figure~\ref{fig: apd-rho-compare}.
For both APD systems, at a given temperature, when we increase shear rate the spatial distribution of particles as well as identity becomes homogeneous throught the film indicated by the flattening of the curve and thus the effect of substrate quality is insignificant at high shear rates.
As shown, for $\delta \approx 0$ the difference in the profiles at small shear rates vanishes at very high shear rate.
For $\delta > 0$, where the system is in fluid state, the profile is independent of $\dot \gamma$ (except for small differences close to the walls at small shear rates). On the other hand, $\epsilon_i^{\rm eff}$ distributions show that at high shear rates peak position is shifted to a lower value.
The mean value $\left< \epsilon_i^{\rm eff} \right>$ at different shear rates is plotted in figure~\ref{fig: apd-eeff-mean-compare} for APD systems. Slightly below the transition temperature, $\left< \epsilon_i^{\rm eff} \right>$ roughly remains constant for small shear rates and decreases beyond a threshold value as indicated in the figure for $\delta \approx -0.1$.
The decrease of $\left< \epsilon_i^{\rm eff} \right>$ with increasing shear rate is observed for all the temperatures considered. However, at high temperature (see $\delta \approx 1.5$ in the figure) the shear rate dependence of $\left< \epsilon_i^{\rm eff} \right>$ is rather weak. It is expected that the curves for relatively small $\delta$ (i.e., close to transition temperature $T^\ast$) will approach that of the high temperature of the respective systems at sufficiently high $\dot{\gamma}$ values, where the applied mechanical shear dominates over the interparticle interaction and the local identity ordering is completely destroyed.
Between the two APD systems no qualitative difference is observed, but only a quantitative difference, i.e., for a given shear rate $\left< \epsilon_i^{\rm eff} \right>$ is relatively higher for APD-system with selective walls and hence a relatively large $\dot\gamma$ where $\left< \epsilon_i^{\rm eff} \right>$ drops for $\delta \approx -0.1$. The observed decrease in mean effective interation parameter with increasing shear rate is associated with the structural changes in the system and is consequently reflected in the macroscopic response of the system discussed in the following section.
\begin{figure}[ht]
\begin{center}
\includegraphics*[width=0.45\textwidth]{figs1/fig-epsilon-eff-nsW-sW-compare1.pdf}
\caption{Two scale plot showing the variation of mean effective interaction parameter $\left< \epsilon_i^{\rm eff} \right>$ for APD system with nsW (left scale) and sW (right scale) obtained at different values of $\delta$ indicated in the figure. Open(close) symbols represent APD-nsW (APD-sW) system. Asterix marked points ($\dot\gamma\approx$ 0.013 and 0.017 for ns-W and s-W respectively) in the $\delta=-0.1$ curves indicates the shear rate above which $\left< \epsilon_i^{\rm eff} \right>$ decreases.}
\label{fig: apd-eeff-mean-compare}
\end{center}
\end{figure}
\section{Macroscopic response}
\label{sec: macro-response}
When mechanical shear is introduced to simple fluids, a wide variety of intersting effects such as shear induced ordering and phase transition, and non-Newtonian rheological behavior, i.e., shear thinning/thickening are observed.\cite{hess1980,hess1985,hess1994,hess1997,evans1984,evans2004} For the present APD fluid system also, the rheological behavior or the macroscopic response under shear is studied. In particular, we calculate shear viscosity $\eta_{xy}$ defined as
\begin{equation}
\eta_{xy}=\frac{\sigma_{xy}}{\dot\gamma}~,
\label{eq: shear-visco}
\end{equation}
with $\sigma_{xy}$ shear stress,
in order to see whether there is any qualitative and quantitative differences among the systems. For Newtonian fluid shear viscosity is independent of shear rate, i.e.,
\begin{equation}
\eta_{xy}\sim\dot{\gamma}^0~.
\end{equation}
On the other hand, for non-Newtonian fluid
\begin{equation}
\eta_{xy}\sim \dot{\gamma}^\alpha~,
\end{equation}
in general, with $\alpha<0$ for shear-thinning and $\alpha>0$ for shear-thickening.
\begin{figure}[ht]
\begin{center}
\includegraphics*[width=0.45\textwidth]{figs1/fig-shear-visco-nsw-with-config1.png}
\caption{Shear viscosity $\eta$ against the imposed steady shear rate $\dot\gamma$ for APD system with non-selective wall at four different temperatures indicated in the figure. Subfigures maked (a)-(c) are the corresponding steady state configurations for the $\bigotimes$ marked points in the curve for $\delta=-0.1$.}
\label{fig: shear-visco-nsw}
\end{center}
\end{figure}
\begin{figure}[ht]
\begin{center}
\includegraphics*[width=0.45\textwidth]{figs1/visco-compare1.pdf}
\caption{Comparison of shear viscosity $\eta_{xy}$ profiles among different systems indicated in the figure for (a) $\delta=-0.1$, (b) $\delta=0$, and (c) $\delta=0.1$. Vertical dashed line is guide to eye to indicate the onset shear thinning for APD system with ns-Walls.}
\label{fig: shear-visco-compare}
\end{center}
\end{figure}
Typical shear viscosity profiles of the APD system (with non-selective substrates) close to transition temperature is shown in figure~\ref{fig: shear-visco-nsw}. For $T<T^*$ $(\delta=-0.1~\text{in the figure})$, it is observed that the shear viscosity $\eta_{xy}$ decreases with increasing $\dot\gamma$ (shear-thinning) which is followed by an increase in $\eta_{xy}$ on further increasing $\dot\gamma$ (shear-thickening) and finally a decrease again at high $\dot\gamma$ values (shear-thinning).
The shear-thickening happens in a relatively short range of $\dot\gamma$, and the onset of shear-thickening shifts towards smaller $\dot\gamma$ values as $T\rightarrow T^*$, e.g., the onset is observed at $\dot\gamma\approx0.013$ for $\delta=-0.1$ as indicated in the figure. Further, it is interesting to note that the onset of shear-thickening corresponds to the decrease in average interaction parameter $\left< \epsilon_i^{\rm eff} \right>$, see figure~\ref{fig: apd-eeff-mean-compare} for $\delta=-0.1$ (nsW), and thus reflects the structural as well as local particle identity ordering changes with increasing shear rate. Steady-state configurations of the systems around the shear-thickening region is shown in figure~\ref{fig: shear-visco-nsw} as inset. It is evident from the configurations that before the onset of shear-thickening no major identity reorganization takes places and thus $\left< \epsilon_i^{\rm eff} \right>$ is roughly constant (e.g., see figure~\ref{fig: apd-eeff-mean-compare} for $\delta<0$). Comparison of subfigures -(a) and -(b) of figure~\ref{fig: shear-visco-nsw} clearly shows that a strong spatial and identity ordering happens during the shear-thickening, where we see particle clustering (formation of denser phase composed predominantly of high-$\epsilon_i^{\rm eff}$ particles). Similar behavior is observed during shear-thickening in colloidal systems.\cite{brown2014} The second shear-thinning corresponds to gradual melting and hence decrease in $\left< \epsilon_i^{\rm eff} \right>$ leading to a completely (identity) disordered fluid phase. When $\delta=0$ only shear-thinning is observed, and for $\delta>0$ shear-viscosity very weakly depends on the shear rate, i.e., $\eta_{xy}\sim \dot\gamma^0$. \\
In figure~\ref{fig: shear-visco-compare}, we show comparison among the systems at three different temperatures. Both shear-thinning and shear-thickening is observed for temperature slightly below $T^*$. As shown in figure~\ref{fig: shear-visco-compare}(a), at very small shear rates both APD-nsW and 1C systems show $\eta_{xy} \sim \dot{\gamma}^0$, and first shear-thinning ($\eta_{xy} \sim \dot{\gamma}^{-\alpha}$) sets in above a threshold value $\dot\gamma_c \approx 0.0025$ for APD-nsW (with exponent $\alpha \approx 1$) and $\dot\gamma \approx 0.004$ for 1C. On the other hand, for APD-sW system first shear-thinning is observed for shear rate upto $\dot\gamma \approx 0.01$ with exponent $\alpha\approx 0.33$. Comparing the exponent $\alpha$ it is clear that the shear-thinning is relatively strong for APD-nsW system.
Upon increasing $\dot\gamma$ further shear-thickening is observed followed by second shear-thinning where the decrease of $\eta_{xy}$ with $\dot\gamma$ is first seen for APD-nsW and 1C systems followed by APD-sW system; however, all the systems have $\alpha\approx 0.78$. We note that as $\delta\rightarrow 0$ the onset of first shear-thinning shifts towards smaller value of $\dot\gamma$. At $\delta\approx 0$, the shear-thinning exponent $\alpha \approx 0.42$, see figure~\ref{fig: shear-visco-compare}(b). As $\delta$ becomes positive and gets larger (i.e. fluid phase further away from $T^*$) the exponent $\alpha\rightarrow 0$ and show Newtonian-like behavior for all systems.
\section{Summary and discussion}
\label{sec: conclusion}
We have studied the effect of linear flow field on the spatial and particle identity ordering, density profiles, and macroscopic response of a multi-component fluid model. All the particles in the system are different in the sense that each particle has different interaction strength $\epsilon_i$ drawn from a uniform distribution. In equilibrium, this multi-component fluid model system exhibit local self-organization and clustering of particles according to their identity, i.e. a non-random state, leading to different dynamical properties for different particle types. When the system is subjected to steady flow field, depending on the shear rate, it displays various interesting features. We focus on temperature close to the solidification transition and investigated the system at various shear rates considering non-selective and selective substrate types. The results are briefly summarized below.
The particle density profiles of the two APD systems is significantly different, e.g., in contrast to APD (ns-wall), the density profile for APD (s-Wall) has maxima at the center and minima (or particle depletion) near the substrates. As we increase the shear rate particle density becomes homogeneous and no difference is seen between the two APD-systems in the bulk, however close to the substrate significant difference in the density is observed enven at high shear rates. It is interesting to note that the density profiles of APD(ns-wall) and 1C systems are very similar.\\
Further study of radial distribution function reveal that this homogenization is related to the shear-induced melting, where higher order peaks (corresponding to hexagonal packing) in $g(r)$ disappears at high shear rates.
Apart from structural changes, one of the main concern of this study is to understand how the neighborhood identity ordering is affected due to shear. As detailed in section~\ref{sec: identity-ordering}, the neighborhood identity ordering is characterized through $\epsilon_i^{\rm eff}$. The laterally averaged $\epsilon_i^{\rm eff}$ profiles closely follow that of the corresponding density profiles, where the value of $\epsilon_i^{\rm eff}$ is higher in higher density regions, see figure~\ref{fig: apd-eeff-compare1}. At high shear rates, in consistent with the homogenization observed, the profile is flat and $\epsilon_i^{\rm eff}\rightarrow 2.5$ in all the regions (except for small deviations near the substrates). Furthermore, the $\epsilon_i^{\rm eff}$ distribution show that the mean of the distribution is shifted to lower values with increasing shear rates and hence $\left<\epsilon_i^{\rm eff}\right>$ also decreases. We found only quantitative difference between the systems with selective and non-selective substrates, i.e., $\left<\epsilon_i^{\rm eff}\right>$ for system with selective-walls is consistently higher in the range of $\dot\gamma$ considered. It is clear from this observation that the local particle identity ordering is affected by the quality of substrates and it is reflected in the global average quantity (i.e., selective substrate type leads to higher value of $\left<\epsilon_i^{\rm eff}\right>$).
Finally, we looked at the behavior of shear viscosity, $\eta_{xy}$, for both systems. Below the transiton temperature both energy polydispersed and 1C systems show shear-thinning as well as shear-thickening regimes, and shear-thickening is more pronounced with temperature farther away from the transition point. Shear-thickening is also associated with transient increase of $\left<\epsilon_i^{\rm eff}\right>$. Shear-thickening is followed by shear-thinning again leading to shear-induced melting at high shear rates. Exponent characterising shear-thinning has braod distribution and depends on temperature. It is important to point out that APD fluid with selective (non-selective) substrate has the shear viscosity below (or above depending on $T$) that of 1C fluid indicating that by tuning substrate or wall properties one can tune shear viscosity of this multi-component fluid. In conclusion, we present a comprehensive study on the energy polydispersed fluid models in the presence of flow field and we believe that this model, apart from biological systems, is also very relevant to industrial mixing of materials containing many energetically different component.
\begin{acknowledgements}
LS acknowledge fruitful discussion with Yitzhak Rabin.
\end{acknowledgements}
|
\section{Introduction}
Hyperspectral (HS) cameras are able to acquire images with very high spectral resolution. However, the fundamental compromise between signal-to-noise ratio, spatial resolution, and spectral resolution means that their spatial resolution is usually low~\cite{shaw2003reviewSpecImaging}. Multispectral (MS) devices, on the other hand, are able to achieve a much higher spatial resolution since they contain only a small number of spectral bands.
An approach that attempts to circumvent the physical limitations of imaging sensors consists in combining HS and MS images (MSI) of the same scene to obtain images with high spatial and spectral resolution, in a multimodal image fusion problem~\cite{yokoya2017HS_MS_fusinoRev}, commonly referred to as \emph{hyperspectral super resolution}.
Different algorithms have been proposed to solve this problem. Early approaches were based on component substitution or on multiresolution analysis, which attempt to extract high-frequency spatial details from the MSI and combine them with the HS image (HSI)~\cite{carper1990componentSubstitutionPansharpening,liu2000MRA_HS_MS_fusion}.
Subspace-based formulations have later received a significant amount of interest as they explore the natural representation of the pixels in an HSI as the linear combination of a small number of spectral signatures~\cite{yokoya2017HS_MS_fusinoRev,Keshava:2002p5667,borsoi2018superpixels1_sparseU}.
Different algorithms have been proposed following this approach using, e.g., Bayesian formulations~\cite{hardie2004hs_ms_fusionMAP} or sparse representations on learned dictionaries~\cite{wei2015hs_ms_fusionBayesianSparse}, and different kinds of matrix factorization formulations employing sparse and spatial regularizations~\cite{kawakami2011hs_ms_fusionMFsparse,simoes2015HySure}, or estimating both the basis vectors and their coefficients blindly/unsupervisedly from the images~\cite{yokoya2012coupledNMF}.
The natural representation of HSIs and MSIs as 3-dimensional tensors has been successfully exploited for hyperspectral unmixing, denoising~\cite{imbiriba2018ULTRA_V,veganzones2015tensorDecompMultisngularHSIs,imbiriba2018_ULTRA,liu2012denoisingHSIsTensorCPD} and super-resolution. Superior super-resolution performance and exact recovery guarantees have been obtained using this formulation~\cite{kanatsoulis2018hyperspectralSRR_coupledCPD,prevost2020coupledTucker_hyperspectralSRR_TSP}.
The image fusion problem was formulated in~\cite{kanatsoulis2018hyperspectralSRR_coupledCPD} as a coupled tensor approximation problem. Assuming that the high resolution (HR) image admits a low-rank canonical polyadic decomposition (CPD), the problem was solved using an alternating optimization strategy. The recovery of the correct HR image (HRI) was shown to be guaranteed provided that the CPD of the MSI is identifiable, and state of the art performance was achieved.
A recent work extended this approach by assuming the high resolution images to follow a block term decomposition (BTD), which shows a closer connection to the physical mixing model when compared to the CPD~\cite{zhang2019hyperspectralSRR_coupledBTD}.
A simpler approach was later proposed in~\cite{kanatsoulis2018hyperspSRR_lowRankMatrixTensor} by requiring only the computation of one CPD of the MSI and a singular value decomposition (SVD) of the HSI.
A Tucker decomposition-based approach was later considered in~\cite{prevost2019coupledTucker_hyperspectralSRR,prevost2020coupledTucker_hyperspectralSRR_TSP}. Closed form SVD-based algorithms were proposed for the image fusion problem, achieving results comparable to~\cite{kanatsoulis2018hyperspectralSRR_coupledCPD} at a very small computational complexity; exact recovery guarantees were also provided. A coupled Tucker approximation was also considered in~\cite{li2018hs_ms_fusionTensorFactorization} using an alternating optimization approach and employing a sparsity regularization on the elements of the core tensor. {This work was later extended by incorporating the piecewise smoothness of the reconstructed image tensor by using a Total Variation regularization along each of its modes~\cite{xu2020HSMSfusion_1DTV}.}
Another approach considered the CPD of non-local similar patch tensors to explore the non-local redundancy of the image~\cite{xu2019nonlocalTensorDecompImageFusion}.
{A different non-local approach was also proposed in~\cite{dian2019hyperspectralSRR_tensorMultiRank} by using a recent definition of the nuclear norm of order-3 tensors.}
Most existing algorithms, however, share a common limitation: they assume that the HSI and the MSI are acquired under the same conditions.
However, despite the short revisit cycles provided by the increasing number of optical satellites orbiting the Earth (e.g. Sentinel, Orbview, Landsat and Quickbird missions), the number of platforms carrying both HS and MS sensors is still considerably limited~\cite{eckardt2015DESIS_satelliteSpecs,kaufmann2006EnMAP_satelliteSpecs}. This makes combining HS and MS observations acquired on board of different satellites of great interest to obtain HRIs~\cite{hilker2009fusion_MODIS_denseTimeSeries_exp2,emelyanova2013fusion_MODIS_denseTimeSeries_exp}.
Images acquired at different time instants can be impacted by, e.g., illumination, atmospheric or seasonal changes. This may result in significant variations between the HSI and the MSI~\cite{borsoi2020variabilityReview}, negatively impacting traditional image fusion algorithms.
Recently, a method was proposed to combine HSIs and MSIs accounting for seasonal {(inter-image)} spectral variability~\cite{Borsoi_2018_Fusion}. Using a low-rank matrix formulation, the set of spectral basis vectors of the HRIs underlying the HS and the MS observations are allowed to be different from each other, with variations introduced by a set of multiplicative scaling factors~\cite{imbiriba2018glmm}. This algorithm led to significant performance improvements when the HSI and MSI are subject to spatially uniform seasonal or acquisition variations. However, it does not account for spatially localized changes commonly seen in practical scenes~\cite{borsoi2020variabilityReview}. Moreover, the algorithm in~\cite{Borsoi_2018_Fusion} presented high computation times and does not offer any theoretical guarantees.
In this paper, we propose a tensor-based image fusion formulation that accounts for localized spatial and spectral changes between the HSI and MSI. A general observation model is considered, in which the HRI underlying the MSI admits an additive variability term to account for changes between the scenes. Studying the general identifiability of this model, we show that this variability term can only be identified in general up to its smooth structure (which is defined according to the degradation operators). To introduce additional a priori information and mitigate the ambiguity associated with the proposed model, both the HRI and the additive perturbations are assumed to have low multilinear rank (i.e., to admit a Tucker decomposition). Two algorithms are then proposed, one totally algebraic and another based on an optimization procedure. Theoretical guarantees for the exact recovery of the HRI are provided for both. Simulation results show that the proposed optimization-based algorithm yields superior performance at a considerably lower computational cost when compared to~\cite{Borsoi_2018_Fusion}, especially when spatially localized variability is considered.
\section{Tensors -- background} \label{sec:background}
\subsection{Notation and definitions}
\label{sec:back_defs}
An order-3 tensor $\mathcal{T}\in\amsmathbb{R}^{N_1\times N_2 \times N_3}$ is an $N_1\times N_2 \times N_3$ array whose elements are indexed by~$[\mathcal{T}]_{n_1,n_2,n_3}$. Each dimension of \textcolor{darkgreen}{an order-3 tensor} is called a \emph{mode}. A mode-$k$ \emph{fiber} of tensor $\mathcal{T}$ is the one-dimensional subset of $\mathcal{T}$ which is obtained by fixing \textcolor{darkgreen}{all but one of the three modes -- the $k$-th dimension}. Similarly, a \emph{slab} or \emph{slice} of a tensor $\mathcal{T}$ is a matrix whose elements are the two-dimensional subset of $\mathcal{T}$ obtained by fixing all but two of its modes. Operator $\operatorname{vec}(\cdot)$ represents the standard matrix column-major vectorization, or tensor vectorization. The (left) pseudo-inverse of matrix $\boldsymbol{X}$ is denoted by $\boldsymbol{X}^\dagger$. We denote scalars by lowercase ($x$) or uppercase ($X$) plain font, vectors and matrices by lowercase ($\boldsymbol{x}$) and uppercase ($\boldsymbol{X}$) bold font, respectively, and order-3 tensors by calligraphic plain font ($\mathcal{T}$) or using the blackboard Greek alphabet ($\mathcal{\Theta}$). In the following, we review some useful operations of multilinear (tensor) algebra that will be used in the rest of the manuscript (see, e.g.,~\cite{cichocki2015tensor,kolda2009tensor} for more details).
\begin{definition}
\noinden
{\color{red}The \textbf{mode-$k$ product} between a tensor $\mathcal{T}$ and a matrix $\cb{B}$ produces a tensor $\mathcal{U}$ that is evaluated such that each mode-$k$ fiber of $\mathcal{T}$ is multiplied by $\cb{B}$.
For instance, the mode-2 product between $\mathcal{T} \in \amsmathbb{R}^{N_1\times N_2\times N_3}$ and $\cb{B} \in \amsmathbb{R}^{M_2\times N_2}$ produces a tensor
$\mathcal{U}\in\amsmathbb{R}^{N_{1}\times M_2\times N_{3}}$, denoted by
$\mathcal{U}=\mathcal{T}\times_2\cb{B}$.}
{\color{magenta}
Its elements are accessed as $[\mathcal{U}]_{n_1,m_2,n_3}=\sum_{i=1}^{N_2}[\mathcal{T}]_{n_1,i,n_3} [\cb{B}]_{m_2,i}\ , \ m_2=1,\ldots,M_2$.}
\end{definition}
{\color{magenta}Note that the mode-$k$ product has the following properties:
\begin{align}
\mathcal{T} \times_i\boldsymbol{A} \times_j\boldsymbol{B} &= \mathcal{T} \times_j\boldsymbol{B} \times_i \boldsymbol{A}\,, \ i \neq j \,,
\label{eq:mode_k_prop1}
\\
\mathcal{T} \times_k \boldsymbol{A} \times_k\boldsymbol{B} &= \mathcal{T}\times_k\big(\boldsymbol{A}\boldsymbol{B} \big) \,.
\label{eq:mode_k_prop2}
\end{align}}
\begin{definition}
\textcolor{darkgreen}{The full \textbf{multilinear product} is denoted by $\big\ldbrack\mathcal{T};\boldsymbol{B}_{1},\boldsymbol{B}_{2},\boldsymbol{B}_{3}\big\rdbrack$, and consists of a series of successive mode-$k$ products, for {$k\in\{1,2,3\}$, between a tensor $\mathcal{T}$ and matrices $\boldsymbol{B}_1$, $\boldsymbol{B}_2$ and $\boldsymbol{B}_3$,} respectively, and {is} expressed as $\mathcal{T}\times_1\boldsymbol{B}_{1}\times_2\boldsymbol{B}_{2}\times_3\boldsymbol{B}_{3}$.}
\end{definition}
\begin{definition}
The \textbf{mode-$k$ matricization} of an order-3 tensor $\mathcal{T}\in\amsmathbb{R}^{N_1\times N_2\times N_3}$, denoted by $\unfold{\boldsymbol{T}}{k}$, arranges its mode-$k$ fibers to be the columns of the resulting matrix $\unfold{\boldsymbol{T}}{k}\in\amsmathbb{R}^{N_k\times N_{\ell}N_{m}}$, $k,\ell,m\in\{1,2,3\}$, $k \neq \ell \neq m$, where the $n_k$-th row of $\unfold{\boldsymbol{T}}{k}$ consists of the vectorization of the slice of $\mathcal{T}$ obtained by fixing the index of the $k$-th mode of $\mathcal{T}$ as $n_k$.
\end{definition}
\begin{definition}
We define by $\operatorname{tSVD}_R(\boldsymbol{X})$ the operator which returns a matrix containing {\color{red} the $R$ left singular vectors associated with the largest singular values of the matrix $\boldsymbol{X}$}
\end{definition}
\subsection{Tensor decompositions}
{The Tucker decomposition is able to represent an order-3 tensor $\mathcal{T}\in\amsmathbb{R}^{N_1\times N_2\times N_3}$ compactly, using a set of factor matrices given by $\boldsymbol{B}_{i}\in\amsmathbb{R}^{N_i\times K_i}$, $i\in\{1,2,3\}$ and a small core tensor $\mathcal{G}\in\amsmathbb{R}^{K_1\times K_2\times K_3}$, as~\cite{kolda2009tensor}
\begin{align}
\mathcal{T} &= \big\ldbrack \mathcal{G};\boldsymbol{B}_{1},\boldsymbol{B}_{2},\boldsymbol{B}_{3} \big\rdbrack \,.
\label{eq:tensorDef_Tucker}
\end{align}
The tuple $(K_1,K_2,K_3)$ is called the multilinear rank of $\mathcal{T}$. Each value $K_i$ is also equal to the rank of the mode-$i$ unfolding of~$\mathcal{T}$~\cite{cichocki2015tensor}.}
{\color{magenta}
The Tucker decomposition
allows the rank along each mode of the tensor to be different (i.e., $K_i\neq K_j$)~\cite{sidiropoulos2017tensor}. This property can be very useful since it allows one to set a higher rank to specific modes of the decomposition in order to adequately represent the data diversity while still keeping the \mbox{model low rank.}
}
The matricizations and vectorization of a tensor $\mathcal{T}$ following the Tucker decomposition~\eqref{eq:tensorDef_Tucker} are given by~\cite{kolda2009tensor}:
\begin{align}
\operatorname{vec}(\mathcal{T}) &= (\boldsymbol{B}_{3} \otimes \boldsymbol{B}_{2} \otimes \boldsymbol{B}_{1} ) \operatorname{vec}(\textcolor{magenta}{\mathcal{G}}) \,,
\label{eq:gen_tensor_vec}
\\
\unfold{\boldsymbol{T}}{1} &= \boldsymbol{B}_{1} \unfold{\textcolor{magenta}{\boldsymbol{G}}}{1} (\boldsymbol{B}_{3} \otimes \boldsymbol{B}_{2})^\top \,,
\label{eq:gen_tensor_mat_1}
\\
\unfold{\boldsymbol{T}}{2} &= \boldsymbol{B}_{2} \unfold{\textcolor{magenta}{\boldsymbol{G}}}{2} (\boldsymbol{B}_{3} \otimes \boldsymbol{B}_{1})^\top \,,
\label{eq:gen_tensor_mat_2}
\\
\unfold{\boldsymbol{T}}{3} &= \boldsymbol{B}_{3} \unfold{\textcolor{magenta}{\boldsymbol{G}}}{3} (\boldsymbol{B}_{2} \otimes \boldsymbol{B}_{1})^\top \,.
\label{eq:gen_tensor_mat_3}
\end{align}
{The Tucker decomposition can be computed using fast algorithms such as the high-order SVD~\cite{lathauwer2000multilinearHOSVD}.}
{Another classic tensor decomposition is the Canonical Polyadic Decomposition (CPD)~\cite{kolda2009tensor}. The CPD enjoys uniqueness properties under very mild conditions, and its rank can exceed the dimensions of the tensor. However, it is also more difficult to compute, and the same rank value is used to represent all modes of the tensor.}
{\color{magenta}
The Block Term Decomposition (BTD) generalizes CPD and Tucker, and allows us to combine benefits from both approaches~\cite{lathauwer2008tensor_BTD2_uniqueness}.}
Specifically, the BTD of an order-3 tensor $\mathcal{T}$ is defined as a sum of rank-$(K_{1,r},K_{2,r},K_{3,r})$ terms as~\cite{lathauwer2008tensor_BTD2_uniqueness}:
\begin{align} \label{eq:tensor_BTD_i}
\mathcal{T} & = \sum_{r=1}^R \big\ldbrack\mathcal{G}_r;\boldsymbol{B}_{r,1},\boldsymbol{B}_{r,2},\boldsymbol{B}_{r,3} \big\rdbrack \,,
\end{align}
where $\mathcal{G}_r$, $\boldsymbol{B}_{r,1}$, $\boldsymbol{B}_{r,2}$, and $\boldsymbol{B}_{r,3}$, for $r=1,\ldots,R$, are the core tensors and the factors corresponding to each mode of $\mathcal{T}$. {Differently from the Tucker decomposition, the BTD additionally requires the selection of parameter $R$ (number of blocks), and can be more costly to compute.}
{\color{magenta}
However, the BTD benefits from uniqueness results which, although not as strong than those of the CPD, are still interesting for many applications~\cite[Section~5]{lathauwer2008tensor_BTD2_uniqueness}.}
{These results will prove very important to derive the recoverability guarantees for the algorithm in Section~\ref{sec:alg2_optimizationBased}.
Note that in the CPD can be viewed as a special case of the BTD with $R=1$.}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/diagram_tmp2b2.pdf}
\vspace{-0.7cm}
\caption{{Proposed imaging model. The observed HSI and MSI are acquired under different conditions or at different time instants. Their underlying HRIs, represented by tensors $\mathcal{Z}_h$ and $\mathcal{Z}_m$, respectively, can be different from each other due to the effect of spectral and acquisition variations, as well as scenery changes. The changes occurring between the HRIs $\mathcal{Z}_h$ and $\mathcal{Z}_m$ are represented using an additive tensor $\mathcal{\Psi}$ which captures variability and changes, and both the HRI and the variability tensors are assumed to have low Tucker rank.}}
\label{fig:gen_diagram}
\end{figure}
\section{Proposed model and its undeterminacies}
\label{sec:imaging_mdl_identif}
\subsection{The imaging model}
Let an HSI with high spectral resolution and low spatial resolution be represented as an order-3 tensor $\mathcal{Y}_h\in\amsmathbb{R}^{N_1\times N_2\times L_h}$, where $N_1$ and $N_2$ are the spatial and $L_h$ the spectral dimensions.
Similarly, an MSI with high spatial and low spectral resolution is denoted by an order-3 tensor $\mathcal{Y}_m\in\amsmathbb{R}^{M_1\times M_2\times L_m}$, where $M_1>N_1$ and $M_2>N_2$ are the spatial and $L_m<L_h$ the spectral dimensions.
Both the HSI and the MSI are assumed to be degraded versions of a tensor $\mathcal{Z}\in\amsmathbb{R}_+^{M_1\times M_2\times L_h}$, with high spectral and spatial resolutions.
This degradation process is commonly described as~\cite{kanatsoulis2018hyperspectralSRR_coupledCPD,prevost2019coupledTucker_hyperspectralSRR,prevost2020coupledTucker_hyperspectralSRR_TSP,li2018hs_ms_fusionTensorFactorization}:
\begin{align}
\mathcal{Y}_h &= \mathcal{Z} \times_1 \boldsymbol{P}_1 \times_2 \boldsymbol{P}_2 + \mathcal{E}_h \,,
\label{eq:tensor_obs_mdl_hsi_novariability}
\\
\mathcal{Y}_m &= \mathcal{Z} \times_3 \boldsymbol{P}_3 + \mathcal{E}_m \,,
\label{eq:tensor_obs_mdl_msi_novariability}
\end{align}
where tensors $\mathcal{E}_m\in\amsmathbb{R}^{M_1\times M_2\times L_m}$ and $\mathcal{E}_h\in\amsmathbb{R}^{N_1\times N_2\times L_h}$ represent additive noise.
The matrix $\boldsymbol{P}_3\in\amsmathbb{R}^{L_m\times L_h}$ contains the spectral response functions (SRF) of each band of the multispectral sensor, and matrices $\boldsymbol{P}_1\in\amsmathbb{R}^{N_1\times M_1}$ and $\boldsymbol{P}_2\in\amsmathbb{R}^{N_2\times M_2}$ represent the spatial blurring and downsampling in the hyperspectral sensor, which we assume to be separable for each spatial dimension as previously done in, e.g.,~\cite{kanatsoulis2018hyperspectralSRR_coupledCPD,prevost2019coupledTucker_hyperspectralSRR,prevost2020coupledTucker_hyperspectralSRR_TSP,li2018hs_ms_fusionTensorFactorization}.
{Note that since the mode-$k$ product obeys~\eqref{eq:mode_k_prop1}, the choice of the ordering of $\boldsymbol{P}_1$ and $\boldsymbol{P}_2$ in~\eqref{eq:tensor_obs_mdl_hsi_novariability} does not affect the result.}
To make notation more convenient, we also denote the (linear) spatial and spectral degradation operators more compactly as
\begin{align}
\mathscr{P}_{1,2}(\mathcal{T}) &= \mathcal{T} \times_1 \boldsymbol{P}_1 \times_2 \boldsymbol{P}_2 \,,
\label{eq:operatorNotation_a}
\\
\mathscr{P}_{3}(\mathcal{T}) &= \mathcal{T} \times_3 \boldsymbol{P}_3 \,.
\label{eq:operatorNotation_b}
\end{align}
Most previous works consider that $\mathcal{Y}_h$ and $\mathcal{Y}_m$ are acquired under the same conditions, implicitly assuming that no variability {or changes} occur between the images.
{However, when the HSI and MSI are not acquired from the same mission/instrument and at the same time, the scene which underlies the (degraded) observations $\mathcal{Y}_h$ and $\mathcal{Y}_m$ can be subject to significant changes, referred to as {\color{magenta} inter-image} variability\footnote{\color{magenta} Here, variability should not be confused with the spectral variability considered in \cite{borsoi2020variabilityReview}, which focuses on inter-pixel endmembers variations.}, which include spatial and spectral variations as illustrated in Fig.~\ref{fig:gen_diagram}. Spectral variations originate from different \textcolor{magenta}{chemical}, atmospheric, illumination or seasonal conditions between the scenes~\cite{somers2011variabilityReview,Zare-2014-ID324-variabilityReview,borsoi2020variabilityReview}, and are typical even under short acquisition time differences. Spatial variations, on the other hand, occur due to, e.g., some regions of the scene being affected unequally by seasonal effects (which are strongly material-dependent~\cite{borsoi2020variabilityReview}) or due to the sudden insertion/removal of an object~\cite{liu2019reviewCD_GRSM}. Spatial variations can be very prominent when large acquisition time differences are considered.
These effects are not accounted for in the majority of the existing algorithms, what motivates the development of more flexible models.
}
Recently, spatially uniform spectral variability has been considered in~\cite{Borsoi_2018_Fusion}. The image fusion problem was formulated as a matrix factorization problem, and the (multiplicative) spectral variability as well as the spatial coefficients were estimated from the observed images.
However, this work still did not address two fundamental problems: 1) How to account for both spatial and spectral variability and 2) what theoretical guarantees can be offered for the recovery of the HRI and (possibly) of the variability factors under these more challenging conditions.
To address these issues, we adopt a more general approach by considering two different HRIs $\mathcal{Z}_h\in\amsmathbb{R}_+^{M_1\times M_2\times L_h}$ and $\mathcal{Z}_m\in\amsmathbb{R}_+^{M_1\times M_2\times L_h}$, both with high spectral and spatial resolutions, underlying the observed HSI and the MSI, respectively. This leads to the following extension of model~\eqref{eq:tensor_obs_mdl_hsi_novariability}--\eqref{eq:tensor_obs_mdl_msi_novariability}:
\begin{align}
\mathcal{Y}_h &= \mathscr{P}_{1,2}(\mathcal{Z}_h) + \mathcal{E}_h \,,
\label{eq:tensor_obs_mdl_hsi_0}
\\
\mathcal{Y}_m &= \mathscr{P}_{3}(\mathcal{Z}_m) + \mathcal{E}_m \,.
\label{eq:tensor_obs_mdl_msi_0}
\end{align}
{\color{magenta} The tensors $\mathcal{Z}_h$ and $\mathcal{Z}_m$ \textcolor{darkgreen}{represent} the underlying HRIs of the observed scene under different acquisition conditions and at possibly different times. To account for inter-image variability, the HRIs are related to each other as follows:}
\begin{align}
\mathcal{Z}_m = {\mathcal{Z}_h} + \mathcal{\Psi},
\label{eq:hr_hsi_msi_add_mdl_0}
\end{align}
where $\mathcal{\Psi}\in\amsmathbb{R}^{M_1\times M_2\times L_h}$ is an additive variability tensor representing changes between the scenes.
{\color{magenta}By introducing $\mathcal{\Psi}$, model~\eqref{eq:hr_hsi_msi_add_mdl_0} makes the inter-image variability between the HSI and the MSI explicit.}
Considering the variability model~\eqref{eq:hr_hsi_msi_add_mdl_0} along with~\eqref{eq:tensor_obs_mdl_hsi_0}--\eqref{eq:tensor_obs_mdl_msi_0}, we obtain the following observation model for the acquired HSI and MSI:
\begin{align}
\mathcal{Y}_h &= \mathscr{P}_{1,2}({\mathcal{Z}_h}) + \mathcal{E}_h \,,
\label{eq:tensor_obs_mdl_hsi}
\\
\mathcal{Y}_m &= \mathscr{P}_{3}({\mathcal{Z}_h} + \mathcal{\Psi}) + \mathcal{E}_m.
\label{eq:tensor_obs_mdl_msi_add}
\end{align}
{\color{magenta}In the following, $\mathcal{Z}_h$ and $\mathcal{\Psi}$ will be referred to as the HRI and the variability tensor, respectively.}
\subsection{The image fusion problem and its undeterminacies}
The image fusion problem in this case consists in recovering~$\mathcal{\Psi}$ and ${\mathcal{Z}_h}$ from the observed images $\mathcal{Y}_h$ and $\mathcal{Y}_m$. More precisely,
\begin{align} \label{eq:HS_MS_fus_var_prob}
\Bigg\{
\begin{split}
& \text{find } {\mathcal{Z}_h}\in\Omega_Z \text{ and } \mathcal{\Psi}\in\Omega_{\Psi},
\\
& \text{such that equations~\eqref{eq:tensor_obs_mdl_hsi}--\eqref{eq:tensor_obs_mdl_msi_add} are satisfied.}
\end{split}
\end{align}
The sets $\Omega_Z\subseteq\amsmathbb{R}^{M_1\times M_2\times L_h}$ and $\Omega_{\Psi}\subseteq\amsmathbb{R}^{M_1\times M_2\times L_h}$ denote prior information about the HRI and the variability factor, respectively.
Since the number of unknowns is significantly greater than the number of observations, problem~\eqref{eq:HS_MS_fus_var_prob} is severely ill-posed and additional \emph{a priori} information about the structure of ${\mathcal{Z}_h}$ and $\mathcal{\Psi}$ must be introduced through the sets $\Omega_Z$ and $\Omega_{\Psi}$ in order to obtain a stable recovery.
Common information that has been used to construct $\Omega_Z$ includes spatial \mbox{(piecewise-)} smoothness~\cite{simoes2015HySure}, low matrix (spectral) rank~\cite{yokoya2012coupledNMF,Borsoi_2018_Fusion}, low tensor rank~\cite{kanatsoulis2018hyperspectralSRR_coupledCPD,prevost2020coupledTucker_hyperspectralSRR_TSP,xu2019nonlocalTensorDecompImageFusion}, and non-local spatial information~\cite{li2018hs_ms_fusionTensorFactorization}.
The choice of prior information in $\Omega_Z$ and $\Omega_{\Psi}$ turns to the question of whether assuming additional structure over the pair $({\mathcal{Z}_h},\mathcal{\Psi})$ makes these variables \emph{identifiable} from the observations $(\mathcal{Y}_{h},\mathcal{Y}_{m})$.
Recent works in HS-MS image fusion advocates for a low-rank tensor model~\cite{kanatsoulis2018hyperspectralSRR_coupledCPD,prevost2020coupledTucker_hyperspectralSRR_TSP,prevost2019coupledTucker_hyperspectralSRR}. However, the case at hand is more challenging because of the additional variability $\mathcal{\Psi}$, which makes the model more ambiguous.
In many inverse problems such as matrix or tensor factorization, dictionary learning and blind deconvolution, identifiability of the underlying variables often can only be defined up to some fundamental ambiguities.
Transformation groups and equivalence classes~\cite{li2016identifiabilityBilinearInverseProbs} can be used to precisely define which sets of solutions can generate each possible observations.
These ideas can be leveraged to characterize some of the fundamental ambiguities associated with the model~\eqref{eq:tensor_obs_mdl_hsi}--\eqref{eq:tensor_obs_mdl_msi_add}, and to provide insights into the development of efficient algorithms.
First, we will show that the presence of $\mathcal{\Psi}$ makes the model fundamentally ambiguous, as the content in ${\mathcal{Z}_h}$ cannot be easily distinguished from that of $\mathcal{\Psi}$. Moreover, we will define an equivalence class that characterizes the sets of images ${\mathcal{Z}_h}$ and factors $\mathcal{\Psi}$ which are certain to result in different observed HSI and MSI.
This gives us insight into what kind of structure from these variables can be recovered from $(\mathcal{Y}_{h},\mathcal{Y}_{m})$.
{To proceed, let us first denote by $\mathscr{A}:(\mathcal{Z}_h,\mathcal{\Psi})\mapsto(\mathcal{Y}_h,\mathcal{Y}_m)$ the operator which describes the degradation process in equations~\eqref{eq:tensor_obs_mdl_hsi}--\eqref{eq:tensor_obs_mdl_msi_add}.}
{By representing operators $\mathscr{P}_{1,2}$, $\mathscr{P}_{3}$ and $\mathscr{A}$ in matrix form as $\widetilde{\boldsymbol{P}}_{1,2}\in\amsmathbb{R}^{N_1N_2L_h \times M_1M_2L_h}$, $\widetilde{\boldsymbol{P}}_{3}\in\amsmathbb{R}^{M_1M_2L_m \times M_1M_2L_h}$ and $\boldsymbol{A}\in\amsmathbb{R}^{(N_1N_2L_h+M_1M_2L_m)\times 2M_1M_2L_h}$, respectively, we can write the model~\eqref{eq:tensor_obs_mdl_hsi}--\eqref{eq:tensor_obs_mdl_msi_add} in the noiseless case ($\mathcal{E}_h = \mathcal{E}_m = \mathcal{0}$) equivalently as: }
\begin{align}
\begin{bmatrix}
\operatorname{vec}(\mathcal{Y}_h) \\ \operatorname{vec}(\mathcal{Y}_m)
\end{bmatrix}
= \underbrace{
\begin{bmatrix}
\widetilde{\boldsymbol{P}}_{1,2} & \cb{0} \\
\widetilde{\boldsymbol{P}}_{3} & \widetilde{\boldsymbol{P}}_{3}
\end{bmatrix}
}_{\boldsymbol{A}}
\begin{bmatrix}
\operatorname{vec}(\mathcal{Z}_h) \\ \operatorname{vec}(\mathcal{\Psi})
\end{bmatrix} \,.
\label{eq:thm1_mtxForm_1}
\end{align}
{Define also the equivalence relation $\sim_{{Z,\Psi}}$ based on operator $\mathscr{A}$ as follows:
\begin{align}
(\mathcal{Z}_h,\mathcal{\Psi}) \sim_{{Z,\Psi}} (\mathcal{Z}_h',\mathcal{\Psi}') \iff \mathscr{A}(\mathcal{Z}_h,\mathcal{\Psi}) - \mathscr{A}(\mathcal{Z}_h',\mathcal{\Psi}') =\mathcal{0}
\label{eq:thm_prel2_eq_rel}
\end{align}
and its associated equivalence class (EC) $[(\mathcal{Z}_h,\mathcal{\Psi})]_{\sim_{{Z,\Psi}}}$ as
\begin{align}
[(\mathcal{Z}_h,\mathcal{\Psi})]_{\sim_{{Z,\Psi}}} & = \big\{\mathcal{X}\in\Omega_Z\times\Omega_{\Psi}\,:\,\mathcal{X} \sim_{{Z,\Psi}} (\mathcal{Z}_h,\mathcal{\Psi})\big\}.
\label{eq:thm_prel2_eq_class}
\end{align}
Now, we are ready to present the following result.}
\begin{theorem} \label{thm:identifiability_ECc_psi}
Suppose that the observation noise is zero (i.e., $\mathcal{E}_h=\mathcal{E}_m=\mathcal{0}$) and that $\Omega_Z=\amsmathbb{R}^{M_1\times M_2\times L_h}_+$ and $\Omega_{\Psi}=\amsmathbb{R}^{M_1\times M_2\times L_h}$.
Then, given a set of HSI and MSI observations $(\mathcal{Y}_h,\mathcal{Y}_m)${, the following is verified:}
\begin{itemize}
\item[a)] {If operator $\mathscr{A}$ has nontrivial nullspace (e.g., if $L_m<L_h$ or if $N_1N_2<M_1M_2$), then the pair $(\mathcal{Z}_h,\mathcal{\Psi})$ cannot be uniquely identified from the observations $(\mathcal{Y}_h,\mathcal{Y}_m)$.}
\item[b)] There is only one (unique) equivalence class $[(\mathcal{Z}_0,\mathcal{\Psi}_0)]_{\sim_{{Z,\Psi}}}$ containing HR images and scaling factors $({\mathcal{Z}_h},\mathcal{\Psi})$ that can generate~$(\mathcal{Y}_h,\mathcal{Y}_m)$ according to model~\eqref{eq:tensor_obs_mdl_hsi}--\eqref{eq:tensor_obs_mdl_msi_add}. In other words, $({\mathcal{Z}_h},\mathcal{\Psi})$ can be identified uniquely up to $[(\mathcal{Z}_0,\mathcal{\Psi}_0)]_{\sim_{{Z,\Psi}}}$.
\end{itemize}
\end{theorem}
\begin{proof}
{\textbf{Proof of a):}} {Due to the special structure of matrix $\boldsymbol{A}$ in~\eqref{eq:thm1_mtxForm_1}, is is clear that $\operatorname{rank}(\boldsymbol{A})=\operatorname{rank}(\widetilde{\boldsymbol{P}}_{1,2})+\operatorname{rank}(\widetilde{\boldsymbol{P}}_{3})\leq 2M_1M_2L_h$. Thus, if $\mathscr{A}$ has nontrivial nullspace then either $\mathscr{P}_{1,2}$ or $\mathscr{P}_{3}$ have nontrivial nullspace.}
If the operator $\mathscr{P}_{1,2}$ has nontrivial nullspace, then we can find ${\mathcal{Z}_h}$, ${\mathcal{Z}_h'}$, different from one another, such that
\begin{align}
\mathscr{P}_{1,2}({\mathcal{Z}_h}) & = \mathscr{P}_{1,2}({\mathcal{Z}_h'}) \,,
\label{eq:thm1_i}
\end{align}
implying that $\mathcal{Y}_h=\mathcal{Y}_h'$. Now, we can always find $\mathcal{\Psi}$ and $\mathcal{\Psi}'$ satisfying
\begin{align}
{\mathcal{Z}_h} + \mathcal{\Psi} & = {\mathcal{Z}_h'} + \mathcal{\Psi}',
\label{eq:thm1_ii}
\end{align}
which implies that $\mathcal{Y}_m=\mathcal{Y}_m'$ (i.e., the model is not identifiable).
Similarly, if operator $\mathscr{P}_{3}$ has nontrivial nullspace, then suppose we select ${\mathcal{Z}_h}={\mathcal{Z}_h'}$. This makes $\mathcal{Y}_h=\mathcal{Y}_h'$. Then, we can select $\mathcal{\Psi}$ and $\mathcal{\Psi}'$, distinct from one another, satisfying
\begin{align}
\mathscr{P}_{3}(\mathcal{\Psi}) - \mathscr{P}_{3}(\mathcal{\Psi}') = \mathcal{0} \,,
\label{eq:thm1_iii}
\end{align}
where $\mathcal{0}$ is the tensor of zeros. Since ${\mathcal{Z}_h}={\mathcal{Z}_h'}$, this leads to:
\begin{align}
\mathscr{P}_{3}({\mathcal{Z}_h}) & = \mathscr{P}_{3}({\mathcal{Z}_h'})
\nonumber\\
& = \mathscr{P}_{3}({\mathcal{Z}_h'}) + \mathscr{P}_{3}(\mathcal{\Psi}) - \mathscr{P}_{3}(\mathcal{\Psi}') \,,
\label{eq:thm1_iv}
\end{align}
which also implies that $\mathcal{Y}_m=\mathcal{Y}_m'$ (i.e., the model is not identifiable).
{\textbf{Proof of b):}} Note that since there is no additive noise, $\mathscr{A}$ is a linear operator and, thus, satisfies the following relation:
\begin{align}
\mathscr{A}({\mathcal{Z}_h},\mathcal{\Psi}) - \mathscr{A}({\mathcal{Z}_h'},\mathcal{\Psi}') =
\mathscr{A}({\mathcal{Z}_h}-{\mathcal{Z}_h'},\mathcal{\Psi}-\mathcal{\Psi}') \,.
\end{align}
By inspecting the definition of the equivalence relation in~\eqref{eq:thm_prel2_eq_rel}, it can be seen that the equivalence class in~\eqref{eq:thm_prel2_eq_class} is characterized by the
kernel of the operator $\mathscr{A}$, and can also be written as:
\begin{align}
[({\mathcal{Z}_h},\mathcal{\Psi})]_{\sim_{{Z,\Psi}}}
& = \big\{({\mathcal{Z}_h},\mathcal{\Psi}) + \mathcal{X} \,:\, \mathcal{X} \in \ker(\mathscr{A})\big\}.
\label{eq:thm_prel2_eq_class2}
\end{align}
Now, suppose that we have two sets of HR images and variability factors belonging to different ECs, i.e., $({\mathcal{Z}_h},\mathcal{\Psi})\in[(\mathcal{Z}_0,\mathcal{\Psi}_0)]_{\sim_{{Z,\Psi}}}$, $({\mathcal{Z}_h'},\mathcal{\Psi}')\in[(\mathcal{Z}_0',\mathcal{\Psi}_0')]_{\sim_{{Z,\Psi}}}$, with $[(\mathcal{Z}_0,\mathcal{\Psi}_0)]_{\sim_{{Z,\Psi}}}\neq[(\mathcal{Z}_0',\mathcal{\Psi}_0')]_{\sim_{{Z,\Psi}}}$. Comparing the observations $(\mathcal{Y}_h,\mathcal{Y}_m)$ and $(\mathcal{Y}_h',\mathcal{Y}_m')$ generated by elements of each EC, we have:
\begin{align}
(\mathcal{Y}_h,\mathcal{Y}_m)
&= \mathscr{A}(({\mathcal{Z}_h},\mathcal{\Psi}) + \mathcal{X})
\nonumber \\
& \neq \mathscr{A}(({\mathcal{Z}_h'},\mathcal{\Psi}') + \mathcal{X}')
= (\mathcal{Y}_h',\mathcal{Y}_m') \,,
\end{align}
for all $\mathcal{X},\mathcal{X}'\in\ker(\mathscr{A})$. Thus, elements selected from different ECs will always lead to different observations, ensuring that the EC is identifiable.
\end{proof}
Intuitively, {item a)} shows that variations in $\mathcal{\Psi}$ that occur in the nullspace of operators $\mathscr{P}_{1,2}$ and $\mathscr{P}_{3}$ are not reflected in the corresponding observations $(\mathcal{Y}_h,\mathcal{Y}_m)$ in such a way that they can be differentiated from possible changes in ${\mathcal{Z}_h}$.
More generally, changes of $\mathcal{\Psi}$ and ${\mathcal{Z}_h}$ that occur in the nullspace of matrix $\boldsymbol{A}$ do not affect $(\mathcal{Y}_h,\mathcal{Y}_m)${, which is clear from~\eqref{eq:thm1_mtxForm_1}}. This notion can be further extended by noting that changes occurring in the column space of $\boldsymbol{A}$ will certainly lead to different observations, which is made precise in {item b)}.
{Item b) in} Theorem~\ref{thm:identifiability_ECc_psi} allows us to characterize the ambiguities in the model {in more detail}. However, it is important to consider the characteristics of $\ker(\mathscr{A})$ in our problem to better understand the recoverability of the variability factor $\mathcal{\Psi}$. Let us consider the model in~\eqref{eq:thm1_mtxForm_1}, and two sets of variables $({\mathcal{Z}_h},\mathcal{\Psi})\sim_{{Z,\Psi}}({\mathcal{Z}_h'},\mathcal{\Psi}')$, belonging to the same EC.
It can be seen that to generate the same observations, the HR images need to satisfy ${\mathcal{Z}_h}-{\mathcal{Z}_h'}\in\ker(\mathscr{P}_{1,2})$, while the variability factors have to satisfy $\mathscr{P}_{3}({\mathcal{Z}_h}-{\mathcal{Z}_h'})=-\mathscr{P}_{3}(\mathcal{\Psi}-\mathcal{\Psi}')$. Therefore, the general form of the difference between the variability factors inside each equivalence class is of the form:
\begin{align}
\mathcal{\Psi}-\mathcal{\Psi}' = -\underbrace{({\mathcal{Z}_h}-{\mathcal{Z}_h'})}_{\in\ker(\mathscr{P}_{1,2})} + \underbrace{\mathcal{X}}_{\in\ker(\mathscr{P}_{3})} \,.
\label{eq:psi_EC_general_form}
\end{align}
The set of all possible $\mathcal{\Psi}-\mathcal{\Psi}'$ satisfying~\eqref{eq:psi_EC_general_form} is the sum of $\ker(\mathscr{P}_{3})$ and $\ker(\mathscr{P}_{1,2})$, which is given by $\ker(\mathscr{P}_{3}\circ\mathscr{P}_{1,2})$. We can readily see that $\mathcal{\Psi}$ cannot be recovered from the observations.
Only the spectrally degraded variability factors $\mathscr{P}_{3}(\mathcal{\Psi})$ can be uniquely recovered (which comes ``for free'' with the recovery of ${\mathcal{Z}_h}$ since it can be computed as $\mathscr{P}_{3}(\mathcal{\Psi})=\mathcal{Y}_m-\mathscr{P}_{3}({\mathcal{Z}_h})$). This makes it sufficient to study the capability of an algorithm to recover ${\mathcal{Z}_h}$ in our model. Since the matrices $\boldsymbol{P}_i$, $i\in\{1,2,3\}$ are essentially low-pass filtering and downsampling operators, their nullspaces intuitively encode high-frequency information along each tensor mode. Thus, only the smooth structure of $\mathcal{\Psi}$ can be identified uniquely from observations~$(\mathcal{Y}_h,\mathcal{Y}_m)$, since otherwise we cannot separate the effects of $\mathcal{\Psi}$ from ${\mathcal{Z}_h}$.
We also note that each EC in~\eqref{eq:thm_prel2_eq_class}, which contains all factors $\mathcal{\Psi}$ whose difference lies in the nullspace of the combined operator $\mathscr{P}_{1,2}\circ\mathscr{P}_{3}$, is strictly larger than if we considered changes that occur in the nullspace of each of these operators individually (i.e., $\mathscr{P}_{1,2}$ and $\mathscr{P}_{3}$).
Theorem~\ref{thm:identifiability_ECc_psi} guarantees that tensors belonging to different ECs will result in different observations, which is the minimal requirement for having identifiable ${\mathcal{Z}_h}$ and $\mathscr{P}_{3}(\mathcal{\Psi})$. However, the coresponding inverse problem still remains ill-posed as the number of {\color{red}unknowns} is greater than the number of observations.
Thus, stronger identifiability conditions cannot be obtained unless we provide stricter a priori characterizations of the sets $\Omega_Z$ and $\Omega_{\Psi}$.
\subsection{A Low-Multilinear-Rank Model}
\label{sec:low_rank_model}
One possible condition that can be imposed on the structures of both $\Omega_{\Psi}$ and $\Omega_Z$ is the low-rank tensor model. This kind of structure makes it possible to obtain identifiability and exact recovery guarantees for problem~\eqref{eq:HS_MS_fus_var_prob}, where spatial and spectral variabilities are present. Moreover, it also makes the problem well-posed and easier to solve since the number of unknowns becomes smaller than the amount of available data.
Suppose that ${\mathcal{Z}_h}$ and $\mathcal{\Psi}$ have multilinear ranks $(K_{Z,1},K_{Z,2},K_{Z,3})$ and $(K_{\Psi,1},K_{\Psi,2},K_{\Psi,3})$, respectively. This means that they can be represented as
\begin{align}
{\mathcal{Z}_h} &= \big\ldbrack\mathcal{G}_{Z}; \boldsymbol{B}_{Z,1}, \boldsymbol{B}_{Z,2}, \boldsymbol{B}_{Z,3} \big\rdbrack \,,
\label{eq:tucker_mdl_Z}
\\
\mathcal{\Psi} &= \big\ldbrack\mathcal{G}_{\Psi}; \boldsymbol{B}_{\Psi,1}, \boldsymbol{B}_{\Psi,2}, \boldsymbol{B}_{\Psi,3} \big\rdbrack \,,
\label{eq:tucker_mdl_Psi}
\end{align}
where $\boldsymbol{B}_{Z,i}\in\amsmathbb{R}^{M_i\times K_{Z,i}}$, $\boldsymbol{B}_{\Psi,i}\in\amsmathbb{R}^{M_i\times K_{\Psi,i}}$, $i\in\{1,2\}$, $\boldsymbol{B}_{Z,3}\in\amsmathbb{R}^{L_h\times K_{Z,3}}$, $\boldsymbol{B}_{\Psi,3}\in\amsmathbb{R}^{L_h\times K_{\Psi,3}}$ are the factor matrices and $\mathcal{G}_{Z}\in\amsmathbb{R}^{K_{Z,1}\times K_{Z,2}\times K_{Z,3}}$, $\mathcal{G}_{\Psi}\in\amsmathbb{R}^{K_{\Psi,1}\times K_{\Psi,2}\times K_{\Psi,3}}$ are the core tensors.
Our objective is to study the identifiability and exact recovery of these variables given the observation model in~\eqref{eq:tensor_obs_mdl_hsi}--\eqref{eq:tensor_obs_mdl_msi_add}.
Using this model{, and applying the definition of the multilinear product and the properties of the mode-$k$ product defined in Section~\ref{sec:back_defs}}, the noiseless case of the degradation model~\eqref{eq:tensor_obs_mdl_hsi}--\eqref{eq:tensor_obs_mdl_msi_add} can be written as
\begin{align}
\mathcal{Y}_h ={}& \big\ldbrack\mathcal{G}_{Z}; \boldsymbol{P}_1 \boldsymbol{B}_{Z,1}, \boldsymbol{P}_2 \boldsymbol{B}_{Z,2}, \boldsymbol{B}_{Z,3} \big\rdbrack \,
\label{eq:obs_mdl_hsi_lowrank}
\\
\mathcal{Y}_m ={}& \big\ldbrack\mathcal{G}_{Z}; \boldsymbol{B}_{Z,1}, \boldsymbol{B}_{Z,2}, \boldsymbol{P}_3 \boldsymbol{B}_{Z,3} \big\rdbrack
\nonumber \\
& + \big\ldbrack\mathcal{G}_{\Psi}; \boldsymbol{B}_{\Psi,1}, \boldsymbol{B}_{\Psi,2}, \boldsymbol{P}_3 \boldsymbol{B}_{\Psi,3} \big\rdbrack \,.
\label{eq:obs_mdl_msi_lowrank_btd}
\end{align}
Note that we can represent the multispectral image model in~\eqref{eq:obs_mdl_msi_lowrank_btd} equivalently using a standard Tucker model as:
\begin{align}
\mathcal{Y}_m ={}& \big\ldbrack \mathcal{C}_{m}; \boldsymbol{C}_{m,1}, \boldsymbol{C}_{m,2}, \boldsymbol{P}_3 \boldsymbol{C}_{m,3} \big\rdbrack \,,
\label{eq:obs_mdl_msi_lowrank_hightucker}
\end{align}
where $\boldsymbol{C}_{m,i}$, $i\in\{1,2,3\}$ and $\mathcal{C}_m$ are the factor matrices and the core tensor of the MSI, which satisfy:
\begin{align}
\mathcal{C}_{m} &= \mathcal{G}_{Z}\oplus\mathcal{G}_{\Psi} \,, \label{eq:tucker_mdl_msi_var_ext_core}
\\
\boldsymbol{C}_{m,i} &= \big[\boldsymbol{B}_{Z,i} \, \boldsymbol{B}_{\Psi,i}\big],\,\,\, i\in\{1,2,3\} \,, \label{eq:tucker_mdl_msi_var_ext_factors}
\end{align}
{\color{red}where for two tensors $\mathcal{A}$ and $\mathcal{B}$, the binary operation $\mathcal{A}\oplus\mathcal{B}$ returns a block-diagonal tensor whose diagonal blocks are $\mathcal{A}$ and $\mathcal{B}$.}
{The model in equations~\eqref{eq:tucker_mdl_Z}--\eqref{eq:tucker_mdl_msi_var_ext_factors} will be subsequently used in Sections~\ref{sec:alg1_algebraic} and~\ref{sec:alg2_optimizationBased} to develop two image fusion algorithms, one algebraic (faster, but with stringent rank constraints) and another based on an optimization procedure (which allows for higher rank values). In each case, a new algorithm will be presented followed by its recoverability guarantees. It should be noted that in practice, the HRI $\mathcal{Z}_h$ and the variability tensor $\mathcal{\Psi}$ can have high rank. Nevertheless, we will perform only a coupled tensor approximation, with which we are able to capture most of the energy of the data even with insufficient ranks. In practice, however, higher rank models will be preferred to ensure the data is well represented and to avoid the presence artifacts in the reconstructed HRI (which will be achieved with the method of Section~\ref{sec:alg2_optimizationBased}).}
\section{An algebraic algorithm}
\label{sec:alg1_algebraic}
Considering the model in Section~\ref{sec:low_rank_model}, the image fusion problem consists in estimating the factors and core tensor $\mathcal{G}_Z$, $\boldsymbol{B}_{Z,i}$, $i\in\{1,2,3\}$. However, if the values composing the multilinear rank of ${\mathcal{Z}_h}$ are sufficiently low, those variables can be computed by solving the following coupled system of equations:
\begin{align}
\begin{cases}
\mathcal{Y}_h &= \big\ldbrack\mathcal{G}_{Z}; \boldsymbol{C}_{h,1}, \boldsymbol{C}_{h,2}, \boldsymbol{B}_{Z,3} \big\rdbrack
\\
\mathcal{Y}_m &= \big\ldbrack\mathcal{C}_{m}; \boldsymbol{C}_{m,1}, \boldsymbol{C}_{m,2}, \boldsymbol{P}_3 \boldsymbol{C}_{m,3} \big\rdbrack
\\
\boldsymbol{C}_{h,i} &= \boldsymbol{P}_i \boldsymbol{B}_{Z,i}, \,\,\, i\in\{1,2\}
\\
\boldsymbol{C}_{m,i} &= \big[\boldsymbol{B}_{Z,i},\boldsymbol{B}_{\Psi,i}\big], \,\,\, i\in\{1,2,3\}
\end{cases},
\label{eq:alg1_coupled_eqs_formulation}
\end{align}
where $\boldsymbol{C}_{h,i}$, $i\in\{1,2\}$ denote the spatial factor matrices of the HSI, and the HRI is obtained from the solution of~\eqref{eq:alg1_coupled_eqs_formulation} as ${\mathcal{Z}_h}=\ldbrack\mathcal{G}_{Z};\boldsymbol{B}_{Z,1},\boldsymbol{B}_{Z,2},\boldsymbol{B}_{Z,3}\rdbrack$.
If we suppose that $K_{Z,i}+K_{\Psi,i}\leq N_i$, $i\in\{1,2\}$,~\eqref{eq:alg1_coupled_eqs_formulation} can be solved using an efficient, algebraic approach detailed in Alg.~\ref{alg:alg1}, which we call CT-STAR (Coupled Tucker decompositions for hyperspectral Super-resoluTion with vARiability). {The basic intuition behind this algorithm is to use the correspondence between the mode-1 and mode-2 matricizations of the HSI and MSI in order to separate the HRI from the variability tensor when computing its factor matrices. More details will be provided in the following.}
\begin{algorithm} [thb]
\footnotesize
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\caption{\mbox{Algebraic image fusion method (CT-STAR)}\label{alg:alg1}}
\Input{\mbox{Images $\mathcal{Y}_h$, $\mathcal{Y}_m$ ranks $K_{Z,i}$, $K_{\Psi,i}$, $i\in\{1,2,3\}$}}
\Output{HRI ${\widehat{\mathcal{Z}}_h}$, spectrally degraded variability factors $\mathscr{P}_{3}(\widehat{\mathcal{\Psi}})$}
Check if $K_{Z,i}+K_{\Psi,i}\leq N_i$, $i\in\{1,2\}$ \;
Compute $\widehat{\boldsymbol{C}}_{h,3} = \operatorname{tSVD}_{K_{Z,3}}(\unfold{\boldsymbol{Y}_{\!h}}{3})$ \;
Compute $\widehat{\boldsymbol{C}}_{m,i} = \operatorname{tSVD}_{K_{Z,i}+K_{\Psi,i}}(\unfold{\boldsymbol{Y}_{\!m}}{i})$ for $i\in\{1,2\}$ \;
Compute $\widetilde{\boldsymbol{Q}}_i$, for $i\in\{1,2\}$, as $\widetilde{\boldsymbol{Q}}_i = \big(\boldsymbol{P}_i \widehat{\boldsymbol{C}}_{m,i}\big)^{\dagger} \operatorname{tSVD}_{K_{Z,i}}(\unfold{\boldsymbol{Y}_{\!h}}{i})$\;
Compute $\widetilde{\boldsymbol{C}}_{m,i}=\widehat{\boldsymbol{C}}_{m,i} \widetilde{\boldsymbol{Q}}_i$, for $i\in\{1,2\}$ \;
Compute $\widehat{\mathcal{G}}_Z$ by solving
$(\widehat{\boldsymbol{C}}_{h,3} \otimes \boldsymbol{P}_2\widetilde{\boldsymbol{C}}_{m,2} \otimes \boldsymbol{P}_1\widetilde{\boldsymbol{C}}_{m,1} )\operatorname{vec}(\mathcal{G}_Z)=\operatorname{vec}(\mathcal{Y}_h)$\;
Compute ${\widehat{\mathcal{Z}}_h}=\ldbrack\widehat{\mathcal{G}}_{Z}; \widetilde{\boldsymbol{C}}_{m,1}, \widetilde{\boldsymbol{C}}_{m,2}, \widehat{\boldsymbol{C}}_{h,3}\rdbrack$ \;
Compute $\mathscr{P}_{3}(\widehat{\mathcal{\Psi}})=\mathcal{Y}_m-{\widehat{\mathcal{Z}}_h}\times_3\boldsymbol{P}_3$ \;
\end{algorithm}
It is important to note that CT-STAR does not \textcolor{magenta}{enforce} the block diagonal structure of the core tensor of the MSI (described in~\eqref{eq:tucker_mdl_msi_var_ext_core}). The following theorem gives a constructive proof of exact recovery conditions from which Alg.~\ref{alg:alg1} is derived.
\begin{theorem} \label{thm:exacr_rec_alg1}
Suppose that the HRI ${\mathcal{Z}_h}$ and the variability tensor $\mathcal{\Psi}$ have multilinear ranks $(K_{Z,1},K_{Z,2},K_{Z,3})$ and $(K_{\Psi,1},K_{\Psi,2},K_{\Psi,3})$, respectively, that $\mathcal{Y}_h$ and {\color{red}$\mathcal{Y}_m$} admit Tucker decompositions as denoted in~\eqref{eq:alg1_coupled_eqs_formulation}, that the observation noise is zero (i.e. $\mathcal{E}_h=\mathcal{0}$, $\mathcal{E}_m=\mathcal{0}$), and that
\begin{align}
& \operatorname{rank}(\boldsymbol{P}_i \boldsymbol{B}_{Z,i}) = K_{Z,i}\,, \,\,\, i\in\{1,2\}
\label{eq:thm3_rank_cond1}
\\
& \operatorname{rank}(\boldsymbol{P}_i \boldsymbol{B}_{\Psi,i}) \leq K_{\Psi,i}\,, \,\,\, i\in\{1,2\}
\label{eq:thm3_rank_cond2}
\\
& \operatorname{rank}(\unfold{\boldsymbol{Y}_{\!h}}{i}) = K_{Z,i}\,, \,\,\, i\in\{1,2,3\}
\label{eq:thm3_rank_cond3}
\\
& \operatorname{rank}(\unfold{\boldsymbol{Y}_{\!m}}{i}) = K_{Z,i}+K_{\Psi,i} \leq N_i\,, \,\,\, i\in\{1,2\}
\label{eq:thm3_rank_cond4}
\end{align}
Then, if all columns in $\boldsymbol{P}_i \boldsymbol{B}_{Z,i}$ are linearly independent from those in $\boldsymbol{P}_i \boldsymbol{B}_{\Psi,i}$, for $i\in\{1,2\}$, Algorithm~\ref{alg:alg1} exactly recovers ${\mathcal{Z}_h}$ from the observations.
\end{theorem}
\begin{proof}
Let us compute matrices $\widehat{\boldsymbol{C}}_{m,i}$, $i\in\{1,2\}$, and $\widehat{\boldsymbol{C}}_{h,i}$, $i\in\{1,2,3\}$ as the left-singular vectors associated with the non-zero singular values of $\unfold{\boldsymbol{Y}_{\!m}}{i}$, $i\in\{1,2\}$ and $\unfold{\boldsymbol{Y}_{\!h}}{i}$, $i\in\{1,2,3\}$, respectively.
Then, due to~\eqref{eq:thm3_rank_cond3}--\eqref{eq:thm3_rank_cond4} and to the non-uniqueness of matrix decomposition, these matrices satisfy:
\begin{align}
\widehat{\boldsymbol{C}}_{h,i} &= \boldsymbol{P}_i \boldsymbol{B}_{Z,i} \boldsymbol{Q}_{h,i}, \,\,\, i\in\{1,2\}
\label{eq:thm3_4a}
\\
\widehat{\boldsymbol{C}}_{h,3} &= \boldsymbol{B}_{Z,3} \boldsymbol{Q}_{h,3}
\label{eq:thm3_4b}
\\
\widehat{\boldsymbol{C}}_{m,i} &= \boldsymbol{C}_{m,i} \boldsymbol{Q}_{m,i}, \,\,\, i\in\{1,2\}
\label{eq:thm3_4c}
\end{align}
for invertible matrices $\boldsymbol{Q}_{h,i}\in\amsmathbb{R}^{K_{Z,i}\times K_{Z,i}}$, $i\in\{1,2,3\}$ and $\boldsymbol{Q}_{m,i}\in\amsmathbb{R}^{(K_{Z,i}+K_{\Psi,i})\times (K_{Z,i}+K_{\Psi,i})}$, $i\in\{1,2\}$.
Now, the main problem caused by the presence of variability is that matrices $\boldsymbol{Q}_{m,1}$ and $\boldsymbol{Q}_{m,2}$ preclude us from distinguishing the factors $\boldsymbol{B}_{Z,1}$ and $\boldsymbol{B}_{Z,2}$, associated with ${\mathcal{Z}_h}$, from $\boldsymbol{B}_{\Psi,1}$ and $\boldsymbol{B}_{\Psi,2}$, associated with $\mathcal{\Psi}$, using only information available in the MSI. These two become mixed in the spatial factors $\widehat{\boldsymbol{C}}_{m,i}$.
Nonetheless, consider the relationship between spatial degradation of the factors estimated from the MSI and the spatial factors of the HSI:
\begin{align}
\boldsymbol{P}_i \widehat{\boldsymbol{C}}_{m,i} &= \boldsymbol{P}_i \boldsymbol{C}_{m,i} \boldsymbol{Q}_{m,i}
\label{eq:thm3_5a}
\\
&= \big[\boldsymbol{P}_i \boldsymbol{B}_{Z,i}, \, \boldsymbol{P}_i \boldsymbol{B}_{\Psi,i}\big] \boldsymbol{Q}_{m,i} \,.
\label{eq:thm3_5b}
\end{align}
Now, let us compute matrices $\widetilde{\boldsymbol{Q}}_i\in\amsmathbb{R}^{(K_{Z,i}+K_{\Psi,i})\times K_{Z,i}}$, $i\in\{1,2\}$ such that
\begin{align}
& \widehat{\boldsymbol{C}}_{h,i} = \boldsymbol{P}_i \widehat{\boldsymbol{C}}_{m,i} \widetilde{\boldsymbol{Q}}_i \,.
\label{eq:thm3_6a}
\end{align}
By partitioning the following matrix product as $\boldsymbol{Q}_{m,i}\widetilde{\boldsymbol{Q}}_i=[\overline{\boldsymbol{Q}}_{Z,i}^\top,\overline{\boldsymbol{Q}}_{\Psi,i}^\top]^\top$,~\eqref{eq:thm3_6a} can be represented as
\begin{align}
\widehat{\boldsymbol{C}}_{h,i} &= \boldsymbol{P}_i \boldsymbol{B}_{Z,i} \boldsymbol{Q}_{h,i}
\nonumber \\
&= \boldsymbol{P}_i \boldsymbol{B}_{Z,i} \overline{\boldsymbol{Q}}_{Z,i} + \boldsymbol{P}_i \boldsymbol{B}_{\Psi,i} \overline{\boldsymbol{Q}}_{\Psi,i} \,.
\label{eq:thm3_7a}
\end{align}
Since all columns in $\boldsymbol{P}_i \boldsymbol{B}_{Z,i}$ are linearly independent from those in $\boldsymbol{P}_i \boldsymbol{B}_{\Psi,i}$, equality~\eqref{eq:thm3_6a} will be satisfied if and only if the result of the product $\widehat{\boldsymbol{C}}_{m,i} \widetilde{\boldsymbol{Q}}_i$ (i.e., the right hand side of~\eqref{eq:thm3_7a}) does not contain any nontrivial linear combination of the columns of $\boldsymbol{P}_i \boldsymbol{B}_{\Psi,i}$. Thus, $\overline{\boldsymbol{Q}}_{\Psi,i}=\cb{0}$ and $\overline{\boldsymbol{Q}}_{Z,i}=\boldsymbol{Q}_{h,i}$ due to~\eqref{eq:thm3_rank_cond1} and~\eqref{eq:thm3_rank_cond3}.
This allows us to ``separate'' the variability and image factors as
\begin{align}
\widetilde{\boldsymbol{C}}_{m,i} &= \widehat{\boldsymbol{C}}_{m,i} \widetilde{\boldsymbol{Q}}_i
\nonumber\\
& = \boldsymbol{B}_{Z,i} \boldsymbol{Q}_{h,i} \,,
\end{align}
for $i\in\{1,2\}$.
Now, consider the vectorization of the HSI as:
\begin{align}
(\widehat{\boldsymbol{C}}_{h,3} \otimes \boldsymbol{P}_2\widetilde{\boldsymbol{C}}_{m,2} \otimes \boldsymbol{P}_1\widetilde{\boldsymbol{C}}_{m,1} )\operatorname{vec}(\mathcal{G}_Z) = \operatorname{vec}(\mathcal{Y}_h) \,.
\label{eq:thm3_7b}
\end{align}
Since {\color{magenta} the matrix in the left hand side of~\eqref{eq:thm3_7b} has full column rank}, $\widehat{\mathcal{G}}_Z$ can be uniquely recovered from this equation, and will satisfy $\widehat{\mathcal{G}}_Z=\mathcal{G}_Z\times_1\boldsymbol{Q}_{h,1}^{-1}\times_2\boldsymbol{Q}_{h,2}^{-1}\times_3\boldsymbol{Q}_{h,3}^{-1}$.
The HRI and the spectrally degraded scaling factors are then finally recovered as:
\begin{align}
\begin{split}
{\widehat{\mathcal{Z}}_h} &= \ldbrack\widehat{\mathcal{G}}_{Z}; \widetilde{\boldsymbol{C}}_{m,1}, \widetilde{\boldsymbol{C}}_{m,2}, \widehat{\boldsymbol{C}}_{h,3}\rdbrack
\\
&= \ldbrack \mathcal{G}_{Z}; \boldsymbol{B}_{Z,1}, \boldsymbol{B}_{Z,2}, \boldsymbol{B}_{Z,3}\rdbrack = \mathcal{Z}
\end{split}
\end{align}
and
\begin{align} \label{eq:recovering_low_psi1}
\widehat{\mathcal{\Psi}\times_3\boldsymbol{P}_3}=\mathcal{Y}_m-{\widehat{\mathcal{Z}}_h}\times_3\boldsymbol{P}_3 \,,
\end{align}
which completes the proof.
\end{proof}
{Note that CT-STAR does not use the spectral degradation operation $\boldsymbol{P}_3$ to recover the HRI $\mathcal{Z}_h$, what makes it a spectrally ``blind'' algorithm.}
The CT-STAR algorithm is {also} fast (see Section~\ref{sec:comp}), but only works for the cases where the ranks of the spatial modes of ${\mathcal{Z}_h}$ are smaller than the dimensions of the HSI, which is quite restrictive. {This can make CT-STAR unsuited to process real images which can have high rank values, and motivates the search for a method with more flexibility in the selection of the ranks.} Moreover, both Alg.~\ref{alg:alg1} and Theorem~\ref{thm:exacr_rec_alg1}, in considering model~\eqref{eq:alg1_coupled_eqs_formulation}, made no assumptions about the (block diagonal) structure of the core tensor of the MSI. Although this led to more freedom from a modeling perspective, the recoverability conditions turned out to be restrictive.
{In the following section, we will explore the block diagonal structure of $\mathcal{C}_m$ using an optimization-based algorithm to address these limitations.}
\section{An optimization-based algorithm}
\label{sec:alg2_optimizationBased}
In this section, we pursue a different approach. Assume that model~\eqref{eq:obs_mdl_hsi_lowrank}--\eqref{eq:obs_mdl_msi_lowrank_btd} holds and that the values forming the multilinear ranks of both ${\mathcal{Z}_h}$ and $\mathcal{\Psi}$ are sufficiently low so that $\mathcal{Y}_m$ admits a block term decomposition (BTD) in the noiseless case~\cite{lathauwer2008tensor_BTD2_uniqueness}. We can then use uniqueness results thereof to guarantee the identifiability of ${\mathcal{Z}_h}$ under less restrictive conditions.
Let us consider the image fusion problem as the solution to the following optimization problem:
\begin{align}
\mathop{\min}_{\boldsymbol{\Theta}} & \,\, J(\boldsymbol{\Theta}) \triangleq \Big\|\mathcal{Y}_h - \big\ldbrack {\mathcal{G}}_{Z}; \boldsymbol{P}_1 {\boldsymbol{B}}_{Z,1}, \boldsymbol{P}_2 {\boldsymbol{B}}_{Z,2}, {\boldsymbol{B}}_{Z,3} \big\rdbrack \Big\|_F^2
\nonumber\\
& \hspace{-1ex} + {\lambda}\,\bigg\|\mathcal{Y}_m - \sum_{\iota\in\{Z,\Psi\}} \big\ldbrack {\mathcal{G}}_{\iota}; {\boldsymbol{B}}_{\iota,1}, {\boldsymbol{B}}_{\iota,2}, \boldsymbol{P}_3 {\boldsymbol{B}}_{\iota,3} \big\rdbrack \bigg\|_F^2
\label{eq:opt_prob_fus_btd}
\end{align}
where $\boldsymbol{\Theta}=\{{\mathcal{G}}_{\iota},{\boldsymbol{B}}_{\iota,i}:\iota\in\{Z,\Psi\}, \, i\in\{1,2,3\}\}$ {and $\lambda\in\amsmathbb{R}_+$ is a fixed parameter which balances the contribution of each term in the cost function. In the following, we first describe a procedure to solve problem~\eqref{eq:opt_prob_fus_btd} in Section~\ref{sec:CBSTAR_opt}, and later provide exact recovery guarantees in Section~\ref{sec:CBSTAR_recovery}.}
\begin{algorithm} [thb]
\footnotesize
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\caption{\mbox{Optimization-based image fusion (CB-STAR)}\label{alg:alg2}}
\Input{\mbox{Images $\mathcal{Y}_h$,$\mathcal{Y}_m$ ranks $K_{Z,i}$,$K_{\Psi,i}$, $i\in\{1,2,3\}${, iterations~$F$}}}
\Output{HRI ${\widehat{\mathcal{Z}}_h}$, spectrally degraded variability factors $\mathscr{P}_{3}(\widehat{\mathcal{\Psi}})$}
Initialize $\boldsymbol{\Theta}^{(0)}$ according to Section~\ref{sec:alg2_initialization}\;
\While{Stopping criteria is not satisfied}{
{Compute $\mathcal{G}_Z$ and $\boldsymbol{B}_{Z,i}$, $i\in\{1,2,3\}$ by solving~\eqref{eq:opt_subp_Z} with Algorithm~\ref{alg:alg3} in Appendix~\ref{sec:appendix1}, using $F$ iterations\;}
Compute $\boldsymbol{B}_{\Psi,1}$, $\boldsymbol{B}_{\Psi,2}$, $\boldsymbol{P}_3\boldsymbol{B}_{\Psi,3}$ and $\mathcal{G}_{\Psi}$ by solving \eqref{eq:HOSVD_optPsi} using the high-order SVD with rank $(K_{\Psi,1},K_{\Psi,2},K_{\Psi,3})$\;
}
Compute ${\widehat{\mathcal{Z}}_h}=\ldbrack\widehat{\mathcal{G}}_{Z}; \widehat{\boldsymbol{B}}_{Z,1}, \widehat{\boldsymbol{B}}_{Z,2}, \widehat{\boldsymbol{B}}_{Z,3}\rdbrack$ \;
Compute $\mathscr{P}_{3}(\widehat{\mathcal{\Psi}})=\mathcal{Y}_m-{\widehat{\mathcal{Z}}_h}\times_3\boldsymbol{P}_3$ \;
\end{algorithm}
\subsection{Optimization}
\label{sec:CBSTAR_opt}
{In order to minimize the cost function in~\eqref{eq:opt_prob_fus_btd}, we consider a block coordinate descent strategy, which successively minimizes $J$ w.r.t. $\mathcal{Z}_h$ (i.e., $\mathcal{G}_Z$ and $\boldsymbol{B}_{Z,i}$, $i\in\{1,2,3\}$) and w.r.t. $\mathcal{\Psi}$ (i.e., $\mathcal{G}_{\Psi}$ and $\boldsymbol{B}_{\Psi,i}$, $i\in\{1,2,3\}$), while keeping the remaining variables fixed.}
The optimization procedure is detailed in Alg.~\ref{alg:alg2}, which we call CB-STAR (Coupled Block term decompositions for hyperspectral Super-resoluTion with vARiability).
\begin{color}{orange}
\subsubsection{Optimizing w.r.t. $\mathcal{Z}_h$}
The optimization problem w.r.t. $\mathcal{Z}_h$ can be written as:
\begin{align} \label{eq:opt_subp_Z}
\min_{\mathcal{G}_Z,\boldsymbol{B}_{Z,i}} \,\,
& \Big\|\mathcal{Y}_h - \big\ldbrack {\mathcal{G}}_{Z}; \boldsymbol{P}_1{\boldsymbol{B}}_{Z,1}, \boldsymbol{P}_2{\boldsymbol{B}}_{Z,2}, {\boldsymbol{B}}_{Z,3}\big\rdbrack \Big\|_F^2
\nonumber \\
+\lambda & \Big\|\textcolor{magenta}{\mathcal{Y}}_0 - \big\ldbrack {\mathcal{G}}_{Z}; {\boldsymbol{B}}_{Z,1}, {\boldsymbol{B}}_{Z,2}, \boldsymbol{P}_3{\boldsymbol{B}}_{Z,3} \big\rdbrack \Big\|_F^2 \,,
\end{align}
where $\textcolor{magenta}{\mathcal{Y}}_0=\mathcal{Y}_m-\big\ldbrack {\mathcal{G}}_{\Psi}; {\boldsymbol{B}}_{\Psi,1}, {\boldsymbol{B}}_{\Psi,2}, \boldsymbol{P}_3{\boldsymbol{B}}_{\Psi,3}\big\rdbrack$. This is a variability-free, Tucker-based image fusion problem. We propose to solve~\eqref{eq:opt_subp_Z} using a block coordinate descent strategy w.r.t.
$\mathcal{G}_Z$ and $\boldsymbol{B}_{Z,i}$, $i\in\{1,2,3\}$, with a small number of iterations~$F$. This procedure is detailed in Alg.~\ref{alg:alg3} and in Appendix~\ref{sec:appendix1}.
An approximate closed form solution to~\eqref{eq:opt_subp_Z} can also be computed efficiently using the SCOTT algorithm~\cite{prevost2020coupledTucker_hyperspectralSRR_TSP}.
\end{color}
\subsubsection{Optimizing w.r.t. $\mathcal{\Psi}$}
This optimization problem can be written equivalently as
\begin{align} \label{eq:HOSVD_optPsi}
\min_{\mathcal{G}_{\Psi},\boldsymbol{B}_{\Psi,i},\boldsymbol{X}_2} \,\, \Big\|\textcolor{magenta}{\mathcal{Y}}_1 - \big\ldbrack {\mathcal{G}}_{\Psi}; {\boldsymbol{B}}_{\Psi,1}, {\boldsymbol{B}}_{\Psi,2}, \boldsymbol{X}_2 \big\rdbrack \Big\|_F^2 \,,
\end{align}
where $\textcolor{magenta}{\mathcal{Y}}_1=\mathcal{Y}_{m}-\ldbrack {\mathcal{G}}_{Z}; {\boldsymbol{B}}_{Z,1}, {\boldsymbol{B}}_{Z,2}, \boldsymbol{P}_3 {\boldsymbol{B}}_{Z,3}\rdbrack$ and $\boldsymbol{X}_2=\boldsymbol{P}_3 {\boldsymbol{B}}_{\Psi,3}$. This problem can be solved by computing the high-order SVD of $\textcolor{magenta}{\mathcal{Y}}_1$ with rank $(K_{\Psi,1},K_{\Psi,2},K_{\Psi,3})$~\cite{lathauwer2000multilinearHOSVD}. Note that problem~\eqref{eq:HOSVD_optPsi} only returns $\boldsymbol{X}_2=\boldsymbol{P}_3 {\boldsymbol{B}}_{\Psi,3}$ instead of $\boldsymbol{B}_{\Psi,3}$. This is not a problem since the variations of $\boldsymbol{B}_{\Psi,3}$ in the nullspace of $\boldsymbol{P}_3$ are not identifiable.
\subsubsection{Initialization}
\label{sec:alg2_initialization}
Since this optimization problem is non-convex, the choice of initialization can have a significant impact on the performance of the algorithm. This can be particularly prominent in this algorithm since the model considered in~\eqref{eq:tensor_obs_mdl_hsi}--\eqref{eq:tensor_obs_mdl_msi_add} allows for a significant amount of ambiguity.
Fortunately, for practical scenes, we can consider a simple strategy to provide a reasonably accurate initial guess.
In the noiseless case, the following relation is satisfied:
\begin{align} \label{eq:initialization_psi_1}
\mathscr{P}_{1,2}(\mathcal{Y}_m)
& = \mathscr{P}_{1,2}\big(\mathscr{P}_{3}({\mathcal{Z}_h} + \mathcal{\Psi})\big)
\\ \nonumber
& = \mathscr{P}_{3}(\mathcal{Y}_h) + \mathscr{P}_{1,2}\big(\mathscr{P}_{3}(\mathcal{\Psi})\big) \,.
\end{align}
Thus, we can obtain a spatially and spectrally degraded version of $\mathcal{\Psi}$ directly from the HSI and MSI simply as:
\begin{align}
\widetilde{\mathcal{\Psi}}
& = \mathscr{P}_{1,2}(\mathcal{Y}_m) - \mathscr{P}_{3}(\mathcal{Y}_h)
\nonumber \\
& = \mathscr{P}_{1,2}\big(\mathscr{P}_{3}(\mathcal{\Psi})\big) \,.
\label{eq:initialization_psi_2}
\end{align}
Then, if $\mathcal{\Psi}$ is smooth, we can spatially upscale $\widetilde{\mathcal{\Psi}}$ using some form of interpolation (e.g., bicubic), leading to $\mathscr{P}_{3}(\mathcal{\Psi}^{(0)})$.
Finally, we can use the Tucker decomposition of $\mathscr{P}_{3}(\mathcal{\Psi}^{(0)})$ to initialize $\mathcal{G}_{\Psi}$, $\boldsymbol{B}_{\Psi,i}$, $i\in\{1,2,3\}$. We call this the \emph{interpolation} initialization.
Another option is to try to invert~\eqref{eq:initialization_psi_2} using the pseudoinverse of operator $\mathscr{P}_{1,2}$, which can be computed using properties of the tensor vectorization and Kronecker product:
\begin{align}
\big(\mathscr{P}_{1,2}\big)^{\dagger} =
(\cdot) \times_1 \boldsymbol{P}_1^\dagger \times_2 \boldsymbol{P}_2^\dagger \,,
\end{align}
where $\boldsymbol{X}^{\dagger}$ denotes the pseudoinverse of $\boldsymbol{X}$. the initialization can be then computed as $\mathscr{P}_{3}(\mathcal{\Psi}^{(0)})=\widetilde{\mathcal{\Psi}}\times_1 \boldsymbol{P}_1^\dagger \times_2 \boldsymbol{P}_2^\dagger$. We call this the \emph{pseudoinverse} initialization.
The initialization of $\boldsymbol{B}_{Z,i}$, $i\in\{1,2,3\}$ can then be performed as ${\boldsymbol{B}}_{Z,3} = \operatorname{tSVD}_{K_{Z,3}}(\unfold{\boldsymbol{Y}_{\!h}}{3})$ and ${\boldsymbol{B}}_{Z,i} = \operatorname{tSVD}_{K_{Z,i}}(\unfold{\boldsymbol{X}}{i})$ for $i\in\{1,2\}$, where $\mathcal{X}=\mathcal{Y}_m-\mathscr{P}_{3}(\mathcal{\Psi}^{(0)})$.
\subsection{Exact Recovery}
\label{sec:CBSTAR_recovery}
Suppose that $K_{Z,i}=K_{\Psi,i}$, $i\in\{1,2,3\}$, without loss of generality, so that the MSI follows a standard BTD as considered in~\cite{lathauwer2008tensor_BTD2_uniqueness}. Then we have the following result regarding the identifiability of the proposed algorithm.
\begin{theorem} \label{thm:exacr_rec_alg2}
Suppose that $K_i\equiv K_{Z,i}=K_{\Psi,i}$, $i\in\{1,2,3\}$, that the observations are noise free (i.e., $\mathcal{E}_h=\mathcal{0}$, $\mathcal{E}_m=\mathcal{0}$), that $\{\mathcal{G}_{\iota},\boldsymbol{B}_{\iota,i}:\iota\in\{Z,\Psi\},\,i\in\{1,2,3\}\}$ are drawn from some joint absolutely continuous distribution, and that the following conditions on the dimensions hold:
\begin{align}
M_1 \ge 2 K_1 & \mbox{ and } M_2 \ge 2 K_2
\label{eq:dimension_conditions_thm_alg2_c}
\\
K_{Z,3} \leq \min & \big\{ N_1 N_2, \, K_{Z,1}K_{Z,2} \big\}
\label{eq:dimension_conditions_thm_alg2_ef}
\end{align}
and either one of the following:
\begin{align}
\begin{cases}
\begin{cases}
K_3 > K_1+K_2-2 \,, & \text{or} \\
|K_1-K_2| > K_3-2 \,, & \\
\end{cases}
\\
\mbox{ and } L_m \ge 2 K_3
\end{cases}
\label{eq:dimension_conditions_thm_alg2_a}
\end{align}
or
\begin{align}
\begin{split}
K_1=K_2 \,,\,\,\, K_3 \geq 3 \,\,\,\mbox{and}\,\,\, K_3 < L_m \,,
\end{split}
\label{eq:dimension_conditions_thm_alg2_b}
\end{align}
is satisfied. Then, the solution to optimization problem~\eqref{eq:opt_prob_fus_btd} satisfies {$\widehat{\mathcal{Z}}_h=\mathcal{Z}_h$} almost surely.
\end{theorem}
\begin{proof}
Since there is no additive noise, the optimal solution to optimization problem~\eqref{eq:opt_prob_fus_btd} will necessarily make both terms in the cost function equal to zero. This implies that
\begin{align}
\mathcal{Y}_h & = \ldbrack \widehat{\mathcal{G}}_{Z}; \boldsymbol{P}_1\widehat{\boldsymbol{B}}_{Z,1}, \boldsymbol{P}_2\widehat{\boldsymbol{B}}_{Z,2}, \widehat{\boldsymbol{B}}_{Z,3} \rdbrack \,,
\\
\mathcal{Y}_m & = \sum_{\iota\in\{Z,\Psi\}} \ldbrack \widehat{\mathcal{G}}_{\iota}; \widehat{\boldsymbol{B}}_{\iota,1}, \widehat{\boldsymbol{B}}_{\iota,2}, \boldsymbol{P}_3 \widehat{\boldsymbol{B}}_{\iota,3} \rdbrack \,,
\label{eq:recovered_MSI_BTD_opt}
\end{align}
where $\{\widehat{\mathcal{G}}_{\iota},\widehat{\boldsymbol{B}}_{\iota,i}:\iota\in\{Z,\Psi\},\,i\in\{1,2,3\}\}$ denotes a solution to~\eqref{eq:opt_prob_fus_btd}.
Since the dimension conditions~\eqref{eq:dimension_conditions_thm_alg2_c} and either~\eqref{eq:dimension_conditions_thm_alg2_a} or~\eqref{eq:dimension_conditions_thm_alg2_b} are satisfied and the core tensor and factor matrices are drawn from joint absolutely continuous distributions, the BTD decomposition of the MSI in~\eqref{eq:recovered_MSI_BTD_opt} is essentially unique according to Theorems~5.1 and~5.5 in~\cite{lathauwer2008tensor_BTD2_uniqueness}. This means that the following conditions are satisfied:
\begin{align}
\widehat{\boldsymbol{B}}_{Z,i} & = \boldsymbol{B}_{\iota_1,i} \boldsymbol{Q}_{\iota_1,i}, \,\, i\in\{1,2\}
\label{eq:ambig_thm2_1}
\\
\widehat{\boldsymbol{B}}_{\Psi,i} & = \boldsymbol{B}_{\iota_2,i} \boldsymbol{Q}_{\iota_2,i}, \,\, i\in\{1,2\}
\label{eq:ambig_thm2_2}
\\
\boldsymbol{P}_3\widehat{\boldsymbol{B}}_{Z,3} & = \boldsymbol{P}_3 \boldsymbol{B}_{\iota_1,3} \boldsymbol{Q}_{\iota_1,3}
\label{eq:ambig_thm2_3}
\\
\boldsymbol{P}_3\widehat{\boldsymbol{B}}_{\Psi,3} & = \boldsymbol{P}_3 \boldsymbol{B}_{\iota_2,3} \boldsymbol{Q}_{\iota_2,3}
\label{eq:ambig_thm2_4}
\\
\widehat{\mathcal{G}}_{Z} & = \mathcal{G}_{\iota_1} \times_1 \boldsymbol{Q}_{\iota_1,1}^{-1} \times_2 \boldsymbol{Q}_{\iota_1,2}^{-1} \times_3 \boldsymbol{Q}_{\iota_1,3}^{-1}
\label{eq:ambig_thm2_5}
\\
\widehat{\mathcal{G}}_{\Psi} & = \mathcal{G}_{\iota_2} \times_1 \boldsymbol{Q}_{\iota_2,1}^{-1} \times_2 \boldsymbol{Q}_{\iota_2,2}^{-1} \times_3 \boldsymbol{Q}_{\iota_2,3}^{-1}
\label{eq:ambig_thm2_6}
\end{align}
{\color{magenta} To account for the possible permutation of the two BTD terms, indexes $\iota_1$ and $\iota_2$ can be either $(\iota_1,\iota_2)=(Z,\Psi)$ or $(\iota_1,\iota_2)=(\Psi,Z)$,} and $\boldsymbol{Q}_{Z,i}$, $\boldsymbol{Q}_{\Psi,i}$, $i\in\{1,2,3\}$ are invertible matrices of appropriate size, {\color{magenta} which account for rotational {and scaling} ambiguities of the model.}
{\color{magenta} Let us consider the mode-3 unfolding of the spectrally degraded HSI:}
\begin{align*}
\boldsymbol{P}_3\unfold{{\boldsymbol{Y}_{\!h}}}{3} & = \boldsymbol{P}_3\widehat{\boldsymbol{B}}_{Z,3} \underbrace{\unfold{\mathop{\widehat{\boldsymbol{G}}_{Z}}}{3} \big(\boldsymbol{P}_2\widehat{\boldsymbol{B}}_{Z,2} \otimes \boldsymbol{P}_1\widehat{\boldsymbol{B}}_{Z,1}\big)^\top}_{\,\widehat{\!\boldsymbol{X}}}
\\
& = \boldsymbol{P}_3{\boldsymbol{B}}_{Z,3} \underbrace{\unfold{\mathop{{\boldsymbol{G}}_{Z}}}{3} \big(\boldsymbol{P}_2{\boldsymbol{B}}_{Z,2} \otimes \boldsymbol{P}_1{\boldsymbol{B}}_{Z,1}\big)^\top}_{\boldsymbol{X}} \,,
\end{align*}
where $\,\widehat{\!\boldsymbol{X}}$ and ${\boldsymbol{X}}$ are generically full row rank by the assumption~\eqref{eq:dimension_conditions_thm_alg2_ef} on the ranks and on the dimensions.
Therefore, we have
\begin{align}
\operatorname{span}\big( \boldsymbol{P}_3\widehat{\boldsymbol{B}}_{Z,3} \big) = \operatorname{span}\big( \boldsymbol{P}_3{\boldsymbol{B}}_{Z,3} \big) \,.
\label{eq:thm2_btd_span_eq}
\end{align}
However, since $K_3<L_m$ and due to the distributional assumptions on the factor matrices, we have that, generically, matrices $\boldsymbol{P}_3\boldsymbol{B}_{Z,3}$ and $\boldsymbol{P}_3\boldsymbol{B}_{\Psi,3}$ both have rank $K_3$, and matrix $\begin{bmatrix}
\boldsymbol{P}_3 {\boldsymbol{B}}_{Z,3} & \boldsymbol{P}_3\boldsymbol{B}_{\Psi,3}
\end{bmatrix}$ is full column rank (i.e., it has rank greater than $K_3$). Therefore, the subspaces spanned by $\boldsymbol{P}_3\boldsymbol{B}_{Z,3}$ and $\boldsymbol{P}_3\boldsymbol{B}_{\Psi,3}$ are different, and it is not possible to have $\boldsymbol{P}_3 {\boldsymbol{B}}_{Z,3} = \boldsymbol{P}_3\boldsymbol{B}_{\Psi,3} \boldsymbol{S}$, for any matrix $\boldsymbol{S}$.
Thus,~\eqref{eq:thm2_btd_span_eq} implies that the equation $\boldsymbol{P}_3\widehat{\boldsymbol{B}}_{Z,3}=\boldsymbol{P}_3\boldsymbol{B}_{\Psi,3}\boldsymbol{Q}_{\Psi,3}$ is not possible (i.e., it cannot be satisfied for any matrix $\boldsymbol{Q}_{\Psi,3}$). This ensures, due to \eqref{eq:ambig_thm2_3} (and to the essential uniqueness of the MSI BTD), that we must have $\iota_1=Z$ in~\eqref{eq:ambig_thm2_1},~\eqref{eq:ambig_thm2_3} and~\eqref{eq:ambig_thm2_5}, (i.e., the factor matrices related to $\mathcal{\Psi}$ can not fit the HSI), which shows that the correct permutation of the BTD terms is selected.
Finally, since $\,\widehat{\!\boldsymbol{X}}$ is generically full row rank, (or since $\unfold{{\boldsymbol{Y}_{\!h}}}{3}$ has rank $K_3$), we have that $\widehat{\boldsymbol{B}}_{Z,3}=\boldsymbol{B}_{Z,3}\boldsymbol{S}$ for some $\boldsymbol{S}$, and~\eqref{eq:ambig_thm2_3} ensures that $\boldsymbol{S}=\boldsymbol{Q}_{Z,3}$. This means that $\widehat{\boldsymbol{B}}_{Z,3}=\boldsymbol{B}_{Z,3}\boldsymbol{Q}_{Z,3}$ (almost surely), and, using~\eqref{eq:ambig_thm2_1} and~\eqref{eq:ambig_thm2_5} and $\iota_1=Z$, the reconstructed image consequently satisfies
\begin{align}
{\widehat{\mathcal{Z}}_h} & = \big\ldbrack \widehat{\mathcal{G}}_{Z}; \widehat{\boldsymbol{B}}_{Z,1}, \widehat{\boldsymbol{B}}_{Z,2}, \widehat{\boldsymbol{B}}_{Z,3} \big\rdbrack
\nonumber \\
& = \big\ldbrack {\mathcal{G}}_{Z}; {\boldsymbol{B}}_{Z,1}, {\boldsymbol{B}}_{Z,2}, {\boldsymbol{B}}_{Z,3} \big\rdbrack
= {\mathcal{Z}_h},
\end{align}
(almost surely), which concludes the proof.
\end{proof}
By taking the block diagonal structure of $\mathcal{C}_m$ into account, Theorem~\ref{thm:exacr_rec_alg2} obtains generally less restrictive recovery conditions. Comparing Theorems~\ref{thm:exacr_rec_alg1} and~\ref{thm:exacr_rec_alg2}, we can see that: 1) conditions~\eqref{eq:thm3_rank_cond3} and~\eqref{eq:dimension_conditions_thm_alg2_ef} are equivalent; 2) the conditions for the spectral ranks are not directly comparable but are similarly restrictive for both theorems; and, most notably, 3) The constraint~\eqref{eq:thm3_rank_cond4} on spatial ranks is much more restrictive when compared to the one in~\eqref{eq:dimension_conditions_thm_alg2_c}, required by Theorem~\ref{thm:exacr_rec_alg2}. {This shows that, although computationally more demanding, CB-STAR has more flexibility and may be able to deliver \textcolor{darkgreen}{better performance} when compared to CT-STAR for images with complex spatial content.}
\section{{Computational complexity}}\label{sec:comp}
{The computational complexity of the algorithms is given as follows. The total cost involved with the main operations in CT-STAR (Alg.~\ref{alg:alg1}) are the following: 1) computation of the truncated SVDs in steps~2 and~3, which requires $\mathcal{O}(\max\{K_{Z,1},K_{Z,2}\}M_1M_2L_m+K_{Z,3}N_1N_2L_h)$ flops, 2) computation of the equations in steps~4 and~5, which requires~$\mathcal{O}(\max\{K_{Z,1},K_{Z,2}\}N_1N_2L_h+M_1N_1K_{Z,1}+M_2N_2K_{Z,2})$ flops, and solution of the linear equation in step~6, which requires~$\mathcal{O}\big(L_hN_1N_2(K_{Z,1}K_{Z,2}K_{Z,3})^2\big)$ flops.}
{ For CB-STAR (Alg.~\ref{alg:alg2}), the total cost involved in each iteration is 1) solving problem~\eqref{eq:opt_subp_Z} in step~3 using Alg.~\ref{alg:alg3}, whose main costs are based on the solution to Sylvester equations, and are given by $\mathcal{O}\big(F(M_1^3+M_2^3+L_h^3+(K_{Z,2}K_{Z,3})^3)\big)$ (where we assume $K_{Z,1}\geq K_{Z,3}$ for simplicity), and 2) computing the high-order SVD in step~4, which costs~$\mathcal{O}\big(\max_i\{K_{\Psi,i}\}M_1M_2L_m\big)$ flops.
}
\section{Experiments}
\label{sec:results}
In this section, the performance of the proposed approach is illustrated through numerical experiments considering both synthetic and real data containing spatial and spectral variability.
All simulations were coded in MATLAB and run on a desktop with a 4.2 GHz Intel Core i7 and 16GB RAM.
\subsection{Experimental Setup} \label{sec:sim_setup}
We compared CT-STAR and CB-STAR to both matrix and tensor factorization-based algorithms. Among the matrix factorization-based methods, we considered the HySure~\cite{simoes2015HySure} and CNMF~\cite{yokoya2012coupledNMF} methods, the FuVar~\cite{Borsoi_2018_Fusion} method, which accounts for spectral variability, and the multiresolution analysis-based GLP-HS algorithm~\cite{aiazzi2006GLP_HS}. We also considered the {LTMR~\cite{dian2019hyperspectralSRR_tensorMultiRank},} STEREO~\cite{kanatsoulis2018hyperspectralSRR_coupledCPD} and SCOTT~\cite{prevost2020coupledTucker_hyperspectralSRR_TSP} algorithms, {which are tensor-based image fusion methods.}
The real {HRIs and MSIs}, which were acquired at different time instants but at the same spatial resolution, were pre-processed as described in~\cite{simoes2015HySure}. {This consisted} in the manual removal of water absorption and low-SNR bands, followed by the normalization of all bands of the {HRIs and MSIs} such that the 0.999 intensity quantile {corresponded} to a value of~1.
Afterwards, the {HRIs were} denoised (as described in~\cite{roger1996denoisingHSI}) to yield the high-SNR reference image~${\mathcal{Z}_h}$~\cite{yokoya2017HS_MS_fusinoRev}.
The observed HSIs $\mathcal{Y}_{h}$ were then generated from ${\mathcal{Z}_h}$ by applying a separable degradation operator, with $\boldsymbol{P}_1=\boldsymbol{P}_2$ (a Gaussian filter with unity variance followed by a subsampling with a decimation factor of two\footnote{Details on how to construct $\boldsymbol{P}_1$ and $\boldsymbol{P}_2$ can be found in~\cite{prevost2020coupledTucker_hyperspectralSRR_TSP}.}). Gaussian noise was also added to obtain an SNR of 30dB.
The observed MSIs $\mathcal{Y}_{m}$ were generated by adding noise to the reference MSI to obtain an SNR of 40dB. The spectral response function $\boldsymbol{P}_3$ was obtained from calibration measurements and known a priori\footnote{Available for download \href{https://earth.esa.int/web/sentinel/user-guides/sentinel-2-msi/document-library/-/asset_publisher/Wk0TKajiISaR/content/sentinel-2a-spectral-responses}{here}.}.
{Note that the HRI $\mathcal{Z}_h$ is not known by the algorithms, and is just used to assess their performance during the experiments using quantitative metrics.}
The parameters of the algorithms were selected as follows.
We selected the ranks and regularization parameters for HySure, CNMF, FuVar and LTMR according to the original works~\cite{simoes2015HySure,yokoya2012coupledNMF,Borsoi_2018_Fusion,dian2019hyperspectralSRR_tensorMultiRank}. For STEREO, we selected the rank in the interval $[5,80]$ which led to the best reconstruction results. Similarly, for SCOTT and for the proposed algorithms, we selected the spatial ranks in the intervals $[10,80]$ and the spectral ranks in the interval $[2,30]$, which led to the best reconstruction results. {For simplicity, we also set $\lambda=1$ for CB-STAR.} The spatial and spectral degradation operators (or, equivalently, the blurring kernels for HySure, CNMF and FuVar) were assumed to be known a priori for all methods.
The BCD procedure in Alg.~\ref{alg:alg2} was performed until the relative change in the objective function value was smaller than~$10^{-3}$. {At each iteration of Alg.~\ref{alg:alg2}, \textcolor{darkgreen}{we ran Alg.~\ref{alg:alg3} for} one single inner iteration (i.e., $F=1$), which resulted in good experimental performance with moderate execution times.} Both the \emph{interpolation} and the \emph{pseudoinverse} initializations described in Section~\ref{sec:alg2_initialization} were considered, but only the first one (which performed better) is shown in the visual results.
To evaluate the quality of the reconstructed images ${\widehat{\mathcal{Z}}_h}$, we considered four quantitative metrics, which were previously used in~\mbox{\cite{yokoya2017HS_MS_fusinoRev,simoes2015HySure,Borsoi_2018_Fusion}}.
The first metric is the peak signal to noise ratio (PSNR), defined as
\begin{align}
{\rm PSNR}
{}={} \frac{10}{L} \sum_{\ell=1}^L
\log_{10} \Bigg(
\frac{M_1 M_2 \operatorname{E}\!\big\{\!\max\big({[\mathcal{Z}_h]_{:,:,\ell}}\big)\big\}}{\big\|{[\mathcal{Z}_h]_{:,:,\ell} -[\widehat{\mathcal{Z}}_h]_{:,:,\ell}\big\|_F^2}} \Bigg)
\nonumber\,,
\end{align}
where $\operatorname{E}\{\cdot\}$ denotes the expectation operator.
The second metric is the Spectral Angle Mapper (SAM):
\begin{align}
{\rm SAM}
{}={} \frac{1}{M_1 M_2} \sum_{n,m} \arccos \Bigg( \frac{{[\mathcal{Z}_h]_{:,n,m}^\top[\widehat{\mathcal{Z}}_h]_{:,n,m}}}
{\big\|{[\mathcal{Z}_h]_{:,n,m}}\big\|_2\big\|{[\widehat{\mathcal{Z}}_h]_{:,n,m}}\big\|_2} \Bigg)
\nonumber\,.
\end{align}
The ERGAS~\cite{wald2000qualityERGAS} metric provides a global statistical measure of the quality of the fused data, and is defined as:
\begin{align}
{\rm ERGAS}
{}={} \frac{M_1M_2}{N_1N_2} \sqrt{\!\frac{10^4}{L_h} \!
\sum_{\ell=1}^{L_h} \frac{\big\|{[\mathcal{Z}_h]_{:,:,\ell}
-[\widehat{\mathcal{Z}}_h]_{:,:,\ell}}\big\|_F^2}{(\cb{1}^\top{[\mathcal{Z}_h]_{:,:,\ell}}\cb{1}/(M_1M_2))^2}}
\nonumber \,.
\end{align}
The last metric is the average of the bandwise UIQI~\cite{wang2002qualityUIQI}, which evaluates image distortions caused by loss of correlation and by luminance and contrast distortion, with value approaching one as ${\widehat{\mathcal{Z}}_h}$ approaches ${\mathcal{Z}_h}$.
We also evaluate the reconstructed images visually, by displaying true- and pseudo-color representations of the visual and infrared spectra of~${\widehat{\mathcal{Z}}_h}$ (corresponding to the wavelengths $0.45$, $0.56$ and $0.66~\mu m$, and $0.80$, $1.50$ and~$2.20~\mu m$, respectively). Due to space limitations, we only display the results of FuVar, {LTMR,} STEREO, SCOTT, CT-STAR and CB-STAR, since these are the methods which performed best, and the ones which were conceptually closest to our approach.
The spectrally degraded additive factors $\mathscr{P}_{3}(\widehat{\mathcal{\Psi}})$ estimated by CT-STAR and CB-STAR are also evaluated visually, through pseudo-color representations of its visible and infrared spectra, and by the norm (over all bands) of each of its pixels.
\begin{table}[ht!]
\caption{Results - synthetic example (``rk'' stands for ``ranks'')} \vspace{-5pt}
\centering
\renewcommand{\arraystretch}{1.15}
\setlength\tabcolsep{3.5pt}
\resizebox{\linewidth}{!}{
\pgfplotstabletypeset[header=false,
col sep = tab,
columns/0/.style={string type},
every head row/.style={before row={\hline}, after row=\hline},
columns/0/.style={string type, column name={Algorithm}, column type/.add={@{}@{\,}}{}},
columns/1/.style={column name={SAM}, column type/.add={@{\,}|@{\,}}{}},
columns/2/.style={column name={ERGAS}, column type/.add={@{\,}|@{\,}}{}},
columns/3/.style={column name={PSNR}, column type/.add={@{\,}|@{\,}}{}},
columns/4/.style={column name={UIQI}, column type/.add={@{\,}|@{\,}}{}},
columns/5/.style={column name={time}, fixed, precision=2, column type/.add={@{\,}|@{\,}}{}},
columns ={0,1,2,3,4,5},
every row 9 column 1/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 13 column 1/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 13 column 2/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 13 column 3/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 7 column 4/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 8 column 4/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 9 column 4/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 10 column 4/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 11 column 4/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 12 column 4/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 13 column 4/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 6 column 5/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
]{tables/tab_synthetic_MC.txt}}
\label{tab:ex0_bis}
\end{table}
\subsection{Examples -- Synthetic data}
To evaluate the {proposed} algorithms in a controlled scenario, we first considered a simulation with synthetic data.
The tensors ${\mathcal{Z}_h}$ and $\mathcal{\Psi}$, of dimensions $100\times100\times200$, were generated following the Tucker model, with uniformly distributed entries on the interval $[0,1]$ and ranks $(10,10,5)$ and $(5,5,3)$, respectively.
The spectral response function $\boldsymbol{P}_3\in\amsmathbb{R}^{10\times200}$ was constructed by uniformly averaging groups of $20$ bands, and the rest of the simulation setup was the same as described in Section~\ref{sec:sim_setup}. For this experiment, we initialized CB-STAR with the results of CT-STAR.
{\textcolor{darkgreen}{We considered} two examples. First, \textcolor{darkgreen}{we compared} the proposed method to other state-of-the-art algorithms, for different choices of rank. Afterwards, \textcolor{darkgreen}{we evaluated} the sensitivity of CT-STAR and CB-STAR to the presence of additive noise. In both cases, we report average results of a Monte Carlo simulation with 100 \textcolor{magenta}{noise} realizations.}
\subsubsection{Comparison to other algorithms}
{For this comparison, we set the} ranks of STEREO and SCOTT were $50$ and $(60,60,5)$, respectively. {\textcolor{darkgreen}{We also ran} the proposed methods with four different rank values (smaller, equal, and larger than the true data ranks)}, indicated in Table~\ref{tab:ex0_bis}.
The results in Table~\ref{tab:ex0_bis} show that the proposed methods yielded significant improvements when compared to the other algorithms, which is expected since this dataset was generated according to the model~\eqref{eq:tensor_obs_mdl_hsi}--\eqref{eq:tensor_obs_mdl_msi_add}. Moreover, the performance of both CT-STAR and CB-STAR as a function of the ranks was similar, with the best results {observed} when the rank was the same as the ground truth, but with similar performance when the rank was underestimated. {This indicates that CT-STAR and CB-STAR can still perform well when the data rank is higher than the one specified for the model.} When the selected rank was overestimated, the performance of the proposed methods degraded more sharply (with a more prominent decrease for CB-STAR), indicating that the ranks should not be much greater than the true values in order to obtain the best performance.
\begin{color}{orange}
\subsubsection{Effect of noise}
The exact recovery results obtained in Theorems~\ref{thm:exacr_rec_alg1} and~\ref{thm:exacr_rec_alg2} assume a noiseless observation model, which is not the case in real applications.
To illustrate how the performance of the proposed methods is affected by the presence of additive noise~$\mathcal{E}_h$ and~$\mathcal{E}_m$, \textcolor{darkgreen}{we evaluated} the quantitative reconstruction metrics for different SNRs, varying from 0dB up to the noiseless case (SNR = $\infty$). For simplicity, we set $K_Z$ and $K_{\Psi}$ as the correct ranks of $\mathcal{Z}_h$ and $\mathcal{\Psi}$, respectively. The results are shown in Table~\ref{tab:exp_synth_SNRs}. It can be seen that the performance of both CT-STAR and CB-STAR decreased for low SNRs. Moreover, CB-STAR achieved better results for high SNRs ($\geq$40dB), but was surpassed by CT-STAR for low SNR scenarios ($\leq$30dB), whose results were considerably better for extremely noisy scenarios (0dB). Moreover, for SNRs found in typical scenes ($\geq$20dB), \textcolor{darkgreen}{these results indicate that} both methods are able to obtain satisfactory results.
Finally, we note that since the ranks of $\mathcal{Z}_h$ and $\mathcal{\Psi}$ \textcolor{darkgreen}{satisfied the} requirements of Theorems~\ref{thm:exacr_rec_alg1} and~\ref{thm:exacr_rec_alg2}, exact reconstruction was obtained for SNR~=~$\infty$ (i.e., the original and reconstructed images were equal up to machine precision).
The theoretical results presented in this work \textcolor{darkgreen}{made use of two hypotheses} to guarantee the recovery of the HRI, namely, that the image and variability factors have low multilinear rank, and that there is no observation noise. However, those hypotheses are often not strictly satisfied in practice. Nonetheless, CT-STAR and CB-STAR only perform a low-rank approximation of the data. Thus, they can still perform well when the true rank of the image is larger than the one specified to the algorithm and under the presence of noise, as illustrated in the previous simulations (even though exact recovery is not guaranteed in those cases). In the following section, we will evaluate the performance of the proposed methods with real HSIs and MSIs, which do not necessarily come from the specified low-rank models.
\end{color}
\begin{table}[th]
\renewcommand{\arraystretch}{1.15}
\setlength\tabcolsep{3.5pt}
\centering
\caption{{Performance of the proposed methods for different SNRs}} \vspace{-5pt}
\resizebox{\linewidth}{!}{
\begin{tabular}{c|cccccccccc} \hline
\multicolumn{10}{c}{CT-STAR} \\\hline
SNR & $0$ & $10$ & $20$ & $30$ & $40$ & $60$ & $80$ & $100$ & $\infty$ \\\hline
SAM & $9.622$ & $3.191$ & $1.197$ & $0.647$ & $0.370$ & $0.033$ & $0.003$ & $0$ & $0$ \\
ERGAS & $15.33$ & $6.948$ & $2.005$ & $0.887$ & $0.436$ & $0.040$ & $0.004$ & $0$ & $0$ \\
PSNR & $17.18$ & $23.76$ & $34.59$ & $41.72$ & $48.35$ & $69.23$ & $89.35$ & $109.4$ & $297.4$ \\
UIQI & $0.501$ & $0.856$ & $0.987$ & $0.998$ & $0.999$ & $1$ & $1$ & $1$ & $1$ \\
\hline
\multicolumn{10}{c}{CB-STAR} \\\hline
SNR & $0$ & $10$ & $20$ & $30$ & $40$ & $60$ & $80$ & $100$ & $\infty$ \\\hline
SAM & $36.89$ & $14.47$ & $4.567$ & $1.272$ & $0.295$ & $0.023$ & $0.002$ & $0$ & $0$ \\
ERGAS & $52.83$ & $17.26$ & $5.130$ & $1.433$ & $0.342$ & $0.028$ & $0.003$ & $0$ & $0$ \\
PSNR & $7.74$ & $17.32$ & $27.66$ & $38.17$ & $50.54$ & $71.99$ & $91.95$ & $111.9$ & $265.9$ \\
UIQI & $0.157$ & $0.590$ & $0.927$ & $0.994$ & $1$ & $1$ & $1$ & $1$ & $1$ \\
\hline
\end{tabular}}
\label{tab:exp_synth_SNRs}
\end{table}
\begin{figure}[thb]
\centering
\begin{minipage}{0.36\linewidth}
\centering
\includegraphics[width=1\linewidth]{figures/HSI_MSI_4b_IsabellaLake_sub1.pdf}
Lake Isabella
\end{minipage}
\,\,
\begin{minipage}{0.42\linewidth}
\centering
\includegraphics[width=1\linewidth]{figures/HSI_MSI_lockwood_s1.pdf}
Lockwood
\end{minipage}
\caption{Hyperspectral and multispectral images with a small acquisition time difference used in the experiments.}
\label{fig:ex3_HSIs_MSIs_a}
\end{figure}
\begin{figure}[thb]
\centering
\begin{minipage}{0.28\linewidth}
\centering
\includegraphics[width=1\linewidth]{figures/HSI_MSI_Playa-crop}
Ivanpah Playa
\end{minipage}
\,\,
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1\linewidth]{figures/HSI_MSI_Tahoe_better_reg.pdf}
Lake Tahoe A
\end{minipage}
\,\,
\begin{minipage}{0.32\linewidth}
\centering
\includegraphics[width=1\linewidth]{figures/HSI_MSI_Tahoe_changed.pdf}
Lake Tahoe B
\end{minipage}
\caption{Hyperspectral and multispectral images with a large acquisition time difference used in the experiments.}
\label{fig:ex3_HSIs_MSIs_b}
\end{figure}
\subsection{Example -- Real data}
In this example, we evaluated the algorithms using real HS and MS images acquired at different time instants, thus presenting different acquisition and seasonal conditions.
The reference hyperspectral and multispectral images, with a pixel size of 20~m, acquired by the AVIRIS and by the Sentinel-2A instruments, respectively, were originally considered in~\cite{Borsoi_2018_Fusion}.
Four sets of image pairs were available. Two of which contained images acquired less than three months apart (thus containing moderate variability). The other two contained images acquired with a time difference of more than one year (thus containing more significant variability).
The HSI and MSI contained $L_h=173$ and $L_m=10$ bands, respectively.
The selected ranks for the tensor-based methods are shown in Table~\ref{tab:exp_ranks_values}.
{Although the ranks of CB-STAR \textcolor{darkgreen}{satisfied Theorem~\ref{thm:exacr_rec_alg2} only for} the Ivanpah Playa image pair, \textcolor{darkgreen}{this did not have} a negative impact on its performance, as will be shown in the following.}
\subsubsection{\textcolor{magenta}{Rank} {sensitivity analysis}}
{
Before proceeding, we evaluate how sensitive the performance of the proposed methods is to the selection of the ranks by plotting the PSNR as a function of each of the ranks $K_{Z,i}$ and $K_{\Psi,i}$, while keeping the others fixed.
For simplicity, we kept the spatial ranks equal to each other (i.e., $K_{Z,1}=K_{Z,2}$ and $K_{\Psi,1}=K_{\Psi,2}$), and only show the results for the Lockwood image (to be described in Section~\ref{sec:res_realdata_moderate}) due to space limitations\footnote{Additional results are available on the supplemental material.}. The results, shown in Fig.~\ref{fig:sensitivity_ex6}, indicate that CT-STAR performs well when $K_{\Psi,1}=K_{\Psi,2}$ are small, and performs well for values of $K_{Z,3}$ which are not small. The optimal $K_{Z,1}=K_{Z,2}$ were relatively large, but a drop in performance was observed when they approach their upper limit $N_i-K_{\Psi,i}$, $i\in\{1,2,\}$. The performance of CB-STAR increased steadily with $K_{Z,1}=K_{Z,2}$ and for small values of $K_{Z,3}$, but decreased more sharply when all values $K_{Z,i}$, $i\in\{1,2,3\}$ were large. The variability ranks $K_{\Psi,i}$, on the other hand, did not affect the results too much when $K_{\Psi,1}=K_{\Psi,2}$ were sufficiently large.
}
\begin{figure}
\centering
\begin{minipage}{0.45\linewidth} \centering
\includegraphics[width=\linewidth]{figures/sensit/ex6_sensit_CTSTAR_rankZ.pdf}
\end{minipage}
\begin{minipage}{0.45\linewidth} \centering
\includegraphics[width=\linewidth]{figures/sensit/ex6_sensit_CTSTAR_rankPsi.pdf}
\end{minipage}
\begin{minipage}{0.45\linewidth} \centering
\includegraphics[width=\linewidth]{figures/sensit/ex6_sensit_CBSTAR_rankZ.pdf}
\end{minipage}
\begin{minipage}{0.45\linewidth} \centering
\includegraphics[width=\linewidth]{figures/sensit/ex6_sensit_CBSTAR_rankPsi.pdf}
\end{minipage}
\caption{{Sensitivity analysis of the proposed methods Lockwood image. Top: PSNR of CT-STAR as a function of the ranks of $\mathcal{Z}_h$ (left) and $\mathcal{\Psi}$ (right). Bottom: PSNR of CB-STAR as a function of the ranks of $\mathcal{Z}_h$ (left) and $\mathcal{\Psi}$ (right)}}
\label{fig:sensitivity_ex6}
\end{figure}
\begin{table}[ht]
\renewcommand{\arraystretch}{1.15}
\setlength\tabcolsep{3.5pt}
\centering
\caption{Ranks of the tensor-based algorithms used in the experiments} \vspace{-5pt}
{\small
\begin{tabular}{c|cccc} \hline
& CT-STAR: & CB-STAR: & SCOTT: & STEREO: \\
& $K_Z,K_{\Psi}$ & $K_Z,K_{\Psi}$ & $K$ & $K$ \\ \hline
Lockwood & \makecell{$(30,30,8)$,\\$(3,3,2)$} & \makecell{$(70,70,5)$,\\$(40,40,3)$} & $(60,60,5)$ & $50$\\
Lake Isabella & \makecell{$(30,30,8)$,\\$(3,3,2)$} & \makecell{$(50,50,5)$,\\$(40,40,3)$} & $(60,60,5)$ & $50$\\
Lake Tahoe & \makecell{$(30,30,10)$,\\$(3,3,1)$} & \makecell{$(35,35,9)$,\\$(50,50,4)$} & $(40,40,7)$ & $30$\\
Ivanpah Playa & \makecell{$(16,16,8)$,\\$(3,3,2)$} & \makecell{$(40,40,4)$,\\$(40,40,5)$} & $(30,30,30)$ & $10$\\ \hline
\end{tabular}}
\label{tab:exp_ranks_values}
\end{table}
\subsubsection{Moderate variability}
\label{sec:res_realdata_moderate}
The first pair of images considered in this example contained $80\times80$ pixels and were acquired over the region surrounding Lake Isabella, on 2018-06-27 and on 2018-08-27. The second pair of images contained $80\times100$ pixels and was acquired near Lockwood, on 2018-08-20 and on 2018-10-19. A true color representation of the HSI and MSI for this example can be seen in Fig.~\ref{fig:ex3_HSIs_MSIs_a}.
Due to the relatively small difference between the acquisition dates of both images, the HSI and MSI look similar. However, there are slight differences between them, as seen in the overall color hue of the images and in the upper right part of the Lake Isabella HSI.
The quantitative performance metrics of all algorithms are shown in Tables~\ref{tab:ex_lockwood} and~\ref{tab:ex_isabellalake}, while the reconstructed images are presented in Figs.~\ref{fig:ex_lockwood_Z_Psi_est} and~\ref{fig:ex_isabellalake_Z_Psi_est}.
\begin{table}[ht!]
\caption{Results - Lockwood} \vspace{-5pt}
\centering
\renewcommand{\arraystretch}{1.15}
\setlength\tabcolsep{3.5pt}
{\small
\pgfplotstabletypeset[header=false,
col sep = tab,
columns/0/.style={string type},
every head row/.style={before row={\hline}, after row=\hline},
columns/0/.style={string type, column name={Algorithm}, column type/.add={@{}@{\,}}{}},
columns/1/.style={column name={SAM}, column type/.add={@{\,}|@{\,}}{}},
columns/2/.style={column name={ERGAS}, column type/.add={@{\,}|@{\,}}{}},
columns/3/.style={column name={PSNR}, column type/.add={@{\,}|@{\,}}{}},
columns/4/.style={column name={UIQI}, column type/.add={@{\,}|@{\,}}{}},
columns/5/.style={column name={time}, fixed, precision=2, column type/.add={@{\,}|@{\,}}{}},
columns ={0,1,2,3,4,5},
every row 8 column 1/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 3 column 2/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 8 column 3/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 8 column 4/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 9 column 4/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 6 column 5/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
]{tables/tab_lockwood.txt}}
\label{tab:ex_lockwood}
\end{table}
\begin{table}[ht!]
\caption{Results - Isabella Lake} \vspace{-5pt}
\centering
\renewcommand{\arraystretch}{1.15}
\setlength\tabcolsep{3.5pt}
{\small
\pgfplotstabletypeset[header=false,
col sep = tab,
columns/0/.style={string type},
every head row/.style={before row={\hline}, after row=\hline},
columns/0/.style={string type, column name={Algorithm}, column type/.add={@{}@{\,}}{}},
columns/1/.style={column name={SAM}, column type/.add={@{\,}|@{\,}}{}},
columns/2/.style={column name={ERGAS}, column type/.add={@{\,}|@{\,}}{}},
columns/3/.style={column name={PSNR}, column type/.add={@{\,}|@{\,}}{}},
columns/4/.style={column name={UIQI}, column type/.add={@{\,}|@{\,}}{}},
columns/5/.style={column name={time}, fixed, precision=2, column type/.add={@{\,}|@{\,}}{}},
columns ={0,1,2,3,4,5},
every row 8 column 1/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 3 column 2/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 8 column 3/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 8 column 4/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 6 column 5/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
]{tables/tab_isabellalake.txt}}
\label{tab:ex_isabellalake}
\end{table}
The quantitative results show that CB-STAR achieved the overall best results for this example, outperforming the other methods in all metrics except in ERGAS, where it performed very similarly to FuVar (which yielded the best results for this metric). CT-STAR, on the other hand, performed similarly to STEREO and SCOTT, being limited by the more stringent constraints on the image ranks.
The visual inspection of the results indicates that CB-STAR provides reconstructions closest to the ground truth when compared to the remaining methods. Although FuVar also provided good results, it yielded a slightly worse representation of the road in the left part of the Lockwood HSI, as well as more aberrations in the color of the light-brown regions in the middle of the Isabella Lake scene (which are not seen in the results of CB-STAR). {LTMR,} STEREO and SCOTT, not being able to account for variability, yielded slight color aberrations in the reconstructions, which are most clearly seen in the central part of the Isabella Lake image, while CT-STAR produced significant artifacts due to the stringent rank constraints.
The estimated factors $\mathscr{P}_{3}(\widehat{\mathcal{\Psi}})$ are in agreement with the localized changes observed in Fig.~\ref{fig:ex3_HSIs_MSIs_a}, particularly in the upper-central area of the Isabella Lake image pair, which is subject to local illumination changes.
The computation times of the algorithms show a large difference between that of FuVar and those of the other algorithms, which indicates that CB-STAR achieves better results at a significantly smaller computational complexity.
\subsubsection{Significant variability}
The remaining image pairs used in this example were acquired over the Ivanpah Playa and over Lake Tahoe area. The Ivanpah Playa image pair contained $80\times128$ pixels and was acquired on 2015-10-26 and on 2017-12-17. For the Lake Tahoe region, we considered two different image pairs (``A'' and ``B''), both with $100\times80$ pixels, the first one acquired on 2014-10-04 and on 2017-10-24, and the second one acquired on 2014-09-19 and on 2017-10-24. A true color representation of the HSI and MSI for this example can be seen in Fig.~\ref{fig:ex3_HSIs_MSIs_b}.
Due to the considerable difference between the acquisition date/time of the HSI and MSI, significant differences can be found between them. For the Ivanpah Playa images, there are large variations between the sand colors in the central part of the image. For the Lake Tahoe region, significant differences are observed in both image pairs, with differences in the color hues of the ground and of the crop circles for the image pair A, and also a large change in the water level of the lake in the image pair B.
The quantitative performance metrics of all algorithms are shown in Tables~\ref{tab:ex_playa},~\ref{tab:ex_tahoe}, and~\ref{tab:ex_tahoeChanged}, while the reconstructed images are presented in Figs.~\ref{fig:ex_playa_Z_Psi_est},~\ref{fig:ex_tahoe_Z_Psi_est} and~\ref{fig:ex_tahoeChanged_Z_Psi_est}.
\begin{table}[ht!]
\caption{Results - Ivanpah Playa} \vspace{-5pt}
\centering
\renewcommand{\arraystretch}{1.15}
\setlength\tabcolsep{3.5pt}
{\small
\pgfplotstabletypeset[header=false,
col sep = tab,
columns/0/.style={string type},
every head row/.style={before row={\hline}, after row=\hline},
columns/0/.style={string type, column name={Algorithm}, column type/.add={@{}@{\,}}{}},
columns/1/.style={column name={SAM}, column type/.add={@{\,}|@{\,}}{}},
columns/2/.style={column name={ERGAS}, column type/.add={@{\,}|@{\,}}{}},
columns/3/.style={column name={PSNR}, column type/.add={@{\,}|@{\,}}{}},
columns/4/.style={column name={UIQI}, column type/.add={@{\,}|@{\,}}{}},
columns/5/.style={column name={time}, fixed, precision=2, column type/.add={@{\,}|@{\,}}{}},
columns ={0,1,2,3,4,5},
every row 3 column 1/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 8 column 2/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 8 column 3/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 3 column 4/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 7 column 5/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
]{tables/tab_playa.txt}}
\label{tab:ex_playa}
\end{table}
\begin{table}[ht!]
\caption{Results - Lake Tahoe A} \vspace{-5pt}
\centering
\renewcommand{\arraystretch}{1.15}
\setlength\tabcolsep{3.5pt}
{\small
\pgfplotstabletypeset[header=false,
col sep = tab,
columns/0/.style={string type},
every head row/.style={before row={\hline}, after row=\hline},
columns/0/.style={string type, column name={Algorithm}, column type/.add={@{}@{\,}}{}},
columns/1/.style={column name={SAM}, column type/.add={@{\,}|@{\,}}{}},
columns/2/.style={column name={ERGAS}, column type/.add={@{\,}|@{\,}}{}},
columns/3/.style={column name={PSNR}, column type/.add={@{\,}|@{\,}}{}},
columns/4/.style={column name={UIQI}, column type/.add={@{\,}|@{\,}}{}},
columns/5/.style={column name={time}, fixed, precision=2, column type/.add={@{\,}|@{\,}}{}},
columns ={0,1,2,3,4,5},
every row 3 column 1/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 8 column 2/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 8 column 3/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 8 column 4/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 9 column 4/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 5 column 5/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
]{tables/tab_tahoe.txt}}
\label{tab:ex_tahoe}
\end{table}
\begin{table}[ht!]
\caption{Results - Lake Tahoe B} \vspace{-5pt}
\centering
\renewcommand{\arraystretch}{1.15}
\setlength\tabcolsep{3.5pt}
{\small
\pgfplotstabletypeset[header=false,
col sep = tab,
columns/0/.style={string type},
every head row/.style={before row={\hline}, after row=\hline},
columns/0/.style={string type, column name={Algorithm}, column type/.add={@{}@{\,}}{}},
columns/1/.style={column name={SAM}, column type/.add={@{\,}|@{\,}}{}},
columns/2/.style={column name={ERGAS}, column type/.add={@{\,}|@{\,}}{}},
columns/3/.style={column name={PSNR}, column type/.add={@{\,}|@{\,}}{}},
columns/4/.style={column name={UIQI}, column type/.add={@{\,}|@{\,}}{}},
columns/5/.style={column name={time}, fixed, precision=2, column type/.add={@{\,}|@{\,}}{}},
columns ={0,1,2,3,4,5},
every row 8 column 1/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 8 column 2/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 8 column 3/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 8 column 4/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
every row 5 column 5/.style={postproc cell content/.style={ @cell content/.add={$\bf}{$} }},
]{tables/tab_tahoeChanged.txt}}
\label{tab:ex_tahoeChanged}
\end{table}
The quantitative results show that \mbox{CB-STAR} achieved again the overall best results for this example, outperforming the remaining algorithms in most metrics, except in the SAM and UIQI for the Ivanpah Playa HRI and in the SAM of the Lake Tahoe A HRI. Moreover, there was a stronger gap between the performance of the methods that consider variability and the remaining algorithms. \mbox{CT-STAR}, although better than {LTMR,} STEREO and SCOTT, performed significantly worse than CB-STAR due to its stringent constraints on the image ranks.
The visual inspection of the results again indicates that CB-STAR provides reconstructions closest to the ground truth when compared to the remaining methods. Although FuVar also provided good results {(as it accounts for spectral variability)}, the reconstructions by CB-STAR were closer to the ground truth, as can be observed in the color shades of the upper part of the Ivanpah Playa image and of the crop circles of the Lake Tahoe~A image, and especially in the overall colors in the more uniform regions containing soil and water and vegetation in the Lake Tahoe B image. {However, FuVar results showed slightly sharper edges in some regions (e.g., around the crop circles in the Lake Tahoe images), which occurs due to the use of a Total Variation spatial regularization. Nonetheless, a spatial regularization can also be incorporated to the CB-STAR cost function in~\eqref{eq:opt_prob_fus_btd} to achieve a similar effect.}
{LTMR,} STEREO and SCOTT, on the other hand, produced significant artifacts in all reconstructed images. CT-STAR also produced significant artifacts, which as in the previous example are due to the stringent rank constraints{, with reconstructions visually worse than those of CB-STAR and FuVar}.
The estimated factors $\mathscr{P}_{3}(\widehat{\mathcal{\Psi}})$ were in close agreement with the variability seen in the scenes, notably in the sand region of the central part of the Ivanpah Playa image (which lies at the botton of a hill), and in the regions near the lake in the Lake Tahoe A and B images, which undergo variations in the water level. Moreover, the overall amplitude of the variables was significantly larger than in the previous example, in which the differences between the images were more moderate.
The computation times of all algorithms were similar to those observed in the previous example, except for that of CB-STAR, which was higher since it underwent a larger number of iterations for the data in this example. Nonetheless, the computation times of CB-STAR were still considerably smaller than those of FuVar.
\begin{figure*}
\centering
\resizebox{0.85\linewidth}{!}
{
\includegraphics[height=10ex]{figures/Zhat/ex6_lockwood_recIm_vis_IR.pdf}
\includegraphics[height=9.85ex]{figures/estimPsis/ex6_lockwood_both_algs_P3Psi_vis_norms2.pdf}
}
\vspace{-0.3cm}
\caption{Left: Visible (top) and infrared (bottom) representation of the true and estimated versions of the Lockwood HSI. Right: Spectrally degraded additive scaling factors $\mathscr{P}_{3}(\mathcal{\Psi})$ estimated by CT-STAR and CB-STAR.}
\label{fig:ex_lockwood_Z_Psi_est}
\end{figure*}
\begin{figure*}
\centering
\resizebox{0.85\linewidth}{!}
{
\includegraphics[height=10ex]{figures/Zhat/ex5_isabellalake_recIm_vis_IR.pdf}
\includegraphics[height=9.9ex]{figures/estimPsis/ex5_isabellalake_both_algs_P3Psi_vis_norms2.pdf}
}
\vspace{-0.3cm}
\caption{Left: Visible (top) and infrared (bottom) representation of the true and estimated versions of the Isabella Lake HSI. Right: Spectrally degraded additive scaling factors $\mathscr{P}_{3}(\mathcal{\Psi})$ estimated by CT-STAR and CB-STAR.}
\label{fig:ex_isabellalake_Z_Psi_est}
\end{figure*}
\begin{figure*}
\centering
\resizebox{0.85\linewidth}{!}
{
\includegraphics[height=10ex]{figures/Zhat/ex3_playa_recIm_vis_IR.pdf}
\includegraphics[height=9.95ex]{figures/estimPsis/ex3_playa_both_algs_P3Psi_vis_norms2.pdf}
}
\vspace{-0.3cm}
\caption{Left: Visible (top) and infrared (bottom) representation of the true and estimated versions of the Ivanpah Playa HSI. Right: Spectrally degraded additive scaling factors $\mathscr{P}_{3}(\mathcal{\Psi})$ estimated by CT-STAR and CB-STAR.}
\label{fig:ex_playa_Z_Psi_est}
\end{figure*}
\begin{figure*}
\centering
\resizebox{0.85\linewidth}{!}
{
\includegraphics[height=10ex]{figures/Zhat/ex4_tahoeA_recIm_vis_IR.pdf}
\includegraphics[height=9.9ex]{figures/estimPsis/ex4_tahoe_both_algs_P3Psi_vis_norms2.pdf}
}
\vspace{-0.3cm}
\caption{Left: Visible (top) and infrared (bottom) representation of the true and estimated versions of the Lake Tahoe A HSI. Right: Spectrally degraded additive scaling factors $\mathscr{P}_{3}(\mathcal{\Psi})$ estimated by CT-STAR and CB-STAR.}
\label{fig:ex_tahoe_Z_Psi_est}
\end{figure*}
\begin{figure*}[thb]
\centering
\resizebox{0.85\linewidth}{!}
{
\includegraphics[height=10ex]{figures/Zhat/ex4_tahoeB_recIm_vis_IR.pdf}
\includegraphics[height=9.9ex]{figures/estimPsis/ex4_tahoe_changed_both_algs_P3Psi_vis_norms2.pdf}
}
\vspace{-0.3cm}
\caption{Left: Visible (top) and infrared (bottom) representation of the true and estimated versions of the Lake Tahoe B HSI. Right: Spectrally degraded additive scaling factors $\mathscr{P}_{3}(\mathcal{\Psi})$ estimated by CT-STAR and CB-STAR.}
\label{fig:ex_tahoeChanged_Z_Psi_est}
\end{figure*}
\section{Conclusions}
\label{sec:conclusions}
In this paper, we proposed a novel framework for multimodal (hyperspectral and multispectral) image fusion accounting for spatially and spectrally localized changes. We first studied the general identifiability of the considered model, which becomes more ambiguous due to the presence of changes. Then, assuming that the high resolution image and the variation factors admit a Tucker decomposition, two new algorithms were proposed -- one being purely algebraic (which was computationally more efficient), and another based on an optimization procedure (which allowed for more relaxed specification of the multilinear ranks). Theoretical guarantees for the exact recovery of the high resolution image were provided for both algorithms. The proposed optimization-based algorithm achieved superior experimental performance in the presence of spectral and spatial variations between the images, while also exhibiting a smaller computational cost.
\appendices
\section{Optimization of cost function~\eqref{eq:opt_subp_Z}}
\label{sec:appendix1}
{To solve~\eqref{eq:opt_subp_Z}, we consider a block coordinate descent strategy (as in, e.g.,~\cite{lathauwer2008tensor_BTD3_algorithms}), which successively minimizes the cost function in~\eqref{eq:opt_subp_Z} with respect to each of the variables $\mathcal{G}_Z$, $\boldsymbol{B}_{Z,i}$, $i\in\{1,2,3\}$,} while keeping the remaining ones fixed.
Let us denote by $J\big(\boldsymbol{X}\,|\,\boldsymbol{\Theta}_{\setminus\{\boldsymbol{X}\}}\big)$ the cost function $J$ in which all variables but $\boldsymbol{X}$ are fixed.
Note that we apply the QR factorization after computing each of the factor matrices $\boldsymbol{B}_{\iota,j}$ to constrain them to be unitary at all the iterations, as performed in~\cite{lathauwer2008tensor_BTD3_algorithms}. This normalization prevents convergence issues by avoiding under/over-flow and keeping these matrices well-conditioned. {Starting from an initialization $\mathcal{G}_Z$, $\boldsymbol{B}_{Z,i}$, $i\in\{1,2,3\}$, this procedure is repeated for $F$ inner iterations.}
\begin{algorithm} [thb]
\footnotesize
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\caption{\mbox{Block coordinate descent to solve~\eqref{eq:opt_subp_Z}}
\label{alg:alg3}}
\Input{\mbox{Tensors $\mathcal{Y}_h$, $\mathcal{X}_0$, $\mathcal{G}_Z$ matrices $\boldsymbol{B}_{Z,i}$,$i\in\{1,2,3\}${, iterations~$F$}}}
\Output{$\mathcal{G}_Z$ matrices $\boldsymbol{B}_{Z,i}$,$i\in\{1,2,3\}$}
\For{$i=1,\ldots,F$}{
Minimize $J(\mathcal{G}_Z|\boldsymbol{\Theta}_{\setminus\{\mathcal{G}_Z\}}^{(i)})$ w.r.t. $\mathcal{G}_Z$\;
Minimize $J(\boldsymbol{B}_{Z,1}|\boldsymbol{\Theta}_{\setminus\{\boldsymbol{B}_{Z,1}\}}^{(i)})$ w.r.t. $\boldsymbol{B}_{Z,1}$\;
Minimize $J(\boldsymbol{B}_{Z,2}|\boldsymbol{\Theta}_{\setminus\{\boldsymbol{B}_{Z,2}\}}^{(i)})$ w.r.t. $\boldsymbol{B}_{Z,2}$\;
Minimize $J(\boldsymbol{B}_{Z,3}|\boldsymbol{\Theta}_{\setminus\{\boldsymbol{B}_{Z,3}\}}^{(i)})$ w.r.t. $\boldsymbol{B}_{Z,3}$\;}
\end{algorithm}
\subsubsection{Optimizing w.r.t. $\boldsymbol{B}_{Z,i}$, $i\in\{1,2,3\}$}
To save space, we present only the case where $i=1$. The extension to $i\in\{2,3\}$ is straightforward.
Note that the cost function $J\big(\boldsymbol{B}_{Z,1}\,|\,\boldsymbol{\Theta}_{\setminus\{\boldsymbol{B}_{Z,1}\}}\big)$ can be equivalently reformulated using the mode-$1$ matricization as
\begin{align} \label{eq:opt_subp_1_i}
J\big(& \boldsymbol{B}_{Z,1}\,|\,\boldsymbol{\Theta}_{\setminus\{\boldsymbol{B}_{Z,1}\}}\big) = \big\|\unfold{\boldsymbol{Y}_{\!h}}{1}^\top - \boldsymbol{X}_2 \boldsymbol{B}_{Z,1}^\top\boldsymbol{P}_1^\top\big\|_F^2
\nonumber\\ &
+ {\lambda} \big\|\unfold{\boldsymbol{Y}_{\!m}}{1}^\top - \boldsymbol{X}_3 - \boldsymbol{X}_1 \boldsymbol{B}_{Z,1}^\top\big\|_F^2 \,,
\end{align}
where matrices $\boldsymbol{X}_1$, $\boldsymbol{X}_2$ and $\boldsymbol{X}_3$ are given by
\begin{align}
\boldsymbol{X}_1 &= \big({\boldsymbol{B}}_{Z,3} \otimes \boldsymbol{P}_3{\boldsymbol{B}}_{Z,2}\big) \unfold{\boldsymbol{G}_Z}{1}^\top \,,
\\
\boldsymbol{X}_2 &= \big(\boldsymbol{P}_2{\boldsymbol{B}}_{Z,3} \otimes {\boldsymbol{B}}_{Z,2}\big) \unfold{\boldsymbol{G}_Z}{1}^\top \,,
\\
\boldsymbol{X}_3 & = \big( \boldsymbol{P}_3 {\boldsymbol{B}}_{\Psi,3} \otimes {\boldsymbol{B}}_{\Psi,2}\big)\unfold{\boldsymbol{G}_{\Psi}}{1}^\top {\boldsymbol{B}}_{\Psi,1}^\top \,.
\end{align}
Computing the derivative of~\eqref{eq:opt_subp_1_i} and setting it equal to zero results in the following expression:
\begin{align}
\begin{split}
{\lambda}\boldsymbol{X}_1^\top\boldsymbol{X}_1 {\boldsymbol{B}}_{Z,1}^\top
& + \boldsymbol{X}_2^\top\boldsymbol{X}_2 {\boldsymbol{B}}_{Z,1}^\top\boldsymbol{P}_1^\top\boldsymbol{P}_1
\\
= {\lambda} & \boldsymbol{X}_1^\top(\unfold{\boldsymbol{Y}_{\!m}}{1}^\top - \boldsymbol{X}_3)
+ \boldsymbol{X}_2^\top\unfold{\boldsymbol{Y}_{\!h}}{1}^\top\boldsymbol{P}_1 \,.
\end{split}
\end{align}
This is a Sylvester equation that can be directly solved using existing software with, e.g., the Hessenberg-Schur or the Bartels-Stewart algorithms (see~\cite{simoncini2016solversSylvesterEq} and references therein).
\subsubsection{Optimizing w.r.t. $\mathcal{G}_{Z}$}
Cost function $J\big(\mathcal{G}_{Z}\,|\,\boldsymbol{\Theta}_{\setminus\{\mathcal{G}_{Z}\}}\big)$ can be equivalently reformulated using the tensor vectorization as
\begin{align} \label{eq:opt_subp_4_i}
J\big(\mathcal{G}_{Z}\,|\,\boldsymbol{\Theta}_{\setminus\{\mathcal{G}_{Z}\}}\big)
={} & \big\|\operatorname{vec}(\mathcal{Y}_h) - \boldsymbol{X}_2 \,\boldsymbol{g}_Z \,\big\|_F^2
\nonumber \\ &
+ {\lambda} \big\|\operatorname{vec}(\mathcal{Y}_m) - \boldsymbol{x}_3 - \boldsymbol{X}_1 \, \boldsymbol{g}_Z \, \big\|_F^2 \,,
\end{align}
where $\boldsymbol{g}_Z=\operatorname{vec}(\mathcal{G}_{Z})$ is the vectorization of the core tensor, and $\boldsymbol{X}_1$, $\boldsymbol{X}_2$ and $\boldsymbol{x}_3$ are given by
\begin{align}
\boldsymbol{X}_1 & = \big(\boldsymbol{P}_3 \boldsymbol{B}_{Z,3} \otimes \boldsymbol{B}_{Z,2} \otimes \boldsymbol{B}_{Z,1} \big) \,,
\\
\boldsymbol{X}_2 & = \big(\boldsymbol{B}_{Z,3} \otimes \boldsymbol{P}_2 \boldsymbol{B}_{Z,2} \otimes \boldsymbol{P}_1 \boldsymbol{B}_{Z,1} \big) \,,
\\
\boldsymbol{x}_3 & = \big(\boldsymbol{P}_3\boldsymbol{B}_{\Psi,3} \otimes \boldsymbol{B}_{\Psi,2} \otimes \boldsymbol{B}_{\Psi,1} \big) \operatorname{vec}(\mathcal{G}_{\Psi}) \,.
\end{align}
The solution that minimizes~\eqref{eq:opt_subp_4_i} can be computed through the normal equations, which can be written as
\begin{align}
({\lambda}\boldsymbol{X}_1^\top\boldsymbol{X}_1 + \boldsymbol{X}_2^\top\boldsymbol{X}_2) \boldsymbol{g}_Z {}={} & {\lambda}\boldsymbol{X}_1^\top(\operatorname{vec}(\mathcal{Y}_m)-\boldsymbol{x}_3)
\nonumber \\
& + \boldsymbol{X}_2^\top\operatorname{vec}(\mathcal{Y}_h) \,.
\end{align}
As shown in~\cite{prevost2020coupledTucker_hyperspectralSRR_TSP}, this set of equations can be alternatively interpreted as a generalized Sylvester equation, for which efficient solvers can be used.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:introduction}
In this article, we revisit the $\Psi$--process introduced in \cite{candy} and studied in \cite{Junge}. This point process on $[0,1]$ can be defined as follows. Suppose that $N_t$ is the counting function for a Poisson process on $[0, \infty)$ with intensity $e^t.$ Suppose at time $0,$ the process is started at some discrete point configuration on $[0,1].$ A point is added to the configuration on $[0,1]$ at each time that $N_t$ jumps according to a Markovian rule we will now describe.
At each time $t,$ the point configuration partitions $[0,1]$ into intervals, whose lengths we denote by
\(
I_1^{(t)},
I_2^{(t)},
\ldots,
I_{N_t+n_0}^{(t)}
\).
Define the size-biased empirical distribution function
\[
\Atilde[t](x) = \sum_{i=1}^{N_t+n_0} |I_i^{(t)}| \one[|I_i^{(t)}|\le x].
\]
At a jump time $t,$ select an interval $\mathcal{I}_t$ with length $\Atilde[t_{-}]^{-1}(u)$ uniformly at random, where $u$ is sampled from a law on $(0,1]$ with distribution function $\Psi$. Finally add a point to $\mathcal{I}_t$ uniformly at random.
Different choices of $\Psi$ produce substantially different behavior, and some canonical choices of $\Psi$ produce processes with alternative descriptions (we will elaborate on this some later; see \cite{candy} for further discussion). The case $\Psi(u) = u$ we call the \emph{uniform} process. In this case, intervals are selected with probability proportional to length and subdivided uniformly, so that the evolution of interval lengths $\{I_i^{(t)}\}$ has the same law as the spacings between points which are added independently and uniformly in $[0,1]$ with rate $e^t.$
The motivating example for the $\Psi$--process is the \emph{max--$k$ process}, for $k \in \mathbb N,$ which is given by $\Psi(u) = u^k.$ This process can also be described by executing the following rule whenever $N_t$ jumps: choose $k$ independent and uniformly distributed points in $[0,1].$ Select the point the lands in the largest interval, breaking ties uniformly, and add it to the existing point configuration. Analogously, in the \emph{min--$k$ process} for $k \in \mathbb N,$ given by $\Psi(u) = 1 - (1-u)^k,$ one instead selects the point that lands in the smallest interval. While max--$k$ and min--$k$ processes have this alternative description for $k \in \mathbb N,$ as $\Psi$--processes they are well--defined for any real $k > 0.$
The max--$k$ process is an example of a \emph{choice} algorithm, in which a constant number of equivalent random choices are presented to an agent, which then employs a heuristic to choose between them. There are a wide variety of problems in which this has been employed, see \cite{MRS01} for a survey, and which are collectively named the \emph{power--of--choice} paradigm. While we will not elaborate on this body of work here, let us mention there has been subsequent developments on the power--of--choice in equidistribution of points \cite{Ohadfromthefuture}.
In \cite{candy}, we show that after appropriately rescaling, the distribution of interval lengths in the $\Psi$--process converges to a deterministic limit under some mild assumptions (we give the formal statement in Theorem \ref{thm:candy} below). Most importantly, we will suppose the following.
\begin{assumption} \label{ass:arsch}
There are constants $c_{\Psi}>0$ and $\kappa_\Psi\in[1,\infty)$ so that
\[
1-\Psi(u) \ge c_{\Psi}(1-u)^{\kappa_\Psi} \quad \text{ for all } u\in(0,1).
\]
\end{assumption}
\noindent This assumption ensures that large intervals are subdivided frequently enough. While perhaps not optimal, it can not be removed entirely; for example in the simple case that $\Psi(u) \equiv 1$ on the interval $[1-u_0,1],$ it follows the associated $\Psi$--process will cease subdividing the longest interval once its length is less than $u_0,$ which precludes any deterministic scaling limit for the empirical distribution of interval lengths.
We define
\[
\A[t](x) = \Atilde[t](e^{t}x)
\]
for all $t,x \geq 0.$ We have the following theorem:
\begin{theorem}[\cite{candy}]
\label{thm:candy}
Assume that $\Psi$ is continuous and satisfies Assumption \ref{ass:arsch}. Then there is an absolutely continuous probability measure $\mu^\Psi$ on $(0,\infty)$ with mean 1, independent of the initial configuration, such that $\A[t]$ converges pointwise to the function $F^{\Psi}(x) = \int_0^x y\,\mu^\Psi(dy)$, almost surely as $t\to\infty$. Furthermore, the Lebesgue-Stieltjes measures $\frac{1}{z}d\A[t](z)$ almost surely converge weakly to $\mu^\Psi$ as $t\to\infty$. The function $F^\Psi$ is characterized by the equation
\[
\frac{dF(x)}{dx} = x\int_x^\infty \frac{1}{z}\,d\Psi(F(z)).
\]
\end{theorem}
The proof of Theorem~\ref{thm:candy} in \cite{candy} relies on the solvability of a certain non-linear, non-local evolution equation together with the method of \emph{asymptotic pseudotrajectories}, or \emph{Kushner-Clark method}, for stochastic approximation algorithms, applied in an infinite-dimensional setting (see \cite{BeLeRa2002} for an earlier appearence of this method). Note that explicit convergence rates are, in general, unknown.
The max--$k$ and min--$k$ processes satisfy the assumptions of Theorem~\ref{thm:candy} for any $k \geq 1.$ The most substantial difference seen between these processes is in the tail behavior of $F^{\Psi}(x).$ For the max processes, $1-F^{\Psi}(x) \sim C_kxe^{-k x}$ as $x \to \infty$ (\cite[Proposition 9.2]{candy}) for some constant $C_k > 0$. For the min processes with $k > 1$, $1-F^{\Psi}(x) \sim c_k x^{-1/(k-1)}.$ Nonetheless, we conjecture in \cite{candy} that the empirical distribution of points almost surely weakly converges to the uniform distribution on $[0,1]$ as $t\to\infty$, for all $\Psi$--processes to which Theorem~\ref{thm:candy} applies, in spite of the heavy--tailed behavior. A $\Psi$--process with this property is said to be \emph{equidistributed almost surely}.
In \cite{Junge}, Junge shows this for some $\Psi$--processes.
\begin{theorem}[\cite{Junge}]
Suppose that $\Psi$ is $C^2[0,1],$ and let $\psi = \Psi'.$ If for some $\delta \in (0,1]$ and all $z \geq 0,$
\[
|z \psi'(F^{\Psi}(z)) (F^{\Psi})'(z) - \psi(F^{\Psi}(z))| \leq (2-\delta) \psi(F^{\Psi}(z)),
\]
then the $\Psi$--process is equidistributed almost surely as $t \to \infty.$
\label{thm:Junge}
\end{theorem}
This condition is satisfied by the max--$2$ process but surprisingly not by any max--$k$ process for $k \geq 3,$ nor any min--$k$ process. It is however satisfied by certain combinations of max--$k$ and min--$k$ process, including ones which slightly favor picking a smaller interval.
In this paper, we remove the additional technical condition of \cite{Junge} and weaken the smoothness assumptions on $\Psi:$
\begin{theorem}
Suppose that $\Psi$ is $C^1[0,1]$ and satisfies Assumption \ref{ass:arsch}.
Then the $\Psi$--process is equidistributed almost surely as $t \to \infty.$
\label{thm:main}
\end{theorem}
\noindent This in particular shows that all max--$k$ and min--$k$ processes equidistribute for $k \geq 1$.
The method we use is a development on top of \cite{Junge}, and we outline this method and how our method differs from \cite{Junge} (see Section \ref{sec:formalism} for a formal overview of the proof). To show equidistribution, it suffices to show that for any $\alpha \in (0,1)$ the asymptotic fraction of points less than $\alpha$ is $\alpha.$ In turn, it should suffice to show that the empirical distribution of intervals contained in $(0,\alpha)$ is asymptotically the same as the global interval distribution. In this case, heuristically, conditioned on picking an interval of some length, the probability that interval is in $(0,\alpha)$ is just $\alpha,$ on account of the total length of the interval being an $\alpha$ fraction of the whole; hence points would be added to $(0,\alpha)$ with asymptotic rate $\alpha e^{t}.$ Some smoothing of the interval lengths would be required to formalize such an argument, and in fact it is possible to bypass the issue entirely (see the proof of Theorem \ref{thm:main} in Section \ref{sec:formalism}).
To show that the empirical distribution of intervals in $(0,\alpha)$ has the same limit as the global distribution, we use what might be described as a linearization procedure. We probabilistically describe the restriction of the $\Psi$--process to $(0,\alpha)$ as one in which potential points are added with rate $e^{t}$, but are thinned at a rate which is a functional of the global interval distribution of the $\Psi$--process (this is one interpretation of Proposition \ref{prop:semimartingale} or \cite[Proposition 3]{Junge} -- as a side comment, this is one place where $\Psi \in C^1[0,1]$ represents a natural technical barrier, as the description becomes problematic for rougher $\Psi$). Knowing that the global interval distribution converges, we may seek to replace this functional by one that depends only on the \emph{limiting} interval distribution. Having done so, we arrive at an \emph{analytic} description of an interval fragmentation process in which intervals evolve independently of one another.
The idea to make this linearized comparison is a key idea in the analysis of \cite{Junge}, and we weaken the smoothness assumptions on $\Psi$ required (see Proposition \ref{prop:continuity} and c.f.\ \cite[Proposition 5(V)]{Junge}). Having made the reduction, the final major step is to show that the (deterministic) linearized evolution of the size--biased distribution functions converges to $F^{\Psi}.$ This is the origin of the technical hypothesis in Theorem \ref{thm:Junge}, and a new argument for this convergence is the major development in this article on Junge's and one which we believe may be of independent interest.
\subsection*{Growth--fragmentation equations}
The resulting linearized equation for the evolution of the (rescaled) size--biased distribution $\F = (\F[t])_{t \geq 0}$ is
\begin{equation} \label{eq:RGa}
\F[t](x)
= {\F[0]}(e^{-t}x) + \int_0^t (e^{s-t}x)^2 \left[\int_{e^{s-t}x}^\infty \frac{R(z)}{z}\,d\F[s](z)\right]\,ds \quad \text{for all } t,x \geq 0,
\end{equation}
where $R$
is some locally integrable non--negative Borel function (for application to the $\Psi$--process, $R = \psi \circ F^{\Psi}$).
We need to show that solutions of this equation converge for essentially arbitrary initial conditions as $t \to \infty$ to the limit:
\[
F^R(x) = \frac{1}{Z^R} \int_0^x y\exp\left( -\int_1^x R(y)\,dy \right)\,dx,
\]
using the usual Riemann integration convention that for $x < 1,$ $\int_1^x R(y)\,dy = -\int_x^1 R(y)\,dy.$
The constant $Z^R$ is a normalization so that $F^R$ is a distribution function, and for $F^R$ to be a distribution function we must assume:
\begin{equation}\label{eq:ZR}
Z^R = \int_0^\infty y\exp\left( -\int_1^x R(y)\,dy \right)\,dx < \infty.
\end{equation}
Besides the finiteness of $Z^R,$ we make no further assumptions on $R.$ For the application to $\Psi$--processes, it is an automatic corollary of the theory in \cite{candy} that $Z^R$ is finite.
The evolution equation \eqref{eq:RGa} is an integrated and scaled form of a very--well studied partial integro--differential equation, the growth--fragmentation equation, which originated in questions arising in mathematical biology (see e.g.\ \cite{Perthame,EngelNagel,CCF1,CCF2}).
Following \cite{BertoinWatsonJFA}, the growth--fragmentation equation is given by
\begin{equation} \label{eq:gfeq}
\partial_t u_t(x) + \partial_x( u_t(x)c(x) ) = \int_x^\infty u_t(y)k(y,x)dy - u_t(x)K(x).
\end{equation}
The quantity $u_t(x)$ represents the density of particles of varying masses $x.$ Each particle grows with a rate $c: (0,\infty) \to (0,\infty)$ that depends on its mass. The particles fragment into smaller masses, and $k(y,x)$ represents the rate at which particles of mass $x$ result from fragmentation of a particle of mass $y > x.$ The term $K(x)$ represents the rate of destruction of particles of mass $x,$ as a result of fragmentation.
In our application, $c(x) = x$ and $k(y,x) = 2xR(y)/y,$ and $K(x) = xR(x).$ The special choice of $c(x) = x$ is is called the \emph{self--similar fragmentation equation} in \cite{CCM} (caveat emptor: \cite{CCM} use a different meaning of self--similar fragmentation kernel than \cite{BCG} or \cite{BertoinWatsonJFA}).
A common assumption in the literature is the assumption of \emph{mass--conservative fragmentations}, so that no mass is created or destroyed upon fragmentation of a particle (\cite{Bertoin,BertoinWatsonAAP,BertoinWatsonJFA,BCG,CCM}. Analytically, this corresponds to
\[
K(y) = \frac{1}{y}\int_0^y x k(y,x) dx, \quad \text{ for all } y > 0.
\]
Our process \emph{does not} satisfy this assumption. Rather, we assume \emph{population--conservative fragmentations}, that is
\[
K(y) = \int_0^y k(y,x) dx, \quad \text{ for all } y > 0.
\]
This corresponds to the evolution of mass--biased random sampling of a fragment from a mass--conservative fragmentation. Such fragmentation equations have been called \emph{conservative} in the literature (c.f.\ \cite{Bertoin, Bouguet}), but as there is an obvious risk of confusion, we shall enforce the lengthier \emph{population--conservative} terminology. Besides appearing in \cite{Bouguet} explicitly, it also appears in quite general form in the work of \cite{BertoinWatsonJFA} and also in a special case (the TCP process) in \cite{Chafai,Bardet}.
The population--conservative growth--fragmentation equation (PCGFE) has been used to model the size of a biological cell, which grows in time and then undergoes mitosis (fragments) into two daughter cells, after which point \emph{one} of the resulting daughters is chosen as a representative of the two. This naturally corresponds to a piecewise deterministic Markov process (PDMP), which grows deterministically and then jumps down. It has been called a \emph{cell process} in the literature (\cite{Bouguet,Bertoin}). This process appears in \cite{Bouguet} (take $\tau(x)=x, \beta(x) = xR(x),$ and $Q(x,dy) = 2y\,dy$) as does it appear in \cite{BertoinWatsonJFA} ($\bar{k}(x,y) = 2yR(x)/x$ and $\underline{c}(x) = 0$).
Our main task is to show that solutions of PCGFEs tend to equilibrium in the large--time limit. This we do by a probabilistic method, which essentially has three components. The first part of the work is to construct the cell process. While general existence theorems for PDMPs exist (for example \cite{Davis1984}), they are generally formulated under the assumption of some type of uniform control on the jump rate. As we do not assume such a feature, we must show that the cell process is well--posed and has the usual desirable properties, c.f.\ Proposition \ref{prop:markov}.
The second part of the work is to show that PCGFE solutions can be represented by the semigroup of the associated cell process (Proposition \ref{prop:rgb}, c.f.\ Sections \ref{sec:diff_mu}, \ref{sec:construction_markov_process}, \ref{sec:Lpi_and_uniqueness}). Under additional continuity and boundedness assumptions on $R,$ as assumed in \cite{Bouguet, BertoinWatsonJFA}, the usual $C_b(\mathbb{R})$--semigroup theory, such as that which is found in \cite{EthierKurtz}, applies. We assume very little boundedness and smoothness of $xR(x),$ and so in fact the $C_b(\mathbb{R})$--semigroup formulation is nonsense (in fact, even under the additional assumption that $R$ be continuous, the $C_b(\mathbb{R})$--setting is substantially complicated by the lack of boundedness on $xR(x)$). Hence, we switch to an $\Lpi$--setting (with $\pi=dF^{R}$) to develop the semigroup theory. This $\Lpi$--semigroup setting is common in the analysis literature on growth--fragmentation equations, see for example \cite{Mischler2016}. Moreover, we must use some of the central ideas from \cite{Mischler2016} to develop this $\Lpi$--semigroup theory for the cell process.
The final part is to show the resulting cell process is ergodic (see Section \ref{sec:ergodicity}). This we do by appealing to the regenerative structure of the cell process. While we do not show a rate, as is common in much of the literature (\cite{Bouguet,BertoinWatsonJFA,BCG,CCM, Mischler2016}, we emphasize that the argument we give still works under minimal assumptions on $R.$ The argument is also conceptually simple, from the probabilistic viewpoint.
To our knowledge, the only explicit existing work on ergodicity of PCGFEs is in \cite{Bouguet,BertoinWatsonJFA}, which would only apply to regularly varying and continuous $R$ (see \cite[Assumption 2.1]{Bouguet} and \cite[Assumption 6.1]{BertoinWatsonJFA}). The method of \cite{Mischler2016} could in principle apply, but the details are only worked out for some explicit cases. Let us also mention that a central idea of \cite{BertoinWatsonJFA} is something like a Girsanov transform, that allows certain mass--conservative growth--fragmentation equations (MCGFEs) to be related to other PCFGEs. This could in priniciple allow results to be transfered between MCGFEs and PCGFEs. We give some further discussion of our contributions to the theory of growth--fragmentation equations in Section \ref{sec:growth-fragmentation}.
\section*{Acknowledgements}
This work was initiated while the second author was invited professor (\emph{professeur invit\'e}) at Universit\'e Paris-Sud. The authors thank this institution for the support and hospitality. They would also like to thank the CRM at Universit\'e de Montr\'eal for their hospitality during a second meeting. The authors also are indebted to Itai Benjamini, Jean Bertoin, B\'en\'edicte Haas, Matt Junge, Jean-Claude Lootgieter, Olivier Raimond and Lorenzo Zambotti for helpful conversations regarding this work.
\section{Formalism}
\label{sec:formalism}
In this section, we will give the formal setup of the $\Psi$--process and the approach to Theorem \ref{thm:main}.
Let $N_t^{\alpha}$ be the number of points that land in $[0,\alpha].$ To prove Theorem~\ref{thm:main}, it suffices to show that $e^{-t}N_t^{\alpha} \to \alpha.$
We use the convention that boldface letters represent processes, i.e.\,they are time dependent. This parameter will appear as a subscript when referring to the process at a fixed time, e.g. $\A = (\A[t])_{t \geq 0}.$
We define the process $(\Aatilde, \Aaatilde, \Atilde)$ to be the size-biased empirical distributions of interval lengths contained in $([0,\alpha], [\alpha,1],[0,1])$, respectively. An interval which straddles the point $\alpha$ contributes whatever fraction of its length landed on either side of $\alpha$ to $\Aatilde$ or $\Aaatilde$ respectively. For every $t \geq 0,$ let $g_t^\alpha : \left\{ 1,2,\dots,N_t \right\} \to [0,1]$ be defined by
\[
g_t^\alpha(j) = \frac{| I_j^{(t)} \cap [0,\alpha]|}{|I_j^{(t)}|}.
\]
We then define
\[
\Aatilde[t](x) = \sum_{j=1}^{N_t} g_t^\alpha(j) |I_j^{(t)}| \cdot \one[|I_j^{(t)}| \leq x],
\]
for all $t,x \geq 0$ and define $\Aaatilde$ by $\Aaatilde = \Atilde - \Aatilde.$
By comparing with the case of $\alpha = 1,$ in which $g_t^\alpha = 1,$ we see $d\Aatilde[t]$ is absolutely continuous with respect to $d\Atilde[t]$ for all $t.$ Moreover, the Radon--Nikodym derivative $\frac{d\Aatilde[t]}{d\Atilde[t]}(x)$ is equal to the fraction of mass of intervals of length $x$ that are contained in $[0,\alpha].$ In particular, we may take $\sup_{t \geq 0, x\in [0,1]} \left| \frac{d\Aatilde[t]}{d\Atilde[t]}(x)\right| \leq 1.$
It is now possible to give a semimartingale decomposition of $\Aatilde.$
\begin{proposition}
\label{prop:semimartingale}
Suppose that the intervals $\left\{ I^{(0)}_j : 1 \leq j \leq n_0 \right\}$ have distinct lengths.
Suppose that $\Psi$ is continuous. %
The process $(\Aatilde, \Atilde)$ satisfies the equation
\[
\Aatilde[t](x)
=
\Aatilde[0](x)+
\int_0^t e^s x^2 \int_x^\infty
\frac{d\Aatilde[s]}{d\Atilde[s]}(z)
\cdot\frac{ d\Psi(\Atilde[s](z))}{z}
%
%
\,ds + \Matilde[t](x),
\]
for all $t,x \geq 0,$ where $\Matilde$ is a martingale.
\end{proposition}
\begin{remark}\label{rem:abscontinuity}
If in addition $\Psi$ is absolutely continuous, then one can rewrite the expression on the right-hand side in the above proposition. First note that for any real numbers $a < b,$ and any distribution function $F,$ we have $\int \one[(a,b)]d(\Psi \circ F) \leq |\Psi(F(b)) - \Psi(F(a))|.$ It follows for any $\epsilon>0$ there is a $\delta > 0$ so that for any open set $U$ with $\int_U dF < \delta$ then $\int_U d(\Psi \circ F) \leq \epsilon.$ So, by regularity $d\Psi \circ F$ is absolutely continuous with respect to $dF.$ Hence, by the Radon--Nikodym theorem, we can also write for any $x > 0$ and any $s \geq 0,$
\[
\int_x^\infty
\frac{d\Aatilde[s]}{d\Atilde[s]}(z)
\cdot\frac{ d\Psi(\Atilde[s](z))}{z}
=
\int_x^\infty
\frac{d\Psi(\Atilde[s])}{d\Atilde[s]}
\cdot\frac{d\Aatilde[s]}{d\Atilde[s]}
\cdot
\frac{d\Atilde[s](z)}{z}
=
\int_x^\infty
\frac{d\Psi(\Atilde[s])}{d\Atilde[s]}(z)
\cdot\frac{d\Aatilde[s](z)}{z}.
\]
If in addition $\Atilde[s]$ is absolutely continuous, then we may take for all $z \in [0,1]$
\begin{equation}\label{eq:jungian}
\frac{d\Psi(\Atilde[s])}{d\Atilde[s]}(z)
=\psi(\Atilde[s](z)),
\quad
\text{where}
\quad
\Psi(z) = \int_0^z \psi(u)\,du.
\end{equation}
Note that without the assumption that $\Atilde[s]$ is absolutely continuous, we may no longer
have \eqref{eq:jungian}. This can be seen for example in the simple case that
\[
\Atilde[s](z) =
\frac{1}{2}\one[z \geq 1/3]
+\frac{1}{2}\one[z \geq 2/3]
\quad
\text{and}
\quad
\Psi(u)=u^2.
\]
In particular \cite[Proposition 3]{Junge} is not correct as claimed.
%
Nonetheless, the results of \cite{Junge} appear to be fixable by using Proposition \ref{prop:semimartingale}.
%
\end{remark}
\begin{proof}
Under the assumption of distinct starting lengths, and from the definition of the splitting rule for the intervals (i.e.\ uniformly), almost surely all intervals $\{I^{(t)}_j : 1 \leq j \leq N_t+n_0\},$ have distinct lengths at all times $t \geq 0.$
Define a Poisson random measure $\Pi$ on $[0,\infty) \times [0,1]^2$ with intensity $e^t \,dt \otimes d\Psi(u) \otimes dv.$ Let $\ell_t(u) = \Atilde[t^{-}]^{-1}(u)$ for all $u \in [0,1],$ with $\Atilde[t^{-}]^{-1}$ the generalized right--continuous inverse of $\Atilde[t^{-}],$ i.e.\ for all $p \in (0,1],$
\begin{equation}\label{eq:quantile}
\Atilde[t^{-}]^{-1}(p) = \inf \left\{ x : \Atilde[t^{-}](x) > p \right\}.
\end{equation}
As the lengths are distinct, we may unambiguously define a bounded non--negative measurable function $k_t^{\alpha}$ so that $k_t^{\alpha}(\ell_t(u)) = g_t^{\alpha}(j)$ where $1 \leq j \leq N_t + n_0$ is the index so that $|I_j^{(t)}| = \ell_t(u).$
At a point $(s,u,v) \in \Pi$ the interval with length $\ell_s(u)$ is split into two intervals, one of fraction $v$ length of the whole. This affects the empirical distribution function $\Atilde$ by replacing an atom at $x=\ell_s(u)$ by two atoms at $x = v\ell_s(u)$ and at $x=(1-v)\ell_s(u).$ Hence if we let $h(v,\ell,x) = v \one[\ell v \leq x] + (1-v) \one[ \ell (1-v) \leq x],$
then we have the identity that for any $x \in [0,1]$ and $t \geq 0$
\[
\begin{aligned}
&\Aatilde[t](x) = \Aatilde[0](x) +
\sum_{(s,u,v) \in \Pi, s\leq t}
\tilde{B}^{\alpha}_s(u,v,x),
\quad
\text{where}
\\
&\tilde{B}^{\alpha}_s(u,v,x)
=
\ell_s(u) k_t^{\alpha}(\ell_s(u)) \one[ \ell_s(u) > x] h(v,\ell_s(u),x).
\end{aligned}
\]
We wish to form a semimartingale decomposition, and so we will integrate $\tilde{B}^{\alpha}$ against the intensity of $\Pi.$
We observe that for $x < \ell_s(u),$
\[
\int_0^1 h(v,\ell_s(u),x)\,dv = \left( \frac{x}{\ell_s(u)} \right)^2,
\]
and therefore
\[
\int_0^1
\int_0^1
\tilde{B}^{\alpha}_s(u,v,x)\,dv
d\Psi(u)
=
x^2
\int_0^1
\frac{k_t^{\alpha}(\ell_s(u))}{\ell_s(u)} \one[ \ell_s(u) > x]
d\Psi(u).
\]
By the explicit definition of $\Atilde[t^{-}]^{-1}$ in \eqref{eq:quantile}, for any $a < b \in \mathbb{R},$ with $F=\Atilde[t^{-}],$
\[
\left\{ u \in \mathbb{R} : F^{-1}(u) \in (a,b) \right\}
\subseteq
[F(a),F(b))
\subseteq
\left\{ u \in \mathbb{R} : F^{-1}(u) \in [a,b] \right\}
\]
From the continuity of $\Psi$ it follows that
\[
\int \one\{u : F^{-1}(u) \in (a,b]\}
\,d\Psi(u)
=
\Psi(F(b)) - \Psi(F(a))
=
\int \one\{ u \in (a,b]\}
d\Psi(F(u)).
\]
By a monotone class argument, it follows for all non--negative Borel measurable functions $f$ and continuous $\Psi,$
\[
\int
f(\Atilde[s^{-}]^{-1}(u))\,d\Psi(u)
=
\int
f(u)\,d\Psi(\Atilde[s^{-}](u)).
\]
Thus we conclude that
\[
\int_0^1
\int_0^1
\tilde{B}^{\alpha}_s(u,v,x)\,dv
d\Psi(u)
=
x^2
\int_x^\infty
\frac{d\Aatilde[s^{-}]}{d\Atilde[s^{-}]}(z)
\cdot
\frac{d\Psi(\Atilde[s^{-}](z))}{z}
\]
As we then integrate the expression in time, we may freely replace the $s^{-}$ by $s$
to conclude the proof.
\end{proof}
\noindent We also define $(\Aa,\Aaa,\A)$ and $\Ma$ by letting $\Aa[t](x) = \Aatilde[t](e^{t}x)$ for all $t,x \geq 0$ and similarily for $\Aaa,\A,$ and $\Ma.$ After this time change, we can write
\begin{equation}
\label{eq:dynamics}
\Aa[t](x)
=
\Aa[0](x)+
\int_0^t (e^{s-t}x)^2 \int_{e^{s-t}x}^\infty
\frac{d\Aa[s]}{d\A[s]}(z)
\cdot\frac{ d\Psi(\A[s](z))}{z}\,ds
%
+ \Ma[t](x),
\end{equation}
for all $t,x \geq 0.$ This motivates the study of the operator on pairs of time evolving distribution functions that appears as the drift term in this decomposition. To formalize this operator, we recall some notation from \cite{candy,Junge}.
Define the space $L^1_{\text{loc}}$ of locally integrable functions $f:[0,\infty)\to\mathbb{R}$, endowed with the following canonical metric $d_{\Lloc}$,
\[
d_{\Lloc}(f,g) = \sum_{k=1}^\infty 2^{-k} \wedge \int_0^k |f(x)-g(x)|\,dx,
\]
which makes $L^1_{\text{loc}}$ into a complete separable metric space. Define the subspace $\mathcal{D}\subsetL^1_{\text{loc}}$ of subdistribution functions by
\[
\mathcal{D} = \left\{
F : [0, \infty] \to [0,1],\text{ c\`{a}dl\`{a}g, increasing}
\right\}.
\]
In this paper we will reserve the term \emph{distribution function} for cumulative distribution functions of probability measures. Recall that convergence under $d_{\Lloc}$ of distribution functions is equivalent to vague convergence:
\begin{lemma}[Lemma 2.3 of \cite{candy}]
\label{lem:L1loc_pointwise}
For $F,F_1,F_2,\ldots\in\mathcal{D}$, $F_n\to F$ with respect to $d_{\Lloc}$ if and only if $F_n(x)\to F(x)$ at every point of continuity $x$ of $F$.
\end{lemma}
\noindent Lemma~\ref{lem:L1loc_pointwise} and Helly's selection theorem imply in particular that $(\mathcal{D},d_{\Lloc})$ is a compact metric space.
In \cite{candy,Junge}, the following norm played an important role. Define for $f\inL^1_{\text{loc}}$:
\[
\candy[f] = \int_0^\infty x^{-2}|f(x)|\,dx\in[0,\infty].
\]
Also define a subspace $\mathcal{D}_1\subset\mathcal{D}$ by
\[
\mathcal{D}_1 = \left\{
F\in\mathcal{D}: \candy[F] \leq 1
\right\}.
\]
Using $d_{x^{-2}},$ the space $\mathcal{D}_1$ becomes a complete metric space.
\begin{lemma}[Lemma of \cite{candy}]
\label{lem:complete_candy}
The metric space $(\mathcal{D}_1,d_{x^{-2}})$ is complete.
\end{lemma}
We now define the space $\mathcal B([0,\infty),L^1_{\text{loc}})$ of Borel measurable maps from $[0,\infty)$ to $L^1_{\text{loc}}$.
We endow this space with the topology of locally uniform convergence, which we denote by the symbol $ \overset{\DTSpace}{\to} $. Then $\F^{(n)} \overset{\DTSpace}{\to} \F$ as $n\to\infty$ if and only if for all compact $K \subseteq [0,\infty)$ and all $t > 0,$
\[
\lim_{n \to \infty} \sup_{0 \leq s \leq t} \int\limits_K | \F[s]^{(n)}(x) - \F[s](x) |\,dx = 0.
\]
The subspaces $\mathcal{X},\mathcal{X}_1\subset\mathcal B([0,\infty),L^1_{\text{loc}})$ are defined by
\[
\mathcal{X} = \mathcal B([0,\infty),\mathcal{D}),\quad \mathcal{X}_1 = \mathcal B([0,\infty),\mathcal{D}_1) \subset \mathcal{X}.
\]
Since $\mathcal{D}$ and $\mathcal{D}_1$ are closed subsets of $L^1_{\text{loc}}$, $\mathcal{X}$ and $\mathcal{X}_1$ are closed subsets of $\mathcal B([0,\infty),L^1_{\text{loc}})$.%
The spaces of continuous maps $C([0,\infty),\mathcal{D})$ and $C([0,\infty),L^1_{\text{loc}})$ are closed subsets of $\mathcal{X}$ and $B([0,\infty),L^1_{\text{loc}})$, respectively. Furthermore, the topology on these spaces can be metrized to make them complete separable metric spaces.
For measures $\mu$ and $\nu$, write $\mu\ll \nu$ if $\mu$ is absolutely continuous w.r.t.~$\nu$. Define a space $\mathcal{R}$ by
\[
\mathcal{R} = \left\{ (\F,\G) \in \mathcal{X} \times \mathcal{X} : d\G[t] \ll d\F[t]\,\forall t\ge0 \right\}.
\]
For \emph{absolutely continuous} $\Psi$ we define an operator
\(
\mathscr{C} \colon \mathcal{R} \to \mathcal{X}
\)
by the formula
\begin{align}\label{eq:cop}
\mathscr{C}( \F,\G)_t(x)
= {\G[0]}(e^{-t}x) + \int_0^t (e^{s-t}x)^2 \left[\int_{e^{s-t}x}^\infty
\frac{d\G[s]}{d\F[s]}(z)\cdot \frac{d(\Psi({\F[s]}(z)))}{z}
%
\right]\,ds
\end{align}
for all $t,x \geq 0.$ Note that the integral over $z$ is well defined since the Radon-Nikodym derivative $d\G[s]/d\F[s]$ is defined $d\F[s]$-almost surely and the measure $d\Psi\circ \F[s]$ is absolutely continuous w.r.t.~$d\F[s]$ (see Remark~\ref{rem:abscontinuity}).
This allows us to rewrite \eqref{eq:dynamics} as
\(
\Aa = \mathscr{C}( \A, \Aa) + \Ma.
\)
We would like this operator to be continuous. This is not true in general as a map from $\mathcal{R} \to \mathcal{X}.$ However, the operator is continuous at certain specific points.
\begin{proposition}
\label{prop:continuity}
Let $(\F^n,\G^n)$ be a sequence in $\mathcal{R}$ and let $(\F,\G)\in\mathcal{R}$. Suppose furthermore
\begin{enumerate}
\item $\Psi \in C^1[0,1],$
\item $\F^n \overset{\DTSpace}{\to} \F$ and $\G^n \overset{\DTSpace}{\to} \G$,
\item the map $(t,x) \mapsto \F[t](x)$ is continuous,
\item $\left\{ \F[t]: t \geq 0 \right\}$ is a tight family of distribution functions.
\end{enumerate}
Then, $(\F,\G)\in\mathcal{R}$ and
\[
\mathscr{C}(\F^n,\G^n) \overset{\DTSpace}{\to} \mathscr{C}(\F,\G).
\]
\end{proposition}
The proof of Proposition~\ref{prop:continuity} is postponed to Section~\ref{sec:proof_continuity}.
In the case that $\alpha = 1,$ \eqref{eq:dynamics} becomes
\(
\A = \mathscr{C}( \A, \A) + \M.
\)
In the notation of \cite{candy}, we write $\mathscr{S}^{\Psi} : \mathcal{X} \to C( [0,\infty), L^1_{\text{loc}}),$ which is given by $\mathscr{S}^{\Psi}(\F) = \mathscr{C}(\F,\F),$ so that $\A = \mathscr{S}^{\Psi}(\A) + \M.$
We also introduce the following family for $\mathscr{S}^{\Psi}$:
\[
\mathfrak{F} = \{\F \in \mathcal{X}_1: \F = \mathscr{S}^{\Psi}(\F), \forall t \geq 0: \F[t](+\infty) = 1 \text{ and } \{\F[t]\}_{t\geq 0} \text{ tight} \}.
\]
Here, we recall that a family of distribution functions $\{\F[\beta]\}_{\beta \in X}$ on $[0,\infty)$ is tight if for all $\epsilon > 0,$ there is an $N > 0$ sufficiently large such that, for every $\beta \in X,$ $\F[\beta](N) > 1-\epsilon.$
Of particular importance, as a consequence of Theorem~\ref{thm:candy} (more specifically, \cite[Lemma 3.5]{candy}), for all distributions $\Psi \in C^1[0,1],$ we have that there is a stationary fixed point $\F^* \in \mathfrak{F}$ which has $F^*_t = F^{\Psi}$ for all $t \geq 0.$ Following the insight of \cite{Junge} and adapting his formalism, we additionally introduce the operator
\(
\mathscr{C}^* \colon \mathcal{X}^* \to \mathcal{X}
\)
given by
\begin{align*}
&\mathcal{X}^* = \{\F\in \mathcal{X}_1: d\F[t] \ll dF^\Psi\,\forall t\ge0,\,\sup_{t\ge0}\|\tfrac{d\F[t]}{dF^\Psi}\|_{\operatorname{L}^\infty(dF^\Psi)} < \infty\}\\
&\mathscr{C}^*(\F) = \mathscr{C}(\F^*,\F),\quad \F\in\mathcal{X}^*
\end{align*}
and we introduce the family of fixed points
\[
\mathfrak{F}^* = \{\F \in \mathcal{X}^*: \F = \mathscr{C}^*(\F), \forall t \geq 0: \F[t](+\infty) = 1 \text{ and } \{\F[t]\}_{t\geq 0} \text{ tight} \}.
\]
The key contribution of this article is to show that in fact every solution in $\mathfrak{F}^*$ has the same $t \to \infty$ limit, $F^{\Psi},$ which we do by study of a related piecewise deterministic Markov process in Section~\ref{sec:growth-fragmentation}. The following proposition is a special case of Proposition~\ref{prop:convergence_RFPSpace}:
\begin{proposition}
\label{prop:ergodicity}
Let $\F \in \mathfrak{F}^*$. Then $d\F[t]$ converges in total variation to $dF^{\Psi}.$ Furthermore, if $\mathcal M$ is a tight family of probability distributions on $(0,\infty)$, the convergence is uniform on those $\F$ which furthermore satisfy $d\F[0] \in \mathcal M$.
\end{proposition}
It is now relatively simple to complete the proof of Theorem~\ref{thm:main}. We need the following simple facts about the dynamics which are easy adaptations of arguments in \cite{candy,Junge}.
\begin{proposition}
\label{prop:evo_prop}
For $\Psi \in C^1[0,1],$ and any $\alpha \in (0,1],$ the following hold almost surely:
\begin{enumerate}[(i)]
\item The collection of distribution functions $\{\alpha^{-1}\Aa[t]\}_{t \geq 0}$ is tight. \label{p:tight}
\item The family $\{\alpha^{-1}\Aan\}_{n \geq 0}$ defined by $\Aan[t] = \Aa[t+n]$ for every $t\ge0$ is \emph{asymptotically equicontinuous}, i.e.\,for any $K$ compact
\[
\lim_{\delta \to 0}
\limsup_{t_0 \to \infty}
\sup_{\substack{s,t \geq t_0 \\ |s-t| \leq \delta} }
\int\limits_K \left| \Aa[s](x) - \Aa[t](x) \right|\,dx=0.
\] \label{p:aec}
\item The noise vanishes in the limit, i.e.~$\Man \overset{\DTSpace}{\to} 0$ as $n \to \infty$, where $\Man[t] = \Ma[t+n] - T_t \Ma[n]$ for every $t\ge0$, where $T_tf(x) = f(e^tx).$ \label{p:noise}
\end{enumerate}
\end{proposition}
\begin{proof}
This is a simple modification of \cite[Proposition 7.2]{candy} or \cite[Proposition 5]{Junge}.
\end{proof}
With these properties in hand, it is now elementary to prove the following convergence:
\begin{proposition}
\label{prop:candyass}
For each $\alpha \in [0,1],$ we have that
\[
\lim_{t \to \infty} \Aa[t] = \alpha F^{\Psi}
\]
almost surely.
\end{proposition}
\begin{proof}
Let $\alpha\in(0,1]$ (the case $\alpha=0$ is trivial).
Let $\left\{ t_k \right\}_{k=0}^\infty$ be an arbitrary sequence of times converging to $\infty$. It is enough to show that $d_{\Lloc}(\alpha^{-1} \Aa[t_k],F^{\Psi})\to0$ along a subsequence. Fix $\epsilon > 0$. Let $\mathcal{H}$ be the $d_{\Lloc}$ closure of $\left\{ \alpha^{-1}\Aa[t] \right\}_{t \geq 0}.$ From Proposition~\ref{prop:evo_prop}(\ref{p:tight}) and the fact that for each $F \in \mathcal{H},$
\[
\candy[F] \leq \sup_{t \geq 0}\alpha^{-1}e^{-t}N_t < \infty,
\]
we have that $\mathcal{H}$ is tight as a family of distributions on $\left( 0,\infty \right).$ As a consequence, Proposition~\ref{prop:ergodicity} yields that there exists $T>0$, such that
\begin{equation}
\label{eq:suplate}
\sup_{\F\in\mathfrak{F}^*: \F[0] \in \mathcal H}d_{\operatorname{TV}}(d\F[T],\pi) < \epsilon.
\end{equation}
Define $n_k = t_k - T$ for all $k \in\mathbb{N}$ (assume w.l.o.g.~that $t_k\ge T$ for all $k$). For $s\ge0$, let $\Aas = (\Aas[t])_{t \geq 0}=(\Aa[t+s])_{t \geq 0}$ be the shifted process. By Proposition~\ref{prop:evo_prop}(\ref{p:aec}), the family $\left\{ \Aank \right\}_{0}^\infty$ is asymptotically equicontinuous. Hence by \cite[Lemma 7.3]{candy}, we may extract a subsequence $\left\{ \Aankt \right\}_{0}^\infty$ which converges in $\mathcal{X}$ to $\alpha \F^{\alpha}$ for some $\F^{\alpha}$. In particular, with $ t'_k = n'_k + T$, we have for all $k$ sufficiently large,
\begin{equation}
\label{eq:itslate}
d_{\Lloc}(\alpha^{-1} \Aa[ t'_k], \F[T]^{\alpha}) =
d_{\Lloc}(\alpha^{-1} \Aankt[T], \F[T]^{\alpha}) < \epsilon.
\end{equation}
Furthermore, we have
\[
\alpha^{-1}\Aankt = \mathscr{C}\left(
\Ankt, \alpha^{-1}\Aankt
\right)
+ \Mankt \overset{\DTSpace}{\to} \F^{\alpha}.
\]
Since $\Aa[t]$ has Radon-Nikodym derivative bounded by 1 w.r.t.~$\A[t]$ for every $t\ge0$ and $\Ankt \to \F^*$ as $k\to\infty$, $\F[t]^\alpha$ has Radon-Nikodym derivative bounded by $\alpha^{-1}$ w.r.t.~$F^\Psi$ for every $t\ge0$. Hence, $\F^\alpha \in \mathcal{X}^*$ and $(\F^*,\F^\alpha)\in\mathcal R$.
By Proposition~\ref{prop:evo_prop}(\ref{p:noise}), $\Mankt \overset{\DTSpace}{\to} 0.$ By Theorem~\ref{thm:candy}, we have that $\A[t] \to F^{\Psi}$ almost surely, and hence $\Ankt \overset{\DTSpace}{\to} \F^*,$ which is jointly continuous in time and space. Then by Proposition~\ref{prop:continuity}, we have that
\[
\F^{\alpha} = \mathscr{C}(\F^*, \F^{\alpha}) = \mathscr{C}^*(\F^{\alpha}).
\]
By definition of the space $\mathcal H$, $\F[t]^\alpha \in \mathcal H$ for every $t\ge0$, in particular, the family $\{\F[t]^\alpha\}_{t\geq0}$ is a tight family of distribution functions of probability measures, and so $\F^\alpha\in \mathfrak{F}^*$.
Hence, \eqref{eq:suplate} implies
\begin{equation}
\label{eq:itsverylate}
d_{\operatorname{TV}}\left( dF_T^\alpha, \pi \right) < \epsilon.
\end{equation}
The $d_{\Lloc}$ distance of two distribution functions on $[0,\infty)$ can easily be estimated in terms of the total variation distance of the underlying measures. By \eqref{eq:itslate} and \eqref{eq:itsverylate}, we then get that there is some monotone increasing function $f : \mathbb{R}^+ \to \mathbb{R}^+$ with $\lim_{x \to 0}f(x) = 0$ so that
\(
d_{\Lloc}(\alpha^{-1} \Aa[t'_k],\pi) < \epsilon + f(\epsilon)
\).
As $\epsilon$ can be made arbitrarily small, the proof is complete.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{thm:main}]
Recall that we wish to show that $e^{-t}N_t^{\alpha}\to \alpha$ almost surely as $t\to\infty$, for every $\alpha\in[0,1]$. The key is the following chain of inequalities:
\begin{align}
\label{eq:sandwich}
e^{-t}(N_t^{\alpha}-1) \leq \candy[\Aa[t]] \leq e^{-t}N_t^{\alpha},
\end{align}
which are easy consequences of the following identity for size-biased distribution functions \cite[Lemma 2.2]{candy}:
\[
\candy[F] = \int_0^\infty x^{-1} dF(x),
\]
This identity furthermore implies $\candy[\A[t]] = e^{-t}N_t$.
Fix $\alpha\in[0,1]$. Recall that $\candy[F^{\Psi}] = 1$ \cite[Lemma~3.5]{candy}. Together with Proposition \ref{prop:candyass} and Fatou's lemma, we have
\[
\liminf_{t\to\infty} \candy[\Aa[t]] \geq \alpha
\]
almost surely. By symmetry under reversing the interval $[0,1],$ it also follows that $\liminf_{t\to\infty} \candy[\Aaa[t]] \geq (1-\alpha).$ However, by the above, almost surely,
\[
\lim_{t \to \infty} \candy[\Aa[t]]+\candy[\Aaa[t]]
=
\lim_{t \to \infty} \candy[\A[t]]
=
\lim_{t \to \infty} e^{-t}N_t
= 1,
\]
whence
\[
\lim_{t\to\infty} \candy[\Aa[t]] = \alpha
\]
almost surely. Equation~\eqref{eq:sandwich} now gives
\[
\lim_{t\to\infty} e^{-t}N_t^{\alpha} = \alpha,
\]
which was to be proven.
\end{proof}
\section{The cell process}
\label{sec:growth-fragmentation}
This section can be read independently of the remainder of the article. Its goal is to prove Proposition~\ref{prop:ergodicity}, but in order to simplify notation and to make logical dependencies clearer we state a more general result, Proposition~\ref{prop:convergence_RFPSpace}. Throughout the section, we fix a non-negative function $R\in L^1_{\text{loc}}.$ For the application to $\Psi$--processes, we will set $R = \psi \circ F^{\Psi}$, however, the results established in this section will not require any boundedness or continuity properties of the function $R$.
Define the operator $\mathscr{R}$ given by
\[
\mathscr{R}(\F)_t(x)
= {\F[0]}(e^{-t}x) + \int_0^t (e^{s-t}x)^2 \left[\int_{e^{s-t}x}^\infty \frac{R(z)}{z}\,d\F[s](z)\right]\,ds,
\]
for all $\F \in \mathcal{X}_{\mathscr{R}} = \{\F\in\mathcal{X}: d\F[s] \ll \mathrm{Leb}\}$, where $\mathrm{Leb}$ is Lebesgue measure on $\mathbb{R}_+$. In parallel with previous definitions, define
\[
\mathfrak{R} = \{\F \in \mathcal{X}_{\mathscr{R}}: \F = \mathscr{R}(\F), \forall t \geq 0: \F[t](+\infty) = 1,\ \F[t](0)=0 \text{ and }
\{\F[t]\}_{t\geq 0} \text{ tight} \}.
\]
Throughout the section, we will work under the following assumption, which will be satisfied in the cases in which we are concerned.
\begin{assumption}\
\label{ass:stat}
There is a stationary element $\F \in \mathfrak{R},$ that is $\F[t] = F$ for some distribution function $F$ and all $t \geq 0.$ Then $F$ is the distribution function of a probability measure, which we denote by $\pi$.
\end{assumption}
\noindent This assumption is equivalent to \eqref{eq:ZR}, as we shall show below in Lemma \ref{lem:R}.
We also define a subspace of $\mathfrak{R}$ by
\[
\mathfrak{R}_{\operatorname{ac}} = \{\F \in \mathfrak{R}: \sup_{t \geq 0} \| \tfrac{d\F[t]}{d\pi}\|_{\operatorname{L}^\infty(\pi)} < \infty \}.
\]
These naturally arise in our application. Our goal in this section is to show the following proposition:
\begin{proposition}
\label{prop:convergence_RFPSpace}
Under Assumption~\ref{ass:stat}, every $\F\in\mathfrak{R}_{\operatorname{ac}}$ satisfies
$$
d_{\operatorname{TV}}(d\F[t],\pi)\to 0,\quad\text{as $t\to\infty$.}
$$
Furthermore, if $\mathcal M$ is a tight family of probability distributions on $(0,\infty)$, the convergence is uniform on the set $\{\F\in\mathfrak{R}_{\operatorname{ac}}: d\F[0] \in \mathcal M\}$.
\end{proposition}
We now describe the global structure of the proof, which is split over several subsections. Let us first define the operator $\mathscr{L}: C_c^1(0,\infty) \to \Lpi$ by,
\begin{equation}
\label{eq:RG}
\mathscr{L} f(x) = xf'(x) + xR(x)\left[\int_0^x\frac{2u}{x^2}f(u)\,du - f(x)\right],
\end{equation}
where the fact that $\mathscr{L} f \in \Lpi$ easily follows from \eqref{eq:integral_R}.
We start by showing in Section~\ref{sec:diff_mu} that for every $F\in \mathfrak{R}$, the family of measures $\mu_t = d\F[t]$ solves the equation
\begin{equation}
\label{eq:diff_mu}
(\mu_t,f) = (\mu_0,f) + \int_0^t
\left( \mu_s,\mathscr{L} f \right)\,ds,\quad \forall f \in C_c^1(0,\infty),\ \forall t \geq 0,
\end{equation}
where $(\mu,f):= \int f\,d\mu$ for every $\mu$ and $f$.
This motivates Section~\ref{sec:construction_markov_process}, where we construct a piecewise deterministic Markov process, \emph{the cell process}, and identify $\mathscr{L}$ as its infinitesimal generator (in a certain sense). This process can be seen as describing the evolution of the tagged fragment in a certain growth--fragmentation process (see \cite{Bertoin}), but we will not exploit this relation further. We also show in that section that $\pi$ is an invariant measure of the Markov process.
The goal of the next section, Section~\ref{sec:Lpi_and_uniqueness}, is to show that every solution $(\mu_t)_{t\ge0}$ to \eqref{eq:diff_mu} satisfies $\mu_t = \mu_0 P_t$, i.e.~that $\mu_t$ is the law of the Markov process at time $t$, when the starting point is distributed according to $\mu_0$. The arguments required for this are of analytic nature. We first show that the transition operator of the Markov process defines a strongly continuous contraction semigroup $(P_t)_{t\ge0}$ on $\Lpi$. The main part is then to prove that the space $C_c^1(0,\infty)$ is a core of the domain of its generator. For this, we adapt ideas from the growth-fragmentation literature (especially the decomposition approach of \cite{Mischler2016}). This then allows to prove that every solution $(\mu_t)_{t\ge0}$ to \eqref{eq:diff_mu} is uniquely determined from its initial condition and indeed given by $\mu_t = \mu_0P_t$.
Having established that $\mu_t$ is the law of the Markov process at time $t$ starting from the law $\mu_0$, the last step is to show that this process is ergodic. This is done in Section~\ref{sec:ergodicity} using purely probabilistic arguments based on coupling of regenerative processes. Again, no additional assumptions on $R$ are necessary, ergodicity follows from the mere existence of a stationary probability, together with certain irreducibility properties. In particular, no Lyapunov functions are used. With everything in place, Section~\ref{sec:proof_prop} wraps up the proof of Proposition~\ref{prop:convergence_RFPSpace}.
To finish this section, let us collect some consequences of Assumption~\ref{ass:stat} which will be of use later. First, under Assumption~\ref{ass:stat}, it is easy to check that $F$ is in fact continuously differentiable in $(0,\infty)$ and satisfies the identity
\begin{equation}
\label{eq:Fprime}
\frac{F'(x)}{x} = \int_x^\infty R(z)\,\frac{F'(z)}{z}\,dz
\end{equation}
for every $x > 0.$ Solving \eqref{eq:Fprime} for $F'(x)/x$ (see also the proof of Corollary 9.3 in \cite{candy}) then shows that for every $0<x\le y<\infty$,
\begin{equation}
\label{eq:Fprime_exp}
\frac{F'(y)}{y} = \frac{F'(x)}{x} \exp\left(-\int_x^y R(z)\,dz\right).
\end{equation}
In particular, since $F'$ cannot be zero everywhere by the assumptions on $F$, $F'(x)> 0$ for all $x\in(0,\infty)$. In particular, $\pi$ is equivalent to Lebesgue measure on $(0,\infty)$.
\begin{lemma} \label{lem:R}
Assumption \ref{ass:stat} holds if and only if \eqref{eq:ZR} holds.
\end{lemma}
\noindent We shall not use this Lemma in the development that follows. It is only provided to show that \ref{ass:stat} is a relatively simple condition.
\begin{proof}
Taking $x=1$ in \eqref{eq:Fprime_exp} shows that
\[
F'(y) = y F'(1)\exp\left(-\int_1^y R(z)\,dz\right)
\]
Hence on integrating both sides over all $y$, we conclude that
\[
1 = F'(1) \int_0^\infty y \exp\left(-\int_1^y R(z)\,dz\right)\,dy
= F'(1)Z^R.
\]
As $F'$ is everywhere positive, it follows that $Z^R < \infty.$
On the other hand, if $Z^R < \infty,$ then we can define the probability density
\[
f(x) = x\exp\left(-\int_1^x R(y)\,dy\right)/Z^R, \quad \text{ for all } x > 0.
\]
Letting $F$ be its cumulative distribution function, it is elementary to see that $\F=(\F[t])_{t\geq 0}$ with $\F[t] = F$ for all $t \geq 0$ is a stationary element of $\mathfrak{R}.$
\end{proof}
Also note that the positivity of $F'$ together with \eqref{eq:Fprime} implies the following:
\begin{equation}
\label{eq:R_positivity}
\forall x\in(0,\infty): \int_x^\infty R(y)\,dy > 0.
\end{equation}
Actually, the last integral can be shown to be infinite, but we will not use this fact explicitly.
Furthermore, multiplying both sides in \eqref{eq:Fprime} by $x$, then integrating over all $x \geq 0$ and applying Fubini's theorem, we can additionally conclude that
\begin{equation}
\label{eq:integral_R}
\int_0^\infty z R(z)\,\pi(dz) = 2\int_0^\infty \,\pi(dx) = 2.
\end{equation}
\subsection{Proof of (\ref{eq:diff_mu})}
\label{sec:diff_mu}
The goal of this section is to prove the following proposition:
\begin{proposition}
\label{prop:markov}
For every $\F \in \mathfrak{R},$ $f \in C_c^1(0,\infty)$ and all $t \geq 0,$ the following holds:
\[
(d\F[t],f) = (d\F[0],f) + \int_0^t
\left( d\F[s],\mathscr{L} f \right)\,ds,
\]
i.e.~the family of measures $(d\F[t])_{t\ge0}$ solves \eqref{eq:diff_mu}.
%
\end{proposition}
\begin{proof}
Throughout the proof, fix $\F \in \mathfrak{R}$ and $f \in C_c^1(0,\infty)$. Define the semigroup $(T_t)_{t\ge0}$ by $T_tf(x) = f(e^t x)$. Furthermore, define the operator
\[
{\tilde{\mathscr{L}}} f(x) = \mathscr{L} f(x) - xf'(x) = xR(x)\left[\int_0^x\frac{2u}{x^2}f(u)\,du - f(x)\right].
\]
We now first show the following fact:
For all $t \geq 0,$
\begin{equation}
\label{eq:diff_aux}
(d\F[t], f) = (d\F[0],T_tf) + \int_0^t
\left( d\F[s],{\tilde{\mathscr{L}}} T_{t-s}f \right)\,ds,
\end{equation}
This equation will follow directly from the definition of $\mathfrak{R}$. Using the boundary conditions on $f,$ we can write
\[
(d\F[t], f)
=\int_0^\infty f(x)\,dF_t(x)
=-\int_0^\infty f'(x) F_t(x)\, dx.
\]
We may now use the definition of $\mathscr{R}$ to write
\[
(d\F[t], f)
=
-\int\limits_0^\infty F_0(e^{-t}x) f'(x)\,dx
-\int\limits_0^\infty f'(x)\int\limits_0^t (e^{s-t}x)^2 \left[\int_{e^{s-t}x}^\infty \frac{R(z)}{z}\,d\F[s]\right]\,ds
\,dx.
\]
The first term we integrate by parts and change variables. To the second term, we apply Fubini to bring the $t$ integral to the outside and the $x$ integral to the inside (the justification for the change of order of integration follows readily from $\F = \mathscr{R}(\F)$). This gives
\[
(d\F[t], f)
=
\int\limits_0^\infty f(e^{t}x)\,dF_0(x)
-
\int\limits_0^t
\int\limits_0^\infty
\frac{R(z)}{z}
\int_0^{e^{t-s}z}
f'(x) (e^{s-t}x)^2
\,dx
d\F[s](z)
ds.
\]
Finally, integrating the $x$ integral by parts, we arrive at
\[
(d\F[t], f)
=
\int\limits_0^\infty f(e^{t}x)\,dF_0(x)
+
\int\limits_0^t
\int\limits_0^\infty
\frac{R(z)}{z}
\biggl[
\int_0^{z}
2xf(e^{t-s}x)\,dx
-z^2f(e^{t-s}z)\biggr]
d\F[s](z)
ds,
\]
which is \eqref{eq:diff_aux}.
We now claim that the map $t \mapsto (d\F[t], f)$ is continuously differentiable. To see this, we begin by noting that the equation $t \mapsto (d\F[t],f)$ is continuous by \eqref{eq:diff_aux}. Moreover, for fixed $f,$ the function
\[
H(s,t) = (d\F[s],{\tilde{\mathscr{L}}} T_{t-s}f)
\]
is continuous in $s$ and $t,$ and in fact differentiable in $t$ as can be seen by differentiating under the integral sign in $(d\F[s],{\tilde{\mathscr{L}}} T_{t-s}f).$ Specifically, we have that
\[
\frac{|{\tilde{\mathscr{L}}}[T_{b}f](x)-{\tilde{\mathscr{L}}}[T_{b}f](x)|}{|b-a|}
\leq xR(x)C_f,
\]
for all $a,b \in (0,\infty)$ some constant $C_f$ depending only on $f.$ Thus, the differentiation is justified, and we have that
\[
\partial_t H(s,t) = (d\F[s],{\tilde{\mathscr{L}}} \partial_t(T_{t-s}f))
=(d\F[s],{\tilde{\mathscr{L}}} T_{t-s}[xf']),
\]
which is continuous in $s.$ Hence, we get from \eqref{eq:diff_aux} that $t \mapsto (d\F[t], f)$ is continuously differentiable and moreover
\begin{align*}
\partial_t (d\F[t], f)
&= (d\F[0],T_t[xf']) +
\int_0^t
\left( d\F[s],{\tilde{\mathscr{L}}} T_{t-s}[xf'] \right)\,ds
+
\left( d\F[t],{\tilde{\mathscr{L}}} f \right) \\
\intertext{The first two terms can be combined by \eqref{eq:diff_aux}, however, to give:}
\partial_t (d\F[t], f)
&=(d\F[t], xf')+( d\F[t],{\tilde{\mathscr{L}}} f ) \\
&=\left(d\F[t], \mathscr{L} f \right).
\end{align*}
Hence, on integrating this expression, we arrive at the statement of the proposition.
\end{proof}
\subsection{Construction of the cell process}
\label{sec:construction_markov_process}
Corresponding to the operator $\mathscr{L},$ we construct a \emph{piecewise deterministic Markov process (PDMP)} $X = (X_t)_{t\ge0}$ on $(0,\infty),$ which we refer to as the \emph{cell process}. A PDMP is a Markov process that almost surely has a finite number of jumps in any finite time interval and which moves deterministically between the jumps. In our case, denote by $(x,t)\mapsto \Phi(x,t)$ the flow generated by the vector field $x\mapsto x\frac{\partial}{\partial x}$. Then $\partial_t \Phi(x,t) = \Phi(x,t)$ and $\Phi(x,0) = x$, so that $\Phi(x,t) = xe^{t}$.
Simply stated, the cell process flows from $x$ along the integral curve $xe^{t}$ until some random time (regulated by the \emph{jump rate} $r(x) = xR(x)$) at which point the process jumps downwards by a multiplicative factor whose distribution is the size bias of $\operatorname{Unif}[0,1].$ This procedure then restarts, flowing upwards and jumping downwards.
It will be useful to define the process precisely, making use of the special features of the process which allow for a simple construction. The state space of our Markov process will be the interval $[0,\infty)$, the point $0$ being an absorbing state. The process will be defined on a probability space $(\Omega,\Pr)$ supporting an iid sequence $(U_1,U_2,\ldots)$ of r.v.~uniformly distributed in $(0,1)$. The process will be denoted by $X^x = (X^x_t)_{t\ge0}$, where $x\in [0,\infty)$ is the starting point. The definition goes as follows.
If the starting point is $x = 0$, we set $X^x_t = 0$ for all $t\ge0$. Suppose now that $x\in (0,\infty)$.
Define for $z\in(0,\infty)$ the \emph{survivor function} $\surv_z(t)$ given by
\[
\surv_z(t) = \exp\left( - \int_0^t T_s[r](z) ds \right) = \exp\left(- \int_z^{ze^t} R(u)\,du\right).
\]
Note that $\surv_z(t) > 0$ for all $z\in(0,\infty)$ and $t\ge0$, by the local integrability of $R$.
We construct a sequence of pairs $(\tau_k,J_k)_{k=1}^\infty$, where $\tau_k$ will be the time of the $k$-th jump, and $J_k$ the multiplicative factor of the jump of the process.
These random variables also depend on $x$, but we suppress this from the notation.
The construction is done recursively: Set $\tau_0 := 0$ and $Y_0=0$. Let $k\in\mathbb N=\{1,2,\ldots,\}$. We define $(\tau_k,J_k,Y_k)$ as follows:
\begin{itemize}
\item If $\tau_{k-1} = \infty$, then $\tau_k = \infty$ as well, otherwise
\[
\tau_k = \tau_{k-1} + \surv_{Y_{k-1}}^{-1}(U_{2k-1}),
\]
where $Y_k = x e^{\tau_k} \prod_{j=1}^k J_j$ and $ \surv_{x}^{-1}$ is the generalized inverse of the function $\surv_x$. Note in particular that since $J_j > 0$ for all $j$ and $\surv_{x}(t) > 0$ for all $x>0$ and $t\ge0$, we have $\tau_k > \tau_{k-1}$ for all $k$.
\item $J_k = \sqrt{U_{2k}},$ i.e.~$J_k$ takes values in $(0,1)$ and its law has density $u\mapsto 2u$ on $(0,1)$. We can and will assume that $\prod_{k=1}^\infty J_k = 0$ for all $\omega\in \Omega$.
\end{itemize}
We then define the process $X^x = (X^x_t)_{t\ge0}$ by
\begin{equation}
\label{eq:def_X}
X^x_t =
\begin{cases}
xe^t\prod_{j=1}^k J_j = e^{t-\tau_k} Y_k & \text{if $\tau_k \le t < \tau_{k+1}$ for some $k\in\mathbb N_0$}\\
0 & \text{if $t\ge \zeta \coloneqq \lim_{k\to\infty} \tau_k$}.
\end{cases}
\end{equation}
Note that in particular, $X^x_{\tau_k} = Y_k$ for all $k\in\mathbb N_0$.
Finally, since $J_j >0$ for all $j$, we note the equivalence (if the starting point $x \ne 0$):
\begin{equation}
\label{eq:equivalence}
X^x_t = 0 \quad\text{iff}\quad t\ge \zeta.
\end{equation}
\emph{From now on, following common usage, we will rather work with a single stochastic process $X = (X_t)_{t\ge0}$ defined on a measurable space $(\Omega,\mathcal A)$, endowed with a family of probability measures $(\Pr_x)_{x\ge0}$, such that under $\Pr_x$, $X$ has the law of $X^x$. }We note that $(\Pr_x)_{x\ge0}$ is a probability kernel, i.e. it is measurable in $x$, since $X^x$ is a measurable function of the random variables $(\tau_k,J_k)_{k=1}^\infty$, which are easily seen by induction to be measurable w.r.t.~$x$; the measurability of $\Pr_x$ then is a consequence of (part of) Fubini's theorem.
Denote by $\mathscr{F} = (\mathscr{F}_t)_{t\ge0}$ the natural filtration of the process $X$. Further, set for all $t,x\ge0$ and every bounded Borel function $f:[0,\infty)\to\mathbb{R}$:
\begin{equation}
\label{eq:P_t}
P_tf(x) \coloneqq \mathbb E_x[f(X_t)].
\end{equation}
Finally, let $\mathcal{C}_{0,b}$ be the space of bounded continuous functions on $[0,\infty)$ that are $0$ at $0,$ made into a Banach space using the supremum norm.
\begin{proposition}
\label{prop:X_Markov}
\
\begin{enumerate}[(i)]
\item
The process $X$ is a homogeneous strong Markov process on $[0,\infty)$ with paths in the Skorohod space $D([0,\infty))$ of c\`{a}dl\`{a}g functions on $[0,\infty)$ and whose semigroup of transition operators is $(P_t)_{t\ge0}$. In other words, for every $x\ge0$, every $\mathscr{F}$-stopping time $\tau$ and every bounded Borel function $f:[0,\infty)\to\mathbb{R}$, we have
\[
\mathbb E_x[f(X_{\tau+t})\,|\,\mathscr{F}_\tau] \mathbf{1}_{\tau<\infty} = P_tf(X_\tau)\mathbf{1}_{\tau<\infty}.
\]
\end{enumerate}
Moreover, under Assumption~\ref{ass:stat}, we have the following:
\begin{enumerate}[(i)]
\setcounter{enumi}{1}
\item The probability measure $\pi$ is invariant for the Markov process $X$.
\item For every $f\in C_c^1(0,\infty)$, we have
$$
\lim_{t\to\infty} t^{-1}(P_t f - f) = \mathscr{L} f,\quad\text{in $\Lpi$}.
$$
\end{enumerate}
\end{proposition}
\begin{proof}
Point (i) is proven in \cite[Chapter 25]{Davis1993} for general PDMP under the assumption that for every $x\in [0,\infty)$, the expected number of jumps of the process started from $x$ is finite in every finite time interval. This assumption is
difficult to verify \emph{a priori} (and possibly false), as the jump rate may not be bounded near 0. We therefore localize the construction, i.e.~we construct a sequence of approximating processes through an appropriate family of stopping times which satisfy this assumption.
Assume $x\in(0,\infty)$, the case $x=0$ being trivial. Then define for every $n\in\mathbb N$, $n> 1/x$,
\[
T^n = \inf\{t: X_t \le 1/n\},\quad\text{and}\quad X^n_t = \begin{cases} X_t & \text{if $t<T^n$}\\ 0 & \text{otherwise}\end{cases}.
\]
Since the process can decrease only by jumps, we have the following fact:
\begin{equation}
\label{eq:Tn_tauk}
\text{If $T^n < \infty$, then there exists $k\in\mathbb N$, such that $T^n = \tau_k$.}
\end{equation}
Inspecting the construction of a PDMP in \cite[Chapter 24]{Davis1993}, we can then identify the process $X^n$ as a PDMP on $\{0\}\cup(1/n,\infty)$ which informally evolves as follows:
\begin{itemize}
\item In $[1/n,\infty)$, it moves according to the flow $\Phi$,
\item jumps at rate $r$,
\item according to the jump kernel $Q(x,A) = \Pr(xJ_1\mathbf{1}_{xJ_1 \ge 1/n} \in A)$, where $J_1$ is defined as above.
\item The state $0$ is an absorbing state, i.e. the process does not move and the jump rate is 0.
\end{itemize}
To be precise, in \cite{Davis1993}, the state space of the PDMP is supposed to be an open subset of $\mathbb{R}^d$, but this can be remedied by considering the enlarged state space $(-\varepsilon,\varepsilon)\cup(1/n,\infty)$ for small enough $\varepsilon>0$, and make all states in $(-\varepsilon,\varepsilon)$ absorbing. In Lemma~\ref{lem:existence_n} below, we show that the process $X^n$ satisfies the assumption of finite expected number of jumps from \cite{Davis1993}.
Theorem~25.5 in \cite{Davis1993} then states that $X^n$ is indeed a homogeneous strong Markov process (with c\`{a}dl\`{a}g paths), i.e.\ if we define $P^n_t f(x) = \mathbb E_x[f(X^n_t)]$ and denote by $\mathscr{F}^n = (\mathscr{F}^n_t)_{t\ge0}$ the natural filtration of $X^n$, then for every $\mathscr{F}^n$-stopping time $\tau^n$,
\begin{equation*}
\mathbb E_x[f(X^n_{\tau^n+t})\,|\,\mathscr{F}^n_{\tau^n}] \mathbf{1}_{\tau^n<\infty} = P_tf(X^n_{\tau^n})\mathbf{1}_{\tau^n<\infty}.
\end{equation*}
Note that if $\tau$ is a $\mathscr{F}$-stopping time, then $\tau\wedge T^n$ is a $\mathscr{F}^n$-stopping time. Furthermore, we have $\mathscr{F}^n_{\tau\wedge T^n} = \mathscr{F}_{\tau\wedge T^n}$. Finally, observe that $X^n$ depends on $\mathscr{F}_\tau$ only through $\mathscr{F}_{\tau\wedge T^n}$, since $X^n_t = 0$ for $t\ge T^n$. Combining these observations gives for every $\mathscr{F}$-stopping time $\tau$,
\begin{equation}
\label{eq:markov_Xn}
\mathbb E_x[f(X^n_{\tau\wedge T^n+t})\,|\,\mathscr{F}_\tau] \mathbf{1}_{\tau\wedge T^n<\infty} = P_tf(X^n_{\tau\wedge T^n})\mathbf{1}_{\tau\wedge T^n<\infty}.
\end{equation}
We now decompose \eqref{eq:markov_Xn} according to whether $\tau < T^n$ or $\tau \ge T^n$. The left-hand side then equals,
\begin{align*}
\mathbb E_x[f(X_{\tau+t})\,|\,\mathscr{F}_\tau] \mathbf{1}_{\tau < T^n} + f(0) \mathbf{1}_{T^n \le \tau < \infty},
\end{align*}
and the right-hand side equals
\begin{align*}
P_tf(X_\tau) \mathbf{1}_{\tau < T^n} + f(0) \mathbf{1}_{T^n \le \tau < \infty},
\end{align*}
Passing to the limit $n\to\infty$ then gives
\begin{equation}
\label{eq:markov_X_pre}
\mathbb E_x[f(X_{\tau+t})\,|\,\mathscr{F}_\tau] \mathbf{1}_{\tau < \lim_{n\to\infty} T^n} = P_tf(X_\tau) \mathbf{1}_{\tau < \lim_{n\to\infty} T^n}
\end{equation}
We claim that $X_t = 0$ for all $t\ge\lim_{n\to\infty}T^n$. Indeed, by \eqref{eq:Tn_tauk}, $(T^n)_{n\ge1}$ is a subsequence of $(\tau_k)_{k\ge1}$, hence
$$\lim_{n\to\infty} T^n = \lim_{k\to\infty} \tau_k = \zeta.$$
The claim now follows from \eqref{eq:equivalence}.
Hence, adding $f(0) \mathbf{1}_{\zeta \le \tau < \infty}$ to both sides of \eqref{eq:markov_X_pre} yields
\[
\mathbb E_x[f(X_{\tau+t})\,|\,\mathscr{F}_\tau] \mathbf{1}_{\tau < \infty}= P_tf(X_\tau) \mathbf{1}_{\tau < \infty},
\]
which was to be proven.
It remains to show that $X$ is c\`{a}dl\`{a}g. Looking at the definition of $X$ in \eqref{eq:def_X}, this is obvious on $[0,\zeta)$, since only finitely many jumps occur in every interval $[0,a]$ for $a<\zeta$. Since $X_t = 0$ for $t\ge \zeta$ (by \eqref{eq:equivalence}), it remains to show that $X$ has a left-hand limit at $0$. But since $\tau_k$ is strictly increasing in $k$ and the function $t\mapsto e^t$ is continuous, we have
\[
\lim_{t\uparrow \zeta} X_t = \lim_{k\to\infty} X_{\tau_k} = xe^\zeta \prod_{j=1}^\infty J_j = 0 = X_\zeta,
\]
so that $X$ is in fact continuous at $\zeta$. This finishes the proof of point (i).
We now prove point (ii), i.e.~that $\pi$ is an invariant measure for the process $X$. For this, it is enough to show that there exists a function $\eta:\mathbb{R}_+\to \mathbb{R}_+$ with $\eta(t) \to 0$ as $t\to 0$ and such that for every Borel function $g:[0,\infty)\to [0,1]$,
\begin{equation}
\label{eq:to_show}
\forall t>0: \left|\int_0^\infty (P_t g(x) - g(x))\, \pi(dx)\right| \le t \eta(t).
\end{equation}
Indeed, if \eqref{eq:to_show} holds, then telescoping and using the semigroup property yields for every $t>0$, every $g$ as above and every $n\ge 1$,
\begin{align*}
\left|\int_0^\infty (P_t g(x) - g(x))\, \pi(dx)\right|
&= \sum_{k=0}^{n-1} \left|\int_0^\infty (P_{t/n}P_{kt/n} g(x) - P_{kt/n} g(x))\, \pi(dx) \right|
\end{align*}
Note that $P_s g$ takes values in $[0,1]$ for all $s\ge 0$ if $g$ does so. Applying \eqref{eq:to_show} then shows that
\[
\left|\int_0^\infty (P_t g(x) - g(x))\, \pi(dx)\right| \le n \frac t n \eta\left(\frac t n\right)
\]
and letting $n\to\infty$ shows that
\[
\int_0^\infty P_t g(x)\, \pi(dx) = \int_0^\infty g(x)\, \pi(dx),
\]
where $g$ was any Borel function with values in $[0,1]$. This implies that $\pi$ is an invariant measure.
We now show \eqref{eq:to_show}. Fix a Borel function $g:[0,\infty)\to [0,1]$. We let $X' = (X'_t)_{t\ge0}$ be the process that equals $(X_t)_{t \geq 0}$ up to and including the first jump and then grows exponentially afterwards. Formally,
\begin{equation}
\label{eq:Xprime}
X'_t =
\begin{cases}
X_0 e^t,\quad t <\tau_1\\
X_0 e^t J_1,\quad t \ge \tau_1
\end{cases},
\end{equation}
and note that $X'_t = X_t$ for all $t < \tau_2$. Hence, for every $x\in (0,\infty)$ and $t>0$,
\begin{equation}
\label{eq:comparison_X_Xprime}
\big|\mathbb E_x[g(X_t)] - \mathbb E_x[g(X'_t)]\big| \le \Pr_x(\tau_2 \le t),
\end{equation}
using that $g$ takes values in $[0,1]$.
We now claim:
\begin{align}
\label{eq:gen1} \forall t>0: \left|\int_0^\infty [\mathbb E_x[g(X'_t)] - g(x)]\,\pi(dx)\right| &\le 8t^2\\
\label{eq:gen2} \text{and}\qquad \Pr_\pi(\tau_2 \le t) /t &\to 0,\quad t\to0.
\end{align}
Then \eqref{eq:gen1} and \eqref{eq:gen2} together with \eqref{eq:comparison_X_Xprime} readily imply \eqref{eq:to_show}.
We prove \eqref{eq:gen1} and \eqref{eq:gen2} by direct calculation. Recall that $\pi(dx) = F'(x)\,dx$, with $F$ from Assumption~\ref{ass:stat}. Recalling the definition of the survival function $\surv_x$, we can rewrite \eqref{eq:Fprime_exp} as
\begin{equation}
\label{eq:xes}
\forall s\ge 0,\,x\in(0,\infty): F'(x) \surv_x(s) = e^{-s} F'(xe^s),
\end{equation}
and recalling that $\pi(dx) = F'(x)\,dx$, this gives for every bounded Borel function $h:(0,\infty)\to\mathbb{R}$,
\begin{equation}
\label{eq:xes2}
\forall s\ge 0: \int_0^\infty h(xe^s) S_x(s)\,\pi(dx) = e^{-2s} \int_0^\infty h(x)\,\pi(dx),
\end{equation}
where we applied first \eqref{eq:xes}, then a change of variables $xe^s\mapsto x$. This formula will be used in several places.
We now show \eqref{eq:gen1}. We decompose
\begin{align}
\mathbb E_x[g(X'_t)]
&= \mathbb E_x[g(xe^t)\mathbf{1}_{t < \tau_1}] + \mathbb E_x[g(xe^tJ_1)\mathbf{1}_{t\ge \tau_1}]\nonumber\\
&= g(xe^t)S_x(t) + \mathbb E[g(xe^tJ_1)] (1-S_x(t)) \label{eq:decomp},
\end{align}
where we used the fact that $\tau_1$ and $J_1$ are independent. Integrating w.r.t.~$\pi$, we have for the first term by \eqref{eq:xes2},
\begin{align}
\int_0^\infty g(xe^t)S_x(t)\,\pi(dx) = e^{-2t} \int_0^\infty g(x) \,\pi(dx). \label{eq:decomp1}
\end{align}
As for the second term on the RHS of \eqref{eq:decomp}, we have again by \eqref{eq:xes2},
\begin{align}
\int_0^\infty \mathbb E[g(xe^tJ_1)] (1-S_x(t))\,\pi(dx) &= \int_0^\infty (\mathbb E[g(xe^tJ_1)] - e^{-2t} \mathbb E[g(xJ_1)])\,\pi(dx). \label{eq:decomp2}
\end{align}
Now, by the definition of the random variable $J_1$, we have for every $x\in(0,\infty)$,
\begin{align}
\mathbb E[g(xe^tJ_1)] - e^{-2t} \mathbb E[g(xJ_1)]
&= \int_0^{e^t} \frac{2u}{(e^t)^2} g(xu)\,du - e^{-2t}\int_0^x 2u g(xu)\,du \nonumber \\
&= e^{-2t} \int_1^{e^t} 2u \,g(xu)\,du \nonumber\\
&= 2 e^{-2t} \int_0^t e^{2s} g(xe^s)\,ds. \label{eq:decomp3}
\end{align}
Integrating against $\pi$ and using Fubini's theorem, we get
\begin{align}
\int_0^\infty \int_0^t e^{2s} g(xe^s)\,ds\,\pi(dx)
&= \int_0^t e^{2s} \int_0^\infty g(xe^s)\,\pi(dx) \,ds \nonumber\\
&= \int_0^t \int_0^\infty g(x)\frac 1 {S_{xe^{-s}}(s)}\,\pi(dx). \label{eq:decomp4}
\end{align}
Since $S_{xe^{-s}}(s)$ is decreasing in $s$ and equals 1 at 0, we get from \eqref{eq:decomp4}, using that $g$ takes values in $[0,1]$,
\begin{align}
\left|\int_0^\infty \int_0^t e^{2s} g(xe^s)\,ds\,\pi(dx) - t \int_0^\infty g(x)\,\pi(dx)\right|
& \le t \int_0^\infty \left(\frac 1 {S_{xe^{-t}}(t)}-1\right)\,\pi(dx)\nonumber\\
& = t (e^{2t}-1) \label{eq:decomp5},
\end{align}
where the last equality follows from \eqref{eq:xes2} applied to $h(x) = 1/S_{xe^{-t}}(t)$ and the fact that $\pi$ is a probability measure. Collecting \eqref{eq:decomp2}, \eqref{eq:decomp3} and \eqref{eq:decomp5}, we get
\begin{align}
\left|\int_0^\infty \mathbb E[g(xe^tJ_1)] (1-S_x(t))\,\pi(dx) - 2t e^{-2t} \int_0^\infty g(x)\,\pi(dx)\right| \le 2t(1-e^{-2t}) \label{eq:decomp6}
\end{align}
Equations \eqref{eq:decomp}, \eqref{eq:decomp1} and \eqref{eq:decomp6} now yield, using that $g$ takes values in $[0,1]$,
\begin{align*}
\left|\int_0^\infty [\mathbb E_x[g(X'_t)] - g(x)]\,\pi(dx)\right| \le |(1+2t)e^{-2t} - 1|+ 2t(1-e^{-2t}) \le 8t^2
\end{align*}
using standard estimates on the exponential function, in particular $1 \ge (1+x)e^{-x}\ge 1-x^2$. This is \eqref{eq:gen1}.
It remains to prove \eqref{eq:gen2}. Recall that $r(x) = xR(x)$ for all $x > 0$ and $\frac{d}{dt} S_x(t) = r(xe^t)S_X(t).$ By definition of the process $X$, we have for every $x\in(0,\infty)$,
\begin{align*}
\Pr_x(\tau_2 \le t) = \int_0^t r(xe^s)S_x(s) \mathbb E[1-S_{xe^sJ_1}(s)]\,ds,
\end{align*}
where the expectation is meant with respect to $J_1$.
Integrating against $\pi$ and using \eqref{eq:xes2}, we get
\begin{align}
\Pr_\pi(\tau_2 \le t)
&= \int_0^t \left(\int_0^\infty r(xe^s)S_x(s)\mathbb E[1-S_{xe^sJ_1}(s)]\,\pi(dx)\right)ds\nonumber\\
&= \int_0^t e^{-2s} \left(\int_0^\infty r(x)\mathbb E[1-S_{xJ_1}(s)]\,\pi(dx)\right)ds\nonumber\\
&\le t \int_0^\infty r(x)\mathbb E[1-S_{xJ_1}(t)]\,\pi(dx), \label{eq:pr_tau2}
\end{align}
since the function $S_y(s)$ is decreasing in $s$ for every $y\in(0,\infty)$.
Now, $\mathbb E[1-S_{xJ_1}(t)] \to 0$ as $t\to\infty$, for every $x\in (0,\infty)$. Furthermore by \eqref{eq:integral_R},
\[
\int_0^\infty r(x)\,\pi(dx) = 2.
\]
By dominated convergence, we thus get
\begin{equation}
\label{eq:pr_tau2_2}
\int_0^\infty r(x)\mathbb E[1-S_{xJ_1}(t)]\,\pi(dx) \to 0,\quad t\to 0.
\end{equation}
Equation \eqref{eq:gen2} now follows from \eqref{eq:pr_tau2} and \eqref{eq:pr_tau2_2}.
We now prove point (iii). Fix a function $f\in C_c^1(0,\infty)$. We want to show that $t^{-1} (P_t f - f)$ converges to $\mathscr{L} f$ in $\Lpi$. Recalling the definition of $(X_t')_{t\ge0}$ from above, it is enough by \eqref{eq:gen2} to show the following convergence:
\begin{equation}
\label{eq:conv_Lpi}
\frac{\mathbb E_x[f(X_t')] - f(x)}{t} \to \mathscr{L} f(x),\quad\text{in $\Lpi$}.
\end{equation}
By \eqref{eq:decomp}, we can decompose
\begin{align}
\label{eq:decomp10}
\frac{\mathbb E_x[f(X_t')] - f(x)}{t} = \frac{f(xe^t) - f(x)}t S_x(t) + (\mathbb E[f(xe^tJ_1)] - f(x)) \frac{1-S_x(t)}{t}.
\end{align}
Since $f'\in C_c(0,\infty)$ and $S_x(t) \uparrow 1$ as $t\to0$, the first term on the right-hand side of \eqref{eq:decomp10} converges to $xf'(x)$ in $\mathrm L^\infty$ as $t\to0$ and therefore in $\Lpi$. As for the second term, first observe that $\mathbb E[f(xe^tJ_1)] - f(x)$ converges to $\mathbb E[f(xJ_1)]-f(x)$ in $\mathrm L^\infty$ as $t\to0$, by the continuity of $f$. We claim that $(1-S_x(t))/t$ converges to $xR(x)$ in $\Lpi$. First, note that by the Lebesgue differentiation theorem the convergence holds Lebesgue-a.e., hence $\pi$-a.e. by the equivalence of the two measures. Second, we have convergence of the $\Lpi$-norms:
\begin{align*}
\int \frac{1-S_x(t)}{t}\,\pi(dx)
&= \frac{1 - e^{-2t}} t \int \pi(dx) && \text{by \eqref{eq:xes2}}\\
&\to 2 && \text{as $t\to 0$}\\
&= \int x R(x)\,\pi(dx) && \text{by \eqref{eq:integral_R}}.
\end{align*}
By a theorem of Riesz known amongst probabilists as Scheff\'e's lemma (\cite{Kusolitsch}!), using the positivity of $1-S_x(t)$, we then get that $(1-S_x(t))/t$ converges indeed to $xR(x)$ in $\Lpi$. To summarize, we have proven the following convergence in $\Lpi$ as $t\to0$:
\[
\frac{\mathbb E_x[f(X_t')] - f(x)}{t} \to xf'(x) + xR(x) (\mathbb E[f(xJ_1)]-f(x)) = \mathscr{L} f(x),
\]
where the last equality follows from the definition of $J_1$. This is exactly \eqref{eq:conv_Lpi} which was to be shown.
\end{proof}
The following lemma appeared in the above proof of Proposition~\ref{prop:X_Markov}:
\begin{lemma}
\label{lem:existence_n}
For every $t\ge0$ and every $x>1/n$, we have
$$\mathbb E_x\left[\sum_{k=1}^\infty \mathbf{1}_{\tau_k \le t\wedge T^n}\right] < \infty.$$
\end{lemma}
\begin{proof}
Fix $t_*\ge0$. By \eqref{eq:def_X}, we have $X_t \le xe^t \le xe^{t_*}$ for all $t\le t_*$. In particular, for each $k\in\mathbb N$, conditioned on $\tau_1,\ldots,\tau_k$ and $J_1,\ldots,J_k$, on the event $\tau_k \le t_*\wedge T^n$, the difference $\tau^n_{k+1}-\tau^n_k$ is stochastically dominated from below by a random variable with survivor function
\[
E^n(t)
= \exp\left(- \sup_{x \in[1/n,xe^{t_*}]} \int_x^{e^tx} R(u)\,du\right) \quad \text{for all } 0 \leq t \leq t_*,
\]
and such a random variable is non-degenerate since $E^n(t) > 0$ for all $t$ by the local integrability of $R$. Hence, the sum $\sum_{k=1}^\infty \mathbf{1}_{\tau_k \le t\wedge T^n}$ is stochastically bounded from above by the number of points in the interval $[0,t_*]$ of a renewal process with interarrival times distributed according to the survivor function $E^n$. But this quantity has finite expectation, which finishes the proof.
\end{proof}
\subsection{Uniqueness of solutions to (\ref{eq:diff_mu})}
\label{sec:Lpi_and_uniqueness}
{\emph{Everywhere in this section, we work under Assumption~\ref{ass:stat}.}}
The goal of this section is to prove the following result:
\begin{proposition}
\label{prop:uniqueness}
Suppose that $(\mu_t)_{t \geq 0}$ is a family of Borel probability measures absolutely continuous with respect to $\pi$ satisfying that $t \mapsto \mu_t$ is Borel measurable and so that
\[
\sup_{t \geq 0} \left\|\tfrac{d\mu_t}{d\pi}\right\|_{\operatorname{L}^\infty(\pi)} < \infty.
\]
Suppose further that $(\mu_t)_{t\ge0}$ is a solution to \eqref{eq:diff_mu}, i.e., for all $f \in C_c^1(0,\infty)$ and $t\ge0$,
\[
(\mu_t, f) - (\mu_0,f) = \int_0^t (\mu_s, \mathscr{L} f)\,ds.
\]
Then $\mu_t = \mu_0 P_t$ for all $t\ge0$.
%
\end{proposition}
As mentioned above, the proof of Proposition~\ref{prop:uniqueness} will heavily rely on semigroup theory. The results from the previous section will be an important ingredient as well. We start with the following lemma:
\begin{lemma}
\label{lem:semigroup}
$(P_t)_{t\ge0}$ is a strongly continuous contraction semigroup on $\Lpi.$
\end{lemma}
\begin{proof}
By virtue of having shown that $\pi$ is invariant for $P_t$ (part (ii) of Proposition~\ref{prop:X_Markov}) and by virtue of the positivity of $P_t,$ we have that for $f \in \Lpi$,
\[
\Lpin[ P_tf ]
\leq
\Lpin[ P_t |f| ]
=(\tfrac{d\pi}{dx}, P_t|f|)
=(\tfrac{d\pi}{dx},|f|)
= \Lpin[ f ].
\]
It follows that $P_t$ is contraction semigroup on $\Lpi$. Now define the space
\[
L_0 =\left\{
f \in \Lpi~:~\lim_{t \to 0} \Lpin[ P_t f - f] = 0
\right\}
\]
By part (iii) of Proposition~\ref{prop:X_Markov}, it contains $C_c^1(0,\infty)$ and is thus dense in $\Lpi$. Moreover, this space is necessarily closed (see \cite[(1.3)]{DynkinMPBookI}), and hence in fact $L_0 = \Lpi.$ This shows that $(P_t)_{t\ge0}$ is also strongly continuous and finishes the proof of the lemma.
\end{proof}
It follows from Lemma~\ref{lem:semigroup} and classical semigroup theory \cite[Theorems~1.1 and~1.3]{DynkinMPBookI} that there is a subspace $D_{\RG} \subset \Lpi$ and an operator $\mathscr{L} : D_{\RG} \to \Lpi$, called the infinitesimal generator of the semigroup $(P_t)_{t\ge0}$, so that:
\begin{enumerate}[(i)]
\item $D_{\RG}$ is the space of all $f \in \Lpi$ such that $\lim_{t \to 0}t^{-1}(P_t f - f)$ exists in $\Lpi$, moreover, this limit equals $\mathscr{L} f$.
\item
For all $\lambda > 0,$ the map $(\lambda - \mathscr{L})^{-1}$ given by $\int_0^\infty e^{-\lambda t}P_t\,dt$ is a bounded linear operator from $\Lpi$ to $D_{\RG}.$
\item
We have that $P_t$ maps $D_{\RG} \to D_{\RG}$ and for all $f \in D_{\RG},$ we have that
\begin{equation}
\label{eq:Pt_RG}
P_t f - f = \int_0^t P_s \mathscr{L} f\,ds = \int_0^t \mathscr{L} P_s f\,ds.
\end{equation}
\end{enumerate}
By part (iii) of Proposition~\ref{prop:X_Markov}, we see that there is no conflict of notation: for $f\in C_c^1(0,\infty)$, the operator $\mathscr{L}$ defined above indeed coincides with the operator $\mathscr{L}$ defined in \eqref{eq:RG}. In particular, $C_c^1(0,\infty) \subset D_{\RG}$. The key to the proof of Proposition~\ref{prop:uniqueness} is now the following proposition:
\begin{proposition}
%
The space $C_c^1(0,\infty)$ is a core for $\mathscr{L};$ that is, for any $f \in D_{\RG},$ there is a sequence $\left\{ f_n \right\}_{n=1}^\infty \subset C_c^1(0,\infty)$ so that $f_n \to f$ and $\mathscr{L} f_n \to \mathscr{L} f$ in $\Lpi$, as $n\to\infty$.
\label{prop:core}
\end{proposition}
To show this, we will employ an idea used widely in the growth-fragmentation process literature (see for example \cite{Mischler2016} or \cite[Chapter 9]{Banasiak2006}). We decompose the operator $\mathscr{L}$ into two pieces $\mathscr{B}$ and $\mathscr{K}$ where $\mathscr{B}$ is roughly a differential part and $\mathscr{K}$ is in some sense a perturbation. Here we define, for all $f \in C_c^1(0,\infty),$
\begin{align*}
\mathscr{B} f &= xf'(x) -xR(x)f(x) \\
\mathscr{K} f &= xR(x) \int_0^x \frac{2u}{x^2} f(u)\,du.
\end{align*}
With this notation, we have that $\mathscr{L} = \mathscr{B} + \mathscr{K}.$ We can view $\mathscr{K}$ as a perturbation because it is in fact a bounded linear operator on $\Lpi:$
\begin{lemma}
$\mathscr{K}$ extends uniquely to an operator on $\Lpi.$ Moreover,
for all $f \in \Lpi,$ we have that
\[
\Lpin[\mathscr{K} f] \leq 2\Lpin[f].
\]
\label{lem:rgk}
\end{lemma}
\begin{proof}
It suffices to show the estimate in the statement of the Lemma for all $f\in C_c^1(0,\infty).$ For these $f$, we have that by the Fubini-Tonelli theorem,
\begin{align*}
\Lpin[\mathscr{K} f]
&\leq \int_0^\infty xR(x) \int_0^x \frac{2u}{x^2} |f(u)|\,du\pi(dx) \\
&\leq \int_0^\infty 2u |f(u)| \int_u^\infty \frac{R(x)}{x}\pi(dx) \\
&= \int_0^\infty 2|f(u)|\pi(du),\qquad\text{by \eqref{eq:Fprime}}.
\end{align*}
This proves the lemma.
\end{proof}
In light of this, there is a sense in which $\mathscr{B}$ and $\mathscr{L}$ can be compared analytically. Further, for $\mathscr{B},$ the core statement we wish to prove is relatively straightforward.
\begin{proposition}
Let $(Q_t)_{t\ge0}$ be the semigroup
\[
Q_tf(x) = \mathbb{E}_x f(X_t) \one[ \tau_1 > t] = f(xe^t)S_x(t),
\]
for all bounded Borel measurable $f.$ Then
\begin{enumerate}[(i)]
\item $Q_t$ is a strongly continuous contraction semigroup on $\Lpi.$
\item For all $f \in C_c^1(0,\infty),$ we have that
\[
\lim_{t \to 0} t^{-1}(Q_t f - f) = \mathscr{B} f,\quad\text{in $\Lpi$.}
\]
\end{enumerate}
Point (ii) says that $C_c^1(0,\infty)$ is a subset of the domain of the infinitesimal generator of the semigroup $(Q_t)_{t\ge0}$ and that $\mathscr{B}$ coincides with this generator on $C_c^1(0,\infty)$. We therefore denote without risk of ambiguity by $\mathscr{B}$ as well this infinitesimal generator and by $D_{\mathscr{B}}\subset \Lpi$ its domain.
\begin{enumerate}[(i)]\setcounter{enumi}{2}
\item For any $\lambda > 0,$ the resolvent $(\lambda - \mathscr{B})^{-1} = \int_0^\infty e^{-\lambda t} Q_t\,dt$ maps $\Lpi \to D_{\mathscr{B}}$ and has operator norm
\[
\|(\lambda - \mathscr{B})^{-1}\| < \frac{1}{2+\lambda}.
\]
\item $C_c^1(0,\infty)$ is a core for $\mathscr{B}.$
\end{enumerate}
\label{prop:rgb}
\end{proposition}
\begin{proof}
We begin by observing that for $f \geq 0,$ we have that $Q_t f \leq P_t f,$ at all points. Hence we have that for all $f\in \Lpi,$
\[
\Lpin[ Q_t f ]
\leq \Lpin[ Q_t |f| ]
\leq \Lpin[ P_t |f| ]
\leq \Lpin[ f ].
\]
Hence $Q_t$ is indeed a contraction semigroup on $\Lpi.$ Strong continuity follows from the fact that for $f\in C_c(0,\infty)$, we have $Q_tf(x) = f(xe^t)S_x(t) \to f(x)$ in $\mathrm L^\infty$, hence in $\Lpi$.
And hence, we also have it in $\Lpi,$ by the same argument as in Lemma~\ref{lem:semigroup}.
This proves point (i).
The proof of point (ii) is contained in the proof of part (iii) of Proposition~\ref{prop:X_Markov}.
We now get to the main part of the proof, which is the proof of the resolvent estimate (iii) and that $C_c^1(0,\infty)$ is a core for $\mathscr{B}$ (point (iv)). We will prove them both together. For the statement about the core, we use the following condition from~\cite{EthierKurtz}.
\begin{lemma}[Chapter 1, Proposition 3.1 of \cite{EthierKurtz}]
\label{lem:eth}
Let $A$ generate a strongly continuous contraction semigroup on a Banach space $L.$ Then a subspace $V \subset D_A$ is a core for $A$ if and only if $V$ is dense in $L$ and the range of $\lambda - A$ restricted to $V$ is dense in $L$ for some $\lambda > 0.$
\end{lemma}
The density of $C_c^1(0,\infty) \subset \Lpi$ is immediate. Hence we must show that $\lambda - \mathscr{B}$ has dense range for some $\lambda > 0$ to complete the proof.
Fix $\lambda > 0.$ Suppose we would like to solve the equation
\begin{align}
\label{eq:diffeq}
g = (\lambda - \mathscr{B})f = \lambda f - xf' + xR(x)f(x),
\end{align}
for some function $g$.
This is nothing but a first order linear differential equation, and hence we have the formal solution
\[
f(x) = \frac{1}{\mu(x)}
\int_x^\infty \frac{g(y)}{y}\mu(y)\,dy
\]
where $\mu(y) = y^{-\lambda}\exp(-\int_1^y R(x)\,dx).$ Note that we can in fact express $\mu$ in terms of $\pi$ since by \eqref{eq:Fprime_exp} we have that
\[
R(z) = - \frac{d}{dz} \log\left( z^{-1} d\pi/dz \right).
\]
Hence we may formally express the solution to \eqref{eq:diffeq} by
\begin{align}
\label{eq:resol}
f(x) = \frac{x^{1+\lambda}}{\tfrac{d\pi}{dx}}\int_x^\infty \frac{g(y)}{y^{2+\lambda}}\,\pi(dy)
\end{align}
(note that $d\pi/dx$ is positive and continuous on $(0,\infty)$, see the discussion around \eqref{eq:Fprime}).
When $g \in C_c(0,\infty),$ then this is a well-defined absolutely continuous solution to \eqref{eq:diffeq}. Hence, $(\lambda-\mathscr{B})^{-1} g = f\inD_{\mathscr{B}}$ and %
\begin{align}
\Lpin[(\lambda-\mathscr{B})^{-1}g]
& = \Lpin[f] \nonumber\\
&\leq \int_0^\infty x^{1+\lambda}\int_x^{\infty} \frac{|g(y)|}{y^{2+\lambda}}\,\pi(dy)\,dx \nonumber\\
&=\int_0^\infty \frac{|g(y)|}{y^{2+\lambda}}\int_0^{y} x^{1+\lambda}\,dx\,\pi(dy) \nonumber\\
&=\frac{1}{2+\lambda}\int_0^\infty{|g(y)|}\,\pi(dy) \nonumber\\
&=\frac{1}{2+\lambda}\Lpin[g]\label{eq:resolventbound}
\end{align}
By density, we conclude the desired estimate on $\|(\lambda - \mathscr{B})^{-1}\|$ and thus prove point (iii).
Fixing $g \in C_c(0,\infty),$ we will now show that $f$ can be approximated by a sequence $f_n \in C_c^1(0,\infty)$ so that $(\lambda-\mathscr{B})f_n \to g$ in $\Lpin.$ We start by truncating the support of $f.$ By \eqref{eq:resol}, $f(x) = 0$ for large enough $x$, so that it is enough to truncate near the origin. For $n \in \mathbb{N},$ let $\rho_n$ be a $C_c^1(0,\infty)$ bump function that is $1$ for $x > 3n^{-1}$ and $0$ for $x < 2n^{-1}.$ We can choose these functions so their derivatives are bounded uniformly by $O(n)$ on $[2n^{-1},3n^{-1}].$ In particular, we have that $|\rho_n'(x)| \leq Cx^{-1}$ for some constant $C>0.$ We have,
\begin{align}
\label{eq:rgbtrunc}
\Lpin[ (\lambda - \mathscr{B})(f-f\rho_n)] &
\leq
\int\limits_0^\infty |\lambda f(x) - xf'(x)|(1-\rho_n(x))\,\pi(dx)\\
&+
\int\limits_0^\infty \left|xf(x)\rho_n'(x) \right|\,\pi(dx) \nonumber \\
&+\int\limits_0^\infty xR(x)|f(x)|(1-\rho_n(x))\,\pi(dx). \nonumber
\end{align}
We wish apply dominated convergence to all of these integrals to get their convergence to 0. We just need to exhibit suitable dominators. Observe that $\lambda f(x) - xf'(x) = g(x) - xR(x)f(x),$ and $\Lpi[g] < \infty$. Hence, for the first and last integrals, it suffices to show that $\int_0^\infty xR(x)|f(x)|\pi(dx) < \infty.$ Moreover, since $f(x)$ vanishes for large $x$, it is enough to show that this integral converges near zero.
From \eqref{eq:resol}, we have that $|f(x)|\tfrac{d\pi}{dx} < x^{1+\lambda}C_g$ for all $x \geq 0$ and some constant $C_g.$ Thus it suffices to show that $x^2R(x)$ is integrable at $0.$ Recall from Assumption~\ref{ass:stat} that $\int_0^\infty xR(x)\tfrac{d\pi}{dx}\,dx < \infty.$
Furthermore, by \eqref{eq:Fprime_exp}, we have for all $x \le 1,$
\[
\frac{d\pi}{dx} \geq cx,
\]
where $c \coloneqq (d\pi/dx)(1)>0$, by the discussion around \eqref{eq:Fprime_exp}. Hence,
\[
\int_0^1 cx^2 R(x) \leq \int_0^\infty xR(x)\frac{d\pi}{dx}\,dx < \infty,
\]
which was to be proven.
For the second integral of \eqref{eq:rgbtrunc}, since $|\rho_n'(x)| \leq Cx^{-1},$ we can dominate the integrand by $C|f|$, since $\Lpin[f]<\infty$.
We now argue that it is possible to smooth $f \rho_n$ slightly so that it is in $C_c^1(0,\infty).$ This is relatively straightforward by density. Just observe that $(f\rho_n)'$ is a compactly supported $\Lpi$ function. Hence we can choose $f_n \in C_c^1(0,\infty)$ so that
\[
\max\left(\int_0^\infty x|(f\rho_n)'(x) - f_n'(x)|\,\pi(dx) , \|f_n - f\rho_n\|_\infty \right) < \frac{1}{n}.
\]
It now follows immediately that
\(
(\lambda - \mathscr{B})(f_n - f\rho_n) \to 0
\)
in $\Lpin.$ As we also have that
\(
(\lambda - \mathscr{B})(f\rho_n - f) \to 0
\)
in $\Lpin,$ the proof is complete.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prop:core}]
From Lemma~\ref{lem:eth} and Proposition~\ref{prop:rgb} part (iv), there is a $\lambda >0$ so that $\lambda - \mathscr{B}$ has dense range when restricted to $C_c^1(0,\infty).$ Then by Proposition~\ref{prop:rgb} part (iii), the resolvent operator $(\lambda - \mathscr{B})^{-1} : \Lpi \to D_{\mathscr{B}}$ has operator norm bounded by $(2+\lambda)^{-1}.$ Hence, by Lemma~\ref{lem:rgk}, we have that
\[
I - \mathscr{K}(\lambda - \mathscr{B})^{-1}
\]
is invertible on $\Lpi$ with bounded inverse.
We now show that the range of $\lambda - \mathscr{L}$ on $C_c^1(0,\infty)$ is dense in $\Lpi.$ Fix $f \in \Lpi,$ and let $h \in \Lpi$ have that $f = (I - \mathscr{K}(\lambda - \mathscr{B})^{-1})h.$ By Proposition~\ref{prop:rgb}, the operator $\lambda - \mathscr{B}$ is dense on $C_c^1(0,\infty)$ and so for every $\epsilon > 0$ we can find a $g \in C_c^1(0,\infty)$ so that $\Lpin[ (\lambda - \mathscr{B})g - h] < \epsilon.$ Then we have that
\[
\Lpin[ (\lambda - \mathscr{B} - \mathscr{K})g - f] =
\Lpin[ (I - \mathscr{K}(\lambda - \mathscr{B})^{-1})((\lambda - \mathscr{B})g - h)] < 2\epsilon,
\]
which completes the proof.
\end{proof}
We can now finally pass to the proof of the uniqueness result for \eqref{eq:diff_mu}:
\begin{proof}[Proof of Proposition~\ref{prop:uniqueness}]
By Proposition~\ref{prop:core}, for any $f \in D_{\RG},$ there is a sequence $f_n \in C_c^1(0,\infty)$ so that
\(
f_n \to f
\)
and $\mathscr{L} f_n \to \mathscr{L} f$ in $\Lpi.$ Since
$\|\tfrac{d\mu_t}{d\pi}\|_{\operatorname{L}^\infty(\pi)}
$
is uniformly bounded, we have that $(\mu_s, \mathscr{L} f_n) \to (\mu_s, \mathscr{L} f)$ uniformly in $s \in [0,\infty).$ By taking limits, we obtain:
\begin{equation}
\label{eq:diff_mu_full}
\forall f \in D_{\RG}: (\mu_t, f) - (\mu_0,f) = \int_0^t (\mu_s, \mathscr{L} f)\,ds.
\end{equation}
It remains to conclude from \eqref{eq:diff_mu_full} that $\mu_t = \mu_0 P_t$ for all $t\ge0$.
This follows from a standard argument (see e.g.~\cite[Theorem 1.3]{DynkinMPBookI}), which we recall for convenience.
Set $\nu_t = \mu_0 P_t$ for all $t\ge0$. Then we have for all $f\in D_{\RG}$ and $t\ge0$,
\begin{align*}
(\nu_t, f) - (\nu_0,f)
&= (\nu_0, P_t f - f)\\
&= (\nu_0, \int_0^t P_s \mathscr{L} f\,ds) && \text{ by \eqref{eq:Pt_RG}}\\
&= \int_0^t (\nu_0,P_s\mathscr{L} f)\,ds\\
&= \int_0^t (\nu_s,\mathscr{L} f)\,ds.
\end{align*}
Hence, $(\nu_t)_{t\ge0}$ satisfies \eqref{eq:diff_mu_full}. By linearity, the difference $(\mu_t - \nu_t)_{t\ge0}$ then satisfies \eqref{eq:diff_mu_full} as well.
It now suffices to show that for
all $f \in C_c^1(0,\infty)$ and all $\lambda >0,$
\[
\int_0^\infty
e^{-\lambda t}
(\mu_t-\nu_t,f)
\,dt
= 0,
\]
for then it follows that $(\mu_t-\nu_t,f)=0$ for all $t \geq 0$ and thus $\mu_t = \nu_t$ for all $t\ge0$. Let $g = (\lambda -\mathscr{L})^{-1} f \in D_{\RG} \subset \Lpi.$ Then
\begin{equation} \label{eq:una}
\int_0^\infty
e^{-\lambda t}
(\mu_t-\nu_t,f)
\,dt
=
\int_0^\infty
e^{-\lambda t}
(\mu_t-\nu_t,(\lambda - \mathscr{L})g)
\,dt.
\end{equation}
Since $(\mu_t-\nu_t)_{t\ge0}$ solves \eqref{eq:diff_mu_full}, a primitive of $(\mu_t-\nu_t,\mathscr{L} g)$ in $t$ is $(\mu_t-\nu_t,g).$ Note that, by assumption for all $t \geq 0,$
\[
|(\mu_t, g)| = |(\pi,g \tfrac{d\mu_t}{d\pi})|
\leq
\|g\|_{\operatorname{L}^1(\pi)}
\cdot
\sup_{t \geq 0}\|\tfrac{d\mu_t}{d\pi}\|_{\operatorname{L}^\infty(\pi)}
< \infty.
\]
Note also that for all $t \geq 0,$ by Lemma \ref{lem:semigroup}
\[
|(\nu_t, g)|
= |(\mu_0,P_t g)|
\leq \int_0^\infty P_t(|g|) \tfrac{d\mu_0}{d\pi} d\pi
\leq \|g\|_{\operatorname{L}^1(\pi)}
\cdot
\sup_{t \geq 0}\|\tfrac{d\mu_t}{d\pi}\|_{\operatorname{L}^\infty(\pi)}
< \infty.
\]
Combining the previous two display equations, we conclude that $(\mu_t-\nu_t,g)$ is bounded uniformly in $t.$ Thus for any $\lambda > 0,$ we have using integration by parts that
\[
\int_0^\infty
e^{-\lambda t}
(\mu_t-\nu_t,\mathscr{L} g)
\,dt
=
\int_0^\infty
\lambda e^{-\lambda t}
(\mu_t-\nu_t, g)
\,dt.
\]
We conclude by splitting the right hand side of \eqref{eq:una} and applying the previous display that
\begin{align*}
\int_0^\infty
e^{-\lambda t}
(\mu_t-\nu_t,f)
\,dt
&=
\int_0^\infty
e^{-\lambda t}
(\mu_t-\nu_t,\lambda g)
\,dt
+
\int_0^\infty
-\lambda
e^{-\lambda t}
(\mu_t-\nu_t,g)
\,dt \\
&= 0,
\end{align*}
which was to be proven.
\end{proof}
\subsection{Ergodicity}
\label{sec:ergodicity}
Recall the definition of the Markov process $X = (X_t)_{t\ge0}$ from Section~\ref{sec:construction_markov_process} and its associated semigroup $(P_t)_{t\ge0}$.
The main goal of this section is to prove the following proposition:
\begin{proposition}
\label{prop:ergodicity_gf}
Under Assumption~\ref{ass:stat}, the Markov process $X$ is ergodic, i.e.~for every initial distribution $\nu$, we have
\begin{equation}
\label{eq:ergodicity}
d_{\operatorname{TV}}\left(\nu P_t, \pi\right) \to 0,\quad \text{as $t\to\infty$.}
\end{equation}
Moreover, if $\mathcal M$ is a tight family of distributions, the convergence is uniform on $\mathcal M$, i.e.
\[
\sup_{\nu\in\mathcal M} d_{\operatorname{TV}}\left(\nu P_t, \pi\right) \to 0,\quad \text{as $t\to\infty$.}
\]
\end{proposition}
We could prove this using tools from the theory of piecewise deterministic processes from Costa and Dufour \cite{Costa2008}, which ultimately relies on Meyn and Tweedie \cite{Meyn1993}. However, we prefer a more elementary approach using regeneration which is possible without too many technicalities due to the fact that points are hit almost surely (and are thus \emph{non-polar}).
Define for $x\in(0,\infty)$,
\[
H_{x} = \inf\{t>0: X_t = x\}.
\]
We will need the following three preliminary lemmas:
\begin{lemma}
\label{lem:couple_small_time}
For every $t\ge0$ and $x\in(0,\infty)$, denote by $\nu^x_t$ the law of $(X_{s+t})_{s\ge0}$ under $\Pr_{xe^{-t}}$. Then for every $x\in(0,\infty)$,
\[
d_{\operatorname{TV}}(\nu^x_t,\nu^x_0) \to 0,\quad\text{as $t\to 0$}.
\]
\end{lemma}
\begin{proof}
Since $X$ is a Markov process, it suffices to show that the law of $X_t$ under $\Pr_{xe^{-t}}$ converges to $\delta_x$ in total variation. But since $X_t = x$ on the event $\tau_1 > t$ under $\Pr_{xe^{-t}}$, we have,
\[
\Pr_{xe^{-t}}(X_t = x) \ge \Pr_{xe^{-t}}(\tau_1 > t) = S_{xe^{-t}}(t) = \exp\left(-\int_{xe^{-t}}^x R(z)\,dz\right),
\]
and this goes to 1 as $t\to0$.
This proves the claim.
\end{proof}
\begin{lemma}
\label{lem:hit_points}
For any $x\in(0,\infty)$ and $I\subset (0,x]$ compact, there exist $T>0$, such that
\[
\inf_{y\in I} \Pr_y(H_x < T) > 0.
\]
\end{lemma}
\begin{proof}
Fix $x\in(0,\infty)$. Similarly to the proof of Lemma~\ref{lem:couple_small_time}, we have for every $y\in (0,x]$,
\[
\Pr_y(H_x = \log(x/y)) \ge \Pr_y(\tau_1 > \log(x/y)) = \exp\left(-\int_y^x R(z)\,dz\right).
\]
Hence, if $I\subset (0,x]$ is compact and $y_0 = \min I > 0$, then for every $T > \log(x/y_0)$,
\[
\inf_{y\in I} \Pr_y(H_x < T) \ge \exp\left(-\int_{y_0}^x R(z)\,dz\right) > 0.
\]
This shows the result.
\end{proof}
\begin{lemma}
\label{lem:H_abs_cont}
Suppose Assumption~\ref{ass:stat} is in place. Then for any $x\in(0,\infty)$, the law of $H_x$ under $\Pr_x$ has a non-zero absolutely continuous component.
\end{lemma}
\begin{proof}
Fix $x\in(0,\infty)$. Let $f:[0,\infty)\to[0,1]$ be Borel. Then
\begin{align*}
\mathbb E_x[f(H_x)]
&\ge \mathbb E_x[f(H_x)\one[\tau_1 \le H_x < \tau_2]].
\end{align*}
By the definition of the process $X$, we have equality of the events
\[
\{\tau_1 \le H_x < \tau_2\} = \{\tau_1 \le -\log(J_1) < \tau_2\},
\]
and on both events we have $H_x = -\log(J_1)$. Applying the strong Markov property at time $\tau_1$, we then get
\[
\mathbb E_x[f(H_x)\one[\tau_1 \le H_x < \tau_2]]= \mathbb E_x[\mathbb E_{xe^{\tau_1}J_1}[\tau_1 > -\log(J_1)] f(-\log(J_1))\one[-\log(J_1) \ge \tau_1]].
\]
Now note that $\mathbb E_{xe^{\tau_1}J_1}[\tau_1 > -\log(J_1)] = S_{xe^{\tau_1}J_1}(-\log(J_1)) > 0$, $\Pr_x$-almost surely. Collecting all the previous equations then gives, for some r.v.~$W>0$,
\[
\mathbb E_x[f(H_x)] \ge \mathbb E_x[f(-\log(J_1))W \one[-\log(J_1) \ge \tau_1]].
\]
Now note by definition of $J_1$, the law of the r.v. $-\log(J_1)$ is absolutely continuous w.r.t.~Lebesgue measure and has full support in $(0,\infty)$. Also, $J_1$ and $\tau_1$ are independent random variables. Furthermore, by \eqref{eq:R_positivity}, we have $\Pr_x(\tau_1 < \infty) = 1-S_x(\infty) > 0$ (this is the only place where Assumption~\ref{ass:stat} is used). The statement easily follows.
\end{proof}
We can now proceed to the proof of the main result from this section.
\begin{proof}[Proof of Proposition~\ref{prop:ergodicity_gf}]
We first prove the pointwise convergence result \eqref{eq:ergodicity}. This will be shown using a classical result on coupling of regenerative processes. Define a sequence of random times by
\[
\tau_0 = 0,\quad \forall n\ge0:\tau_{n+1} = \inf\{t>\tau_n:X_t = 1\}.
\]
Suppose for the moment we have established the following: for every initial distribution $\nu$,
\begin{equation}
\label{eq:regeneration}
\Pr_\nu(\forall n:\tau_n < \infty) = 1.
\end{equation}
Then by the strong Markov property, the sequence $(\tau_n)_{n\ge1}$ is a sequence of regeneration times for the process $X$, in particular, $(\tau_n)_{n\ge1}$ is a renewal process with a delay distribution depending on $\nu$. Furthermore, by Lemma~\ref{lem:H_abs_cont}, the interarrival time $\tau_{n+1}-\tau_n$, $n\ge1$, which has the law of $H_1$ under $\Pr_1$, has a non-zero absolutely continuous component. This implies \cite[Chapter 10, Theorem 3.3]{ThorissonBook} that
\[
d_{\operatorname{TV}}\left(\mu P_t,\nu P_t\right)\to 0,
\]
for all initial distributions $\mu$ and $\nu$. In particular, this is the case for $\mu = \pi$, which satisfies $\pi P_t = \pi$ (Proposition \ref{prop:X_Markov} (ii)) for all $t \geq 0$ and which proves the pointwise convergence result \eqref{eq:ergodicity}.
It remains to show \eqref{eq:regeneration}. This will rely on Lemma~\ref{lem:hit_points} together with Birkhoff's ergodic theorem. We seperate the proof into several steps.
1) By the ergodic theorem, we have for every Borel $A\subset(0,\infty)$,
\begin{equation}
\label{eq:birkhoff}
\frac 1 t \int_0^t \mathbf{1}_{X_s\in A}\,ds \to \mathbb E_\pi[\mathbf{1}_{X_0\in A}\,|\,\mathscr I],\quad \text{$\Pr_\pi$-almost surely},
\end{equation}
where $\mathscr I$ is the \emph{invariant $\sigma$-field}, i.e.~the $\sigma$-field of events invariant under time shifts. Fixing a version of the conditional expectation on the right-hand side, this gives for every Borel $A\subset(0,\infty)$:
\begin{equation}
\label{eq:birkhoff2}
\text{for $\pi$-a.e. $x$: } \frac 1 t \int_0^t \mathbf{1}_{X_s\in A}\,ds \to \mathbb E_\pi[\mathbf{1}_{X_0\in A}\,|\,\mathscr I],\quad \text{$\Pr_x$-almost surely},
\end{equation}
2) We now show that $\mathscr I$ is trivial\footnote{This seems to be a general property of positive Harris recurrent Markov processes (maybe under some extra irreducibility assumption), but we could not find a proper reference.} under $\Pr_\pi$ (although one could work around it). Taking expectations in \eqref{eq:birkhoff2} and using dominated convergence gives for every Borel $A\subset(0,\infty)$,
\[
\text{for $\pi$-a.e. $x$: } \frac 1 t \int_0^t \Pr_x(X_s\in A)\,ds \to \pi(A),
\]
In fact, the above statement is true for every $x$, which is an easy consequence of Lemma~\ref{lem:couple_small_time} and the fact that $\pi$ has full support in $(0,\infty)$ (see the discussion after \eqref{eq:Fprime_exp}). This implies
\[
\text{for all $x\in(0,\infty)$: } d_{\operatorname{TV}}\left(\frac 1 t \int_0^t \delta_xP_s\,ds, \pi \right) \to 0,
\]
which, by \cite[Chapter 6, Theorem 5.2]{ThorissonBook} proves triviality of $\mathscr I$ under $\Pr_\pi$.
3) By the triviality of $\mathscr I$ under $\Pr_\pi$, Equations \eqref{eq:birkhoff} and \eqref{eq:birkhoff2} hold with the right-hand side replaced by $\pi(A)$. Using again Lemma~\ref{lem:couple_small_time}, this implies the following for every Borel $A\subset(0,\infty)$:
\begin{equation}
\label{eq:birkhoff3}
\text{for all $x\in(0,\infty)$: } \frac 1 t \int_0^t \mathbf{1}_{X_s\in A}\,ds \to \pi(A),\quad \text{$\Pr_x$-almost surely.}
\end{equation}
4) Since $\pi$ has full support, we have $\pi([1/2,1]) > 0$. It follows from \eqref{eq:birkhoff3} that for every starting distribution $\nu$, the interval $[1/2,1]$ is visited a positive proportion of times $\Pr_\nu$-almost surely. Together with Lemma~\ref{lem:hit_points} and the strong Markov property, this easily allows to show that the point $1$ is visited an infinite number of times $\Pr_\nu$-almost surely. This is exactly \eqref{eq:regeneration}, which was to be shown.
\medskip
We have shown \eqref{eq:ergodicity}. We now show that the convergence is uniform on tight families of distributions on $(0,\infty)$. Define for all $t \geq 0$ and all $x > 0,$
\[
m_t(x) = d_{\operatorname{TV}}\left(
\delta_x P_t, \pi
\right).
\]
As $P_t$ is a contraction in the total variation distance, this function is monotone decreasing in $t$. Further, by \eqref{eq:ergodicity}, it converges to $0$ pointwise as $t \to \infty.$ We claim that the convergence is uniform on compact sets. For this, it is enough to show the same for the function
\[
\widetilde m_t(x) = \begin{cases}
m_{t-\log x}(x), & x\le e^t\\
1, & \text{otherwise}.
\end{cases}
\]
This function converges again pointwise to 0 as $t\to\infty$ and is monotone decreasing in $t$. Further, using Lemma~\ref{lem:couple_small_time} it is easy to show that $\widetilde m_t(x)$ is continuous in $x$ for every $x < e^t$. By Dini's theorem, $\widetilde m_t$ therefore converges to $0$ as $t\to\infty$, uniformly on compact sets, and the same easily follows for $m_t$.
Now let $\mathcal M$ be a tight family of distributions on $(0,\infty)$. Fix $\epsilon > 0.$ We wish to show that $\sup_{\nu\in\mathcal M} d_{\operatorname{TV}}\left( \nu P_t, \pi \right)<\epsilon$ for every $t$ large enough. By tightness of $\mathcal{M}$ we can find a compact set $K \subset \left( 0,\infty \right)$ such that for any $\nu \in \mathcal{M},$ $\nu(K^c) < \epsilon/2.$ We then bound
\begin{align*}
d_{\operatorname{TV}}\left( \nu P_t, \pi \right)
&\leq
\int\limits_0^\infty m_t(x)\,\nu(dx) \\
&\leq
\int\limits_K m_t(x)\,\nu(dx) + \epsilon/2 \\
&\leq \sup_{x \in K} m_t(x) + \epsilon/2.
\end{align*}
The first term on the right-hand side is smaller than $\epsilon/2$ for large $t$ by the uniform convergence on compact sets of $m_t$. Since $\epsilon$ was arbitrary, this shows that $\sup_{\nu\in\mathcal M} d_{\operatorname{TV}}\left( \nu P_t, \pi \right)\to 0$ goes to $0$ as $t\to\infty$, which was to be proven.
\end{proof}
\subsection{Proof of Proposition~\ref{prop:convergence_RFPSpace}}
\label{sec:proof_prop}
We now wrap up the results from the previous sections for the proof of Proposition~\ref{prop:convergence_RFPSpace}:
\begin{proof}[Proof of Proposition~\ref{prop:convergence_RFPSpace}]
Let $\F\in\mathfrak{R}_{\operatorname{ac}}$ and define $\mu_t = d\F[t]$. By Proposition~\ref{prop:markov}, the family of measures $(\mu_t)_{t\ge0}$ satisfies \eqref{eq:diff_mu} for every $f\in C_c^1(0,\infty)$. Furthermore, by definition of the space $\mathfrak{R}_{\operatorname{ac}}$, the family $(\mu_t)_{t\ge0}$ satisfies the first assumption of Proposition~\ref{prop:uniqueness}. Proposition~\ref{prop:uniqueness} then shows that $\mu_t = \mu_0 P_t$ for all $t\ge0$. The result then follows from Proposition~\ref{prop:ergodicity_gf}.
\end{proof}
\section{Continuity of $\mathscr{C}$}
\label{sec:proof_continuity}
Recall the notion of locally uniform convergence on the space $\mathcal{X} = \mathcal B([0,\infty),\mathcal{D})$ of Borel measurable maps from $[0,\infty)$ to $\mathcal{D}$: $\F^{(n)} \overset{\DTSpace}{\to} \F$ as $n\to\infty$ if and only if for all compact $K \subseteq [0,\infty)$ and all $T> 0,$
\[
\lim_{n \to \infty} \sup_{0 \leq t \leq T} \int\limits_K | \F[t]^{(n)}(x) - \F[t](x) |\,dx = 0.
\]
For clarity, we define a finer topology on $\mathcal{X}$ by saying $\F^n \overset{\DTSSpace}{\to} \F$ if for all $T > 0$
\[
\lim_{n \to \infty}
\sup_{0 \leq s \leq T}
\sup_{0 \leq x}
| \F[s]^n(x) - \F[s](x)| \to 0,
\]
which is metrizable in a similar way as was done for $d_{\Lloc}.$ We let $\mathcal{Y}$ denote the resulting metric space.
\begin{lemma}
Suppose that $\F^n \overset{\DTSpace}{\to} \F,$ where $(t,x) \mapsto \F[t](x)$ is a continuous map from $[0,\infty)^2 \to [0,1]$ and for each $t,$ $\F[t](+\infty)=1$, then $\F^n \overset{\DTSSpace}{\to} \F.$
\label{lem:conv_upgrade}
\end{lemma}
\begin{proof}
Fix $T>0$ and $\epsilon>0$ and define the sequence of functions $f_n(t) = F_t(n)$ on $[0,T]$. By assumption, the functions $f_n$ are continuous and converge pointwise to $1$ as $n\to\infty$. By monotonicity, Dini's theorem implies that the convergence holds uniformly on $[0,T]$. Hence, we can find a $K > 0$ sufficiently large that $\F[t](K) > 1 - \epsilon/2$ for all $0 \leq t \leq T.$ By the continuity of $(t,x) \mapsto \F[t](x),$ we have that there exists a $\delta > 0$ sufficiently small that $|x-y| < \delta$ imply that $|\F[t](x) - \F[t](y)| < \epsilon/2,$ for all $t\in[0,T]$ and $x,y\in[0,K]$. Using that $\F[t](x) > 1-\epsilon/2$ for all $x\ge K$, it is easily shown that
\begin{align}
\label{eq:equicont}
|\F[t](x) - \F[t](y)| < \epsilon\quad\text{$\forall t\in[0,T],\,x,y\ge 0$ s.t. $|x-y| < \delta$}.
\end{align}
Now let $t\in[0,T]$ and $x\ge0$ such that $\F[t]^n(x) \ge \F[t](x) + 2\epsilon$. Note that necessarily, $x\le K$, otherwise $\F[t]^n(x) > 1+\epsilon$. Then by monotonicity of $\F[t]^n(x)$ in $x$ and \eqref{eq:equicont}, we have $\F[t]^n(y) > \F[t](y)+\epsilon$ for all $x < y < x+\delta.$ Hence,
\begin{align*}
\forall t\in[0,T]:\int_0^{K+\delta} |\F[t]^n(x) - \F[t](x)|\,dx \ge \epsilon\delta.
\end{align*}
Thus, since $\sup_{0\le t\le T} \int_0^{K+\delta} |\F[t]^n(x) - \F[t](x)|\,dx\to 0$ as $n\to\infty$, we must have
\[
\limsup_{n\to\infty} \sup_{0\le t\le T} \sup_{x\ge0} \F[t]^n(x) - \F[t](x) \le 2\epsilon.
\]
A similar argument shows that
\[
\liminf_{n\to\infty} \inf_{0\le t\le T} \inf_{x\in [\delta,K]} \F[t]^n(x) - \F[t](x) \ge -2\epsilon.
\]
By the monotonicity of $\F[t]^n$ and the fact that $\F[t](x) \ge 1-\epsilon$ for all $x\ge K$, we also have
\[
\liminf_{n\to\infty} \inf_{0\le t\le T} \inf_{x\ge K} \F[t]^n(x) - \F[t](x) \ge -3\epsilon.
\]
Furthermore, since $\F[t]^n\ge 0$ and $\F[t](x) \in [0,\epsilon]$ for all $x\in[0,\delta]$ by \eqref{eq:equicont}, we have
\[
\liminf_{n\to\infty} \inf_{0\le t\le T} \inf_{x\in [0,\delta]} \F[t]^n(x) - \F[t](x) \ge -\epsilon.
\]
Altogether, this shows that
\[
\limsup_{n\to\infty} \sup_{0\le t\le T} \sup_{x\ge0} |\F[t]^n(x) - \F[t](x)| \le 3\epsilon.
\]
As $\epsilon$ was arbitrary, we have that $\F^n \overset{\DTSSpace}{\to} \F.$
\end{proof}
This improved convergence allows us to control the Radon--Nikodym derivatives.
\begin{lemma}
Suppose that $\Psi \in C^1[0,1]$ with derivative $\psi.$
Suppose that $\F^n \overset{\DTSSpace}{\to} \F,$ where $(t,x) \mapsto \F[t](x)$ is a continuous map from $[0,\infty)^2 \to [0,1]$ with $\left\{ \F[t] : t \geq 0 \right\}$ tight.
%
Then
\[
\lim_{n \to \infty} \sup_{0 \leq s \leq T} \int_0^\infty
\left|
\frac{d\Psi(\F[s]^n)}{d\F[s]^n}(x)
-\psi(\F[s](x))
\right|
\,d\F[s]^n(x)
=0
\]
\label{lem:rnderiv}
\end{lemma}
\begin{proof}
It is enough to prove the statement with $\psi(\F[s](x))$ replaced by $\psi(\F[s]^n(x))$, as is easily shown using the uniformity in the definition of the convergence $\F^n \overset{\DTSSpace}{\to} \F$.
Define for any $a \in [0,1]$ and any $\delta > 0$
\[
\psi_\delta(a) = \sup_{\substack{|a-x| < \delta \\ |a - y| < \delta}}
\frac{\Psi(x)-\Psi(y)}{x-y},
\]
where $x,y \in [0,1].$ %
For any distribution function $F$ all of whose atoms are at most $\delta/2,$ there is a $\eta_0$ sufficiently small such that for all $a \geq 0$ and all $0 < \eta < \eta_0$
\[
F(a+\eta)-F(a) < \delta.
\]
Then by definition of $\psi_\delta,$
\begin{align*}
\int_a^{a+\eta} \psi_\delta(F(x))\,dF(x)
&\geq
\int_a^{a+\eta}\frac{\Psi(F(a+\eta)) - \Psi(F(a))}{F(a+\eta) - F(a)}\,dF(x)\\
&=\Psi(F(a+\eta)) - \Psi(F(a))
\end{align*}
It follows by taking unions that for any $U = (a,b]$ with $a<b$
\[
\int_U \psi_\delta(F(x))\,dF(x) > \int_U \frac{d\Psi(F)}{dF}(x)\,dF(x),
\]
and hence by a monotone class argument we have $\psi_\delta(F(x)) \geq \frac{d\Psi(F)}{dF}(x)$ $dF$--almost everywhere. %
From the convergence of $\F^n \overset{\DTSSpace}{\to} \F,$ the continuity of $\F$ and the tightness of $\F[t],$ for any $\delta >0$ there is an $n_0$ sufficiently large and a $\eta_0$ sufficiently small that for $n \geq n_0$ and $0 < \eta <\eta_0,$
\[
\sup_{a \geq 0} \sup_{0 \leq s \leq T} (\F[s]^n(a+\eta) - \F[s]^n(a)) < \delta.
\]
Hence by the same argument as above, for any $0 \leq s \leq T$
\[
\frac{d\Psi(\F[s]^n)}{d\F[s]^n}(x) \leq \psi_\delta(\F[s]^n(x)).
\]
For any $\epsilon > 0$, we can pick $\delta$ sufficiently small that
$\psi_\delta(p) \leq \psi(p) + \epsilon$ for all $p \in [0,1].$
Then,
\[
\limsup_{n \to \infty} \sup_{0 \leq s \leq T} \int_0^\infty
\left(
\frac{d\Psi(\F[s]^n)}{d\F[s]^n}(x)
-\psi(\F[s]^n(x))
\right)_+ \,d\F[s]^n(x)
\leq \epsilon.
\]
Taking $\epsilon \to 0$ shows this $\limsup$ is at most 0.
An almost identical proof using
\[
\psi^{-}_\delta(a) = \inf_{\substack{|a-x| < \delta \\ |a - y| < \delta}}
\frac{\Psi(x)-\Psi(y)}{x-y},
\]
shows
\[
\limsup_{n \to \infty} \sup_{0 \leq s \leq T} \int_0^\infty
\left(
\frac{d\Psi(\F[s]^n)}{d\F[s]^n}(x)
-\psi(\F[s]^n(x))
\right)_{-} \,d\F[s]^n(x)
\leq 0.
\]
By what was mentioned at the beginning of the proof, this implies the statement of the lemma.
\end{proof}
Using this additional strengthening of the mode of convergence, we now show the continuity of $\mathscr{C}$ at those limit points at which the first time evolving distribution function is continuous.
\begin{proof}[Proof of Proposition~\ref{prop:continuity}]
Let $(\F^n,\G^n)_{n\ge1}$ and $(\F,\G)$ be as in the statement of the proposition.
We must show that for any fixed $T, K >0,$
\[
\lim_{n \to \infty}
\sup_{0 \leq t \leq T}
\int\limits_0^K
|\mathscr{C}(\F^n,\G^n)_t(x)
-\mathscr{C}(\F,\G)_t(x)|\,dx = 0.
\]
We bound $\mathscr{C}(\F^n,\G^n)_t(x)-\mathscr{C}(\F,\G)_t(x)$ pointwise by the following pieces:
\begin{align}
|\mathscr{C}(\F^n&,\G^n)_t(x)-\mathscr{C}(\F,\G)_t(x)| \nonumber \\
&\leq |\G[0]^n(e^{-t}x) - \G[0](e^{-t}x)| \label{eq:cop1} \\
&+\int_0^t (e^{s-t}x)^2\int_{e^{s-t}x}^\infty
{\left|\frac{d\Psi(\F[s]^n)}{d\F[s]^n}(z) - \psi(\F[s](z))\right|}
\frac{d\G[s]^n(s)}{z}
\,ds \label{eq:cop2} \\
&+\int_0^t (e^{s-t}x)^2
\left|\int_{e^{s-t}x}^\infty
\frac{\psi(\F[s](z))}{z}
(
d\G[s]^n(z)
-d\G[s](z)
)
\right|\,ds.
\label{eq:cop3}
\end{align}
The first piece \eqref{eq:cop1} converges to $0$ simply by definition of $\G^n \overset{\DTSpace}{\to} \G.$ As for the second piece \eqref{eq:cop2}, we can bound
\begin{align*}
\int_0^t (e^{s-t}x)^2&\int_{e^{s-t}x}^\infty
{\left|\frac{d\Psi(\F[s]^n)}{d\F[s]^n}(z) - \psi(\F[s](z))\right|}
\frac{d\G[s]^n(s)}{z}\\
&\leq
\sup_{\substack{0 \leq s \leq t \\ 0 \leq z}}
\left|\frac{d\Psi(\F[s]^n)}{d\F[s]^n}(z) - \psi(\F[s](z))\right|
\int_0^t (e^{s-t}x)\int_{e^{s-t}x}^\infty
\,d\G[s]^n(z)ds \\
&\leq
x
\sup_{\substack{0 \leq s \leq t \\ 0 \leq z}}
\left|\frac{d\Psi(\F[s]^n)}{d\F[s]^n}(z) - \psi(\F[s](z))\right|.
\end{align*}
Hence by Lemma~\ref{lem:rnderiv}, this converges to $0$ in $\mathcal{X}.$
Finally, for the last piece \eqref{eq:cop3}, we let, for any $s,u \geq 0,$
\(
Q_s^n(u) = \int_u^\infty \frac{\psi(\F[s](z))}{z}\,d\G[s]^n(z),
\)
and let $Q$ denote the same with $\G^n$ replaced by $\G.$ By weak convergence of $\G[s]^n$ to $\G[s],$ at every point of continuity of $\G[s],$ we have that $Q_s^n(u) \to Q_s(u)$ as $n \to \infty.$
Hence, we have that at fixed $t,$
\begin{align*}
\int_0^K
\int_0^t (e^{s-t}x)^2
&\left|\int_{e^{s-t}x}^\infty
\frac{\psi(\F[s](z))}{z}
(
d\G[s]^n(z)
-d\G[s](z)
)
\right|\,ds\,dx \\
&=
\int_0^K
\int_0^t (e^{s-t}x)^2
\left|Q^n_s(e^{s-t}x) - Q_s(e^{s-t}x)\right|\,ds\,dx \\
&=
\int_0^t
\int_0^K
(e^{s-t}x)^2
\left|Q^n_s(e^{s-t}x) - Q_s(e^{s-t}x)\right|\,dx\,ds \\
&=
\int_0^t
e^{t-s}
\int_0^{e^{t-s}K}
u^2
\left|Q^n_s(u) - Q_s(u)\right|\,du\,ds.
\end{align*}
Let $H(s,t)^n$ be given by
\[
H(s,t)^n = \int_0^{e^tK} x^2 \left|Q^n_s(x) - Q_s(x)\right|\,dx.
\]
Observe that we can bound $Q_s^n(u) \leq \frac{\|\psi\|_\infty}{u}.$ Hence by dominated convergence, for every fixed $s,t$ with $0\leq s \leq t,$
\[
\lim_{n \to \infty} H(s,t)^n = 0.
\]
Further, we have that $H(s,t)^n$ can be bounded uniformly by $2e^{2t}K^2\|\psi\|_\infty.$ Hence again by dominated convergence, we have that
\begin{align*}
\sup_{0 \leq t \leq T}
\int_0^K
\int_0^t (e^{s-t}x)^2
&\left|\int_{e^{s-t}x}^\infty
\frac{\psi(\F[s](z))}{z}
(
d\G[s]^n(z)
-d\G[s](z)
)
\right|\,ds\,dx \\
&\leq \int_0^T e^{T} H(s,T)^n\,ds,
\end{align*}
which converges to $0$ by dominated convergence.
\end{proof}
\printbibliography
\end{document}
|
\section{Mining Docstrings}
\subsection{Mining Argument Schemas}
Sphinx and numpydoc pre-parse the documentation
into a list of argument tuples of the form
\python{(name, short_desc, long_desc)}.
The CNL parser uses the grammar in Figure~\ref{fig:type_grammar_full} to
extract per-argument schemas from the \python{short_desc} fields.
The start symbol of the grammar, \bnf{start}, splits the docstring into
three parts:
the type, or a sequence of possible types (encoded in JSON using the \json{anyOf} keyword); an optional flag; and the default value.
\subsection{Mining Constraints}
For inter-argument constraints, the CNL parser first extracts complete sentences from the long description.
Next, it uses a set of regular expression rules to flag possible candidate constraints, for example, sentences containing the word `only'.
On each candidate it discovers, the CNL parser uses the grammar in \Cref{fig:constraint_grammar_full}.
Unfortunately, there is great variety in how docstrings express conditional hyperparameter constraints.
Since our CNL grammar does not anticipate the syntax of this example, it cannot fully extract this constraint.
However, it does detect the presence of some constraint, and it puts a placeholder into the raw JSON Schema with a \json{TODO} that a human domain expert can then fill in later to further enhance it.
\begin{figure}
\begin{lstlisting}[language=bnf]
start ::= seq optional default ("." | ",")?
seq ::= (type ","?)+ ("or " type)?
type ::= "int" | "integer" | "float" | "double" | "boolean" | "bool" | "string" | "str"
| "None" | "Ignored" | "callable" | "dict" | "type"
| obj | array | enum
optional ::= (", optional")?
default ::= ","? ("default" ("=" | ":")? val
| "(" "default" ("=" | ":")? val ")"
| val "by " "default"
| "or " val "(" "default" ")")?
obj ::= "object" | "RandomState" "instance" | "returns an instance of self"
array ::= atype (shape)?
atype ::= "list" | "array" | "tuple" | "array_like" | "array-like"
| "numpy" "array" | "sparse" "matrix" | "scipy.sparse" | "scipy" "sparse"
| "{"? atype ("or " | ",") atype "}"?
shape ::= ","? "of "? ("shape" | "size")? "="? vtuple ("or " shape)?
enum ::= "{" val (","? "or "? ("an " | "a ")? val)* "}"
| ("string" | "str") ","? enum
| "["? val ("|" val)+ "]"?
vtuple ::= ("(" | "[") val ("," val)* ","? ("]" | ")") | "None"
val ::= NAME | NUMBER
\end{lstlisting}
\caption{\label{fig:type_grammar_full}CNL grammar for parsing per-argument schemas.}
\end{figure}
\begin{figure}
\begin{lstlisting}[language=bnf]
start ::= only when cond
only ::= "only" ("used" | "effective" | "compatible" | "significant" | "available" | "applies")?
when ::= "when" | "if" | "with" | "in" | "for"
cond ::= atom | cond ("and" | "or") cond
atom ::= NAME compare seq | "the"? seq NAME ("is used")?
seq ::= val (("," val)* ","? ("and" | "or") val)?
compare ::= "==" | "=" | ">" | "<" | ">=" | ">=" | "is set to" | "is"
val ::= NUMBER | NAME
\end{lstlisting}
\caption{\label{fig:constraint_grammar_full}CNL grammar for parsing inter-argument constraints.}
\end{figure}
\section{Results}
This section measures the effectiveness of our extractor tool on two
datasets: complete and curated. The \emph{complete dataset} comprises
119\xspace operators from three different libraries: 115\xspace from
scikit-learn~\cite{buitinck_et_al_2013}, 2 from
XGBoost~\cite{chen_guestrin_2016}, and 2 from
LightGBM~\cite{ke_et_al_2017}.
For XGBoost and LightGBM we considered both the regressor and the classifier.
For scikit-learn we filter the classes revealed by \python{sklearn.utils.testing.all_estimators()} to obtain estimators and transformers.
We exclude classes that are abstract, or that are meta-estimators.
Further we examine their method resolution order, confirm the existence of \python{fit}, and \python{predict} or \python{transform}, and confirm their signatures.
Finally, we exclude some classes known to be deprecated, e.g., \python{Imputer}, and some known to be intended only to be used by other classes, e.g., \python{ExtraTree}.
The \emph{curated dataset} comprises 42\xspace operators with manually curated schemas: 38\xspace
from scikit-learn, 2 from XGBoost, and 2 from LightGBM. Whereas the
complete dataset allows us to evaluate the robustness and coverage of
the tool overall, the curated dataset allows us to compare the
extracted schemas against a ground truth for reference.
Finally, we used the generated schemas to find three-steps pipelines for 15 OpenML datasets with Lale and compare the results with with auto-sklearn.
\subsection{Complete Dataset Evaluation}\label{apx:complete}
Table~\ref{table:global_results} presents the results of the extractor
executed on the complete dataset. For each library, it reports the
number of extracted arguments, types, default values, ranges, and constraints.
In addition, Table~\ref{table:global_results} explicitly shows the contributions of the CNL parser and the schema refiner for each category to illustrate the advantages of combining the two strategies.
Overall, we were able to mine $94\%$ of the~$1,867$ argument types (including the input/output schemas of the \python{fit}, \python{transform}, and \python{predict} methods of all the operators).
Missing argument types mostly come from unsupported or under-specified data structures, e.g., \python{object}, \python{dict}, or \python{callable}.
Even when mined, it is not clear how an AI automation tool would
instantiate such arguments during search.
We extracted a default value for $64\%$ of the arguments but default values are not always relevant, e.g., for the input/output type of the \python{fit} or \python{predict} method.
Table~\ref{table:global_results} shows that, even if default values
are often documented, the schema refiner can extract a lot of
additional information, e.g., via dynamic analysis.
Since ranges are not consistently documented, the range analysis is solely based on the schema refiner.
We found a valid range for~$50\%$ of the~$792$ relevant arguments, i.e., numeric arguments (\json{integer} or \json{number}) or \json{string} arguments used to captures \json{enum} values.
However, ranges are not required for all of these arguments.
In fact we observe that, even in the curated schemas, ranges are only defined for a few arguments to produce valid search spaces for hyperparameter search tools.
Finally, we detected~$120$ constraints but only~$43$ were converted into valid JSON Schema. The remaining~$77$ generate \json{TODO} warnings that can be manually inspected.
\subsection{Curated Dataset Evaluation}\label{apx:curated}
Next, we compare the result of the extractor with a set of manually curated schemas.
Results are presented in Table~\ref{table:curated_results}.
Compared to Table~\ref{table:global_results} we focus on the
arguments to the operator's constructor (\python{__init__}),
leaving aside the arguments and return values of its \python{fit},
\python{predict}, or \python{transform} methods
For each category we report the reference number, the generated
number, and the number of matches between the two sets. We also report
the corresponding precision, recall, and $F_1$ score.
The extractor is able to detect all the~$452$ arguments, which indicates that all the arguments are consistently documented across the three libraries.
In the curated set, all arguments have a correct type and a default value.
The extractor correctly mined the type for~$81\%$ of the arguments and the default value for~$97\%$.
The extractor also found a valid range for~$81\%$ of the~$103$ defined ranges in the curated set.
We categorized a range as valid if the extractor returns an interval
included in the range defined in the curated schema.
Compared to the complete dataset, these counts are relatively low because the curated dataset includes the XGBoost and LightGBM operators that are both more complex and less documented.
Mismatches between extracted and curated default values are due to values that cannot be represented in JSON Schemas: the default value of the \python{missing} argument of the XGBoost operators and the \python{missing_values} of \python{SimpleImputer} is \python{nan}, which is replaced by \python{None} in the curated schemas.
Additionally, our extractor found inconsistencies between the documentation and the code.
For instance, the documentation for the \python{categories} argument of \python{OneHotEncoder} is:
\begin{lstlisting}[language=desc, numbers=none]
categories : 'auto' or a list of lists/arrays of values, \
default='auto'.
\end{lstlisting}
but the \python{OneHotEncoder.__init__} method gives a default value
\python{categories=None}.
Arguments types are often complex \emph{union} types allowing multiple choices for one argument. For example, \python{n_jobs} can often be either \python{None} or an \json{integer}.
To further investigate the discrepancies between generated and curated types, we analyzed the numbers of \emph{type values} and \emph{enum values}.
The \emph{type values} analysis reports the number of \emph{terminal} types found in the schemas, i.e, \json{boolean}, \json{integer}, \json{number}, \json{string}, or \json{enum}.
The \emph{enum values} analysis reports the number of members found in each \json{enum} list.
\begin{center}
\begin{tabular}{l r r r r r r}
& reference & generated & match &precision & recall & $F_1$\\
\toprule
type values & 631 & 611 & 525 & 0.86 & 0.83 & 0.85\\
enum values & 426 & 269 & 238 & 0.88 & 0.56 & 0.68\\
\bottomrule
\end{tabular}
\end{center}
We observe that for both analyses the precision is relatively high:
$86\%$ for type values and $88\%$ for enum values, which suggests that
extracted data are mostly correct.
However, for enum values the generated number is significantly lower than the curated number: the extractor only found $56\%$ of the enum values in the curated dataset.
This is mostly due to arguments that are under-specified as
\python{string} in the documentation and for which the schema
refiner can not find a suitable enumeration.
For example, the documentation of the \python{criterion} argument
of \python{GradientBoostingClassifier} is:
\begin{lstlisting}[language=desc, numbers=none]
criterion: string, optional (default="friedman_mse")
\end{lstlisting}
but the valid enumeration is
\json{['friedman_mse', 'mse', 'mae']}.
These under-specifications are relatively common when the enum value is only one possible choice in a complex union type.
For example, the documentation of the \python{max_features} argument
of \python{DecisionTreeClassifier} is:
\begin{lstlisting}[language=desc, numbers=none]
max_features: int, float, string or None, optional \
(default=None)
\end{lstlisting}
but again the valid enumeration is \json{['auto', 'sqrt', 'log2']}.
Finally, the extractor detected~$50$ of the~$65$ constraints of the curated set.
Among the detected constraints, $20$ are converted into valid JSON Schema and~$90\%$ of these match the curated schemas.
The mismatches are due to complex constraints that are merged in the curated schemas.
For instance, the description of the \python{power_t} argument of \python{MLPClassifier} contains:
\begin{lstlisting}[language=desc, numbers=none]
It is used in updating effective learning rate when
the learning_rate is set to 'invscaling'.
Only used when solver='sgd'.
\end{lstlisting}
which is captured in the curated schemas as: "\python{power_t} can only differ from its default value $0.5$ if \python{solver == 'sgd'} and \python{learning_rate == 'invscaling'}", or in JSON Schema:
\begin{lstlisting}[language=json, numbers=none]
'anyOf': [
{'type': 'object',
'properties': {'power_t': {'enum': [0.5]}}},
{'type': 'object',
'properties': {
'learning_rate': {'enum': ['invscaling']},
'solver': {'enum': ['sgd']}}}]}
\end{lstlisting}
but the extractor is only able to extract the second condition on the solver (\python{solver == 'sgd'}).
The results for the constraints show that, even if we are able to flag
most of the constraints, our CNL parser is not the best tool to
extract meaningful information from the constraint candidate.
The language used to express the constraints is far less constrained
than the one used for the type description. An obvious direction for
future work is thus to try classic natural language understanding
techniques.
\subsection{AutoML Pipelines}\label{apx:autoai}
The selected datasets comprise 5 simple classification tasks (test accuracy~$>90\%$ in all our experiments) and 10 relatively complex tasks (test accuracy~$< 90\%$).
For all the tasks we start from the same three-step pipeline with both generated --- lale-gen in \Cref{tab:accuracy_openml} --- and curated schema --- lale-cur in \Cref{tab:accuracy_openml}.
For comparison, we used auto-sklearn~\cite{feurer_et_al_2015} --- autoskl in \Cref{tab:accuracy_openml} --- as a baseline.
All tasks were configured with the same resource constraints: one hour
of optimization time and a timeout of six minutes per trial.
Lale then uses the search spaces defined in the schemas, the topology of the pipeline, and off-the-self optimizers such as Hyperopt~\cite{komer_bergstra_eliasmith_2014}, to find the best candidate.
\begin{lstlisting}[language=python]
preprocessors = [ NoOp, MinMaxScaler, StandardScaler, Normalizer, RobustScaler]
features_extractors = [ NoOp, PCA, PolynomialFeatures, Nystroem]
classifiers = [ GaussianNB, GradientBoostingClassifier, KNeighborsClassifier,
RandomForestClassifier, ExtraTreesClassifier,
QuadraticDiscriminantAnalysis, PassiveAggressiveClassifier,
DecisionTreeClassifier, LogisticRegression, XGBClassifier,
LGBMClassifier, SVC ]
lale_pipe = make_pipeline( make_choice(*preprocessors),
make_choice(*features_extractors),
make_choice(*classifiers) )
\end{lstlisting}
For each experiment, we used a $66\%-33\%$ validation-test split, and a 5-fold cross validation on the validation split during optimization.
Experiments were run on a 32 cores (2.0GHz) virtual machine with 128GB memory.
\Cref{tab:accuracy_openml} shows the accuracy results (mean and standard deviation across 5 independent runs) and the number of runs for each experiments (where ``ok'' indicates a successful run and ``ko'' indicates an aborted run).
We observe that Lale with our auto-generated schemas achieves accuracies (88.1\%) that are comparable to auto-sklearn (88.0\% with auto) and Lale with curated schemas (88.5\%).
However, the number of aborted runs show that side-constraints play a key role during the optimization process.
Only 4.0\% of the runs using curated schemas were aborted, compared to 21.7\% with generated schemas (and 6.7\% for auto-sklearn).
\input{table_autoai_results}
\section{Conclusion}\label{sec:conclusion}
This paper presents a tool that mines Python docstrings of ML
libraries to extract hyperparameter schemas for HPO. The extracted
schemas include names, types, defaults, and descriptions, ranges and
distributions for continuous hyperparameters, enumerations of
constants for categorical hyperparameters, and constraints for
conditional hyperparameters.
\section{Mining Docstrings}\label{sec:docstrings}
This section describes the CNL Parser component of the overview
diagram in Figure~\ref{fig:overview}. A CNL (controlled natural
language) is a natural language (e.g., English) with some
amount of structure (e.g., the input format for Sphinx and numpydoc,
adopted by scikit-learn and other ML libraries).
The CNL parser starts by reading the docstring from the Python file,
such as the one on the left of \Cref{fig:lr_hp}. It uses Sphinx
and numpydoc to extract the docstring of methods \python{__init__}
(class constructor), \python{fit} (for training), and
\python{predict} or \python{transform} (for using an operator
after training) of the operator class.
Sphinx and numpydoc pre-parse this information
into a list of argument tuples of the form
\python{(name, short_desc, long_desc)} as well as
descriptions of the return values of the methods.
Given the list of argument tuples, the
CNL parser uses two hand-crafted CNL grammars to extract per-argument
schemas and inter-argument constraints, respectively.
\paragraph{Mining Argument Schemas.}
The CNL parser uses a grammar (see \Cref{fig:type_grammar_full} in the appendix) to
extract per-argument schemas from the \python{short_desc} fields and inter-argument constraints that appear in the \python{long_desc}.
The example of \Cref{fig:lr_hp} illustrates the main difficulties for mining and
extraction. The parser needs to ignore noise such as whitespace,
string quotes, or the trailing backslash~(\textbackslash).
In addition, even within the
same operator, the documentation uses different ways to express the
same thing, e.g., enumerating values in curly braces \lstinline!{...}!
vs.\ using \texttt{or}.
Overall, these difficulties arise from the `N' in
CNL: docstrings use natural language.
Noise is easy to handle using filtering during
lexical analysis. For the other difficulties, our grammar takes
advantage of the `C' in CNL: docstrings use controlled language to the
extent that they follow the conventions encouraged by Sphinx and
numpydoc.
The grammar thus specifies multiple syntactic alternatives to capture different ways to express the
same thing (e.g., \bnf{enum} or \bnf{default}).
\paragraph{Mining Constraints.}
For inter-argument constraints, the CNL parser first extracts complete
sentences from the long description. Next, it uses regular
expressions to flag possible candidate constraints, for example,
sentences containing the word `only'. Then it parses each candidate
using a grammar that captures common patterns (see
\Cref{fig:constraint_grammar_full} in the appendix).
Unfortunately, there is great variety in how docstrings express conditional
hyperparameter constraints.
Our grammar is only a first attempt to extract meaningful information.
When the CNL parser fails to parse a sentence flagged as a potential constraint, it puts a
placeholder into the schema with a \json{TODO}
that a human can fill in later.
Having mined both per-argument schemas and inter-argument constraints,
the last step of the CNL parser is to assemble all the pieces into a
single raw schema.
The resulting JSON Schema is machine-readable and captures the
information in a format suitable both for validation and for search.
\section{Introduction}\label{sec:intro}
Machine-learning practitioners use libraries of \emph{operators}:
reusable implementations of estimators (such as logistic regression,
LR) and transformers (such as principal component analysis, PCA).
Training an operator fits its \emph{parameters} (learnable
coefficients such as LR weights or PCA eigenvectors) to a dataset.
Besides parameters, most operators also have \emph{hyperparameters}:
arguments that must be configured before training, such as the choice
of LR solver or the number of PCA components. Python libraries for
machine learning (ML) such as scikit-learn~\cite{buitinck_et_al_2013} tend
to have good human readable documentation for hyperparameters.
Unfortunately, this documentation is usually not easily machine
readable.
ML practitioners can configure hyperparameters either by hand or by
using an HPO (automated hyperparameter optimization) tool such as
hyperopt-sklearn~\cite{komer_bergstra_eliasmith_2014} or
auto-sklearn~\cite{feurer_et_al_2015},
or the grid search or randomized search from
scikit-learn. A \emph{hyperparameter schema}
specifies which hyperparameters are categorical and which continuous,
which values or ranges are valid, and conditional hyperparameter
constraints.
\medskip
\begin{wrapfigure}{r}{0.55\textwidth}
\centerline{\includegraphics[scale=0.45]{figures/overview_2.pdf}}
\vspace*{-2mm}
\caption{\label{fig:overview}Overview of our mining approach.}
\end{wrapfigure}
Python has recently emerged as the dominant ML language and many ML
libraries adopt scikit-learn style conventions for interoperability
(including
PyTorch~\cite{skorch},
pandas~\cite{sklearn-pandas},
Spark~\cite{spark-sklearn},
statsmodels~\cite{statsmodels}, and
TensorFlow~\cite{keras-sklearn}).
This
paper proposes and demonstrates an approach for mining hyperparameter schemas from the
Python file implementing an ML operator. The approach, shown in
Figure~\ref{fig:overview}, mines the docstring and refines the
resulting schema via dynamic analysis of the implementation. Using the
Python file as a single source of truth simplifies maintenance when a
new library version adds new features or deprecates old ones.
Furthermore, since the Python file is written by library developers and the
documentation is widely read by library users, it is a reliable source of
truth.
Our approach outputs hyperparameter schemas in \emph{JSON Schema},
which is a type description language for JSON
documents~\cite{pezoa_et_al_2016}. JSON Schema is widely adopted for
web APIs, cloud management, and document databases, among other
domains, and there are abundant public resources for learning and
using it. JSON Schema is independent from specific AI automation
tools and recent work has demonstrated that it can be converted to
specifications for popular such tools~\cite{baudart_et_al_2020}. We found
JSON Schema to have just the right expressiveness for hyperparameters
including categoricals and conditional constraints. Furthermore, we
found JSON Schema easy to extend with additional meta-data such as
distributions for continuous hyperparameters.
This paper makes the following contributions:
(1)~Mining Python docstrings to extract hyperparameter schemas including constraints.
(2)~Using dynamic analysis to obtain additional information about hyperparameters beyond the docstrings.
(3)~Reconciling hyperparameter metadata into a single machine-readable schema in JSON Schema format.
We evaluate our approach on 119\xspace automatically mined
hyperparameter schemas for ML operators and 42\xspace hand-curated schemas.
We make both datasets publicly available (\url{https://github.com/IBM/lale/tree/master/lale/lib/}\textsf{\small(autogen$|$sklearn)}).
Overall, we hope this paper contributes towards making HPO tools
easier to use, more reliable, and more effective.
\section{Problem Statement}\label{sec:problem}
This paper is about solving the problem of mining hyperparameter
specifications from a Python docstring and turning them into a JSON
Schema. To make things concrete, \Cref{fig:lr_hp} show an example input and the desired corresponding output of this mining problem.
The left side of \Cref{fig:lr_hp} shows an excerpt of class
\python{sklearn.linear_model.LogisticRegression} with its docstring. A
\emph{docstring} is a string literal that documents a specific class
or function definition. The HTML documentation for scikit-learn and other
popular ML libraries is auto-generated from their docstrings. For
this to work, the docstrings follow conventions understood by the HTML
generation tool, in this case, the numpydoc
extension~\cite{numpydoc_2008} for Sphinx~\cite{brandl_2008}. In
other words, docstrings are written in a controlled natural language
(CNL) \cite{kuhn_2014}: controlled, since they follow numpydoc
conventions, and natural, since they are human-readable even before
being converted into HTML. In practice, while docstrings suffice for
HTML generation, they exhibit variability and typos that make schema
extraction non-trivial.
This paper proposes an extractor that converts the docstring not to
HTML but to JSON Schema. The right side of
\Cref{fig:lr_hp} shows the schema for two categorical arguments
\python{solver} and \python{penalty} and one continuous
arguments \python{C}.
Like most ML libraries, scikit-learn encodes categorical hyperparameters via Python string constants as opposed to Python enums, but only the values
mentioned in the documentation are valid.
The example also contains a conditional hyperparameter constraint:
the value of \python{solver} implies which values are valid for
\python{penalty}. We can express this implication by taking the logically
equivalent form
\mbox{\small$\neg\textit{\textsf{premise}}\vee\textit{\textsf{conclusion}}$}
and using the JSON Schema keywords \json{anyOf} and~\json{not}.
\begin{figure}[t]
\begin{minipage}{0.49\textwidth}
\input{lr_code}
\end{minipage}
\begin{minipage}{0.51\textwidth}
\input{lr_schema}
\end{minipage}
\caption{Simplified excerpt of the scikit-learn code for the LogisticRegression estimator~(left) and the corresponding hyperparameters schema (right).}
\label{fig:lr_hp}
\vspace*{-4mm}
\end{figure}
\section{Refining Mined Meta-Data}\label{sec:refining}
This section describes the Schema Refiner component of the overview
diagram in Figure~\ref{fig:overview}. This component uses dynamic
analysis on the Python code to make additional observations, which
it combines with heuristics and overrides to turn the raw schema
from the CNL Parser into a refined schema for HPO tools.
\paragraph{Dynamic analysis for default values.}
Non-algorithmic defaults complicate the analysis of types
and values.
This occurs, for example, when an argument default is appropriated for
purposes other than parameterization.
To illustrate, it has become
relatively commonplace within scikit-learn to advise users of upcoming
changes in defaults for important arguments by setting the default
value in the constructor's signature to \python{'warn'}
(e.g., Figure~\ref{fig:lr_hp} (left)) to trigger a warning
message.
Our dynamic analysis creates an instance by calling the constructor
\python{__init__()} without passing any explicit arguments, then
calls \python{fit} on the resulting instance --- which might assign the argument to its actual algorithmic default value --- and finally
introspects the instance for these (possibly altered) default values
and their types.
\paragraph{Dynamic analysis via runtime exception testing.}
We use the following techniques to harvest good values and filter bad
values for constructor arguments.
\noindent
\emph{Bad values:}
Defaulting all other arguments, we give a deliberately bad value
to the argument under test and capture the exception. This
exception text usually reports the bad value which is easily
distinguished in the message. Frequently, the exception text also
reports valid choices for the argument value, using a range of
syntax that we can parse.
\noindent
\emph{Greedy harvesting:}
We allow argument values that are valid for one operator to be tested
for the same argument name in a different operator. This occasionally
discovers valid values, particularly for under-documented classes.
\noindent
\emph{Sampling:}
Defaulting all other arguments, an argument's range is sampled for
testing for valid values. If categorical, all values are tested.
Failed values are filtered out for the class. The message
received for deliberately false values can help to disambiguate
the complaint in the case where it is not known a priori if
the tested value is good or bad.
\noindent
\emph{Bounds testing:}
With the caveat that some bounds may depend upon
data and the values of other arguments, the bounds of continuous
ranges can be tested individually for validity and
exclusivity, which can be expressed in JSON Schema
via e.g., \json{'exclusiveMinimum': True}.
\paragraph{Argument Overrides and Relevance to HPO.}
We provide for a dictionary of overrides that allow the automatically
extracted types and ranges to be replaced with user-specified values,
or for the parameter to be excluded from optimization.
E.g., suppressing the \python{'mae'} choice for \python{'criterion'} on tree regressors because of prolonged fit times, or custom bounds for numeric parameters.
The \python{ForOptimizer} suffix indicates that these are not hard constraints.
This step also sets the \python{distribution} field
(e.g.\ \python{loguniform}) and the \python{relevantToOptimizer}
list, omitting irrelevant arguments such as verbosity.
\section{Related Work}\label{sec:related}
The most closely related work is jDoctor, which mines javadoc comments
to extract method pre- and post-conditions~\cite{blasi_et_al_2018}.
The results of mining are similar to schemas in that they can capture
argument ranges and even some constraints. But jDoctor focuses on
Java, whereas we focus on Python code without static type annotations
and with string constants. Furthermore, jDoctor focuses on testing,
whereas we focus on AutoML.
Our schema refiner uses dynamic analysis on Python code to augment the
information extracted from docstrings. Fuzz
testing, also known as \emph{fuzzing}, is a well-established approach
for finding software defects by generating random
inputs~\cite{miller_fredriksen_so_1990}. While our schema refiner is
inspired by fuzzing, its goal is not to find defects but to
extract schemas.
The primary contribution of this paper is the documentation mining, not
the chosen output format. We could have used different formats to
express hyperparameter schemas.
Python~3 introduces optional type annotations that can be checked
statically~\cite{vanrossum_lehtosalo_langa_2014}. Unfortunately, since
Python~3 types lack intersection types, string constant types, and
conditional constraints, they are less suitable for HPO.
PCS is a file format for specifying hyperparameter schemas for the
SMAC tool~\cite{hutter_ramage_2015}. PCS is well-suited for HPO and
JSON Schema can be converted to PCS~\cite{baudart_et_al_2020}.
\section{Results}\label{sec:results}
This sections measures the effectiveness of our extractor tool with three experiments.
(1)~We mined the schema of 119\xspace operators from three different libraries: 115\xspace from
scikit-learn~\cite{buitinck_et_al_2013}, 2 from
XGBoost~\cite{chen_guestrin_2016}, and 2 from
LightGBM~\cite{ke_et_al_2017}.
(2)~We compared the schemas of 42\xspace operators with manually curated schemas: 38\xspace
from scikit-learn, 2 from XGBoost, and 2 from LightGBM.
(3)~We used the generated schemas to find three-steps pipelines for 15\xspace OpenML datasets with Lale~\cite{baudart_et_al_2020}, an Auto-ML
library that uses hyperparameter search spaces in JSON Schema.
\input{table_global_results}
\paragraph{Complete dataset.}~\\
\Cref{table:global_results} presents the results of the extractor
executed on the complete dataset (see also \Cref{apx:complete}).
For each category, we report the number mined by the CNL parser and the corrections made by the schema refiner ($\textit{parser} + \textit{refiner}$).
For the constraints we report the number of valid constraints and the number of detected constraints ($\textit{valid}/\textit{detected}$).
Overall, we were able to mine $94\%$ of the~$1,758$ argument types (including the input/output schemas of the \python{fit}, \python{transform}, and \python{predict} methods of all the operators).
We extracted a default value for $64\%$ of the arguments but default values are not always relevant, e.g., for the input/output type of the \python{fit} or \python{predict} method.
We found a valid range for~$50\%$ of the~$790$ relevant arguments, i.e., numeric arguments (\json{integer} or \json{number}) or \json{string} arguments used to captures \json{enum} values.
Finally, we detected~$120$ constraints but only~$43$ were converted into valid JSON Schema.
\paragraph{Curated dataset.}
\Cref{table:curated_results} presents the results of the comparison (see also \Cref{apx:curated}).
For this experiment, we focused on the arguments to the operator's constructor.
The extractor correctly mined the type for~$81\%$ of the arguments and the default value for~$97\%$.
\input{table_curated_results}
The extractor also found a valid range for~$81\%$ of the~$103$ defined ranges in the curated set, and~$75\%$ of the distributions.
Finally, the extractor detected~$50$ of the~$65$ constraints of the curated set.
Among the detected cons\-traints, $20$ are converted into valid JSON Schema and~$90\%$ of these match the curated schemas.
\paragraph{Auto-ML pipelines.}
To demonstrate the use of our schemas, we use Lale pipelines of the
form \python{preprocessor >> feature_extractor >> classifier}. Then,
we let AutoML automatically select each step from a predefined set of
operators (see details in \Cref{apx:autoai}) and tune its
hyperparameters based on our extracted schemas.
For comparison, we used auto-sklearn~\cite{feurer_et_al_2015} with the
same resource constraints as a baseline: 1h of optimization time and a
timeout of 6mn per trial.
Note that in this comparison, both the framework and the
hyperparameter schemas differ.
The results show that Lale with our auto-generated schemas achieves
similar accuracies as auto-sklearn, a state-of-the-art tool.
|
\section{Introduction}
\vspace{-0.3cm}
Generative Adversarial Networks (GANs)~\cite{goodfellow2014generative} have achieved great success in learning high dimensional probability distributions. The capability of GANs advances many important and useful applications, like high quality image generation~\cite{karras2017progressive,brock2018large}, image-to-image translation~\cite{isola2017image,wang2018high}, image editing~\cite{perarnau2016invertible, gu2019mask}, and so on. However, GANs still suffer from the unstable training process.
Many recent works~\cite{radford2015unsupervised,mao2017least,arjovsky2017wasserstein,gulrajani2017improved,miyato2018spectral} focus on stabilizing the training process of GANs. Despite their success, there are still two main issues in current GANs. First, the learned distribution always contains massive low quality samples. We refer to this phenomenon as the \emph{low quality} problem. Second, the learned distribution misses certain areas of real data distribution. We refer to this phenomenon as the \emph{missing modes} (also known as mode dropping) problem. We have carefully studied these two problems and observed that they may be caused by the inaccurate gradient direction from the discriminator.
Ideally, the gradient direction of the discriminator to the generator will make the generated data distribution as close to the real data distribution as possible. However, we found that in practice, the discriminator cannot achieve the theoretical optimal point due to insufficient iteration number or limited fitting ability. In this situation, the gradient direction given by the discriminator is sometimes inaccurate. Previous methods~\cite{gulrajani2017improved,miyato2018spectral} try to solve the inaccurate gradient problem with the Lipschitz constraint for the discriminator. But the Lipschitz constraint will further limit the capability of the discriminator so that it still can not provide accurate gradient for the generated data. More recent work LGAN~\cite{zhou2019lipschitz} proved that penalizing Lipschitz constant
guarantees the gradient for generated data is accurate, but it requires the discriminator need to be optimal before each iteration of generator training, which is impractical for GANs training.
To address these issues, we adopt an auxiliary model to help the generated data get the accurate gradient direction. We propose to build an explicitly prior to the real data distribution for GANs, which we call PriorGANs. Specifically, we choose to adopt the Gaussian Mixture Model (GMM) to establish a probabilistic model for the real data distribution. Equipped with this model, we can address both issues. For the low quality issue, we can apply the model to estimate the probability of each generated image, the probability can also indicate image quality. By encouraging the low quality samples to get a higher probability in the model, low quality samples will get a gradient direction to the real data nearby. Therefore the low quality problem could be solved. We call this the \emph{quality loss}.
For the missing modes issue, we can take full advantage of the GMM model to divide the real data into several groups. Each group may represent a specific mode. Then we can get the densities of the real data in each group. During the training process of GANs, we can estimate the densities of the generated data in each group. If the density of generated data in a group is lower (or higher) than real data, we will increase (or decrease) the sampling frequency of the real data in that group for the next training iteration. We call this the \emph{resampling strategy}. It will help the generative model solve the missing modes problem
The proposed PriorGAN is simple, flexible, and general. Our experiments demonstrate that it is suitable for various types of GANs, and it outperforms state-of-the-art GAN frameworks, including LSGAN, WGAN-GP, SNGAN and StyleGAN, on CIFAR-10, FFHQ, LSUN-cat and LSUN-bird datasets by large margins.
In summary, our key contributions are:
\begin{enumerate}
\vspace{-0.1cm}
\item We demonstrate the \emph{low quality} and \emph{missing modes} problems in current GANs and demonstrate the cause of these two problems on toy examples.
\vspace{-0.1cm}
\item We propose \emph{real data prior} for GANs(PriorGANs), equipped with the prior, we propose the \emph{quality loss} and \emph{resampling strategy} help current GANs solve these problems.
\vspace{-0.1cm}
\item Our experiments demonstrate that PriorGANs outperform state-of-the-art GAN frameworks on various datasets by large margins.
\vspace{-0.1cm}
\end{enumerate}
\iffalse
\fi
\begin{comment}
\begin{figure*}[t]
\centering
\includegraphics[width=1.0\columnwidth]{problem.pdf}
\vspace{-0.7cm}
\caption{Two problems in GAN training, the imprecise boundary problem and mode missing problem.}
\label{fig:gan_problem}
\vspace{-0.2em}
\end{figure*}
\end{comment}
\vspace{-0.3cm}
\section{Related Work}
\vspace{-0.3cm}
Over the past years, GANs have achieved significant attention in the deep learning community. Massive works are proposed to improve the performance of GANs. In this section, we will briefly review them. Also, we will review some classical algorithms about modeling data distribution.
\noindent \textbf{Generative Adversarial Nets}.
Early works about GANs try to solve the unstable training process of GANs, many works focus on proposing new loss functions to solve this problem. For example, LSGAN~\cite{mao2017least} replace the original BCE loss with the proposed least square loss. WGAN~\cite{arjovsky2017wasserstein} proposes the Wasserstein distance and weight clipping to achieve the Lipschitz condition. Later work began to notice the Lipschitz condition. Some typical works are as follows:
WGAN-GP~\cite{arjovsky2017wasserstein} further introduce the gradient penalty to achieve the Lipschitz condition. SNGAN~\cite{miyato2018spectral} proposes Spectral Normalization to achieve Lipschitz condition for training GANs. More recent works~\cite{dong2019margingan,NIPS2019_8560,cao2019multi} design task-specific loss functions for GANs. Meanwhile, some works aim to improve the diversity of the GANs, for instance, VEEGAN~\cite{srivastava2017veegan} and CVAE-GAN~\cite{bao2017cvae} proposes to reconstruct the images in the pixel level to avoid the missing modes problems. MS-GAN~\cite{mao2019mode} proposes a loss term that maximizes the ratio of the distance between generated images with respect to the corresponding latent codes. Besides new loss functions, there are also some works~\cite{radford2015unsupervised, karras2017progressive, zhang2018self, karras2019style} focus on using different network structures and training strategies for training. In contrast to these methods, our method proposes to build explicit real data prior to solving these problems.
\noindent \textbf{Build Prior for Data Distribution}.
There are various methods that aim to build the prior for underlying data distribution. For example, Principle Component Analysis (PCA) ~\cite{turk1991face}, Independent Component Analysis (ICA)~\cite{hyvarinen2000independent}, and the Gaussian Mixture Model (GMM)~\cite{xu1996convergence, richardson2018gans}, all assume the data distribution follow a simple assumption. But they have difficulty modeling complex patterns of irregular distributions. Later works, such as the Hidden Markov Model (HMM)~\cite{starner1997real}, Markov Random Field (MRF)~\cite{ranzato2010generating}, and restricted Boltzmann machines (RBMs)~\cite{salakhutdinov2009deep}. limiting their results on texture patches, digital numbers, due to a lack of effective feature representations. Instead of building the model directly on the pixel space, we leverage the successful feature presentation of the CNNs~\cite{He_2016_CVPR,szegedy2016rethinking} and build the model on these feature representations.
\vspace{-0.3cm}
\section{Methods}
Before we introduce our methods, we first give a brief introduction and our findings about current GANs. In the vanilla GAN, the discriminator D and the generator G play the following two-player minimax game with value function $V(D, G)$:
\begin{equation}
\label{eq:minimaxgame-definition}
\min_{G} \max_D V(D, G) = \mathbb{E}_{x \sim p_{\text{r}}(x)}[\log D(x)] + \mathbb{E}_{x \sim G(z)}[\log (1 - D(x))].
\end{equation}
suppose $p_g(x) = G(z)$, for G fixed, the optimal discriminator $D$ for $V(D, G)$ is:
\begin{equation}
\label{eq:optimal-D}
D^*_G(\bm{x}) = \frac{p_\text{r}(x)}{p_\text{r}(x) + p_g(x)}
\end{equation}
According to the analysis in GAN~\cite{goodfellow2014generative}, the global minimum of $V(D^*_G, G)$ is achieved if and only if $p_g = p_r$. But in the practical training process of GAN, the optimal discriminator $D_G^*(x)$ is usually hard to achieve due to insufficient training iterations or limited network capability. In this situation, the gradient is probably incorrect for reaching the optimal $p_g$.
To demonstrate this, we train a basic GAN~\cite{goodfellow2014generative} on two typical toy distributions. As illustrated in Figure~\ref{fig:toy_vis}, The green points are the real data, the blue points are the generated data, the discriminator is trained based on current real and generated data distribution. The red arrows denote the gradient of the discriminator for each generated data. In Figure~\ref{fig:toy_vis}(a), some generated data are out of the real data distribution. But the gradients of these low quality generated points are not correct and they can't move to the real data distribution. So these samples stay in the situation of \emph{low quality} problem. In Figure~\ref{fig:toy_vis}(c), we notice that the real data distribution has two main regions, but the gradient of the generated points have the same direction to one region of the real data. This seems to be an intrinsic cause of the \emph{missing modes} problem.
\begin{figure*}[t]
\centering
\includegraphics[width=1.0\columnwidth]{vis.pdf}
\vspace{-0.3cm}
\caption{An example of \emph{low quality} problem(a,b) and \emph{missing modes} problem(c,d) in GANs training on toy examples, (a) and (c) show the gradient of vanilla GAN, (b) and (d) show the gradient of our proposed PriorGAN.}
\label{fig:toy_vis}
\vspace{-0.3cm}
\end{figure*}
To solve these problems, we propose to establish a prior model on real data to help the discriminator solve the inaccurate gradient problem. Equipped with the prior, we can solve the problem of \emph{low quality} and \emph{missing modes}.
In the next section, we will first introduce how to build real data prior. Then we present how to apply the prior as a loss term in GAN training to improve the quality of generated images. After that, we show how to use this prior to estimate the diversity of the real and generated data distribution. Finally, we introduce the resampling strategy to improve the diversity of the generated samples.
\subsection{Build Prior for Real Data}
Building an explicit prior to the high dimensional data is very difficult and challenging, especially for high-resolution images. This is because the real data distribution is extremely complicated to be captured by a simple prior. Previous method~\cite{richardson2018gans} try to leverage the Gaussian Mixture Model(GMM) model to build a model on the image patch level. However, the image resolution is restricted to be very small and the generated results are really blurry. Instead of directly building the GMM model on the pixel level, we take full advantage of the recent success of deep convolutional networks and build the GMM model on the features extracted by the networks. The deep feature is suitable for building the GMM model from two aspects. First, the deep feature is relatively low dimensional, which helps avoid the overfitting problem. Second, the deep feature is very representative and explicitly represents the content of the image. Denote the input image as $x$, the deep feature extractor function as $F(\cdot)$. We use a pre-trained network as feature extractor and the extracted feature dimension is $d$. Suppose the GMM model has $M$ Gaussian components, the mean vector and covariance matrix for $i$-th Gaussian component are $\mu_i$ and $\Sigma_i$, respectively. Given these parameters, the density function of a Gaussian Mixture Model takes the form
\begin{equation}
p(x|\lambda) = \sum_{i=1}^{M} w_i \mathcal{N}(F(x)|\mu_i, \Sigma_i), x \in P_r
\label{eqn:GMM_probability}
\end{equation}
$P_r$ denotes real image space, $w_i$ is the mixture weights, which satisfy the constraint that $\sum_{i=1}^{M} w_i = 1$. And $\mathcal{N}(F(x)|\mu_i, \Sigma_i)$ is a $d$-variate Gaussian function of the form. The complete Gaussian mixture model is parameterized by the mean vectors, covariance matrices and mixture weights from all component densities. These parameters are collectively represented by the notation, $\lambda = \{w_i, \mu_i, \Sigma_i\}$. To estimate these parameters, we adopt the expectation-maximization (EM) algorithm~\cite{dempster1977maximum} to iteratively update them with the loss function
\begin{equation}
L= \sum_{x \in P_r} -\log p(x|\lambda).
\vspace{-0.2cm}
\end{equation}
\subsection{Improving Quality with Prior}
Given the estimated GMM model, for a generated image $x \in P_g$, we can adopt $p(\mathbf{x}|\lambda)$ in Equation.~\ref{eqn:GMM_probability} to estimate its probability. If its probability is high, it indicates that it is close to some of the real samples and the quality is high, otherwise, it is far from real samples and the quality is poor. So the generator should pay more attention to these poor quality samples. However, previous analysis shows that these poor quality generated samples may get inaccurate gradient from the discriminator. To solve this problem, we propose to utilize the prior to restrict these poor quality samples to get a higher probability in the built GMM model. This will help the poor quality samples to have an accurate gradient direction towards the real data distribution. Formally, given an generated image $x = G(z)$, $G$ is the generator, the loss for the generator is:
\begin{equation}
\label{eqn:quality_score}
\mathcal{L}_{Q} =
\begin{cases}
-\log p(F(G(z))| \lambda) & \textit{if } p(F(G(z))| \lambda) < \theta\\
0 & \textit{otherwise}
\end{cases},
\end{equation}
where $\theta$ is the quality threshold. The quality loss $\mathcal{L}_{Q}$ mainly focuses on the low quality generated results and will give a right gradient direction for these low quality generated samples, as illustrated in Figure~\ref{fig:toy_vis}(b). During the training process of GANs, we give a loss weight $\delta$ to this quality loss and apply it as an auxiliary term to update the generator besides the GAN loss.
\subsection{Improving Diversity with Prior}
Current GANs often suffer from the \emph{missing modes} problem, the learned distribution from GANs may miss some certain regions of the real data distribution. Therefore the key is how we can estimate the diversity of real data distribution and find these missing regions in the learned distribution. One possible solution is to apply the built GMM model. Specifically, if we establish a GMM model to represent a real data distribution, then we can apply the $M$ Gaussian components in the GMM model to divide the real data distribution into $M$ groups and estimate the frequency of the real samples in each group. A sample belonging to which Gaussian component is decided by the minimum distance to the center of each Gaussian component, as we shown in Figure~\ref{fig:prior}. Suppose the number of the real samples on the $i$-th Gaussian component is $n_i^r$, then the frequency of the real samples on the $i$-th Gaussian component is $f_i^r = {n_i^r}/{\sum_i^M{n_i^r}}$. And we have the vector $[f_1^r, f_2^r,\cdots, f_M^r]$ to represent the diversity of the real samples. Similarly, we apply the same GMM model to get the frequency of generated samples, which is denoted as $[f_1^g, f_2^g,\cdots, f_M^g]$.
\begin{figure*}[t]
\centering
\includegraphics[width=1.0\columnwidth]{split.pdf}
\vspace{-0.6cm}
\caption{Visualization of learned GMM-based prior on FFHQ and LSUN-cat dataset, each column are images randomly selected from the same Gaussian component. We can observe that images within the same Gaussian component share the same attributes, like hat, hair, sunglasses for the FFHQ or eyes, color, posture for the LSUN-cat.}
\label{fig:prior}
\vspace{-0.2cm}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=1.0\columnwidth]{resample.pdf}
\vspace{-0.7cm}
\caption{illustration of the pipeline of our resampling strategy.}
\label{fig:resample}
\vspace{-0.4cm}
\end{figure*}
The frequency difference between the real and generated samples can indicate whether the missing modes problem occurs. If the learned distribution misses a certain region $S$ of the real data distribution, and $S$ belongs to the $i$-th Gaussian component of the built GMM model, then $i$-th learned data frequency $f_i^g$ will be smaller than the real data frequency $f_i^r$. So the diversity difference between real data distribution and learned distribution can be represented by the frequency distances. Suppose the diversity distance between real data distribution $p_r$ and learned distribution $p_g$ is denoted as
\begin{equation}
d(p_r,p_g) = [f_1^r, f_2^r,\cdots, f_M^r] - [f_1^g, f_2^g,\cdots, f_M^g].
\end{equation}
So $d(p_r, p_g)$ is a vector with $M$ elements, the $i$-th element is $d_i(p_r,p_g)$. So the overall diversity differences score (DDS) is
\begin{equation}
DDS(p_r, p_g) = \sum_i^M |d_i(p_r,p_g)|.
\end{equation}
An ideal learned distribution should satisfy $DDS(p_r, p_g)=0$. However, the $DDS$ is non-differentiable and cannot be directly used as a loss function. So we optimize it in an indirect way and proposed a resampling strategy. Considering a typical missing modes phenomenon where the real data distribution has two regions $S_1$ and $S_2$, the generated samples only lies around the $S_1$, current discriminator can't force the generated samples around $S_1$ to move to $S_2$ since the discriminator consider the generated samples around $S_1$ are "real enough". To overcome this problem, we propose to change the decision boundary of the discriminator. If we sample fewer samples in $S_1$ and more samples in $S_2$ for training the discriminator, the decision boundary of the discriminator will be changed. The discriminator will encourage the generated samples to move to $S_2$ and solve the \emph{missing modes} problem, as illustrated in Figure~\ref{fig:toy_vis}(d). We call this process the resampling strategy. This strategy will change the real data distribution in the training procedure to help the generator overcome the missing modes problem. Formally, we use a new sampling frequency $f^{r'}_i$ to sample the real samples in the $i$-th Gaussian component for training. $f^{r'}_i$ is decided by the original frequency and the diversity distances:
\begin{equation}
f^{r'}_i = max(f_{i}^{r} + \alpha d_i(p_r,p_g), 0),
\end{equation}
Then we normalized the sampling frequency, $f^{r'}_i = {f^{r'}_i}/ {\sum_i^M f^{r'}_i}$. Such that the new sampling frequency is between $0$ and $1$, $\alpha$ is a hyper parameter which decides the resample weight. For $f_g^i < f_r^i$, we increase the sampling frequency of region $i$ and decrease it if otherwise. The resampling strategy is also illustrated in Figure~\ref{fig:resample}.
\iffalse
In high dimensions, it's not easy to get $p_r(x)$, $p_g(x)$ for a given sample $x$. so we use our estimated GMM model to get the approximate number. In details, if we split the real data distribution to $M$ subspace, calculate the frequency of real data according to the identity of nearest Gaussian Kernel, we denote the frequency as $\mathbf{F_{r}_i}, i={1,2,..M}$, and the frequency for generated data as $\mathbf{F_{g}_i}, i={1,2,..M}$, these can be seen as
the summation of $p_r$ and $p_g$ in each subspace, and it satisfied:
\begin{equation}
\sum_{\mathbf{i}=1}^\mathbf{M} \mathbf{F_{r}_i} = \sum_{\mathbf{i}=1}^\mathbf{M} \mathbf{F_{g}_i} = 1
\end{equation}
we denote the difference between these two frequency distribution as $\mathbf{D(r,g)}$:
\begin{equation}
\mathbf{D(r,g)} = \sum_{\mathbf{i}=1}^\mathbf{M} |\mathbf{F_{r}_i}-\mathbf{F_{g}_i}|
\end{equation}
We named it DIV-score. An ideal generated distribution should satisfy $\mathbf{D(r,g)}=0$. And a low DIV-score is a necessary condition which indicates that generated distribution fits the real data distribution.
But in practical application, the DIV-score might not be very small to some GAN models, this usually indicates modes missing. To narrow the gap between $p_r$ and $p_g$, instead of sampling from real data distribution directly, we aim to balance the sampling procedure to remedy the missing mode.
Specifically, we assume the default sampling frequency for $\mathbf{i}$-th Gaussian component is $\mathbf{F_r_i}$, we update it to $\mathbf{F_n_i}$:
\begin{equation}
\mathbf{F_{n}^{i}} = \mathbf{F_{r}^{i}} + (\mathbf{F_{r}^{i}} - \mathbf{F_{g}^{i}}) * K, K>0
\end{equation}
K is a hyper parameter which decides the resample weight. For $F_g^i < F_r^i$, we increase the sampling frequency of region $i$ and decrease it if $F_g^i > F_r^i$.
\fi
Above we introduce our methods for improving quality and diversity with our proposed prior. This prior can be integrated into various frameworks of GANs, such as WGAN-GP~\cite{gulrajani2017improved}, LSGAN~\cite{mao2017least}, SNGAN~\cite{miyato2018spectral}, and even the StyleGAN~\cite{karras2019style}. This flexible combination with existing methods can help these methods get further improvement.
\section{Experiments}
In this part, we conduct experiments on a large variety of generative models trained on different datasets including CIFAR-10~\cite{krizhevsky2009learning}, LSUN~\cite{yu2015lsun}, and FFHQ~\cite{karras2019style} datasets to evaluate the performance of our proposed methods.
\textbf{Datasets.}
CIFAR-10 is a widely used dataset for image generation. We directly apply the original image size $32 \times 32$ for training. LSUN~\cite{yu2015lsun} is widely used for high-resolution image generation, we choose two categories in LSUN for training and evaluation: cat and bird, which we call LSUN-cat and LSUN-bird, respectively. LSUN-cat contains $1.6$ million images and LSUN-bird contains $2.2$ million images, we follow the settings in StyleGAN~\cite{karras2019style} to resize all the images to $256 \times 256$ resolution. FFHQ~\cite{karras2019style} is a recently high-resolution face dataset contains $70000$ images, we resize them to $256 \times 256$ resolution for training, which we call FFHQ-256.
\textbf{Evaluations.}
To evaluate the quality of an image set, as we mentioned earlier, the probability of each image in our GMM model can indicate its quality. So we use the mean of normalized image density in Equation~\ref{eqn:GMM_probability} to evaluate the quality of an image set, calculated as: $ \frac{1}{K} \sum_{\mathbf{i}=1}^K norm( log ( p(x_i|\lambda)))$, $K$ is the number of images in the image set\footnote[1]{More implementation details please refer to supplementary material.}. We named it Quality Score(QS), higher QS indicates better image quality. And we use our proposed DDS to evaluate the diversity of the generative model, FID~\cite{heusel2017gans} to measure both quality and diversity.
For a fair comparison, we would use $5k$ or $50k$ genearted images to calculate the FID, denoted as FID(5K) and FID(50K) respectively. $10k$ images are used to calculate QS and DDS for the CIFAR-10 dataset. For the FFHQ, LSUN-cat, and LSUN-bird datasets, we use $50k$ images to calculate the result of the FID, QS and DDS.
\textbf{Implementation Details.}
For each baseline, we follow training settings in their experiments. We only add quality loss and the resampling strategy during the training of each baseline. In order to build the prior for each dataset, we apply ResNet101~\cite{He_2016_CVPR} or InceptionV3~\cite{szegedy2016rethinking} to extract the features for all the real images. The features we use are from the last average pooling layer. Without special notes, we use ResNet101 to extract the features.
Then we build the prior for each dataset. Specifically, we set the number of Gaussian component to $35$,$70$, $7$, and $7$, the resample weight $\alpha$ to $3$,$7$,$7$,$7$ for the CIFAR-10, FFHQ, LSUN-cat, and LSUN-bird, respectively, and the quality loss weight $\delta$ set to $0.1$. More details could be found in the supplementary material.
\begin{table}[t]
\centering
\small{
\begin{tabular}[t]{c|ccc}
\hline
Method & Inception score $\uparrow$ & FID(5k)$\downarrow$ & FID(50k)$\downarrow$ \\
\shline
Real data & 11.24$\pm$.12 & 8.63$\pm$.0 & - \\
\hline
DCGAN~\cite{radford2015unsupervised}& 6.64$\pm$.14 & - & - \\
LR-GANs~\cite{yang2017lr} & 7.17$\pm$.07 & - & - \\
Warde-Farley et al.~\cite{warde2016improving} & 7.72$\pm$.13 & - & - \\
WGAN-GP~\cite{gulrajani2017improved} & 7.86$\pm$.08 & 40.2$\pm$.0 & - \\
LGAN~\cite{zhou2019lipschitz} & 8.03$\pm$.03 & - & 15.64$\pm$.07 \\
SNGAN~\cite{miyato2018spectral} & 8.22$\pm$.05 & 21.7$\pm$.21 & 13.78$\pm$.11 \\
$L^{10}$ BWGAN~\cite{adler2018banach} & 8.31$\pm$.07 & - & - \\
QAGAN~\cite{NIPS2019_8560} & 8.37$\pm$.04 & - & 13.91$\pm$.11 \\
WGAN-ALP~\cite{terjek2019virtual} & 8.34$\pm$.06 & - & 12.96$\pm$.35 \\
PriorSNGAN(\textbf{ours}) & \textbf{8.60$\pm$.19} & \textbf{18.19$\pm$.06} & \textbf{ 11.94$\pm$.18} \\
\shline
\end{tabular}
\vspace{0.1cm}
\caption{\label{tab:incepscores}Inception scores and FIDs with unconditional image generation on CIFAR-10. FID$(50k)$ means calculating the score with $50k$ generated images, otherwise with $5k$ generated images.
}
\vspace{-0.5cm}
}
\end{table}
\begin{comment}
\begin{table}[t]
\centering
\small{
\begin{tabular}[t]{lcc}
\toprule
Method & Inception score $\uparrow$ & FID(50k)$\downarrow$ \\
\midrule
Real data & 11.24$\pm$.12 & \\
\midrule
DCGAN~\cite{radford2015unsupervised}& 6.64$\pm$.14 & \\
LR-GANs~\cite{yang2017lr} & 7.17$\pm$.07\\
Warde-Farley et al.~\cite{warde2016improving} & 7.72$\pm$.13 & \\
WGAN-GP~\cite{gulrajani2017improved} & 7.86$\pm$.08 & 19.65 \\
LGAN~\cite{zhou2019lipschitz} & 8.03$\pm$.03 & 15.64$\pm$.07 \\
SNGAN~\cite{miyato2018spectral} & 8.22$\pm$.05 & 13.78$\pm$.11 \\
$L^{10}$ BWGAN~\cite{adler2018banach} & 8.31$\pm$.07 & - \\
WGAN-ALP~\cite{terjek2019virtual} & 8.34$\pm$.06 & 12.96$\pm$.35 \\
PriorSNGAN(\textbf{ours}) & \textbf{8.60$\pm$.19} & \textbf{ 11.94$\pm$.18} \\
\bottomrule
\end{tabular}
\vspace{0.1cm}
\caption{\label{tab:incepscores}Inception scores and FIDs with unconditional image generation on CIFAR-10. FID$(50k)$ means calculating the score with $50k$ images otherwise with $5k$ generated images.
}
}
\end{table}
\end{comment}
\subsection{Comparison with state-of-the-art Methods}
In this section, we compare our results with other state-of-the-art methods. We conduct the experiments on the CIFAR-10 dataset. The basic framework is an unconditional GAN, the generator and discriminator network structure are the same as the ResNet structure adopted in the baseline SNGAN~\cite{miyato2018spectral}. We use $5k$ randomly generated images to calculate the Inception score ten times and report the mean and variance. For the FID, there are two different settings to calculate the result, some methods~\cite{miyato2018spectral, heusel2017gans} use $5k$ generated images and $10k$ real images to calculate, while the others~\cite{parimala2019quality,terjek2019virtual} use $50k$ generated and $50k$ real images to calculate. Using $50k$ generated images can get a better result. For a fair comparison, we report both results and denote them as the FID($5k$) and the FID($50k$). As shown in Table~\ref{tab:incepscores}, our method (PriorSNGAN) outperforms the baseline SNGAN and other state-of-the-art GANs. These results confirm the effectiveness of our proposed real data prior to GAN training.
\begin{table}[t]
\centering
\begin{tabular}{c|ccc|ccc|ccc}
\hline
\multirow{2}{*}{ } &
\multicolumn{3}{c|}{SNGAN } &
\multicolumn{3}{c|}{WGAN-GP } &
\multicolumn{3}{c}{LSGAN } \\
\cline{2-10}
& FID$\downarrow$ & QS$\uparrow$ & DDS$\downarrow$ & FID$\downarrow$& QS$\uparrow$ & DDS$\downarrow$ & FID$\downarrow$ & QS$\uparrow$ & DDS$\downarrow$ \\
\shline
Baseline & 13.78 & 0.429 & 0.488 & 19.65 & 0.410 & 0.532 & 34.90 & 0.367 & 0.545 \\
+Prior & 11.94 & 0.512 & 0.380 & 18.16 & 0.439 & 0.475 & 30.24 & 0.406 & 0.487 \\
\shline
\end{tabular}
\vspace{0.14cm}
\caption{Comparison of FID, QS, and DDS metrics on three different GANs: SNGAN, WGAN-GP and LSGAN.}
\label{table:model_score}
\vspace{-0.5cm}
\end{table}
\subsection{Prior for Various GANs}
Our approach enjoys a high degree of flexibility and can be integrated into various kinds of GAN frameworks. We choose some variants of GAN as baseline frameworks, such as LSGAN~\cite{mao2017least}, WGAN-GP~\cite{gulrajani2017improved}, SNGAN~\cite{miyato2018spectral}, and then integrate our method into these frameworks for comparison. For a fair comparison, we follow the same settings and architecture of GANs and only apply our quality loss and the resampling strategy, which are denoted as +Prior.
We conducted experiments on the CIFAR-10 dataset, Table~\ref{table:model_score} reports a comparison of our method with baseline models. We can observe that our methods outperform the corresponding baseline GAN frameworks in terms of FID, QS, and DDS metrics, which demonstrate the superiority of our proposed framework.
\begin{figure*}[t]
\centering
\includegraphics[width=1.0\columnwidth]{visualization_results.pdf}
\vspace{-0.4cm}
\caption{Visualization results of our generated images on FFHQ-256, LSUN-cat and LSUN-bird datasets, our method can generate not only high quality but also diverse images.}
\label{fig:vis_results}
\vspace{-0.1cm}
\end{figure*}
\begin{table}[t]
\centering
\begin{tabular}{c|ccc|ccc|ccc}
\hline
\multirow{2}{*}{ } &
\multicolumn{3}{c|}{FFHQ-256 } &
\multicolumn{3}{c|}{LSUN-cat } &
\multicolumn{3}{c}{LSUN-bird } \\%&
\cline{2-10}
& FID$\downarrow$ & QS$\uparrow$ & DDS$\downarrow$ & FID$\downarrow$& QS$\uparrow$ & DDS$\downarrow$ & FID$\downarrow$ & QS$\uparrow$ & DDS$\downarrow$ \\
\shline
StyleGAN & 11.69 & 0.703 & 0.250 & 9.66 & 0.486 & 0.203 & 10.26 & 0.513 & 0.136 \\
+Prior(R101) & 8.79 & 0.724 & 0.199 & 8.25 & 0.526 & 0.162 & 9.04 & 0.530 & 0.109 \\
+Prior(InV3)& 7.12 & 0.749 & 0.170 & 7.21 & 0.553 & 0.151 & 8.44 & 0.554 & 0.086 \\
\shline
\end{tabular}
\vspace{0.2cm}
\caption{Comparison of FID, QS, and DDS metrics on three different datasets: FFHQ-256, LSUN-cat, and LSUN-bird.}
\label{table:stylegan_score}
\vspace{-0.2cm}
\end{table}
\begin{table}[t]
\centering
\begin{tabular}{c|ccc|ccc|ccc
\hline
\multirow{2}{*}{ } &
\multicolumn{3}{c|}{FFHQ-256 } &
\multicolumn{3}{c|}{LSUN-cat } &
\multicolumn{3}{c}{LSUN-bird } \\%&
\cline{2-10}
& FID$\downarrow$ & QS$\uparrow$ & DDS$\downarrow$ & FID$\downarrow$& QS$\uparrow$ & DDS$\downarrow$ & FID$\downarrow$ & QS$\uparrow$ & DDS$\downarrow$ \\
\shline
StyleGAN & 11.69 & 0.703 & 0.250 & 9.66 & 0.486 & 0.203 & 10.26 & 0.513 & 0.136 \\% & 4.44 & 0.658 & 0.582 \\
StyleGAN +$\mathcal{L}_Q$ & 9.75 & 0.732 & 0.229 & 8.85 & 0.531 & 0.188 & 9.17 & 0.534 & 0.123 \\
StyleGAN +RS & 10.00 & 0.715 & 0.196 & 9.11 & 0.505 & 0.149 & 9.79 & 0.521 & 0.095 \\% & 4.33 & 0.660 & 0.584 \\
\shline
\end{tabular}
\vspace{0.14cm}
\caption{Ablation study of our proposed prior with only quality loss and only resampling strategy.}
\label{table:ablation_study}
\vspace{-0.25cm}
\end{table}
\subsection{Prior for Various Datasets}
We also try to apply our proposed prior for various datasets, such as FFHQ-256, LSUN-cat and LSUN-bird. StyleGAN~\cite{karras2019style} recently achieve great success for high-resolution image generation. In this section, we try to combine our method with StyleGAN and evaluate the results on these datasets. We apply two different pre-trained models, ResNet101~\cite{He_2016_CVPR} and InceptionV3~\cite{szegedy2016rethinking} to extract deep features for building the GMM model. We call them +Prior(R101) and +Prior(InV3), respectively. The quantitative results are shown in Table~\ref{table:stylegan_score}. We can observe that our approach can effectively improve the quality(QS), diversity(DDS), and the overall score(FID) over the strong baseline StyleGAN by large margins on all datasets. And the real data prior build on the feature extracted with InceptionV3 model achieves better results than ResNet101, one possible reason is that the evaluation methods are also based on InceptionV3 model. Some visualization results are shown in Figure~\ref{fig:vis_results}.
\begin{table}[h!]
\begin{minipage}{\columnwidth}
\begin{minipage}[t]{0.3\columnwidth}
\small
\centering
\begin{tabular}{c|ccc}
\hline
$M$ & FID$\downarrow$ & QS$\uparrow$ & DDS$\downarrow$ \\
\shline
5 & 7.26 & 0.548 & 0.156 \\
7 & 7.21 & 0.553 & 0.151 \\
10 & 7.32 & 0.552 & 0.159 \\
20 & 7.36 & 0.549 & 0.163 \\
50 & 7.45 & 0.543 & 0.177 \\
\shline
\end{tabular}
\vspace{0.08cm}
\makeatletter\def\@captype{table}\makeatother\caption{PriorGAN results on LSUN-cat dataset with different number of Gaussian components $M$.}
\label{table:kernel-densities}
\vspace{-0.3cm}
\end{minipage}
\quad
\begin{minipage}[t]{0.3\columnwidth}
\small
\centering
\begin{tabular}{c|ccc}
\hline
$\delta$ & FID$\downarrow$ & QS$\uparrow$ & DDS$\downarrow$ \\
\shline
0.03 & 8.31 & 0.509 & 0.191 \\
0.07 & 7.82 & 0.529 & 0.201 \\
0.10 & 7.66 & 0.550 & 0.195 \\
0.15 & 8.11 & 0.578 & 0.220 \\
0.2 & 8.29 & 0.586 & 0.227 \\
\shline
\end{tabular}
\vspace{0.08cm}
\makeatletter\def\@captype{table}\makeatother\caption{PriorGAN results on LSUN-cat dataset with different quality loss weight $\delta$.}
\label{table:quality-loss-weight}
\vspace{-0.3cm}
\end{minipage}
\quad
\begin{minipage}[t]{0.3\columnwidth}
\small
\centering
\begin{tabular}{c|ccc}
\hline
$\alpha$ & FID$\downarrow$ & QS$\uparrow$ & DDS$\downarrow$ \\
\shline
1 & 8.76 & 0.499 & 0.177 \\
3 & 8.35 & 0.506 & 0.162 \\
5 & 8.30 & 0.514 & 0.143 \\
7 & 8.08 & 0.521 & 0.149 \\
9 & 8.16 & 0.515 & 0.145 \\
\shline
\end{tabular}
\vspace{0.2cm}
\makeatletter\def\@captype{table}\makeatother\caption{PriorGAN results on LSUN-cat dataset with different resample weight $\alpha$. }
\label{table:resample-weight}
\vspace{-0.3cm}
\end{minipage}
\end{minipage}
\end{table}
\subsection{Ablation Study}
To validate the effectiveness of our proposed two key methods: the quality loss and resampling strategy, we apply additional experiments on FFHQ-256, LSUN-cat, LSUN-bird dataset. The quality loss $\mathcal{L}_Q$ and resampling strategy are applied independently to StyleGAN. The resulting models are called StyleGAN +$\mathcal{L}_Q$ and StyleGAN +RS. The results are shown in Table~\ref{table:ablation_study}. We can observe that adding the loss function $\mathcal{L}_Q$ mainly improve QS, which indicates the loss function $\mathcal{L}_Q$ mainly solve the \emph{low quality} problem. Adding the resampling strategy can get a lower DDS, which indicates that the resampling strategy mainly solves the \emph{missing modes} problem. This further validates the effectiveness of our proposed methods.
\subsection{Hyper Parameter Analysis}
In this section, we conduct experiments to investigate the sensitiveness of the number of Gaussian components $M$, quality loss weight $\delta$, and resample weight $\alpha$ in our proposed method. All the experiments are applied to the LSUN-cat dataset.
To reduce the computation cost, we apply PCA on the feature extracted from InceptionV3 model and keep 98\% of the total variance by default. First, we explore how the number of Gaussian components $M$ affects the results. We set $M$ to 5,7,10,20,50 and test the results on LSUN-cat dataset, we show the results in Table~\ref{table:kernel-densities}. The results are not very sensitive to $M$ and we achieve the best performance when setting it to 7. Besides, we set $M$ to 7 and explore how quality loss weight $\delta$ and resample weight $\alpha$ affects the performance, the results are shown in Table ~\ref{table:quality-loss-weight},~\ref{table:resample-weight}.
\section{Conclusion}
In this paper, we analyze current GANs suffer from low quality and missing modes problem. To tackle this problem, we propose a novel method which builds a GMM prior on deep feature space, by using this GMM prior, we propose quality loss to increase generated image quality and resampling strategy to enrich the diversity. Experiments have shown our method can easily insert into any GAN baseline and outperforms state-of-the-art by large margins.
\section*{Broader Impact}
Our PriorGAN is a general method for image generation. From the positive aspect, many generation tasks like image-to-image translation, image editing will benefit from it. We hope our work will serve as a solid baseline and can ease future research in generative models. From the negative aspect, our method can also be applied to some image manipulation techniques, which may cause severe trust issues and security concerns in our society.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
Eclipsing binary systems provide the ability to determine absolute fundamental parameters of stellar components, calibrate stellar evolution models, measure the orientation and dynamical changes in orbits, and derive distances without the need for calibration. For benchmark-grade eclipsing binaries, a precision in fundamental stellar parameters (masses and radii) of a few percent can be obtained \citep{torres2010}. All these scientific goals can only be achieved through solving the inverse problem: models are created to simulate synthetic observables given a set of input parameters, and the input parameters can then be varied to determine which values result in synthetic observables that best represent the observations. The parameter space, however, is complex and often not smooth, making this task expensive and time-consuming.
PHOEBE is one of many modeling codes which operates by discretizing the gravitationally distorted surface of each stellar component, populating the discretized grid with local quantities from atmosphere tables, and integrating over visible elements at any given time in the orbit to synthesize model observables. Until now, all releases since the 2.0 release \citep{prsa2016, horvat2018, jones2020} have been focused on adding advanced physics and precision to the forward-model and leaving the inverse problem to the user.
Due to the complexity of the parameter space, different combinations of observations available, and the wide-range of system morphologies and relevant physics effects, each individual system often requires a manual treatment in order to obtain the most accurate and precise model and uncertainties possible. Furthermore, it is often necessary to utilize multiple algorithms in order to increase the chance of determining the optimal solution and produce reliable and realistic uncertainties on the modeled parameters. Having to ``wrap'' several algorithms, each with its own interface, around the forward-model can present a steep learning curve and effort-overhead to solving these systems. To address this, PHOEBE now includes a common interface to interact with a variety of algorithms, some off-the-shelf and others custom built, that are useful for solving the inverse problem.
To complicate matters, the available forward-model codes each have their own interfaces and often different parameterization, making it even more difficult to compare the forward-models or choose the code most appropriate for a given scenario. PHOEBE 2 has been designed and developed with precision, robustness, and flexibility as the main priorities -- but this comes at the cost of computational efficiency. In many cases, other codes may be just as capable of modeling a given system but at a fraction of the time cost. Learning the intricacies of each of these codes presents yet another learning curve and overhead. PHOEBE now incorporates several other publicly available codes as the forward-model, converting parameterization and units for both inputs and outputs as necessary. This brings several large benefits, including the abilities to: easily compare the synthetic models created by different codes, choose the code most appropriate and efficient for a specific system, and transition to more expensive but precise models throughout solving the inverse problem.
Here we outline a general framework for solving the inverse problem with PHOEBE. We divide the available solver algorithms incorporated in this PHOEBE release into three categories: ``estimators'', ``optimizers'', and ``samplers''. Section \ref{sec:estimators} details the ``estimators'' now available in PHOEBE that estimate system parameters from observations alone, without the need for a forward-model. These estimators are designed to cheaply determine the parameter space of a particular system given any of the data available. Section \ref{sec:merit_function} then discusses the merit function from the available forward-models. This merit function is then used for both ``optimizers'' (Section \ref{sec:optimizers}) and ``samplers'' (Section \ref{sec:samplers}) to find the optimal solution and determine robust posteriors (Section \ref{sec:posteriors}). Section \ref{sec:ui} then presents the new user interface introduced in this release, allowing easy access to most of the inverse-problem functionality from either a web browser or a dedicated desktop client. See Appendix \ref{app:symbols} for descriptions of the symbols used for system parameters throughout the paper.
This paper is accompanied by the 2.3 release of PHOEBE, available from \url{http://phoebe-project.org} and \url{https://github.com/phoebe-project/phoebe2}. Scripts to reproduce all figures in this paper are available at \url{http://phoebe-project.org/publications/2020Conroy+}.
\section{Estimators}\label{sec:estimators}
It is often useful and computationally prudent to estimate the values of as many parameters as possible without having to compute multiple physical models. The 2.3 release of PHOEBE incorporates several algorithms which act directly on the observations themselves, with the goal to provide fast initial solutions for a subset of the relevant system parameters without significant user supervision, including:
\begin{description}
\item[RV periodogram, LC periodogram] computes the periodogram for light curves or radial velocity curves, respectively, and proposes an orbital period, $P_\mathrm{orb}$;
\item[RV geometry] fits a simple Keplerian orbit to propose values for: $t_{0, \mathrm{supconj}}$, $e$, $\omega_0$, $v_\gamma$, and either $q$ and the projected orbital semi-major axis, $a_\mathrm{orb} \sin i$, or the projected per-component semi-major axes, $a_\mathrm{comp} \sin i$ ;
\item[LC geometry] estimates phases of eclipse mimima, ingress, and egress and uses that information to propose values for: $t_{0, \mathrm{supconj}}$, $e$, and $\omega_0$; and
\item[EBAI] uses a trained artificial neural network on the input light curves to propose values for: $t_{0, \mathrm{supconj}}$, $T_\mathrm{eff, 2} / T_\mathrm{eff, 1}$, $(R_\mathrm{equiv, 1} + R_\mathrm{equiv, 2}) / a_\textrm{orb}$, $e \sin \omega_0$, $e \cos \omega_0$, and $i$.
\end{description}
For all estimators that work in phase-space (as opposed to time-space), the input data can optionally be binned to maintain the efficiency of the estimators on large data sets.
\subsection{Periodograms}
The most common observational data sets used when modeling eclipsing star systems are light curves and radial velocity curves. For the majority of eclipsing systems (those that do not exhibit significant time-dependent changes), these data can be phase-folded on the orbital period of the system. If the period is well known and the orbital parameters do not change between successive orbits, then the system can be sampled in phase-space instead.
PHOEBE 2.3 contains wrappers around two common periodograms -- Box Least Squares (BLS, for light curves) and Lomb Scargle (for light curves or radial velocities) -- as implemented by \texttt{astropy.timeseries}\footnote{requires astropy v3.2 or later.}\citep{2013A&A...558A..33A}. The wrapper takes any number of light curves or radial velocities as input, in addition to several of the advanced options from astropy, including: automatic or manual sampling in period/frequency, the option to change the objective function for BLS, and the ability to set the proposed eclipse durations for BLS. If running on multiple light curves, each light curve is normalized by dividing by either the median or maximum flux value before sending to the periodogram algorithm. Radial velocities across all requested data sets are combined and then normalized by the absolute maximum value for the primary and secondary star independently, with the secondary then mirrored. After running the wrapper, the periodogram itself is returned and can be plotted, and the peak-period is proposed for adoption. Especially for near-circular systems with similar eclipses, these algorithms often find a strong signal at aliases of the true orbital period, so an easy interface is provided for adopting any factor of the proposed peak period.
\subsection{Radial Velocity Geometry}\label{sec:rv_geometry}
Analytical radial velocities (RVs) are used as an estimator for the eccentricity ($e$), argument of periastron ($\omega$), systemic velocity ($v_{\gamma}$), and the time of superior conjunction ($t_{0, \mathrm{supconj}}$), as well as the mass ratio ($q$) and projected orbital semi-major axis ($a_\mathrm{orb} \sin i$) in the case where RVs are provided for both components or the single-component projected semi-major axis ($a_\mathrm{comp, i} \sin i$) in the case where RVs are only provided for either one of the stellar components. All estimates are based on the per-component radial velocity equations as a function of the true anomaly ($\vartheta$) \citep{prsa2018}:
\begin{eqnarray}\label{rvs_eq}
RV_1 (\vartheta) & = & \frac{2\pi a_1 \sin i}{P_\mathrm{orb}\sqrt{1-e^2}}[e\cos\omega + \cos(\omega + \vartheta)] + v_{\gamma} \nonumber \\
& = & \frac{2\pi q a \sin i}{P_\mathrm{orb}(1+q)\sqrt{1-e^2}}[e\cos\omega + \cos(\omega + \vartheta)] + v_{\gamma}, \\
RV_2 (\vartheta) & = & - \frac{2\pi a_2 \sin i}{P_\mathrm{orb}\sqrt{1-e^2}}[e\cos\omega + \cos(\omega + \vartheta)] + v_{\gamma} \nonumber \\
& = & - \frac{2\pi a \sin i}{P_\mathrm{orb}(1+q)\sqrt{1-e^2}}[e\cos\omega + \cos(\omega + \vartheta)] + v_{\gamma}.
\end{eqnarray}
We assume that the period ($P_\mathrm{orb}$) is known and use the phase-folded radial velocity curves as input. The data is smoothed with a low-pass Savitzky-Golay filter to remove high frequency noise.
The first two parameters that are estimated in parallel are the mass ratio and systemic velocity. From Eq.~\ref{rvs_eq} we can derive analytical expressions that hold at any point in the orbit for both parameters:
\begin{equation}
v_{\gamma} = \frac{RV_1 (\vartheta) + q RV_2 (\vartheta)}{1+q},
\end{equation}
\begin{equation}
q = \frac{RV_1 (\vartheta) - v_{\gamma}}{-RV_2 (\vartheta) + v_{\gamma}}.
\end{equation}
As we can see, $v_{\gamma}$ and $q$ appear in both expressions, which means we cannot uniquely determine them independently of each other. For that purpose we compute them iteratively, starting from a crude estimate of $q$ as the ratio of primary and secondary RV. To maximize the efficiency, we only estimate $q$ and $v_{\gamma}$ in the RV points that correspond to the maximum primary RV. If only one RV is available, $q$ cannot be reliably estimated and in this case only $v_{\gamma}$ is estimated as the midpoint of the available RV.
With $q$ and $v_{\gamma}$ estimated, we can compute the projected semi-major axis, $a\sin i$. We first estimate the semi-amplitudes from the available observations in each RV:
\begin{equation}
K_i = 0.5 [\mathrm{max}(RV_i-v_{\gamma}) - \mathrm{min}(RV_i - v_{\gamma})]
\end{equation}
where $i = 1, 2$ refers to the primary and secondary component, respectively. The corresponding projected semi-major axes are then:
\begin{equation}
a_i\sin i = \frac{K_i P_\mathrm{orb}}{2\pi} \sqrt{1-e^2}
\end{equation}
Initially, we assume $e=0$. If both RVs are available, $a\sin i = a_1\sin i + a_2\sin i$.
Once $q$, $v_{\gamma}$ and $a\sin i$ are estimated, we fix them in the analytical radial velocities and fit for eccentricity and argument of periastron with a least-squares algorithm. We compute several solutions by iterating over a small grid of starting points with the combinations $e_0 = [0, 0.4]$ and $\omega_0 = [0, \pi/2, \pi]$ which ensures the entire feasible parameter space is covered and we iterate over the least squares solution until the values converge within a tolerance or the maximum number of iterations is reached. After each iteration, the value of $a\sin i$ is recomputed with the updated eccentricity $e$ and the solution is used as an initial point in the next iteration.
Finally, the phase of superior conjunction is estimated as the phase at which $v = v_{\gamma}$ and then converted to the time of superior conjunction, $t_{0, \mathrm{supconj}}$.
An example of the performance of the estimator is given in Figure \ref{fig:rv_geometry}.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{figure_rv_geometry.pdf} \\
\caption{Analytic Keplerian radial velocity model fitted to a synthetic set of observations with estimated values for eccentricity ($e$), argument of periastron ($\omega_0$), mass ratio ($q$), projected semi-major axis ($a_\mathrm{orb} \sin i$), and the time of superior conjunction ($t_{0,\mathrm{supconj}}$) from only the phased radial velocity observations.}
\label{fig:rv_geometry}
\end{figure}
\subsection{Light Curve Geometry}\label{sec:lc_geometry}
PHOEBE 2.3 introduces an estimator based on fitting a two-Gaussian model to a phased light curve to evaluate the eccentricity and argument of periastron, as the two best-constrained parameters from light curve geometry. For a detailed description of the model refer to \citet{mowlavi2017}. Here, we will focus on its implementation in PHOEBE.
The two-Gaussian model is a fast, analytical model that fits a composite function of a constant, cosine term, two Gaussians, or any combination of these to the observed data, resulting in seven different models. The simplest model consists only of a constant term ($C$), while the most complex one of a constant, two Gaussian functions and a cosine term. Each Gaussian is characterized by its central position ($\mu$), amplitude ($A$), and root mean square (RMS) width ($\sigma$). The cosine term is only characterized by its amplitude ($A_{\mathrm{ell}}$) as the model assumes its period is equal to half the light curve period (0.5 in the case of a phase-folded light curve) and its trough coincides with either the position of the primary or secondary eclipse. Figure \ref{fig:lc_geometry} shows an example case with the best fit model and the resulting eclipse phases.
The two-Gaussian model is sensitive to the initial values of the model parameters. Thus, to ensure its convergence, PHOEBE first estimates the eclipse positions, widths and depths using a simple algorithm that searches for the minimum of the light curve and isolates data in its vicinity that cross the median flux in phase ($\varphi$) space. All seven models are then fitted with a least-squares optimizer and the best fit is chosen as the one with the highest Bayesian Information Criterion \citep[BIC;][]{bic1978}. We determine the eclipse parameters following the prescriptions in \citet{mowlavi2017}, where the positions of the eclipses coincide with the central positions of the Gaussians ($\varphi_i = \mu_i$), the widths of the eclipses are a factor of the Gaussian RMS widths ($w_i = 5.6 \sigma_i$) and the depths are computed as the constant term minus the flux at eclipse positions ($d_i = C - \mathrm{flux}(\varphi_i)$).
The estimator computes the time of superior conjunction, eccentricity and argument of periastron as:
\begin{description}
\item[time of superior conjunction]
\begin{equation}
t_{0, \mathrm{supconj}} = t_{0, \mathrm{supconj}, \mathrm{orig}} + \varphi_1 P_\mathrm{orb}
\end{equation}
\item[eccentricity]
\begin{equation}
e = \left[\sin^2\left(\frac{\psi-\pi}{2}\right) + \left(\frac{w_2-w_1}{w_2+w_1}\right)^2\cos^2\left(\frac{\psi-\pi}{2}\right)\right]^{1/2}
\end{equation}
\item[argument of periastron]
\begin{eqnarray}
\omega_1 & = & \arcsin\left(\frac{1}{e}\frac{w_2-w_1}{w_2+w_1}\right) \\
\omega_2 & = & \arccos\left(\frac{\sqrt{1-e^2}}{2e\tan(\psi-\pi)}\right) \\
\omega & = & \left\{
\begin{array}{@{}cc@{}}
\omega_2, & \text{if}\ \omega_1 >= 0 \\
2\pi - \omega_2, & \text{if}\ \omega_1 < 0
\end{array}\right.
\end{eqnarray}
where $\psi = \pi + 2\arctan\frac{e\cos\omega}{\sqrt{1-e^2}}$ and is computed iteratively through solving $2\pi\Delta\Phi = \psi - \sin\psi$, where $\Delta\Phi$ is the separation between the two eclipses in phase space.
\end{description}
In addition, the phases of ingress and egress and eclipse widths and depths are computed and exposed to the user. These are then used to also propose an optional phase-mask based on the eclipse edges and a padding of 30\% the eclipse width (see Section \ref{sec:solver_times} and Figure \ref{fig:lc_geometry_mask}).
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{figure_lc_geometry.pdf}
\caption{Best of seven trial two-Gaussian models used to determine the phases of eclipse minima, ingress, and egress (blue and orange vertical lines representing the primary and secondary eclipses, respectively) as well as the input binned synthetic observations. These are then used to estimate values for eccentricity ($e$), argument of periastron ($\omega_0$), and the time of superior conjunction ($t_{0,\mathrm{supconj}}$).}
\label{fig:lc_geometry}
\end{figure}
\subsection{Artificial Neural Network (\texttt{\sc ebai})}\label{sec:ebai}
\texttt{\sc ebai} is a back-propagating artificial neural network specifically implemented to recover fundamental parameters from phased light curves of eclipsing binary systems \citep{prsa2008, 2019ascl.soft08018P}. For the implementation within PHOEBE 2.3, we retrain the network on detached systems using the two-Gaussian model (see Section \ref{sec:lc_geometry}), using 201 input units, 40 hidden units and 5 output units, learning rate parameter $\eta=0.2$, 33,235 exemplars and $4 \times 10^6$ iterations.
To use \texttt{\sc{ebai}} as an estimator, the user can pass any number of light curve data sets, which are normalized and fitted with a two-Gaussian model, automatically introducing a phase-shift if necessary to ensure the eclipse is at phase zero as required by the training set. This analytic representation is then normalized to unity, sampled at the same 201 phase-points as the training set, and is passed to the two matrix transformations determined from the \texttt{\sc ebai} training (Figure \ref{fig:ebai}). The results directly from \texttt{\sc ebai} are then converted as necessary to the proposed values for: $T_\mathrm{eff, 2} / T_\mathrm{eff, 1}$, $(R_\mathrm{equiv, 1} + R_\mathrm{equiv, 2}) / a_\textrm{orb}$, $e \sin \omega_0$, $e \cos \omega_0$, and $i$, in addition to $t_{0, \mathrm{supconj}}$ computed from the applied phase-shift.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{figure_ebai.pdf} \\
\caption{Analytic two-Gaussian representation of the automatically binned phased light curve sent to ebai to estimate $t_{0, \mathrm{supconj}}$, $T_\mathrm{eff, ratio}$, $R_\mathrm{equiv, fractional~sum}$, $e \sin \omega_0$, $e \cos \omega_0$, and $i$.}
\label{fig:ebai}
\end{figure}
\section{The Merit Function}\label{sec:merit_function}
Any optimization (see Section \ref{sec:optimizers}) or sampling (Section \ref{sec:samplers}) that requires the forward-model to be computed, requires a definition of the merit function to compare the synthetic model to the observations. The components of the merit function (lnprobability) and terms that have been adopted within PHOEBE are explained below:
\begin{description}
\item[lnpriors] the log-priors term is defined as the log probability of drawing the current face values, $p$, from the provided prior probability distribution functions, $\pi$. When priors are not provided or applicable, this term will be constant and so will not contribute to the optimization/sampling. Priors are discussed in more detail in Section \ref{sec:priors}.
\begin{equation}\label{eq:lnpriors}
\mathrm{lnpriors} = \sum_\mathrm{priors} \mathrm{ln}( \mathrm{P}(p \mid \Pi))
\end{equation}
\item[residuals] observations ($y_{o}$) $-$ synthetic model ($y_{m}$), whether those are fluxes or radial velocities, after any phase masking (see Section \ref{sec:solver_times}). By default, the synthetic model is computed at the exact times of the (masked) observations. However, the user can also compute the model at a list of custom times. In that case, the synthetic model is linearly interpolated to the times (or phases) of the observations and then residuals are determined. It is the responsibility of the user to ensure the model is sufficiently sampled so that linear interpolation does not cause any issues. See Section \ref{sec:forward_model} for details about the forward-model itself, times of the synthetic model, phase-masking, and interpolation.
\begin{equation}
\mathrm{residuals} = y_{o}(t_\mathrm{masked}) - y_{m}(t_\mathrm{masked})
\end{equation}
\item[chi2 ($\chi^2$)] $\chi^2$ is defined here to be the sum of squares of the residuals of all data points across all enabled ``datasets'' over the squares of the provided per point uncertainties (if provided - although note that some algorithms require uncertainties to be provided) plus an additional term to handle uncertainty underestimation from the provided observation uncertainties, $\sigma_\mathrm{o}$. This uncertainty underestimation ($\sigma_\mathrm{lnf}$) term can be optimized or sampled in the case where the provided uncertainties are believed to be underestimated. By default, $\sigma_\mathrm{lnf} = - \inf$, which cancels the additional term in the denominator. See Section \ref{sec:noise_nuisance} for more about this ``noise-nuisance'' parameter.
\begin{equation}\label{eq:chi2}
\chi^2 = \sum_\mathrm{datasets} \frac{(y_{o} - y_{m})^2}{\sigma^2} + \mathrm{ln}(\sigma^2),
\end{equation}
%
where
%
\begin{equation}\label{eq:chi2sigma}
\sigma^2 = \sigma_\mathrm{o}^2 + y_{m}^2 \mathrm{e}^{(2 \sigma_\mathrm{lnf})}
\end{equation}
\item[lnlikelihood] the log-likelihood is, by definition:
\begin{equation}
\mathrm{lnlikelihood} = - 0.5 \chi^2
\end{equation}
\item[lnprobability] the log-probability is the default merit function used within PHOEBE (note that, technically, the negative log-probability is used for optimizers which minimize a merit function as the optimal model is found by maximizing the log-probability). When priors are not provided or applicable, this essentially becomes analogous to $\chi^2$. The log-probability is defined as the sum of the log-priors and the log-likelihood:
\begin{equation}\label{eq:lnprobability}
\mathrm{lnprobability} = \mathrm{lnpriors} + \mathrm{lnlikelihood}
\end{equation}
\item[lnposteriors] the log-posteriors, similar to the log-priors, describe the probability of drawing the current face values, $p$, given the resulting posterior probability distribution functions. These posteriors (see Section \ref{sec:posteriors}) are often determined from sampling the log-probability with an algorithm such as MCMC (see Section \ref{sec:mcmc}).
\begin{equation}
\mathrm{lnposteriors} = \sum_\mathrm{posteriors} \mathrm{ln}( \mathrm{P}(p \mid \mathrm{posterior~pdf}))
\end{equation}
\end{description}
Unless overriding with a custom merit function, all optimizers and samplers within PHOEBE use the log-probability as defined above. Additionally, if the user would like to implement their own optimizer or sampler outside of PHOEBE, all of the quantities above are exposed to the user for any given forward-model.
\subsection{Parameterization}\label{sec:parameterization}
Depending on the available observations (single light curve, multiple light curves, single-lined radial velocities, double-lined radial velocities, etc.), previously known information, and the orthogonality (or lack thereof) of the parameter-space, it is advantageous to be able to choose a specific parameterization. For example, due to the parameterization of the Roche model, PHOEBE 1 (based on Wilson-Devinney) parameterizes the orbit by the orbital period ($P_\mathrm{orb}$), mass-ratio ($q$), and semi-major axis ($a_\mathrm{orb}$) (along with eccentricity ($e$) and orientation parameters such as inclination, and argument of periastron). The individual stellar masses ($M_1$ and $M_2$) can be computed from these parameters and Kepler's third law, but cannot directly be fixed or adjusted.
PHOEBE now provides a flexible framework for scenarios like these. For example, by default, PHOEBE adopts a similar parameterization to the legacy version but also exposes read-only parameters for the masses that are derived by Kepler's third law:
\begin{equation}
M_1 = \frac{4 \pi a^3}{G P^2} \frac{1}{1+q}, \quad
M_2 = \frac{4 \pi a^3}{G P^2} (1+q)
\end{equation}
This system of five parameters ($M_1$, $M_2$, $P$, $a$, $q \equiv M_2/M_1$) can be manipulated from the python interface such that any two of the five parameters are ``constrained'' as read-only, based on Kepler's third law. Setting or fitting the individual masses is then possible by choosing two other parameters to be read-only instead.
There are many other similar parameterization choices in a binary model in which one parameterization may be preferred over another due to the orthongonality within the local parameter space or due to the availability of known information in advance. Table \ref{table:constraints} shows a list of the applicable constraints included in the 2.3 release of PHOEBE.
\begin{center}
\begin{table}
\begin{tabular}{ |c|c||c| }
\hline
Default Free Parameter(s) & Other Inputs & Default Read-Only Parameter(s) \\
\hline \hline
$P_\mathrm{orb}$, $q$, $a_\mathrm{orb}$ & & $M_1$, $M_2$ \\
\hline
$P_\mathrm{orb}$ & & $f_\mathrm{orb}$ (frequency) \\
\hline
$e$, $\omega_0$ & & $e \sin \omega_0$, $e \cos \omega_0$ \\
\hline
$a_\mathrm{orb}$, $i_\mathrm{orb}$ & & $a_\mathrm{orb} \sin i_\mathrm{orb}$ \\
\hline
$t_{0,\mathrm{supconj}}$ & $P\mathrm{orb}$, $\dot{P}_\mathrm{orb}$ $e$, $\omega_0$, $\dot{\omega}$, $t_0$ & $t_{0,\mathrm{perpass}}$, $t_{0,\mathrm{ref}}$ \\
\hline
$t_{0,\mathrm{perpass}}$ & $P_\mathrm{orb}$, $\dot{P}_\mathrm{orb}$, $t_0$ & $M$ (mean anomaly) \\
\hline
$F$ (synchronicity) & $P_\mathrm{orb}$ & $P_\mathrm{rot}$ \\
\hline
$P_\mathrm{rot}$ & & $f_\mathrm{rot}$ (frequency) \\
\hline
$M$, $R_\mathrm{equiv}$ & & logg \\
\hline
$a_\mathrm{orb}$ & $q$ & $a_\mathrm{comp}$ \\
\hline
$a_\mathrm{orb}$ & $i_\mathrm{orb}$, $q$ & $a_\mathrm{comp} \sin i_\mathrm{orb}$ \\
\hline
pitch, $i_\mathrm{orb}$ & & $i_\mathrm{comp}$ \\
\hline
yaw, $\Omega_\mathrm{orb}$ (asc.~node) & & $\Omega_\mathrm{comp}$\\
\hline
$A_v$, $R_v$ & & $E(B-V)$\\
\hline
contacts: $R_\mathrm{equiv}$ & $q$ & $FF$ (fillout factor), $\Omega$ (equipotential) \\
\hline
\multicolumn{3}{c}{Optional Constraints} \\
\hline
$T_{\mathrm{eff},1}$, $T_{\mathrm{eff},2}$ & & $T_\mathrm{eff, 2} / T_\mathrm{eff, 1}$ \\
\hline
$R_{\mathrm{equiv},1}$, $R_{\mathrm{equiv},2}$, $a_\mathrm{orb}$ & & $R_\mathrm{equiv,2} / R_\mathrm{equiv,1}$, $(R_\mathrm{equiv, 1} + R_\mathrm{equiv, 2}) / a_\textrm{orb}$\\
\hline
semidetached: $R_\mathrm{equiv}$ & & $R_\mathrm{equiv, crit}$ \\
\hline
$\pi$ (parallax) & & $d$ (distance) \\
\hline
\end{tabular}
\caption{Built-in constraints in PHOEBE (the top of the table before the break are included in systems by default whereas the bottom can optionally be added). The parameters in the right column are, by default, read-only and derived from the parameters in the two left columns. The read-only status of any parameter in the right column can be replaced with one in the left column (as long as any given parameter is ``constrained'' by a single expression), allowing for a large combination of parameterization choices. Parameters in the middle column are included in the expression itself, but are not currently available to be chosen as the read-only parameter.}
\label{table:constraints}
\end{table}
\end{center}
\subsection{Priors}\label{sec:priors}
PHOEBE allows creating and attaching distributions (as \texttt{\sc distl}\footnote{https://github.com/kecnry/distl} distribution objects) to any parameter represented by a float-value in the system. \texttt{\sc distl} is a Python package which handles defining, converting, storing, and manipulating various probability distributions. The version currently packaged with PHOEBE supports the following univariate distribution types: uniform (boxcar), Gaussian, delta, histogram, and samples (which generates a Kernel Density Estimation -- KDE -- under-the-hood around an input set of samples to allow drawing from the same distribution as the samples while not drawing from the set directly). Additionally, the following multivariate distribution types are supported: multivariate gaussians, multivariate histograms, multivariate samples which each contain the known covariances between parameters.
Once defined and attached to a parameter, these distributions can then be plotted directly, or propagated through the constraint logic discussed in Section \ref{sec:parameterization} (see Figure \ref{fig:priors_constraints}). Any distribution attached to a ``free'' parameter can have its value randomly drawn and set for the respective distribution. Any distribution (whether attached to an adjustable or ``read-only'' parameter) can provide the probability of drawing the current face value of the parameter. When passed to an optimization or sampling algorithm that supports priors, these probabilities are then included in the merit function (see Equations \ref{eq:lnpriors} and \ref{eq:lnprobability}).
\begin{figure}
\subfloat{%
\includegraphics[width=0.33\columnwidth]{figure_priors_sma.pdf}
}
\subfloat{%
\includegraphics[width=0.33\columnwidth]{figure_priors_incl.pdf}
}
\subfloat{%
\includegraphics[width=0.33\columnwidth]{figure_priors_asini.pdf}
}
\caption{A uniform distribution placed on the semi-major axis (left) and a gaussian distribution placed on the inclination (middle). These distributions can then be propagated through the logic discussed in Section \ref{sec:parameterization} to get the resulting distribution on $a \sin i$ (right).}
\label{fig:priors_constraints}
\end{figure}
This flexibility allows for numerous useful scenarios. For example, it is possible to set a prior distribution on $a \sin i$ (e.g., known from the literature) while marginalizing over a defined range in both $a$ and $i$. This also allows for setting a prior on any observationally-constrained parameter that is not directly a dataset observable, including $\log g$ or mass from spectra reduction, for instance. As will be discussed in Section \ref{sec:posteriors} it is also possible to sample with one parameterization ($e$ and $\omega_0$, for example) but then expose the posteriors and uncertainties in another ($e \sin \omega_0$ and $e \cos \omega_0$).
\subsection{Angle-Wrapping}\label{sec:angle_wrapping}
All parameters and distributions that correspond to angles (argument of periastron, longitude of ascending node, etc.) include automatic angle-wrapping (with the exception of inclination which is limited to the range between $0$ and $\pi$ to avoid ambiguity with the longitude of ascending node). For example, if setting a value, either manually or within an optimizer or sampler, outside the $[0, 2 \pi]$ range, the value will automatically be mapped back onto the range between $[0, 2 \pi]$. This is particularly useful if the true value is near the wrapping point -- as the optimizer or sampler can explore the parameter space continuously.
This can cause some issues within optimizing or sampling when the merit function is insensitive to one of the parameters. For example, for a circular system, the argument of periastron has no effect on the observables. If allowed to wrap continuously, the sampler will continue to randomly walk indefinitely. To avoid this, the per-wrapped angles are limited to within $\pi$ from the central value of the initializing distribution (see Section \ref{sec:samplers}), essentially placing an uninformative prior and preventing the sampler from wandering across the wrapped space.
\subsection{Forward-Model}\label{sec:forward_model}
The ``forward-model'' consists of the synthetic model for a fixed set of input parameters, which is then compared to the observations through the merit function as described in detail in Section \ref{sec:merit_function}.
\subsubsection{Phase-Masking, Exposure Times, Undersampling, and Interpolation}\label{sec:solver_times}
By default, PHOEBE will compute the forward-model at the timestamps of the observations. For convenience purposes, PHOEBE also supports excluding specific phases of the data to improve computational efficiency. For example, during initial exploration and optimizations of basic geometric and orbital parameters, it can be useful to exclude the out-of-eclipse section of light curves (see Figure \ref{fig:lc_geometry_mask}). These data should then be re-introduced into the merit function for further optimization and determination of posteriors from all available observations. To allow for this, PHOEBE implements a very basic ``phase-masking'' which can be enabled and disabled. Note however that any outlier-removal or more complex masking (in time-space, for example), will currently need to be done outside of PHOEBE. When phase-masking is being used within optimizing or sampling, PHOEBE will exclude those times from both the forward-model and the residuals in the merit function.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{figure_lc_geometry_mask.pdf}
\caption{The same light curve data set as in Figures \ref{fig:lc_geometry} and \ref{fig:ebai} masked to the phases of ingress and egress (plus an additional 30\% padding) as proposed by the light curve geometry estimator (see Section \ref{sec:lc_geometry}).}
\label{fig:lc_geometry_mask}
\end{figure}
Additionally, it is often useful to ``downsample'' the observations to save on computation time or, when not optimizing or sampling, to ``oversample'' for plotting purposes or to predict how the model behaves outside of the limits of observations. To accomplish all of these use cases, PHOEBE allows overriding the computation times of each data set to a custom array of times. When optimizing or sampling, PHOEBE will, by default, automatically take the shorter of these two arrays (the masked observation times or the requested computation times). Within the merit function, the synthetic model is then interpolated onto the times of the masked observations. In cases where the system is not time-dependent, PHOEBE will allow for ``extrapolating'' outside the time-bounds of the synthetic model by interpolating in phase-space, but will not allow doing so when the system is time-dependent (for example: if any time-derivative is non-zero, including the presence of apsidal motion, or at least one of the components has spots and is asynchronously rotating).
Lastly, PHOEBE supports oversampling the model over the exposure time for each requested synthetic time. The resulting synthetic flux at the mid-time is then exposed as the mean of the oversampled fluxes. Long-cadence Kepler data, for example, has a 29.4 minute exposure time which results in phase-smoothing of the photometry. Setting the appropriate exposure time and enabling oversampling allows the model to reproduce this effect.
\subsubsection{Limb-Darkening}\label{sec:ld_mode}
PHOEBE 2.0 introduced automatic interpolated limb-darkening by default \citep[see][Section 5.2.3]{prsa2016}, in addition to manual coefficients for linear, logarithmic, quadratic, square-root, and power laws. PHOEBE 2.2 \citep{jones2020} then included the ability to automatically query coefficients for any of the built-in laws, on a per-surface-element basis.
Optimizing or sampling in PHOEBE 2.3 supports both interpolated and ``lookup'' limb-darkening, but does not currently support fitting manually set coefficients. Interpolated limb-darkening and per-element lookup is only supported within the PHOEBE backend itself -- for other backends (see Section \ref{sec:alternate_backends}), the coefficients are queried from the atmosphere tables in PHOEBE based on the mean stellar parameters.
\subsubsection{Alternate Computation Backends}\label{sec:alternate_backends}
PHOEBE has been designed to be as robust and accurate as possible \citep[for specific implementation details, see ][]{prsa2016, horvat2018, jones2020} -- but this does come at the cost of computational efficiency. Any inefficiency is magnified when a large number of forward-model instances needs to be generated by an optimizer or sampler. Depending on the specific system being studied, substantial time can be saved by ``disabling'' certain effects and making simplifying assumptions. PHOEBE does support some of these assumptions -- irradiation is expensive and can be disabled when negligible or spherical stars can replace Roche-distorted stars, for instance. Other codes may be better suited or more optimized for a given system. In some cases, higher-order effects that require the use of PHOEBE can be ignored while searching the whole parameter space and optimizing orbital parameters, for example, but may still be necessary to include when attempting to get precise estimates for the final reported values.
For this reason, and to enable comparing codes against each other, PHOEBE includes wrappers to the forward-model component of several other public codes, including: PHOEBE legacy \citep[based on Wilson-Devinney;][]{prsa2005, wilson1971, wilson1979, wilson2008, wilson2014}, \texttt{\sc ellc} \citep{maxted2016}, and \texttt{\sc jktebop} \citep{southworth2004, southworth2007, southworth2009, southworth2011}. These forward-models are designed to be ``drop-in'' replacements to the forward-model provided by PHOEBE, with a minimal number of additional options that are specific to that code. Figure \ref{fig:backends_compare} shows a comparison of several forward-models computed through PHOEBE on the same set of parameters. There are some notable discrepancies between the models caused by the different assumptions in each code and limitations on the ability to translate parameterizations and outputs of each code. However, this interface provides the ability to conduct these comparisons and choose the most appropriate model for any given system. Table \ref{table:backends} shows an overview of the implemented features, limitations, and assumptions of each of the backends currently available.
\begin{figure}
\subfloat{%
\includegraphics[width=0.5\columnwidth]{figure_backends_compare.pdf}
}
\subfloat{%
\includegraphics[width=0.5\columnwidth]{figure_backends_compare_semidetached.pdf}
}
\caption{Comparison of several supported backends run through PHOEBE on the same system (detached system on the left and semi-detached system on the right). Although not identical, this is useful to compare the assumptions between different models and to conveniently take advantage of models that are most efficient for a particular science-case or system. Slight vertical scaling differences are likely due to the assumptions in each code as well as the assumptions in the flux-scaling described in Section \ref{sec:flux_scaling}.}
\label{fig:backends_compare}
\end{figure}
These wrappers translate from the parameterization used within PHOEBE to the parameterization used within the requested forward-model, call the external code, and translate the results into the correct units used by PHOEBE. It is important to note that this translation may not always be exact and does add some time-cost -- a native optimization and sampling wrapped around the individual codes will always perform faster. However, regardless of the chosen forward-model, all backends can also benefit from some of the functionality built into the PHOEBE frontend, including: flexible parameterization via ``constraints'' (see Section \ref{sec:parameterization}), phase-masking, exposure times, and time-interpolation (see Section \ref{sec:solver_times}), limb-darkening coefficient lookups (see Section \ref{sec:ld_mode}), flux-scaling via passband luminosities and third light in light curves (some natively, see Section \ref{sec:flux_scaling}), systemic velocities and per-component radial velocity offset (see Section \ref{sec:rv_offset}), noise nuisance to handle observational uncertainty underestimation (see Section \ref{sec:noise_nuisance}), gaussian processes (see Section \ref{sec:gaussian_processes}), and all of the optimization and sampling methods implemented in PHOEBE. If using any of these backends or disabling physical effects within the PHOEBE backend, it is always good practice to occasionally compare the resulting model with one computed without these assumptions.
Whenever publishing work that makes use of these alternate backends, we encourage citing the external code as appropriate.
\begin{center}
\begin{table}
\small
\begin{flushleft}
\begin{tabular}{ r||p{3cm}|p{3cm}|p{3cm}|p{3cm}|| }
$ $ & PHOEBE 2 & PHOEBE 1 \newline(``legacy'') & ellc & jktebop \\
\hline \hline
Supported Versions & 2.3+ & 1.0 & 1.8.2+ & v40+ \\
\hline
LCs & yes & yes & yes & yes \\
LCs (absolute fluxes) & yes & yes & re-scaled & re-scaled \\
RVs (dynamical) & yes & yes & yes & yes \\
RVs (with R-M) & yes & yes & yes (no irrad.) & no \\
Spectral Line Profiles & forward-model & no & no & no \\
Orbits & forward-model & no & no & no \\
Access to Underlying Meshes & forward-model & forward-model & no & no \\
\hline
Detached Systems & yes & yes & yes & yes \\
Semi-detached Systems & yes & yes & yes & no \\
Contact Systems & yes & yes & no & no \\
\hline
(Native) Parameterization & period, \newline mass-ratio, \newline semi-major axis, \newline synchronicities, \newline equivalent radii, \newline Teffs, \newline eccentricity ($e$), \newline arg.~of per.~($\omega_0$) & period, \newline mass-ratio, \newline semi-major axis, \newline synchronicities, \newline equipotentials, \newline Teffs, \newline eccentricity ($e$), \newline arg.~of per.~($\omega_0$) & period, \newline mass-ratio, \newline \newline synchronicities, \newline fractional radii, \newline SB ratio, \newline $\sqrt{e} \cos \omega_0$, \newline $\sqrt{e} \sin \omega_0$ & period, \newline mass-ratio, \newline sum of fractional radii, \newline ratio of radii, \newline SB ratio, \newline $e \cos \omega_0$, \newline $e \sin \omega_0$ (or $e \& \omega_0$)\\
\hline
Surface Distortion & roche, \newline rotating star, \newline sphere & roche & roche, \newline sphere, \newline roche\_v, \newline poly1p5, \newline poly3p0, \newline love & biaxial-spheroid for ellipsoidal and irradiation contributions, sphere for eclipse shapes \\
\hline
Asynchronous Rotation & yes & yes & yes & no \\
\hline
Atmospheres & blackbody, \newline Castelli-Kurucz, \newline phoenix & blackbody (planckint), \newline Castelli-Kurucz (atmx) & none & none \\
\hline
Number of Supported Passbands & $\sim 30$ & $\sim 30$ & 0 & 0 \\
\hline
Irradiation & Horvat/Lambert, \newline Wilson & Wilson & Lambert & biaxial-spheroid \\
\hline
Limb-Darkening & interpolated, lin, log, quad, sqrt, power & lin, log, sqrt & lin, log, quad, sqrt, power & lin, log, quad, sqrt \\
\hline
Gravity Brightening/Darkening & yes & yes & yes & yes \\
\hline
Spots & circular & circular & circular & no \\
\hline
R{\o}mer Delay & yes (optional) & no & yes & no \\
\hline
Apsidal Motion & yes & yes & yes & no \\
\hline
Period Time-Derivative & yes & yes & no & no \\
\hline
Spin-Orbit Misalignment & yes & no & limited & no \\
\hline
Extinction/Reddening & yes & no & no & no \\
\hline \hline
\end{tabular}
\end{flushleft}
\caption{Comparison of available physics and features in the forward-model ``backends'' as wrapped by PHOEBE. Note that some additional capabilities may be available in the native versions of the codes but are excluded here if they are not easily mapped from PHOEBE's parameterization.}
\label{table:backends}
\end{table}
\end{center}
\subsubsection{Flux Scaling (Passband Luminosity, Distance, \& Third Light)}\label{sec:flux_scaling}
The surface fluxes (formally computed at a distance of 1\,m from the equivalent point source) calculated directly by the PHOEBE 2 backend are defined as \citep[cf.][]{prsa2018}:
\begin{equation}\label{eq:F_backend_phoebe}
F_{\mathrm{pb}|w} = \mathcal P_{\mathrm{int}|w} \sum_\mathrm{stars} S_\mathrm{rel} \sum_\mathrm{elems} \langle I_{\mu, \mathrm{abs} |w} \rangle \mathscr{V} \mu \Delta A,
\end{equation}
where $|w$ denotes the flux weighting scheme (energy or photon counts), $\mathcal P_\mathrm{int|w}$ is the suitably weighted integral of the passband transmission function, ``stars'' are the components defined in the system, $S_\mathrm{rel}$ is the per-star scaling factor that translates between absolute and relative units, $\langle I_{\mu, \mathrm{abs}|w} \rangle$ are passband-averaged, per-surface-element specific emergent intensities interpolated from the model atmosphere tables, $\mathscr{V}$ is surface element visibility function, and $\mu \Delta A$ are projected surface areas of each element.
For components where a passband luminosity is provided by the user, $S_\mathrm{rel}$ is determined as:
\begin{equation}\label{eq:S_rel}
S_\mathrm{rel} = \frac{L_{\mathrm{pb, rel}|w}}{L_{\mathrm{pb, abs}|w}}.
\end{equation}
Note that these absolute and relative luminosities are defined at time $t_0$ and exclude any extrinsic or aspect-dependent effects (distance, irradiation, spots, pulsations, boosting, etc). Absolute luminosities are determined from absolute normal intensities from the tables:
\begin{equation}\label{eq:L_pb_abs_phoebe}
L_\mathrm{pb, abs|w} = \pi \mathcal P_\mathrm{int|w} \sum_\mathrm{elems} \langle I_{\perp,\mathrm{abs}|w} \rangle \mathcal{D}_\mathrm{int} \Delta A,
\end{equation}
where $\langle I_{\perp, \mathrm{abs}|w} \rangle$ are the passband-averaged absolute emergent normal intensities interpolated from the model atmosphere tables, $\mathcal{D}_\mathrm{int}$ are the integrals of the limb darkening functions over all angles, and $\Delta A$ are the surface areas of each element \citep{prsa2018}.
Since version 2.2 \citep{jones2020}, PHOEBE has supported multiple modes for handling passband luminosities and scaling of light curve fluxes:
\begin{description}
\item[absolute] intensities are used and integrated directly as retrieved from atmosphere and passband tables, resulting in absolute fluxes. In this case, the scaling factor, $S_\mathrm{rel}$, in Equations (\ref{eq:F_backend_phoebe}) and (\ref{eq:S_rel}) is, by definition, unity.
\item[dataset-scaled] if observations are provided, \textit{``dataset-scaled''} allows to automatically scale the resulting absolute fluxes to the data using a least squares algorithm. This is particularly useful to optimize parameters for a normalized light curve without having to worry about correlations with the luminosity. Although convenient, when determining uncertainties from sampling (Sections \ref{sec:samplers} and \ref{sec:posteriors}), it is advised to instead marginalize over the passband luminosities in order to avoid underestimated uncertainties.
\item[component-coupled] a passband luminosity ($L_{\mathrm{pb, rel}|w}$) for one of the components is provided by the user, and intensities are internally scaled such that this prescribed intrinsic luminosity at time $t_0$ is achieved (Equation \ref{eq:S_rel}). This same scaling factor is then used for both components in Equation (\ref{eq:F_backend_phoebe}).
\item[dataset-coupled] if multi-band photometry is available, \textit{``dataset-coupled''} allows using the same scaling factor across multiple light curves to preserve any color information -- adopting $S_\mathrm{rel}$ for both components in one data set from those defined in another data set, and therefore allowing $L_{\mathrm{pb, rel}|w}$ to be computed from Equation (\ref{eq:S_rel}).
\item[decoupled] decoupling allows for setting the desired luminosities of stars independently, ignoring any scaling information from the atmosphere and passband tables. Here the scaling factors for each star in Equations (\ref{eq:F_backend_phoebe}) and (\ref{eq:S_rel}) are allowed to be independent of each other.
\end{description}
For most ``alternate backends'' (see Section \ref{sec:alternate_backends} and Table \ref{table:backends}), where scaling the intensities or passing passband luminosities directly is not possible, the returned fluxes or magnitudes need to be re-scaled into the units used by PHOEBE by estimating (using the same assumptions as above) the total passband flux from the individual passband luminosities:
\begin{equation}\label{eq:F_backend_other}
F_\mathrm{alt.~backend,~scaled} = F_\mathrm{alt.~backend} \sum_\mathrm{stars} \frac{L_{\mathrm{pb, rel}|w}}{4 \pi}
\end{equation}
When using the \textit{``decoupled''} mode, the desired luminosities, $L_{\mathrm{pb, rel}|w}$ per-star can be used directly to estimate this flux scaling without the need for computing absolute luminosities. But for any other mode, one or more relative passband luminosities must be computed internally (via Equation \ref{eq:S_rel}) by first determining the absolute luminosities of each component. As PHOEBE Legacy follows similar logic to PHOEBE 2 and takes passband luminosities as input; the respective $L_{\mathrm{pb, rel}|w}$ are passed directly and the returned fluxes do not need to be rescaled with Equation (\ref{eq:F_backend_other}). However, unless using \textit{``decoupled''}, \textit{``component-coupled''} (which is supported natively by PHOEBE Legacy), or \textit{``dataset-scaled''}, the absolute luminosites still need to be estimated to handle the appropriate coupling of the relative luminosities. In any of these cases, two methods are introduced in PHOEBE 2.3 to estimate these absolute luminosities:
\begin{description}
\item[PHOEBE meshes at $t_0$] a PHOEBE mesh is created at time $t_0$ using the Roche model for surface distortion and is then populated with the appropriate intensities from the given atmosphere and passband tables and integrated over the entire surface to calculate the intrinsic passband luminosity (Equation \ref{eq:L_pb_abs_phoebe}). As building the mesh is one of the more computationally expensive steps in PHOEBE, estimating luminosities in this way can largely negate the speed benefits of using these ``alternate backends'', but is more accurate for any significant surface distortion.
\item[Stefan-Boltzmann approximation] the mean stellar values for $R_\mathrm{equiv}$, $T_\mathrm{eff}$, logg, and abundance are used to query the selected atmosphere and passband tables for the ``mean'' $\langle I_{\perp|, \mathrm{abs}|w} \rangle$ and the integral of the limb-darkening model $\mathcal{D}_\mathrm{int}$. The absolute passband luminosity for each star is then approximated as:
\begin{equation}
L_{\mathrm{pb, abs}|w} = 4 \pi R_\mathrm{equiv}^2 \langle I_{\perp,\mathrm{abs}|w} (T_\mathrm{eff}, \mathrm{logg}, \mathrm{abun})\rangle \mathcal{D}_\mathrm{int}(T_\mathrm{eff}, \mathrm{logg}, \mathrm{abun}, \mathrm{ld}) \mathcal P_\mathrm{int|w}
\end{equation}
%
This essentially treats each star as a uniform sphere for the purposes of computing the absolute luminosity (and therefore the scaling factor). For cases where distortion is minimal or the exact scaling is not important, using this approximation will be significantly faster than requiring a mesh to be built.
\end{description}
It is important to note that, with either method, Equation (\ref{eq:F_backend_other}) makes assumptions in estimating passband flux levels from these luminosities, and so the scaled fluxes cannot be exact. Furthermore, for backends which accept surface brightness ratio instead of absolute effective temperatures, these are estimated as the ratio of the calculated passband luminosities over the square of their respective equivalent radii. And lastly, the normalizations natively used by these codes differ slightly, both from each other and from PHOEBE. \texttt{\sc ellc} normalizes the exposed light curves by the integrated flux over the irradiated surfaces of each component \citep{maxted2016}, whereas \texttt{\sc jktebop} normalizes to the magnitude of the system out-of-eclipse at quadrature (John Southworth, private communication). The effect of these different treatments on PHOEBE's rescaling can be seen in Figure \ref{fig:backends_compare}. Despite these drawbacks, this implementation allows for simple support of passband-level effects that may not be supported natively by these other codes and allow for an easier drop-in replacement.
To be consistent, the fluxes returned from all backends do not include contributions from third-light and distance (they are disabled even for codes that would otherwise natively support these effects to ensure consistency). In the case where third light is provided in fractional (instead of flux) units, we first convert to flux units from the scaled luminosities of each star:
\begin{equation}\label{eq:l3_frac_flux}
l3_\mathrm{flux} = \frac{l3_\mathrm{frac}}{1-l3_\mathrm{frac}} \sum_\mathrm{stars} \frac{L_{\mathrm{pb, rel}|w}}{4 \pi}
\end{equation}
where ``stars'' are again the components in the system, not including the source of the extraneous third light. This uses the same uniform and spherical approximation as was used for the luminosity to flux translation in Equation (\ref{eq:F_backend_other}) and also requires absolute luminosities to be estimated through either of the two methods mentioned above whenever not using the \textit{``decoupled''} mode.
If using the \textit{``dataset-scaled''} mode, the fluxes are instead scaled to the observations -- before the inclusion of third light and so acting directly on absolute fluxes (i.e.~$L_{\mathrm{pb, rel}|w} \equiv L_{\mathrm{pb, abs}|w}$) as passband luminosity scaling and distance are arbitrary in this case -- by determining a flux-scale factor, $S_\mathrm{ds}$, via least squares:
\begin{equation}\label{eq:ds_l3_flux}
F_\mathrm{synthetic,~dataset-scaled} = S_\mathrm{ds} F_\mathrm{backend} + l3_\mathrm{flux}
\end{equation}
For any case where $l3_\mathrm{frac}$ is provided instead, the conversion to $l3_\mathrm{flux}$ will also include this same scale factor:
\begin{equation}\label{eq:ds_l3_frac}
F_\mathrm{synthetic,~dataset-scaled} = S_\mathrm{ds} \left( F_\mathrm{backend} + \frac{l3_\mathrm{frac}}{1-l3_\mathrm{frac}} \sum_\mathrm{stars} \frac{L_{\mathrm{pb, abs}|w}}{4 \pi} \right)
\end{equation}
PHOEBE 2.3 also introduces support for using \textit{``dataset-scaled''} on multiple light curves which are coupled together using the \textit{``dataset-coupled''} mode. In this case, Equations (\ref{eq:ds_l3_flux}) and (\ref{eq:ds_l3_frac}) are used, as appropriate, but fitting for a single $S_\mathrm{ds}$ for all the coupled data sets simultaneously.
In all other (non-\textit{``dataset-scaled''}) modes, distance and third light are then included on top of the returned fluxes from the backend:
\begin{equation}
F_\mathrm{synthetic} = \frac{F_\mathrm{backend}}{d^2} + l3_\mathrm{flux}
\end{equation}
where $F_\mathrm{backend}$ is either adopted from PHOEBE (Equation \ref{eq:F_backend_phoebe}) or any alternate backend after rescaling, if necessary (Equation \ref{eq:F_backend_other}).
\subsubsection{Systemic Velocity and Per-Component Radial Velocity Offsets}\label{sec:rv_offset}
Similarly to distance and third light being included on top of the respective backend fluxes, radial velocities directly from any backend exclude any offsets (even for backends that do natively support systemic velocities). The final synthetic radial velocities are then defined as:
\begin{equation}
RV_\mathrm{synthetic} = RV_\mathrm{backend} + RV_\gamma + RV_\mathrm{offset}
\end{equation}
where $RV_\gamma$ is the constant barycentric systemic velocity added to all RVs (for all components across all data sets), whereas $RV_\mathrm{offset}$ allows for per-dataset and per-component offsets to the velocities.
These per-component offsets are particularly useful for hot stars, where offsets in spectral lines are commonplace due to being created in different levels of the atmospheres \citep{underhill1994, harmanec2002, shenar2018}. The magnitude of this offset can vary between the two components in the system but also between radial velocity data sets which were reduced from different spectral lines. By building this into the model itself, these offsets can be marginalized over and any correlations (with the mass-ratio or systemic velocity $RV_\gamma$, for example) can be properly accounted for in the posteriors.
\subsection{Noise Nuisance}\label{sec:noise_nuisance}
Determining posteriors and uncertainties of physical parameters of the system depends strongly on the observational uncertainties. Unfortunately, these are often known to be underestimated, so it is important that any such underestimation does not propagate through to a biased solution or underestimated model uncertainties.
In the cases where the quoted observational uncertainties may be underestimated by a constant factor (such as in \textsl{Kepler} data; cf.~\citealt{jenkins2017}), we can introduce the uncertainty scaling factor directly into observational uncertainties in the definition of the merit function (see also Equations \ref{eq:chi2} and \ref{eq:chi2sigma}):
\begin{equation}
\sigma^2 = \sigma_\mathrm{o}^2 + y_{m}^2 \mathrm{e}^{(2 \sigma_\mathrm{lnf})}.
\end{equation}
This can be particularly useful as a ``nuisance parameter'' while sampling to determine posteriors (see Sections \ref{sec:samplers} and \ref{sec:posteriors}). By marginalizing over this underestimation factor, any degeneracies between the factor itself and the physical model parameters can be encoded in the resulting posteriors\footnote{See the example for fitting a line to data with underestimated uncertainties in the \texttt{\sc emcee} online documentation at \url{https://emcee.readthedocs.io/en/v3.0.0/tutorials/line/}.}.
We refer the reader to \citet{hogg2010}, which provides a practical overview to handling observational uncertainties in several different situations.
\subsection{Gaussian Processes}\label{sec:gaussian_processes}
A Gaussian process (GP) is a random process in which each data point is drawn from the assigned random variable where the joint distribution of all variables is Gaussian. GPs allow modeling the noise correlations (such as serial correlation and/or heteroscedasticity) on top of the astrophysical signal described by the noise-free forward-model, and thus replace the need to fit functions (such as polynomials) to the data for the purpose of detrending. Within PHOEBE, GPs are applied to the forward-model and, as such, are available for any of the supported computation backends with any choice of the kernel (which describes the covariances between adjacent points). PHOEBE implements GPs via \texttt{\sc celerite}: a Python implementation of GPs that aims to be computationally efficient and scalable to a large number of data points \citep{celerite}. PHOEBE 2.3 supports any combination of Matern $3/2$ and Simple Harmonic Oscillator kernels, but intentionally excludes the white noise term (``jitter'') natively supported by \texttt{\sc celerite} to avoid conflicting with the noise nuisance parameter separately implemented within PHOEBE (Section \ref{sec:noise_nuisance}).
After the original forward-model is computed and interpolated onto the observation times (if necessary), the residuals between the forward-model and observations are determined. These residuals and the input parameters for the kernels are then passed to \texttt{\sc celerite}, which then provides the GP component of the model. The forward model and GP contribution are then added, resulting in the final model (which in turn is used to compute the residuals or merit function for optimizers or samplers). Note that GPs require the final forward-model to be exposed at the exact times of the observations although the physical forward model is still computed at the requested times or phases as discussed in Section \ref{sec:solver_times}.
Figure \ref{fig:GPs} showcases an example of including synthetic observations with additional superimposed trends and noise. Here the inclusion of GPs accounts for the residuals between the observations and the synthetic forward model.
While GPs would ideally account only for the noise component (stochastic \emph{and} correlated), their high degree of freedom can lead to overfitting the data and accounting for some of the signal. This can manifest in excess GP spectral power at the orbital period and, in consequence, it leads to degenerate solutions and it affects parameter posteriors. It is thus important to properly marginalize (see Section \ref{sec:samplers}) over GP parameters to avoid, or at least quantify, the level of overfitting.
\begin{figure}
\subfloat{%
\includegraphics[width=0.5\columnwidth]{figure_GPs_times.pdf}
}
\subfloat{%
\includegraphics[width=0.5\columnwidth]{figure_GPs_phases.pdf}
}
\caption{Forward-model both without (red dashed lines) and with (green solid lines) the Gaussian processes. Two Gaussian process kernels (a Matern 3/2 kernel and a simple harmonic oscillation kernel) are shown in time-space (left) and phase-space (right).}
\label{fig:GPs}
\end{figure}
\subsection{Parallelization}
The PHOEBE backend itself is parallelized at the per-time level via a Message Passing Interface (MPI) implementation. Here the initial setup is done by a single processor which then sends individual time stamps to the remaining processors to compute the requested synthetic observables at those times. The main processor then compiles and orders the results and exposes the model. The wrappers around other backends generally support parallelization at the per-dataset level -- allowing each processor to compute the observables for all times, but for a single data set. However, within optimizers or samplers that support that capability, the parallelization within the forward-models is disabled in favor of parallelizing at the per-model level via either MPI or multiprocessing. As this has less overhead, it is often more efficient.
It should be noted that for Markov Chain Monte Carlo (MCMC; see Section \ref{sec:mcmc}), for example, this loses its advantages if there are more processors available than requested walkers. Whenever this is the case, PHOEBE automatically switches to run MCMC in serial with each iteration parallelized per-time.
In addition, if not using MPI, PHOEBE will make use of a multiprocessing pool across all available processors for any backend that supports multiprocessing (currently emcee and dynesty).
\section{Optimizers}\label{sec:optimizers}
PHOEBE 2.3 includes wrappers around several optimization algorithms from \texttt{scipy.optimize}\footnote{PHOEBE 2.3 requires \texttt{scipy} 1.2+.} \citep{2020SciPy-NMeth}, including: Nelder-Mead \citep{nelder-mead}, Powell, and conjugate gradient. These optimizers can be quite efficient at improving a model fit once already in the vicinity of the optimal solution -- found via estimators or a full global parameter space search, for example.
PHOEBE allows for choosing which parameters to be adjusted by the optimizer as well as the ability to define priors, if desired. The respective algorithm is then called by passing the negative log-likelihood (see Section \ref{sec:merit_function}) to the appropriate \texttt{\sc scipy} algorithm. The optimization results in the maximum a priori (MAP) solution if priors are defined, or the maximum likelihood estimation (MLE) otherwise. It is important to note that, irrespective of whether the user provides priors or not, the parameter limits, wrapping limits, and any failed models still act as uninformative priors which penalize the merit function.
As with estimators (Section \ref{sec:estimators}), PHOEBE then exposes the proposed values for each of the adjusted parameters along with the value of the merit function before and after optimization as well as any diagnostic values returned by \texttt{\sc scipy}, allowing the user to decide which, if any, of the proposed values to adopt. Figure \ref{fig:optimizer_nm} shows the results of a Nelder-Mead optimization on several parameters starting at the proposed values from estimators. After just a few hundred iterations, the residuals and $\chi^2$ show that the model has improved significantly.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{figure_optimizer_nm.pdf}
\caption{Model improvement after adopting proposals from estimators (dashed red lines) and after running Nelder-Mead (solid green lines).}
\label{fig:optimizer_nm}
\end{figure}
\section{Samplers}\label{sec:samplers}
Samplers are not designed for finding the global solution, rather, their purpose is to explore the shape of the local parameter space and provide robust posteriors and uncertainties that expose the underlying degeneracies between various parameters.
\subsection{Markov Chain Monte Carlo (emcee)}\label{sec:mcmc}
Markov Chain Monte Carlo (MCMC) is a Bayesian method to sample the parameter space in order to estimate the posterior density function -- including any correlations between the sampled parameters -- by using the log probability as the merit function (see Section \ref{sec:merit_function}). Affine-Invariant MCMC algorithms \citep{goodmanweare2010} sample this same space based on the \textit{ensemble} of workers and generally has a shorter burn-in period and requires less user-tuning \citep{foreman-mackey2013}, and is therefore a good candidate to use whenever the forward-model is computationally expensive and the probability space may not be orthogonal.
The \texttt{\sc emcee}\footnote{PHOEBE 2.3 has been tested with \texttt{\sc emcee} v3.0 (and does not support earlier versions).} \citep{foreman-mackey2013, emceev3} python package is an Affine-Invariant Markov Chain Monte Carlo Python implementation that has been widely adopted by the field. PHOEBE includes a wrapper around \texttt{\sc emcee}, exposing most of the basic functionality along with the convenience to easily add or remove sampled parameters and use complex distributions for priors and the initial sample, while handling the merit function and parallelization issues. However, as with any algorithm, it is still important to understand \texttt{\sc emcee} itself, how best to adjust the inputs, and how to interpret the results. \citet{foreman-mackey2013} provides a good overview and discussion and the online documentation for \texttt{\sc emcee} (\url{https://emcee.readthedocs.io}) provides several helpful examples.
When preparing an \texttt{\sc emcee} run through PHOEBE, the user can define the distributions to be used for the initial sample (which defines which parameters will be included in the samples) in any preferred parameterization and, optionally, distributions for the priors. If priors are not provided, uninformative distributions are still adopted based on the limits of the individual parameters and angle wrapping limits (see Section \ref{sec:angle_wrapping}). Furthermore, if any individual model fails for any reason, the merit function (see Section \ref{sec:merit_function}) will return $- \infty$, essentially acting as a prior on the allowed parameter space due to unphysical systems or limitations of the backend itself. The causes for each of these rejected samples can optionally be exposed to the user, as shown in Figure \ref{fig:emcee_failed_samples}.
\begin{figure}
\centering
\includegraphics[width=1.0\textwidth]{figure_emcee_failed_samples.pdf}
\caption{An example corner plot showing the posteriors from an emcee run along with the positions in the parameter space of ``failed/rejected samples''. The legend shows that some of the values proposed by emcee were rejected due to Roche overflow or a failure to lookup appropriate limb darkening coefficients, for example. This plot is particularly useful to diagnose any issues, especially if the posteriors seem to be pushing against a boundary.}
\label{fig:emcee_failed_samples}
\end{figure}
Although the initial sample and the priors can be the same set of distributions, it is generally good practice to start \texttt{\sc emcee} in an N-dimensional hyperball (i.e.~via Gaussian distributions on multiple parameters) around the best-known solution from optimization (see Section \ref{sec:optimizers}). Alternatively, it is possible to sample directly from the priors themselves (although this is generally less efficient and more susceptible to getting stuck in local solutions -- see \citet{foreman-mackey2013} for a discussion on this topic). Priors, on the other hand, can either be informative -- uncertainties from an external analysis or from the literature -- or uninformative -- conservative uniform distributions to prevent the chains from ``wandering'' into parameter space that is clearly incorrect or unphysical, and are included in the merit function as described in Section \ref{sec:merit_function}.
In addition to distributions, the user can set options for the number of processors, number of walkers, and number of iterations. After the run is complete (or at intermediate steps, as requested by the user), the user can view the progress of the chains and the log-probability versus iteration and adjust the burnin and thinning parameters as necessary. If the run is completed but has not yet converged, PHOEBE allows continuing an \texttt{\sc emcee} run from the existing chains.
In addition to the full chains, the acceptance fractions and autocorrelation times are exposed to the user. As a rule of thumb, \citet{foreman-mackey2013} suggests acceptance fractions to all be between 0.2 and 0.5 and enough iterations to cover roughly ten autocorrelation times. PHOEBE then automatically determines defaults for thinning and burn-in based on the autocorrelation times of the chains:
\begin{equation}
\mathrm{burnin} = F_\mathrm{burnin} \max( \tau_\mathrm{autocorr})
\end{equation}
and
\begin{equation}
\mathrm{thin} = F_\mathrm{thin} \min( \tau_\mathrm{autocorr})
\end{equation}
where $F_\mathrm{burnin}$ and $F_\mathrm{thin}$ are user-defined scaling factors that default to 2 and 0.5, respectively, and $\tau_\mathrm{autocorr}$ are the estimated autocorrelation times, per-parameter, as exposed by \texttt{\sc emcee}.
Once the chains sufficiently cover the parameter space and are considered to be converged, the user can then adjust the thinning, burn-in, and/or a cutoff in log-probability to apply to the chains returned by \texttt{\sc emcee} and generate a multivariate distribution of the resulting posteriors (see more in Section \ref{sec:posteriors}).
\subsection{Dynamic Nested Sampling (dynesty)}\label{sec:nested_sampling}
Nested sampling is another Bayesian method for sampling the parameter space. Unlike MCMC, which samples from an initial distribution, walks the parameter space, and is ``penalized'' by the priors in the merit function, nested sampling continually samples directly from the priors themselves which are broken into slices \citep[see ][for an in-depth comparison and discussion]{speagle2020}. Nested sampling, unlike MCMC, is capable of exploring and exposing multi-modal posteriors. However, with wide uninformative priors, nested sampling can become prohibitively expensive. \textit{Dynamic} nested sampling somewhat alleviates this expense by dynamically changing the number of samples per iteration as the algorithm converges to the final posterior. Additionally, unlike MCMC which is allowed to wander outside the initial sampling distributions as long as the priors allow it, dynamic nested sampling will never consider any solution outside the original parameter-space defined by the priors.
The \texttt{\sc dynesty}\footnote{PHOEBE 2.3 has been tested with \texttt{\sc dynesty} v1.0.} code is a dynamic nested sampling Python package. As with \texttt{\sc emcee}, it is very useful to first understand the details and intricacies of dynamic nested sampling and the \texttt{\sc dynesty} code. \citet{speagle2020} provides detailed discussion and the online documentation (\url{https://dynesty.readthedocs.io}) gives numerous examples.
Since \texttt{\sc dynesty} samples directly from the priors, PHOEBE excludes the priors from the log-likelihood. Instead, PHOEBE makes use of \texttt{\sc distl} (see Section \ref{sec:priors}) to transform the defined priors into prior transforms which map a random value between 0 and 1 onto the drawn values of each individual parameter, and passes these to \texttt{\sc dynesty}. It is important to note that this transformation requires ignoring any possible covariances between parameters in the priors (i.e.~if priors were provided as multivariate distributions they would be first flattened into univariate distributions before being passed to \texttt{\sc dynesty}).
After calling \texttt{\sc dynesty} from within PHOEBE, the user has access to all of the output arrays from \texttt{\sc dynesty} itself to access or plot any diagnostic figures. The samples from \texttt{\sc dynesty} can then be transformed into posterior distributions by accounting for their respective weights.
\section{Posteriors \& Uncertainties}\label{sec:posteriors}
The results from any sampler can be accessed as a posterior distribution -- representing not only the uncertainties on each of the sampled parameter values, but also the correlations (or covariances) between all of the sampled parameters. When these distributions are sufficiently gaussian, the posteriors can easily be converted into a multivariate gaussian distribution which represents the posteriors as just the means and a covariance matrix. Figure \ref{fig:posteriors_mvsamples_mvgaussian} shows an example corner plot of the posteriors of several parameters obtained from \texttt{\sc emcee} samples and their conversion into a multivariate gaussian.
\begin{figure}
\subfloat{%
\includegraphics[width=0.5\columnwidth]{figure_posteriors_mvsamples.pdf}
}
\subfloat{%
\includegraphics[width=0.5\columnwidth]{figure_posteriors_mvgaussian.pdf}
}
\caption{An example corner plot of the (burn-in and thinning applied) posteriors directly from emcee results (left). These can optionally be translated into a multivariate guassian distribution (right) which, when appropriate, represents the same information with only the means and covariance matrix. This is particularly useful to include the means and covariances in a publication or to reuse the posteriors as priors.}
\label{fig:posteriors_mvsamples_mvgaussian}
\end{figure}
As these distributions are \texttt{\sc distl} distribution objects, they can easily be manipulated into any of the supported distribution types, saved to a file, plotted, expose symmetric or asymmetric uncertainties to include in published results, or used for further sampling. For example, posteriors from one sampling run can be adopted and used as either the initial sampling distribution or priors for another sampling run. This can be useful to optimize and determine posteriors for one data set, but maintain that information in the log-likelihood for another data set or when new observations become available. As with distributions used for priors (Section \ref{sec:priors}), these posteriors can be propagated through the ``constraints'' and exposed in any desired parameterization (Section \ref{sec:parameterization}). For example, Figure \ref{fig:posteriors_parameterization} shows the \texttt{\sc emcee} samples for eccentricity and argument of periastron which can then be propagated to $e \sin \omega_0$ and $e \cos \omega_0$ which are easier to represent as a multivariate gaussian and report in published scientific results. This allows for independent choices for the combination of parameters for sampling, priors, and posteriors, giving the full flexibility needed based on the available observations and known information.
\begin{figure}
\subfloat{%
\includegraphics[width=0.5\columnwidth]{figure_posteriors_ew.pdf}
}
\subfloat{%
\includegraphics[width=0.5\columnwidth]{figure_posteriors_ecs.pdf}
}
\caption{An example corner plot of the eccentricity and argument of periastron posteriors from an emcee run (left). Although the distributions are correct, these are by nature both non-gaussian and highly correlated, making it difficult to parameterize the results for inclusion in a publication, for example. These posteriors can be propagated through the respective constraints (see Section \ref{sec:parameterization}) to instead expose the posteriors in $e \sin \omega$ and $e \cos \omega$ which are more orthogonal and can often conveniently be converted to gaussian posteriors.}
\label{fig:posteriors_parameterization}
\end{figure}
It is important to note that the degeneracies and uncertainties determined by any sampler are computed under the assumption that all remaining parameters are held fixed at their face values with no uncertainties, leading to uncertainty underestimation or possibly even an incorrect solution. For example, if the value of the orbital period is believed to be well-known and therefore left fixed during sampling, the resulting posteriors -- and therefore uncertainties -- on all other sampled parameters will exclude any degeneracies with respect to the orbital period, and will therefore likely be underestimated. Likewise, in the case where the assumed fixed value is incorrect, the remaining sampled parameters may be incorrect as well. In effect, all parameters that are not sampled are treated as if they have a delta function for a prior and are known to infinite precision. In a perfect world, all parameters would be sampled or marginalized over, but practically that is not feasible with tens or even hundreds of possible parameters.
Lastly, as with any distribution, PHOEBE now allows propagating posteriors through the forward-model itself. By running the forward-model over a number of samples from the posterior distributions, PHOEBE can compute and expose these individual models or the median model and 1-, 3-, or 5- sigma spreads in flux or radial velocity. Figure \ref{fig:posteriors_sample_from} shows an example of the 3-sigma spread in the synthetic model when drawn from the posteriors from an \texttt{\sc emcee} run.
\begin{figure}
\subfloat{%
\includegraphics[width=0.5\columnwidth]{figure_posteriors_sample_from_lc.pdf}
}
\subfloat{%
\includegraphics[width=0.5\columnwidth]{figure_posteriors_sample_from_rv.pdf}
}
\caption{Any distribution (in this case the posteriors from the emcee run) can be propagated through the forward-model. Here the three-sigma uncertainties on the light curve (left) and radial velocity curves (right) and their residuals (bottom) are shown, depicting that the spread in the observations, given their uncertainties, are well represented by the spread in the model caused by the uncertainties in the posteriors.}
\label{fig:posteriors_sample_from}
\end{figure}
\section{User Interface}\label{sec:ui}
The PHOEBE 2.3 release also introduces a user interface, capable of all of the functionality of both the forward-model and inverse problem described above, including basic plotting functionality, and is available for download at \url{http://phoebe-project.org/clients}. The user interface is designed with a server--client model, splitting the codebase into three distinct roles: the Python package itself, the user interface client, and a lightweight webserver which can either run on the same machine or remotely from the client and is responsible for running commands sent from the user interface client via the Python package. This design allows for the flexibility of installing the server on a remote machine with more resources and either installing the user interface locally or accessing a hosted version through the web browser\footnote{a publicly available version is currently hosted at \url{http://ui.phoebe-project.org}.}.
With or without the user interface, PHOEBE also allows the user to conveniently export the forward-model or inverse-problem jobs into a standalone script, which can then be run on a high performance cluster, and the resulting file can be imported locally to view the progress or final results. This allows for running all interactive and visual components of the fitting process on a local machine while offloading computationally expensive tasks to a remote machine that may not be capable of hosting a webserver or plotting graphics.
\section{Conclusion}
The 2.3 release of PHOEBE builds on past releases and introduces a general framework for data fitting running several common algorithms to address the inverse problem, and an interface for defining and dealing with complex distributions. It also introduces the ability to run the forward-model through several other publicly available codes in addition to PHOEBE, and a web-based user interface.
Although far from a black-box automated fitting pipeline, this common interface aims to ease the learning curve and effort required to employ multiple algorithms and codes while attempting to obtain both accurate and precise parameter values and uncertainties for eclipsing binary systems. As additional algorithms and forward-models continue to be adopted by the field, we plan to incorporate them into PHOEBE within this same framework.
PHOEBE is an open-source project under the GPL3 license and is hosted at \url{http://phoebe-project.org} and \url{https://github.com/phoebe-project/phoebe2}. Contributions and feedback are welcome.
\acknowledgments
The development of PHOEBE is possible through the NSF AAG grants \#1517474 and \#1909109 and NASA 17-ADAP17-68, which we gratefully acknowledge.
We thank John Southworth and Pierre Maxted for their permission and discussions regarding their codes, \texttt{\sc jktebop} and \texttt{\sc ellc}.
DJ acknowledges support from the State Research Agency (AEI) of the Spanish Ministry of Science, Innovation and Universities (MCIU) and the European Regional Development Fund (FEDER) under grant AYA2017-83383-P. DJ also acknowledges support under grant P/308614 financed by funds transferred from the Spanish Ministry of Science, Innovation and Universities, charged to the General State Budgets and with funds transferred from the General Budgets of the Autonomous Community of the Canary Islands by the Ministry of Economy, Industry, Trade and Knowledge.
KH gratefully acknowledges support from NASA ADAP grant 18-ADAP18-228.
DRH gratefully acknowledges the support of the Australian Government Research Training Program.
MA acknowledges support from the FWO-Odysseus program under project G0F8H6N.
\vspace{5mm}
\software{ PHOEBE \citep{prsa2016, horvat2018, jones2020},
distl (\url{https://github.com/kecnry/distl}),
EBAI \citep{prsa2008, 2019ascl.soft08018P},
ellc \citep{maxted2016},
jktebop \citep{southworth2004, southworth2007, southworth2009, southworth2011},
celerite \citep{celerite},
emcee \citep{foreman-mackey2013, emceev3},
dynesty \citep{speagle2020},
schwimmbad \citep{2017JOSS....2..357P},
numpy \citep{numpy},
scipy \citep{2020SciPy-NMeth},
astropy \citep{2013A&A...558A..33A},
matplotlib \citep{2007CSE.....9...90H},
corner \citep{corner}
}
|
\section{Introduction}
Satellite Earth observation (EO) data can provide critical, relevant, and timely information in support of agricultural monitoring and food security \citep{BeckerReshef2019,NationalAcademies2019}. EO can enable crop yield assessment and forecasting \citep{Becker2010,Franch2015,Franch2017,Lobell2013,Porter2017}, inform agricultural commodities markets, enable early warning and mitigation of impending crop shortages, and inform decisions about subsidies in times of crisis such as extreme drought \citep{Dorward2011} or outbreaks of disease or pests. However, the lack of up-to-date, high-resolution cropland and crop type maps is a critical barrier to agricultural and food security assessments and operational rapid response to humanitarian crises. These data are especially lacking in regions dominated by smallholder agriculture that are most vulnerable to food insecurity (e.g., sub-saharan Africa).
Machine learning techniques have commonly been used to identify cropland in EO data. The majority of studies use tree-based classifiers (primarily random forests or decision trees) or neural network/deep learning methods (primarily recurrent or convolutional neural networks). However, prior methods have typically been applied to small and/or spatially homogeneous areas \citep{RNN_Camargue, RNN_winter, Skakun2019}, require large training datasets (hundreds of thousands of labels) \citep{breizhcrops2020,Wang2020,lstm_crop_identification, RF_france}, or use EO inputs with insufficient spatial resolution for reliably detecting smallholder farms which often have field sizes $<$1 ha \citep{GFSAD,Teluguntla2018,Oliphant2019,CopernicusLandCover,Samasse2020,Waldner2017}. In many regions, particularly developing countries and regions dominated by smallholder farming, the lack of readily accessible ground truth data is a major barrier for training and validating machine learning techniques for cropland classification \citep{Coutu2020}.
In this paper, we demonstrate a novel method for rapidly generating cropland maps for areas with little to no available labeled data over a large heterogeneous area. This is achieved by leveraging global and local datasets. The global dataset contains tens of thousands of crowdsourced labels from diverse geographies \citep{GeoWiki} to learn general features for identifying cropland globally. These global examples supplement a much smaller dataset of curated, local examples (manually labeled using photointerpretation of high-resolution satellite imagery) to learn features specific to cropland in the target mapping region. Combined, these datasets provide sufficient examples to train a Long Short Term Memory network (LSTM) \citep{LSTM} to predict the presence of cropland in each pixel given a monthly time-series of Sentinel-2 multispectral observations.
The COVID-19 global pandemic has severely impacted the global food system, and the threat of food shortages and rising food prices has required countries at risk of food insecurity to rapidly design and enact aid programs in response. In the West African country of Togo, the government announced a program to boost national food production in response to the COVID-19 crisis by distributing aid to farmers, but found that high-resolution spatial information about the distribution of farms across Togo (which primarily consist of smallholder farms under 1 ha) was a critical piece of missing information for designing this program. Using our method, we were able to generate an accurate high-resolution (10 m) map of cropland in Togo for 2019 in under 10 days to respond to the Togolese government's immediate need for information about the distribution of farms across the country. We showed that our model outperforms existing publicly-available cropland maps for Togo \citep{GFSAD, CopernicusLandCover} while also being higher resolution and more recent. We have delivered this product to the Togolese government where it is being used to help inform decision-making in response to the COVID-19 pandemic, thus demonstrating a successful transition of machine learning research to operational rapid response for a real humanitarian crisis. We have made our map as well as all data and code used in this study publicly available to facilitate future research and operational systems in data-sparse regions.
\begin{figure}
\centering
\includegraphics[width=0.75\linewidth]{diagrams/label_distribution.pdf}
\caption{Distribution of labeled examples in training (a) and test (b) datasets.}
\label{fig:labels}
\end{figure}
\section{Methods}
\subsection{Labeled Data}
\subsubsection{Crowdsourced Labels}
We leveraged a large (35,866 labels) crowd-sourced dataset with labels of ``crop'' or ``non-crop'' from diverse, globally-distributed locations \citep{GeoWiki} to help the model learn features that are useful for detecting a wide range of cropland globally. Details about sample selection, labeling procedures, and quality assessment/control measures can be found in \citet{Bayas2017}. Each location was labelled by multiple labelers; to turn these labels into binary labels, we took the mean of the labels for each location, with a threshold of 0.5 to define a point as ``crop'' or ``non-crop.''
\subsubsection{Active Labeling}
To generate training labels, experts in photointerpretation of agricultural land cover (authors Becker-Reshef, Barker, Hosseini, Kerner, Munshell, and Paliyam) drew polygons over pixels with labels of ``crop'' or ``non-crop.'' We used a combination of basemaps for photointerpretation: a SkySat 72cm composite of January-March 2019 \citep{Planet}, PlanetScope composites of April-July and July-October 2019 \citep{Planet}, and Google Earth Pro basemaps (comprised primarily of high-resolution Maxar images). We used QGIS and Google Earth Pro to draw polygons over basemaps. The locations of labels were chosen to be distributed across different agroecological zones and non-crop land cover types observed across Togo. The central pixel within each polygon was used for training. We obtained 394 crop and 194 non-crop labels for the initial training dataset to supplement the 43 examples within Togo in the \citet{GeoWiki} dataset (631 examples total). After training the model with the hand-labeled and \citet{GeoWiki} dataset, we collected additional labeled examples (294 crop and 394 non-crop) by analyzing the predicted map and identifying areas of model confusion (e.g., transient vegetation on shorelines that the model might confuse with crops). This allowed us to maximize the utility of each training example, as it focused on places where the model was weakest rather than adding redundancy to the dataset. This resulted in a final training dataset of 1,319 hand-labeled examples, visualized spatially in Figure \ref{fig:labels}a.
\subsection{EO Data}
We used Copernicus Sentinel-2 surface reflectance (Level 2A) observations that corresponded to label locations as input to the model. We exported 160 m $\times$ 160 m patches for each labeled pixel using Google Earth Engine (GEE) \citep{GORELICK201718}, then extracted the closest pixel within the patch to the label location. To construct a cloud-free timeseries representation of each pixel, we used the algorithm from \citet{CloudFreeSentinel}, which finds the least cloudy pixel within a defined time period. We used 12 30-day time periods to construct a monthly timeseries spanning 360 days for each labeled pixel. We used all spectral bands except for B1 (coastal aerosol) and B10 (cirrus SWIR), and added the normalized difference vegetation index (NDVI), computed as $\text{NDVI} = \frac{\text{B08}-\text{B04}}{\text{B08}+\text{B04}}$. All bands (which range from 10 to 60m resolution) were upsampled to 10 m during GEE export. We used observations acquired between March of Year N and March of Year N+1 where N is the year the labels were created for. Thus for \citet{GeoWiki}, we used observations acquired March 2017-March 2018 and for our hand-labeled dataset we used observations acquired March 2019-March 2020.
\subsection{Model} \label{section:model}
We trained a one layer multi-headed LSTM model (Figure \ref{fig:lstm}) to predict whether a pixel contained cropland (1) or not (0). The input was expressed as a timeseries $\bm{X} = \{\bm{x}_{1}, \bm{x}_{2}, ... \bm{x}_{12}\}$, where each timestep $\bm{x}_{i} \in \mathbb{R}^{1\times14}$ consisted of the least-cloudy composite over a 30-day period of 11 optical bands (all bands except B01 and B10) plus NDVI. The LSTM hidden layer had 64 units. A classifier and a sigmoid activation were used on the final hidden output to map the hidden layer output (logits) to a value that can be interpreted as the posterior probability of crop presence in the pixel.
We performed a grid search that optimized AUC ROC on the validation set to choose values for two hyperparameters: number of linear layers in the classifier (1 or 2) and whether or not a dropout of 0.2 should be applied between each timestep in the LSTM.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{diagrams/model.png}
\caption{A multi-headed LSTM based model for pixel-wise crop identification. The model has two classification layers: a global linear classifier to classify instances outside Togo, and a local linear classifier to classify instances within Togo.}
\label{fig:lstm}
\end{figure}
To allow the model to focus on examples within Togo while also learning from the global \citet{GeoWiki} dataset, we employed a multi-headed LSTM model, where one classifier was used for all examples within Togo (local instances) and another was used for all other examples (global instances). All instances were shuffled during training: a batch could therefore contain a combination of global and local data instances. To combine the results from both classifiers, we used the following loss function to train the model:
\begin{equation}
\mathcal{L} = \frac{W}{\alpha}\mathcal{L}_{\textrm{global}} + \mathcal{L}_{\textrm{local}}
\end{equation}
Where $\mathcal{L}_{\textrm{global}}$ and $\mathcal{L}_{\textrm{local}}$ are the binary cross-entropy losses for the global and local data instances respectively, and $\alpha$ is a weighting parameter. $W$ is a batch-specific value, used to weight the ratio of global and local instances in a batch:
\begin{equation}
W = \frac{\mbox{Number of global instances in batch}}{\mbox{Number of local instances in batch}}
\end{equation}
In our experiments, we used $\alpha = 10$. We used the Adam optimizer with default parameters to tune weights during training \citep{Kingma2014}. To determine when to stop training, we used early stopping with a patience of 10 using a validation set consisting of 20\% of the combined \citet{GeoWiki} and hand-labeled Togo datasets. We implemented the model and experiments using PyTorch \citep{pytorch} in python. All code is publicly available at \url{https://github.com/nasaharvest/togo-crop-mask}.
\begin{table*}
\caption{Results for each method evaluated on the test set. The best results are in bold.}
\label{tab:experimental-results}
\begin{tabular}{cccccc}
\toprule
Description &Accuracy &AUC &Precision &Recall &F1\\
\midrule
\citet{CopernicusLandCover} (Copernicus 100m) & 0.67 & 0.74 & 0.54 & 0.24 & 0.33\\
Single-headed LSTM (hand-labeled only) & 0.71 & 0.90 & 0.59 & \textbf{0.93} & 0.72\\
\citet{GFSAD} (GFSAD 30m) & 0.74 & - & 0.62 & 0.62 & 0.62 \\
Support Vector Machine (hand-labeled only) & 0.79 & 0.88 & 0.65 & 0.84 & 0.73 \\
Random Forest (hand-labeled only) & 0.81 & 0.90 & 0.66 & 0.88 & \textbf{0.76} \\
Multi-headed LSTM (hand-labeled + \citet{GeoWiki}) & \textbf{0.83} & \textbf{0.91} & \textbf{0.81} & 0.68 & 0.74\\
\bottomrule
\end{tabular}
\label{experimental-setup-results}
\end{table*}
\section{Results}
\label{sec:results}
We selected points for a test dataset by randomly sampling 350 points within the Togo country boundaries, constrained by a minimum distance of 50 m between samples (Figure \ref{fig:labels}b). Four experts (authors Barker, Kerner, Munshell, and Nakalembe) labeled each example. Labels were determined by majority vote in order to mitigate label noise in the absence of groundtruth validation, and ties were discarded. The consensus test dataset contained 106 crop and 200 non-crop examples (306 total test examples).
We compared two experimental setups to evaluate the contribution of the \citet{GeoWiki} dataset:
\begin{itemize}
\item A single-headed LSTM trained on only the hand-labeled Togo data
\item A multi-headed LSTM trained on both the hand-labeled Togo and \citet{GeoWiki} data
\end{itemize}
In both cases, we used a model with hyperparameters determined by the grid search described in \ref{section:model}. For the single headed model, this consisted of a single classification layer in the classifier \textit{without} dropout in the LSTM. For the multi-headed model, this consisted of two classification layers in the classifier \textit{with} dropout in the LSTM.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{diagrams/togo_full.pdf}
\caption{Cropland maps of Togo generated using our method, \citet{CopernicusLandCover}, and \citet{GFSAD}. White boxes indicate locations of insets shown in Figure \ref{fig:comparisons}.}
\label{fig:fullmap}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.9\linewidth]{diagrams/comparisons.pdf}
\caption{Qualitative comparison of cropland maps produced using our multi-headed and single-headed LSTM methods, \citet{CopernicusLandCover}, and \citet{GFSAD} for three example regions, with SkySat 72 cm basemap (April-March composite) for reference. Locations of each inset indicated by white boxes in Figure \ref{fig:fullmap}.}
\label{fig:comparisons}
\end{figure*}
We compared our results to two publicly-available cropland maps of Togo. One map is Global Food Security-support Analysis Data (GFSAD) cropland extent map of Africa \citep{GFSAD}. This map has 30 m resolution and was produced using two pixel-based supervised classifiers, a Random Forest and a Support Vector Machine (SVM), and one object-based classifier, a Recursive Hierarchical Image Segmentation classifier. The GFSAD map does not provide cropland probabilities, but a discrete class assigned to each pixel of 0 (water), 1 (non-cropland), or 2 (cropland). The most recent GFSAD product is from 2015. Prior to our map, the GFSAD map is to our knowledge the highest resolution cropland map publicly available for Togo. Figure \ref{fig:fullmap} (right) shows the GFSAD land cover map for Togo. The second map we compared our results to is the Copernicus Land Cover cropland map \citep{CopernicusLandCover}. The Copernicus map specifies the fraction of each pixel covered by crops (using values 0-100) at 100 m resolution and was produced using a combination of random forests and expert rules with satellite data inputs from the PROBA-V sensor. The most recent Copernicus map is for 2018. To enable quantitative comparison with our results, we treated crop fraction as analogous to crop probability and computed all performance metrics using our test dataset (Table \ref{tab:experimental-results}); however, we emphasize that this may not be how the map was intended to be used and qualitative comparisons may be more faithful to the intended use of the \citet{CopernicusLandCover} map. Figure \ref{fig:fullmap} (center) shows the Copernicus crop fraction map for Togo. A direct comparison between our method and \citet{GFSAD} and \citet{CopernicusLandCover} would require running all three methods on the same data (i.e., 10m Sentinel-2 observations paired with our labeled examples). Since neither the code for these methods nor sufficient information for reproducing them faithfully are available, we additionally performed two baseline comparisons to Random Forest and SVM classifiers which are the basis for the \citet{GFSAD} and \citet{CopernicusLandCover} methods.
For the Random Forest, we used 100 estimators with no maximum depth (nodes were expanded until all leaves were pure, or until the leaves contained fewer than 2 samples). For the SVM, we used a radial basis function kernel. We implemented both methods using Scikit-learn in python \citep{scikit-learn}.
To assess performance, we computed the area under the receiver operating characteristic (ROC) curve (AUC score), accuracy (0.5 threshold on posterior probabilities), precision (user's accuracy), recall (producer's accuracy), and F1-score (harmonic mean between precision and recall). We did not compute AUC for GFSAD, which only provides the discrete predicted class. The results are reported in Table \ref{tab:experimental-results}. Figure \ref{fig:fullmap} (left) shows our predicted cropland map.
To qualitatively compare our method compared with \citet{GFSAD} and \citet{CopernicusLandCover}, we show each map compared to the SkySat 72cm basemap (true color) for several example locations in Figure \ref{fig:comparisons}. Each region in Figure \ref{fig:comparisons} contains primarily cropland, though the level of vegetation in non-crop areas varies in each image since Togo's climate ranges from dry savanna in the north to tropical in the south. While all maps detect most fields in the second row, the \citet{GFSAD} and \citet{CopernicusLandCover} maps failed to detect most of the fields in the first and third rows.
Our resulting map of cropland probabilities at 10 m resolution as well as all hand-labeled training and testing labels are publicly available at: \url{https://doi.org/10.5281/zenodo.3836629}.
\section{Discussion}
The multi-headed model trained on the combined hand-labeled and \citet{GeoWiki} dataset had the highest score for most performance metrics. While AUC score is a measure of model performance across a range of possible thresholds on posterior probability, the decision boundary when using the sigmoid function as the output activation layer is 0.5. While in practice a different threshold could be chosen, it is not clear how to construct a large and diverse enough validation set that could be used to reliably select an alternate threshold over very large, heterogeneous regions (e.g., the country of Togo). Thus, the improved accuracy and precision of the multi-headed model (using a threshold of 0.5) - in addition to its improved AUC score - made it preferable. Additionally, since the multi-headed model included training data from prior years (2017) and global instances of cropland \citep{GeoWiki}, we expect that this model would generalize better to potential domain shift in future years than the single-headed model trained with only one year of local data. Figure \ref{fig:comparisons} shows that the multi-headed LSTM and single-headed LSTM have similar patterns of detection, but overall the multi-headed LSTM has less extreme probabilities (values closer to 0.5 than 0 or 1). This suggests that the global examples used to train the multi-headed LSTM act like a regularizer for the classifier by moderating the confidence of predictions and preventing overfitting.
We found that the \citet{CopernicusLandCover} map had the lowest performance for the test dataset, with an F1 score significantly lower than other methods, though we note that these results could underestimate the performance because the map reports the fraction of each pixel covered by crops, not probability, in each pixel. This lower performance may, in part, be due to the coarse spatial resolution of EO data used to produce the map and the growth in cropland area that has likely occurred since the map was produced in 2015. Similarly, the lower performance of the \citet{GFSAD} could also, in part, be due to lower spatial resolution of input data (30 m) and changes in cropland area that might have occurred between 2018 (when the most recent \citet{GFSAD} map was produced) and 2019.
\subsection{Sensitivity to number of local examples}
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{diagrams/auc_acc_mean_3.png}
\caption{A plot of AUC ROC and accuracy scores of the multi-headed model as the size of the hand-labeled dataset size increases. Each score is calculated from a mean of 3 runs with different random seeds.}
\label{fig:auc_from_size}
\end{figure}
To assess the sensitivity of the model to the frequency and number of local (Togo) examples in the training dataset and provide a guide for future labeling campaigns, we measured the performance of the multi-headed model on the test set as a function of the number of hand-labeled examples used in training (Figure \ref{fig:auc_from_size}). Although the performance of the model increased as the size of the hand-labeled dataset increased, the model performs well even with very little labeled data. This suggests this technique may be applied to other data-sparse regions with a comparable or fewer number of labels as we created for this study.
\subsection{Consensus in photointerpreted labels}
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth]{diagrams/consensus.pdf}
\caption{Tetrahedron showing number of examples (of 350 total) for which each pair of labelers (indicated by initials) agreed.}
\label{fig:consensus}
\end{figure}
As described in Section \ref{sec:results}, we created a test dataset by randomly sampling 350 points within Togo and having four experts label each point as ``crop'' or ``non-crop'' using photointerpretation of high-resolution satellite imagery. For each point, we chose the label assigned by the majority of labelers (ties were discarded). Figure \ref{fig:consensus} shows the number of examples for which each pair of labelers chose the same label. Agreement between pairs of labelers ranged from 229-278 (65\%-79\%) of the total points and all labelers agreed unanimously on 181 (52\%) of the points. This level of consensus was surprisingly low, and while our photointerpretation-based dataset is a high-quality surrogate for ground-truth labels, it underscores the difficulty for even experts to determine land cover types from high-resolution satellite imagery, particularly in developing countries. This should be considered in the design of future photointerpretation-based labeling efforts, especially crowdsourced efforts in which labelers may not be experts. We plan to conduct a future study to assess agreement between labels determined by experts for the same points from photointerpretation vs. field observation to better characterize the level of accuracy to expect from photointerpreted labels for different regions and land cover types.
\subsection{Lessons learned for rapid response}
The urgent need for a cropland map of Togo that captured smallholder farms (usually $<$1 ha) required us to leverage our research methods to rapidly deliver a result for operational use in less than 10 days. During this process, we gathered several ``lessons learned'' that we share to help prepare other researchers to leverage their methods for rapid humanitarian response in the future, as well as encourage community practices that facilitate rapid response.
\begin{itemize}
\item Good software engineering practices (such as developing modular and well-documented code) is often prioritized in operational rather than research settings. However, developing research methods motivated by their use for rapid humanitarian response in the future requires integration of operational practices into research workflows. We found that developing research methods using operational software engineering practices enabled us to quickly execute experiments and apply the method to large geographic areas when the need for rapid response arose.
\item Use of cloud services (in this case AWS) allowed us to easily scale our model (particularly during country-wide inference) and enable faster computation in a short period of time. Additionally, the use of EC2 instances enabled processes to be run, monitored, and modified by a distributed team rather than a single user.
\item File transfer/download remains a bottleneck for rapid response workflows. Transferring data from storage (e.g., via Google Earth Engine, the Sentinel-2 AWS S3 bucket, or USGS Earth Explorer) to compute (e.g., AWS) constituted a significant part of the time to deliver our cropland map. There is a community need for better tools to enable searching, querying, and subsetting of geospatial data on cloud platforms.
\item Given the complexity of deep learning architectures such as Inception-v3 \citep{Szegedy2016} or ResNet \citep{He2016} designed for image classification, it might seem that classification of land cover or crops in remote sensing data---which often exhibit high intraclass variance and low inter-class variance---would also require complex architectures. However, we found that a relatively simple, shallow architecture achieved good results for this task, and has the benefit of being faster in training/inference, requiring fewer examples, and being less susceptible to overfitting than more complex architectures.
\item The GFSAD \citep{GFSAD} and Copernicus \citep{CopernicusLandCover} cropland maps were easy to find and download and were well-documented, which enabled us to quickly benchmark our map against existing maps. To give others this same benefit with our cropland map, we have made our cropland map, code, and training labels publicly available.
\end{itemize}
\section{Conclusion}
Cropland maps are critical inputs for decision-makers to rapidly design and enact policies during humanitarian crises, e.g., to deliver aid or boost productivity for farmers, yet unavailability of high-resolution, up-to-date cropland maps in most countries globally precludes their use for rapid response. The development of cropland maps, in turn, is limited by the availability of ground data that can be used for training and validating machine learning classifiers.
We present a method for cropland classification in regions with little to no available ground data that uses a multi-headed LSTM network to learn global and local features for identifying cropland based on multispectral time-series Sentinel-2 observations. We used this method to create a cropland map of Togo at 10 m resolution in fewer than 10 days to assist the Togolese government in decision-making about aid distribution to farmers during the COVID-19 crisis. This demonstrated the successful transition of machine learning research to operational use in rapid response for a real humanitarian crisis, and we provided lessons learned from this experience to facilitate this transition for other researchers in the future. Additionally, our approach gives an example of how to leverage existing readily-available labeled datasets and reduce the number of data points to be collected in field campaigns, which can be costly as well as infeasible due to regional insecurity or travel restrictions (e.g., during the COVID-19 pandemic). In future work, we plan to investigate leveraging our multi-headed LSTM approach for crop type classification, subsetting the Geo-Wiki dataset based on geographical or agroecological zones, and including Sentinel-1 synethic aperture radar data in the model input to mitigate the effects of clouds during the growing season. Additionally, we will apply our method for cropland mapping in other data-sparse African countries.
\begin{acks}
We would like to thank Tanya Harrison, Zara Khan, and Charlie Candy from Planet, Inc., for making SkySat and PlanetScope basemaps available for this study. We are also grateful to Noel Gorelick from Google Earth Engine for helping to accelerate our Google Earth Engine processing workflow in order to meet this rapid response timeline. Finally, we would like to thank Steffen Fritz, Linda See, and the rest of the Geo-Wiki team for providing the publicly-available, global, crowdsourced cropland dataset critical to our method.
\end{acks}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
Satellite Earth observation (EO) data can provide critical, relevant, and timely information in support of agricultural monitoring and food security \citep{BeckerReshef2019,NationalAcademies2019}. EO can enable crop yield assessment and forecasting \citep{Becker2010,Franch2015,Franch2017,Lobell2013,Porter2017}, inform agricultural commodities markets, enable early warning and mitigation of impending crop shortages, and inform decisions about subsidies in times of crisis such as extreme drought \citep{Dorward2011} or outbreaks of disease or pests. However, the lack of up-to-date, high-resolution cropland and crop type maps is a critical barrier to agricultural and food security assessments and operational rapid response to humanitarian crises. These data are especially lacking in regions dominated by smallholder agriculture that are most vulnerable to food insecurity (e.g., sub-saharan Africa).
Machine learning techniques have commonly been used to identify cropland in EO data. The majority of studies use tree-based classifiers (primarily random forests or decision trees) or neural network/deep learning methods (primarily recurrent or convolutional neural networks). However, prior methods have typically been applied to small and/or spatially homogeneous areas \citep{RNN_Camargue, RNN_winter, Skakun2019}, require large training datasets (hundreds of thousands of labels) \citep{breizhcrops2020,Wang2020,lstm_crop_identification, RF_france}, or use EO inputs with insufficient spatial resolution for reliably detecting smallholder farms which often have field sizes $<$1 ha \citep{GFSAD,Teluguntla2018,Oliphant2019,CopernicusLandCover,Samasse2020,Waldner2017}. In many regions, particularly developing countries and regions dominated by smallholder farming, the lack of readily accessible ground truth data is a major barrier for training and validating machine learning techniques for cropland classification \citep{Coutu2020}.
In this paper, we demonstrate a novel method for rapidly generating cropland maps for areas with little to no available labeled data over a large heterogeneous area. This is achieved by leveraging global and local datasets. The global dataset contains tens of thousands of crowdsourced labels from diverse geographies \citep{GeoWiki} to learn general features for identifying cropland globally. These global examples supplement a much smaller dataset of curated, local examples (manually labeled using photointerpretation of high-resolution satellite imagery) to learn features specific to cropland in the target mapping region. Combined, these datasets provide sufficient examples to train a Long Short Term Memory network (LSTM) \citep{LSTM} to predict the presence of cropland in each pixel given a monthly time-series of Sentinel-2 multispectral observations.
The COVID-19 global pandemic has severely impacted the global food system, and the threat of food shortages and rising food prices has required countries at risk of food insecurity to rapidly design and enact aid programs in response. In the West African country of Togo, the government announced a program to boost national food production in response to the COVID-19 crisis by distributing aid to farmers, but found that high-resolution spatial information about the distribution of farms across Togo (which primarily consist of smallholder farms under 1 ha) was a critical piece of missing information for designing this program. Using our method, we were able to generate an accurate high-resolution (10 m) map of cropland in Togo for 2019 in under 10 days to respond to the Togolese government's immediate need for information about the distribution of farms across the country. We showed that our model outperforms existing publicly-available cropland maps for Togo \citep{GFSAD, CopernicusLandCover} while also being higher resolution and more recent. We have delivered this product to the Togolese government where it is being used to help inform decision-making in response to the COVID-19 pandemic, thus demonstrating a successful transition of machine learning research to operational rapid response for a real humanitarian crisis. We have made our map as well as all data and code used in this study publicly available to facilitate future research and operational systems in data-sparse regions.
\begin{figure}
\centering
\includegraphics[width=0.75\linewidth]{diagrams/label_distribution.pdf}
\caption{Distribution of labeled examples in training (a) and test (b) datasets.}
\label{fig:labels}
\end{figure}
\section{Methods}
\subsection{Labeled Data}
\subsubsection{Crowdsourced Labels}
We leveraged a large (35,866 labels) crowd-sourced dataset with labels of ``crop'' or ``non-crop'' from diverse, globally-distributed locations \citep{GeoWiki} to help the model learn features that are useful for detecting a wide range of cropland globally. Details about sample selection, labeling procedures, and quality assessment/control measures can be found in \citet{Bayas2017}. Each location was labelled by multiple labelers; to turn these labels into binary labels, we took the mean of the labels for each location, with a threshold of 0.5 to define a point as ``crop'' or ``non-crop.''
\subsubsection{Active Labeling}
To generate training labels, experts in photointerpretation of agricultural land cover (authors Becker-Reshef, Barker, Hosseini, Kerner, Munshell, and Paliyam) drew polygons over pixels with labels of ``crop'' or ``non-crop.'' We used a combination of basemaps for photointerpretation: a SkySat 72cm composite of January-March 2019 \citep{Planet}, PlanetScope composites of April-July and July-October 2019 \citep{Planet}, and Google Earth Pro basemaps (comprised primarily of high-resolution Maxar images). We used QGIS and Google Earth Pro to draw polygons over basemaps. The locations of labels were chosen to be distributed across different agroecological zones and non-crop land cover types observed across Togo. The central pixel within each polygon was used for training. We obtained 394 crop and 194 non-crop labels for the initial training dataset to supplement the 43 examples within Togo in the \citet{GeoWiki} dataset (631 examples total). After training the model with the hand-labeled and \citet{GeoWiki} dataset, we collected additional labeled examples (294 crop and 394 non-crop) by analyzing the predicted map and identifying areas of model confusion (e.g., transient vegetation on shorelines that the model might confuse with crops). This allowed us to maximize the utility of each training example, as it focused on places where the model was weakest rather than adding redundancy to the dataset. This resulted in a final training dataset of 1,319 hand-labeled examples, visualized spatially in Figure \ref{fig:labels}a.
\subsection{EO Data}
We used Copernicus Sentinel-2 surface reflectance (Level 2A) observations that corresponded to label locations as input to the model. We exported 160 m $\times$ 160 m patches for each labeled pixel using Google Earth Engine (GEE) \citep{GORELICK201718}, then extracted the closest pixel within the patch to the label location. To construct a cloud-free timeseries representation of each pixel, we used the algorithm from \citet{CloudFreeSentinel}, which finds the least cloudy pixel within a defined time period. We used 12 30-day time periods to construct a monthly timeseries spanning 360 days for each labeled pixel. We used all spectral bands except for B1 (coastal aerosol) and B10 (cirrus SWIR), and added the normalized difference vegetation index (NDVI), computed as $\text{NDVI} = \frac{\text{B08}-\text{B04}}{\text{B08}+\text{B04}}$. All bands (which range from 10 to 60m resolution) were upsampled to 10 m during GEE export. We used observations acquired between March of Year N and March of Year N+1 where N is the year the labels were created for. Thus for \citet{GeoWiki}, we used observations acquired March 2017-March 2018 and for our hand-labeled dataset we used observations acquired March 2019-March 2020.
\subsection{Model} \label{section:model}
We trained a one layer multi-headed LSTM model (Figure \ref{fig:lstm}) to predict whether a pixel contained cropland (1) or not (0). The input was expressed as a timeseries $\bm{X} = \{\bm{x}_{1}, \bm{x}_{2}, ... \bm{x}_{12}\}$, where each timestep $\bm{x}_{i} \in \mathbb{R}^{1\times14}$ consisted of the least-cloudy composite over a 30-day period of 11 optical bands (all bands except B01 and B10) plus NDVI. The LSTM hidden layer had 64 units. A classifier and a sigmoid activation were used on the final hidden output to map the hidden layer output (logits) to a value that can be interpreted as the posterior probability of crop presence in the pixel.
We performed a grid search that optimized AUC ROC on the validation set to choose values for two hyperparameters: number of linear layers in the classifier (1 or 2) and whether or not a dropout of 0.2 should be applied between each timestep in the LSTM.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{diagrams/model.png}
\caption{A multi-headed LSTM based model for pixel-wise crop identification. The model has two classification layers: a global linear classifier to classify instances outside Togo, and a local linear classifier to classify instances within Togo.}
\label{fig:lstm}
\end{figure}
To allow the model to focus on examples within Togo while also learning from the global \citet{GeoWiki} dataset, we employed a multi-headed LSTM model, where one classifier was used for all examples within Togo (local instances) and another was used for all other examples (global instances). All instances were shuffled during training: a batch could therefore contain a combination of global and local data instances. To combine the results from both classifiers, we used the following loss function to train the model:
\begin{equation}
\mathcal{L} = \frac{W}{\alpha}\mathcal{L}_{\textrm{global}} + \mathcal{L}_{\textrm{local}}
\end{equation}
Where $\mathcal{L}_{\textrm{global}}$ and $\mathcal{L}_{\textrm{local}}$ are the binary cross-entropy losses for the global and local data instances respectively, and $\alpha$ is a weighting parameter. $W$ is a batch-specific value, used to weight the ratio of global and local instances in a batch:
\begin{equation}
W = \frac{\mbox{Number of global instances in batch}}{\mbox{Number of local instances in batch}}
\end{equation}
In our experiments, we used $\alpha = 10$. We used the Adam optimizer with default parameters to tune weights during training \citep{Kingma2014}. To determine when to stop training, we used early stopping with a patience of 10 using a validation set consisting of 20\% of the combined \citet{GeoWiki} and hand-labeled Togo datasets. We implemented the model and experiments using PyTorch \citep{pytorch} in python. All code is publicly available at \url{https://github.com/nasaharvest/togo-crop-mask}.
\begin{table*}
\caption{Results for each method evaluated on the test set. The best results are in bold.}
\label{tab:experimental-results}
\begin{tabular}{cccccc}
\toprule
Description &Accuracy &AUC &Precision &Recall &F1\\
\midrule
\citet{CopernicusLandCover} (Copernicus 100m) & 0.67 & 0.74 & 0.54 & 0.24 & 0.33\\
Single-headed LSTM (hand-labeled only) & 0.71 & 0.90 & 0.59 & \textbf{0.93} & 0.72\\
\citet{GFSAD} (GFSAD 30m) & 0.74 & - & 0.62 & 0.62 & 0.62 \\
Support Vector Machine (hand-labeled only) & 0.79 & 0.88 & 0.65 & 0.84 & 0.73 \\
Random Forest (hand-labeled only) & 0.81 & 0.90 & 0.66 & 0.88 & \textbf{0.76} \\
Multi-headed LSTM (hand-labeled + \citet{GeoWiki}) & \textbf{0.83} & \textbf{0.91} & \textbf{0.81} & 0.68 & 0.74\\
\bottomrule
\end{tabular}
\label{experimental-setup-results}
\end{table*}
\section{Results}
\label{sec:results}
We selected points for a test dataset by randomly sampling 350 points within the Togo country boundaries, constrained by a minimum distance of 50 m between samples (Figure \ref{fig:labels}b). Four experts (authors Barker, Kerner, Munshell, and Nakalembe) labeled each example. Labels were determined by majority vote in order to mitigate label noise in the absence of groundtruth validation, and ties were discarded. The consensus test dataset contained 106 crop and 200 non-crop examples (306 total test examples).
We compared two experimental setups to evaluate the contribution of the \citet{GeoWiki} dataset:
\begin{itemize}
\item A single-headed LSTM trained on only the hand-labeled Togo data
\item A multi-headed LSTM trained on both the hand-labeled Togo and \citet{GeoWiki} data
\end{itemize}
In both cases, we used a model with hyperparameters determined by the grid search described in \ref{section:model}. For the single headed model, this consisted of a single classification layer in the classifier \textit{without} dropout in the LSTM. For the multi-headed model, this consisted of two classification layers in the classifier \textit{with} dropout in the LSTM.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{diagrams/togo_full.pdf}
\caption{Cropland maps of Togo generated using our method, \citet{CopernicusLandCover}, and \citet{GFSAD}. White boxes indicate locations of insets shown in Figure \ref{fig:comparisons}.}
\label{fig:fullmap}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.9\linewidth]{diagrams/comparisons.pdf}
\caption{Qualitative comparison of cropland maps produced using our multi-headed and single-headed LSTM methods, \citet{CopernicusLandCover}, and \citet{GFSAD} for three example regions, with SkySat 72 cm basemap (April-March composite) for reference. Locations of each inset indicated by white boxes in Figure \ref{fig:fullmap}.}
\label{fig:comparisons}
\end{figure*}
We compared our results to two publicly-available cropland maps of Togo. One map is Global Food Security-support Analysis Data (GFSAD) cropland extent map of Africa \citep{GFSAD}. This map has 30 m resolution and was produced using two pixel-based supervised classifiers, a Random Forest and a Support Vector Machine (SVM), and one object-based classifier, a Recursive Hierarchical Image Segmentation classifier. The GFSAD map does not provide cropland probabilities, but a discrete class assigned to each pixel of 0 (water), 1 (non-cropland), or 2 (cropland). The most recent GFSAD product is from 2015. Prior to our map, the GFSAD map is to our knowledge the highest resolution cropland map publicly available for Togo. Figure \ref{fig:fullmap} (right) shows the GFSAD land cover map for Togo. The second map we compared our results to is the Copernicus Land Cover cropland map \citep{CopernicusLandCover}. The Copernicus map specifies the fraction of each pixel covered by crops (using values 0-100) at 100 m resolution and was produced using a combination of random forests and expert rules with satellite data inputs from the PROBA-V sensor. The most recent Copernicus map is for 2018. To enable quantitative comparison with our results, we treated crop fraction as analogous to crop probability and computed all performance metrics using our test dataset (Table \ref{tab:experimental-results}); however, we emphasize that this may not be how the map was intended to be used and qualitative comparisons may be more faithful to the intended use of the \citet{CopernicusLandCover} map. Figure \ref{fig:fullmap} (center) shows the Copernicus crop fraction map for Togo. A direct comparison between our method and \citet{GFSAD} and \citet{CopernicusLandCover} would require running all three methods on the same data (i.e., 10m Sentinel-2 observations paired with our labeled examples). Since neither the code for these methods nor sufficient information for reproducing them faithfully are available, we additionally performed two baseline comparisons to Random Forest and SVM classifiers which are the basis for the \citet{GFSAD} and \citet{CopernicusLandCover} methods.
For the Random Forest, we used 100 estimators with no maximum depth (nodes were expanded until all leaves were pure, or until the leaves contained fewer than 2 samples). For the SVM, we used a radial basis function kernel. We implemented both methods using Scikit-learn in python \citep{scikit-learn}.
To assess performance, we computed the area under the receiver operating characteristic (ROC) curve (AUC score), accuracy (0.5 threshold on posterior probabilities), precision (user's accuracy), recall (producer's accuracy), and F1-score (harmonic mean between precision and recall). We did not compute AUC for GFSAD, which only provides the discrete predicted class. The results are reported in Table \ref{tab:experimental-results}. Figure \ref{fig:fullmap} (left) shows our predicted cropland map.
To qualitatively compare our method compared with \citet{GFSAD} and \citet{CopernicusLandCover}, we show each map compared to the SkySat 72cm basemap (true color) for several example locations in Figure \ref{fig:comparisons}. Each region in Figure \ref{fig:comparisons} contains primarily cropland, though the level of vegetation in non-crop areas varies in each image since Togo's climate ranges from dry savanna in the north to tropical in the south. While all maps detect most fields in the second row, the \citet{GFSAD} and \citet{CopernicusLandCover} maps failed to detect most of the fields in the first and third rows.
Our resulting map of cropland probabilities at 10 m resolution as well as all hand-labeled training and testing labels are publicly available at: \url{https://doi.org/10.5281/zenodo.3836629}.
\section{Discussion}
The multi-headed model trained on the combined hand-labeled and \citet{GeoWiki} dataset had the highest score for most performance metrics. While AUC score is a measure of model performance across a range of possible thresholds on posterior probability, the decision boundary when using the sigmoid function as the output activation layer is 0.5. While in practice a different threshold could be chosen, it is not clear how to construct a large and diverse enough validation set that could be used to reliably select an alternate threshold over very large, heterogeneous regions (e.g., the country of Togo). Thus, the improved accuracy and precision of the multi-headed model (using a threshold of 0.5) - in addition to its improved AUC score - made it preferable. Additionally, since the multi-headed model included training data from prior years (2017) and global instances of cropland \citep{GeoWiki}, we expect that this model would generalize better to potential domain shift in future years than the single-headed model trained with only one year of local data. Figure \ref{fig:comparisons} shows that the multi-headed LSTM and single-headed LSTM have similar patterns of detection, but overall the multi-headed LSTM has less extreme probabilities (values closer to 0.5 than 0 or 1). This suggests that the global examples used to train the multi-headed LSTM act like a regularizer for the classifier by moderating the confidence of predictions and preventing overfitting.
We found that the \citet{CopernicusLandCover} map had the lowest performance for the test dataset, with an F1 score significantly lower than other methods, though we note that these results could underestimate the performance because the map reports the fraction of each pixel covered by crops, not probability, in each pixel. This lower performance may, in part, be due to the coarse spatial resolution of EO data used to produce the map and the growth in cropland area that has likely occurred since the map was produced in 2015. Similarly, the lower performance of the \citet{GFSAD} could also, in part, be due to lower spatial resolution of input data (30 m) and changes in cropland area that might have occurred between 2018 (when the most recent \citet{GFSAD} map was produced) and 2019.
\subsection{Sensitivity to number of local examples}
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{diagrams/auc_acc_mean_3.png}
\caption{A plot of AUC ROC and accuracy scores of the multi-headed model as the size of the hand-labeled dataset size increases. Each score is calculated from a mean of 3 runs with different random seeds.}
\label{fig:auc_from_size}
\end{figure}
To assess the sensitivity of the model to the frequency and number of local (Togo) examples in the training dataset and provide a guide for future labeling campaigns, we measured the performance of the multi-headed model on the test set as a function of the number of hand-labeled examples used in training (Figure \ref{fig:auc_from_size}). Although the performance of the model increased as the size of the hand-labeled dataset increased, the model performs well even with very little labeled data. This suggests this technique may be applied to other data-sparse regions with a comparable or fewer number of labels as we created for this study.
\subsection{Consensus in photointerpreted labels}
\begin{figure}
\centering
\includegraphics[width=0.7\linewidth]{diagrams/consensus.pdf}
\caption{Tetrahedron showing number of examples (of 350 total) for which each pair of labelers (indicated by initials) agreed.}
\label{fig:consensus}
\end{figure}
As described in Section \ref{sec:results}, we created a test dataset by randomly sampling 350 points within Togo and having four experts label each point as ``crop'' or ``non-crop'' using photointerpretation of high-resolution satellite imagery. For each point, we chose the label assigned by the majority of labelers (ties were discarded). Figure \ref{fig:consensus} shows the number of examples for which each pair of labelers chose the same label. Agreement between pairs of labelers ranged from 229-278 (65\%-79\%) of the total points and all labelers agreed unanimously on 181 (52\%) of the points. This level of consensus was surprisingly low, and while our photointerpretation-based dataset is a high-quality surrogate for ground-truth labels, it underscores the difficulty for even experts to determine land cover types from high-resolution satellite imagery, particularly in developing countries. This should be considered in the design of future photointerpretation-based labeling efforts, especially crowdsourced efforts in which labelers may not be experts. We plan to conduct a future study to assess agreement between labels determined by experts for the same points from photointerpretation vs. field observation to better characterize the level of accuracy to expect from photointerpreted labels for different regions and land cover types.
\subsection{Lessons learned for rapid response}
The urgent need for a cropland map of Togo that captured smallholder farms (usually $<$1 ha) required us to leverage our research methods to rapidly deliver a result for operational use in less than 10 days. During this process, we gathered several ``lessons learned'' that we share to help prepare other researchers to leverage their methods for rapid humanitarian response in the future, as well as encourage community practices that facilitate rapid response.
\begin{itemize}
\item Good software engineering practices (such as developing modular and well-documented code) is often prioritized in operational rather than research settings. However, developing research methods motivated by their use for rapid humanitarian response in the future requires integration of operational practices into research workflows. We found that developing research methods using operational software engineering practices enabled us to quickly execute experiments and apply the method to large geographic areas when the need for rapid response arose.
\item Use of cloud services (in this case AWS) allowed us to easily scale our model (particularly during country-wide inference) and enable faster computation in a short period of time. Additionally, the use of EC2 instances enabled processes to be run, monitored, and modified by a distributed team rather than a single user.
\item File transfer/download remains a bottleneck for rapid response workflows. Transferring data from storage (e.g., via Google Earth Engine, the Sentinel-2 AWS S3 bucket, or USGS Earth Explorer) to compute (e.g., AWS) constituted a significant part of the time to deliver our cropland map. There is a community need for better tools to enable searching, querying, and subsetting of geospatial data on cloud platforms.
\item Given the complexity of deep learning architectures such as Inception-v3 \citep{Szegedy2016} or ResNet \citep{He2016} designed for image classification, it might seem that classification of land cover or crops in remote sensing data---which often exhibit high intraclass variance and low inter-class variance---would also require complex architectures. However, we found that a relatively simple, shallow architecture achieved good results for this task, and has the benefit of being faster in training/inference, requiring fewer examples, and being less susceptible to overfitting than more complex architectures.
\item The GFSAD \citep{GFSAD} and Copernicus \citep{CopernicusLandCover} cropland maps were easy to find and download and were well-documented, which enabled us to quickly benchmark our map against existing maps. To give others this same benefit with our cropland map, we have made our cropland map, code, and training labels publicly available.
\end{itemize}
\section{Conclusion}
Cropland maps are critical inputs for decision-makers to rapidly design and enact policies during humanitarian crises, e.g., to deliver aid or boost productivity for farmers, yet unavailability of high-resolution, up-to-date cropland maps in most countries globally precludes their use for rapid response. The development of cropland maps, in turn, is limited by the availability of ground data that can be used for training and validating machine learning classifiers.
We present a method for cropland classification in regions with little to no available ground data that uses a multi-headed LSTM network to learn global and local features for identifying cropland based on multispectral time-series Sentinel-2 observations. We used this method to create a cropland map of Togo at 10 m resolution in fewer than 10 days to assist the Togolese government in decision-making about aid distribution to farmers during the COVID-19 crisis. This demonstrated the successful transition of machine learning research to operational use in rapid response for a real humanitarian crisis, and we provided lessons learned from this experience to facilitate this transition for other researchers in the future. Additionally, our approach gives an example of how to leverage existing readily-available labeled datasets and reduce the number of data points to be collected in field campaigns, which can be costly as well as infeasible due to regional insecurity or travel restrictions (e.g., during the COVID-19 pandemic). In future work, we plan to investigate leveraging our multi-headed LSTM approach for crop type classification, subsetting the Geo-Wiki dataset based on geographical or agroecological zones, and including Sentinel-1 synethic aperture radar data in the model input to mitigate the effects of clouds during the growing season. Additionally, we will apply our method for cropland mapping in other data-sparse African countries.
\begin{acks}
We would like to thank Tanya Harrison, Zara Khan, and Charlie Candy from Planet, Inc., for making SkySat and PlanetScope basemaps available for this study. We are also grateful to Noel Gorelick from Google Earth Engine for helping to accelerate our Google Earth Engine processing workflow in order to meet this rapid response timeline. Finally, we would like to thank Steffen Fritz, Linda See, and the rest of the Geo-Wiki team for providing the publicly-available, global, crowdsourced cropland dataset critical to our method.
\end{acks}
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
\input{01_introduction}
\section{Related Work}
\input{02_related_work}
\section{FHIR Transformation}
\input{03_FHIR}
\section{Patient Diagnosis Prediction}
\input{04_prediction_model}
\section{Experiments and Results}
\input{05_experiments}
\section{Conclusion, Limitations, and Future Work}
\input{06_conclusion}
\begin{comment}
\begin{acks}
\end{acks}
\end{comment}
\bibliographystyle{ACM-Reference-Format}
\subsection{FHIR}
The FHIR format is a new format to exchange EHR data. The FHIR format represents data in containers that are consistent, scalable, and hierarchical. This facilitates the exchange of data between participants in health ecosystems. Moreover, the format does not ensure semantic consistency, which allows the format to be dynamically adjusted and implemented by any health system.
Every recorded event along a patient's admission is considered as a resource in the FHIR format. These resources contain multiple attributes. The ``Medication Dispense'' resource for example contains the trade name, the generic name or the medication ingredients as attributes. Each attribute is defined by a specific data structure and type, creating the hierarchical structure of the resources. These structures and requirements are explicitly documented on the FHIR website \cite{hl7_fhir}.
Each FHIR resource type belongs to a specific thematic category, which in turn belongs to a certain level depending on how sensitive and widespread the resource type is. The levels define how consistently the resource types are defined and how tightly they are governed. Resources types from the top levels are the most widespread and support the most common health care transactions.
\subsection{MIMIC-III to FHIR Transformation}
To be able to combine the MIMIC-III dataset with other EHR datasets in future applications, we mapped it to the FHIR format. Consequently, all primary MIMIC-III tables were converted to the appropriate FHIR resource type, if a corresponding FHIR resource type existed \footnote{No correspoding FHIR resource type were found for the 'callout', 'transfers' or 'drg\_codes' tables.}, see Table \ref{table:fhir_mapping} for the corresponding mapping.
\begin{table}[!ht]
\centering
\footnotesize
\begin{tabular}{|c l l|}
\hline
& MIMIC-III table & FHIR Resource Type \\ [0.5ex]
\hline
1 & patients & patient\\
2& admissions & encounter \\
3& diagnosis\_icd & encounter\\
4& icustay & enounter\\
5&cptevents & claim\\
6¬eevents & diagnosticReport\\
7&inputevents\_cv & medicationDispense\\
8&inputevents\_mv & medicationDispense\\
9&prescriptions & medicationRequest\\
10&chartevents & observation\\
11&datetimeevents & observation\\
12&labevents & observation\\
13&caregivers & practitioner\\
14&procedures\_icd & procedure\\
15&procedureevents\_mv & procedure\\
16µbiology & specimen\\
17&outputevents & specimen\\
18&service & serviceRequest\\
19&callout & \\
20&transfers & \\
21&drgcodes & \\
\hline
\end{tabular}
\label{table:fhir_mapping}
\caption{MIMIC-III Data Mapping to FHIR Resources}
\end{table}
The MIMIC-III data tables correspond to level 3 and 4 FHIR resource types. Instead of saving the FHIR resources as hierarchical containers, the resources are saved as single level containers that can easily be read into Python Pandas DataFrames.
The FHIR format allows resources to be saved in multiple file formats such as XML, JSON and Turtle. The proposed Python package saves the FHIR resources as JSON files, as these can be easily read by the Python programming language. Each MIMIC-III table was therefore converted to a collection of FHIR resource objects, saved as a GZIP compressed JSON file.
We provide a visual representation of the admission table entries mapped to the encounter resource type with greater details in the Appendix.
The python package can be downloaded from its online repository~\footnote{https://github.com/leopold-franz/MIMIC-III\_FHIR\_Transformation}. The next steps for its use are as follows: (i) import the transform function from the mimic\_fhir\_transformation.py, and (ii) call the transform function with these inputs: \textit{input\_path} of the original MIMIC-III table CSV file, \textit{output\_path} of where the collection of FHIR resources should be saved as a JSON file. Note that by adding the '.gz' extension the function can read compressed CSV files and save compressed JSON files. The function then saves the FHIR collection as a file and returns a dataframe with the flat hierarchy FHIR resources as rows. This allows the function call to be directly incorporated into any python pre-processing step. The MIMIC-III FHIR collections can then be combined with FHIR collections from other EHR datasets and used to train robust machine learning models.
\subsection{DeepObserver}
DeepObserver is a deep learning model inspired by \cite{kemp2019baselinemimic}. In this model, all numerical observations from the MIMIC-III Chartevents table, that contains vital sign measurements and laboratory results, are used for training instead of using only a few variables.
\subsubsection{Data \& Pre-processing}
The Chartevents table contains more than 330 million measurements. These measurements belong to different observation types, for example some Chartevent entries denote arterial blood pressure measurements and others denote body temperature data. Depending on the observation type, the measurements might not be numerical values, instead short strings or DateTime objects. As a first step of pre-processing, all values from non-numerical observation types were filtered out and only remaining 450 observation types with numerical values were used. Models in \cite{kemp2019baselinemimic}, which has similar pre-processing steps, also uses categorical string values converted into numeric embeddings. Next, the time difference between each measurement and the corresponding admission discharge time is computed. Subsequently, all observations are grouped together into 4 time interval \textit{bins} similar to \cite{kemp2019baselinemimic}. The last three bins correspond to the three consecutive 8h intervals before the discharge time point and all values before the last 24h of an admission are grouped together in the first bin, see Figure \ref{figure:deepobserver_binning}.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth]{figures/deepObserver_bins.png}
\caption{Grouping numerical observations into bins.}
\label{figure:deepobserver_binning}
\end{figure}
Measurements of the same observation type in the same bin are averaged together. The resulting data of a single admission corresponds to 450 numeric observation types each with 4 time bin values. All values of the same type are then normalized to Z-values.
The input for the DeepObserver model corresponds to an array of size (450,4) filled with normalized numerical observations, representing a patient admission. The model is trained on 29'590 admissions, as not every MIMIC-III patient admission has numerical entries in the Chartevents table.
DeepObserver is a supervised model that predicts multiple disease diagnoses.
Each admission representation is therefore passed to the model with corresponding disease diagnosis labels. There are two locations to record the admission diagnoses within the MIMIC-III dataset.
First location corresponds to a free-text entry in the Admission table and the second corresponds to the Diagnoses ICD table, where each admission is attributed with ICD-9-CM codes that denote the official diagnosis codes for hospital utilization in the United states. There exists 15'073 ICD-9-CM codes, of which the MIMIC-III dataset contains 6'984. Due to their high number, these ICD-9-CM codes are grouped and mapped to 281 meaningful CCS categories. "The Clinical Classifications Software (CCS) is a tool for clustering patient diagnoses and procedures into a manageable number of clinically meaningful categories." \cite{icd9_ccs_mapping}. There are 284 meaningful CCS diagnosis categories.
Each admission representation is labeled with an array of size 281. Every entry in this array is a Boolean item indicating whether the admission is diagnosed with an diagnosis code of the corresponding CCS category. A positive entry signifies that the patient is diagnosed with the corresponding disease category for the admission.
\subsubsection{Sampling} A challenge with working with medical data is that very often we face an unbalanced ratio of positive to negative labels. For example some rare CCS categories appear in less than 10 admissions of the MIMIC-III dataset. This means that at least 29'580 other admissions have a negative binary label for this CCS category. If a model uses all of the training data, it will therefore learn to predict a negative label with a high probability. One solution is to under-sample the majority class, which means that negative samples are excluded from the dataset until an equal number positive and negative samples remains, and then split the dataset into train (80\%), validation (10\%) and test (10\%) partitions. Unfortunately, an equal number of positive and negative samples cannot be reached for each label by under-sampling an extremely unbalanced multi-label dataset. Therefore an iterative train-test-split method \cite{multilabelstrat} is used that maintains the original label distribution in the train, validation and test sets. This method first calculates the label distribution in the original dataset and then calculates the label frequency needed in the desired partitions to maintain the label distribution. Next, in an iterative manner the sample with the rarest label in the original dataset is moved to the partition needing the rare label most to reach the required distribution.
\subsubsection{Model Architecture}
The DeepObserver model consists of a four layer neural network with supplementary dropout layers \cite{srivastava2014dropout} as regularization. The first layer reduces the admission representations dimension by learning patterns across the time dimension. Learning these time behaviours for each observation type is extremely important, and therefore different neural network layer architectures were evaluated for the first layer. The first architecture evaluated was a simple Fully Connected Neural Network (FCNN) layer. The second architecture tested was a Convolutional Neural Network (CNN) layer with filters of size (1,4) spanning across the time dimension. Finally, a simple Recurrent Neural Network (RNN) was evaluated, which is considered effective at learning patterns from sequences. Because this sequence is very short, no Long Short Term Memory (LSTM) or GRU (Gated Recurrent Unit) RNN architectures were considered. The second and third layer of the DeepObserver Network are all fully connected neural network layers of size 512. The last layer is a fully connected neural network layer with an output size equal to the number of CCS diagnosis categories to predict and with a sigmoid activation function that transforms the each output into label probabilities.
\subsubsection{Training}
The DeepObserver model was trained for three epochs with the Adam optimizer \cite{kingma2014adam}, a learning rate of $2e-5$ and a batch size of 32. Increasing the number of epochs did not improve the model. The hardware used for training and evaluating DeepObserver was a machine with one GPU having 11GiB of memory and 2 CPUs nodes with 40GB of RAM.
\subsection{ClinicalBERT\_Multi}
We develop an adaptation over the ClinicalBERT model \cite{clinicalbert} called ClinicalBERT\_Multi, which predicts disease diagnoses instead of readmission. The ClinicalBERT architecture is able to transform clinical notes into embeddings that can provide interpretable clinical insights. Similar to \cite{clinicalbert}, our adapted model is trained only using discharge summaries and clinical notes from the first 3 days of an admission. A model trained on notes from the first 3 days of admission can predict disease diagnoses early on in the admission, when the patient is still in the hospital. This early model is therefore evaluated on clinical notes from the first 3 and first 2 days of admission. Early predictions are exceedingly valuable in a clinical setting, as they allow patient treatments to be adjusted according to the predicted diagnoses while the patients are still in the hospital.
\subsubsection{Data Pre-processing}
\label{section:clinicalbert_preprocessing}
The clinical notes are first pre-processed, lower cased and tokenized. During pre-processing some key abbreviations are replaced (e.g., 'dr.' -> 'doctor'), and some superfluous characters are removed (e.g., the new line character '\textbackslash n').
The ClinicalBERT\_Multi model is not trained with all clinical notes that exist per admission. This differs from the DeepObserver model that does use all available numerical observations per admission. Instead, similar to the original ClinicalBERT \cite{clinicalbert}, the clinical notes dataset is split into three subsets are as follows:
\begin{enumerate}
\item[$D_{disch}$] contains all MIMIC-III discharge summaries. Discharge summaries contain a summary of the admission, a written form of the diagnosis, a list of prescribed medications and even summaries of other clinical notes, for example a radiology report summary. Each admission has only one discharge summary.
\item[$D_{3days}$] contains all MIMIC-III clinical notes except discharge summaries from the first three days of admission. Clinical Notes other than discharge summaries address a specific issue or service and can be more comprehensive.
\item[$D_{2days}$] contains all MIMIC-III clinical notes except discharge summaries from the first two days of admission. This subset is only used as an additional evaluation subset of the model trained on $D_{3days}$.
\end{enumerate}
$D_{disch}$ and $D_{3days}$ are then partitioned into 80\% train, 10\% validation and 10\% test sets using the same iterative train-test-split method \cite{multilabelstrat} as DeepObserver. One ClinicalBERT\_Multi model then trains on the train set of $D_{disch}$ and another ClinicalBERT\_Multi model is trained on the train partition of $D_{3days}$. A model trained to represent notes from the first three days of admission can also represent notes from the first two days of admission. It is therefore not necessary to train an extra model using only notes from the first two days. Consequently all clinical notes in $D_{2days}$, that also exist in the test partition of $D_{3days}$ are saved as a test partition of $D_{2days}$.The $D_{2days}$ subset is then only used as an additional test set of the model trained on $D_{3days}$. In Table \ref{table:subtasks}, information on the different subsets is shown.
All notes from $D_{3days}$ or $D_{2days}$ belonging to the same admission are concatenated together to form one long text string for each admission.
The ClinicalBERT architecture \cite{clinicalbert} can however only take in 512 inputs at a time. Therefore, long text sequences are split into multiple smaller text chunks. The text chunk strings are then converted to token sequences using the BERT tokenizer. The BERT tokenizer splits a string into word units and additionally splits rare or long word units into subword units. Furthermore, this tokenizer adds an extra token at the beginning of each sequence, which after being embedded by BERT can be used to classify the whole sequence.
Our models have a different task than the original ClinicalBERT model. The binary readmission labels are therefore transformed to diagnosis array labels, that are identical to the DeepObserver labels.
\subsubsection{Model Architecture}
The ClinicalBERT\_Multi model has the same base architecture as the original ClinicalBERT model, which consists of the standard BERT architecture with an extra fully connected neural network layer on top.
To understand BERT we must understand how attention and transformers work. Attention assigns weights to every input feature based on the features importance for the task at hand. By visualizing these weights, the learned importance of the input tokens can be shown. This ability makes the model output interpretable which is highy relevant in the clinical setting where clinicians demand models to be interpretative. Transformer encoders then use this attention to simultaneously process every element of a sequence into a sequence of deep embedding representations. BERT is based on 12 stacked transformer encoders \cite{vaswani2017attention} with 12 attention heads. The number of attention heads in BERT defines how many attention mechanisms are used per transformer, thereby defining how many filters are learned per transformer.
The last layer after BERT has an output size equal to the number of labels and a sigmoid activation function, which transforms each output into a probability. Consequently, the last layer acts as a classification layer. It is also where the ClinicalBERT\_Multi model differs from the original ClinicalBERT model, as the output size of ClinicalBERT\_Multi's classification layer is not equal to one but to the number of diagnosis categories to predict.
BERT now takes in a text chunk token sequence $\vec{T}$ as input and returns the embeddings of every token in the sequence. As mentioned in the pre-processing step a classification token is inserted at the beginning of every sequence. The last fully connected layer with weights $W$ then uses this embedded classification token $h_{\text{CLS}}$ output by BERT to compute a diagnosis probability $P$ for each CCS category, see equation \ref{equation:classification layer}.
\begin{equation}
P(\verb|CCS category 1| = 1 | h_{\text{CLS}}) = \sigma(Wh_{CLS})
\label{equation:classification layer}
\end{equation}
Each probability determines whether the admission will be diagnosed with a disease from the corresponding CCS category.
\begin{equation}
ClinicalBERT(\vec{T}) = \begin{bmatrix} P(\verb|CCS category 1| = 1 | h_{\text{CLS}}) \\ P(\verb|CCS category 2| = 1 | h_{\text{CLS}}) \\ ... \\ P(\verb|CCS category 281| = 1 | h_{\text{CLS}}) \end{bmatrix}
\label{equation:clinicalbert_input_output}
\end{equation}
Equation \ref{equation:clinicalbert_input_output} shows the ClinicalBERT\_Multi input token sequence $\vec{T}$ and text chunk probability array output.a
Multiple text chunks belong to the same admission, therefore the text chunk probabilities of the same admission and same CCS category are combined to a single admission CCS probability using Equation \ref{equation:chunk_probability_combination} from \cite{clinicalbert}, where $h_{admission}$ represents the admission representation.
\begin{equation}
P(\verb|CCS category x| | h_{admission}) = \frac{P^n_{max} + P^n_{mean}*n/c}{1 + n/c}
\label{equation:chunk_probability_combination}
\end{equation}
This calculated admission level diagnosis category probability improves on the individual disease category probabilities. $P^n_{max}$ is the maximum probability of all text chunks belonging to the same admission and is used to give additional weight to certain text chunks that have a high probability. Analogously $P^n_{mean}$ is the mean probability of all text chunks and is used to attenuate noise from a certain text chunks. $n$ represents the number of text chunks that belong to the corresponding admission. $c$ is a scaling factor that balances the weight given to the maximum probability and the mean probability. $c=2$ was used like \cite{clinicalbert}
Results presented in this paper are based on the admission level diagnosis category probabilities instead of on the text chunk diagnosis probabilities, as the admission level probabilities have shown to "consistently outperform predictions on each subsequence individually by 3-8\%"\cite{clinicalbert}. The resulting admission CCS probabilities can be converted into positive or negative predictions by setting a certain threshold for each CSS category. Any admission CCS probability over the CCS category threshold is converted to a positive prediction otherwise the admission CCS probability is converted to negative prediction. \footnote{Model performance can be evaluated without setting a threshold, which is why such a threshold was not chosen.}
\subsection{Training}
The pre-trained model from \cite{clinicalbert} was used and fine-tuned to the diagnosis prediction task for one epoch with the Adam optimizer \cite{kingma2014adam} using the standard learning rate of $2e-5$ and a batch size of 8.
The training and evaluation of the ClinicalBERT\_Multi model was done on a Google Cloud Platform Virtual Machine Instance was setup with 8 CPUs, each having 30GB of memory, 1 NVIDIA Tesla K80 GPU and a 128GB SSD harddisk.
\subsection{Experiments}
The first experiment aims to evaluate the best performing DeepObserver model. Therefore evaluating which architecture for the first layer works best. The different DeepObserver models are DeepObserver FCNN, DeepObserver CNN and DeepObserver RNN, where the second part of the name defines what architecture is used for the first layer.
The second experiment evaluates the performance of different models predicting a specific CCS category: CCS Category 98, which corresponds to Hypertension, also known as high blood pressure. This particular category was chosen as it is the most frequent CCS category in the MIMIC-III dataset, appearing 21'139 times. Results of this experiment are labelled with 'CCS Cat:98' in the task column of Table \ref{table:results}. The models evaluated in for this subtask are DeepObserver CNN, ClinicalBERT\_Multi and additionally ClinicalBERT\_Binary. ClinicalBERT\_Binary is a replica of the original ClinicalBERT model only the task is changed to predicting CCS category 98, which means that each binary readmission label is changed to a binary label defining whether the admission is diagnosed with the CCS category 98.
The goal of the third experiment is to compare the performance of the ClinicalBERT\_Binary and ClinicalBERT\_Multi when evaluated on the different data subsets $D_{disch}$, $D_{3days}$, $D_{2days}$. This experiment therefore evaluates how well the models perform when predicting the diagnosis categories at different admission time points, i.e. after 2 days of admission, after 3 days or after discharge.
The fourth and main experiment compares the predictive performance of all models predicting all diagnosis categories simultaneously. Therefore comparing ClinicalBERT\_Multi using Discharge Summaries, ClinicalBERT\_Multi using Clinical notes from the first 3 days, ClinicalBERT\_Multi using Clinical notes from the first 2 days, DeepObserver FCNN, DeepObserver CNN and DeepObserver RNN and the SHiP model on the multi-label prediction task.
\subsection{Results}
\label{section:results}
In Table \ref{table:results}, we present the evaluation of all models using AU-ROC, AU-PR, and Recall at $precision=80\%$ (Recall@Prec80). Note that the AU-ROC score is only representative of the predictive performance of models evaluated on balanced datasets such as the ClinicalBERT\_binary and the original ClinicalBERT. Therefore, to be able to compare all models, the AU-PR scores are compared, as well as the recall@prec80 scores. All multi-label classification models are marked with $ ^{b} $ to signify that the micro averaged scores are shown.
In the case of disease diagnosis support systems for multi-morbid patients, it is important to develop models with low miss rates. High recall scores are therefore required to make sure a few diseases are missed. Low precision scores are not very problematic, as high numbers of false alarms will be disregarded by doctors with a high precision rate.
The AU-PR score of a random classifier is equal to the ratio of positive samples in the evaluation dataset. The ratio of positive CCS Cat.:98 samples in the test partition of the Diagnoses task is 0.071. Therefore, if ClinicalBERT\_Multi and DeepObserver reach AU-PR scores above 0.071 in the CCS Cat.:98 , they outperform a random classifier. Meanwhile, the ClinicalBERT\_Binary model is tested with an equal amount of positive and negative samples, due to the original ClinicalBERT model under-sampling negative samples. Consequently, the AU-PR score of ClinicalBERT\_Binary model should be above 0.5 to outperform the random classifier. Note that the AU-PR scores of all models assessed on the CCS Cat.:98 task can still be compared. The ratio of positive labels in the Diagnoses task is 0.043, which means any Micro-AU-PR score in the Diagnoses task above this value outperforms a random classifier.
\begin{table}[!ht]
\centering
\begin{tabular}{l c c c c}
\hline
Model & Dataset & \# of Pat. & \# of Adm. & Avg. chunks \# \\ [0.5ex]
\hline
ClinicalBERT & $D_{disch}$ & 24'742 & 29'974 & 4.89\\
\_Binary& $D_{3days}$ & 25'103 & 30'476 & 8.54\\
& $D_{2days}$ & & & 6.28\\
ClinicalBERT & $D_{disch}$ & 27'238 & 33'684 & 4.64\\
\_Multi& $D_{3days}$ & 27'564 & 34'152 & 7.31\\
& $D_{2days}$ & & & 5.51\\
\hline
\end{tabular}
\caption{\footnotesize{Details of ClinicalBERT datasets. `\# of Pat.' and `\# of Adm.' show how many patients and admissions are in the train partition of the data subsets. `Avg. chunks \#' indicates the average number of text chunks per admission in the evaluation partition.}}
\label{table:subtasks}
\end{table}
\begin{table}[!ht]
\footnotesize
\centering
\begin{tabular}{l l l c c c}
\toprule
\textbf{Model} & \textbf{Task} & \textbf{$t_{pred}$} & \textbf{AU-ROC} & \textbf{AU-PR} & \textbf{Recall@} \\
& & & & & \textbf{Prec80} \\ [0.5ex]
\midrule
\textbf{SHiP $ ^{a} $} & \textit{Diagnoses} &\textit{Discharge} & \textit{0.897} & \textit{0.352} &\textit{ -} \\
\midrule
\textbf{DeepObserver} & Diagnoses & Discharge & 0.897$ ^{b} $ & 0.367$ ^{b} $ & 0.039$ ^{b} $\\
\hspace{8mm} \textbf{FCNN} & & & & & \\
\midrule
\textbf{DeepObserver} & Diagnoses & Discharge & 0.866$ ^{b} $ & 0.231$ ^{b} $ & 0.000$ ^{b} $\\
\hspace{8mm} \textbf{RNN} & & & & & \\
\midrule
\textbf{DeepObserver} & Diagnoses & Discharge & 0.90$ ^{b} $ & 0.372$ ^{b} $ & 0.039$ ^{b} $\\
\cmidrule(lr{1em}){2-6}
\hspace{8mm} \textbf{FCNN} & CCS Cat: 98 & Discharge & 0.699 & 0.585 & 0.001\\
\midrule
\textbf{ClinicalBERT\_} & CCS Cat.: 98 & Discharge & \textbf{0.873} & \textbf{0.862} & \textbf{0.803}\\
\cmidrule(lr{1em}){3-6}
\hspace{8mm} \textbf{Binary} & & 3 Days & 0.751 & 0.726 & 0.219\\
\cmidrule(lr{1em}){3-6}
& & 2 Days & 0.745 & 0.720 & 0.239\\
\midrule
\textbf{ClinicalBERT\_} & Diagnoses & Discharge & 0.919$^{b}$ & 0.408$^{b}$ & 0.08$1^{b}$\\
\cmidrule(lr{1em}){3-6}
\hspace{8mm} \textbf{Multi} & & 3 Days & \textbf{0.921}$ ^{b} $ & \textbf{0.426}$ ^{b} $ & \textbf{0.117}$ ^{b} $\\
\cmidrule(lr{1em}){3-6}
& & 2 Days & 0.708$ ^{b} $ & 0.202$ ^{b} $ & 0.061$ ^{b} $\\
\midrule
\textbf{ClinicalBERT\_} & CCS Cat.: 98 & Discharge & 0.739 & 0.626 & 0.032\\
\cmidrule(lr{1em}){3-6}
\hspace{8mm} \textbf{Multi} & & 3 Days & 0.723 & 0.613 & 0.082\\
\cmidrule(lr{1em}){3-6}
& & 2 Days & 0.666 & 0.558 & 0.003\\[1ex]
\bottomrule
\end{tabular}
\caption{\footnotesize{Predictive qualities of deep learning for disease diagnoses. The prediction time point is represented by $t_{pred}$. The \textbf{Bold values} are the values of the best performing model per task. $ ^{a} $ The Sequential, Hierarchical, and Pretrained Model results are retrieved from \cite{kemp2019baselinemimic} paper, which diagnoses multiple ICD-9 codes. $^{b}$ indicates that the values are micro averaged evaluation values. $ ^{c} $ The ClinicalBERT results are retrieved from the \cite{clinicalbert} paper. \textit{Results in italics are retrieved from the original papers}. Any non italicized values are results from models proposed in this paper.}}
\label{table:results}
\end{table}
Next, we summarize the results of the predictive performance of models and the results of the experiments.
All models proposed in this paper perform better than a random guess, as they all have an AU-PR score well above the AU-PR score of a random classifier. The smallest performance improvement compared to a random classifier for models evaluated on an imbalanced dataset is the ClinicalBERT\_Multi model. It predicts the disease diagnosis codes after \textbf{2} days of admission with a performance improvement of nearly 370\% (0.202/0.043 -1). The ClinicalBERT\_Binary model is the only model trained for this paper, which trains with an equal ratio of positive and negative samples. It shows significant improvement from the random classifier of 50\%.
The results of the first experiment show that the best performing DeepObserver model is the DeepObserver CNN model with a Micro-AU-PR score of 0.372. This indicates that the CNN filters are best at capturing time dependent patterns in these circumstances. The lower performance of the FCNN model is probably due to the higher number of trainable weights (500\% more weights). The lower performance of the DeepObserver RNN model can be explained by the low number of time steps along the time dimension the RNN can learn patterns from.
For the second experiment, we compare the models performances of models predicting the CCS Cat.:98 task. The best performing model is ClinicalBERT\_Binary predicting the diagnosis of the 98th CCS category at the Discharge time point. It achieves an AU-PR score of 0.862 and an exceptionally high recall@prec80 of 0.803. The ClinicalBERT\_Multi model trained to predict all diagnosis categories at Discharge is evaluated when predicting CCS Cat.:98 and achieves an AU-PR score of 0.626. This shows that fine-tuning the ClinicalBERT architecture on a balanced single diagnosis code dataset compared to on all diagnosis categories achieves a superior performance.
Results from the third experiment show that the prediction time point, which directly influences the amount of input data, has an effect on the predictive performance of models. The AU-PR scores of ClinicalBERT, ClinicalBERT\_Binary and ClinicalBERT\_Multi models improve when evaluating clinical notes of the first 3 days compared to evaluating clinical notes from the first 2 days. The difference in input data amount can be found in the 'Avg. chunks \#' column of Table \ref{table:subtasks}. All previously mentioned models show a higher AU-PR score when making predictions after 3 days of admission compared to when making predictions after 2 days of admission. Furthermore, ClinicalBERT\_Binary and ClinicalBERT\_Multi evaluated on task 'CCS Cat.: 98' have a better AU-PR score when using the discharge diagnoses than when using all clinical notes from the first 3 days of admission. Interestingly, the ClinicalBERT\_Multi sub-model predicting diagnoses after 3 days (AU-PR = 0.426) slightly outperforms ClinicalBERT\_Multi trained and evaluated with discharge summaries(AU-PR = 0.408). This could be due to the higher amount of input data when making predictions using clinical notes of the first 3 days of admission compared to making predictions using discharge summaries. The difference in input data is seen in column 'Avg. chunks \#' of table \ref{table:subtasks}.
The results of the fourth experiment show that the model with the highest predictive performance for the Diagnoses task is the ClinicalBERT\_Multi model predicting disease diagnosis categories using notes of the first 3 days of admission. It has an AU-PR score of 0.426. It slightly outperforms the DeepObserver CNN model making predictions after discharge (AU-PR = 0.372). This performance difference can be explained by the richer information in clinical notes when compared to the information in the numerical observations. Note that the high accuracy of the DeepObserver CNN model remains a significant achievement, as the data from the DeepObserver model is not human generated and does not depend on quality of clinical notes written by caregivers. Furthermore, comparing the results of the ClinicalBERT\_Multi model predicting CCS categories after 3 days with the results of the external SHiP model (AU-PR = 0.352) shows that ClinicalBERT\_Multi has a higher AU-PR score. The SHiP model differs from the CLinicalBERT\_Multi task by the following two points: (1) it uses all notes from an admission to make diagnosis predictions, which should help the SHiP model to achieve a higher predictive score, (2) the SHiP model predicts ICD-9-CM codes, which are harder to predict as there are more of them, instead of CCS categories.
As a summary, the DeepObserver CNN, which uses numerical observations, easily outperforms a random classifier when predicting multiple diagnosis categories simultaneously. Moreover, ClinicalBERT\_Multi, which uses clinical notes, has an even higher predictive performance than DeepObserver CNN when predicting multiple diagnosis categories and is can even provide good predictions when predicting the diagnoses early on in the admission.
\subsection{ClinicalBERT Interpretation}
By visualizing the self-attention heads from BERT, the diagnoses predictions of ClinicalBERT\_Binary and ClinicalBERT\_Multi can be interpreted. The attention function takes as input a key, a query and a value vector. Using these vectors it computes a distribution over all keys for each query and then multiplies these distribution weights with the values. In the case of self-attention the queries, keys and values are each constructed by multiplying the input tokens with a set of learned weights. The attention weight distribution for a query vector $q$ of length $ d $, and a set of keys $K$ each also of the length $ d $ is therefore:
\begin{equation}
\text{Attention}(Q, K, V) = \text{softmax}\left( \dfrac{QK^\top}{\sqrt{d}}\right)V
\end{equation}
The self-attention heads therefore act as filters over the input token sequence.
\begin{figure}[!ht]
\centering
\includegraphics[width=0.45\textwidth,keepaspectratio]{figures/one_attention_heads_diagnosis_98_shortInput.png}
\caption{Self-attention weights aligned to key and query tokens from the ClinicalBERT\_Binary model trained on discharge summaries.}
\label{figure:attention_short_input}
\end{figure}
An attention weight with a high value signifies that the interaction between the query and key token is predictive about the task at hand. In Figure \ref{figure:attention_short_input}, the weights of a single self-attention head are visualized. As you can see a the failure token is given a high weight.
Being able to interpret the results from ClinicalBERT\_Multi sub-models is important to gain an understanding of the model predictions and better informing medical practitioner's decision-making.
|
\section{INTRODUCTION}
Behavioral economists have argued for years that the traditional model of \emph{homo economicus}---%
an
agent who is always rational and behaves optimally---is misguided.
There is a lot of experimental work backing up their claims
(see, e.g., \cite{Thaler15}). Recent work has argued that perhaps the
behavior that we observe can best be explained by thinking of agents
as rational (i.e., trying to behave optimally), but not able to due to
computational limitations; that is, they are doing the best they can,
given their computational limitations.
In this paper, following a tradition that goes back Rubinstein
\citeyear{rub85} and Neyman \citeyear{ney85}, we model computationally
bounded agents as probabilistic finite automata (PFAs). We can think
of the number of states of the automaton as a proxy for how
computationally bounded the agent is. Neyman \citeyear{ney85} showed
that cooperation can arise if PFAs play a finitely-repeated prisoner's
dilemma; work on this topic has continued to attract attention (see
Papadimitriou and Yannakakis \citeyear{PY94} and the references
therein). Wilson \citeyear{W02} considered a decision problem where an
agent must decide whether nature is in state 0 or state 1, after
getting signals that are correlated with nature's state. She
characterized an optimal $n$-state PFA for making this decision, and
showed that it exhibited ``human-like'' behavior; specifically, it
ignored evidence (something a Bayesian would never do), and exhibited
what could be viewed as a first-impression bias and confirmation bias.
Halpern, Pass, and Seeman \citeyear{HPS12} considered a similar
problem in a dynamic setting, where the state of nature could change
(slowly) over time. Again, they showed that a simple PFA both
performed well and exhibited the kind of behavior humans exhibited in
games studied by Erev, Ert, and Roth \citeyear{EER10}.
We continue this line of work, and try to understand
the behavior of computationally bounded agents playing a multi-armed
bandit (e.g., playing slot machines in Las Vegas). Our first step in
doing this is to understand the extent to which optimal play can be
approximated by a PFA without worrying about the number of states
used. There are a number of notions of optimal that we could
consider. Here we focus on arguably the simplest one: we compare the
expected average payoff of the automaton after it runs for $N$ steps
to the expected average payoff of always pulling the
optimal arm of the bandit.
We also assume that the possible payoff of each arm is either 1 or 0
(i.e. success or failure), so the expected payoff of an arm is just
the probability of getting a 1.
There are well-known protocols that use
Bayesian methods (e.g., \emph{Thompson Sampling} \cite{Thompson33})
that approach optimal play in the limit; however, these approaches are
computationally expensive. We show that they have to be. No approach
that can be implemented by a PFA can perform optimally. Indeed, for
all PFAs, there exists an $\epsilon>0$ such that as the number of
steps gets large, the ratio of the expected payoff of the automaton to
the expected payoff of the optimal arm is at most $1-\epsilon$. That
is, a PFA must be off by some $\epsilon > 0$ from optimal play
(although we can make $\epsilon$ as small as we like by allowing
sufficiently many states). Among families of finite automata that
have near-optimal payoff, we are interested in ones that (a) make
efficient use of their states (so, for a fixed number $M$ of states,
have high expected payoff), (b) converge to near-optimal behavior
quickly, and (c) use simple ``human-like'' heuristics.
A standard approach to dealing with multi-armed bandit problem is one
we call \emph{explore-then-exploit}. We simply test each arm $N$ times
(where $N$ is a parameter), and from then on play the best arm (i.e.,
the one with the highest average reward). If the bandit has $K$ arms,
then we need roughly $O(N^K\log(N)\log(K))$ states, since we need to
keep track of
the possible tuples of outcomes of the tests, as well as two counters,
one to keep track of which arm is being tested, and the other to keep
track of how many times we have played it.
We can greatly reduce the
number of states by essentially using an elimination tournament.
We first compare arm 1 against arm 2, eliminate the worse arm, run the
winner against arm 3, eliminate the worse arm, run the winner against
arm 4, and so on. The way we compare arm $i$ and $j$ is straightforward:
we alternate playing $i$ and $j$ and use a counter to keep track of
the relative number of successes of $i$. If the counter hits an
appropriate threshold $M$ (so that $i$ has had $M$ more successes than
$j$), $i$ is the winner; if the counter hits $-M$, then $j$ is the winner.
To do this, we need ${K \choose 2}2(2M+1) \sim 2K^2M$ states: we need to keep
track of which arms are being played, which arm is currently moving,
and the counter. In choosing $M$, we need to balance out the desire
not to mistakenly eliminate a good arm (which is more likely to
happen the smaller that $M$ is) with the desire not to ``waste'' too much
time in finding the right arm (since the payoff while we are doing
that may not be so high, particularly if we are playing two arms whose
success probabilities are equal but not very high).
We deal with this by stopping a comparison after an expected number
$N$ of steps.
(We implement this by stopping
the comparison
with probability $1/N$, which does not
require any extra states.)
As we shall see, this approach, which we call the
\emph{elimination tournament}, does extremely well.
The \emph{$\epsilon$-greedy} protocol is a slight variant of this
approach: Again, we test for the first $N$ steps, and then play the
current best arm with probability $1-\epsilon$ and a random arm with
probability $\epsilon$.
But this requires infinitely many states,
since we must keep track of the fraction of successes for all arms to
determine the current best arm.
Clearly neither approach is optimal. With positive probability, both
explore-then-exploit and the
elimination-tournament protocols will choose a non-optimal arm; from
then on it is not
getting the optimal reward. The $\epsilon$-greedy protocol gets a
non-optimal reward with (roughly) probability $\epsilon$. While we can
make all these approaches arbitrarily close to optimal by choosing the
parameters $N$, $\epsilon$, and $M$ appropriately,
they do not satisfy our
third criterion: they don't seem to be what people are doing.
The $\epsilon$-greedy approach and explore-then-exploit require
an agent to keep track of large amounts of information, while the
elimination tournament alternates between arms at every
step, which may have nontrivial costs.
(Imagine a gambler in Las Vegas who wants to compare two arms that are
at opposite ends of a large room. Will he really walk back and forth?)
We instead consider an approach that takes as its starting point
earlier work by Rao
\citeyear{Rao17}, who considered only two-armed bandits, where,
just as for us,
each arm has a payoff in $\{0,1\}$.
She defined a family of PFAs
that act like ``approximate Bayesians''. More precisely, each arm has
an associated rank that represents a coarse estimate of the arm's
payoff probability.
Rao plays the arms repeatedly (using complicated rules to
determine which arm to play next) in order to estimate
the success probability of each arm, and then chooses the best arm.
While we use ranks, we use them in a very different way from Rao. We
take as our inspiration Simon's notion of \emph{satisficing}
\cite{Simon56}. The idea is that an arm will be accepted if its
success probability is above some threshold. In the words of
Gigerenzer and Gaissmaier \citeyear{GG15}: ``Set an aspiration
level, search through alternatives sequentially, and stop search as soon
as an alternative is found that satisfies the level.''
(We remark that the importance of the aspiration level goes back to
the 1930s in the psychology literature, and has been studied at length
since then; see, e.g., the highly-cited work of Lewin et al. \citeyear{LDFS44}.)
But how do we determine the aspiration level? This is a nontrivial
issue. Selten \citeyear{Selten98} and Simon \citeyear{Simon82}
(both Nobel prize winners) discuss this issue at length. As
Gigerenzer and Gaissmaier \citeyear{GG15} observe, ``The aspiration
level need not be fixed, but can be dynamically adjusted to feedback.''
In our setting, it is relatively straightforward: we use an \emph{optimism
bias} \cite{Sharot11}. We start with a high aspiration level
(success probability) $p$, and run a tournament as above between each arm
$k$ and a ``virtual arm'' that has success probability $p$. Since
this is a virtual arm, we are essentially comparing the
performance of each arm $k$ to our expectation. If arm $k$ does not
meet our expectation, then we go to the next arm. If no arm meets our
expectation, we adjust the aspiration level according to this feedback,
by lowering it.
This requires $KMm$ states, where $M$, as before, is the counter used
to keep track of the relative performance of the arm being tested and
$m$ is the number of ranks. We call this the
\emph{aspiration-level} approach.
We get good performance by taking $m \sim K$, so the aspiration-level
approach uses essentially the same
number of states as the elimination tournament. Moreover, as
we show by simulation, its performance approach degrades
gracefully as the number of states decreases. Even with
relatively few states, it compares quite favorably to the
$\epsilon$-greedy approach and to Thompson Sampling, although they
require infinitely many states. More importantly from our perspective,
the aspiration-level approach is quite human-like.
We have already mentioned how it incorporates satisficing, the
adjustment of expectations according to feedback, and an optimism
bias. But there is more. Whereas the elimination-tournament approach
treats the two arms that it is comparing
symmetrically, the aspiration-level approach does not.
If the virtual arm wins, it just
means that we try another arm. Moreover, especially initially, we
expect the virtual arm to win because we start out with a high
aspiration level. On the other hand, if an actual arm wins, that is
the arm we use from then on. Thus, we want to be relatively quick to
reject an arm, and slow to accept. This can be be viewed as a
\emph{negativity bias} \cite{KH72}: negative outcomes have a greater
effect than positive outcomes. The focus on
recent behavior can be viewed as implementing an \emph{availability
heuristic} \cite{TK73}: people tend to heavily weight their judgments
toward more recent or available information. Finally, a short run of
good luck can have a significant influence, causing an arm to be
played for a long time (or even played forever, if it is enough to get
it accepted). People are well-known to label some arms as ``lucky''
and keep playing them long after the evidence has indicated
otherwise.
This can also be viewed as an instance of the \emph{status
quo bias} \cite{SZ88}: people are much more likely to stick with the
current state of affairs (provided they think it is reasonably good).
\section{MULTI-ARMED BANDITS}
This section provides the necessary background for the rest of the paper. In particular, we (1) briefly review multi-armed bandits, (2) define the notion of optimality we consider, and (3) prove that a PFA cannot be optimal.
\subsection{THE MULTI-ARMED BANDIT PROBLEM}
The multi-armed bandit (MAB) problem is a standard way of modeling the
tradeoff between exploitation and exploration. An agent has $K$ arms
that she can pull. Each arm offers a set of possible rewards, each
obtained with some probability. The agent does not know the
probabilities in advance, but can learn them by playing the arm
sufficiently often. Formally, a \emph{$K$-armed bandit} is a tuple
distribution over rewards for arm \(k\). Let $\mu_k$ be the expected
reward of arm \(i\), for $i=k,\ldots, K$. The best expected reward of
$B$ is denoted \(\mu^*_B = \max_k\{\mu_k\}\).
We assume for simplicity in this paper that the possible rewards of an
arm are either 0 or 1. With this assumption, $\mu_k$ is the
probability of getting a 1 with arm $k$. We can easily modify the
protocol to deal with a finite set of possible rewards, as long as the
set of possible rewards is known in advance. We also assume for now
that the distributions $R_k$ do not vary over time.
\subsection{OPTIMAL PROTOCOLS FOR MAB PROBLEMS}
We are interested in protocols that play MABs (almost)
optimally. Formally, a \emph{protocol} is a (possibly randomized)
function from history to actions. We focus on one particular simple
notion of optimality here, which informally amounts to approaching the
average reward of the best arm. To make this precise, given a protocol
$P$, let $a_t^{P,B}$ be a random variable that denotes the arm played
by protocol $P$ at the $t$th step. Thus, $\mu_{a_t^{P,B}}$ is the
expected reward of arm $a_t^{P,B}$. It is easy to see that the
expected cumulative reward of protocol $P$
when run for $N$ steps on MAB $B$ is $Cum(P,B, N) =
\sum_{t=1}^N\mathbb{E}[\mu_{a^{P,B}_t}]$. Since the reward for playing
the optimal arm of MAB $B$ for $N$ steps is $N\mu^*_B$, the
\emph{expected regret} is the difference between the cumulative reward
of $P$ and the optimal reward: $Reg(P,B,N) = N \mu^*_B -
Cum(P,B,N)$. Finally, the \emph{average $N$-step regret} of $P$ on $B$
is $AReg(P,B,N) = Reg(P,B,N)/N$. We say that $P$ is \emph{optimal} if
$\lim_{N \rightarrow \infty} AReg(P,B,N) = 0$ for all MABs $B$.
As we observed in the introduction, neither
explore-then-exploit nor the $\epsilon$-greedy protocol is
optimal in this
sense. There are Bayesian approaches that are optimal. We briefly
discuss one: Thompson Sampling \cite{Thompson33}. Roughly speaking,
at each step, this protocol computes the probability of each arm being
optimal, given the observations. It then chooses arm $k$ with a
probability proportional to its current estimate that $k$ is the
optimal arm. It is not hard to show that, with probability 1, the
probability of a non-optimal arm being chosen goes to 0. (By way of
contrast, the probability of a non-optimal arm being chosen at any
given step with the $\epsilon$-greedy protocol is a constant: at least
$(K-1)\epsilon/K$, if there are $K$ arms.)
As shown by Kaufman, Kordan, and Munos \citeyear{KKM12}, Thompson
Sampling is optimal in an even stronger sense than what we have
considered so far. Taking TS to denote Thompson Sampling, not only do
we have $\lim_{N\rightarrow \infty} Reg(TS,B,N)/N = 0$, but there is a
constant $c^*_B$ (that depends on the MAB $B$, but has been completely
characterized) such that $\lim_{N\rightarrow \infty} Reg(TS,N)/\log(N)
= c^*_B$. Moreover, this is optimal; as shown by Lai and Robbins
\citeyear{LR85}, for all protocols $P$ satisfying a minimal technical
condition, we must have $\lim_{N\rightarrow \infty} G(P,B,N)/\log(N)
\ge c^*_B$. That means that Thompson Sampling approaches optimal
behavior as quickly as possible, and its cumulative regret grows only
logarithmically. We mention this because we will be comparing the
performance of our approach to that of Thompson Sampling later.
\subsection{PROBABILISTIC FINITE AUTOMATA AND NON-OPTIMALITY}
As we said in the introduction, we are interested in resource-bounded agents playing MABs, and we model resource-boundedness using PFAs. A PFA is just like a deterministic finite automaton, except that the transitions are probabilistic. We also want our automata to produce an output (an arm to pull, or no arm), rather than accepting a language, so, technically, we are looking at what have been called \emph{probabilistic finite automata with output} or \emph{probabilistic transducers}. (This is also the case for all the earlier papers that considered PFAs playing games or making decisions, such as \cite{HPS12,PY94,rub85,W02}.) Formally, a PFA with output is a tuple \((Q, q_0, \Sigma, O, \gamma, \delta)\), where
\begin{itemize}
\item $Q$ is a finite set of \emph{states};
\item $q_0 \in Q$ is the initial state;
\item $\Sigma$ is the input alphabet (in our case this will
consist of the observations ``arm $k$ had reward $j$'' for $j
\in\{0,1\}$);
\item $O$ is the output alphabet (in our case this will be
``$k$'', which is interpreted as playing arm $k$, for $k \in\{ 1,
\ldots, K\}$);
\item $\gamma: Q \rightarrow \Delta(O)$ is a probabilistic action function (as usual, $\Delta(X)$ denotes the set of probability distributions on $X$);
\item $\delta : Q \times \Sigma \rightarrow \Delta(Q)$ is a probabilistic transition function.
\end{itemize}
Intuitively, the automaton starts in state $q_0$ and plays an arm according to distribution $\gamma(q_0)$. It then observes the outcome $o$ of pulling the arm (an element of $\Sigma$) and then transitions to a state $q'$ (according to $\delta(q_0,o)$). It then plays arm $\gamma(q')$, and so on.
It is easy to see that the explore-then-exploit protocol can be
implemented by a finite automaton. On the other hand, the
$\epsilon$-greedy protocol and Thompson Sampling cannot. That is
because they keep track of the total number of times each arm $k$ was
played, and the fraction of those times that a reward of 1 was
obtained with $k$. This requires infinitely many states.
We claim that no protocol implemented by a PFA can be optimal. To prove this, we need some definitions.
\begin{definition}
A $K$-arm MAB $B = (\mu_1, \ldots, \mu_K)$ is \emph{generic} if (1) $\mu_B^* < 1$, (2) $\mu_i \ne \mu_j$ for $i \ne j$, and (3) if $K=2$, then $\min(\mu_1,\mu_2) > 0$.
\end{definition}
Note the if we put the obvious uniform distribution on the set of $K$-armed bandits (identifying a $K$-armed bandit with a $K$-vector of real numbers), then the set of generic MABs has probability 1.
\begin{definition}
$B' = (\mu_1', \ldots, \mu_K')$ is a \emph{permutation} of $B =
(\mu_1, \ldots, \mu_K)$ if there is some permutation $\rho$ of the
indices such that $\mu_k = \mu'_{\rho(k)}$.
\end{definition}
\begin{theorem}\label{thm:non-optimal} For all PFAs $M$ and all generic MABs $B$, there exists some $\epsilon_{M,B} > 0$ (that, as the notation suggests, depends on both $M$ and $B$) and an MAB $B'$ that is a permutation of $B$ such that $\lim_{N\rightarrow \infty} Reg(M,B',N)/N \ge \epsilon_{M,B}$.
\end{theorem}
Before giving the proof, we can explain why we must consider generic
MABs and permutations. To understand why we consider permutations,
suppose that $M$ always plays arm 1. If it so happens that arm 1 is
the best arm for $B$, then $M$ gets the optimal reward with input $B$.
But it will not get the optimal reward for a permutation of $B$ for
which arm 1 is not the best arm. It is not hard to see that if
$\mu^*_B =1$, then there exists a PFA $M$ that gets the optimal reward
given input $B$ or any of its permutations: $M$ just plays an arm
until it does not get a payoff of 1, then goes on to the next arm.
Sooner or later $M$ will play an arm that always gets a reward of 1. A
similar PFA also gets the optimal reward if $K=2$ given an input $B =
(\mu_1,\mu_2)$ such that $\mu_k = 0$ for some arm $k$: it alternates
between the arms until it finds an arm that gives reward 1, and sticks
with that arm. Finally, if $\mu_1 = \cdots = \mu_K$, then no matter
what arm $M$ plays, it will get the optimal reward on $B$ and all of
its permutations. The requirement that all $\mu_k$s are distinct is
actually stronger than we need, but since slight perturbations of the
rewards of an arm suffice to make all rewards distinct, we use it here
for simplicity.
\begin{proof}
Given a PFA $M$ and a nontrivial MAB $B$, there are two possibilities:
(1) there is some state $q$ that can be reached from the start state
$q_0$ with positive probability and an arm $k$ such that, after
reaching state $q$,
$M$ plays arm $k$ from then on, no matter what it observes; (2)
there is no such state $q$. Note that the first case is what happens
with explore-then-exploit. After the exploration phase, the
same arm is played over and over. The second case is more like
Thompson Sampling or $\epsilon$-greedy; there is always some positive
probability that a given arm $k$ will be played.
For case (1), let $o_1, \ldots, o_T$ be a sequence of observations
that, with positive probability, leads $M$ to a state $q$ after which
it always plays arm $k$. If the arm that $M$ plays in state $q$ is not
the best arm of $B$, let $\delta = \mu^*_B - \mu_k$, and let
$\delta_{M,B}$ be the probability with which $o_1, \ldots, o_T$ is
observed when running $M$ on input $B$. Clearly, $\lim_{N\rightarrow
\infty} Reg(M,B,N)/N \ge \delta \delta_{M,B}$. And if $\mu_k =
\mu^*$, consider a permutation $B' = (\mu_1', \ldots, \mu_K')$ such
that $\mu_j' = 0$ if and only if $\mu_j = 0$ (i.e., the permutation is
the identity on all arms $j$ such that $\mu_j = 0$) such that $\mu_k'
\ne \mu^*_{B'} = \mu^*_B$. It is still the case that $o_1, \ldots,
o_T$ can be observed with some positive probability $\delta_{M,B'}$
when running $M$ on input $B'$. Taking $\delta' = \mu^*_B - \mu_j$,
we have $\lim_{N\rightarrow \infty} Reg(M,B,N)/N \ge \delta'
\delta_{M,B'}$.
For case (2), no matter what state $q$ $M$ is in, with some probability $\epsilon_q > 0$, $M$ plays a non-optimal arm at $q$ or moves to another state $q'$ and plays a non-optimal arm there. Let $\epsilon_M^* = \min_q \epsilon_q$. Since $M$ has only finitely many states, $\epsilon^* > 0$. Given as input an MAB $B$, let $\delta_B$ be the difference between the $\mu^*_B$ and the probability that the second-best arm returns 1. (Here we are using the fact that all arms have different probabilities of returning 1.) Let $X_{M,B,T}$ be a random variable that represents the reward received on the $T$th step that $M$ is run on input $B$. Our discussion shows that, for all $T$, we must have $E(X_T + X_{T+1}) \le 2\mu^*_B - \epsilon^*_M\delta_B$, since with probability at least $\epsilon^*_M$, one of $X_T$ or $X_{T+1}$ is at least $\delta_B$ less than $\mu^*_B$. Since $Reg(M,B,2N)$ = $2N\mu^*_B - \sum_{T=1}^{2N} X_{M,B,T} \ge 2N\mu^*_B - N(2\mu^*_B - \epsilon^*_N\delta_B)$, it follows that $Reg(M,B,2N)/2N \ge \epsilon^*_B\delta_B/2$. This gives us the desired result.
\end{proof}
\section{AN ALMOST-OPTIMAL FAMILY OF PFAS FOR MAB PROBLEMS}
In this section, we introduce the aspiration-level protocol more formally.
We start by reviewing
Rao's \citeyear{Rao17} approach to dealing with 2-armed bandits, since
our approach uses some of the same ideas.
\subsection{RAO'S APPROACH}
With only finitely many states, a PFA cannot keep track of the exact
success rate of each arm in an MAB. Thus, it needs to keep a finite
representation of the success rate. Rao's idea was to use a finite set
of possible ranks to encode the agent's belief about the relative
goodness of each arm. There are $m$ possible ranks, $\{1,\ldots,
m\}$, where $m$ is a parameter of the protocol. Thus, Rao's PFA has
$m^2$ possible states, which have the form $(r_1,r_2)$ (since Rao
considers only 2-armed bandits), where $r_1, r_2 \in \{1,\ldots,
m\}$.
Rao assumes that the initial state of the PFA has the form $(n,n)$ for some $n \in \{1,\ldots,m\}$; the exact choice does not matter. Thus, initially, the two arms are assumed to be equally good. Of course, if an agent has some prior reason to believe that one arm is better than the other, then the initial state can encode this belief.
The action function $\gamma$ is defined as follows: If the
higher-ranked arm has the highest possible rank ($m$) and the other
arm does not, then the higher-ranked arm is played. Otherwise,
similar in spirit to Thompson Sampling, the next arm to play is chosen
according to a probability that depends on the difference between the
ranks of the arms ($|r_1 - r_2|$) and how far the arm's ranks are from
average ($|r_1 - m/2| + |r_2 - m/2|$). The two numbers are then
combined using two further parameters (called $\alpha$ and $C_\alpha$
by Rao) of the protocol. We refer the reader to \cite{Rao17} for the
technical detail and intuition.
Finally, the transition function $\delta$ is defined as follows: the rank of the arm last played goes up with some probability (if it is not already $m$) if a payoff of 1 is observed and goes down with some probability (if it is not already 1) if a payoff of 0 is observed. The rank of the arm not played does not changed. The exact probability of a state change depends on a quantity that Rao calls the \emph{inertia}, which is determined by the ranks of the arms, and two other parameters of the protocol, called $\beta$ and $C_t$ by Rao. Intuitively, the inertia characterizes the resistance to a change in rank. The less frequently an arm has been played, the higher its associated inertia will be, so its rank is updated with a lower probability. Again, we refer the reader to \cite{Rao17} for details.
\subsection{THE ASPIRATION-LEVEL PROTOCOL}
We want to define a family of PFAs for $K$-armed MABs. We continue to
use Rao's idea of associating with each arm a rank. The naive
extension would thus require $O(m^K)$ states. For large $K$, this is
quite unreasonable.
So we assume that the PFA focuses only one arm at a time, comparing it
to a ``virtual'' arm whose success probability can be thought of as
the agent's \emph{aspiration level} \cite{LDFS44}. The first arm
that meets the agent's aspirations is the arm that is played from then
on. As we mentioned in the introduction, this can be viewed as
\emph{satisficing} \cite{Simon56}.
Not only does this approach use significantly fewer states, it seems
more like what people do.
Rao's protocol has another feature that renders it an implausible
model of human behavior. It uses a number of parameters ($m, \alpha,
C_\alpha, \beta, C_t$) to trade off exploitation and exploration; the
best choice of parameter settings depends on the application domain.
Moreover, these parameters are combined in a nontrivial way (using,
for example, exponentiation). It is hard to believe that people would
take the trouble (or have enough experience) to learn the appropriate
parameter settings for a particular domain, nor are they likely to be
willing to do the computations needed to use them.
We thus significantly simplify the action function and transition
function.
\commentout{
The action function follows human-like heuristics: we start
considering two arms at a time, we run an elimination tournament,
trying to find the better of the two and eliminating the other one.
We do this essentially by always playing the best arm until the
tournament is over (after roughly $N$ steps), and then eliminating
the worse arm. But there is a slight twist.
We continue to play the arm for $n$ steps after it is no longer the
best arm (where $n$ is a parameter to be set) to
see if it becomes the best option again; intuitively, the bad
performance may just be a temporary run of bad luck. People are
well-known to exhibit a \emph{status quo} bias: they tend to stick
with what they already have \cite{SZ88}. We can think of $n$ as
measuring how long people are willing to stick with a possibly bad
choice. In this case, it is easy to show analytically and empirically
that having $n > 0$ makes things better (see below). If the arm does
not become the arm of highest rank within $n$ steps, we switch to the
arm with the current highest rank. (If there is a tie, we play the arm
of highest rank with the lowest index.) To ensure that every arm is
explored seriously, when we first consider an arm, we
``optimistically'' give it the maximum rank $m$. This ensures it will
be played until we get evidence showing that it is not as good as some
other arms.
}
As we said, we use the idea of a tournament, but we play the current
arm against a ``virtual arm'', whose success probability is determined
by the aspiration level, which is rank. If there are $m$ ranks, then
a rank of $r \in \{1, \ldots, m\}$ can be thought of as representing
the interval of probability $[(r-1)/m, r/m]$. We thus take the
success probability of a virtual arm with aspiration level $r$ to be
$(r-.5)/m$, the midpoint of the interval. If we compare arm $i$ to
the virtual arm using a counter. Suppose that we get a success with
arm $i$ (i.e., 1 is observed). Since we expect the virtual arm to
have a success with
probability $(r-.5)/m$, we increase the counter by 1 with probability
$1-(r-.5)/m$ (since this is the probability that the virtual arm had a
failure, so that arm $i$ had one more success than the virtual arm),
and leave the counter unchanged with probability $(r-.5)/m$ (since,
with this probability, both the virtual arm and arm $i$ had a
success). Similarly, if there is a failure with arm $i$, we decrease
the counter with probability $(r-.5)/m$ and leave it unchanged with
probability $1-(r-.5)/m$.
We use two thresholds $M_1$ and $M_2$ to decide when to end the
comparison. If the counter reaches $M_1$, then we declare the current
arm $i$ being considered to
have won the tournament; intuitively, its success probability is
higher than that
of the virtual arm. From then on we play arm $i$. If the counter
reaches $M_2$, then the virtual arm has won the comparison. We
(temporarily) eliminate arm $k$, and compare the virtual arm to arm
$k+1$ if $k < K$. We discuss what happens if $k=K$ shortly, but first note
that there is no analogue to the parameter $N$ of the
elimination-tournament protocol here. The concern in the
elimination-tournament protocol is that we are comparing two arms $i$
and $j$ that
have roughly equal, but not very good success probabilities. Then the
tournament will go on for a long time, but not give a high reward.
With the aspiration-level protocol, if arm $k$ has a success
probability that is essentially
the same as that of the virtual arm, although the comparison may go on
for a long time, the agent is getting a cumulative reward that
essentially matches expectations, so there is no pressure to stop the
comparison.
If $k=K$, then the virtual arm did better than all arms with this
aspiration level. That means that our expectations are too high, so we lower
the aspiration level from $r$ to $r-1$, and retest all arms.
As discussed in the introduction, we do not assume that $M_1 = M_2$.
The implications of an arm $i$ winning the comparison against the virtual
arm are much different than the implications of the virtual arm
winning. In the former case, we play arm $i$ from then on; in the
latter case, we just continue looking for another (hopefully better)
arm. Because the implications are so different, it turns out that we
want to take $M_1$ significantly larger than $M_2$. (Our experiments
suggest that $M_1 = 20$ and $M_2 =-3$ are good choices, along with
$m=100$; see Section~\ref{sec:apirationparameters}.)
One other issue: if the actual best success probability is low
(say, .2) and there are 100 ranks, it will take a long time before the
aspiration level is set appropriately. During this time, the
cumulative regret is increasing. To speed up the process of finding
the ``right'' aspiration level, we can do a quick preprocessing phase to
find the right range, and then explore more carefully. Specifically,
if $m=100$, in the preprocessing phase, when we reset the rank, we
decrease it by 10 (in general, we decrease it by $\sqrt{m}$) rather
than decreasing it by 1. We also use smaller values of $M_1$ and
$M_2$ (say, $M_1=5$ and $M_2 = -1$ rather than $M_1 = 20$ and $M_2=
-3$). If an arm $i$ beats the virtual arm when the aspiration level
$r= 60$, we go back to the previous setting of aspiration level
$r=70$, and do a more careful search starting from there, now
decreasing the aspiration level by 1, and using $M_1 = 20$ and
$M_2=-3$. This preprocessing phase allows us to home in on the
appropriate expectations quickly. Again, besides being more
efficient, this seems to be the type of thing that people do.
\commentout{
There is another complication: as we said, we consider only $k$ of the
$K$ arms at any time. We view these arms as playing a tournament; the
one perceived to be worst will be eliminated. We consider each subset
of $k$ arms for $N$ rounds. At the end of $N$ rounds, we eliminate
the lowest-ranking arm (if there are ties, we eliminate the
lowest-ranking arm of the lowest index), and replace it with another
arm, which initially has rank $m$, so that it will be explored.
This approach uses significantly fewer states if $k \ll K$; moreover,
people seem to focus on a small number of options at any given
time. They can feel overwhelmed if they have too many choices. Indeed,
having too many choices has been associated with unhappiness
\cite{Schwartz04}; Iyenagar and Lepper \citeyear{IL00} showed that
people are more likely to purchase goods when offered a limited set of
options and report greater subsequent satisfaction with their
selections.
There is a final feature of our protocol that we have already hinted
at above: it implements a tournament. Our goal is to identify the
optimal arm and play it as often as possible. If we continually swap
in a new arm every $N$ steps, we are almost certain to spend a lot of
time playing arms that are not very good. We want to completely
eliminate ``bad'' arms from consideration. To do this, every $N$
steps, we eliminate the arm with the lowest rank. That means that
after $(K-1)N$ steps, we are guaranteed to be down to one arm. Thus,
our approach has some of the spirit of the \emph{explore then exploit}
strategy; we essentially explore for $(K-1)N$ steps (in expectation),
and then choose the best arm. By using a sequence of tournaments to
eliminate arms, we can so this in a state-efficient way. Of course,
there is always the risk that we may end up with the wrong arm. But it
does seem quite close to what people do. People (at least most
people!) prefer to eliminate options and eventually to home in on the
``best'' one, rather than continually reconsidering options.
}
\commentout{
We have one final improvement. Before running the elimination
tournament, we have a phase where we test each arm to get an estimate
of the success probability of the best arm.
Whereas before, we viewed rank $i$ as representing an interval of
probability of roughly $i/m$ (so that a rank of 1 represents a
success probability in the interval $[0,1/m]$, a rank of 2 represents a
success probability in the interval $[1/m,2/m]$, and so on)
if the ranks of the best arm is $r^*$, we now view the ranks as
representing much smaller intervals, that range roughly $(r^* - 1)/m$ to
$(r^*+.5)/m$. For example, if we have 10 ranks, and the highest rank
among the final $k-1$ arms is 7, then we reinterpret the 10 ranks as
ranging from .6 to .75. so that rank $k$ now represents a
success probability in the interval $[.6 + (k/10)(.15)]$. This allows
us to make much
finer
distinctions among the remaining arms. The probabilities $u_i$ and
$d_i$ of increasing and decreasing the ranks are then changed to
reflect this interpretation.
Again, this seems human-like: once we
have narrowed things down, we look at the remaining candidates more
carefully.
After this pre-processing phase,
we can then get down to a single best arm by running an elimination
tournament.
}
With this background, we are ready to define our family of
PFAs.
For ease of presentation, we do not use a preprocessing phase.
Formally, we have a family $M_{K, m, M_1,M_2} =
(Q_{K,m,M_1,M_2}, q_m,\Sigma_K, O_K, \gamma_{K},
\delta_{K,m,M_1,M_2})$ of PFAs, indexed by 4 parameters: \(K\) is the
total number of arms,
\(m\) is the number of possible ranks for each arm, and $M_1$ and
$M_2$ are the upper and lower thresholds for the counter.
We assume that $K$ is given as part of
the input; we discuss how $m$, $M_1$, and $M_2$ are chosen in the
next section. Not only do we have fewer parameters than Rao, as we
shall see, they are easier to set (and easier to explain and
understand). In more detail, the components of the tuple are as
follows:
\begin{itemize}
\item A state $q \in Q_{K,m,M_1,M_2}$ has the form $(r,k, c)$,
where $1 \le r \le m$, $1 \le k \le K$, and $-M_2 < c \le M_1$.
Intuitively, a state $(r,k,c)$ says that the current aspiration level
is $r$, we are testing arm $k$, and the counter that keeps track of
the relative success rate of arm $k$ compared to the virtual arm is at $c$.
\item We take the initial state $q_0$ to be $(m, 1, 0)$:
we start by setting the aspiration level to $m$ (the highest level
possible), testing arm 1, and have the counter at 0.
\item $\Sigma_K$ consists of observations of the form $(k,h)$, where
$k \in \{1, \ldots, K\}$ and $h \in \{0,1\}$. We observe the
outcome of playing arm $k$, which is a reward of either 0
or 1.
\item $O_K = \{1, \ldots, K\}$: we can play any arm.
\item The action function $\gamma_{K}$ at a state $(r,k,c)$
plays arm $k$.
\item The transition function $\delta_{K,m,M_1,M_2}$ proceeds as
follows. In state $(r,k,c)$,
if $c=M_1$, the state does not change. (We have chosen $i$ as the arm
to play from then on.) If $c < M_1$, given an observation $(h,k)$, if
$h=1$ (a success was observed), the new state is $(r,i,c')$, where
$c'=c+1$ with
probability $1-(r-.5)/m$, and otherwise $c' = c$. If $h=0$ and $c>
M_2+1$, then the new state is $(r,i,c')$, where $c'=c-1$ with
probability $(r-.5)/m$, and otherwise is unchanged. If $c = M_2+1$,
then with probability $(r-.5)/m$, the new state is $(r-1,1,0)$ (the
aspiration level is lowered and we start over comparing the virtual
arm to all the arms, starting with arm 1); otherwise the state is unchanged.
\end{itemize}
\section{EXPERIMENTS}
\subsection{PERFORMANCE METRICS}
We use simulations to test the performance of various protocols. In
the simulations, we consider an MAB $B$ with $K$ arms, whose
true success probabilities are uniformly distributed in [0,$\alpha$],
where $\alpha$ is a random number in [0,1]. If we had just assumed
that the success probabilities were uniformly distributed in [0,1],
then the probability of there being
an arm in the [0.9,1] interval is $1-0.9^K$, which is approximately
0.995 for $K=50$. Indeed, the probability of there being an arm in the
interval $[.99,1]$, is about 0.4. Not only does this seem
unreasonable in practice, this assumption would make it too easy to
set the right aspiration level in our approach (i.e., it would hide
some real-world difficulty).
The assumption that the success probabilities are bounded by $\alpha$ for a
randomly-chosen $\alpha$ seems more reasonable. While assuming that
the success probabilities are \emph{uniformly} distributed in
$[0,\alpha]$ may not be so reasonable, our results remain essentially
unchanged even if the success probabilities are chosen adversarially,
and the uniform distribution is much easier to generate.
We focus on two metrics when it comes to measuring the performance of a
protocol:
(1) the expected cumulative regret of a protocol $P$ as a function of the
number of steps played (which roughly depends on how long it
takes to find the best arm) and (2) the expected average regret in the limit
(i.e., $\lim_{N \rightarrow \infty} AReg(P,B,N)$), which essentially measures the gap
between the success probability of the arm chosen by protocol $P$ and
the success probability of the optimal arm of $B$. We take the
expectation over MABs $B$ generated as discussed above. Essentially,
we want a protocol that gets to the best arm quickly and
accurately.
\subsection{PARAMETER SETTINGS IN THE ASPIRATION-LEVEL PROTOCOL}\label{sec:apirationparameters}
There are three parameter settings for the aspiration-level
protocol: the number of ranks $m$, and the thresholds $M_1$ and $M_2$
for winning and losing a comparison against the virtual arm.
We examine the effect of different choices here.
\commentout{
Intuitively, the larger $m$ is, the more closely the ranks approximate
the true success probabilities of the arms. We verified this by
performing simulations with one arm. That is, we played one arm
repeatedly, starting it at rank $m$, using the rules describe above to
update the rank. At each step $t$, we compared $r_t/m$, where $r_t$
is the current rank of the arm, to the true success probability $\mu$
of the arm. The smaller the difference $\mu - r_t/m$, the closer the
estimate is to the true probability. Figure \ref{fig:1} shows the
difference between the true probability and the estimated probability
for different values of $m$ in the case of an arm with $\mu =
.8$. (The results are qualitatively similar for other success
probabilities.) The results are the average of 1000 repetitions. The
difference starts out negative, because we start the arm with rank
$m$, but quickly converges to more or less 0. With small $m$, the
intervals are coarser, so the graph is not as smooth, but the
convergence is faster, since there are fewer steps needed to get from
$m$ to $.8m$. On the other hand, with large $m$, once the rank gets
to $.8m$, it stays very close; the graph is less ``jumpy''. These
results confirm that the ranks do approximate probability well, given
our simple update rule.
\begin{figure}[h]
\includegraphics[width=6.5cm]{m.png}
\caption{Difference from true success rate for different choices of $m$.}
\label{fig:1}
\end{figure}
}
The larger $m$ is, the finer distinctions we will be able to make
between the arms that we are testing. Roughly speaking, if the
virtual arm has rank $r$ and the virtual arm performed better than all
arms when the aspiration level was $r+1$, we would expect that all
arms have success probability less than $(r+.5)/m$, and that an arm
with success probability greater than $(r-.5)/m$ will beat the virtual
arm. However, this arm can have probability as much as $1/m$ less than
the arm with highest success probability. By taking $m$ larger, we
thus minimize the expected gap between the success probability of the
arm chosen and the best arm.
We consider an MAB $B$ with $K=50$ arms and run simulations. As
expected, the larger $m$ is, the smaller the gap, but there are
diminishing returns. Figure \ref{fig:1} shows that, with other
parameters fixed ($M_1=20, M_2=3$), there is significant improvement in
going from $m=50$ to $m=100$; but the marginal improvement drops off
quickly. This is no significant difference between $m=100$ and larger
values such as $m=200$ or $m=500$. The corresponding gaps between the
success probability of the arm chosen and the success
probability of the optimal arm of $B$, averaged over 100 repetitions,
are 0.020, 0.007, 0.0068, 0.0065, respectively. In addition, since we
start optimistically by initializing the aspiration level at the
highest possible rank, when $m$ is larger, it takes longer to get the
right aspiration level and hence the cumulative regret is larger, as
shown in Figure \ref{fig:1}. Considering both performance metrics as
mentioned above, we choose $m=100$ in the later simulations.
\begin{figure}[h]
\includegraphics[width=6.5cm]{m_choice.png}
\caption{Cumulative regret for different $m$.}
\label{fig:1}
\end{figure}
Once we fix $m$, we now examine the choices of $M_1$ and $M_2$.
The parameters $M_1$ and $M_2$ determine the conditions of winning and
losing: if counter gets to $M_1$, then the current arm beats the
``virtual arm" and is therefore chosen as the best arm; if the counter
gets to $-M_2$, then the current arm loses the tournament with the
``virtual arm" and we move to a new arm. If all $K$ arms lose the
tournament, we decrease the aspiration level by 1 and restart
the tournament. We want it to be easier for the ``virtual arm" to win,
since the consequences are lower in that case (the protocol ends if
we declare arm $i$ a winner, whereas we keep going if the ``virtual
arm" is a winner). Therefore, it makes sense to have an asymmetry and
choose $M_1$ greater than $M_2$.
We again consider an MAB $B$ with
$K=50$ arms and $m=100$ fixed. As shown in Figure \ref{fig:2}, the
cumulative regret increases as $M_1$ and $M_2$ get larger. However,
the gap between the success probability of the arm chosen by protocol
$P$ and the success probability of the optimal arm of $B$,
decreases. The corresponding gaps, averaged over 100 repetitions, are
0.014, 0.007, 0.005, 0.004, respectively. Since the number of states
in the aspiration-level protocol is $Km(M_1 + M_2)$, there is a
tradeoff between accuracy and the number of states required. Taking
into account state-efficiency, accuracy, and the expected cumulative
regret, we choose $M_1=20$ and $M_2=3$.
\begin{figure}[h]
\includegraphics[width=6.5cm]{M.png}
\caption{Cumulative regret for different $M_1$ and $M_2$.}
\label{fig:2}
\end{figure}
\commentout{
However, getting an accurate test of small distinctions requires us to
test longer, which means that we have to take $M_1$ and $M_2$ larger.
Since the number
of states in the aspiration-level protocol is $Km(M_1 + M_2+1)$, there
is a tradeoff between accuracy and the number of states required.
The parameters $M_1$ and $M_2$ determine the conditions of winning and
losing: if counter gets to $M_1$, then the current arm beats the
``virtual arm" and is therefore chosen as the best arm; if the counter
gets to $-M_2$, then the current arm loses the tournament with the
``virtual arm" and we move to a new arm. If all $K$ arms lose the
tournament, we decrement the aspiration level by 1 and restart
the tournament. We want it to be easier for the ``virtual arm" to win,
since the consequences are lower in that case (the protocol ends if
we declare arm $i$ a winner, whereas we keep going if the ``virtual
arm" is a winner). Therefore, it makes sense to choose $M_1$ greater
than $M_2$.
After experimenting with different values, we choose
$M_1=20$ and $M_2=3$.
}
Both Figure \ref{fig:1} and Figure \ref{fig:2} show that the
performance of the aspiration-level protocol degrades quite
gracefully as we take
smaller values of $m$, $M_1$, and $M_2$ (which is how we would have to deal
with having
fewer states).
\commentout{
Figure \ref{fig:2} shows that the performance of the
\emph{aspiration-level} protocol degrades quite gracefully as we take
smaller values of $m$, $M_1$, and $M_2$. Averaging 100 repetitions of
the simulation, the corresponding gaps from the optimal arm (i.e., the
slope of the cumulative regret) after 100k steps for each protocol are
0.008, 0.02, 0.03, respectively.
\begin{figure}[h]
\includegraphics[width=6.5cm]{degradation.png}
\caption{Graceful degradation.}
\label{fig:2}
\end{figure}
}
\subsection{PARAMETER SETTINGS FOR THE ELIMINATION TOURNAMENT}
The elimination-tournament protocol has two parameters: $M$ (the point
at which an arm is declared a winner in the two-way comparison) and
$N$ (recall that $1/N$ is the probability that an arm is declared in the
two-way comparison if no arm is dominant and has $M$ more successes than
the other). Thus, after an expected number of at most $N(K-1)$
steps, the elimination-tournament protocol has reduced to one arm. We clearly
want $M$ and $N$ to be large enough to give the protocol time to select a
relatively good arm. However, we don't want to stick with bad arms for
too long, since this will lead to larger cumulative regret.
Figure \ref{fig:3} shows the cumulative regrets for different choices
of $N$ and $M$, for an MAB with $K=50$ arms.
For the choices of $(N,M)$ considered---(1000,10), (1000,20), (1000,
100), (100,10), (100,20)---the gaps, averaged 100 repetitions,
are 0.01, 0.007, 0.006, 0.03, 0.03,
respectively. Both $N=1000, M=20$ and $N=1000, M=100$
give similarly good performance in terms of the expected average
regret, but the latter leads to larger expected cumulative
regret. Therefore, for $K=50$, we choose $N=1000$ and $M=20$.
\begin{figure}[h]
\includegraphics[width=6.5cm]{N_M.png}
\caption{Cumulative regret for different $N$ and $M$.}
\label{fig:3}
\end{figure}
\subsection{COMPARING PROTOCOLS}
Based on the simulations above, to minimize the number of states used
while maintaining relatively good performance, for $K=50$, we choose the
parameters $m=100, M_1 = 20, M_2 = 3$ for the aspiration-level
protocol and $M=20$, $N=1000$ for the elimination-tournament protocol, and
compare these two finite-state protocols to the
$\epsilon$-greedy protocol and Thompson Sampling, which are
infinite-state protocols. With these choices, the
aspiration-level protocol uses 115,000 states, while the
elimination-tournament protocol uses just over 100,000. While this
may seem to be a a lot of states, they can be encoded using 17 bits.
Given the number of neurons in a human brain, this should not be a problem.
We can greatly reduce the cumulative regret for the
aspiration-level protocol by a preprocessing phase, as
suggested earlier. For $K=50$ arms and the aspiration-level
protocol with $m=100, M_1 = 20, M_2 = 3$, we first use a preprocessing
phase to get a rough idea of what the true highest success probability
might be.
We use the parameters suggested earlier, decreasing the aspiration
level by 10 after testing all the arms in the preprocessing, and use
thresholds $M_1' = -5$ and $M_2'= -1$.
We use this two-phase approach for the aspiration-level protocol in
the following simulation.
We consider MABs with $K=50$ arms, and see how
the elimination-tournament protocol, the aspiration-level protocol,
$\epsilon$-greedy, and Thompson sampling perform.
Not surprisingly, Thompson sampling
performs best, and has logarithmic cumulative regret, whereas the
other three protocols have linear cumulative regret.
After 50,000 steps, the
expected difference between the success probability of the arm chosen
and that of the optimal arm for
these protocols are 0.007, 0.008, 0.025, 0.003, respectively.
Interestingly, both the aspiration-level protocol and the
elimination-tournament protocol eventually outperform
$\epsilon$-greedy, although the latter requires
infinitely many states.
\begin{figure}[h]
\includegraphics[width=6.5cm]{compare.png}
\caption{Cumulative regret over time.}
\label{fig:4}
\end{figure}
\commentout{
\subsection{The choice of parameters}
Given the total number of arms $K$, there are many possible ways to choose the values of the parameters $k$, $m$, $n$, and $N$. We examined the effect of different choices using simulations.
Intuitively, the larger $m$ is, the more closely the ranks approximate
the true success probabilities of the arms. We verified this by
performing simulations with one arm. That is, we played one arm
repeatedly, starting it at rank $m$, using the rules describe above to
update the rank. At each step $t$, we compared $r_t/m$, where $r_t$
is the current rank of the arm, to the true success probability $\mu$
of the arm. The smaller the difference $\mu - r_t/m$, the closer the
estimate is to the true probability. Figure \ref{fig:1} shows the
difference between the true probability and the estimated probability
for different values of $m$ in the case of an arm with $\mu =
.8$. (The results are qualitatively similar for other success
probabilities.) The results are the average of 1000 repetitions. The
difference starts out negative, because we start the arm with rank
$m$, but quickly converges to more or less 0. With small $m$, the
intervals are coarser, so the graph is not as smooth, but the
convergence is faster, since there are fewer steps needed to get from
$m$ to $.8m$. On the other hand, with large $m$, once the rank gets
to $.8m$, it stays very close; the graph is less ``jumpy''. These
results confirm that the ranks do approximate probability well, given
our simple update rule.
\begin{figure}[h]
\includegraphics[width=6.5cm]{m.png}
\caption{Difference from true success rate for different choices of $m$.}
\label{fig:1}
\end{figure}
Recall that we introduced $n$ to minimize the probability that a good
arm will be eliminated due to a run of bad luck.
Taking $n>0$ has other benefits. If $n=0$, we play
an arm of highest rank at each step. Once an arm goes below the
current best rank, we stop playing it, and switch to the current best
arm. Suppose we consider $m=10$ and initialize all $k$ arms to be at
the highest rank 10. We start playing arm 1. After a number of steps,
arm 1 will have rank 9, so we will switch to arm 2. If $N$ is
sufficiently large, eventually all arms will have rank 9, and we
repeat the process. Again, if $N$ is sufficiently large, all arms will
eventually have rank 8, and so on. It should be clear that if $n=0$
and there are at least two arms, the ranks will not reflect the true
probability at all! By taking $n > 0$ and not too small, we avoid this
problem, and obtain two other benefits: a good arm is more likely to
be the one played when a phase ends (i.e., after roughly $N$ steps),
since it has the time to recover from a run of bad luck, while the
rank of a bad arm will reflect its true success probability, rather
than just being one below the current best rank. The key point here is
that a \emph{status quo} bias (at least in the context of our family
of PFAs) is rational!
The parameter $N$ determines how long it takes for the PFA to eliminate all but one arm. In particular, after $N(K-1)$ steps, we are down to one arm. Clearly we want $N$ to be large enough to allow the PFA to select a relatively good arm. We need $N > nk$ for good performance, since we want to explore each arm in the focus set before we eliminate any arm. Since we play each arm at least expected number of $n$ times, and perhaps much more, we take $N > 4nk$ to allow for some margin of error. Figure \ref{fig:1} shows that the estimated probability converges to the true probability after $n=100$ steps. Therefore, for the following simulations, we consider $n=100$ and $N=400k$.
Intuitively, with a larger choice of $k$, the best arm is less likely to be
eliminated (since it is less likely to be the worst among $k$ arms the
larger $k$ is). We considered multi-armed bandits $B$ with $K=50$
arms, with the success probability of each arm chosen independently
and uniformly at random. We examined the quality of the final arm
chosen by considering the difference $\mu^*_B - \mu$ between the
success probability $\mu$ of the final arm $i$ chosen and the success
probability $\mu^*_B$ of the optimal arm.
Simulations show that, keeping the parameters $m$, $n$, and $N$ fixed,
the larger
$k$ is, the smaller the difference, but there are diminishing
returns. There is significant improvement in going from $k=2$ to
$k=3$; but the marginal improvement drops off quickly. This is no
significant difference between $k=4$ and larger values such as $k=10$
or $k=20$. The total number of arms does not matter (as long as $K >
k$). We verified this by considering another set of multi-armed
bandits $B$ with $K=20$ arms, each with the success probability chosen
independently and uniformly at random. We observe a similar
improvement in going from from $k=2$ to $k=3$ and a decreasing
marginal improvement once we have $k=4$.
We performed simulations on multi-armed bandits $B$ with $K=50$ arms,
with the success probability of each arm chosen independently and
uniformly at random. Figure \ref{fig:3} shows that if $n$ is fixed at 100,
the performance of our family of PFAs degrades quite gracefully as we
take smaller values of $m$, $k$, and $N$. Averaging 100 repetitions of
the simulation, the corresponding differences from the optimal arm (i.e.,
the slope of the cumulative regret) after 100k steps for each PFAs are
0.0001, 0.0002, 0.0003, 0.001, 0.02, respectively.
\begin{figure}[h]
\includegraphics[width=6.5cm]{degradation.png}
\caption{Graceful degradation.}
\label{fig:2}
\end{figure}
As we noted, for $K=50$, the PFA with $m=20$ and $k=4$ uses
$kC(K,k)m^kK \approx 7*10^{12}$ states. This is how many states are
needed to keep track of which subset of 4 of the 50 arms is currently
being considered, and the ranks of these four arms. The big hit comes
in keeping track of which of the $50 \choose 4$ subsets is being
played. Even if we choose $m=10$ and $k=2$, we need as many as
$2*C(50,2)*10^2*50 \approx 10^{7}$ states. Clearly, we cannot expect
people to use that many states!
Our two-phase algorithm both reduces the number of states and
increases the accuracy of the arm choice.
As we said in the introduction, we first play each arm roughly $N$
times to get an estimate of what the best rank is.
This phase requires $Km^2$ states. Now we use our previous
algorithm with $k=2$, zooming in around the best rank. For $K=50$, if
we use the two-phase algorithm in the PFA with $m=10$, we only need
$2K^2m^2 = 5*10^5$ states to reach the same performance. Simulations
show that, averaged over 100 repetitions, the corresponding
differences from optimal arm (i.e., the slope of the cumulative
regret) after 100k steps for is also roughly 0.02. This two-phase
algorithm is also quite human-like. Gamblers in real-life tend to play
somewhat randomly at the beginning to get a sense of how good each
machine is before doing a more systematic examination.
\subsection{Comparing different protocols}
In the experiments, we again considered multi-armed bandits $B$ with $K=50$ arms, each with the success probability chosen independently, and uniformly at random.
Based on the simulations above, to minimize the number of states used
while maintaining a relatively good performance, we choose the
parameters $m=10, k=2, n=100, N=800$ and use the two-phase algorithm,
so we consider the PFA $M_{50,10,2,100,800}$, and compare it to three
other protocols, explore-then-exploit, $\epsilon$-greedy, and
Thompson Sampling. The PFA $M_{50,10,2,100,800}$ with two-phase
algorithm uses $5*10^5$ states. As discussed earlier,
explore-then-exploit is a finite-state protocol, while
$\epsilon$-greedy
and Thompson Sampling use infinitely many states. We can implement
explore-then-exploit protocol using a ``two-armed
tournament". That is, we run arms 1 and 2 for $N$ steps, eliminate the
worse arm, and then run the winner against arm 3, run the winner
against arm 4, and so on. Thus, at any point, we need to just keep
track of the ranks of two arms: the current best arm and the one being
tested. We also need to keep track of which arms these are and need a
counter from 0 to $N$. Thus, there are approximately $C(K,2)N^3$
states if we implement \emph{explore then exploit} as a tournament. If
we take $N=8$, we get a version of \emph{explore then exploit} that
uses roughly the same number of states as the PFA we considered, so we
compared to that in our simulation.
As shown in Figure \ref{fig:3}, not surprisingly, Thompson Sampling performs best, and has logarithmic error, whereas \emph{explore then exploit}, $\epsilon$-greedy, and our PFA all have linear error rates.
Note that since the PFA with two-phase algorithm spends longer time
exploring, it takes longer time to converge to only one arm. However,
a more important measure of the quality of the final arm chosen is the
difference $\mu^*_B - \mu$ between the success probability $\mu$ of
the final arm $i$ chosen and the success probability $\mu^*_B$ of the
optimal arm. Averaging the outcomes 100 repetitions of the simulation,
after 60,000 steps, the corresponding differences from the optimal arm
for each protocol are 0.006, 0.006, 0.0, 0.001, respectively. Note
that the proposed PFA outperforms \emph{explore then exploit} using
the same number of states; it also outperforms $\epsilon$-greedy,
although the former requires infinitely many states.
\begin{figure}[h]
\includegraphics[width=6.5cm]{compare.png}
\caption{Cumulative regret over time.}
\label{fig:3}
\end{figure}
}
\section{DISCUSSION}
We have introduced two finite-state protocols for playing MABs, the
aspiration-level protocol and the elimination-tournament protocol.
Both perform quite well in practice, while using relatively few states.
In cases where switching between arms incurs a significant
cost, the aspiration-level protocol is a better choice.
Recall that the main motivation for this study was understanding human
behavior.
\commentout{
As we observed, building in a \emph{status quo} bias
improves the performance of our protocol. It also demonstrates other
``human-like'' behavior. For example, we can think of the focus set
(which we introduced in order to decrease the size of the state space)
as an instance of an availability bias \cite{TK73}: the only arms that
are played are those that are ``available''; that is, those in the
focus set. Finally, a run of bad luck with a particular arm (which can
be viewed as a negative outcome for that arm) can result in result in
a good arm being eliminated. This is more likely to happen if $k=2$
than if $k$ is large. With smaller values of $m$, we also get more
variability in the ranks, so again, a run of bad luck can have a
larger impact.
}
The fact that the aspiration-level protocol exhibits such human-like
behavior, including adjusting aspiration levels according to feedback,
an optimism bias, a negativity bias, and a status quo bias, as well as
a focus on recent behavior, suggests that humans are not being so
irrational.
Note that these biases are emphasized if the number of states is
decreased. For example, if an agent decreases $M_2$, the threshold for rejecting
an arm in a two-way comparison with the virtual arm, in response to
having fewer states, this increases the negativity
bias. Decreasing $M_1$ increases the likelihood that an agent will continue
to play an apparently ``lucky arm''. The impact of decreasing $M$
in the elimination-tournament protocol is similar.
The bottom line is that these protocols exhibit
apparently irrational behavior for quite rational reasons! At the
same time, they may be of interest even for those not interested in
modeling human behavior, since they have quite good performance, even
with relatively few states.
We have focused here on a static setting, where the probabilities do
not change over time. We could easily modify our PFA to deal with the
dynamic setting by simply
resetting the tournaments from time to time.
More interestingly, we would like to apply these ideas to
a more game-theoretic setting, such as the wildlife poaching setting
considered by Kar et al.~\citeyear{Kar15}, where rangers are trying to
protect rhinos from poachers. We hope to report on that in future
work.
\subsubsection*{Acknowledgements}
This research was supported by MURI (MultiUniversity Research
Initiative) under grant
W911NF-19-1-0217, by the ARO under grant W911NF-17-1-0592,
by the NSF under grants IIS-1703846 and IIS-1718108, and by a grant
from the Open Philosophy Foundation. We thank Alice
Chen for her preliminary work and comments on an earlier version of
the manuscript. We also thank four anonymous reviewers for their
feedback.
\eject
\bibliographystyle{chicago}
|
\section{Introduction}
For any complex variable $x$, define the shifted-factorial to be
\[(x)_{0}=1\quad \text{and}\quad (x)_{n}
=x(x+1)\cdots(x+n-1)\quad \text{when}\quad n\in\mathbb{N}.\]
In 1997, Van Hamme
\cite[(H.2)]{Hamme} conjectured that
\begin{equation}\label{eq:hamme}
\sum_{k=0}^{(p-1)/2}\frac{(1/2)_k^3}{k!^3}\equiv
\begin{cases} \displaystyle -\Gamma_p(1/4)^4 \pmod{p^2}, &\text{if $p\equiv 1\pmod 4$,}\\[10pt]
0\pmod{p^2}, &\text{if $p\equiv 3\pmod 4$.}
\end{cases}
\end{equation}
Here and throughout the paper, $p$ always denotes an odd prime and
$\Gamma_p(x)$ is the $p$-adic Gamma function. In 2016, Long and
Ramakrishna \cite[Theorem 3]{LR} gave the following extension of
\eqref{eq:hamme}:
\begin{equation}\label{eq:long}
\sum_{k=0}^{(p-1)/2}\frac{(1/2)_k^3}{k!^3}\equiv
\begin{cases} \displaystyle -\Gamma_p(1/4)^4 \pmod{p^3}, &\text{if $p\equiv 1\pmod 4$,}\\[10pt]
\displaystyle -\frac{p^2}{16}\Gamma_p(1/4)^4\pmod{p^3}, &\text{if $p\equiv 3\pmod 4$.}
\end{cases}
\end{equation}
For any complex numbers $x$ and $q$, define the $q$-shifted factorial
as
\begin{equation*}
(x;q)_{0}=1\quad\text{and}\quad
(x;q)_n=(1-x)(1-xq)\cdots(1-xq^{n-1})\quad \text{when}\quad
n\in\mathbb{N}.
\end{equation*}
For simplicity, we also adopt the compact notation
\begin{equation*}
(x_1,x_2,\dots,x_m;q)_{n}=(x_1;q)_{n}(x_2;q)_{n}\cdots(x_m;q)_{n}.
\end{equation*}
Following Gasper and Rahman \cite{Gasper}, define the basic
hypergeometric series $_{r+1}\phi_{r}$ by
$$
_{r+1}\phi_{r}\left[\begin{array}{c}
a_1,a_2,\ldots,a_{r+1}\\
b_1,b_2,\ldots,b_{r}
\end{array};q,\, z
\right] =\sum_{k=0}^{\infty}\frac{(a_1,a_2,\ldots, a_{r+1};q)_k}
{(q,b_1,b_2,\ldots,b_{r};q)_k}z^k.
$$
Then the $q$-Whipple formula due to Andrews \cite{Andrews} and the
$q$-Whipple formula due to Jain \cite{Jain} can be stated as
\begin{align}
& _{4}\phi_{3}\!\left[\begin{array}{cccccccc}
q^{-n}, q^{1+n}, b, -b \\
-q, c, b^2q/c
\end{array};q,\, q \right]
=q^{\binom{n+1}{2}}\frac{(b^2q^{1-n}/c, cq^{-n};q^2)_{n}}
{(b^2q/c, c;q)_{n}}, \label{eq:q-whipple-a}
\\[5pt]
& _{4}\phi_{3}\!\left[\begin{array}{cccccccc}
a, q/a, q^{-n}, -q^{-n} \\
c, q^{1-2n}/c, -q
\end{array};q,\, q \right]
=\frac{(ac, cq/a;q^2)_{n}}
{(c;q)_{2n}}. \label{eq:q-whipple-b}
\end{align}
Recently, Guo and Zudilin \cite[Theorem 2]{GuoZu2} displayed a
$q$-analogue of \eqref{eq:hamme}: for any positive odd integer $n$,
\begin{align}
&\sum_{k=0}^{(n-1)/2}\frac{(q;q^2)_k^2(q^2;q^4)_k}{(q^2;q^2)_k^2(q^4;q^4)_k}q^{2k}
\notag\\[5pt]
&\equiv
\begin{cases} \displaystyle\frac{(q^2;q^4)_{(n-1)/4}^2}{(q^4;q^4)_{(n-1)/4}^2}\pmod{\Phi_n(q)^2}, &\text{if $n\equiv 1\pmod 4$,}\\[10pt]
\displaystyle 0\pmod{\Phi_n(q)^2}, &\text{if $n\equiv 3\pmod 4$.}
\end{cases}
\label{eq:guo-a}
\end{align}
Here and throughout the
paper, $\Phi_n(q)$ stands for the $n$-th cyclotomic polynomial in
$q$:
\begin{equation*}
\Phi_n(q)=\prod_{\substack{1\leqslant k\leqslant n\\
\gcd(k,n)=1}}(q-\zeta^k),
\end{equation*}
where $\zeta$ is an $n$-th primitive root of unity. Further, Guo
\cite[Theorem 1]{Guo-new} provided the following partial
$q$-analogue of \eqref{eq:long}: for any positive integer
$n\equiv3\pmod{4}$,
\begin{align}
\sum_{k=0}^{(n-1)/2}\frac{(q,;q^2)_k^2(q^2;q^4)_k}{(q^2;q^2)_k^2(q^4;q^4)_k}q^{2k}
\equiv[n]\frac{(q^3;q^4)_{(n-1)/2}}{(q^5;q^4)_{(n-1)/2}}\pmod{\Phi_n(q)^3}.
\label{eq:guo-b}
\end{align}
For more $q$-analogues of supercongruences, we refer the reader to
\cite{Guo-rima,Guo-jmaa,Guo-rama,Guo-a2,GS1,GuoZu,LP,NP,Tauraso,WY-a,Zu19}.
Motivated by the work just mentioned, we
shall establish the following result.
\begin{thm}\label{thm-a}
Let $n\equiv 1\pmod 4$ be a positive integer. Then, modulo
$\Phi_n(q)^3$,
\begin{align*}
\sum_{k=0}^{(n-1)/2}\frac{(q;q^2)_k^2(q^2;q^4)_k}{(q^2;q^2)_k^2(q^4;q^4)_k}q^{2k}
\equiv
q^{(n-1)/2}\frac{(q^2;q^4)_{(n-1)/4}^2}{(q^4;q^4)_{(n-1)/4}^2}\bigg\{1+2[n]^2\sum_{i=1}^{(n-1)/4}\frac{q^{4i-2}}{[4i-2]^2}\bigg\}.
\end{align*}
\end{thm}
Obviously, Theorem \ref{thm-a} is an extension of \eqref{eq:guo-a}
for $n\equiv1\pmod{4}$. Letting $n=p$ be a prime and taking $q\to 1$
in this theorem, we obtain the conclusion:
\begin{align}\label{eq:wei}
\sum_{k=0}^{(p-1)/2}\frac{(1/2)_k^3}{k!^3} \equiv
\frac{(1/2)_{(p-1)/4}^2}{\big((p-1)/4\big)!^2}\bigg\{1+\frac{p^{2}}{2}H_{(p-1)/2}^{(2)}-\frac{p^{2}}{8}H_{(p-1)/4}^{(2)}\bigg\}\pmod{p^3},
\end{align}
where the harmonic numbers of $2$-order are given by
\[H_{m}^{(2)}
=\sum_{k=1}^m\frac{1}{k^{2}}.\]
Using the known formula (cf. \cite[Page 7]{Sun}):
\begin{align*}
&H_{(p-1)/2}^{(2)}\equiv0\pmod{p}\quad\text{with}\quad p>3,
\end{align*}
we deduce the following supercongruence from \eqref{eq:wei}.
\begin{cor}\label{cor-a}
Let $p\equiv 1\pmod 4$ be a prime. Then
\begin{align}\label{eq:wei-a}
\sum_{k=0}^{(p-1)/2}\frac{(1/2)_k^3}{k!^3} \equiv
\frac{(1/2)_{(p-1)/4}^2}{\big((p-1)/4\big)!^2}\bigg\{1-\frac{p^{2}}{8}H_{(p-1)/4}^{(2)}\bigg\}\pmod{p^3}.
\end{align}
\end{cor}
For the sake of explaining the equivalence of \eqref{eq:long} for
$p\equiv 1\pmod 4$ and \eqref{eq:wei-a}, we need to verify the
following relation.
\begin{prop}\label{prop-a}
Let $p\equiv 1\pmod 4$ be an odd prime. Then
\begin{align*}
\frac{(1/2)_{(p-1)/4}^2}{\big((p-1)/4\big)!^2}\bigg\{1-\frac{p^{2}}{8}H_{(p-1)/4}^{(2)}\bigg\}\equiv
-\Gamma_p(1/4)^4\pmod{p^3}.
\end{align*}
\end{prop}
The rest of the paper is arranged as follows. By means of the
Chinese remainder theorem for coprime polynomials, a
$q$-supercongruence modulo $(1-aq^n)(a-q^n)(b-q^n)$ will be derived
in Section 2. Then it is utilized to provide a proof of Theorem
\ref{thm-a} in the same section. Finally, the proof of Proposition
\ref{prop-a} will be given in Section 3.
\section{Proof of Theorem \ref{thm-a}}
In order to prove Theorem \ref{thm-a}, we need the following
parameter extension of it.
\begin{thm}\label{thm-b}
Let $n\equiv 1\pmod 4$ be a positive integer. Then, modulo
$(1-aq^n)(a-q^n)(b-q^n)$,
\begin{align}
\sum_{k=0}^{(n-1)/2}\frac{(aq,q/a,q/b,-q/b;q^2)_k}{(q^2,q^2,-q^2,q^2/b^2;q^2)_k}q^{2k}
\equiv \Omega_n(a,b),\label{eq:wei-aa}
\end{align}
where
\begin{align*}
\Omega_n(a,b)&=\frac{(b-q^n)(ab-1-a^2+aq^n)}{(a-b)(1-ab)}\frac{(b/q)^{(1-n)/2}(q^2,b^2q^2;q^4)_{(n-1)/4}}{(q^4,q^4/b^2;q^4)_{(n-1)/4}}
\\[5pt]
&+\frac{(1-aq^n)(a-q^n)}{(a-b)(1-ab)}\frac{(aq^3,q^3/a;q^4)_{(n-1)/2}}{(q^2;q^2)_{n-1}}.
\end{align*}
\end{thm}
\begin{proof}
When $a=q^{-n}$ or $a=q^n$, the left-hand side of \eqref{eq:wei-aa}
is equal to
\begin{align}
\sum_{k=0}^{(n-1)/2}\frac{(q^{1-n},q^{1+n},q/b,-q/b;q^2)_k}{(q^2,q^2,-q^2,q^2/b^2;q^2)_k}q^{2k}
= {_{4}\phi_{3}}\!\left[\begin{array}{cccccccc}
q^{1-n}, q^{1+n}, q/b, -q/b \\
q^2, -q^2, q^2/b^2
\end{array};q^2,\, q^2 \right].
\label{eq:whipple-aa}
\end{align}
According to \eqref{eq:q-whipple-a}, the right-hand side of
\eqref{eq:whipple-aa} can be expressed as
\begin{align*}
(b/q)^{(1-n)/2}\frac{(q^2,b^2q^2;q^4)_{(n-1)/4}}{(q^4,q^4/b^2;q^4)_{(n-1)/4}}.
\end{align*}
Since $(1-aq^n)$ and $(a-q^n)$ are relatively prime polynomials, we
get the following result: Modulo $(1-aq^n)(a-q^n)$,
\begin{align}
\sum_{k=0}^{(n-1)/2}\frac{(aq,q/a,q/b,-q/b;q^2)_k}{(q^2,q^2,-q^2,q^2/b^2;q^2)_k}q^{2k}\equiv
(b/q)^{(1-n)/2}\frac{(q^2,b^2q^2;q^4)_{(n-1)/4}}{(q^4,q^4/b^2;q^4)_{(n-1)/4}}. \label{eq:wei-bb}
\end{align}
When $b=q^{n}$, the left-hand side of \eqref{eq:wei-aa} is equal to
\begin{align}
\sum_{k=0}^{(n-1)/2}\frac{(aq,q/a,q^{1-n},-q^{1-n};q^2)_k}{(q^2,q^2,-q^2,q^{2-2n};q^2)_k}q^{2k}
= {_{4}\phi_{3}}\!\left[\begin{array}{cccccccc}
aq, q/a, q^{1-n}, -q^{1-n} \\
q^2, -q^2, q^{2-2n}
\end{array};q^2,\, q^2 \right].
\label{eq:whipple-bb}
\end{align}
In terms of \eqref{eq:q-whipple-b}, the right-hand side of
\eqref{eq:whipple-bb} can be written as
\begin{align*}
\frac{(aq^3,q^3/a;q^4)_{(n-1)/2}}{(q^2;q^2)_{n-1}}.
\end{align*}
Therefore, we are led to the following conclusion: Modulo $(b-q^n)$,
\begin{align}
\sum_{k=0}^{(n-1)/2}\frac{(aq,q/a,q/b,-q/b;q^2)_k}{(q^2,q^2,-q^2,q^2/b^2;q^2)_k}q^{2k}
\equiv\frac{(aq^3,q^3/a;q^4)_{(n-1)/2}}{(q^2;q^2)_{n-1}}.\label{eq:wei-cc}
\end{align}
It is clear that the polynomials $(1-aq^n)(a-q^n)$ and $(b-q^n)$ are
relatively prime. Noting the $q$-congruences
\begin{align*}
&\frac{(b-q^n)(ab-1-a^2+aq^n)}{(a-b)(1-ab)}\equiv1\pmod{(1-aq^n)(a-q^n)},
\\[5pt]
&\qquad\qquad\frac{(1-aq^n)(a-q^n)}{(a-b)(1-ab)}\equiv1\pmod{(b-q^n)}
\end{align*}
and employing the Chinese remainder theorem for coprime polynomials,
we deduce Theorem \ref{thm-b} from \eqref{eq:wei-bb} and
\eqref{eq:wei-cc}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm-a}]
It is not difficult to see that
\begin{align*}
(q^2;q^2)_{n-1}&=(q^2, q^{n+1}, q^4, q^{n+3};q^4)_{(n-1)/4}
\\[5pt]
&=q^{(n-1)(3n-1)/4}(q^2, q^4, q^{2-2n}, q^{4-2n};q^4)_{(n-1)/4}\\[5pt]
&\equiv
b^{n-1}q^{(1-n^2)/4}(q^{2},q^{4},q^2/b^2,q^4/b^2;q^4)_{(n-1)/4}\pmod{(b-q^n)},
\\[5pt]
(aq^3;q^4)_{(n-1)/2}&=(aq^3;q^4)_{(n-1)/4}(aq^{n+2};q^4)_{(n-1)/4}
\\[5pt]
&\equiv(abq^{3-n};q^4)_{(n-1)/4}(abq^{2};q^4)_{(n-1)/4}
\\[5pt]
&=(-ab)^{(n-1)/4}q^{-(n-1)^2/8}(abq^2,q^2/ab;q^4)_{(n-1)/4}\pmod{(b-q^n)},
\\[5pt]
(q^3/a;q^4)_{(n-1)/2}&=(-b/a)^{(n-1)/4}q^{-(n-1)^2/8}(bq^2/a,aq^2/b;q^4)_{(n-1)/4}\pmod{(b-q^n)}.
\end{align*}
Thus, the $q$-supercongruence \eqref{eq:wei-aa} may be rewritten as
follows: Modulo $(1-aq^n)(a-q^n)(b-q^n)$,
\begin{align*}
&\sum_{k=0}^{(n-1)/2}\frac{(aq,q/a,q/b,-q/b;q^2)_k}{(q^2,q^2,-q^2,q^2/b^2;q^2)_k}q^{2k} \\[5pt]
&\quad\equiv
\frac{(b-q^n)(ab-1-a^2+aq^n)}{(a-b)(1-ab)}\frac{(b/q)^{(1-n)/2}(q^2,b^2q^2;q^4)_{(n-1)/4}}{(q^4,q^4/b^2;q^4)_{(n-1)/4}}
\\[5pt]
&\quad\quad+\frac{(1-aq^n)(a-q^n)}{(a-b)(1-ab)}\frac{(b/q)^{(1-n)/2}(abq^2,bq^2/a,aq^2/b,q^2/ab;q^4)_{(n-1)/4}}{(q^2,q^4,q^2/b^2,q^4/b^2;q^4)_{(n-1)/4}}.
\end{align*}
Letting $b\to 1$, we arrive at the following formula: Modulo
$\Phi_n(q)(1-aq^n)(a-q^n)$,
\begin{align}
&\sum_{k=0}^{(n-1)/2}\frac{(aq,q/a;q^2)_k(q^2;q^4)_k}{(q^2;q^2)_k^2(q^4;q^4)_k}q^{2k}
\notag\\[5pt]
&\:\:\:\equiv
q^{(n-1)/2}\frac{(q^2;q^4)_{(n-1)/4}^2}{(q^4;q^4)_{(n-1)/4}^2}+q^{(n-1)/2}\frac{(1-aq^n)(a-q^n)}{(1-a)^2}
\notag\\[5pt]
&\quad\:\:\times\bigg\{\frac{(q^2;q^4)_{(n-1)/4}^2}{(q^4;q^4)_{(n-1)/4}^2}-\frac{(aq^2,q^2/a;q^4)_{(n-1)/4}^2}{(q^2,q^4;q^4)_{(n-1)/4}^2}\bigg\}.
\label{eq:wei-dd}
\end{align}
By the L'Hospital rule, we have
\begin{align*}
&\lim_{a\to1}\frac{(1-aq^n)(a-q^n)}{(1-a)^2}\bigg\{\frac{(q^2;q^4)_{(n-1)/4}^2}{(q^4;q^4)_{(n-1)/4}^2}
-\frac{(aq^2,q^2/a;q^4)_{(n-1)/4}^2}{(q^2,q^4;q^4)_{(n-1)/4}^2}\bigg\}\\[5pt]
&=2[n]^2\frac{(q^2;q^4)_{(n-1)/4}^2}{(q^4;q^4)_{(n-1)/4}^2}\sum_{i=1}^{(n-1)/4}\frac{q^{4i-2}}{[4i-2]^2}.
\end{align*}
Letting $a\to1$ in \eqref{eq:wei-dd} and utilizing the above limit,
we complete the proof of Theorem~\ref{thm-a}.
\end{proof}
\section{Proof of Proposition \ref{prop-a}}
Via the congruence due to Wang and Pan \cite[Page 6]{Wang}:
\begin{align*}
H_{(p-1)/4}^{(2)}\equiv\frac{\Gamma_p^{''}(1/4)}{\Gamma_p(1/4)}-\bigg\{\frac{\Gamma_p^{'}(1/4)}{\Gamma_p(1/4)}\bigg\}^2\pmod{p},
\end{align*}
where $\Gamma_p^{'}(x)$ and $\Gamma_p^{''}(x)$ are respectively the
first derivative and second derivative of $\Gamma_p(x)$, we obtain
\begin{align}\label{eq:wei-b}
1-\frac{p^{2}}{8}H_{(p-1)/4}^{(2)}\equiv1-\frac{p^{2}}{8}\frac{\Gamma_p^{''}(1/4)}{\Gamma_p(1/4)}+
\frac{p^{2}}{8}\bigg\{\frac{\Gamma_p^{'}(1/4)}{\Gamma_p(1/4)}\bigg\}^2\pmod{p^3}.
\end{align}
In terms of the properties of the $p$-adic Gamma function, we get
\begin{align}
\frac{(1/2)_{(p-1)/4}^2}{\big((p-1)/4\big)!^2}&=\bigg\{\frac{\Gamma_p((1+p)/4)\Gamma_p(1)}{\Gamma_p(1/2)\Gamma_p((3+p)/4)}\bigg\}^2
\notag\\[5pt]
&=\bigg\{\frac{\Gamma_p((1+p)/4)\Gamma_p((1-p)/4)}{\Gamma_p(1/2)}\bigg\}^2
\notag\\[5pt]
&\equiv-\bigg\{\Gamma_p(1/4)+\Gamma_p^{'}(1/4)\frac{p}{4}+\Gamma_p^{''}(1/4)\frac{p^2}{2\times4^2}\bigg\}^2
\notag\\[5pt]
&\quad\times\bigg\{\Gamma_p(1/4)-\Gamma_p^{'}(1/4)\frac{p}{4}+\Gamma_p^{''}(1/4)\frac{p^2}{2\times4^2}\bigg\}^2\pmod{p^3}.
\label{eq:wei-c}
\end{align}
The combination of \eqref{eq:wei-b} and \eqref{eq:wei-c} produces
\begin{align*}
&\frac{(1/2)_{(p-1)/4}^2}{\big((p-1)/4\big)!^2}\bigg\{1-\frac{p^{2}}{8}H_{(p-1)/4}^{(2)}\bigg\}
\\[5pt]
&\quad\equiv-\bigg\{\Gamma_p(1/4)+\Gamma_p^{'}(1/4)\frac{p}{4}+\Gamma_p^{''}(1/4)\frac{p^2}{2\times4^2}\bigg\}^2
\notag\\[5pt]
&\qquad\times\bigg\{\Gamma_p(1/4)-\Gamma_p^{'}(1/4)\frac{p}{4}+\Gamma_p^{''}(1/4)\frac{p^2}{2\times4^2}\bigg\}^2
\\[5pt]
&\qquad\times\bigg\{1-\frac{p^{2}}{8}\frac{\Gamma_p^{''}(1/4)}{\Gamma_p(1/4)}+
\frac{p^{2}}{8}\bigg\{\frac{\Gamma_p^{'}(1/4)}{\Gamma_p(1/4)}\bigg\}^2\bigg\}
\\[5pt]
&\quad\equiv -\Gamma_p(1/4)^4 \pmod{p^3}.
\end{align*}
|
\section{Introduction}
Time-harmonic solutions to the wave equation can be computed in the
frequency or in the time domain. In the frequency domain, a discrete
version of the Helmholtz equation is solved. In the time domain, the
periodic solutions to a discrete wave equation with time-harmonic
forcing term are sought. Frequency domain methods involve less degrees
of freedom. However, the indefinite linear systems resulting from
discretizing the Helmholtz equation are often difficult to
solve. Time-domain methods can be attractive because they require
relatively little memory and are easy to implement if a time-domain
solver is available. The introduction of
\cite{appelo2019waveholtz} contains a recent overview of time-harmonic
wave equation solvers.
Time-domain methods are based on the correspondence between Helmholtz
and wave equations. In this paper we will use as example the damped
wave equation
\begin{equation} \label{eq:damped_wave_equation}
\frac{1}{c^2} \pdpd{^2 u}{t^2} + R \pdpd{u}{t} - \Delta u = f ,
\end{equation}
where $u = u(t,x)$ is the wave field, $f = f(t,x)$ is the forcing
term, $\Delta$ is the Laplacian, $c(x)$ is the spatially dependent
wavespeed, $R$ is a spatially dependent damping coefficient, and
$x$ is in some domain $\Omega$ with Dirichlet and/or Neumann
boundary conditions. Let $u(x,t), f(x,t)$ and $U(x),
F(x)$ be related by
\begin{equation} \label{eq:time_harmonic_u_and_f}
u(x,t) = e^{i \omega t} U(x) , \qquad \text{ and }
f(x,t) = e^{i \omega t} F(x) .
\end{equation}
Then $u$ satisfies a wave equation with forcing term $f$
if and only if $U$ satisfies the following Helmholtz equation with
forcing term $F$
\begin{equation} \label{eq:damped_Helmholtz_equation}
- \Delta U - \frac{\omega^2}{c^2} U + i \omega R U = F .
\end{equation}
This equation is supplemented with Dirichlet and/or Neumann boundary
conditions that carry over from those for
(\ref{eq:damped_wave_equation}).
We briefly review some time-domain approaches.
The most basic time-domain method is derived from the
limiting-amplitude principle. This principle states that solutions
$u(x,t)$ to (\ref{eq:damped_wave_equation}) with zero initial
conditions and forcing term
\begin{equation} \label{eq:basic_idea_forcing_term}
f(x,t) = e^{i \omega t} F(x)
\end{equation}
satisfy
\begin{equation} \label{eq:limiting-amplitude_principle}
u(x,t) = e^{i \omega t} U(x) + O(1) , \qquad t \to \infty
\end{equation}
under certain conditions on the problem, see
\cite{EncyclopediaOfMath_LimitingAmplitudePrinciple} and references
therein. Thus, if $u(x,t)$ is a (numerical) solution to this initial
boundary-value problem, and $T$ is some large time, measured in
periods, then an approximate solution to the Helmholtz equation is
given by
\begin{equation} \label{eq:limiting_amplitude_approx_sol}
e^{- i 2 \pi T} u(x, 2\pi \omega^{-1} T ) .
\end{equation}
We will call this the limiting-amplitude approximate solution for
time $T$.
A more advanced method is the
exact controllability method \cite{bristeau1998controllability}. In
this method the periodicity of the solutions is enforced using
optimization. The starting value for the optimization procedure is
typically some partially converged limiting-amplitude solution.
Recently more insights in and some improvements to this method were
obtained \cite{grote2019controllability}, and its parallel
implementation was studied \cite{grote2020parallel}.
In \cite{appelo2019waveholtz} an optimization approach called
WaveHoltz was introduced. This method again uses a form of
optimization but with a different optimization functional.
An important feature of the methods just described is that they approximate
periodic solutions of a given discrete wave equation. This has
two consequences that are in general not desirable. First, the results
will be negatively affected by both spatial and temporal
discretization errors, while solutions to discrete Helmholtz equations
only have spatial discretization errors. Secondly, the method is
limited to situations in which a time-domain scheme is available, and
not (directly) applicable if one only has a discretization of
(\ref{eq:damped_Helmholtz_equation}) available, or perhaps only a
linear system with similar properties.
In this paper, we will address both of these shortcomings by
developing a new time-domain solver for discrete
Helmholtz equations. In fact the method can be applied to any linear
system
\begin{equation} \label{eq:discrete_damped_Helmholtz}
H U = F ,
\end{equation}
where $H$ is a complex $N \times N$ matrix, such that
\begin{equation} \label{eq:Re_H_requirement}
\text{there is $c>0$ such that $\Re H + c I$ is symmetric positive semidefinite} ,
\end{equation}
and
\begin{equation} \label{eq:Im_H_requirement}
\text{$\Im H$ is symmetric positive semidefinite}
\end{equation}
In equation (\ref{eq:discrete_damped_Helmholtz})
$F$ is a vector in ${\mathbb{C}}^N$ and
$U$ is the unknown, also in ${\mathbb{C}}^N$.
We next describe the steps involved in the construction of the new
time-domain solver.
First an $N \times N$ system of second order ODE's
\begin{equation} \label{eq:general_semi-discrete_equation}
\pdpd{u^2}{t^2} + A u + B \pdpd{u}{t} = f
\end{equation}
and a frequency parameter $\omega$ are constructed. We will look for
time-harmonic solutions with frequency $\omega$ of the system
(\ref{eq:general_semi-discrete_equation}). Note that $\omega$ is in
general {\em not} the physical frequency parameter used to derive
(\ref{eq:discrete_damped_Helmholtz}). It is a computational
parameter, chosen together with the matrices $A$ and $B$, and it depends
on $H$ in a way to be specified. Time-harmonic functions
$u = e^{i\omega t} U$ and $f = e^{i \omega t} F$, with
$U, F \in {\mathbb{C}}^N$, satisfy (\ref{eq:general_semi-discrete_equation}) if
and only if
\begin{equation} \label{eq:time-harmonic_from_semi-discrete}
( - \omega^2 I + i \omega B + A) U = F .
\end{equation}
Therefore we will choose $A,B$ and $\omega$ such that
\begin{equation} \label{eq:choice_of_ABomega_intro}
H = -\omega^2 I + i \omega B + A .
\end{equation}
The system (\ref{eq:general_semi-discrete_equation}) plays
the role of a semi-discrete wave equation.
Secondly, this system is time-discretized.
This is done in such a way that time-harmonic solutions of the
discrete-time system are exactly those of the continuous-time system.
I.e.\ if $u_n, f_n$ are related to $U,F$ by $u_n = e^{i\omega n \Delta
t} U$ and
$f_n = e^{i \omega n \Delta t} F$ then $u_n,f_n$ are solutions to the
discrete-time system if and only if $U,F$ satisfy
(\ref{eq:time-harmonic_from_semi-discrete}).
For this purpose we present two modified leapfrog methods.
So by using non-standard finite-difference derivatives,
there are no time-discretization errors for the time-harmonic
solutions. This is related to ideas from the papers
\cite{holberg1987computational,tam1993dispersion} where it was shown
that errors in finite-difference discretizations can be analysed
precisely in the Fourier domain, and that finite-difference
approximations of derivatives can be designed to minimize such errors
over some range of wavenumbers.
See also \cite{berland2006low} for an optimized time-stepping method.
Having a time-discretization of
(\ref{eq:general_semi-discrete_equation}) at hand, the third step is
to define a map from a right-hand side $F$ in
(\ref{eq:discrete_damped_Helmholtz}) to an approximation for the
solution $U$. For this we follow the idea of equation
(\ref{eq:limiting_amplitude_approx_sol}) with one modification, which
is the inclusion of a smooth window function in the time-harmonic
forcing term.
It is known that the convergence of limiting-amplitude approximate
solutions can be slow in some cases, e.g.\ in case of resonant wave
cavities. Therefore we will not use the approximate solution operator
directly. Instead we propose to use it as a preconditioner in an
iterative solution method for (\ref{eq:discrete_damped_Helmholtz}),
such as GMRES or BiCGSTAB. This is the fourth and last step of our
construction. This use as a preconditioner can also be called Krylov
acceleration, as the method itself already approximates the true solution.
It can be seen as an alternative to the exact controllability and
WaveHoltz methods.
Because it is used this way, we will call the new approximate solution
operator a {\em time-domain preconditioner}.
The above distinction between frequency- and time-domain methods
appears no longer satisfactory
for this method: It is based on time-domain methodology, but solves a
frequency-domain discrete wave equation.
The term time-domain preconditioner was used before in
\cite{zschiedrich2006advanced}. In that work a related problem was
solved, but the resulting method was substantially different, as the
Schr\"odinger equation was used instead of the wave equation and the
resulting time-domain scheme was implicit, and not exactly
solved.
The behavior of the proposed method will be explained using theory and
examples. It is established theoretically under which conditions the
new time-discretizations are stable and that the approximate solutions
converge to the exact solution of (\ref{eq:discrete_damped_Helmholtz})
in the limit of large time parameter $T$. The numerical examples
confirm the converge of the approximate solutions and show that Krylov
acceleration is indeed useful to accelerate the solution
process. Krylov acceleration makes a difference in particular in the
case that a resonant low-velocity zone is present. Parameters of the
method include the large time parameter $T$ used in the
preconditioner, and a parameter for the window
function. Results depend weakly on these parameters, i.e.\ there is a
large set of suitable parameter choices.
A few situations that do not fit in the classical time-domain setting,
but can be handled with the new solver are as follows. First one can
use it with discretizations that have been designed specifically for
the Helmholtz equation. For example finite-difference discretizations
that minimize dispersion errors such as those from
\cite{BabuskaEtAl1995} for the 2-D case and
\cite{stolk2016dispersion,sutmann2007compact,TurkelEtAl2013} for the
3-D case. Our examples are about this application. One can
also imagine situations where the physical time-domain model is
complicated to simulate, but reduces to a relatively simple Helmholtz
equation in the frequency domain. A third use case is in the context
of a multigrid method. In multigrid methods the coarsest level system
still has to be solved by another (non-multigrid) method. It is based
on the linear system one started with and on the choice of
multigrid method. In Helmholtz equations the convergence can be very
sensititve to the choice of the coarse level system and it is
recommended to use certain prescribed discretizations
\cite{StolkEtAl2014,stolk2016dispersion}.
However, the usefulness of the method is not restricted to these
cases, and some of the ideas could also be incorporated into other
time-domain methods.
The contents of the remainder of the paper is as follows. In
section~\ref{sec:method}, the construction of the new solver is
described. In section~\ref{sec:examples} examples of this construction
are given in case $H$ results from certain finite-difference
discretizations. Section~\ref{sec:analysis} contains the theoretical
results. After that, section~\ref{sec:numerical_examples} contains the
numerical examples. We conclude the main text with a discussion
section. Appendices contains some material on Fourier analysis used
in section~\ref{sec:analysis}, and some remarks on the compact-stencil
finite-difference discretization described in section~\ref{sec:examples}.
\section{Methods}
\label{sec:method}
In this section we describe in detail the construction of a
time-domain preconditioner for a matrix $H$ satisfying
(\ref{eq:Re_H_requirement}) and (\ref{eq:Im_H_requirement}).
We recall from the introduction
that there are three main steps:
(i) the definition of a suitable second order system of ODE's of the
form (\ref{eq:general_semi-discrete_equation});
(ii) the definition of a suitable time-integration method for this system of
ODE's;
(iii) the definition of a linear map that produces approximate
solutions based on the limiting-amplitude principle.
Step (ii), the time-integration method, will be discussed first, since
the properties of the time-integration method affect the choice of the
system of ODE's (\ref{eq:general_semi-discrete_equation}).
Then steps (i) and (iii) and the
application of the method as a preconditioner are discussed.
Occasionally we will point forward to section~\ref{sec:analysis}, where
some properties of the time-discretizations and the preconditioner are
proven.
\subsection{Frequency-adapted time discretizations of
(\ref{eq:general_semi-discrete_equation})}
\label{subsec:time-discretization}
The leapfrog or basic Verlet method is a standard method to integrate
equations of the form (\ref{eq:general_semi-discrete_equation}) in
case that $B = 0$. It is obtained, basically, by discretizing the second order
time derivative using standard second order finite differences.
To allow for nonzero $B$, the damping term has to be discretized as well.
A standard way to do this is with central differences
\cite{brunger1984stochastic, schlick2010molecular,
sandvik2018numerical}. This yields the equation
\begin{equation} \label{eq:central_differences_damped_leapfrog_verlet}
\frac{1}{\Delta t^2} \left( u_{n+1} - 2 u_n + u_{n-1}
\right)
+ \frac{1}{2 \Delta t} B \left( u_{n+1} - u_{n-1} \right)
+ A u_n = f_n ,
\end{equation}
from which $u_{n+1}$ can be solved.
Note that $u_n$ denotes the discrete approximation to $u(n \Delta t)$,
and that $u(t) \in {\mathbb{C}}^N$.
This only leads to an explicit method if $B$ is diagonal. To obtain an
explicit method in case $B$ is non-diagonal, the damping term can be
discretized by backward differences, which yields the equation
\begin{equation} \label{eq:backward_differences_damped_leapfrog_verlet}
\frac{1}{\Delta t^2} \left( u_{n+1} - 2 u_n + u_{n-1}
\right)
+ \frac{B}{\Delta t} \left( u_n - u_{n-1} \right)
+ A u_n = f_n .
\end{equation}
Since this method is an order less accurate and has stricter CFL
conditions it is only proposed for the case that $B$ is non-diagonal,
in the other case (\ref{eq:central_differences_damped_leapfrog_verlet}) is preferred.
We will formally define the time-integrators resulting from
(\ref{eq:central_differences_damped_leapfrog_verlet}) and
(\ref{eq:backward_differences_damped_leapfrog_verlet}).
\begin{definition} \label{def:Icd_Ibd}
Let
\begin{equation} \label{eq:K_L_central_differences}
\begin{aligned}
K = {}& \Delta t^2 A
\qquad {}&
L = {}& \Delta t B
\qquad {}&
g_n = \Delta t^2 f_n .
\end{aligned}
\end{equation}
{\em Central differences damped leapfrog} will be defined as the time
integrator given by
\begin{equation} \label{eq:cd_integrator}
u_{n+1} = I_{\rm cd}(u_n,u_{n-1},f_n)
:= \left( I + \tfrac{1}{2}L \right)^{-1}
\left( (2 - K) u_n - (I - \tfrac{1}{2}L) u_{n-1} + g_n \right) .
\end{equation}
{\em Backward differences damped leapfrog} will be defined as the time
integrator given by
\begin{equation} \label{eq:bd_integrator}
u_{n+1}
= I_{\rm bd}(u_n,u_{n-1},f_n)
:= (2 - K - L) u_n - (I - L) u_{n-1} + g_n .
\end{equation}
\end{definition}
The stability of these methods is studied in
section~\ref{sec:analysis}. According to Theorem~\ref{th:stability_cd},
$I_{\rm cd}$ is stable if
\begin{equation} \label{eq:stability_cond_K_L}
\text{$K$ and $L$ are positive semidefinite},
\end{equation}
and
\begin{equation} \label{eq:stability_cond_cd_upper}
\text{$4I - K$ is positive definite} ,
\end{equation}
while $I_{\rm bd}$ is stable if
(\ref{eq:stability_cond_K_L}) is satisfied and
\begin{equation} \label{eq:stability_cond_bd_upper}
\text{$4I - K - 2L$ is positive definite} .
\end{equation}
Conditions (\ref{eq:stability_cond_cd_upper}) and
(\ref{eq:stability_cond_bd_upper}) lead to CFL bounds. This will
be discussed below.
\medskip
As mentioned in the introduction, we look for discretizations such
that the time-harmonic solutions of frequency $\omega$ of the
discrete-time system are exactly those of the continuous time system
(\ref{eq:general_semi-discrete_equation}). Due to discretization
errors this is not the case for
the integrators $I_{\rm cd}$ and $I_{\rm bd}$. In the next
proposition we will show that the time-harmonic solutions to
(\ref{eq:general_semi-discrete_equation}) satisfy recursions of the
same form as (\ref{eq:central_differences_damped_leapfrog_verlet}) and
(\ref{eq:backward_differences_damped_leapfrog_verlet}), but with
different choices of $A,B$. From these recursions modified schemes can derived
that have the desired property.
\begin{proposition} \label{prop:reproduce_symbol_at_omega}
Let $u_n$ and $f_n$ be related to $U,F \in {\mathbb{C}}^N$ by
\begin{equation}
u_n = e^{i \omega n \Delta t} U , \qquad
f_n = e^{i \omega n \Delta t} F
\end{equation}
and let
\begin{equation} \label{eq:define_alpha_beta}
\begin{aligned}
\alpha = {}& \frac{(\Delta t \, \omega)^2}{2 - 2 \cos (\omega \Delta t)}
= \frac{(\Delta t \, \omega)^2}{4 \sin( \frac{\omega \Delta
t}{2})^2} ,
& \qquad \text{ and } \qquad
\beta = {}& \frac{\omega \Delta t}{ \sin( \omega \Delta t)} .
\end{aligned}
\end{equation}
Then $U,F$ satisfy (\ref{eq:time-harmonic_from_semi-discrete}) if
and only if $u_n$, $f_n$ satisfy
\begin{equation} \label{eq:mod_central_differences_damped_leapfrog_verlet}
\frac{1}{\Delta t^2} \left( u_{n+1} - 2 u_n + u_{n-1}
\right)
+ \frac{1}{2 \Delta t} \tilde{B} \left( u_{n+1} - u_{n-1} \right)
+ \tilde{A} u_n = \alpha^{-1} f_n ,
\end{equation}
where
\begin{equation} \label{eq:redefineABf_cd}
\tilde{A} = \alpha^{-1} A , \qquad \text{ and } \qquad
\tilde{B} = \alpha^{-1} \beta B ,
\end{equation}
and if and only if $u_n$ and $f_n$ satisfy
\begin{equation} \label{eq:mod_backward_differences_damped_leapfrog_verlet}
\frac{1}{\Delta t^2} \left( u_{n+1} - 2 u_n + u_{n-1}
\right)
+ \frac{\hat{B}}{\Delta t} \left( u_n - u_{n-1} \right)
+ \hat{A} u_n = \alpha^{-1} f_n ,
\end{equation}
where
\begin{equation} \label{eq:redefineABf_bd}
\hat{A} = \alpha^{-1} A
- \frac{\beta (1-\cos(\omega \Delta t))}{\alpha \Delta t} B ,
\qquad \text{ and } \qquad
\hat{B} = \alpha^{-1} \beta B .
\end{equation}
\end{proposition}
\begin{proof}
To prove the first claim,
$\tilde{A}$, $\tilde{B}$ and $\tilde{c}$
will be constructed such that
\begin{equation} \label{eq:recursion_tilde}
\frac{1}{\Delta t^2} \left( u_{n+1} - 2 u_n + u_{n-1}
\right)
+ \frac{1}{2 \Delta t} \tilde{B} \left( u_{n+1} - u_{n-1} \right)
+ \tilde{A} u_n = \tilde{c} f_n ,
\end{equation}
if and only if (\ref{eq:time-harmonic_from_semi-discrete}).
Inserting $u_n = U e^{i n \omega \Delta t}$ into
(\ref{eq:recursion_tilde}), results in
\begin{equation}
\left[ \frac{2 \cos(\omega \Delta t) - 2}{\Delta t^2}
+ \frac{i \, \sin( \Delta t \omega)}{\Delta t} B
+ \tilde{A}
\right] U e^{i n \omega \Delta t} = \tilde{c} e^{i n \omega \Delta t} F .
\end{equation}
Using the definitions of $\alpha$ and $\beta$ to rewrite the
left-hand side, this is equivalent to
\begin{equation}
\left[ - \frac{\omega^2}{\alpha} + i \frac{\omega}{\beta} \tilde{B}
+ \tilde{A} \right] U = \tilde{c} F .
\end{equation}
Multiplying by $\alpha$ results in
\begin{equation}
\left[ - \omega^2 + i \omega \frac{\alpha}{\beta} \tilde{B}
+ \alpha \tilde{A} \right] U = \alpha \tilde{c} F .
\end{equation}
This is equivalent to (\ref{eq:time-harmonic_from_semi-discrete})
if $\tilde{c} = \alpha^{-1}$ and $\tilde{A}$ and $\tilde{B}$ are defined
as in (\ref{eq:redefineABf_cd}).
To prove the second claim,
$\hat{A}$, $\hat{B}$ and $\hat{c}$
will be constructed such that
\begin{equation} \label{eq:recursion_hat}
\frac{1}{\Delta t^2} \left( u_{n+1} - 2 u_n + u_{n-1} \right)
+ \frac{1}{\Delta t} \hat{B} \left( u_n - u_{n-1} \right)
+ \hat{A} u_n = \hat{c} f_n ,
\end{equation}
if and only if (\ref{eq:time-harmonic_from_semi-discrete}).
Inserting $u_n = U e^{i n \omega \Delta t}$ into
(\ref{eq:recursion_hat}), results in
\begin{equation}
\left[ \frac{2 \cos(\omega \Delta t) - 2}{\Delta t^2}
+ \frac{i \, \sin( \Delta t \omega)}{\Delta t} \hat{B}
+ \frac{1 - \cos( \Delta t \omega)}{\Delta t} \hat{B}
+ \hat{A}
\right] U e^{i n \omega \Delta t} = \hat{c} e^{i n \omega \Delta t} F .
\end{equation}
Using the definitions of $\alpha$ and $\beta$ and multiplying by
$\alpha$ results in the equivalent equation
\begin{equation}
\left[ - \omega^2
+ i \omega \frac{\alpha}{\beta} \hat{B}
+ \frac{\alpha(1 - \cos( \Delta t \omega))}{\Delta t} \hat{B}
+ \alpha \hat{A} \right] U = \alpha \hat{c} F .
\end{equation}
This is equivalent to (\ref{eq:time-harmonic_from_semi-discrete})
if $\hat{c} = \alpha^{-1}$ and $\hat{A}$ and $\hat{B}$ are defined
as in (\ref{eq:redefineABf_bd}).
\end{proof}
Based on the proposition we define the following time integrators.
The time-harmonic solutions with frequency $\omega$ of these
integrators correspond exactly to time-harmonic solutions of
(\ref{eq:general_semi-discrete_equation}).
\begin{definition} \label{def:Iacd_Iabd}
{\em Frequency adapted central differences damped leapfrog}
will be defined as the time integrator given by
\begin{equation} \label{eq:acd_integrator}
u_{n+1} = I_{\rm acd}(u_n,u_{n-1},f_n)
:= \left( I + \tfrac{1}{2}L \right)^{-1}
\left( (2 - K) u_n - (I - \tfrac{1}{2}L) u_{n-1} + g_n \right) ,
\end{equation}
where $K$, $L$ and $g_n$ are given by
\begin{equation} \label{eq:K_L_frequency_adapted_central_differences}
\begin{aligned}
K = {}& \frac{\Delta t^2}{\alpha} A ,
& \qquad
L = {}& \frac{\beta \Delta t}{\alpha} B ,
& \qquad
g_n = {}& \frac{\Delta t^2}{\alpha} f_n
\end{aligned}
\end{equation}
{\em Frequency adapted backward differences damped leapfrog} will be
defined as the time integrator given by
\begin{equation} \label{eq:abd_integrator}
u_{n+1}
= I_{\rm abd}(u_n,u_{n-1},f_n)
:= (2 - K - L) u_n - (I - L) u_{n-1} + g_n .
\end{equation}
where $K$, $L$ and $g_n$ are given by
\begin{equation} \label{eq:K_L_frequency_adapted_backward_differences}
\begin{aligned}
K = {}& \frac{\Delta t^2}{\alpha} A
- \frac{\Delta t \,
\beta (1-\cos(\omega \Delta t))}{\alpha} B \qquad
\\
L = {}& \frac{\beta \Delta t}{\alpha} B ,
\\
g_n = {}& \frac{\Delta t^2}{\alpha} f_n
\end{aligned}
\end{equation}
\end{definition}
The time integrators $I_{\rm cd}$ and $I_{\rm acd}$ are of the same
form with different choices for $K$ and $L$. Therefore for
$I_{\rm acd}$ the stability conditions are again
(\ref{eq:stability_cond_K_L}) and (\ref{eq:stability_cond_cd_upper}),
but now with $K,L$ as in
(\ref{eq:K_L_frequency_adapted_central_differences}). Similarly, for
$I_{\rm abd}$, the stability conditions are
(\ref{eq:stability_cond_K_L}) and (\ref{eq:stability_cond_bd_upper})
with $K,L$ as in
(\ref{eq:K_L_frequency_adapted_backward_differences}).
\subsection{Choice of the semi-discrete system and the parameters
$\omega$ and $\Delta t$}
\label{subsec:choose_semidiscrete}
We now look for a second order system of ODE's of the form
(\ref{eq:general_semi-discrete_equation}), and a parameter $\omega$
such that the time-harmonic solutions to
(\ref{eq:general_semi-discrete_equation}) satisfy $H U = F$. Recall
that $\omega$ is the computational frequency parameter, that is chosen
in the construction of the algorithm, and not the physical frequency
used in the underlying Helmholtz equation. We will
also discuss the choice of time integrator ($I_{\rm acd}$ or $I_{\rm
abd}$) and the choice of the parameter $\Delta t$.
The requirements are that an explicit and stable scheme is obtained.
Because of (\ref{eq:choice_of_ABomega_intro}), we set
\begin{align} \label{eq:A_B_from_H}
A = {}& \Re H + \omega^2 I , \qquad \text{and}
\\
\label{eq:relate_B_ImH}
B = {}& \omega^{-1} \Im H .
\end{align}
where $\omega$ is still to be determined.
It is required that the time-integrator leads to explicit scheme for
the given matrices $A,B$. By (\ref{eq:relate_B_ImH}) the matrix $B$ is
diagonal if and only if
\begin{equation} \label{eq:ImH_diagonal}
\text{$\Im H$ is diagonal} .
\end{equation}
Following the remarks below
(\ref{eq:central_differences_damped_leapfrog_verlet})
the frequency adapted central differences method is used if $\Im H$ is
diagonal, and the frequency adapted backward differences method
otherwise.
Next the stability conditions must be considered.
In case $\Im H$ is diagonal these are given in
(\ref{eq:stability_cond_K_L}) and (\ref{eq:stability_cond_cd_upper}).
Because $\Im H$ is positive semidefinite, $L$ is also positive
semidefinite. To ensure that $K$ is positive semidefinite, we set
\begin{equation} \label{eq:define_omega_ImH_diagonal}
\omega = \sqrt{ - \lambda_{\rm min}(\Re H)} .
\end{equation}
(or to a lower bound for $\sqrt{ - \lambda_{\rm min}(\Re H)}$ if this
value is not exactly known).
The other stability condition (\ref{eq:stability_cond_cd_upper}) is a form of the
well-known CFL condition. It implies that the eigenvalues
$K$ must be less than $4$. For central differences
damped leapfrog integration (not frequency adapted) it implies the condition
\begin{equation}
\Delta t < \frac{2}{\sqrt{\lambda_{\rm max}(A)}} .
\end{equation}
For the frequency adapted variant it implies, by
(\ref{eq:K_L_frequency_adapted_central_differences}) and
(\ref{eq:define_alpha_beta}) the following condition on $\Delta t$
\begin{equation}
\sin( \frac{\omega \Delta t}{2} )^2
< \frac{\omega^2}{\lambda_{\rm max}(A)} ,
\end{equation}
hence
\begin{equation} \label{eq:define_Deltat_ImH_diagonal}
\Delta t
<
\frac{2}{\omega} \arcsin( \frac{\omega}{\sqrt{ \lambda_{\rm max}(A) } } ) .
\end{equation}
The choices and requirements (\ref{eq:A_B_from_H}),
(\ref{eq:define_omega_ImH_diagonal}) and
(\ref{eq:define_Deltat_ImH_diagonal}) define a suitable choice of
parameters of the systems of ODEs and the $I_{\rm acd}$ integrator in
case $\Im H$ is diagonal.
In case that $\Im H$ has nonzero entries outside the diagonal and the $I_{\rm
abd}$ integrator is used, the stability conditions are
(\ref{eq:stability_cond_K_L}) and (\ref{eq:stability_cond_bd_upper})
with $K,L$ as in
(\ref{eq:K_L_frequency_adapted_backward_differences}).
This results in two conditions that both involve $\omega$ and
$\Delta t$.
In this case it is convenient to use $\omega \Delta t$ and $\Delta t$
as parameters instead of $\omega$ and $\Delta t$. The parameter
$\omega \Delta t$ should be between 0 and $\pi$. Given
$\omega \Delta t$ the following expression for $\omega^2$ can be
derived from the condition that $K$ is positive semidefinite and
equation (\ref{eq:K_L_frequency_adapted_backward_differences})
\begin{equation}
\omega^2 = - \lambda_{\rm min}(\Re H)
+ \frac{\beta}{\omega \Delta t} ( 1 - \cos(\omega \Delta t))
\lambda_{\rm max}(\Im H)
\end{equation}
(instead of $\lambda_{\rm min}(\Re H)$ and
$\lambda_{\rm max}(\Im H) $ lower and upper bounds can be used respectively).
From the condition that $4I - K - 2L$ is positive definite we then get
the following scalar condition
\begin{equation} \label{eq:Deltat_bd_full_inequality}
\frac{ (\omega \Delta t)^2} {\alpha}
\left( \frac{\lambda_{\rm max}(\Re H)}{\omega^2} + 1 \right)
+ \frac{\beta}{\alpha} (1 + \cos(\omega \Delta t))
\frac{\lambda_{\rm max}(\Im H)}{ \omega^2} < 4 .
\end{equation}
The following is a stronger inequality than
(\ref{eq:Deltat_bd_full_inequality})
\begin{equation}
\left( \frac{\lambda_{\rm max}(\Re H)}{-\lambda_{\rm min}(\Re H)} + 1
\right) (\omega \Delta t)^2
+ 2 \frac{\lambda_{\rm max}(\Im H)}{-\lambda_{\rm min}(\Re H)}
\omega \Delta t < 4 .
\end{equation}
From here a value of $\omega \Delta t$ can be obtained that satisfies
the conditions by solving a simple quadratic equation.
If a larger value of $\omega \Delta t$ is desired, one can look numerically
for a value as large as possible for which
(\ref{eq:Deltat_bd_full_inequality})
is still satisfied.
\subsection{Time-domain approximate solution operators}
\label{subsec:time-domain_approx_sol}
In this section the time-domain approximate solution operator will be
defined. We also show that a complex approximate solution can be computed
by solving a real time-domain wave equation. This appears to be a standard trick
in the field.
The main novelty is that the formula for the time-harmonic forcing
term (\ref{eq:basic_idea_forcing_term}) is modified so that the
forcing is turned on gradually.
The time-domain approximate solution operators will be denoted by
$S^{\rm acd}_T$ and $S^{\rm abd}_T$, depending on which of the
integrators $I_{\rm acd}$ and $I_{\rm abd}$ is used. They are defined
as follows.
\begin{definition}
Let $P = {\rm acd}$ or $P = {\rm abd}$,
and assume that $I_P$ is an integrator for the system
(\ref{eq:general_semi-discrete_equation})
where $A,B$ and $\omega$ satisfy (\ref{eq:A_B_from_H}).
Let $\chi$ be a $C^\infty$ function that is non-decreasing, equal to
zero for $s \le 0$, and equal to one for $s \ge 1$
and let $T$ be a positive real constant, such that
$n_{\rm steps} := 2\pi \omega^{-1} T / \Delta t$ is an integer.
For $F \in {\mathbb{C}}^N$, let
\begin{equation}
f_n = f(n \Delta t) , \qquad
f(t) = \chi(\frac{t}{2\pi \omega^{-1} T} ) e^{i \omega t} F .
\end{equation}
The {\em time-domain approximate solution operator for $H$} associated with the
integrator $I_P$ is the linear map $S_T^P : {\mathbb{C}}^N \to {\mathbb{C}}^N$ defined by
\begin{equation}
S_T^P F = e^{- i 2\pi T} u_{n_{\rm steps} } ,
\end{equation}
where $u_n$, $n=0,1, \ldots, n_{\rm steps} $ is given by
\begin{equation}
u_{n+1} = I_P (u_n,u_{n-1}, f_n) , \qquad u_0 = 0 .
\end{equation}
\end{definition}
In section~\ref{sec:analysis} the convergence of $S_T^{\rm acd} F$ and
$S_T^{\rm abd} F$ to $H^{-1} F$ will be established under the
assumption that requirements for stability of the the time integrators
discussed in subsection~\ref{subsec:choose_semidiscrete} are
satisfied. Section~\ref{sec:numerical_examples} contains numerical
examples for $S_T^{\rm acd}$.
There is still a lot of freedom to choose the window function $\chi$.
In the numerical examples we will introduce therefore an additional
parameter $\rho$, $0 < \rho \le 1$, such that the window function
is 1 on $[\rho 2\pi \omega^{-1} T , 2\pi \omega^{-1} T]$ and positive but strictly less than one on
$(0,\rho 2\pi \omega^{-1} T )$. We will fix a function $\chi_1$ and set
\begin{equation}
\chi(s) = \chi_\rho(s) = \chi_1(\rho^{-1} s) ,
\end{equation}
The function $\chi_1$ that is used is chosen as a simple sine square
window
\begin{equation}
\chi_1(s) =
\left\{
\begin{array}{ll}
0 & \text{if $s \le 0$}
\\
\sin( \pi s /2)^2 & \text{if $0<s<1$}
\\
1 & \text{if $s \ge 1$}.
\end{array}
\right.
\end{equation}
The parameter $\rho$ will be called the window parameter.
We next show that it is sufficient to solve a real time-domain wave
problem to compute the limiting-amplitude approximate solution. The
argument is given in the continuous case, but is applicable
equally well in the discrete case.
Let $F$ be a complex right-hand side for the Helmholtz equation and
$u(t,x)$ be the solution to (\ref{eq:damped_wave_equation})
with right-hand side $f(t,x) = F(x) e^{i \omega t}$. Assuming the
limiting-amplitude principle holds, cf.\
(\ref{eq:limiting-amplitude_principle}),
an approximate solution to the Helmholtz equation is given by
\begin{equation}
U(x) \approx e^{-i\omega t} u(t,x) , \qquad \text{for some large
$t$} .
\end{equation}
The field $\Re u(t,x)$ can be determined by solving the real wave
equation with real forcing term, i.e.\ with forcing term
\begin{equation}
\Re F(x) e^{i \omega t}
= \cos(\omega t) \Re F(x) -\sin(\omega t) \Im F(x) .
\end{equation}
From (\ref{eq:limiting-amplitude_principle}) it follows that
\begin{equation}
\Re u(t,x) = \Re U(x) \cos(\omega t) - \Im U(x) \sin(\omega t) +
o(1) , \qquad t \to \infty .
\end{equation}
Approximations to $\Re U(x)$ and $\Im U(x)$ can hence be obtained from $\Re
u(t,x)$ by
\begin{equation} \label{eq:real_evaluation}
\begin{aligned}
\Re U(x) \approx {}& \Re u(t,x) ,
\qquad t = T \frac{2 \pi}{\omega}
\\
\Im U(x) \approx {}& \Re u(t,x) ,
\qquad t = \left( T - \frac{1}{4} \right) \frac{2 \pi}{\omega}
\end{aligned}
\end{equation}
with $T$ a large integer. Therefore real time-domain simulation is
sufficient.
In the examples $\Delta t$ is chosen such that the period
$2 \pi \omega^{-1}$ is an integer multiple of $ 4 \Delta t$.
The approximations for $\Re U$ and $\Im U$ in
(\ref{eq:real_evaluation}) are then easy to compute.
\subsection{Time-domain preconditioned GMRES}
The approximate Helmholtz solver can be used as a preconditioner for
iterative methods like GMRES of BiCGSTAB.
Without preconditioning, the system to be solved is $H U = F$.
Applying left-preconditioning means that instead the system
\begin{equation}
P H U = P F
\end{equation}
is solved, where $P = S_T^{\rm acd}$ or $S_T^{\rm abd}$.
The right-preconditioned system is
\begin{equation}
H P V = F .
\end{equation}
The vector $V$ is solved from this system and the solution to the
original problem is then given by $U = P V$. We propose a solution method
where GMRES is applied to the left-preconditioned system.
\section{Examples}
\label{sec:examples}
The examples we consider are finite-difference discretizations
of the damped Helmholtz equation
\begin{equation} \label{eq:damped_Helmholtz_example}
- \Delta U - k^2 \left( 1 - i \frac{ \tilde{R} }{\pi} \right) U = F
.
\end{equation}
Here $\tilde{R}$ is the spatially dependent damping in units of
damping per cycle. We start with standard second order differences and
then apply the method to the discretization from
\cite{stolk2016dispersion}.
The method is not limited to finite-difference
discretizations. Finite-element discretizations can also lead to
linear systems $H U = F$ with $H$ satisfying
(\ref{eq:Re_H_requirement}) and (\ref{eq:Im_H_requirement}).
We will see that in our method, the discrete system is a
discretization of a modified PDE, see (\ref{eq:modified_PDE}) below,
and that the CFL bound for this modified PDE is independent of the
velocity.
\subsection{Second order finite differences}
It is instructive to start with a simple second order
finite-difference discretization. For readability we describe the
two-dimensional case. The degrees of freedom will be denoted by
$U^{(i,j)}$ (in two dimensions), where $i,j$ are in some rectangular
domain $D \subset {\mathbb{Z}}^2$. The matrix $H$ is defined by the equation
\begin{equation}
( H U )^{(i,j)} =
h^{-2} \bigg( 4 U^{(i,j)} - U^{(i-1,j)} - U^{(i+1,j)} - U^{(i,j-1)} -
U^{(i,j+1)} \bigg) - (k^{(i,j)})^2 \left( 1 - i \frac{ \tilde{R}^{(i,j)} }{\pi} \right) U^{(i,j)}
\end{equation}
where $h$ denotes the grid spacing, and Dirichlet boundary conditions
are assumed, i.e.\ $U^{(i,j)} = 0$ if $(i,j) \notin D$.
In this case $\Im H$ is diagonal, and $\omega$ and $\Delta t$ are
chosen based on the values
\begin{equation}
\begin{aligned}
\lambda_{\rm min}(\Re H) = {}& - k_{\rm max}^2
\\
\lambda_{\rm max}(\Re H) = {}& - k_{\rm min}^2 + 8 h^{-2} .
\end{aligned}
\end{equation}
where
\begin{equation}
k_{\rm max} := \max_{(i,j) \in D} k^{(i,j)} , \qquad \text{ and } \qquad
k_{\rm min} := \min_{(i,j) \in D} k^{(i,j)} .
\end{equation}
We find that $\omega = k_{\rm max}$
while $\Delta t$ is chosen from (\ref{eq:define_Deltat_ImH_diagonal}) using that
$\lambda_{\rm max}(A) = 8 h^{-2} + k_{\rm max}^2 - k_{\rm min}^2$.
The scheme in the computational time domain, can be written as
\begin{equation} \label{eq:discrete_scheme_2ndOrderFd}
\begin{aligned}
{}& \frac{\alpha}{\Delta t^2} (u_{n+1}^{(i,j)} - 2 u_n^{(i,j)} + u_{n-1}^{(i,j)} )
+ \frac{\beta}{2 \Delta t} \frac{(k^{(i,j)})^2 \tilde{R}^{(i,j)}}{\pi k_{\rm max}}
( u_{n+1}^{(i,j)} - u_{n-1}^{(i,j)} )
\\
{}& + h^{-2} \bigg( 4 u_n^{(i,j)} - u_n^{(i-1,j)} - u_n^{(i+1,j)} - u_n^{(i,j-1)} -
u_n^{(i,j+1)} \bigg) + ( k_{\rm max}^2 -(k^{(i,j)})^2 ) u_n^{(i,j)} =
f_n^{(i,j)} .
\end{aligned}
\end{equation}
The scheme (\ref{eq:discrete_scheme_2ndOrderFd}) differs substantially
from the standard second order FDTD scheme. In fact, it is a
discretization of the PDE
\begin{equation} \label{eq:modified_PDE}
\pdpd{^2 u}{t^2} - \Delta u
+ \frac{k^2 \tilde{R}}{\pi k_{\rm max}} \pdpd{u}{t}
+ \left( k_{\rm max}^2- k^2 \right) u = f .
\end{equation}
This is not necessarily the case. If the operator $H$ would
be rescaled, multiplying from the left and the right by a diagonal
matrix with entries $c^{(i,j)}$ on the diagonal, then the resulting
computational time domain scheme would be close to a standard FDTD
scheme.
This leads to a different behavior of the CFL bound and the minimum
number of timesteps per period. In standard schemes the CFL bound is
chosen based on the maximum velocity, but the grid spacing based on
the minimum velocity, this results in a rough equality.
\begin{equation}
(\text{\# timesteps per period})
\lesssim \frac{c_{\rm max}}{c_{\rm min}}
(\text{min \# gridpoints per wavelength})
\end{equation}
For the new scheme (\ref{eq:discrete_scheme_2ndOrderFd}) we have
\begin{equation} \label{eq:timesteps_per_period_approx_bound_newscheme}
(\text{\# timesteps per period})
\lesssim (\text{min \# gridpoints per wavelength}) .
\end{equation}
For standard schemes, the CFL bound is locally suboptimal, where the
velocity is below the maximum. For the
modified PDE (\ref{eq:modified_PDE}) on the other hand, the CFL bound is
locally optimal or near-optimal throughout the domain.
\subsection{An optimized finite-difference method}
\label{subsec:optimized-fd}
Next we consider a discretization with a 27 point cubic
stencil (in 3-D) that was described in
\cite{stolk2016dispersion}.
There exists different variants of such compact stencil methods, see
among others \cite{BabuskaEtAl1995,sutmann2007compact,TurkelEtAl2013}.
For each gridpoint, the number of neighbors that the gridpoint
interacts with is relatively small (when compared e.g.\ to
higher order finite elements). Even so, the dispersion errors of these
schemes are also relatively small, as shown in \cite{stolk2016dispersion},
so that these methods can be used with relatively coarse meshes%
\footnote{In some applications relatively coarse meshes are used to
save on computation time. Reduction of dispersion errors is
important in this case. In general, other discretization
errors can also be present, e.g.\ related to the presence
of variable coefficients.}.
The equation discretized in \cite{stolk2016dispersion} was the real
part of (\ref{eq:damped_Helmholtz_example}).
We will discuss here the situation with constant $k$, for variable $k$
we refer to appendix~\ref{app:opt_fd}.
The stencil is the 27 point cube that we will denote with
$\{ -1,0,1\}^3$.
For constant $k$, due to symmetry there are four different
matrix coefficients. The values of these coefficients are $h^{-2} f_s(
\frac{h k}{2\pi} )$, for $s=0,1,2,3$ respectively, where the dimensionless
function $f_s$ describes the coefficient as a function of the
dimensionless quantity $\frac{h k}{2\pi} $. To be precise,
\begin{equation}
(\Re H)^{(i,j,l; p,q,r)} =
\left\{
\begin{array}{ll}
\frac{1}{h^2} f_{|i-p|+|j-q|+|l-r|}( \frac{h k}{2 \pi} )
& \text{if $(i-p,j-q,l-r) \in \{-1,0,1\}^3$}
\\
0 & \text{otherwise}
\end{array}
\right.
\end{equation}
In \cite{stolk2016dispersion} the $f_s$ are precomputed functions, in
a class of piecewise polynomial functions, obtained by optimization to
minimize phase errors.
For the 2-D case, in \cite{BabuskaEtAl1995} explicit
expressions for the optimal functions $f_s$, $s=0,1,2$ were derived.
The imaginary part was discretized as above, i.e.\ it was diagonal
with entries
$\frac{(k^{(i,j,l)})^2 \tilde{R}^{(i,j,l)}}{\pi}$.
Because $\Im H$ is diagonal, the integrator $I_{\rm acd}$ is used.
The discretization in \cite{stolk2016dispersion} has in addition the
property that it was second order and such that
\begin{equation}
\text{$\Re H + k^2 I$ is symmetric positive definite} .
\end{equation}
Thus in case of constant $k$
\begin{equation} \label{eq:omega_k}
\omega = k .
\end{equation}
The upperbound for $\Re H$ that determines the maximum for $\Delta t$
is given by
\begin{equation} \label{eq:upperbound_HoptimFd}
\lambda_{\rm max}(\Re H) = f_0( \frac{k h}{2\pi} )
- 6 f_1( \frac{k h}{2\pi} )
+ 12 f_2( \frac{k h}{2\pi} )
- 8 f_3( \frac{k h}{2\pi} ) .
\end{equation}
This bound is typically somewhat less than the bound associated with
second order finite differences with the same parameters.
It is straightforward to follow the recipe of
subsection~\ref{subsec:choose_semidiscrete}. For this scheme we have
again (\ref{eq:timesteps_per_period_approx_bound_newscheme}), and it
can be considered a discretization of (\ref{eq:modified_PDE}) rather
than of (\ref{eq:damped_wave_equation}).
\section{Analysis}
\label{sec:analysis}
Here theoretical results will be obtained concerning the stability of
the time integrators defined in
subsection~\ref{subsec:time-discretization}, the solutions of these
time-integration schemes and and the convergence of
the approximate solution operators defined in
subsection~\ref{subsec:time-domain_approx_sol}.
The consequences of the stability analysis for the choice of $\Delta
t$ were already discussed in subsection~\ref{subsec:choose_semidiscrete}.
\subsection{Stability}
\label{subsec:stability}
To study the conditions under which (\ref{eq:cd_integrator})
and (\ref{eq:acd_integrator})
are stable, we study the growth of the solutions to the homogeneous recursion
\begin{equation} \label{eq:homogeneous_recursion_cd}
( 1 + \tfrac{1}{2} L ) u_{n+1} + (- 2+K) u_n + (I - \tfrac{1}{2} L) u_{n-1} = 0 .
\end{equation}
Throughout we will assume that $K$ and $L$ are symmetric matrices and
that $L$ is positive semidefinite. The basic idea is to show that the
following energy functions stays bounded
\begin{equation} \label{eq:define_Ecd}
E_{\rm cd}(n-1/2) = \langle u_n - u_{n-1} , (4I - K) (u_n - u_{n-1} ) \rangle
+ \langle u_n + u_{n-1} , K (u_n + u_{n-1} ) \rangle .
\end{equation}
Here $\langle \cdot , \cdot \rangle$ denotes the standard inner
product. If the energy function is coercive, than the solution also
stays bounded. The following theorem, states the stability conditions
that follow from such an analysis. The conditions under (\ref{it:item_ii_stability_cd})
in the theorem were already mentioned above, see equations
(\ref{eq:stability_cond_K_L}), (\ref{eq:stability_cond_cd_upper}).
\begin{theorem} \label{th:stability_cd}
Let $u_n$, $n \in {\mathbb{Z}}$ be a solution to
the homogeneous recursion (\ref{eq:homogeneous_recursion_cd}),
where $K$ and $L$ are real valued symmetric matrices and $L$ is
positive semidefinite.
\begin{enumerate}[(i)]
\item
If $K$ and $4I - K$ are positive definite then $E_{\rm cd}$ is equivalent
to a norm on ${\mathbb{R}}^{2N}$. If $L = 0$ then $E_{\rm cd}$ is conserved and
solutions remain bounded if $t \to \pm \infty$. If $L$ is positive
semidefinite then $\Delta E_{\rm cd}(n) \le 0$
and solutions remain bounded if $t \to \infty$.
\item \label{it:item_ii_stability_cd}
If $K$ is positive semidefinite and $4I - K$ is positive definite
and $K$ has one or more zero eigenvalues then $E_{\rm cd}$ is not equivalent to a
norm. If $L = 0$ then $E_{\rm cd}$ is
conserved and solutions grow at most linearly if $t \to \pm \infty$.
If $L$ is positive semidefinite, then $\Delta E_{\rm cd}(n) \le 0$ and
solutions grow at most linearly if $t \to \infty$.
\end{enumerate}
\end{theorem}
\begin{proof}
To prove the result, a bound is derived for
\begin{equation}
\Delta E_{\rm cd}(n) = E_{\rm cd}(n+1/2) - E_{\rm cd}(n-1/2) .
\end{equation}
Using that $K$ is symmetric in combination with basis rules for
standard products results in
\begin{equation}
\begin{aligned}
\Delta E_{\rm cd}(n) = {}&
\langle u_{n+1} - 2 u_n + u_{n-1} , (4I - K) ( u_{n+1}
- u_{n-1} ) \rangle
\\
{}& +
\langle u_{n+1} - u_{n-1} , K ( u_{n+1} + 2 u_n
+ u_{n-1} ) \rangle .
\end{aligned}
\end{equation}
By using the recursion equation in two places results one obtains the estimate
\begin{equation}
\begin{aligned}
\Delta E_{\rm cd}(n) = {}&
\langle -K u_n - \tfrac{1}{2} L(u_{n+1} - u_{n-1} ) , (4I - K) ( u_{n+1}
- u_{n-1} ) \rangle
\\
{}& +
\langle u_{n+1} - u_{n-1} ,
4 K u_n + K ( -K u_n - \tfrac{1}{2} L(u_{n+1} - u_{n-1} ) ) \rangle
\\
= {}&
- 2 \langle u_{n+1} - u_{n-1} ,
L ( u_{n+1} - u_{n-1} ) \rangle
\\
\le {}& 0 ,
\end{aligned}
\end{equation}
where the last inequality holds because by assumption $L$ is positive semidefinite.
The theorem follows from this result.
\end{proof}
\medskip
To establish stability of backward differences damped leapfrog, we study
the growth of solutions to the recursion
\begin{equation} \label{eq:homogeneous_recursion_bd}
u_{n+1} + (- 2+K + L) u_n + (I - L) u_{n-1} = 0
\end{equation}
using the energy function
\begin{equation}
\begin{aligned}
E_{\rm bd}(n-1/2) = {}& \langle u_n - u_{n-1} , (4I - K) (u_n - u_{n-1} ) \rangle
+ \langle u_n + u_{n-1} , K (u_n + u_{n-1} ) \rangle
\\
{}& - 2 \langle u_n - u_{n-1}, L (u_n - u_{n-1}) \rangle .
\end{aligned}
\end{equation}
\begin{theorem}
Let $u_n$, $n \in {\mathbb{Z}}$ be a solution to
the homogeneous recursion (\ref{eq:homogeneous_recursion_bd}),
where $K$ and $L$ are real valued symmetric matrices.
\begin{enumerate}[(i)]
\item
If $K$ and $4I - K - 2L$ are positive definite, then $E_{\rm bd}$ is equivalent
to a norm on ${\mathbb{R}}^{2N}$. If $L = 0$ then $E_{\rm bd}$ is conserved and
solutions remain bounded if $t \to \pm \infty$. If $L$ is positive
semidefinite then $\Delta E_{\rm bd}(n) \le 0$
and solutions remain bounded if $t \to \infty$.
\item
If instead $K$ is positive semidefinite with one or more zero eigenvalues
and $4I - K - 2L$ is positive definite, then $E_{\rm bd}$ is not equivalent to a
norm. If $L = 0$ then $E_{\rm bd}$ is
conserved and solutions grow at most linearly if $t \to \pm \infty$,
If $L$ is positive semidefinite, then $\Delta E_{\rm bd}(n) \le 0$ and
solutions grow at most linearly if $t \to \infty$.
\end{enumerate}
\end{theorem}
\begin{proof}
Similarly as in (\ref{eq:define_Ecd}) define
\begin{equation} \label{eq:define_Ebd}
\Delta E_{\rm bd}(n) = E_{\rm bd}(n+1/2) - E_{\rm bd}(n-1/2) .
\end{equation}
Using that $K$ is symmetric in combination with basis rules for
standard products results in
\begin{equation}
\begin{aligned}
\Delta E_{\rm bd}(n) = {}&
\langle u_{n+1} - 2 u_n + u_{n-1} , (4I - K) ( u_{n+1}
- u_{n-1} ) \rangle
\\
{}& +
\langle u_{n+1} - u_{n-1} , K ( u_{n+1} + 2 u_n
+ u_{n-1} ) \rangle
\\
{}& - 2 \langle u_{n+1} - u_n, L (u_{n+1} - u_n) \rangle
+ 2 \langle u_n - u_{n-1}, L (u_n - u_{n-1}) \rangle .
\end{aligned}
\end{equation}
By using the recursion equation in two places results one obtains the estimate
\begin{equation}
\begin{aligned}
\Delta E_{\rm bd}(n) = {}&
\langle -K u_n - L(u_n - u_{n-1} ) , (4I - K) ( u_{n+1}
- u_{n-1} ) \rangle
\\
{}& +
\langle u_{n+1} - u_{n-1} ,
4 K u_n + K ( -K u_n - L(u_n - u_{n-1} ) ) \rangle
\\
{}& - 2 \langle u_{n+1} - u_n, L (u_{n+1} - u_n) \rangle
+ 2 \langle u_n - u_{n-1}, L (u_n - u_{n-1}) \rangle
\\
= {}&
- 4 \langle u_{n+1} - u_{n-1} ,
L ( u_n - u_{n-1} ) \rangle
\\
{}& - 2 \langle u_{n+1} - u_n, L (u_{n+1} - u_n) \rangle
+ 2 \langle u_n - u_{n-1}, L (u_n - u_{n-1}) \rangle
\\
={}&
-2 \langle u_{n+1} - u_{n-1}, L ( u_{n+1} - u_{n-1} ) \rangle
\\
\le {}& 0 ,
\end{aligned}
\end{equation}
where the last inequality holds because by assumption $L$ is positive semidefinite.
The theorem follows from this result.
\end{proof}
\subsection{Eigenvalue analysis}
An analysis based on the eigenvalues of $K$ gives insight
in the question whether these conditions are necessary.
Suppose $v$ is an eigenvector of $K$ with eigenvalue
$k$ and simultaneously an eigenvector of $L$ with
eigenvalue $\ell$. One can think of the situation that
$L = \ell I$. We start by looking for solutions to equation
(\ref{eq:homogeneous_recursion_cd})
of the form $u_n = \Lambda^n v$.
The constant $\Lambda$ must then satisfy
\begin{equation} \label{eq:scalar_recursion_cd}
(1 + \ell /2) \Lambda^2 + (-2+k) \Lambda + (1- \ell /2) = 0 .
\end{equation}
The solution to this equation are
\begin{equation} \label{eq:Lambda_pm_stability}
\Lambda_\pm = \frac{-1+k/2}{1+\ell/2}
\pm \frac{1}{1+\ell/2} \sqrt{ -k + k^2/4 + \ell^2/4}
\end{equation}
We consider the growth of solutions, first if $\ell = 0$ (no damping).
Then if $0 < k < 4$ both solutions are complex with
$| \Lambda_\pm| = 1$. If $k=0$ then
(\ref{eq:homogeneous_recursion_cd}) has solutions
$u_n = C_1 v + C_2 n v$. If $k<0$ or $k>4$ at least one of
$\Lambda_\pm$ has modulus larger than 1. Next consider the case
$\ell > 0$ (damping present). If $0 < k < 4$ then $\Lambda_\pm$ can
be complex (small damping) or real (large damping), with
$| \Lambda_\pm | < 1$ in both cases. If $k=0$ or $k=4$ one
$\Lambda _\pm$ is equal to 1 and the other smaller than 1. If $k<0$ or
$k>4$, exponentially growing solutions exist.
For stability, $K$ and $4I - K$ need to be positive semidefinite, but
we will require $4I - K$ to be strictly positive definite (to avoid
the solution with $\Lambda = -1$ while $k=4$, $\ell>0$).
\medskip
Next we study the undamped solutions in case the eigenvectors of $L$
are not those of $K$. In this case, the recursion
(\ref{eq:homogeneous_recursion_cd}) can be written in first order
form. Let
\begin{equation}
y_n = \begin{bmatrix} u_{n-1} \\ u_n \end{bmatrix} , \qquad
g_n = \begin{bmatrix} 0 \\ f_n \end{bmatrix} ,
\end{equation}
and
\begin{equation}
\Xi =
\begin{bmatrix}
0 & I \\
- (I+L)^{-1} (I-L)
& (I+L)^{-1} (2I-K)
\end{bmatrix}
\end{equation}
then the integrators $I_{\rm cd}$ and $I_{\rm acd}$ from
definitions~\ref{def:Icd_Ibd} and~\ref{def:Iacd_Iabd}
can be written as
\begin{equation} \label{eq:system_dim2N}
y_{n+1} = \Xi y_n + g_n .
\end{equation}
The properties of the solutions are directly related to the
eigenvalues, (generalized) eigenvectors and the Jordan blocks of
$\Xi$. We next list some properties of $\Xi$ that relate to
eigenvalues $\xi$ of $\Xi$ with $| \xi | = 1$. These correspond to
undamped solutions. The basic idea is to
show that there are not more of these than there should be.
\begin{theorem}
Assume (\ref{eq:stability_cond_K_L}),
(\ref{eq:stability_cond_cd_upper}). The following properties hold
for the eigenvalues, (generalized) eigenvectors and Jordan blocks of
$\Xi$:
\begin{enumerate}[(i)]
\item \label{it:i}
all eigenvectors are of the form $[u,\xi u]$, with $u \in {\mathbb{C}}^N$;
\item \label{it:ii}
there are no eigenvalues $\xi = -1$;
\item \label{it:iii}
for $|\xi| = 1$, $\xi \neq 1$ there are no Jordan blocks of size $>1$;
\item \label{it:iv}
for $|\xi| = 1$, $\xi \neq 1$ the $u$ in $[u,\xi u]$ is in $\ker
L$, and is an eigenvector of $K$ and
$\xi$ follows from (\ref{eq:Lambda_pm_stability});
\item \label{it:v}
for $\xi = 1$, there are no Jordan blocks of size $>2$,
an eigenvector is of the form
$v = [ u, u]$ with $u \in \ker K$ and if $w$ is a generalized
eigenvector such that $\Xi w = w
+ v$, $v$ an eigenvector, then $v = [u,u]$ with $u$ also in $\ker L$, and
$w = [ w_1 , w_1 + u ]$ with $w_1 \in \ker K$.
\end{enumerate}
\end{theorem}
\begin{proof}
Claim (\ref{it:i}) is obvious.
Regarding (\ref{it:ii}), the equations $\Xi [ u ,\xi u]^T = \xi [ u ,\xi u]^T$
with $\xi = -1$ gives $(4I - K ) u = 0$ which is impossible because
$4I - K$ is positive definite by assumption.
Claim (\ref{it:iii}) follows from the energy growth equations.
Solution to second order recursion would be of the form
$u_n = \xi^n w + n \xi^{n-1} v$, and the dominant term in the energy in the
limit $n \to \infty$ would be
$\langle n (1-\xi) v , (4I - K) n (1- \xi) v \rangle$. This term would
be unbounded as $n \to \infty$ which is impossible, hence any Jordan
block of size $>1$ must be associated with an eigenvalue $\xi = 1$.
In the situation of claim (\ref{it:iv}), solutions are of the form $u_n = \xi^n
v$ with $v \in {\mathbb{C}}^N$. Energy must be conserved hence
$\langle u_{n+1} - u_{n-1} , L (u_{n+1} - u_{n-1}) \rangle = 0$ hence
hence $(\xi - \xi^{-1} ) u \in \ker L$, hence $u \in \ker L$. But then
it follows that $K u = ( - \xi^{-1} + 2 - \xi) u$, so that $u$ is a
simultaneous eigenvector of $K$, and $L$.
The first part of part (\ref{it:v}) follows from the bound on the
energy. Jordan blocks of size $>2$ would lead to solutions of the form
\begin{equation}
u_n = w^{(2)} + n w^{(1)} + n (n-1) v ,
\end{equation}
The dominant contribution to the term
$\langle u_n-u_{n-1} , (4I-K) (u_n-u_{n-1}) \rangle$ would be
$\langle 2(n-1) v , (4I-K) 2 (n-1) v \rangle$, which would be
unbounded, which is not possible.
The second part of (\ref{it:v}) follows from the form of $\Xi$.
If eigenvector is $V = [ v_1, v_2]^T$, $v_j \in {\mathbb{C}}^N$, then $v_1 = v_2$ and
\begin{equation}
-(1-L) v_1 + (2I-K) v_1 = (1+L) v_1
\end{equation}
hence $K v_1 = 0$. The third part of (\ref{it:v}) follows from the
form of $\Xi$. Working out the first line of equation $\Xi w = w + v$
yields $w_2 = w_1 + u$, working out the second line gives $L u = 0$.
\end{proof}
\subsection{Convergence of the approximate solution operators $S_T^{\rm
acd}$ and $S_T^{\rm abd}$}
In this subsection we will write $u( n \Delta t)$ instead of
$u_n$ and $f( n \Delta t)$ instead of $f_n$. Instead of the two-sided
sequences $(u_n)$ and $(f_n)$ we write $u(t)$, $f(t)$, with $t \in \Delta t {\mathbb{Z}}$. The Fourier
transform from functions on $\Delta t {\mathbb{Z}}$ to functions on
${\mathbb{T}}_{2\pi/\Delta t}$ is as defined in
appendix~\ref{app:Fourier_analysis_grid}.
Associated with the integrator $I_{\rm acd}$ is a second order
difference operator $C_{\rm acd}$. We define this operator as follows
\begin{equation} \label{eq:operator_Cacd}
C_{\rm acd}(u)(t) = \frac{\alpha}{\Delta t^2}
\left[ ( 1 + \tfrac{1}{2} L ) u(t+\Delta t) + (- 2+K) u(t) + (I -
\tfrac{1}{2} L) u(t-\Delta t) \right]
\end{equation}
where $K, L$ are as defined in (\ref{eq:K_L_frequency_adapted_central_differences}).
Solutions to the time-integration satisfy
\begin{equation} \label{eq:difference_equation_acd}
C_{\rm acd}(u)(t) = f(t) .
\end{equation}
The left-hand side can be see as the convolution of $u$ with a kernel
$\Gamma_{\rm acd}$ that is a matrix valued function on $\Delta t {\mathbb{Z}}$.
(The explicit expression for
$\Gamma_{\rm acd}$ is easily derived from (\ref{eq:operator_Cacd}).)
In the Fourier domain the convolution becomes a multiplication and we have
\begin{equation}
\widehat{\Gamma}_{\rm acd}(\tau) \widehat{u}(\tau) =
\widehat{f}(\tau) .
\end{equation}
From
proposition~\ref{prop:reproduce_symbol_at_omega} it follows that
\begin{equation} \label{eq:symbol_equality_property_acd}
\widehat{\Gamma}_{\rm acd}(\omega) = - \omega^2 I + i \omega B + A .
\end{equation}
In addition we know that $\widehat{\Gamma}$ is $C^\infty$, because
$\Gamma$ is nonzero only for $t \in \{ -\Delta t, 0 , \Delta t \}$.
By $\Phi_{\rm acd}(t)$, $t \in \Delta t {\mathbb{Z}}$ we will denote the causal
Green's function associated with
(\ref{eq:difference_equation_acd}). It is a function
$\Delta t {\mathbb{Z}} \to {\mathbb{C}}^{N \times N}$
defined as the solution to
\begin{equation}
C_{\rm acd} \Phi_{\rm acd} (t) = I \delta(t) ,
\qquad \text{and} \qquad
\Phi_{\rm acd}(t) = 0 \text{ if } t \le 0 ,
\end{equation}
where $\delta(t)$ is as defined in
appendix~\ref{app:Fourier_analysis_grid}.
Assuming that the conditions in
(\ref{eq:stability_cond_K_L}), (\ref{eq:stability_cond_cd_upper})
hold, $\Phi_{\rm acd}$ grows at most polynomially if $t \to \infty$
(by Theorem~\ref{th:stability_cd}), and
its Fourier transform $\widehat{\Phi}_{\rm acd}$ is well defined as a matrix valued
distribution on the circle ${\mathbb{T}}_{2\pi/\Delta t}$.
We will first show that the approximate solution operator
$S_T^{\rm acd}$ can be expressed as a weighted mean of
$\widehat{\Phi}(\tau)$ around $\tau = \omega$. For this purpose we
define a helper function $\phi_\epsilon$, depending on $\chi$, as
follows
\begin{equation}
\psi_T(t) = \chi(1 - \left| \frac{t}{2 \pi \omega^{-1} T} \right| ) , \qquad \text{and} \qquad
\phi_\epsilon(\tau) = \frac{1}{2 \pi} \mathcal{F} \psi_{1/\epsilon}(\tau) .
\end{equation}
Note that $\phi_\epsilon$ is an approximation to the identity. We have
the following result
\begin{theorem} \label{th:convolution_formula_S_T_acd}
Assume $\phi_\epsilon$ and $\Phi_{\rm acd}$ are as just defined, then
\begin{equation} \label{eq:convolved_Fourier_transform_acd}
S_{T}^{\rm acd} = \phi_{1/T} \ast \widehat{\Phi}_{\rm acd}(\omega) .
\end{equation}
\end{theorem}
In words, $s_T^{\rm acd}$ equals the convolution product of
$\phi_{1/T}$ and $\widehat{\Phi}_{\rm acd}$, evaluated at $\tau =
\omega$. The convolution product is taken on the torus
${\mathbb{T}}_{2\pi/\Delta t}$.
\begin{proof}
Let $\tilde{T} = 2\pi \omega^{-1} T$.
By linearity, the time-domain approximate solution $S_T(F)$ satisfies
\begin{equation}
S_T^{\rm acd}(F) = e^{- i \omega \tilde{T}}
\Delta t \sum_{t \in \Delta t {\mathbb{Z}}, 0 \le t < \tilde{T}} \Phi(\tilde{T}-t)
\chi(t/\tilde{T}) e^{i \omega t} F
\end{equation}
Inserting the definition of $\psi_{T}$ this becomes
\begin{equation}
S_T^{\rm acd}(F) = \left[
\Delta t \sum_{t \in \Delta t {\mathbb{Z}}} \Phi(\tilde{T}-t) \psi_T(\tilde{T}-t)
\chi(t/\tilde{T}) e^{-i \omega (\tilde{T}-t)} \right] F
= \left[
\mathcal{F} ( \psi_T \Phi )(\omega)
\right] F .
\end{equation}
The equality (\ref{eq:convolved_Fourier_transform_acd}) follows because
the Fourier transform of the product is $\frac{1}{2\pi}$ times
the convolution product of the Fourier transforms,
see (\ref{eq:FourierTransformFG}).
\end{proof}
Concerning the convergence of the time-domain approximate solutions to the
correct solutions we have the following result.
\begin{theorem} \label{th:convergence_acd}
If
\begin{equation} \label{eq:H_non-singular}
\text{$H = -\omega^2 I + i \omega B + A$ is non-singular}
\end{equation}
and $A,B,\omega$ and $\Delta t$ are such that $K$, $L$ defined in
(\ref{eq:K_L_frequency_adapted_central_differences}) satisfy
the stability conditions (\ref{eq:stability_cond_K_L}) and (\ref{eq:stability_cond_cd_upper})
then
\begin{equation}
\lim_{T \to \infty} S_T^{\rm acd} = H^{-1} .
\end{equation}
\end{theorem}
\begin{proof}
The Fourier transform $\widehat{\Phi}_{\rm acd}$ satisfies
\begin{equation}
\widehat{\Gamma}_{\rm acd}(\tau) \widehat{\Phi}_{\rm acd}(\tau) = I .
\end{equation}
Note that
$\widehat{\Gamma}_{\rm acd}(\tau)$ is a $C^\infty$ function and
$\widehat{\Phi}_{\rm acd}$ is a distribution and that this equality
is in the sense of distributions.
By the assumption (\ref{eq:H_non-singular}) and the property
(\ref{eq:symbol_equality_property_acd}), the matrix valued function
$\widehat{\Gamma}_{\rm acd}(\tau)$ is non-singular on a neighborhood
of $\tau = \omega$. It follows that
$\widehat{\Phi}_{\rm acd}(\tau)$ is continuous on a neighborhood of
$\tau = \omega$ and that
\begin{equation}
\widehat{\Phi}_{\rm acd}(\omega)
= \widehat{\Gamma}_{\rm acd}(\omega)^{-1}
= H^{-1} .
\end{equation}
Because $\phi_\epsilon$ is an approximation to the identity and
$\widehat{\Phi}(\tau)$ is continuous at $\tau = \omega$,
it follows from theorem~\ref{th:convolution_formula_S_T_acd} and
standard results in integration theory and distribution theory that
\begin{equation}
\lim_{T \to \infty} S_T^{\rm acd} = \widehat{\Phi}(\omega) = H^{-1} . \qedhere
\end{equation}
\end{proof}
\medskip
A similar analysis can be performed for the approximate solution
operator $S_T^{\rm abd}$. The equivalent of (\ref{eq:operator_Cacd})
is
\begin{equation} \label{eq:operator_Cabd}
C_{\rm abd}(u)(t) = \frac{\alpha}{\Delta t^2}
\left[ u(t+\Delta t) + (- 2+K+L) u(t)
+ (I - L) u(t-\Delta t) \right]
\end{equation}
where $K,L$ are as defined in (\ref{eq:K_L_frequency_adapted_backward_differences}).
A causal Green's function $\Phi_{\rm abd}$ is defined satisfying
\begin{equation}
C_{\rm abd} \Phi_{\rm abd} (t) = I \delta(t) ,
\qquad \text{and} \qquad
\Phi_{\rm abd}(t) = 0 \text{ if } t \le 0 ,
\end{equation}
The following theorems are proved in the same ways
as theorems~\ref{th:convolution_formula_S_T_acd} and
\ref{th:convergence_acd}
\begin{theorem} \label{th:convolution_formula_S_T_abd}
Assume $\phi_\epsilon$ and $\Phi_{\rm abd}$ are as just defined, then
\begin{equation} \label{eq:convolved_Fourier_transform_abd}
S_{T}^{\rm abd} = \phi_{1/T} \ast \widehat{\Phi}_{\rm abd}(\omega) .
\end{equation}
\end{theorem}
\begin{theorem} \label{th:convergence_abd}
If
\begin{equation} \label{eq:H_non-singular_abd}
\text{$H = -\omega^2 I + i \omega B + A$ is non-singular}
\end{equation}
and $A,B,\omega$ and $\Delta t$ are such that $K$, $L$ defined in
(\ref{eq:K_L_frequency_adapted_backward_differences}) satisfy
the stability conditions (\ref{eq:stability_cond_K_L}) and (\ref{eq:stability_cond_bd_upper})
then
\begin{equation}
\lim_{T \to \infty} S_T^{\rm abd} = H^{-1} .
\end{equation}
\end{theorem}
\section{Numerical examples}
\label{sec:numerical_examples}
In this section we study some examples. The examples all involve an
optimized finite-difference discretization described in
subsection~\ref{subsec:optimized-fd} in two and three dimensions.
To study the 2-D method, a simple implementation in Julia was made.
For the 3-D case a mixed Julia/C implementation was made. In this case
the time stepping was done in C, the computation of coefficients and
the GMRES iteration were done in Julia. In the C implementation, for
each grid point and timestep, 14 real coefficients had to be read, of
which one also had to be written back to. The C implementation used AVX
extensions but was otherwise straightforward. In most computations
double precision numbers were used. The exception was the
time-stepping performed in C. We found that the time-domain preconditioner
could also be run in single precision, since the GMRES solver would
automatically take care of rounding errors in subsequent iterations.
\subsection{2-D examples}
We considered three 2-D examples: A constant velocity model and two
piecewise constant models defined on the unit square. The first of the
two piecewise constant models allows for resonances in the circular
region. On the exterior of the model damping layers of thickness 32
gridpoints were added to simulate an unbounded domain.
The simulations in these examples
were done using a minimum of 6 points per wavelength. In each case
this resulted in 8 timesteps per period.
See
figure~\ref{fig:model_and_sols2D} for non-constant velocity models and
some solutions in these models.
The first objective was to the study the convergence of the
approximate solution $S_T^{\rm acd} F$ to $H^{-1} F$. For this we took models of size
$320\times320$ and $640 \times 640$ (excluding damping layers), and
let the right-hand side be a point source.
The ``taper'' parameter $\rho$ was varied, we considered values in
$\{ 0, 0.25, 0.5, 0.75 \}$. The value $0$ was not exactly zero, in
this case the time-harmonic right-hand side grew from 0 to 1 over one
period.
The convergence (relative difference between approximate and true solutions)
is given in Figure~\ref{fig:large_time_convergence}. Note the
differing axes in the Figure~\ref{fig:large_time_convergence}(b).
The main conclusions from these figures are that the convergence in
the resonant model is relatively poor, and that the choice of $\rho = 0$ (no
tapering) leads to poor convergence. The detailed amount of tapering
is less important. The lower value ($\rho = 0.25$) appears to
perform best unless extremely small relative errors are required.
\newcommand{\addsubfigurelabel}[3]{%
\begin{minipage}{#2}
\begin{center}
#1\\
#3
\end{center}
\end{minipage}}
\begin{figure}
\begin{center}
\addsubfigurelabel{(a)}{6cm}{%
\includegraphics[width=6cm]{ vel2.pdf}}
\addsubfigurelabel{(b)}{6cm}{
\includegraphics[width=6cm]{ vel3.pdf}}
\addsubfigurelabel{(c)}{6cm}{%
\includegraphics[width=6cm]{ sol_vel2_1.pdf}}
\addsubfigurelabel{(d)}{6cm}{
\includegraphics[width=6cm]{ sol_vel3_1.pdf}}
\end{center}
\caption{Two velocity models (a,b) and solutions in these models
(c,d).
A relatively small example of $240 \times 240$ was used so that the
wavefield oscillations are still visible.
}\label{fig:model_and_sols2D}
\end{figure}
\begin{figure}
\begin{center}
\addsubfigurelabel{(a)}{65mm}{%
\includegraphics[width=65mm]{ largeTime_velocity1_loglog.pdf}} \hspace*{5mm}
\addsubfigurelabel{(b)}{65mm}{%
\includegraphics[width=65mm]{ largeTime_velocity2_loglog.pdf}}\\
\addsubfigurelabel{(c)}{65mm}{%
\includegraphics[width=65mm]{ largeTime_velocity3_loglog.pdf}}\\
\end{center}
\caption{Convergence behavior of large time approximate
solutions. Note the different scales in figure (b).}
\label{fig:large_time_convergence}
\end{figure}
GMRES convergence for the three velocity models of size 640 gridpoints
(excluding damping layers) are given in
Figure~\ref{fig:GMRES_convergence}. In each case the number of periods
of the preconditioner was varied.
In Figures~\ref{fig:GMRES_convergence}(a-c) the taper parameter
was 0.25
and both the error indicator from GMRES and the true error are
plotted. In Figure~\ref{fig:GMRES_convergence}(d) the GMRES error was
plotted and the taper parameter was varied.
On the x axis, the number of iterations time the number of periods per
iteration was displayed. This is roughly, but not quite proportional
to the cost, the main difference being that one more preconditioner
application is needed for the right-hand side in the preconditioned
system.
In velocity models 1 and 3, there was no speedup compared to the
direct application of the time-domain solver. Neither was the method
much slower. However, in the resonant velocity model (velocity model
2), the performance using GMRES was substantially better than when the
time-domain solver was used directly.
In the examples about 50 to 100 periods for the preconditioner worked
best. This is approximately the size of the example in wavelenghts.
\begin{figure}
\begin{center}
\addsubfigurelabel{(a)}{65mm}{%
\includegraphics[width=65mm]{ krylov_velocity1_semilogy.pdf}} \hspace*{5mm}
\addsubfigurelabel{(b)}{65mm}{%
\includegraphics[width=65mm]{ krylov_velocity2_semilogy.pdf}}\\
\addsubfigurelabel{(c)}{65mm}{%
\includegraphics[width=65mm]{ krylov_velocity3_semilogy.pdf}} \hspace*{5mm}
\addsubfigurelabel{(d)}{65mm}{%
\includegraphics[width=65mm]{ krylov_velocity2_semilogy_tap.pdf}}
\end{center}
\caption{Preconditioned GMRES convergence for different models and different
parameter choices} \label{fig:GMRES_convergence}
\end{figure}
\subsection{Examples in 3-D}
We also studied the performance of the algorithm in three
dimensions. In this case, the velocity model was the SEG/EAGE Salt
Model. This is a standard, non-resonant synthetic model in the geophysics community.
In this case our goal was to obtain a first estimate of the
computational cost of the method. A minimum of 6 gridpoints per
wavelength was used. The Julia/C code was run on a 2019 MacBook Pro with
a 2.6 GHz 6-core Intel i7 processor and 16 GB main memory. A GMRES
error reduction with a factor $10^{-5}$ was required.
The results are in Table~\ref{tab:seg_eage_salt_example}.
\begin{table} \label{tab:seg_eage_salt_example}
\begin{center}
\begin{tabular}{| l | c | c | c |}
\hline
physical frequency & 2.5 Hz & 4 Hz & 6 Hz\\
problem size & $200 \times 200 \times 106$
& $280 \times 280 \times 130$
& $392 \times 392 \times 165$\\
degrees of freedom & 4.24 e6 & 1.02 e7 & 2.54 e7\\
time steps/period & 8 & 8 & 8 \\
periods/iteration & 25 & 40 & 60 \\
taper parameter & 0.25 & 0.25 & 0.25\\
iterations & 6 & 6 & 6 \\
computation time & 45 s & 166 s & 601 s\\ \hline
\end{tabular}
\end{center}
\caption{Computational results for the SEG/EAGE Salt model}
\end{table}
\section{Generalizations}
As outlined in the introduction, the new time-domain preconditioned
GMRES solver is based on four innovative ideas: A modified
semi-discrete system derived from the discrete Helmholtz equation
(equation (\ref{eq:general_semi-discrete_equation})), a
new time-integration scheme, introduction of a smooth window in the
computation of limiting-amplitude solutions and the use of the
time-domain approximate solution operator as a preconditioner.
Some of these ideas can be used separately. The new semi-discrete
system and the new integration method could be used with the exact
controllability or WaveHoltz methods.
In this section we discuss the
application of Krylov subspace method in combination with a different
semi-discrete equation and time integrator.
We assume the
semi-discrete system is still of the form
(\ref{eq:general_semi-discrete_equation}). Denoting
\begin{equation}
v = \begin{bmatrix} u \\ \pdpd{u}{t}
\end{bmatrix} ,
\qquad \;\;
g = \begin{bmatrix} 0 \\ f
\end{bmatrix} ,
\qquad \;\;
\Xi = \begin{bmatrix}
0 & I \\
-A & -B
\end{bmatrix}
\end{equation}
this can be written in first order form
\begin{equation} \label{eq:general_semi-discrete_first_order}
\pdpd{v}{t} = \Xi v + g .
\end{equation}
The first order system can be discretized for example using a standard
fourth order Runge-Kutta method, which yields
\begin{equation}
\begin{aligned}
v_{n+1} = {}& (I + h \Xi + \frac{h^2}{2} \Xi^2 + \frac{h^3}{6} \Xi^3
+ \frac{h^4}{24} \Xi^4) v_n
\\
{}& + \frac{h}{24} ( 4 + 4 h \Xi + 2 h^2 \Xi^2 + h^3 \Xi^3 ) g_n
+ \frac{h}{12} ( 8 + 4 h \Xi + h^2 \Xi^2) g_{n+1/2}
+ \frac{h}{6} g_{n+1}
\end{aligned}
\end{equation}
Requiring that $v_n = e^{i n \omega \Delta t} V$ and
$g_n = e^{i n \omega \Delta t} G$ leads to the following equation for
$V$ and $G$
\begin{equation}
\begin{aligned}
{}& \bigg[ e^{i\omega \Delta t} I
- I - h \Xi - \frac{h^2}{2} \Xi^2 - \frac{h^3}{6} \Xi^3
- \frac{h^4}{24} \Xi^4 \bigg] V =
\\
{}& \bigg[ \frac{h}{24} ( 4 I + 4 h \Xi + 2 h^2 \Xi^2 + h^3 \Xi^3 )
+ e^{i \omega \Delta /2} \frac{h}{12} ( 8 I + 4 h \Xi + h^2 \Xi^2)
+ e^{i \omega \Delta t} \frac{h}{6} I \bigg] G .
\end{aligned}
\end{equation}
This is the exact discrete Helmholtz equation satisfied by the
periodic solutions of the discrete-time-and-space system.
Based on the experience in this paper, this system could be solved by
GMRES preconditioned by a time-domain preconditioner
defined as in subsection~\ref{subsec:time-domain_approx_sol} but applied to
(\ref{eq:general_semi-discrete_first_order}) solved by RK4.
For different Runge-Kutta integrators, different polynomials in front
of $V$ and $G$ result, but the principle is the same.
\section{Concluding remarks}
In this paper we constructed a time-domain preconditioner for certain
indefinite linear systems, and studied its properties and behavior
analytically and with numerical examples. It should be emphasized
that the method does not compute in the physical time domain. We call
it a time-domain method, because of the similarity with classical
time-domain methods for time-harmonic waves.
For the practical application it would be useful to have further
examples and a comparison with alternatives, and more insight into how
the method behaves for different choices of $H$, e.g.\ for different
discretizations. We will make a few brief remarks in this direction.
First it is clear that the method requires relatively little
memory, compared to
alternatives that use LU (or LDL${}^T$) decompositions, such as
domain-decomposition and direct methods, see
\cite{stolk2017improved,poulson2013parallel} and references in
\cite{gander2019class}.
In the 3-D implementation here, with GMRES with restart $n_{\rm r}
= 10$, most memory was used for the approximately
$n_{\rm r} + 2$ complex double
precision vectors needed for GMRES.
This was in part because the timestepping was done in single precision, and
using real fields. Memory use could be somewhat reduced by setting
$n_{\rm r} = 5$ or using a different iterative method like BiCGSTAB.
For finite element discretizations the cost in general will be
different. These methods often have stricter CFL bounds, and there may
be more nonzero matrix elements, depending on the order of the finite
elements, which leads to larger cost. In case of unstructured meshes
computations in general are less efficient, because there is more
indirect memory access and cache misses. On the other hand, the size
of the grid cells can be adapted to the local velocity which means the
number of grid cells can be smaller.
It is difficult to compare performance of different methods, as
methods are run on different computer systems, with different examples
and implementations are optimized to different degrees. We refer to
\cite{appelo2019waveholtz,CalandraEtAl2013,liu2018solving,
poulson2013parallel,stolk2017improved,taus2019sweeps,WangDeHoopXia2011}
for some alternative methods. We believe that computation times of the
method as outlined are modest, and there is potential for further
improvements, by using GPUs or by reorganizing the code to better
make use of cached data. It remains challenging to get the most out
of modern computer hardware in time-domain finite-difference and
finite-element simulations. We hope that recent developments in this
area, cf.\ \cite{kukreja2016devito}, will lead to further improvements.
\bibliographystyle{abbrv}
|
\section{Introduction }
Topological materials are among the most intensely researched current topics, and encompass not just topological insulators~\cite{kane2005quantum,kane2005z,PhysRevLett.96.106802,bernevig2006quantum, fu2007topological,PhysRevLett.106.106802, hsieh2012topological,tanaka2012experimental,dziawa2012topological,PhysRevB.78.195125}
but also semi-metals with topological nontrivial nodal structures~\cite{burkov2016topological,doi:10.1146/annurev-matsci-070218-010049, doi:10.7566/JPSJ.87.041001, fang2016topological,doi:10.1146/annurev-matsci-070218-010023}. Such nodal materials have provoked widespread studies due to their unconventional density of states and bandstructure, which have lead to new or enhanced avenues for non-linear optical responses, electron tunneling behavior, higher harmonic generation, superconductivity and quantum Hall effects~\cite{RevModPhys.82.3045,doi:10.1146/annurev-conmatphys-031214-014501,RevModPhys.91.015006, ozawa2019topological, doi:10.7566/JPSJ.87.041001,dong2010quantum,pang2015evidence,wang20173d,molina2018surface,zhang2019cyclotron,zhang2019quantum,ang2020universal,bhattacharyya2020evidence}. The potentially intricate topology of nodal loops in 3D space have also inspired their design and realization in metamaterials and lossy or non-reciprocal media~\cite{lu2013weyl,xiao2015synthetic,lin2016photonic,yang2016acoustic,chen2016photonic,chang2017multiple,noh2017experimental,lin2017line,li2018weyl,yang2018ideal,gao2018experimental,lee2018tidal,xia2019observation,zhou2019exceptional,lee2020imaging,cerjan2019experimental,li2019emergence,jia2019observation}.
Despite their seemingly exotic nature, various Nodal line semimetal (NLSM) materials have recently been experimentally characterized, such as PbTaSe$_2$~\cite{bian2016topological, PhysRevB.93.245130}, BiTeI~\cite{crepaldi2012giant}, Mg$_3$Bi$_2$~\cite{doi:10.1021/acs.jpclett.7b02129,chang2019realization,zhou2019experimental}, ZrSiTe and ZrSiSe~\cite{hu2016evidence,PhysRevX.10.011026}, ZrSiS~\cite{schoop2016dirac,Fueaau6459,PhysRevB.100.085137}, BaAgAs~\cite{mardanya2019prediction}, TaN~\cite{PhysRevB.93.241202}, Ca$_3$P$_2$~\cite{chan20163,xie2015potential}, SrAs$_3$~\cite{song2020photoemission,li2018evidence}, CaAgX (X=P,As)~\cite{yamakage2016line,takane2018observation}, ZrGeX$_\text{c}$ (X$_\text{c}$= S, Se, Te) family~\cite{nakamura2019evidence, guo2019electronic}, magnetic semimetals EuB$_6$~\cite{PhysRevLett.124.076403}, spin gapless semimetals~\cite{zhang2020nodal}, the centrosymmetric superconductor SnTaS$_2$ ~\cite{chen2019superconducting} and PbTaS$_2$~\cite{gao2020superconducting}. In the form of closed loops, they have also been observed as nodal links in CaAuAs~\cite{PhysRevB.98.085122}, nodal chains in TiB$_2$ ~\cite{yi2018observation, liu2018experimental} and nodal line networks in RuO$_2$~\cite{jovic2019dirac}. Other materials such as Ti$_3$X (X=Al, Ga, Sn, Pb) family~\cite{PhysRevB.97.235150}, YH$_3$\cite{shao2018nonsymmorphic}, YoCoC$_2$~\cite{xu2019topological}, MnN~\cite{PhysRevMaterials.3.084201}, TiTaSe$_2$~\cite{bian2016drumhead}, CaP$_3$ family~\cite{PhysRevB.95.045136}, ABC-stacked graphdyiyne~\cite{graphdiyne}, layered X$_2$Y (X=Ca, Sr, Ba; Y=As, Sb, Bi)~\cite{niu2017two} have also been theorically proposed to possess
nodal loops.
In some of these NLSMs, the conduction and the valence bands intersect to form potentially very intricate structures, even more intricate than common examples in knot theory and prototypical nodal knot setups~\cite{AlexanderPolynomial,vortexKnot,dennis2017constructing,lee2020imaging,bode2019constructing}. For instance, the material Co$_2$MnGa was theorized~\cite{CMG_theory} to possess a complicated network of 3D band crossings characterized by various types of non-trivial nodal linkages and coupled chains enabled by perpendicular mirror planes. This was subsequently confirmed experimentally~\cite{belopolski2019discovery}. Another proposal of interlocking nodal chains may be realized in carbon networks, consisting of armchair graphene nanoribbon~\cite{LI2020563}. While transport and optical response properties are already well-studied for simpler NLSMs ~\cite{doi:10.1146/annurev-matsci-070218-010023,PhysRevLett.124.076403,PhysRevB.100.085137,yang2019conductivity,guo2019electronic, PhysRevB.99.241102,PhysRevB.95.245113,PhysRevB.102.035164}, those of more complicated nodal \emph{loop} semimetals are still not well understood. Recently, it was proposed that topological nodal linkages can significantly enhance optical response non-linearity and hence HHG (higher harmonics generation)~\cite{PhysRevB.102.035138}. However, the non-linear response does not afford any topological quantization, unlike linear response via the Kubo formula, and a complete understanding of the response properties of realistic nodal loop material necessitates a systematic study of how the geometry of the nodal structure, together with its topology, interplay and lead to various anisotropic and non-linear behavior.
As such, this work shall be concerned with a systematic investigation on how the relative shapes and linkage of nodal loops can lead to various anisotropic and non-linear responses components, and how the various responses come together into a bigger picture that reveals the overall nodal topology and geometry. Following a review of semi-classical response theory in Sect. II, we introduce a canonical model of two nodal loops whose linkage and shapes can be independently tuned. This model serves as an abstraction of the multiple nodal touchings and linkages in realistic nodal materials, for instance Co$_2$MnGa. In Sect. III, we study the various components in the response tensor of our model system, some whose non-linear properties have never been investigated. In Sect. IV, we show how these results help piece together a response surface whose evolution with field strength encapsulates the full information about the response anisotropy and non-linearity, and to some extent the nodal structure and its dispersion. For our model, the direction which the response surface points towards in the high field limit depends on whether the nodal structure is topologically linked.
\section{Non-linear Semi-Classical Optical Response}
To systematically relate the Hopf link response to its detailed momentum-space profile, we first review the theory of non-linear semi-classical response, and next introduce our model with tunable nodal linkage. Similar semi-classical approaches have been highly successful in explaining phenomena like Hall effects and quantum oscillations in diverse settings~\cite{xiao2010berry,price2016measurement,petrides2018six,lee2018electromagnetic,lohse2018exploring}, as well as Bloch oscillations and Berry curvature effects in the context of HHG~\cite{schubert2014sub,liu2017high}. HHG refers to the high harmonics generation induced by strong non-linear interactions when a very intense laser pulse is focused into a sample.
\subsection{Semi-Classical Response Theory}
The non-linear response of nodal materials emerges even at the semi-classical level, where many-body processes are simply encapsulated by a non-equilibrium occupation function $F(\varepsilon(\mathbf{k}-e\mathbf{A(t)}))$ that weighs different contributions to the response optical current~\cite{mikhailov2007non}:
\begin{equation}
\mathbf{J}=\int F(\varepsilon(\mathbf{k}-e\mathbf{A(t)}))\langle\mathbf{k}|\mathbf{\hat J}|\mathbf{k}\rangle d^3\mathbf{k}.
\label{J}
\end{equation}
Here, $F(\varepsilon)=(1+e^{\beta(\varepsilon(\mathbf{k})-\mu)})^{-1}$ is the \emph{equilibrium} Fermi-Dirac occupation function that depends on the energy dispersion $\varepsilon(\mathbf{k})$, and $e\mathbf{A}(t)=e\int_{-\infty}^t\mathbf{E}(t')dt'$ is the impulse on an electron $e$ due to an external electric field $\mathbf{E}(t)$. In frequency space, we can always replace $\bold A$ with $\bold E/i\Omega$. The current operator for a given Hamiltonian $\hat H$ is $\mathbf{\hat J}=\frac{\partial\hat H}{\partial\mathbf{k}}$, which reduces to the dispersion velocity $\langle\mathbf{k}|\mathbf{\hat J}|\mathbf{k}\rangle=\bold v(\mathbf{k})=\frac{\partial \varepsilon(\mathbf k)}{\partial\mathbf{k}}$ in the translation-invariant intra-band case. We express this optical response in material-dependent units of $nev_F\sim10^{12}$ Am$^{-2}$ where $n$ and $v_F$ are the charge carrier density and the Fermi velocity. In real materials, we have $v_F$ to be $2.22\times10^5$ ms$^{-1}$ and $4.0\times10^5$ ms$^{-1}$ in ZrGeSe ~\cite{doi:10.1063/1.5084090} and CaAgAs~\cite{PhysRevB.95.245113} respectively, $n$ to be $3.37\times10^{26}$ m$^{-3}$, $2.98\times10^{26}$ m$^{-3}$ and $1.7\times10^{26}$ m$^{-3}$ in YbCdGe~\cite{PhysRevB.99.241102}, CaCdSn~\cite{PhysRevB.102.035164} and CaAgAs~\cite{PhysRevB.95.245113} respectively. In these NLSMs, the Fermi surface is a simple nodal loop. The optical response of real materials with more sophisticated nodal structure, beyond the simplest nodal loop, have also been computed in Ref. \cite{PhysRevB.102.035138}.
Eq.~\ref{J} holds in the ballistic limit, where it is the exact solution to the semi-classical Boltzmann equation and the semi-classical equations of motions for electronic wavepackets~\cite{mikhailov2007non}. This requires $\Omega\tau\gg 1$, where $\Omega$ is the optical frequency defined by $\mathbf{E}(t)\sim\mathbf{E}(0)e^{i\Omega t}$, and $\tau$ is the relaxation time due to impurity scattering. For a nodal material with relaxation time $\tau\sim 10^{-12}$ to $10^{-13}\,s$, comparable to that in high quality Graphene samples~\cite{TransportTimeGraphene}, $\Omega\tau\gg 1$ can be achieved in the teraHertz regime of $\Omega \sim 50$-$100\,$THz. The relaxation times can be computed from a microscopic model for the scattering processes~\cite{ho2018theoretical}. In this ballistic regime, scattering processes cannot catch up with the much shorter oscillation timescales~\cite{liu2017high,ndabashimiye2016solid}.
According to Eq.~\ref{J}, the response current arises from contributions at $\bold k$ for which the Fermi-Dirac occupation $F(\varepsilon(\mathbf{k}))$ is non-vanishing. As such, it strongly depends on the shape of the (occupied) Fermi region is the Brillouin zone (BZ), particularly its co-dimensionality. This intra-band response is the expectation of $\bold v(\bold k+\bold e \bold A(t))$ within the occupied region that is dynamically shifted by the external electromagnetic impulse. When the occupied region is not a ``blob'' in the BZ but a thin Fermi ``tube'' of nontrivial co-dimension, the expectation of $\bold v(\bold k+\bold e \bold A(t))$ can fluctuate wildly with $\bold A(t)$ due to ``destructive interference'' of $\bold v$ at different $\bold A(t)$ shifts, as previously pointed out in Refs.~\cite{mikhailov2007non,lee2015negative,PhysRevB.102.035138}. In Ref.~\cite{mikhailov2007non}, this observation was first used to explain the non-linear response of Graphene due to its vanishing density of states at small chemical potential. This was generalized in~\cite{lee2015negative} to nodal loop structures i.e. where two bands intersect along a loop in the BZ, where a stronger non-linearity from a characteristic non-linear response curve was reported. In Ref.~\cite{PhysRevB.102.035138}, this non-linear response was further shown to be strongly enhanced in a 3D material with non-trivial nodal topology i.e. when band intersections form loops that are topologically linked.
To more concretely understand why nodal structures with nontrivial linkages have strong non-linear responses, consider small chemical potentials $\mu$ such that the occupied states take the form of nodal ``tubes'' along nodal lines, which are then shifted by the electromagnetic impulse $\int^t\mathbf{E}(t')dt'$ away from the original nodal structure in the BZ. By differentiating Eq.~\ref{J} with respect to the vector potential, we obtain~\cite{PhysRevB.102.035138} for small $\mu$,
\begin{equation}
\frac{\partial J_i}{\partial A_j}\approx2\sum_{\alpha\in\text{NLs}}\frac{\mu}{\hat{e_j}\cdot\mathbf{v_F}}\frac{\partial^2\varepsilon(\mathbf{k}^\alpha+e\mathbf{A})}{\partial k_i \partial k_j}
\label{dJdA}
\end{equation}
where $\mathbf{k}^\alpha$ labels the trajectories of all the nodal lines $\alpha$ in the BZ and $\hat{e}_j\cdot\mathbf{v}_F$ is the component of the Fermi velocity of the $\alpha$-th NL at momentum $\mathbf{k}^\alpha$ along the applied field. In Eq.~\ref{dJdA}, the differential (optical) response tensor of a nodal line structure is given by a sum over the Hessian of the dispersion at the nodal lines shifted by $\mathbf{A}$ in the BZ. Evidently, we expect the response to be non-linear whenever $\frac{\partial J_i}{\partial A_j}$ is significantly non-constant, which has to be the case around nodal crossings with vanishing gap. In particular, in the diagonal sector where $i=j$, $\varepsilon(\mathbf{k}^\alpha+e\mathbf{A})$ has to pass through a singularity whenever two nodal loops are linked in the direction $\bold A$, since the locus of $\mathbf{k}^\alpha+e\mathbf{A}$ for one of the loops $\alpha$ must intersect another loop $\alpha'$ as $\bold A$ increases. This leads to the enhancement of diagonal response non-linearity in topologically linked nodal loops, which can be quantified by the extent of HHG it causes.
However, the abovementioned argument for non-linearity enhancement pertains only to the diagonal $i=j$ sector, where the response is probed in the direction where the applied field displaces the nodal loops. It remains an open question if a nodal linkage also possess signature response behaviors in the other directions, in both longitudinal and transverse sectors. In this work, we shall focus on elucidating the signature contributions to the optical response due to a generic nodal linkage, particularly on how non-topological contributions from the dispersion profile competes with those from the topological linkage. While real nodal materials may also experience response contributions from the Berry curvature and inter-band scattering, these additional contributions are very material-dependent, and do not form the focus of this work, which aims to explore the most generic response behavior expected from nodal linkages.
\subsection{Canonical Hopf link Model}
In this work, we have chosen to focus on the simplest topologically non-trivial nodal linkage, which is the Hopf link - a nodal configuration consisting of two nodal loops that will be linked when they are sufficiently close to each other. By introducing a canonical two-band Hopf link model where the key parameters $r_y$, $r_x$ and $d$ (Fig.~\ref{Hopf_shape}a) are all independently adjustable, we can isolate the various geometric and topological factors that influence its non-linear response. The parameters are:
\begin{itemize}
\item $r_y$: width of the loops in the $k_y$-direction or longitudinal direction along which the two loops are separated;
\item $r_x$: width of the loops in the $k_x$ or $k_z$ transverse directions;
\item $2d$: the displacement between the two loops, centered at $(0,\pm d,0)$. They are linked if $r_y>d$, since the longitudinal separation between them is given by $2(r_y-d)$ (Fig.~\ref{Hopf_shape}).
\end{itemize}
\begin{figure}
\centering
\includegraphics{Fig1.pdf}
\caption{(a) The Fermi surface of the Hopf link in $\mathbf{k}$-space, with the parameters $r_x$, $r_y$, $r_z$ and $d$ as labeled. We assumed $r_x=r_z$. The quantity $2(r_y-d)$ controls the extent of linkage in the Hopf link. (b) The three distinct possibilities of the Hopf structure are: Hopf link ($r_y<d$), nodal chain ($r_y\approx d$) and unlinked nodal loops ($r_y>d$), illustrated here for a non-vanishing $\mu=0.1$.}
\label{Hopf_shape}
\end{figure}
Our two-band nodal Hopf link Hamiltonian takes the form
\begin{eqnarray}H_{\text{hopf}}(\mathbf{k})=h(k_x,k_y,k_z)\sigma_x+g(k_x,k_y,k_z)\sigma_y
\end{eqnarray}
where the functions $h(\mathbf{k})$ and $g(\mathbf{k})$ are obtained by deforming a well-known Hopf link model derived through the Hopf map~\cite{PhysRevB.96.041202,PhysRevB.97.155140,PhysRevLett.118.147003,lee2020imaging}, for which the Hopf link parameters cannot be easily adjusted independently~\footnote{An analogous approach~\cite{lee2017band} with complex singularities also does not allow for independent tuning of parameters.}. The gap $2\epsilon(\mathbf{k})=2\sqrt{g(\mathbf{k})^2+h(\mathbf{k})^2}$ closes along two loops given by the simultaneous solutions of $g(\bold k)=0$ and $h(\bold k)=0$. Explicitly, with the intermediate parameters $\xi=\frac{\cos r_y-1}{\cos r_x-1}\frac{1}{\sin d}-1$ and $\Gamma=\frac{\cos r_x\cos r_y -1}{\cos r_x -1}\frac{1}{\sin d }$, $h(\bold k)$ and $g(\bold k)$ are given by
\begin{widetext}
\begin{subequations}
\begin{align}
h(k_x,k_y,k_z)&=\sin^2(k_z-k_x)-(\xi\cos(k_x+k_z)+\cot d \cos(k_y)+\cos(k_z-k_x)-\Gamma)^2-\sin^2(k_x+k_z)+\sin^2(k_y),\\
g(k_x,k_y,k_z)&=2\sin(k_z-k_x)(\xi\cos(k_x+k_z)+\cot d \cos(k_y)+\cos(k_z-k_x)-\Gamma)-2\sin(k_x+k_z)\sin(k_y).
\end{align}
\end{subequations}
\end{widetext}
At chemical potential $\mu=0$, the two nodal loops lie in the planes $k_x=0$ and $k_z=0$, with a common mirror symmetry axis along the $k_y$ line. The nodal loops each touch this symmetry axis at two points, giving rise to a total of four touching points $\pm(r_y\mp d)$. From Fig.~\ref{Hopf_shape}, $r_y-d$ evidently describes the extent of their linkage. When $r_y>d$, the two nodal loops are linked together, and we call this resulting arrangement to be a Hopf link.
In realistic materials, the chemical potential $\mu$ can often be tuned away from the nodal energy, so as to obtain a finite density of states available for transport. For generic linearly dispersive nodal systems, $\mu$ scales with the thickness of the nodal tube. A large $\mu$ tends to result in thick tubes that inevitably intersects, resulting in a nodal chain even when $r_y\neq d$.
In total, there are three possibilities as illustrated in Fig.~\ref{Hopf_shape}b:
\begin{itemize}
\item Hopf link: The two nodal loops are topologically linked ($r_y>d$ for $\mu\ll 1$),
\item Unlinked nodal loops: The two nodal loops are disjoint and not topologically linked ($r_y<d$ for $\mu\ll 1$),
\item Nodal chain: The two nodal loops touch each other at a point, and are not topologically linked ($r_y\approx d$ for $\mu\ll 1$).
\end{itemize}
Due to the periodicity of its various terms, the Hamiltonian can sometimes admit additional solutions i.e. ``\emph{periodic images}'' within the first BZ. These solutions, analogous to the degenerate valleys in Graphene, occur as two distinct types, as further elaborated in the Appendix. To prevent the confounding ambiguities resulting from the interference of multiple nodal linkages, we shall exclusively study only cases where such periodic images do not exist.
For the rest of this work, we shall set the temperature to a representative value of $10$ K i.e. $\beta:=\frac{eV}{k_BT}=1160$, which can be adjusted to fit the physical temperature of an actual physical scenario by trivially rescaling our canonical model. For definiteness, we choose a chemical potential of $\mu=0.1$ eV, such that the nodal structure consist of thin closed tubes that exhibit distinct crossovers in topology (linked, intersecting, unlinked) as they are shifted due to photon excitations. However, our results shall continue to hold even when typical photon energies are much smaller than $\mu$. We shall study the behavior of the optical response as the three independent parameters $\{d,r_x,r_y\}$ are varied, insofar as the parameters do not admit periodic images, for all three types of cases (linked, unlinked and nodal chain).
\section{Optical Response in various directions}
We next investigate the various diagonal and off-diagonal (Hall) responses of the nodal Hopf link in the principal directions - both longitudinal (along the direction of loop separation $\hat y$) and transverse directions (perpendicular to the direction of loop separation $\hat x$ and $\hat z$).
Consider a sinusoidal time-varying applied electric field signal
$\mathbf{E}(t)=\mathbf{E_0}\cos\Omega t$, which corresponds to the vector potential $\mathbf{A}(t)=\mathbf{A}\sin\Omega t$, where $\mathbf{A}=\mathbf{E_0}/i\Omega$. In the presence of an oscillatory electric field $\mathbf{E}(t)$, the minimally coupled Fermi-Dirac occupation in Eq.~1 will be $F(\varepsilon(\mathbf{k}-e\mathbf{A}(t)))$ which is equivalent to a translation of the Fermi tube of occupied states in the field direction according to the momentum shift $\mathbf{k}\rightarrow\mathbf{k}-e\mathbf{A}(t).$ This picture motivates the $\mathbf{A}$ dependence of $\mathbf{J}$ rather than $\mathbf{E}$. The ballistic criterion $\Omega\sim100$ THz is independent of the nodal energy since the Fermi surface is translationally invariant. Although there are nine possible response components $J_i(A\hat{j})$, $i,j\in{x,y,z}$, some of them necessarily generically vanish or are not independent due to the symmetries of our Hamiltonian (Eq.~3):
\begin{figure}
\includegraphics[width=.6\linewidth]{Fig2.pdf}
\caption{Contribution to the semi-classical response (Eq.~1): In the presence of an illustrative applied field along the $k_y$ direction, the Fermi surface is displaced along the $k_y$-axis in $\mathbf{k}$-space in a manner symmetric about the $k_x=0$ and $k_z=0$ planes. This accounts for the vanishing response $J_x$ and $J_z$.}
\label{vanishing}
\end{figure}
\begin{equation}
J_x(A\hat{y})=0,\quad J_z(A\hat{y})=0,\tag{5a}
\end{equation}
\begin{equation}
J_x(A~ \hat{x})=J_z(A\hat{z}),\tag{5b}
\end{equation}
\begin{equation}
J_z(A\hat{x})=J_x(A\hat{z}),\tag{5c}
\end{equation}
\begin{equation}
J_y(A\hat{x})=-J_y(A\hat{z}).\tag{5d}
\end{equation}
\begin{figure*}
\includegraphics{Fig3.pdf}
\caption{(a) In the presence of an applied transverse field $A\hat{x}$, the Fermi surface of the Hopf link is displaced along the transverse $k_x$-axis in $\mathbf{k}$-space, in turn giving rise to a transverse diagonal optical response $J_x(A\hat{x})$. (b) For the Hopf link, the transverse response weakly depends on $d$, for $r_x=0.8$ and $r_y=1.2>d$. (c) Enhancement of transverse response non-linearity for the Hopf link as $r_x$ increases, for $d=0.5<r_y=1.0$. (d) Enhancement of transverse response non-linearity for the Hopf link with $r_y>d=0.5$ and $r_x=1.0$. (e) Displacement of the Fermi regions of the unlinked nodal loops along the transverse $k_x$ axis in $\mathbf{k}$-space, with the singularities from each loop becoming weaker as the loops are further separated. (f) The transverse response shows a stronger dependence on the $d$ when the nodal loops are no longer topologically linked. As the longitudinal loops separation increases, the response currents become smaller. Here, we contrast the responses of the Hopf link (dashed-dot lines, $d=0.25,0.35$), unlinked nodal loops (dashed lines, $d=1.1,1.3$) and the nodal chain (solid lines $d=0.5,0.9$) for $r_y=0.6$, $r_x=0.4$.
For (b-d), the chosen parameter sets are among those that exhibit the greatest non-linearity in the transverse response. The responses are in units of $env_F$ where $n$ and $v_F$ are the number density and Fermi velocity of the NLSM material.}
\label{Jx}
\end{figure*}
\begin{figure*}
\includegraphics{Fig4.pdf}
\caption{(a) In the presence of an applied longitudinal field $A\hat{y}$, the Fermi surface of the Hopf link is displaced along the longitudinal $k_y$-axis in $\mathbf{k}$-space, in turn giving rise to a longitudinal diagonal optical response $J_y(A\hat{y})$. (b) Enhancement of longitudinal response non-linearity by increasing $d$ in the linked regime with $r_y=1.0>d$, $r_x=0.6$. (c) The shape of the longitudinal response for the Hopf link is independent of the transverse radius $r_x$ but increases in magnitude with $r_x$, as plotted for $r_y=1.2>d=0.8$. (d) Enhancement of longitudinal response non-linearity for the Hopf link by increasing $r_y>d=0.6$ at $r_x=1.0$. (e) Displacement of the Fermi regions of the unlinked nodal loops along the longitudinal $k_y$ axis in $\mathbf{k}$-space, with the singularities from each loop becoming weaker as the loops are further separated. (f) Comparing the longitudinal responses for the Hopf link (dashed-dot lines $d=0.25, 0.35$), unlinked nodal loops (dashed lines $d=1.2, d=1.4$), and the nodal chain (solid lines $d=0.4, d=0.8$) for $r_y=0.6$, $r_x=0.4$. The longitudinal response decreases in magnitude but becomes more non-linear as $d$ increases. This increasing non-linearity is an artifact of the weaker influence of the singularities as the loops are further separated. For (b-d), the chosen parameter sets are among those that exhibit the greatest non-linearity in the transverse response. The responses are in units of $env_F$ where $n$ and $v_F$ are the number density and Fermi velocity of the NLSM material.
}
\label{Jy}
\end{figure*}
Eq.~5a holds since the nodal loops have mirror symmetry about $k_x=0$ and $k_z=0$ planes respectively. As the dispersion velocity $\langle\mathbf{k}|\hat{\mathbf{J}}|\mathbf{k}\rangle$ (Eq.~1) inherits the same symmetry as the Fermi surface - mirror symmetry about the $k_y$ axis, When the Fermi surface is displaced along the longitudinal $k_y$ axis (Fig.~\ref{vanishing}), the resulting contribution by the Fermi tubes on the left of the $k_y$ axis will exactly cancel that due to the Fermi tubes on the right. Eq.~5b and Eq.~5c follow from the symmetrical roles played by transverse momenta $k_z$ and $k_z$. Eq.~5d is true because of the loops are related to each other by interchanging $k_x\leftrightarrow-k_z$. In all, there are only 4 unique currents responses in the following response matrix
\begin{equation}
J_i(A\hat{j})=\begin{pmatrix}J_x(A\hat{x})&J_y(A\hat{x})&J_z(A\hat{x})\\0&J_y(A\hat{y})&0\\J_z(A\hat{x})&-J_y(A\hat{x})&J_x(A\hat{x}).\\
\end{pmatrix}
\end{equation}
Since $A$ scales with $E_0$ for our sinusoidal electric field, Eq. 5 is similar to a conductivity tensor. As we shall show, the diagonal responses are much more affected by the topological linkage of the nodal loops, compared to the non-diagonal response, i.e. Hall responses.
\subsection{Transverse Diagonal Responses $J_x(A\hat{x})$, $J_z(A\hat{z})$}
According to Eq. 1, we generically expect a non-linear semiclassical response since its integral is a highly non-linear function. This is especially the case for a Hopf link, where each nodal loop juts out of the plane of the other nodal loop. As such, it acts as a source of dispersion velocity $\langle\mathbf{k}|\mathbf{\hat{J}}|\mathbf{k}\rangle$, which `destructively interferes' with those due to the nodal loop that encloses this nodal tube~\cite{PhysRevB.102.035138}. When the Fermi surface of the Hopf link starts to translate by the impulse from an applied field (Fig.~\ref{Jx}a), the response, computed from Eq.~1, initially increases sharply followed by a much slower non-monotonic change until the applied field is sufficiently large that the Fermi surface begins to leave the region of influence exerted by the singularity.
Beyond that, the contribution of $\langle\mathbf{k}|\mathbf{\hat{J}}|\mathbf{k}\rangle$ continues to add up, giving a subsequent monotonic response. The extent of non-linearity afforded by each singularity is described by Eq.~2, where the gradient is a sum of the second derivatives of the dispersion. The transverse (perpendicular to the direction of the loop separation) diagonal optical response of the Hopf link was found to observe the following general trends (we plot the most non-linear responses in Figs.~\ref{Jx}b-d):
\begin{enumerate}
\item The response demonstrates a weak dependence on the longitudinal separation $d$ when the nodal loops are topologically linked,
but not true otherwise, as illustrated in Figs.~\ref{Jx}b and f respectively.
\item The extent of response non-linearity is enhanced with larger transverse radius $r_x$ or smaller longitudinal radius $r_y$ i.e. larger loop aspect ratio $r_x/r_y$, as illustrated in Figs.~\ref{Jx}c,d. Correspondingly, the response currents generally decrease with increasing non-linearity.
\end{enumerate}
We can physically understand these trends. Since the relative separation $d$ of the centers of the loops merely changes the relative longitudinal $k_y$ position of the loops, intuitively it should not play a significant role in influencing the transverse response resulting from the transverse translation of the Fermi surface. Indeed, the transverse diagonal response is almost independent of $d$ (Fig.~\ref{Jx}b) in the Hopf link. Yet, this independence from $d$ no longer holds when the loops are no longer topologically linked (Figs.~\ref{Jx}e,f) because further separated loops i.e. sources of singularities correspond to more uniform velocity fields, leading to smaller responses currents. This is unlike the linked cases, where the nodal linkage guarantees the proximity to nodal singularities and topologically ``protects'' the transverse response (previous literature i.e. Ref.~\cite{PhysRevB.102.035138} only reported the topological enhancement of the longitudinal response).
On the other hand, increasing the transverse radius $r_x$ or decreasing the longitudinal radius $r_y$ stretches the aspect ratio of each loop and hence the boundary for ``interference'' (defined by the locus of points where the competition of opposing dispersion velocity vectors terminate) along the transverse $k_x$ direction, giving more room for ``destructive interference'' as the Fermi surface is displaced along the transverse $k_x$ direction. This in turn gives a greater range of $A\hat{x}$ where the response undergoes a non-monotonic change.
\subsection{Longitudinal Diagonal Response $J_y(A\hat y)$}
Unlike in the transverse diagonal response, the longitudinal separations of the loops $2(r_y-d)$ now greatly influences the longitudinal (along the direction of the loop separation) responses of the Hopf link. Like before, these nodal tubes are sources of `destructive interference' for the dispersion velocity $\langle\mathbf{k}|\mathbf{\hat{J}}|\mathbf{k}\rangle$. The semiclassical response is again computed from Eq. 1 as the Fermi surface of the Hopf link displaces along the longitudinal $k_y$ direction (Fig.~\ref{Jy}a). The general trends for the longitudinal diagonal optical response of the Hopf link are:
\begin{enumerate}
\item The extent of response non-linearity is enhanced with larger longitudinal loop separation $d$, as illustrated in Fig.~\ref{Jy}b.
\item Increasing the longitudinal radius $r_y$ also enhances the extent of response non-linearity, as illustrated in Fig.~\ref{Jy}d.
\item However, the transverse radius $r_x$ does not change the shape of the response curves, but the magnitude of the response does increase with $r_x$, as illustrated in Fig.~\ref{Jy}c.
\end{enumerate}
To physically explain these trends, note that as $d$ increases with $r_y$ fixed, the loop separation $2(r_y-d)$ decreases, enhancing the extent of non-linearity due to the closer proximity of one loop with the band singularity of the other. While the non-linearity is still present, as shown in Fig.~\ref{Jy}f, when the nodal loops are no longer topologically linked, this is an artifact from the weaker dispersion field when the loops are separated farther apart. Increasing $r_y$ at fixed $d$ also enhances non-linear response by providing a greater range of values of $A~\hat{y}$ for significant destructive interference to occur.
On the other hand, $r_x$ only stretches the nodal loop in the transverse $k_x$ direction, which plays no role in the `destructive interference' in the longitudinal direction. But yet, $r_x$ increases the circumferential length of the nodal loop, which gives a greater contribution to $\mathbf{J}$ in Eq.~1 and hence a greater response magnitude.
\section{Global aspects of optical response anisotropy}
Having discussed the non-linearity of the response tensor along the principal directions, we now present how the anisotropic response behaves as a whole. Due to non-linearity i.e. $\bold J(\bold A_1+\bold A_2)\neq \bold J(\bold A_1)+\bold J(\bold A_2)$, the responses in directions away from the previously studied principal directions $\bold{\hat{A}}=(1,0,0)$, $(0,1,0)$ and $(0,0,1)$ may behave unexpectedly. In the following, we shall represent the response across all directions in the form of constant $|\bold A|=A_0$ level surfaces in $\bold J$-space (recall that $\bold A(t)=\int_{-\infty}^{t}\mathbf{E}(t')dt'$). We parametrize this surface $\mathcal{S}_A$ with spherical angular coordinates as follows:
\begin{equation}
\mathcal{S}_A=\{A_0\boldsymbol{\hat{A}}=A_0(\cos\theta\cos\phi,\cos\theta\sin\phi,\sin\theta)\}\tag{8}
\end{equation}
where $0\leq\theta\leq\pi,~0\leq\phi\leq 2\pi$. The anisotropy in the optical response of our canonical double nodal loop model (Eqs.~3, 4a,b) due to $\boldsymbol{\alpha}=A_0\boldsymbol{\hat{A}}$ is described by the anisotropy of the smooth $\mathcal{S}_J$ surface embedded in the $J$-space. We represent the azimuthal angle $\phi$ with a colormap defined by $\mathcal{S}_A$ in Fig.~\ref{A0}a, which will be significantly distorted by anisotropy when bijectively mapped onto $\mathcal{S}_J$ (Figs.~\ref{A1} and \ref{A2}).
\begin{figure}
\centering
\includegraphics{Fig5.pdf}
\caption{
(a) To draw the response surfaces, we represent an external field of constant magnitude $A_0$ but arbitrary direction as a sphere $\mathcal{S}_A$ in $A$-space. The sphere is colored by the azimuthal angle $\phi$, and is significantly distorted by anisotropy when bijectively mapped onto the response surfaces $\mathcal{S}_J$ in 3D $J$-space. (b-c) Illustrative examples of the distortion for our model (Eq.~3) with $r_x=0.6,r_y=1.0,\mu=0.1,A_0=1.25$, for $d=0.5$ and $0.7$ respectively ($\mathbf{J}$ in units of $10^{-3}env_F$). (d) Origin of the thicker ``waist'' of (b) vs. (c) in terms of $J_y(A_0 \hat y)$ response. At $A_0=1.25$, the $d=0.5$ case (b) has a response current almost thrice as large.
(e) This difference in $J_y(A_0 \hat y)$ corresponds to a large difference in the thickness of the concave ``waist'' of the response surface, as more clearly illustrated by comparing cross sections of (b,c) at $\theta=\frac{\pi}{2}$. }
\label{A0}
\end{figure}
Insight into the shapes of of the constant $A_0$ response surfaces, as well as their significance with regards to nodal topology and geometry, can be obtained by computing the response in a few high-symmetry directions of $\boldsymbol{\hat{A}}$, such as $\boldsymbol{\hat{A}}=\pm(0,1,0),\pm(1,0,1),\pm(1,0,-1)$ (for notational simplicity, we shall henceforth drop the normalization factor). From Eq. 5a, we know that for a longitudinal external field $\boldsymbol{\hat{A}}=\pm(0,1,0)$, there is vanishing transverse response, i.e. $J_x=J_z=0$ for such $\boldsymbol{\hat A}$. Hence the corresponding point on the $\mathcal{S}_J$ surface must be along the $J_y$ axis, with $|J_y|$ being the width of the surface along the $J_y$-axis. This is also the longitudinal diagonal current discussed in Sect. III. Also, the response surface point corresponding to $\boldsymbol{\hat{A}}=\pm(1,0,1)$ must lie in the $J_x-J_z$ plane, since $J_y=0$ from Eq. 5d. And by Eq. 5b, there is an equal contribution of $J_x$ and $J_z$ from $\boldsymbol{\hat{A}}=\pm(1,0,1)$, so all corresponding points from this orientation of $\boldsymbol{\hat A}$ must lie on the 45 degree diagonals along the $J_x$-$J_z$ axes, as evident in Figs.~\ref{A1} and \ref{A2}. Incidentally, points corresponding to $\boldsymbol{\hat A}=\pm(1,0,-1)$ also lie in the $J_x$-$J_z$ plane despite not being directly protected by Eqs.~5a and 5d, a fine illustration of the non-linearity of $\bold J(\bold A)$. Other high symmetry directions which we shall use in the following discussion include $\boldsymbol{\hat{A}}=~\pm(1,1,0),~\pm(0,1,1),~\pm(1,-1,0),~\pm(0,1,-1)$.
Before starting the detailed analysis of the response surface, we consider a quick example. In Sect. III, we highlighted that a greater longitudinal separation $d$ between two topologically linked nodal loops results in a more non-linear response (Fig.~\ref{Jy}b). This is indeed seen in our corresponding $\mathcal{S}_J$ surfaces. At an illustrative $A_0=1.25$, this non-linear response is approximately $3$ times larger i.e. less non-linear for $d=0.5$ vs $d=0.7$ (Fig.~\ref{A0}d), which corresponds to a proportionally smaller width of the ``waist'' of the response surface in the $\hat y$ direction (Figs.~\ref{A0}b,c). This is more evident when viewed in the cross-sectional plane $\hat{A}=(1,0,1)$ with only $\theta=\frac{\pi}{2}$ points plotted (Fig.~\ref{A0}e). That the general shapes of the $d=0.5$ and $d=0.7$ surfaces appear rather similar (Figs.~\ref{A0}b,c) can be understood from the weak dependence on $d$ for the transverse diagonal responses, as discussed in Sect. III. Clearly, the results from Sect. III alone are inadequate in working out the entire response surface and thus the response surfaces provide a bigger picture and as we will see, shed light on nodal topology and geometry.
\begin{figure*}
\includegraphics{Fig6.pdf}
\caption{
For the three possible nodal configurations (linked, nodal chain, unlinked) of our double loop Hamiltonian, shown here are their corresponding surfaces $\mathcal{S}_J$ for constant chemical potential $\mu=0.1$ under small, medium and large impulses $A_0=0.1,0.75,1.75$ for the linked and chain case, while $A_0=0.1,0.5,1.0$ for the unlinked case. The qualitative evolution of the response surfaces with increasing $A_0$ differ for the different topological nodal configurations, with the large-$A_0$ response surface elongated towards the $J_z\pm J_x$ direction in the topologically linked/unlinked cases.
(a-d) The response surfaces for the Hopf link $r_y=0.6>d=0.15$, $r_x=0.6$ exhibit concavity along directions $\pm(0,1,1)$, $(1,0,\pm1)$ due to the locally depressed responses in these directions. (e-h) The response surfaces for the nodal chain $r_y=0.6=d$, $r_x=0.6$ do not obey the same symmetries as those in the Hopf link since the Fermi surface of the nodal chain breaks reflection symmetry. In particular, for intermediate $A_0$ values, the surface looks like a saddle with asymmetric responses, e.g. $\hat{A}=(0,1,1)$ has a greater response than $\hat{A}=(0,-1,-1)$ at $A_0=0.75$. (i-l) The response surfaces for unlinked nodal rings $r_y=0.6>d=0.15$, $r_x=0.6$ do respect these symmetries too but possess topologies distinct to the topologically linked case. The units of the response currents $\mathbf{J}$ are $10^{-3}env_F$, and parameters are chosen such that no periodic images are present. }
\label{A1}
\end{figure*}
\begin{figure*}
\includegraphics{Fig7.pdf}
\caption{The dependence of the surface morphologies on the aspect ratio parameters $r_x,r_y$ of the nodal Hopf link at constant $\mu=0.1$. Again, their response surfaces $\mathcal{S}_J$ are plotted for small, medium and large impulses $A_0=0.1,0.75$ and $1.75$, with current $\mathbf{J}$ are in units of $10^{-3}env_F$. (a-d) The response surfaces for the Hopf link $r_y=0.4>d=0.15$, $r_x=0.6$. (e-h) The same response surfaces as in Fig.\ref{A1}a-d. (i-l) The response surfaces for the Hopf link $r_y=0.6>d=0.15$, $r_x=0.4$ share several similarities with (e-h), but with a more depressed longitudinal response at $A_0=0.75$ and $A_0=1.75$ and a less depressed response along $\hat{A}=\pm(1,0,1)$ at $A_0=1.75$. Overall, the response surfaces differ for different geometric parameters, but do not change as drastically as when the nodal topology changes (Fig.~6).}
\label{A2}
\end{figure*}
\subsection{Response surfaces and nodal topology}
It is already known that~\cite{PhysRevB.102.035138} nodal linkages can enhance the non-linearity of the response, at least in the longitudinal diagonal direction $J_y(A_0\hat{y})$ (Fig.~4). Further, significant non-linearity is also present in the transverse diagonal directions $J_x(A_0\hat{x})$ and $J_z(A_0 \hat{z})$, although not necessarily enhanced by the topological linkage. Hence we shall expect such non-linearity to be manifested in the constant $A_0$ response current surface too.
Indeed, as shown in Fig.~6, typical linked, touching and unlinked nodal loops exhibit significantly different evolutions of the constant $A_0$ response surfaces as $A_0$ is increased. At small $A_0$, all three cases have ellipsoidal-shaped responses surfaces, testimony to anisotropy of the nodal system, even in the linear (small $A_0$) limit. At very large $A_0$ i.e. $A_0=1.75$ where the Fermi regions have been displaced far from their original positions, the response surfaces are all very anisotropic and large, since minimal cancellation of the velocity field $d\varepsilon/d\bold k$ occurs. Their exact shapes depend on the details of the energy dispersion away from the loops, and are non-universal though decidedly anisotropic. What is most interesting is the intermediate $A_0\approx 0.75$ regime, which for the linked case is around when the Fermi region of one loop crosses the singularity from the other loop. For the linked case, the significant non-linearity of the response around intermediate values of $A$ (see Figs.~3 and 4) suppresses the response current, particularly in the longitudinal directions with ``untwisted colors'' in Fig.~7. As such, this leads to a somewhat ``squeezed'' appearance of the response surface compared to that of the nodal chain or unlinked cases, where the response surface looks comparatively ``puffed up'' during the $A_0$ evolution. Compared to the other cases, the unlinked case shows the least variation in response surface shape during the evolution due to the least amount of cancellation of $d\varepsilon/d\bold k$ during the evolution.
A more detailed characterization of the shape of the response surfaces can be performed by analyzing the high symmetry directions. For instance, for the Hopf linked case, the constant $A_0$ surfaces look similar to ellipsoids (Fig.~\ref{A1}b,~\ref{A2}b,f,j) with the longest axis oriented along the 45 degree diagonal between $J_x>0$ and $J_z>0$ axes, corresponding to the direction $\bold J\propto \pm(1,0,1)$. The three `principal axes' of the response surfaces are thus marked out by the three pair of points $\pm(0,1,0)$, $\pm(1,0,1)$ and $\pm(1,0-1)$, as discussed. This is seen in Figs.~\ref{Appendix: ry06rx06response}m-p where the magnitude for the responses along $\pm(1,0,-1)$ are significantly smaller than along $\pm(1,0,1)$ and $\pm(0,1,0)$. This accounts for its oblate appearance.
As we increase $A_0$ for the linked case, the two faces of the constant $A_0$ surface in Fig.~\ref{A1}c which are characterized by $\pm(1,0,-1)$, exhibit concavity, reminiscent of the shape of a red blood cell. The concavity along the directions $\pm(1,0,-1)$ can be better seen in Fig.~\ref{Appendix: diff}b, where the same response surface is viewed from a lateral direction. This characteristic suppressed response is similar to that in Fig.~\ref{A0}d where it was a consequence from the non-monotonicity of the response. In these cases, the responses along $\pm(1,0,-1)$ increase significantly slower than in other directions (Figs.~\ref{Appendix: ry06rx06response}m-p). As $A_0$ continues to increase, the $\hat A$ directions characterized by $\pm(1,0,1)$ and $\pm(0,1,0)$ also exhibit concavity. Again, this is because the responses along $\pm(1,1,0)$, $\pm(0,1-1)$, $\pm(0,1,1)$, $\pm(1,-1,0)$ (which correspond to the eight corners of the surface in Fig.~\ref{A1}d) grow much quicker than along the principal directions, consistent with the individual response curves in Figs.~\ref{Appendix: ry06rx06response}a-l. This surface (Fig.~\ref{A1}d) also show symmetry consistent with the symmetries in the responses (Figs.~\ref{Appendix: ry06rx06response}a-l). For instance, the constant $A_0$ surface has mirror symmetry about the $J_y=0$ plane. Due to the non-negligible thickness of the nodal loops, certain symmetries of the $\mu=0$ nodal system may be broken. This symmetry breaking is particularly pronounced in the nodal chain case, where the two nodal rings intersect with relatively weak dispersion. Generically, the finitely thick nodal tube can break reflection symmetry in the direction of impulse, i.e. translating the Fermi surface in the $+A_0\hat{y}$ direction results in a different response compared to doing so in the $-A_0\hat{y}$ direction. For small $A_0$, the surfaces are imperfectly rounded blobs (Fig.~\ref{A1}f) and evolve to a saddle shape (Fig.~\ref{A1}g) as $A_0$ grows. This saddle-like shape is more obvious when viewed in a different orientation illustrated in Fig.~\ref{Appendix: diff}c. It can be characterized by 6 out of the 14 points in total: $(0,1,1)$, $(1,1,0)$, $(0,\pm1,0)$, $(0,1,-1)$ and $(-1,1,0)$. Again, we observe asymmetry in the sense that the response grow more than proportionately in the directions $(0,1,1)$, $(1,1,0)$, $(0,1,-1)$ and $(-1,1,0)$ (the negative pair does not grow as fast). As $A_0$ grow, so do the responses corresponding to the remaining 8 points, giving to a surface like in Fig.~\ref{A1}h, with 6 concave sides.
For the unlinked nodal loops, the surface for small $A_0$ is again imperfectly rounded due to linear anisotropy. For intermediate $A_0$, the surface expands into a lemon-like shape due to the relatively small response non-linearity, and can be characterized by the surface directions $\pm(1,1,0),\pm(0,1,1)$ and $\pm(1,-1,0)$. As $A_0$ grow, the responses in the $\pm(1,0,1)$ directions do not grow as fast, and thus the face characterized by this pair of points exhibit concavity. For large $A_0$, the response surface elongates towards the $J_z-J_x$ direction which differs from the linked case where it elongates towards the $J_z+J_x$ direction. This conclusively relates the response surfaces with the topological linkage of the nodal loops.
\subsection{Response surfaces and nodal geometry}
Since the non-linear response current does not correspond to any topologically quantized value, we expect it to be affected by deformations of the nodal structure too. This should apply to both linked and unlinked cases, even if the linked case is more likely to possess strongly non-linear response. Earlier, we have considered circular nodal loops with $r_x=r_y$. How will the surfaces in Fig.~\ref{A1} change as we vary $r_y$ and $r_x$? Varying $r_y$ and $r_x$ add one more layer of complexity in our analysis of these surfaces. To study that, we shall start from the parameters in Fig.~\ref{A1}(a-c) and change $r_y$ and $r_x$ separately, as presented in Fig.~7.
Generally, the surfaces along the same column of Fig.~\ref{A2} will have relatively similar morphologies since they are only slightly distorted from each other and correspond to the same $A_0$ impulse. In particular, the special symmetries in Figs.~\ref{A1}b-d are no longer satisfied as the relative responses along the various directions now behave quite differently. This occurs because a chemical potential of $\mu=0.1$ result in nodal loops with significantly non uniform and asymmetric thickness (Fig.~\ref{A2}i). This further enhances the anisotropy of the response surfaces, particularly at $A_0=0.75$. But at high fields $A_0=1.75$, the surface morphologies of Figs.~\ref{A2}d,h share similar morphological features - suppressed response at $\pm(1,0,1)$ and $\pm(1,0,-1)$. Again, this anisotropic response is due to the varying rates of growth for the responses along the individual directions. For instance, $\pm(1,0,1)$ does not grow as fast as say $\pm(1,1,0)$ in Fig.~\ref{A2}h, but grow at comparable rates in Fig.~\ref{A2}l.
\section{Conclusion}
In this work, we have systematically studied in detail the anisotropic and non-linear optical response of two nodal loops that are linked, unlinked or touching (nodal chain). This system, as parametrized by our canonical two nodal loop model, represents the simplest abstraction of simultaneously occurring nodal loops (linked or unlinked) in nodal materials. First, we studied the effects of nodal geometry and topology individually along various axis directions. Next, we presented constant $A_0$ response surfaces to highlight the anisotropy of the response, and how that global picture can shed light on the overall configuration of the nodal structure. Our findings generalizes existing results on the enhancement of optical response non-linearity by nodal linkages~\cite{PhysRevB.102.035138} to various transverse, Hall and diagonal sectors, and introduces a geometric picture of response non-linearity and anisotropy that will be invaluable in analyzing generic nodal material responses, as well as the engineering of higher-harmonic generation materials for applications like Terahertz radiation generation.
\section{Acknowledgements}
TT is supported by the NSS scholarship by the Agency for Science, Technology and Research (A*STAR), Singapore.
This work is supported by the Singapore Ministry of Education (MOE) Tier 1 start-up grant (R-144-000-435-133).
|
\section{Introduction}
In the late 19th century, the theory of classical mechanics experienced several issues in reporting the physical phenomena of light masses and high velocity microscopic particles. In 1920s, Bohr’s atomic theory \cite{bohr1928quantum}, Heisenberg’s discovery of quantum mechanics \cite{robertson1929uncertainty} and Schrödinger’s \cite{wessels1979schrodinger} discovery of wave mechanics influence the conception of a new field i.e. the quantum mechanics. In 1982, Feynman \cite{feynman1982simulating} stated that quantum mechanical systems can be simulated by quantum computers in exponential time, i.e. better than with classical
computers. Till then, the concept of quantum computing was thought to be only a theoretical possibility, but over the last three decades the research has evolved such as to make quantum computing applications a realistic possibility \cite{wang2012handbook}.
In the last two decades, the field of swarm intelligence has got overwhelming response
among research communities. It is inspired by nature and aims to build decentralized and self-organized systems by collective behavior of individual agents with each other and with their environment. The research foundation of swarm intelligence is constructed mostly upon two families of optimization algorithms i.e. ant colony optimization (Dorigo et at. \cite{dorigo1999ant} and Colorni et al. \cite{colorni1992distributed} 1992); and particle swarm optimization (PSO) (Kennedy \& Eberhart \cite{kennedy1995particle} 1995). Originally, the swarm intelligence is inspired by certain natural behaviors of flocks of birds and swarms of ants.
In the mid 1990s, particle swarm optimization technique was introduced for continuous optimization, motivated by flocking of birds. The evolution of PSO based bio-inspired techniques has been in an expedite development in the last two decades. It has got
attention from different fields such as inventory planning \cite{wang2014modified}, power systems \cite{alrashidi2009survey}, manufacturing \cite{yildiz2009novel}, communication networks \cite{latiff2007energy}, support vector machines \cite{lin2008particle}, to estimate binary inspiral signal \cite{wang2010particle}, gravitational waves \cite{normandin2018particle} and many more. Similar to evolutionary genetic algorithm, it is inspired by simulation of social behavior, where each individual is called particle, and group of individuals is called swarm. In multi-dimensional search space, the position and velocity of each particle represent a probable solution. Particles fly around in a search space seeking potential solution. At each iteration, each particle adjusts its position according to the goal of its own and its neighbors. Each particle in a neighborhood share the information with others \cite{sun2004particle}. Later, each particle keeps the record of best solution experienced so far to update their positions and adjust their velocities accordingly.
\begin{figure*}[!ht]
\centering
\includegraphics[scale=1.25]{psoqpso.eps}
\caption{Particles movement in PSO and QPSO algorithm}
\end{figure*}
Since the first PSO algorithm proposed, the several PSO algorithms have been introduced with plethora of alterations. Recently, the combination of quantum computing, mathematics and computer science have inspired the creation of optimization techniques. Initially, Narayanan and Moore \cite{narayanan1996quantum} introduced quantum-inspired genetic algorithm (QGA) in 1995. Later, Sun et al. \cite{sun2004particle} applied the quantum laws of mechanics to PSO and proposed quantum-inspired particle swarm optimization (QPSO). It is the commencement of quantum-behaved optimization algorithms, which has subsequently made a significant impact on the academic and research communities alike.
Recently, Yuanyuan and Xiyu \cite{yuanyuan2018quantum} proposed a quantum evolutionary algorithm to discover communities in complex social networks. Its applicability is tested on five real social networks and results are compared with classical algorithms. It has been proved that PSO lacks convergence on local optima i.e. it is tough for PSO to come out of the local optimum once it confines into optimal local region. QPSO with mutation operator (QPSO-MO) is proposed to enhance the diversity to escape from local optimum in search \cite{liu2005quantum}. Protopopescu and Barhen \cite{protopopescu2002solving} solved set of global optimization problems efficiently using quantum algorithms. In future, the proposed algorithm can be integrated with matrix product state based quantum classifier for supervised learning \cite{40, 44}.
In this paper, we have combined QPSO with Cauchy mutation operator to add long jump ability for global search and natural selection mechanism for elimination of particles. The results shown that it has great tendency to overcome the problem of trapping into local search space. Therefore, the proposed hybrid QPSO strengthened the local and global search ability and outperformed the other variants of QPSO and PSO due to fast convergence feature.
The illustration of particles movement in PSO and QPSO algorithm is shown in Fig 1. The big circle at center denotes the particle with the global position and other circles are particles. The particles located away from global position are lagged particles. The blue color arrows signify the directions of other particles and the big red arrows point towards the side in which it goes with high probability. During iterations, if the lagged particle is unable to find better position as compared to present global position in PSO, then their impact is null on the other particles. But, in QPSO, the lagged particles move with higher probability in the direction of gbest position. Thus, the contribution of lagged particles is more to the solution in QPSO in comparison with PSO algorithm.
The organization of rest of this paper is as follows: Sect. 2 is devoted to prior work. In Sect. 3, the quantum particle swarm optimization is described. In
Sect. 4, the proposed hybrid QPSO algorithm with Cauchy distribution and natural selection mechanism is presented. The experimental results are
plotted for a set of benchmark problems and compared with several QPSO variants in Sect. 5. The correctness and time complexity are analyzed in Section 6. QPSO-CD is applied to three constrained engineering design problems in Sect. 7.
Finally, Sect. 8 is the conclusion.
\section{Prior Work}
Since the quantum-behaved particle swarm optimization was proposed, various revised variants have been emerged. Initially, Sun et al. \cite{sun2004particle} applied the concept of quantum computing to PSO and developed a quantum Delta potential well model for classical PSO \cite{sun2004global}. It has been shown that the convergence and performance of QPSO are superior as compared to classical PSO. The selection and control of parameters can improve its performance, which is posed as an open problem. Sun et al. \cite{sun2007using} tested the performance of QPSO on constrained and unconstrained problems. It has been claimed that QPSO is a promising optimization algorithm, which performs better than classical PSO algorithms. In 2011, Sun et al. \cite{sun2011quantum} proposed QPSO with Gaussian distribution (GAQPSO) with the local attenuator point and compared its results with several PSO and QPSO counterparts. It has been proved that GAQPSO is efficient and stable with superior features in quality and robustness of solutions.
Further, Coelho \cite{dos2010gaussian} applied GQPSO to constrained engineering problems and shown that the simulation results of GQPSO are much closer to the perfect solution with small standard deviation. Li et al. \cite{li2012improved} presented a cooperative QPSO using Monte Carlo method (CQPSO), where particles cooperate with each other to enhance the performance of original algorithm. It is implemented on several representative functions and performed better than the other QPSO algorithms in context of computational cost and quality of solutions. Peng et al. introduced \cite{peng2013quantum} QPSO with Levy probability distribution and claimed that there are very less chances to be stuck in local optimum.
Researchers have applied PSO and QPSO to real-life problems and achieved optimal solutions as compared to existing algorithms. Ali et al. \cite{90} performed energy-efficient clustering in mobile ad-hoc networks (MANET) with PSO. The similar approach can be followed to analyse and execute mobility over MANET with QPSO-CD \cite{88}. Zhisheng \cite{zhisheng2010quantum} used QPSO in economic load dispatch for power system and proved superior to other existing PSO optimization algorithms. Sun et al. \cite{sun2006qpso} applied QPSO for QoS multicast routing. Firstly, the QoS multicast routing is converted into constrained integer problems and then effectively solved by QPSO with loop deletion task. Further, the performance is investigated on random network topologies. It has been proved that QPSO is more powerful than PSO and genetic algorithm.
Geis and Middendorf proposed PSO with Helix structure for finding ribonucleic acid (RNA) secondary structures with same structure and low energy \cite{geis2011particle}. The QPSO-CD algorithm can be used with two-way quantum finite automata to model the RNA secondary structures \cite{bhatia2018modeling}.
Bagheri et al. \cite{bagheri2014financial} applied the QPSO for tuning the parameters of adaptive network-based fuzzy inference system (ANFIS) for forecasting the financial prices of future market. Davoodi et al. \cite{davoodi2014hybrid} introduced a hybrid improved QPSO with Neldar Mead simplex method (IQPSO-NM), where NM method is used for tuning purpose of solutions. Further, the proposed
algorithm is applied to solve load flow problems of power system and acquired the convergence accurately with efficient search ability. Omkar \cite{omkar2009quantum} proposed QPSO for multi-objective design problems and results are compared with PSO. Recently, Fatemeh et al. \cite{fatemeh2019shuffled} proposed QPSO with shuffled complex evolution (SP-QPSO) and its performance is demonstrated using five engineering design problems. Prithi and Sumathi \cite{prithi2020ld2fa} integrated the concept of classical PSO with deterministic finite automata for transmission of data and intrusion detection. The proposed algorithm QPSO-CD can be used with quantum computational models for wireless communication \cite{10, 20, 30}.
\section{Quantum Particle Swarm Optimization}
Before we explain our hybrid QPSO-CD algorithm mutated with Cauchy operator and natural selection method, it is useful to define the notion of quantum PSO. We assume that the reader is familiar with the concept of classical PSO; otherwise, reader can refer to particle swarm optimization algorithm \cite{kennedy2010particle, shi2001particle}. The specific principle of quantum PSO is given as:
In QPSO, the state of a particle can be represented using wave function $\psi(x, t)$. The probability density function $|\psi(x, t)|^2$ is used to determine the probability of particle occurring in position \textit{x} at any time \textit{t} \cite{sun2004particle, sun2006qpso}. The position of particles is updated according to equations:
\begin{equation}
x_{i, j}(t+1)=p_{i, j} (t) \pm \alpha . |mbest_{i, j}(t)-x_{i, j}(t)|. ln(1/u)
\end{equation}
\begin{equation}
p_{i, j}(t)= (\phi.P_{i, j}(t)+(1-\phi).G_j(t)), (1 \leq i \leq N, 1 \leq j \leq M)
\end{equation}
where each particle must converge to its local attractor $p=(p_1, p_2,..., p_D)$, where \textit{D} is the dimension, \textit{N} and \textit{M} are the number of particles and iterations respectively, $P_{i, j}$ and $G_{j}$ denote the previous and optimal position vector of each particle respectively, $\phi=c_1.r_1/(c_1r_1+c_2r_2)$, where $c_1$; $c_2$ are the acceleration coefficients, $r_1$; $r_2$ and \textit{u} are normally distributed random numbers in (0, 1), $\alpha$ is contraction-expansion coefficient and mbest defines the mean of best positions of particles as:
\begin{equation}
mbest_{i, j}(t)=\dfrac{1}{N} \sum_{i=1, j=1}^{N, M}P_{i, j}(t)=\bigg(\dfrac{1}{N}\sum_{i=1}^{N}P_{i, 1}(t), \dfrac{1}{N}\sum_{i=1}^{N}P_{i, 2}(t), . . ., \dfrac{1}{N}\sum_{i=1}^{N}P_{i, D}(t) \bigg)
\end{equation}
In Eq. (1), $\alpha$ denotes contraction-expansion coefficient, which is setup manually to control the speed of convergence. It can be decreased linearly or fixed. In PSO, $\alpha < 1.782$ to ensure convergence performance of the particle. In QPSO-CD, the value of $\alpha$ is determined by $\alpha$=1-(1.0-0.5)\textit{k}/\textit{M}, i.e. decreases linearly from 1.0 to 0.5 to attain good performance, where \textit{k} is present iteration and \textit{M} is maximum number of iterations.
\section{Hybrid Particle Swarm Optimization}
The hybrid quantum-behaved PSO algorithm with Cauchy distribution and natural selection strategy (QPSO-CD) is described as follows:
The QPSO-CD algorithm begins with the standard QPSO using equations (1), (2) and (3). The position and velocity of particles cannot be determined exactly due to varying dynamic behavior. So, it can only be learned with the probability density function. Each particle can be mutated with Gaussian or Cauchy distribution. We mutated QPSO with Cauchy operator due to its ability to make larger perturbation. Therefore, there is a higher probability with Cauchy as compared to Gaussian distribution to come out of the local optima region. The QPSO algorithm is mutated with Cauchy distribution to increase its diversity, where mbest or global best position is mutated with fixed mutation probability ({\textit{Pr}). The probability density function $(f(x))$ of the standard Cauchy distribution is given as:
\begin{equation}
f(x)=\dfrac{1}{\pi(1+x^2)} ~ ~ ~ ~ -\infty < x < \infty,
\end{equation}
It should be noted that mutation operation is executed on each vector by adding Cauchy distribution random value (\textit{D}(.)) independently such that
\begin{equation}
x^{'}=x+\phi D(.)
\end{equation}
where $x^{'}$ is new location after mutated with random value to \textit{x}. At last, the position of particle is selected and the particles of swarm are sorted on the basis of their fitness values after each iteration. Further, substitute the group of particles having worst fitness values with the best ones and optimal solution is determined. The main objective of using natural mechanism is to refine the capability and accuracy of QPSO algorithm.
\begin{center}
\begin{minipage}{.7\linewidth}
\begin{algorithm}[H]
\caption{QPSO-CD algorithm}
\label{pseudoPSO}
\begin{algorithmic}[1]
\State The swarm is initialized with random numbers distributed uniformly: random $x_i$.
\State Do
\State $\alpha$ decreases linearly from 1.0 to 0.5
\For {\textit{k}=1 to \textit{M} do}
\State $\alpha=1-(1.0-0.5).k/M$
\If {$Pr<rand(0,1)$}
\State Calculate the mbest of the swarm using Eq (3)
\EndIf
\For {\textit{i}=1: to \textit{N} do}
\If {Fitness$(p_i) <$ Fitness$(x_i)$} $p_i=x_i$;
\State $G=argmin(Fitness(p_i)$;
\EndIf
\For{\textit{j}:1 to \textit{D} do}
\State $r_{1}=rand(0, 1); r_{2}=rand(0,1)$;
\State $\phi=c_{1}.r_{1}/(c_{1}r_{1}+c_{2}r_{2})$;
\State $p_{i, j}=\phi.P_{i, j}+(1-\phi). G_j$;
\If {$rand(0,1) < 0.5$}
\State $x_{i,j}=p_{i, j}+\alpha. abs(mbest_{i, j}-x_{i, j}). log (1/u)$
\Else
\State $x_{i,j}=p_{i, j}-\alpha. abs(mbest_{i, j}-x_{i, j}). log (1/u)$
\EndIf
\EndFor
\State $Fx(i)$=Fitness$(x(i,:))$;
\EndFor
\State $[SF, Sx]=sort(Fx)$;
\State $Z=round((N-1)/S)$; \Comment{\textit{S} is selection parameter}
\State $x(Sx((N-Z+1):N))=x(Sx(1:Z))$; \Comment{Sort the particles from best to worst position}
\EndFor
\State Until termination criterion is met
\end{algorithmic}
\end{algorithm}
\end{minipage}
\end{center}
The natural selection method is used to enhance the convergence characteristics of proposed QPSO-CD algorithm, where the fitter solutions are used for the next iteration. The procedure of selection method for \textit{N} particles is as follows:
\begin{equation}
F(X(t))=\{F(x_1(t)), F(x_2(t)), ..., F(x_N(t))\}
\end{equation}
where $X(t)$ is position vector of particles at time \textit{t} and $F(X(t))$ is the fitness function of swarm. Next step is to sort the particles according to their fitness values from best one to worst position such that
\begin{equation}
\begin{split}
F(X^{'}(t))=\{F(x_1^{'}(t)), F(x_2{'}(t)) , ..., &F(x_N{'}(t))\} \\
X^{'}(t)=\{x_1^{'}(t), x_2{'}(t), ..., x_N{'}(t)\}
\end{split}
\end{equation}
In Algorithm 1, $SF$ and $Sx$ are the sorting functions of fitness and position respectively. On the basis of natural selection parameters and fitness values, the positions of swarm particles are updated for the next iteration,
\begin{equation}
\begin{split}
X^{'}(t)=\{x_1^{''}(t), x_2{''}(t), ..., x_S{''}(t)\}, & \\
X^{''}_{k}(t)=\{x_1^{'}(t), x_2{'}(t), ..., x_Z{'}(t)\}
\end{split}
\end{equation}
where $(1 \leq k \leq S)$, \textit{S} denotes the selection parameter, \textit{Z} signifies the number of best positions selected according to fitness values such that $S=N/Z$ and $X^{''}(t)$ is updated position vector of particles. The selection parameter \textit{S} is generally set as 2 to replace the half of worst positions with the half of best positions of particles. It improves the precision of the direction of particles, protect the global searching capability and speed up the convergence.
\section{Experimental results}
The performance of proposed QPSO-CD algorithm is investigated on representative benchmark functions, given in Table 1. Further, the results are compared with classical PSO (PSO), standard QPSO, QPSO with delta potential (QDPSO) and QPSO with mutation operator (QPSO-MO). The details of numerical benchmark functions are given in Table 1.
\begin{table}[!ht]
\centering
\caption{Details of benchmark functions}
\begin{tabular}{|c|c|}
\hline
Test function & Initial range \\
\hline
\multicolumn{2}{|c|}{\textbf{Sphere function}}\\
\hline
$f_1(x)=\sum\limits_{i=1}^{n}x_{i}^2$ & (-100, 100) \\
\hline
\multicolumn{2}{|c|}{\textbf{Rosenbrock function}} \\
\hline
$f_2(x)=\sum\limits_{i=1}^{n}100(x_{i+1}-x_{i}^2)^2+(x_{i}-1)^2$ & (-5.12, 5.12) \\
\hline
\multicolumn{2}{|c|}{\textbf{Greiwank function}}\\
\hline
$f_3(x)=\dfrac{1}{4000} \sum\limits_{i=1}^{n}x_{i}^2-\prod_{i=1}^{n}cos(\dfrac{x_i}{\sqrt{i+1}})+1$& (-600, 600) \\
\hline
\multicolumn{2}{|c|}{\textbf{Rastrigrin function}} \\
\hline
$f_4(x)=\sum\limits_{i=1}^{n}(x_{i}^2-10cos(2\pi x_{i})+10)$ & (-5.12, 5.12) \\
\hline
\end{tabular}
\end{table}
The performance of QPSO has been widely tested for various test functions. Initially, we have considered four representative benchmark functions to determine the reliability of QPSO-CD algorithm. For all the experiments, the size of population is 20, 40 and 80 and dimension sizes are 10, 20 and 30. The parameters for QPSO-CD algorithm are as follows: the value of $\alpha$ decreases from 1.0 to 0.5 linearly, the natural selection parameter \textit{S}=2 is taken, $c_1, c_2$ correlation coefficients are set equal to 2.
The mean best fitness values of PSO, QPSO, QDPSO, QPSO-MO and QPSO-CD are recorded for 1000, 1500 and 2000 runs of each function. Fig. 2 to Fig. 5 depict the performance of functions $f_1$ to $f_4$ with respect to mean best fitness against the number of iterations. In Table 2, \textit{P} denotes the population, dimension is represented by \textit{D} and \textit{G} stands for generation. The numerical results of QPSO-CD shown optimal solution with fast convergence speed and high accuracy. The results shown that QPSO-CD performs better on Rosenbrock function than its counterparts in some cases. When the size of population is 20 and dimension is 30, the results of proposed algorithm are not better than QPSO-MO, but QPSO-CD performs better than PSO, QPSO and QDPSO. The performance of QPSO-CD is significantly better than its variants on Greiwank and Rastrigrin functions. It has outperformed other algorithms and obtained optimal solution (near zero) for Greiwank function. In most of the cases, QPSO-CD is more efficient and outperformed the other algorithms.
\begin{figure}
\centering
\begin{minipage}{.5\textwidth}
\centering
\includegraphics[width=1\linewidth]{f1.eps}
\captionof{figure}{Effectiveness of QPSO-CD for sphere function $f_1$}
\label{fig:test1}
\end{minipage}%
\begin{minipage}{.5\textwidth}
\centering
\includegraphics[width=1\linewidth]{f2.eps}
\captionof{figure}{Effectiveness of QPSO-CD for Rosenbrock function $f_2$}
\label{fig:test2}
\end{minipage}
\end{figure}
\begin{figure} [!ht]
\centering
\begin{minipage}{.5\textwidth}
\centering
\includegraphics[width=1\linewidth]{f3.eps}
\captionof{figure}{Effectiveness of QPSO-CD for Greiwank function $f_3$}
\label{fig:test1}
\end{minipage}%
\begin{minipage}{.5\textwidth}
\centering
\includegraphics[width=1\linewidth]{f4.eps}
\captionof{figure}{Effectiveness of QPSO-CD for Rastrigrin function $f_4$}
\label{fig:test2}
\end{minipage}
\end{figure}
\begin{table}[!ht]
\centering
\caption{Comparison results of Sphere and Rosenbrock functions}
\begin{tabular}{c|c|c|c|c|c|c|c|c|c|c|c|c}
\hline
\multicolumn{3}{c|}{}& \multicolumn{5}{c|}{\textbf{Sphere function}} & \multicolumn{5}{c}{\textbf{Rosenbrock function}}\\
\hline
P & D & G & PSO & QPSO & QDPSO & QPSO-MO & QPSO-CD & PSO & QPSO & QDPSO & QPSO-MO & QPSO-CD \\
\hline
\multirow{3}{*}{20}& 10 & 1000 &0.0 & 4.01e-40 & 1.513e-49 &1.508e-48 & 1.738e-50 & 95.10 & 58.41 & 14.22 & 22.18 & 34.67 \\
& 20 & 1500 & 0.0 & 2.58e-21 & 1.339e-30 & 1.296e-31 & 1.032e-30 & 204.38 & 110.5 & 175.31 & 68.40 & 54.76 \\
& 30 & 2000 & 0.0 & 2.08e-13 & 1.953e-21 & 1.918e-21 & 1.808e-21 & 314.46 & 148.5 & 242.37 & 113.30 & 122.5 \\
\hline
\multirow{3}{*}{40}& 10 & 1000 & 0.0 & 2.73e-67 & 1.087e-73 & 1.146e-51 & 1.154e-72 & 70.28 & 10.42 & 15.86 & 7.985 & 8.843 \\
& 20 & 1500 & 0.0 & 4.84e-28 & 1.397e-42 & 1.417e-42 & 1.237e-41 & 178.98 & 48.45 & 112.46 & 52.93 & 41.77 \\
& 30 & 2000 & 0.0 & 2.02e-25 & 2.850e-30 & 2.471e-28 & 1.946e-23 & 288.58 & 58.32 & 76.42 & 64.19& 58.04 \\
\hline
\multirow{3}{*}{80}& 10 & 1000 & 0.0 & 7.66e-95 & 5.553e-90 & 4.872e-71 & 6.437e-72 & 36.29 & 8.853 & 36.34 & 5.715 & 7.419 \\
& 20 & 1500 & 0.0 & 1.62e-60 & 1.654e-54 & 1.677e-58 & 1.609e-62 & 84.78 & 34.88 & 23.54 &24.45 & 21.78 \\
& 30 & 2000 & 0.0 & 2.05e-44 & 1.042e-40 & 1.131e-42 & 1.128e-41 & 202.58 & 52.17 & 70.81&45.22 & 40.97 \\
\hline
\end{tabular}
\end{table}
\begin{table}[!ht]
\centering
\caption{Comparison results of Greiwank and Rastrigrin functions}
\begin{tabular}{c|c|c|c|c|c|c|c|c|c|c|c|c}
\hline
\multicolumn{3}{c|}{}& \multicolumn{5}{c|}{\textbf{Greiwank function}} & \multicolumn{5}{c}{\textbf{Rastrigrin function}}\\
\hline
P & D & G & PSO & QPSO & QDPSO & QPSO-MO & QPSO-CD & PSO & QPSO & QDPSO & QPSO-MO & QPSO-CD \\
\hline
\multirow{3}{*}{20}& 10 & 1000 & 0.089 & 0.078 & 0.1003 &0.0732 & 0.072 & 5.526 & 5.349 & 4.969 & 4.478 & 4.051 \\
& 20 & 1500 & 0.0300 & 0.2001 & 0.0086 & 0.0189 & 0.0078 & 23.17 & 21.28 & 17.08 & 15.63 & 13.22\\
& 30 & 2000 & 0.0181 & 0.0122 & 0.0544& 0.0103 & 0.0026 & 46.29 & 32.57 & 48.61 & 27.80 & 31.48 \\
\hline
\multirow{3}{*}{40}& 10 & 1000 &0.0826 & 0.055 & 0.048 & 0.0520 & 0.041 & 3.865 & 3.673 & 2.032 & 3.383 & 2.100 \\
& 20 & 1500 & 0.0272 & 0.0149 & 0.0004 & 0.0247& 0.0106 & 15.68 & 14.37 & 10.94 & 11.01 & 10.77\\
& 30 & 2000 & 0.0125 & 0.0117 & 0.0009& 0.0105& 0.0102 & 37.13 & 23.01 & 21.37& 21.01& 21.19\\
\hline
\multirow{3}{*}{80}& 10 & 1000 & 0.0723 & 0.0341 &0.0 & 0.0542& 0.0702 & 2.562 & 2.234 & 0.923 & 2.183& 1.943 \\
& 20 & 1500 & 0.0274 & 0.0189 & 0.0 & 0.0194& 0.0161 & 12.35 & 9.66 & 6.955 & 8.075& 7.021 \\
& 30 & 2000 & 0.0123 & 0.0118 & 0.0 & 0.0082 & 0.0031 & 26.89 & 17.48 & 18.13 &14.99 & 11.73 \\
\hline
\end{tabular}
\end{table}
\section{Correctness and Time Complexity Analysis of a QPSO-CD Algorithm}
In this Section, the correctness and time complexity of a proposed algorithm QPSO-CD is analyzed and compared with the classical PSO algorithm.
\begin{theorem}
The sequence of random variables $\{S_n, n \geq 0\}$ generated by QPSO with Cauchy distribution converges to zero in probability as \textit{\textit{n}} approaches infinity.
\end{theorem}
\begin{proof}
Recall, the probability density function of standard Cauchy distribution and its convergence probability \cite{rudolph1997local} are given as
\begin{equation}
\begin{gathered}
f(s)=\dfrac{1}{\pi (1+s^2)} ~ \text{for} -\infty < s < \infty, \\
P(x \leq S_n \leq y) = \dfrac{1}{\pi} \int_{y}^{x} \dfrac{ds}{(1+s^2)}, ~ \forall x \leq y
\end{gathered}
\end{equation}
Consider a random variable $Q_n$ interpreted as
$$ Q_n = \alpha S_n, ~ \alpha =\dfrac{1}{n^\lambda} $$
where $\lambda$ denotes a fixed positive constant. Correspondingly, the probability density function can be calculated as
\begin{align*}
P(Q_n \leq q) & = P(\alpha S_n \leq q) \\
& = P\left(S_n \leq \dfrac{q}{\alpha}\right)\\
& = \int_{-\infty}^{\dfrac{q}{\alpha}} f(s) ds. \dfrac{dP(q_n \leq q)}{dq}\\
& = \dfrac{1}{\alpha} f\left(\dfrac{q}{\alpha}\right)
\end{align*}
i.e. the probability density function of random variable $Q_n$.
\begin{align*}
P(|Q_n|>\xi) & = P(|S_n|>\xi n^\lambda)\\
& = P(S_n > \xi n^\lambda)+ P(S_n > - \xi n ^\lambda)\\
& = P(\xi n^\lambda < S_n < \infty)+ P(- \infty < S_n < - \xi n^\lambda)
\end{align*}
Using Eq. (9), the probability density function of random variable $Q_n$ becomes
\begin{align*}
P(|Q_n|>\xi) & = \dfrac{1}{\pi}\int_{\xi n^\lambda}^{\infty}\dfrac{ds}{\pi (1+s^2)} + \dfrac{1}{\pi}\int_{\infty}^{-\xi n^\lambda}\dfrac{ds}{\pi (1+s^2)} \\
& = \left[1+\dfrac{1}{\pi} \int_{\xi n^\lambda}^{-\xi n^\lambda}\dfrac{ds}{\pi (1+s^2)}\right]\\
& = 1-\dfrac{1}{\pi} \int_{-\xi n^\lambda}^{\xi n^\lambda}\dfrac{ds}{\pi (1+s^2)}= 0 ~ \text{as} ~ n \rightarrow \infty
\end{align*}
This completes the proof of the theorem.
\end{proof}
\begin{defn}
Let $\{S_n\}$ a random sequence of variables. It is converges to some random variable \textit{s} with probability 1, if for every $\xi > 0$ and $\lambda >0$, there exists $n_1(\xi, \lambda)$ such that $P(|S_n-s|< \xi) > 1-\lambda, \forall n > n_1$ or
\begin{equation}
P\left(\lim_{n \rightarrow \infty} |S_n-s|< \xi\right)=1
\end{equation}
\end{defn}
The efficiency of the QPSO-CD algorithm is evaluated by number of steps needed to reach the optimal region $R(\xi)$. The method is to evaluate the distribution of number of steps needed to hit $R(\xi)$ by comparing the expected value and moments of distribution. The total number of stages to reach the optimal region is determined as $W(\xi)=inf\{n \mid f_n \in R(\xi)\}$. The variance $V(W(\xi))$ and expectation value $E(W(\xi))$ are determined as
\begin{equation}
E(W(\xi))= \sum_{n=0}^{\infty}nx_n,
\end{equation}
\begin{equation}
\begin{split}
V(W(\xi)) & =E(W^2(\xi))-\{E(W(\xi))\}^2\\
& = \sum_{n=0}^{\infty}n^2x_n-\left(\sum_{n=0}^{\infty}nx_n\right)^2
\end{split}
\end{equation}
In fact, the $E(W(\xi))$ depends upon the convergence of $\sum_{n=0}^{\infty}nx_n$. It is needed that $\sum_{j=0}^{\infty}x_n=1$, so that QPSO-CD can converge globally. The number of objective function evaluations are used to measure time. The main benefit of this approach is that it shows relationship between processor and measure time as the complexity of objective function increases. We used Sphere function $f(x)=x^{T}.x$ with a linear constraint $g(x)=\sum^{n}_{j=0}x_j \geq 0$ to compute the time complexity. It has minimum value at 0. The value of optimal region is set as $R(\xi)=R(10^{-4}).$ To determine the time complexity, the algorithms PSO and QPSO-CD are executed 40 times on $f(x)$ with initial scope [-10, 10]$^{N}$, where \textit{N} denotes the dimension. We determine the mean number of objective function evaluations ($W(\xi)$), the variance ($V(W(\xi))$), the standard deviation (SD) ($\sigma_{W(\xi)}$), the standard error (SE) ($\sigma_{W(\xi)}/ \sqrt{40}$) and ratio of mean and dimension ($W(\xi)/N$). The contraction coefficient $\alpha= 0.75$ is used for QPSO-CD and constriction coefficient $\chi=0.73$ for PSO with acceleration factors $c_1=c_2$=2.25.
\begin{table}[!ht]
\centering
\caption{Results of the time complexity for QPSO-CD algorithm}
\begin{tabular}{c|c|c|c|c|c}
\hline
Dimension (N) & Mean & Variance & SD & SE & Mean/N \\
\hline
2 &302.38 & 4164.3 & 64.532 &10.203 & 151.19 \\
3 & 452.18 & 4541.9 & 67.394 &10.655 & 150.72\\
4 & 621.29 & 5208.2 & 72.168 & 11.410 & 155.32\\
5 &755.88 & 6675.0 & 81.701 & 12.918 & 151.17\\
6 & 879.13 & 8523.7 & 92.324 & 14.597 & 146.52\\
7 & 1022.06 & 9575.4 & 97.854 & 15.472 & 146.00\\
8 & 1158.52 & 10269.7 & 101.341 & 16.023 & 144.81\\
9 & 1308.17 & 12053.4 & 109.788 & 17.359 & 145.35 \\
10 & 1459.3 & 12648.3 & 112.465 & 17.782 & 145.93\\
\hline
\end{tabular}
\end{table}
\begin{table}[!ht]
\centering
\caption{Results of the time complexity for PSO algorithm}
\begin{tabular}{c|c|c|c|c|c}
\hline
Dimension (N) & Mean & Variance & SD & SE & Mean/N \\
\hline
2 &691.4 & 17297.5 & 131.52 & 20.795 & 345.7\\
3 & 979.1 & 22281.5& 149.27 & 23.601& 326.3\\
4 & 1167.2 & 24282.9 & 155.72 & 24.638 & 291.8\\
5 & 1328.7 & 21853.7 & 147.83 & 23.373& 265.7\\
6 & 1489.9 & 32008.7& 178.91 & 28.288& 248.3 \\
7 & 1744.3 & 502297 & 224.12 & 35.436 & 249.1 \\
8 & 1978.5 & 41233.3 & 203.06 & 31.106 & 247.3 \\
9 & 2259.1 & 36217.8 & 190.31 & 30.090& 251.0\\
10 & 2604.2 & 43559.8 & 208.71 & 32.999& 260.4\\
\hline
\end{tabular}
\end{table}
\begin{figure*}[!ht]
\centering
\includegraphics[scale=0.6]{Time.eps}
\caption{Time complexity results for PSO and QPSO-CD}
\end{figure*}
\begin{figure*}[!ht]
\centering
\includegraphics[scale=0.6]{Roceffi.eps}
\caption{Comparison of Correlation coefficient of PSP and QPSO-CD}
\end{figure*}
Table 1 and 2 show the statistical results of time complexity test for QPSO-CD and PSO algorithm, respectively. Fig 6 indicates that the time complexity of proposed algorithm increases non-linearly as the dimension increases. However, the time complexity of PSO algorithm increases adequately linearly. Thus, the time complexity of QPSO-CD is lower than PSO algorithm. In Fig 7, QPSO-CD shows a strong correlation between $W(\xi)$ and \textit{N}, i.e. the correlation coefficient is 0.9996. For PSO, the linear correlation coefficient is 0.9939, which is not so phenomenal as that in case of QPSO-CD. The relationship between mean and dimension clearly shows that the value of correlation coefficient is fairly stable for QPSO-CD as compared to PSO algorithm.
\section{QPSO-CD for Constraint Engineering Design Problems}
There exists several approaches for handling constrained optimization problems. The basic principle is to convert the constrained optimization problem to unconstrained by combining objective function and penalty function approach. Further, minimize the newly formed objective function with any unconstrained algorithm. Generally, the constrained optimization problem can be described as in Eq. (13).
The objective is to minimize the objective function \textit{f(x)} subjected to equality $(h_j(x))$ and inequality $(g_i(x))$ constrained functions, where $p(i)$ is the upper bound and $q(i)$ denotes the search space lower bound. The strict inequalities of form $g_i(x) \geq 0$ can be converted into $-g_i(x)\leq 0$ and $h_i(x)$ equality constraints can be converted into inequality constraints $h_i(x) \geq 0$ and $h_i(x) \leq 0$. Sun et al. \cite{sun2007using} adopted non-stationary penalty function to address non-linear programming problems using QPSO. Coelho \cite{dos2010gaussian} used penalty function with some positive constant i.e. set to 5000. We adopted the same approach and replace the constant with dynamically allocated penalty value.
\begin{equation}
\begin{split}
& \min\limits_{x}=f(x) \\
\text{subject to} \\
& g_i(x) \leq 0, i=0, 1,... n-1\\
& h_j(x)=0, j=1,2,...r \\
& p(i)\leq x_i \leq q(i), 1 \leq i \leq m\\
& x=\{x_1,x_2, x_3,..., x_m\}
\end{split}
\end{equation}
Usually, the procedure is to find the solution for design variables that lie in search space upper and lower bound constraints such that $x_i \in [p(i), q(i)]$. If solution violates any of the constraint, then the following rules are applied
\begin{equation}
\begin{split}
& x_i=x_i + \{p(x_i)-q(x_i)\}.~ rand[0,1] \\
& x_i=x_i - \{p(x_i)-q(x_i)\}.~ rand[0,1]
\end{split}
\end{equation}
where rand[0, 1] is randomly distributed function to select value between 0 and 1. Finally, the unconstrained optimization problem is solved using dynamically modified penalty values according to inequality constraints $g_i(x)$. Thus, the objective function is evaluated as
\begin{equation}
F(x)=\left\{
\begin{array}{ll}
f(x)~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~\text{if}~ g_i(x)\leq 0\\
f(x)+y(t). \sum\limits_{i=1}^{n} g_i(x)~ ~ \text{if}~ g_i(x)>0\\
\end{array} \right \}
\end{equation}
where $f(x)$ is the main objective function of optimization problem in Eq. (13), \textit{t} is the iteration number and $y(t)$ represents the dynamically allocated penalty value.
In this Section, QPSO-CD is tested for three-bar truss, tension/compression spring and pressure vessel design problems consisting
different members and constraints. The performance of QPSO-CD is compared and analyzed with the results of PSO, QPSO, and SP-QPSO algorithms as reported in the literature.
\subsection{Three-bar truss design problem}
Three-bar truss is a constraint design optimization problem, which has been widely used to test several methods. It consists cross-section areas of three bars $x_1$ (and $x_3$) and $x_2$ as design variables. The aim of this problem is to minimize the weight of truss subject to maximize the stress on these bars. The structure should be symmetric and subjected to two constant loadings $P_1=P_2=P$ as shown in Fig 8. The mathematical formulation of two design bars ($x_1$, $x_2$) and three restrictive mathematical functions are described as:
\begin{equation}
\begin{split}
\text{Min} ~ f(x)& = (2\sqrt{2}x_1+x_2). l \\
\text{Subject to:} & \\
g_1(x)&=\dfrac{\sqrt{2}x_1+x_2}{\sqrt{2}x_1^{2}+2x_1x_2}P-\sigma \leq 0,\\
g_2(x)&= \dfrac{x_2}{\sqrt{2}x_1^{2}+2x_1x_2}P-\sigma \leq 0 \\
\end{split}
\end{equation}
$$
g_3(x)= \dfrac{1}{x_1+\sqrt{2}x_2}P-\sigma \leq 0, ~ \text{where}$$
$$ 0 \leq x_1, x_2 \leq 1, l=100cm, P=\sigma=2 KN/cm^{2}$$
\begin{figure}[ht]
\centering
\includegraphics[scale=0.75]{bt.eps}
\caption{Structure of Three-bar truss}
\end{figure}
\begin{table}[!ht]
\centering
\caption{Comparison of optimal results for three-bar truss problem}
\begin{tabular}{c|c|c|c|c}
\hline
Variables & PSO & QPSO & SP-QPSO & QPSO-CD \\
\hline
$x_1$ &0.78911058 &0.788649 & 0.788796 & 0.788658\\
$x_2$& 0.40702683 & 0.408322 & 0.407898& 0.40828488\\
$g_1(x)$ & -6.6720e-06 & 1.6313e-07 & 6.4748e-06 & 9.00037e-06 \\
$g_2(x)$ & -1.4655 & -1.4640 & -1.4644 & -1.4640 \\
$g_3(x)$ & -0.5345 &-0.5359 & -0.5354 & -0.5359\\
$f(x)$ & 263.89686 & 263.89584 & 263.89500 & 263.89465\\
\hline
\end{tabular}
\end{table}
The results are obtained by QPSO-CD are compared with its counterparts in Table 6. For three-bar truss problem, QPSO-CD is superior to optimal solutions previously obtained in literature. The difference of best solution obtained by QPSO-CD among other algorithms is shown in Fig 9.
\begin{figure}[H]
\centering
\includegraphics[scale=0.55]{tbar.eps}
\caption{Optimal results of PSO, QPSO, SP-QPSO and QPSO-CD algorithms for three bar truss problem}
\end{figure}
\subsection{Tension/Compression spring design problem}
The main aim is to lessen the volume \textit{V} of a spring
subjected to tension load constantly as shown in Fig 10. Using the symmetry of structure, there are practically three design variables ($x_1, x_2, x_3$), where $x_1$ is the wire diameter, the coil diameter is represented by $x_2$ and $x_3$ denotes the total number of active coils. The mathematical formulation for this problem is described as:
\begin{equation}
\begin{split}
\text{Min} ~ f(x)&= (x_3+2)x_2x_1^2, \\
\text{Subject to:} & \\
g_1(x)&=\dfrac{1-x_2^{3}x_3}{71785x_16{4}} \leq 0,\\
g_2(x)&= \dfrac{4x_2^{2}-x_1x_2}{12566(x_2x_1^{3}-x_1^{4})}+\dfrac{1}{5108x_1^{2}}-1 \leq 0,\\
g_3(x)&= \dfrac{1-140.45x_1}{x_2^2x_3}\leq 0,\\
g_4(x)&= \dfrac{x_2+x_1}{1.5}-1\leq 0,\\
\text{where} & \\
0.05 \leq x_1 & \leq 2, 0.25 \leq x_2, \leq 1.3, 2 \leq x_3 \leq 15, \\
\end{split}
\end{equation}
\begin{figure}[H]
\centering
\includegraphics[scale=0.38]{ts.eps}
\caption{Structure of tension/compression spring}
\end{figure}
\begin{table}[!ht]
\centering
\caption{Comparison of optimal results for tension spring design problem}
\begin{tabular}{c|c|c|c|c}
\hline
Variables & PSO & QPSO & SP-QPSO & QPSO-CD \\
\hline
$x_1$ & 0.0516 & 0.0524 & 0.05 & 0.0513 \\
$x_2$& 0.3542 & 0.2505 & 0.25 & 0.2502 \\
$x_3$ & 11.7942 & 2 & 2 & 2\\
$g_1(x)$ & -2.3006e-02 & 0.93145095 & 0.93034756 & 4.11004e-06 \\
$g_2(x)$ & -5.6059e-03 &-0.17471558 & -0.16568318 & -0.17352479 \\
$g_3(x)$ & -3.9057 & -50.67 & -48.180 & -49.561 \\
$g_4(x)$ & -0.7294 & -0.79986567 & -0.80 & -0.799 \\
$f(x)$ & 0.01305 &0.00275 & 0.00250 & 0.00263\\
\hline
\end{tabular}
\end{table}
It has been observed that QPSO algorithm with Cauchy distribution and natural selection strategy is robust and obtains optimal solutions than
PSO and QPSO, shown in Table 7. The difference between best solutions found by QPSO-CD ($f(x)=0.00263$) and other algorithms for tension spring design problem are reported in Fig 11.
\begin{figure}[H]
\centering
\includegraphics[scale=0.55]{tspg.eps}
\caption{Results of PSO, QPSO, SP-QPSO and QPSO-CD methods for tension spring design problem}
\end{figure}
\subsection{Pressure vessel design problem}
Initially, Kannan and Kramer \cite{kannan1994augmented} studied the pressure vessel design problem with the main aim to reduce the total fabricating cost. Pressure
vessels can be of any shape. For engineering purposes, a cylindrical design capped by hemispherical heads at both ends is widely used \cite{sandgren1990nonlinear}. Fig 12 describes the structure of pressure vessel design problem. It consists four design variables ($x_1, x_2, x_3, x_4$), where $x_1$ denotes the shell thickness $(T_s)$, $x_2$ is used for head thickness ($T_h$), $x_3$ denotes the inner radius (\textit{R}) and $x_4$ represents the length of vessel (\textit{L}). The objective function and constraint equations are described as:
\begin{figure}[H]
\centering
\includegraphics[scale=0.62]{pv.eps}
\caption{Design of pressure vessel}
\end{figure}
\begin{equation}
\begin{split}
\text{Min} ~ f(x)& = 0.6224x_1x_3x_4 + 1.7781x_2x_3^{2}+ 3.166x_1^{2}x_4 \\ & +19.84x_1^{2}x_3, \\
\text{Subject to:} & \\
g_1(x)&= -x_1+0.0193x_3 \leq 0,\\
g_2(x)&= -x_2+0.00954x_3\leq 0,\\
g_3(x)&= -\pi x_3^{2}x_4-\dfrac{4}{3}\pi x_3^{3}+1296000 \leq 0,\\
g_4(x)&= x_4-240 \leq 0,\\
\text{where} & \\
1\times 0.0625 & \leq x_1, x_2 \leq 99 \times 0.0625, 10 \leq x_3 ,x_4 \leq 200 \\
\end{split}
\end{equation}
\begin{table}[!ht]
\centering
\caption{Comparison of optimal results for Pressure vessel design problem}
\begin{tabular}{c|c|c|c|c}
\hline
Variables & PSO & QPSO & SP-QPSO & QPSO-CD \\
\hline
$x_1$ & 0.8125 &0.7783 & 0.7782 & 0.7776\\
$x_2$& 0.4375& 0.3849 & 0.3845 & 0.3848\\
$x_3$ & 42.0984 & 40.3289 & 40.3206 & 40.3278\\
$x_4$ & 176.6365 & 199.8899 & 199.9988 & 199.8865 \\
$g_1(x)$ & -4.500e-15 & 4.777e-05 & -1.242e-05 & 7.2654e-04\\
$g_2(x)$ & -0.035880 & -1.62294e-04 & 1.58523e-04 & -7.2787e-05 \\
$g_3(x)$ & -1.164e-10 & -97.39720071 & -63.63686942 & -0.734359\\
$g_4(x)$ & -63.3634 & -40.1100 & -40.0012 & -40.1135 \\
$f(x)$ & 6059.714
&5886.189 & 5885.268 & 5886.137\\
\hline
\end{tabular}
\end{table}
\begin{figure}[!ht]
\centering
\includegraphics[scale=0.55]{pvdp.eps}
\caption{Optimal results of PSO, QPSO, SP-QPSO and QPSO-CD techniques for pressure vessel design problem}
\end{figure}
The optimal results of QPSO-CD is compared with the SP-QPSO, QPSO and PSO best results noted in the previous work, and are given in Table 8. The best solution obtained from QPSO-CD is better than other algorithms as shown in Fig 13.
\section{Conclusion}
In this paper, a new hybrid quantum particle swarm optimization algorithm is proposed with natural selection method and Cauchy distribution. The performance of the proposed algorithm is experimented on four benchmark functions and the optimal results are compared with existing algorithms. Further, the QPSO-CD is applied to solve engineering design problems. The efficiency of QPSO-CD
is successfully presented with superiority than preceding results
for three engineering design problems: three-bar truss, tension/compression spring and pressure vessel. The efficiency of QPSO-CD algorithm is evaluated by number of steps needed to reach the optimal region and proved that time complexity of proposed algorithm is lower in comparison to classical PSO. In the context of convergence, the experimental outcomes shown that the QPSO-CD converge to get results closer to the superior solution.
\section*{Additional information}
\textbf{Competing interests:} The authors declare no competing interests.
\section*{Acknowledgement}
S.Z. acknowledges support in part from the National Natural Science
Foundation of China (Nos. 61602532), the Natural Science Foundation of Guangdong Province of China (No. 2017A030313378), and the Science and Technology Program of Guangzhou City of China (No. 201707010194).
|
\section{Appendix}
\label{appendix}
\subsection{Additional Evaluation Results}
\begin{table}[H]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
Measure & Monitoring & Analysis \\ \hline \hline
Avg. Response Time & 172.97ms & 136.30ms \\ \hline
Avg. KS-Test & -- & 0.1467 \\ \hline
Avg. Wasserstein Distance & -- & 23.9742 \\ \hline
\end{tabular}
\caption{Aggregated measurements and metrics (across 100 repetitions) about the calibration accuracy for CoCoME's "bookSale" service}
\label{table:cocome:metrics}
\end{table}
\input{externals/wasserstein/wasserstein_cocome_all.tex}
\input{externals/comparisons/wasserstein_teastore_parametric.tex}
\begin{table}[H]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
Measure & Monitoring & PM & NPM \\ \hline \hline
Avg. Response Time & 94.92ms & 86.62ms & 102.25ms \\ \hline
Avg. KS-Test & -- & 0.1267 & 0.1239 \\ \hline
Avg. Wasserstein Distance & -- & 19.1764 & 16.2669 \\ \hline
\end{tabular}
\caption*{PM=Parameterized Model, NPM=Non-Parameterized Model}
\caption{Aggregated measurements and metrics (across 100 repetitions) about the calibration accuracy for TeaStore's "train" service using a load of \textbf{20 users}}
\label{table:teastore:param:20}
\end{table}
\begin{table}[H]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
Measure & Monitoring & PM & NPM \\ \hline \hline
Avg. Response Time & 154.19ms & 184.85ms & 103.87ms \\ \hline
Avg. KS-Test & -- & 0.1609 & 0.2575 \\ \hline
Avg. Wasserstein Distance & -- & 39.6138 & 51.3531 \\ \hline
\end{tabular}
\caption*{PM=Parameterized Model, NPM=Non-Parameterized Model}
\caption{Aggregated measurements and metrics (across 100 repetitions) about the calibration accuracy for TeaStore's "train" service using a load of \textbf{40 users}}
\label{table:teastore:param:40}
\end{table}
\section{Running Example}
\label{Running Example}
We introduce a motivating example that illustrates our approach and was used to evaluate it (see Sec. \ref{evaluation}). The example is part of the TeaStore case study \cite{teastore}: a website to buy different kinds of tea. In this case study, the Recommender component is responsible for calculating the recommendations for a certain shopping cart using the services \textit{‘train’} and \textit{‘recommend’}. The \textit{‘train’} service derives information from the previous orders and prepares the data for the \textit{‘recommend’} service. Because there are different strategies to recommend a list of related items, the developers implemented four versions of \textit{‘recommend’} and \textit{'train'} along different development iterations.
These implementations have different performance characteristics. Performance tests or monitoring can be used to discover these characteristics for the current state, i.e., for the current deployment and the current workload. However, predicting the performance for another state (e.g., different deployment or workload) is expensive and challenging because it requires setting up and performing several tests for each implementation alternative.
In our example, answering the following questions is challenging based on APM: “Which implementation would perform better if the load or the deployment is changed?” or “How well does the \textit{'train'} service perform during yet unseen workload scenarios?” An example for the latter question would be an upcoming offer of discounts, where architects expect an increased number of customers and also a changed behavior of customers in that each customer is expected to order more items.
AbPP can answer these questions faster using simulations instead of the expensive tests if an up-to-date PM is available.
Regardless how the model will be built and updated (reverse engineering extraction or manual/ automatic update), all available approaches recalibrate the whole model by monitoring all parts of the source code instead of recalibrating only the model parts affected by the last changes in source code. For example, the changes in the implementation of \textit{'train'} belong to the last part of the code which is represented as an internal action ‘trainForRecommender’ by modelling the behaviour using SEFF (see Fig. \ref{fig:seff:train}). This means that these changes have only impact on the RD of ‘trainForRecommender’ and all other PMPs are valid (e.g., RD of preprocessor internal action). Recalibrating the whole PMPs loses potential previous manual changes and causes unnecessary monitoring overhead. Additionally, not all the available approaches detect the parametric dependencies (e.g., the RD of ‘trainForRecommender’ is related to the number of ordered items).
\begin{figure*}
\vspace{-0.2 cm}
\centering
\includegraphics[width=0.88\textwidth]{images/MbDevOpsPipline-Final.pdf}
\vspace{-0.2 cm}
\caption{Model-based DevOps pipeline}
\vspace{-0.5 cm}
\label{pipeline}
\end{figure*}
\vspace{-0.1 cm}
\section{Continuous Integration of Performance Model}
\label{CIPM}
This section provides an overview of the \ac{CIPM} approach, before describing how it is embedded in a continuous software engineering approach in \Cref{MbDevOps} and before providing more details on the CIPM activities in Sections VI-VII.
Our approach \ac{CIPM} incrementally extracts and calibrates archi\-tecture-level \acp{PM} with parametric dependencies after each source code commit.
To do so, CIPM updates the PM continuously to keep it consistent with the running system, i.e., the deployed source code and the last measurements.
CIPM consists of four main activities:
\begin{enumerate}
\item Performance model update and adaptive instrumentation: CIPM analyzes the source code changes, updates the architectural PM and static behavior model based on the co-evolution approach \cite{langhammer2017a} and instruments the changed parts of code to calibrate the new/ updated related parts of the architecture as will be discussed in \Cref{adaptive instrumentation}.
\item Monitoring: CIPM collects the required measurements either during testing or executing the system in production.
\item Incremental calibration of \ac{PM} (Sec. \ref{incremental calibration}): CIPM performs the \textit{Dev-time calibration} of behavior, i.e., \ac{PMPs} (Sec. \ref{Dev-time}) considering the parametric dependencies (Sec. \ref{parametric dependencies}) and the \textit{Ops-time calibration}, i.e., updating the deployment and usage parts of \ac{PM} (Sec. \ref{Run-time calibration}).
\item Self-validation:
CIPM automatically starts a simulation and calculates the variation between the simulation results and the monitoring data to show the estimation error before performing \ac{AbPP}.
\end{enumerate}
\label{Preserving the consistency}
To realize these four activities, CIPM extends the co-evolution approach with a monitoring process to keep the PM consistent with the last up-to-date measurements. The static structure and behavior update of PM is already provided by the co-evolution approach. For the adaptive instrumentation, we (a) extend the \ac{VSUM} with an \ac{IM} that describes and manages the instrumentation points and (b) define the consistency rules between the IM and source code. The rules define how \vitruv has to respond to the changes in code by adaptive instrumentation of the changed parts with predefined monitoring probes. For the monitoring, we (c) define a \ac{MM} that describes the resulting monitoring records. These records belong to specific types that are responsible for collecting the necessary monitoring information to calibrate the SEFF elements. Finally, for the incremental calibration and estimation of parametric dependencies, we (d) define the consistency rules between MM and PM, which analyze the monitoring data, calibrate the PMPs and update the deployment and usage parts of PM.
With our approach, we can also handle overlapping commits that occur while the monitoring and calibration of a previous commit are still ongoing by managing multiple copies of the \ac{PM} and keeping track of which model instance belongs to which version of the source code.
\begin{comment}
\end{comment}
\vspace{-0.1 cm}
\section{The Model-based DevOps Pipeline
\label{MbDevOps}
DevOps practices aim to close the gap between the development and operations and to integrate them into one reliable process.
We extend the DevOps practices to integrate and automate the \ac{CIPM} in a \ac{MbDevOps} pipeline. This enables \ac{AbPP} during DevOps-oriented development. The next paragraphs explain the pipeline processes. The lower case numbers refer to the process's number in Fig. \ref{pipeline}.
\begin{comment}
\ac{MbDevOps} integrates CIPM in the development stage to update the architectural PM and instrument the code (\nth{1} CIPM activity). Then, \ac{MbDevOps} calibrates PM (\textit{Dev-time calibration}) using measurements monitored (with CIPM \textit{Monitoring}) by performing only the necessary performance tests. After the self validation of PM (\nth{4} CIPM activity), one will have an up-to-date PM to predict the performance and answer the what-if questions (e.g., how performance would be if we use specific resources). This achieves the first goal of \ac{MbDevOps}: An up-to-date PM can speed up the evaluation of performance requirements and can lower the effort and cost needed to set up the test environment to perform such prediction using performance tests.
\end{comment}
\begin{comment}
After the deployment in production environment, \ac{MbDevOps} monitors the system in specific time intervals during the operation (CIPM \textit{Monitoring}), and updates the extracted PM continuously: potential re-calibration of some PMPs (the first part of CIPM \textit{calibration} activity) and responding to run-time changes, e.g., changes in usage or available resources (\textit{Ops-time calibration}). This will provide an up-to-date run-time PM and achieves the second goal of MbDevOps: supporting the improvement and planning stages of \ac{CSE}; performing model-based experiments to assure the performance will support the improvement, e.g., experiments to support model-based auto scaling. Additionally, having an up-to-date descriptive PM supports the planning of the next version (first stage of \ac{CSE}). This is due to the advantages of models, like increasing the understandability of the current version, modelling and evaluating different design alternatives, and answering what-if questions.
\end{comment}
The \ac{MbDevOps} pipeline starts on the ``development'' side with \(continuous\:integration_{1}\:(CI)\) process \cite{ci1} that merges the source code changes of the developers (cf.~ Fig. \ref{pipeline}). \vitruv responds to the changes in source code with the first CIPM activity: \(updating \:the \:architectural\;PM_{1.1}\) and \(generating\: the\: required\: instrumentation\: points\:in\:IM_{1.2}\) (cf. Sec. \ref{co-evaluation approach}).
Next, \(adaptive\: instrumentation_{2}\) instruments the changed parts of source code using the instrumentation points from \ac{IM} (Sec. \ref{adaptive instrumentation}). The following process is the \(performance\:testing_{3}\), which integrates the second CIPM activity 'Monitoring' to generate the necessary measurements for calibration. The pipeline divides the measurements into training and validation set. Afterwards, the \(Dev{-}time\: calibration_{4}\) (the first part of \nth{3} CIPM activity) enriches the \ac{PM} with \ac{PMPs} using the training set. \Cref{incremental calibration} describes the incremental calibration process whereas \Cref{parametric dependencies} explains how we parametrize the \ac{PMPs} with the influencing input parameters. After the calibration, the pipeline starts the \(self{-}validation\:with\:test\:data_{5}\) process (the \nth{4} CIPM activity), which uses the validation set to evaluate the calibration accuracy.
If the model is deemed accurate, developers can use the resulting \ac{PM} to answer the performance questions using \(AbPP_{6}\). If not, they can change the test configuration to recalibrate \ac{PM}. Answering the performance questions using \ac{AbPP} instead of the test-based performance prediction reduces both the effort and cost of setting up the test environment to perform this prediction.
The ``operations'' side of \Cref{pipeline} starts on the \(continuous\: deployment_{7}\) in the production environment. CIPM \(Monitoring_{8}\) in the production environment generates the required run-time measurements for the next process: \(Ops{-}time\: calibration_{9}\) (the second part of the \nth{3} CIPM activity). The ops-time calibration calibrates and updates both usage model and deployment model (Sec. \ref{Run-time calibration}). Next, the \(self{-}validation\:with\:monitoring\:data_{10}\) (the \nth{4} CIPM activity) validates the estimated \ac{PMPs}. If the \ac{PM} is not accurate enough, the \(adaptive\:recalibration_{11}\) process (Sec. \ref{incremental calibration}) recalibrates the inaccurate model parts using monitoring data.
Finally, the developers can perform more \(model-based\: analysis_{12}\) on the resulting model, e.g., model-based auto scaling. Additionally, having an up-to-date descriptive PM supports the \(development\) \(planning_{13}\). This is due to the advantages of models: increasing the understandability of the current version, modelling and evaluating design alternatives and answering what-if questions.
\vspace{-0.3 cm}
\section{Adaptive Instrumentation}
\label{adaptive instrumentation}
The goal is to instrument the parts of source code, which have been changed and their changes may affect the validity of related PMPs. In our running example, only the source code of 'trainForRecommender' is instrumented to provide the required measurements to update the RD, whereas there is no need to instrument the loop or the code of ‘preprocess’, because these parts of code are not changed, consequently the old estimations of the PMPs remain valid.
To automate the adaptive instrumentation, we use the consistency rules between the source code model and IM, which responds to changes in method body (e.g., adding/ updating statements) by instrumenting the related parts. The rules reconstruct the \ac{SEFF} using a reverse engineering tool \cite{krogmann2012ashort} and map code statements to their \ac{SEFF} element (e.g., internal actions).
Then, they create the required probes (e.g., service probe, internal action probe, loop probe or branch probe) that refer to the \ac{SEFF} elements whose code statements have been changed.
We define the following monitoring record types that are related to the aforementioned probe types using \ac{IRL} (cf. Section~\ref{kieker}).
\begin{itemize}
\item Service call record to monitor the following:\\
- the input parameters properties (e.g., type, value, number of list elements, etc.) that should be considered later as candidates for parametric dependency investigation,\\
- the caller of this service execution to learn the parametric dependency between the input parameters of both caller and callee services.\\
- the allocation context that captures where the component offering this service is deployed.
\item Internal-action record type to monitor the response time of the internal actions.
\item Loop record to monitor the number of loop iterations.
\item Branch record to monitor the selected branch.
\end{itemize}
More details about the records types are in \cite[Chapter~4.3.3]{jaegers}.
Finally, we implemented a model-based instrumentation to generate the instrumented source code as a \vitruv view. This view combines the information from two models: the source code model and the \ac{IM}.
The instrumentation starts with generating the instrumentation code for each probe in the IM according to the probe type. Then it injects the instrumentation codes into a copy of the source code. To detect the correct places for the instrumentation codes, the instrumentation process uses the relations stored in \vitruv correspondence model, i.e., the relation between probes and SEFF elements, the relation between SEFF elements and their source code statements and relation between the original source code and the generated copy (instrumented source code).
\vspace{-0.1 cm}
\section{Incremental Calibration}
\label{incremental calibration}
The following subsections explain the calibration types.
\vspace{-0.1 cm}
\subsection{\ac{Dev-time} calibration}
\label{Dev-time}
The \ac{Dev-time} changes that we consider in this paper are the source code changes that may have an impact on performance, i.e., changes in a method body
On one hand the \textit{incremental} calibration of the SEFF branches, loops or external call arguments requires only to instrument the related source code and to analyze the resulting measurements (e.g., loop iteration number, the selected branch transition, and the values of external call parameters). The goal of this analysis is to detect whether there are dependencies to the service input parameters and express these PMPs sequentially as stochastic expression (cf. \Cref{loop,branches,arguments}).
On the other hand, the incremental calibration of the internal actions with \ac{RD} is challenging because we aim to estimate the \ac{RD} of internal actions incrementally without high monitoring overhead. The existing \ac{RDE} approaches either estimate the RDs at the service level \cite{spinner2015evaluating} or require expensive fine-grained monitoring \cite{Brosig2009Automated,krogmann2009ck}. Therefore, we propose in the following paragraph a light-weight \ac{RDE} process that is based on adaptive instrumentation and monitoring to allow for an incremental \ac{RDE}.
Our incremental \ac{RDE} aims to estimate the \ac{RDs} in the case of adaptive monitoring, i.e., monitoring the changed parts of source code. In our running example, monitoring the code of ‘trainForRecommender’ to reestimate its RD. For this goal, we extend the approach of Brosig et al. \cite{Brosig2009Automated}.
\paragraph {Basis: Non-incremental estimation of resource demands}
Brosig et al. approximate the \ac{RDs} with measured response times in the case of low resource utilization, typically 20\%. Otherwise, they estimate the \ac{RD} of internal action \(i\) (a part of SEFF, see \Cref{palladio}) for resource \(r\) (\(D_{i,r}\)) based on service demand law \cite{menasce2004} shown in equation (1). Here, \(U_{i,r}\) the average utilization of resource r due to executing internal action \(i\) and \(C_{i}\) is the total number of times that internal action \(i\) is executed during the observation period of fixed length \(T\):
\begin{equation}
D_{i,r} = \frac{U_{i,r}}
{{C_{i}}/{T}} = \frac{U_{i,r} \cdot T }
{C_{i}}
\end{equation}
Brosig et al. measure the \(C_{i}\) and estimate \(U_{i,r}\) by using the weighted response time ratios of the total resource utilization, which is not applicable in our adaptive case where not all internal actions are monitored. Therefore, we extend their approach to estimate \(U_{i,r}\) and as a result \(D_{i,r}\) based on the available measurements and the old \ac{RDs} estimations.
\paragraph{Incremental estimation of resource demands}
\label{ird}
Our new approach distinguishes internal actions into two categories based on whether they have been modified in the source code commit preceding the incremental calibration. We denote internal actions whose corresponding code regions have been modified in the preceding source code commit as \ac{MIAs}, e.g., '\textit{trainForRecommender}' in our running example, -- for these code regions, the consistency rules will generate instrumentation probes and the adaptive monitoring produce monitoring results. We denote internal actions whose corresponding code regions have not been changed in the preceding source code commit as \ac{NMIAs}, e.g., '\textit{preprocess}' in our running example -- monitoring data for these code regions has already been observed in a previous iteration and, consequently, we have already an estimation of their RDs.
Based on the fact that the total utilization \(U_{r}\) is measurable and the utilization due to executing \ac{NMIAs} can be estimated based on the old estimations of RDs, we can estimate \(U_{r, MIAs}\) and estimate the RD for each internal action \(i \in MIAs\) accordingly as it will be explained in following paragraphs.
To estimate (\(U_{r, NMIAs}\)), we estimate which internal actions \(nmi \in NMIAs\) are processed in this interval and how many times $nmi$ are called (\(C_{nmi}\)). For that, we analyze the service call records (see \Cref{adaptive instrumentation}) to determine which services are called in an observation period \(T\) and which parameters are passed. Then we traverse the service's control flows (i.e. their SEFFs) to get NMIAs and predict their RD using the input parameters. This requires evaluating branches and loops of the control flow to decide which branch transition we have to follow and how many times we have to handle the inner control flow of loops.
Our calibration, adjusts the new or outdated branches and loops using the monitoring data (as will be described in Sections \ref{loop} and \ref{branches}) before starting this incremental RDE. Thus, we make sure that we can traverse the SEFFs control flow. Consequently, we can sum up the predicted RDs for all calls of the \ac{NMIAs} and divide the result by \(T\) to estimate the \(U_{r,NMIAs}\) based on the utilization law as shown in the equation 2:
\begin{equation}
U_{r,NMIAs} = \frac{ \sum\limits_{nmi \in NMIAs} \: \sum\limits_{k \leq C_{nmi}} { D_{nmi_{k},r}}}{T}
\end{equation}
Accordingly, we estimate the utilization due to executing the \ac{MIAs} (\(U_{r,MIAs}\)) using the measured \(U_{r}\) and the estimated \(U_{r,NMIAs}\) as shown in equation (3):
\begin{equation}
U_{r,MIAs}= U_{r}-U_{r,NMIAs}
\end{equation}
Hence, we can estimate the utilization \(U_{i,r}\) due to executing each internal action \(i \in MIAs\) using the weighted response time ratios as shown in equation (4), where \(R_{i}\) and \(C_{i}\) are the average response time of \(i\) and its throughput. \(R_{j}\) is the average response time of the internal action \(j \in MIAs\) and \(C_{j}\) is the number of executing it in \(T\).
\begin{equation}
U_{i,r} =U_{r,MIAs} . \frac{R_{i} \cdot C_{i}}
{ \sum\limits_{j \in MIAs} {R_{j}} \cdot C_{j} }
\end{equation}
Using \(U_{i,r}\) we can estimate the resource demand for \(i\) (\(D_{i,r}\)) based on the service demand law presented in equation (1).
In the case that the host has multiple processors, our approach uses the average of the utilizations as \(U_{r}\).
Note that we assume that each internal action is dominated by a single resource. If this is not the case, we have to follow the solution proposed by Brosig et al., to measure processing times of individual execution fragments, so that the measured times of these fragments are dominated by a single resource \cite{Brosig2009Automated}.
To differ between the \ac{CPU} demands and disk demands, we suggest detecting the disk-based services in the first activity of CIPM using specific notation or based on the used libraries.
\vspace{-0.1 cm}
\subsection{Ops-time calibration}
\label{Run-time calibration}
The task of the Ops-time calibration is to update the usage models as well as the deployment model according to the run-time measurements.
To achieve that, we use the iObserve approach, which analyzes the measurements and aggregates them to detect changes concerning the deployments and the user behavior.
For this, we extended our monitoring records so that all information required by iObserve is available. This allows us to integrate the usage model extraction and the identification of deployments from iObserve. We do not need the \ac{RAC} from iObserve, because the mapping information is implicitly presented in our monitoring records, e.g., the records that track the execution of a service include explicitly the ID of the associated service in the architectural model.
\vspace{-0.1 cm}
\section{Parametric dependencies}
\label{parametric dependencies}
This work estimates how \ac{PMPs} depend on input data and their properties (such as number of elements in a list or the size of a file).
We begin by estimating the dependencies of branches and loops, because the incremental \ac{RDE} require traversing the SEFF control flow to estimate the utilization of \ac{NMIAs}.
Currently, we investigate a linear, quadratic, cubic and square root relations using Weka library \cite{Hall2009}. However, we are working on optimizing our estimation using genetic algorithms. The following sections explain the analysis that we follow to estimate the parametric dependencies of \ac{PMPs}.
\vspace{-0.1 cm}
\subsection{Resource demands}
\label{rd}
\begin{comment}
As described in \Cref{ird} we first collect the \ac{RDs} of \(NMIAs\) to estimate their resource utilization (\(U_{NMIAs}\)). To achieve that, we iterate through all service call records and traverse their control flow. If we come across a branch, we evaluate the estimated branch transition using the input parameters to decide which branch transition we have to follow in the next step. Similarly, if we get a loop, we evaluate also its predicted iterations count model by the input parameters to estimate how many times we have to handle its inner control follow. We skip the external calls, because we iterate anyway all service calls in the given observation time \(T\). Every time we get a parametrized internal action, we evaluate its RD model.
Consequentially, we can estimate the resource utilization of them (\(U_{NMIAs}\)) as described in \Cref{ird}.
\end{comment}
To learn the parametric dependency between the resource demand of an internal action \(i\) and input parameters \(P\), we first estimate the resource demand on resource \(r\) for each combination of the input parameters (\(D_{i,r}(P)\)) using the proposed incremental RDE as described in \Cref{ird}.
Second, we adjust the estimated \ac{RDs} of an internal action using the processing rate of the resource, where it is executed, to extract the resource demand independently of the resource' processing rate \(D_{i}(p)\).
Third, if the input parameters include enumerations, we perform additional analysis to test the relation between RDs and enumeration values using decision tree. If a relation is found, we build a data set for each enumeration value. Subsequently, we perform the regression analysis as it will be described in the following paragraph. Otherwise, we create one data set for each internal action that includes the estimated \ac{RDs} and their related numeric parameters. The goal is to find the potential significant relations by the regression analysis of the following equation:
\vspace{-0.15cm}
\begin{gather*}
D_{i}(P) = (a*p_0 + b*p_1 + \dotsb + z*p_n + \\
\qquad \qquad a_1*p_0^2 + b_1*p_1^2 + \dotsb + z_1*p_n^2 +\\
\qquad \qquad a_2*p_0^3 + b_2*p_1^3 + \dotsb + z_2*p_n^3 +\\
\qquad \qquad \qquad \qquad a_3*\sqrt{P_0} + b_3*\sqrt{p_1} + \dotsb +z_3*\sqrt{p_n} + \textbf{C}) \vspace{-0.1cm}
\end{gather*}
\( p_{0}\), \( p_{1}\).. \( p_{n}\) are the numeric input parameters and the numeric attributes of objects that are input parameters.
\(a\)... \(z\), \(a_1\)... \(z_1\), \(a_2\)... \(z_2\) and \(a_3\)... \(z_3\) are the weights of the input parameters and their transformations using quadratic, cubic and square root functions. \(C\) is a constant value.
Fourth, we perform the regression analysis algorithm to find the weights of the significant relations and the constant \(C\).
Fifth, we replace the constant value \(C\) with a stochastic expression that describes the empirical distribution of \(C\) value instead of the mean value delivered by the regression analysis. This step is particularly important when no relations to the input parameters are found. In that case, the distribution function will represent the \ac{RD} of internal action better than a constant value. To achieve that, we iterate on the resulting equation that includes the significant parameters and their weights, to recalculate the value of \(C\) for each \ac{RD} value and their relevant parameters. Then, we build a distribution that represents all measured values and their frequency.
Finally, we build the stochastic expression of \ac{RD} that may include the input parameters and the distribution of \(C\).
\vspace{-0.2 cm}\subsection{Loop iterations count}
\label{loop}
To estimate how the number of loop iterations depends on input parameters, we need both the loop iterations' count and the input parameters for each service call. To achieve that, we use our loop records that log the loop iterations' count, every time a loop finishes (see \Cref{adaptive instrumentation}). These records refer to the service call record that contains the input parameters.
The reason why we use additional records to count loop iterations instead of counting the total amount of enclosing service calls, is that the loop may have a nested branch or loop, which does not allow one to infer the correct count of loop iterations \cite{jaegers}.
To estimate the dependency of the loop iteration count on input parameters, we
combine the monitored loop iterations with the integer input parameter into one data
set. To do so, we filter out all non-integer parameters and take into account their integer properties like number of list elements or size of files. Then, we add transformations (quadratic and cubic) of parameters to the data set to test more relations. Finally, we use regression analysis to estimate the weights of the influencing parameters. Due to the restriction that the loop iterations count is an integer number, we have to ensure, that the output value is always an integer value, which is not always the case. Therefore, we have to approximate the non-integer weights or express them as a distribution of integer values. For instance, we can express the value 1.6 using a Palladio distribution function of an integer variable which takes the value \(1\) in \(40\%\) of all cases and the value \(2\) in \(60\%\) of cases.
Similar to the fifth step of parametrized \ac{RDE} in \Cref{rd}, we replace the constant value of the resulting stochastic expression with an integer distribution function. This will be especially useful, when no relation to the input parameters is found.
\begin{comment}
To estimate the parametrized loop iterations count, we need both the loop iterations' count and the input parameters by each execution call. To achieve that, we use the predefined loop records that log the loop iterations' count, every time a loop finishes (see \Cref{adaptive instrumentation}). These records have a reference of the service call record that contains the input parameters.
The reason why we use additional records to count loop iterations instead of counting the total amount of enclosing service execution, is that the loop may have a nested branch or loop, which does not allow one to infer the correct count of loop iterations \cite{jaegers}.
To estimate the dependency of a loop in relation to input parameters, we
combine the monitored loop iterations with the integer input parameter into one data
set. Therefore, we filter out all non-integer parameters and take into account their integer properties like number of list elements or size of files. Then, we add transformations (quadratic and cubic) of parameters to the data set to test more relations. Finally, we use regression analysis to estimate the weights of the influencing parameters. Due to the restriction that the loop iterations count is an integer number, we have to ensure, that the output value is always an integer value, which is not always the case. Therefore, we have to approximate the non-integer weights or express them as a distribution of integer values. For instance, we can express the 1.6 value using Palladio distribution function of an integer variable which takes the value \(1\) in \(40\%\) of all cases and the value \(2\) in \(60\%\) of cases.
Similar to the fifth step of parametrized \ac{RDE} in \Cref{rd}, we replace the constant value of the resulting stochastic expression with an integer distribution function. This will be especially useful, when no relation to the input parameters is found.
\end{comment}
\vspace{-0.2 cm}
\subsection{Branch transitions}
\label{branches}
To estimate the parameterized branch transitions, we use the predefined branch monitoring records that log which branch transitions are chosen in addition to a reference to the enclosing service call.
We monitor each branch instead of predicting the selected transition according to the external call execution enclosed in the branch due to potential nested control flows (e.g., nested branches), where we cannot infer the selected transition
\cite{jaegers}.
The used monitoring records allow us to build a data set for each branch, which includes the branch transitions and the input parameters.
To estimate the potential relations, we use the J48 decision tree of the Weka library, an implementation of the C4.5 decision tree \cite{Quinlan:1993:CPM:152181}. We filter out the non-significant parameters based on cross-validation. Finally, we transform the resulting tree into a boolean stochastic expressions for each branch transition.
If no relation is found, the resulting stochastic expression will be a boolean distribution representing the probability of selecting a branch transition.
\vspace{-0.2 cm}\subsection{External call arguments}
\label{arguments}
This step predicts the parameters of an external call in relation to the input parameters of the calling service.
For each parameter of an external call, we check whether it is constant, identical to one of input parameters, or depend on some of them. To identify the dependencies, we apply linear regression in the case of numeric parameter and build a decision tree in the case of boolean/ enumeration one. For the remaining types of parameters, we build a discrete distribution.
Because the relation between input parameters and external calls parameters may be more complex, we are working on optimizing our estimation using a genetic search similar to the work of Krogmann et al. \cite{krogmann2009ck}.
\section{Conclusion and future works}
\label{Conclusion}
Applying \ac{AbPP} in the agile and DevOps process promises to detect performance problems by simulation instead of the execution in a production environment.
We presented the continuous integration of the architectural performance model in the development process based on the static and dynamic analysis of the changed code. Our approach keeps the PM continuously up-to-date. We also presented the \ac{MbDevOps} pipeline which automates the incremental calibration process.
Our calibration estimates the parametrized PMPs incrementally and uses a novel incremental resource demand estimation based on adaptive monitoring. Moreover, our calibration updates both the usage and deployment models to support model-based run-time performance management like auto-scaling.
We tested the functionality using two case studies.
The evaluation showed the ability to calibrate the \ac{PM} incrementally. We evaluated the accuracy of our calibration by comparing the simulation results with the monitoring data. In these case studies, the incrementally calibrated models and their parametric dependencies were accurate and the used adaptive monitoring significantly reduced the monitoring overhead and the calibration effort.
In this work, we have automated only the part of the proposed \ac{MbDevOps} pipeline that is responsible for testing, incremental calibration and self-validation.
In future work, we aim to automate the whole pipeline to integrate the \ac{CI} process with \vitruv.
Moreover, we will extend the available approaches to integrate existing source code and models into \vitruv platform \cite{leonhardt2015a, Mazkatli2018MiSE}.
Besides, we will extend our parametrized calibration to test for more complex dependencies of external call parameters. It is also planned to extend the monitoring records to update the system model (composition of the components) automatically.
\section{Evaluation}
\label{evaluation}
This section introduces the evaluation goals, setup, environment and the results of the evaluation:
\vspace{-0.3 cm}
\subsection{Evaluation Goals}
The evaluation considers three aspects of our approach:
\begin{itemize}[leftmargin=*]
\item Accuracy of the incremental calibration: We calculate the accuracy of the calibrated models by using the monitoring data as reference and also consider an evolution scenario to include multiple calibration iterations. In particular, we analyze whether the parametric dependencies are recognized by the calibration pipeline.
\item The time the calibration pipeline needs to derive changes and apply them to the model.
\item Quantification of the overhead introduced by the monitoring. This allows us to make statements on whether the approach can be applied in practice.
\end{itemize}
\vspace{-0.4 cm}
\subsection{Experiment Setup}
\label{evaluation:setup}
The general procedure for evaluating our approach is structured as follows:\\
1. We monitor the application over a defined period of
time (e.g., 60 minutes) and perform a load test at the same time to artificially simulate user interactions. \\
2. We run the monitoring in combination with the load test two times. First, we apply a fine-grained monitoring, whose results are passed to the
calibration pipeline for adjusting the architecture model (training set). Second, we only observe service calls (coarse-grained monitoring). These data is used to estimate the accuracy of the calibrated model (validation set). This procedure prevents the monitoring overhead from falsifying the evaluation results. During the calibration we also recorded the execution times of the different parts in the pipeline and the overhead which is caused by the monitoring. \\
3. We adjust the usage model according to the validation set and perform a fixed number of simulations of the updated architecture
model (100 repetitions). A single execution would not be sufficient here, since the simulation samples from stochastic distributions and one result may not be representative. Afterwards we compare the results of the
simulations with the validation set. Both are distributions,
therefore we use the \ac{KS-Test} \cite{KSTest}, the Wasserstein metric \cite{majewski_et_al:LIPIcs:2018:9327} and conventional statistical measures, such as the quartiles, to determine the
similarity. The lower the \ac{KS-Test}, the higher is the accuracy of the updated architecture model. We are aware that the KS test is not the optimal choice in our specific case, which is why we also used alternative metrics such as the Wasserstein distance to confirm the results. The Wasserstein metric, on the other hand, is rather a distance measure that quantifies the effort needed to transfer one distributions into the other. If we want to compare the accuracy of two simulation distributions (using the monitoring data as reference), this metric is very well suited, but as an absolute number it is difficult to interpret. Therefore, we just used this metric as an additional measure to confirm the results.
\begin{comment}
The self-validation executes this step automatically, but only one simulation is triggered. Therefore, we show more reliable results (the results of 100 simulations using an additional validation set) supplementary to the self-validation results.
\end{comment}
\input{externals/seff_booksale.tex}
\vspace{-0.3 cm}
\subsection{Evaluation Environments}
We evaluated our approach using two case studies: \ac{CoCoME}~\cite{cocome} and
TeaStore \cite{teastore}. %
\ac{CoCoME} is a trading system which is designed for the use in supermarkets. It supports several processes like scanning products at a cash desk or processing sales using a credit card. We used a cloud-based implementation of \ac{CoCoME} where the enterprise server and
the database run in the cloud.
\begin{comment}
TeaStore is a micro-service based reference application
\cite{teastore}. It implements a web shop where you can buy
different kinds of tea. Besides, the TeaStore supports user
authentication and automatic recommendations. Amongst others, it is designed to
evaluate approaches in the field of performance modelling \cite{teastore}.
\end{comment}
Additionally, we evaluated our approach using the aforementioned running example of TeaStore case study (cf. \Cref{Running Example}).
Because of the fact that \ac{CoCoME} and TeaStore are not implemented using the \vitruv platform, we instrument the changed parts of the source code manually. In future work, we will integrate earlier approaches \cite{leonhardt2015a} into the current \vitruv platform to import existing source code \cite{leonhardt2015a} into \vitruv VSUM, which allows using the automatic change-based adaptive instrumentation.
\subsection{Model Accuracy Evaluation}
\label{evaluation:accuracy}
First, with the help of CoCoME, we validated the basic functionality and performed initial measurements to assess the accuracy of the adapted models. In the second step, we used the TeaStore case study to evaluate the incremental calibration and ensure that parametric dependencies were considered in the generated models.
\subsubsection{CoCoME case study}
In the evaluation scenario using \ac{CoCoME}, we suppose that the ``bookSale'' service is newly added. According to our assumption, we instrument this service and all services
that are subsequently called by it.
The ``bookSale'' service is responsible for processing a sale after a user submitted his payment. As input, this service receives a list of the purchased products including all information related to the purchase. In the next step, we performed a load test which simulates user purchases. We used the resulting monitoring data to calibrate our initial architecture model.
The service consists of several internal and external actions and two loops. We also monitored and calibrated the behavior of all external actions to make the resulting model more accurate.
Figure \ref{fig:seff:booksale} visualizes the structure of the calibrated "bookSale" service
of \ac{CoCoME}. For reasons of simplification, we focused on the structure and omitted certain details.
As described in the experiment setup, we performed 100 simulations and
compared the distribution of the monitoring with the results of the simulations.
Table \ref{table:cdf} shows the quartiles
for both the monitoring and the simulation results for a single example execution
\input{externals/cdf_cocome_table.tex}
In addition, a KS test value of \textbf{0.1321} and a Wasserstein distance of \textbf{13.7726} were obtained by comparing the two distributions. It is well visible that the two distributions are very close to each for the first, second and third quartile. The discrepancy in the minimum and maximum values of the distributions arises because we limited the time spent for simulation.
The average of the KS tests over all 100 simulation runs
is approximately \textbf{0.1467}, the minimum \textbf{0.1248} and the maximum
\textbf{0.1774}. From these values it follows that our approach provided a model which represents the monitored behavior in our scenario well.
\subsubsection{TeaStore case study}
For the TeaStore case study, we examined the "Recommender" part (cf. \Cref{Running Example}).
\begin{comment}
The Recommender is divided into a service that calculates the
recommendations for a certain shopping cart (recommend service) and a service
that derives information from the previous orders (train service).
\end{comment}
We monitored the "train" service and calibrated it. Furthermore, we modified the application so that the "train" service is executed when a user placed an order. This causes the number of users to have a direct impact on the service response times, as the number of orders in the database grows faster and therefore a higher number of orders have to be processed within the service.
\begin{comment}
Normally the "train" service only receives the past orders and the associated ordered items as input. We
extended it and added a parameter (enum) indicating which Recommender
implementation is used. This allows us to specify a single \ac{SEFF} that is valid for all Recommender implementations.
\end{comment}
Figure \ref{fig:seff:train} shows the simplified layout of the \ac{SEFF}. Except for the last Internal Action, the behavior is exactly the same for all Recommender types.
We also considered an evolution scenario for the TeaStore to evaluate the incremental calibration. This means that the calibration of the models is executed several times. In our considered evolution step, new Recommender implementations are added or existing ones were removed.
In the following, we compared the accuracies of a parameterized and a non-parameterized model.
In the second step, we analyzed the behavior and accuracy of the incrementally calibrated models when executing evolutionary scenarios (e.g., Recommender implementations are added or removed).
In order to confirm the recognition of parametric dependencies, we first monitor the service as previously described in section \ref{evaluation:setup} fine granular (F1) and subsequently coarse granular (C1). We \textbf{only} use the resulting fine granular monitoring data (F1) to calibrate two models: one with parametric dependencies using our incremental calibration and one without parametric dependencies using the distributions of the observed values.
The results confirmed that our calibration was able to detect the parametric dependencies, e.g. the RD of the internal action \textit{trainForRecommender} (the \nth{4} action in train SEFF) is related to the type of recommendation implementation as well as to the number of ordered elements.
To compare the accuracy of the parameterized model with the non-parameterized model, we predict the performance for a what-if scenario, i.e. what the performance would be if the load is changed from 20 concurrent users to 40 concurrent users. To have a baseline, we measure the performance by applying the load of 40 users (performing coarse granular monitoring (C2)). Then we compare the simulation results with the monitoring data, i.e. (C1) as reference for the first load (20 users) and (C2) for the second load.
Table \ref{table:train:kss} shows the comparison of the KS tests and the Wasserstein distances between the parameterized and the non-parameterized model. It can be seen that both models are very accurate for the standard load of 20 users. However, it can be stated that the non-parametric model loses significantly more accuracy than the parameterized model, when the load changes to 40 users.
\input{externals/comparisons/teastore_kstest_table.tex}
\begin{comment}
We confirmed this first impression by using Wasserstein distance as an alternative metric. We included the results in the Appendix in Figure \ref{fig:comp:ws:train} and we also summarized the results there in tabular form (see table \ref{table:teastore:param:20} and table \ref{table:teastore:param:40}).
\end{comment}
In order to evaluate the incremental calibration, we simulated an evolution scenario for the TeaStore application. At the beginning of the scenario, the TeaStore contains a single implementation of the "train" service. Then we perform 15 minutes of fine-granular monitoring and use the resulting data to calibrate the model. After that we monitor the service for another 15 minutes coarse granular. The coarse granular monitoring data is used as reference for the estimation of the accuracy of the calibrated model. In the next step, we add another implementation of the "train" service and again, just like before, we perform a fine granular and a coarse granular monitoring for 15 minutes each. If all implementations of the "train" service have already been added, one of them will be randomly removed again. We repeat this procedure over several iterations and calculate the accuracy of the calibrated model for each iteration. In addition, the database fills up more and more over time, which leads to increasing response times. This adds an extra level of difficulty to the calibration process.
\input{externals/evolution/evolution_table.tex}
We performed 10 consecutive iterations and calculated the accuracy for each of them. As before, we calculated the KS test and confirmed the results using Wasserstein distance.
Table \ref{table:evol} lists the aggregated results. We omitted the values of some iterations because these were always very similar to the previous ones.
The values are slightly higher than before where we did not apply any changes to the application. However, it can be seen that the accuracy of the models is almost constant throughout the evolutionary steps. From this, it can be concluded that the automated calibration can handle changes of the observed services in the TeaStore.
\vspace{-0.3 cm}
\subsection{Calibration Performance}
For both \ac{CoCoME} and TeaStore we computed how much time the individual parts of the pipeline took to complete. For the TeaStore, we also distinguished between the two states within the evolution scenario (before and after the first evolution step). The results are shown in table \ref{table:exec}.
\begin{table}[h]
\centering
\caption{Execution times of the pipeline parts}
\begin{tabular}{|c|c|c|c|}
\hline
Measure & CoCoME & TeaStore (1) & TeaStore (2) \\ \hline \hline
Record Count & 220706 & 294237 & 27085 \\ \hline \hline
Load Records & 4.664s & 5.142s & 0.774s \\ \hline
RD Calibration & 9.168s & 4.996s & 0.789s \\ \hline
UM Adjustment & 2.365s & 0.194s & 0.165s \\ \hline
Self-Validation & 1.729s & 1.440s & 1.327s \\ \hline \hline
\textbf{Total} & 17.213s & 10.826s & 2.199s \\ \hline
\end{tabular}
\label{table:exec}
\end{table}
We can see that the required time strongly depends on the number of monitoring records and the complexity of the observed service. It has already been shown for iObserve that the derivation of the usage model scales linearly with the number of monitoring records \cite{heinrich2015runtime}. In all considered cases, the execution of the whole pipeline took less than 20 seconds.
It can also be observed that the adaptive monitoring limits the number of monitoring records and accordingly the second iteration of the TeaStore monitoring produces far less records, which results in significantly lower execution times.
To gain detailed insights about the performance of the pipeline, we also examined the behavior for an increasing number of monitoring records. For this purpose, we generated monitoring data with the help of \ac{CoCoME} case study and executed the pipeline several times. The results showed that the execution time of the pipeline scales linearly up with the number of monitoring records, plus an additional constant overhead (see Table \ref{table:perf}).
\input{externals/performance_details_table.tex}
\subsection{Monitoring Overhead}
Kieker only introduces a small overhead, the recording and storing of one record takes on average less than 5 microseconds \cite{kieker}. However, due to the fine-grained instrumentation, a single service call can cause a lot of monitoring records.
We also measured the overhead caused by the monitoring. For CoCoME's "bookSale" service the average monitoring overhead was \textbf{1.31ms} (fine-grained), \textbf{252\boldmath$\mu$s} (coarse-grained) and for TeaStore's Recommender.train service the measured overhead on average amounts to \textbf{1.81ms} (fine-grained), \textbf{88\boldmath$\mu$s} (coarse-grained). We can notice that the coarse-grained monitoring has almost no influence on the execution times. The fine-grained monitoring overhead scales up with the complexity of the service. In our case, the overhead of around 1ms had no drastic impact on the performance, since the execution times of the services are significantly higher.
The evolution scenario described for the TeaStore shows very well that the adaptive monitoring helps to significantly reduce the monitoring overhead. In the first iteration we instrument the statements of all SEFF elements shown in Figure \ref{fig:seff:train} - (1), (2), (3) and (4). For each additional iteration we just instrument the changed internal action (4). In this case we saved \textbf{at least 75\%} of the monitoring overhead. This is reflected in the times required for calibration, which are considerably lower compared to the first iteration (see \autoref{table:exec}).
In general there is a trade off between the monitoring overhead and the accuracy of the model, but our approach tries to balance them by using the incremental calibration in combination with the adaptive monitoring. Consequently, the calibration time and the monitoring overhead can be kept as small as possible.
\section{Foundations}
\label{foundation}
\subsection{Palladio}
\label{palladio}
Palladio is an approach to model and simulate architecture-level \acp{PM} and has been used in various industrial projects~\cite{reussner2016a,becker2008a}. Within Palladio, the so-called \ac{PCM} defines a language for describing \acp{PM}: the static structure of the software (e.g. components and interfaces), the behavior, the required resource environment, the allocation of software components and the usage profile. The \ac{PCM} \ac{SEFF} \cite{becker2008a} describes the behavior of a component service on an abstract level using different control flow elements: \textit{internal actions} (a combination of internal computations that do not include calls to required services), \textit{external call actions} (calls to required services), \textit{loops} and \textit{branch actions}. SEFF loops and branch actions include at least one external call. Other loops and branches in the service implementation are ignored and combined into the internal actions to increase the level of abstraction.
To predict the performance measures (response times, \ac{CPU} utilization and throughput) the architects have to enrich the SEFFs with \ac{PMPs}.
Examples of \ac{PMPs} are resource demands (processing amount that internal action requests from a certain active resource, such as a \ac{CPU} or hard disk), the probability of selecting a branch, the number of loop iterations and the arguments of external calls.
Palladio uses the \ac{StoEx} language \cite{koziolek_modeling_2016} to define \ac{PMPs} as expressions that contain random variables or empirical distributions.
\ac{StoEx} allows parameter characterization (e.g. determining NUMBER\_OF\_ELEMENTS, VALUE, BYTESIZE and TYPE) and to express calculations and comparisons (e.g., file.BYTESIZE$<=$5*max.VALUE).
\input{externals/seff_train.tex}
\subsection{Co-evolution approach with \textsc{\textit{Vitruvius}}\xspace}
\label{vitruv}
\label{co-evaluation approach}
The co-evolution approach of Langhammer et al. \cite{langhammer2015a,langhammer2017a} helps software developers and architects to keep the architecture model and the code consistent when their software system evolves. It defines change-driven consistency preservation rules that propagate changes in source code to the architecture model and vice versa using model-based transformations.
These rules are defined based on \vitruv \cite{kramer2013b,burger2014diss}, a view-based framework that encapsulates the heterogeneous models of a system and the semantic relationships between them in a \ac{VSUM} in order to keep them consistent.
\vitruv defines \emph{mappings} and \emph{reactions} languages that describe consistency rules on the metamodel-level. These rules describe the consistency repair logic for each kind of changes, i.e., which and how the artifacts of a metamodel have to be changed to restore the consistency after a change in a related metamodel has occurred.
Using \vitruv, Langhammer et al. implemented their approach to keep Java source code (using an intermediate model \cite{heidenreich2010a}) and a \ac{PCM} model consistent. The defined consistency rules update the structure of a \ac{PCM} model (i.e. components and interfaces) and its behavior (in term of \acsp{SEFF}, but \textit{without} \ac{PMPs}) as a reaction to changes in the source code. Similarly, changes in PCM model are propagated to the Java source code.
\subsection{Kieker}
\label{kieker}
Kieker \cite{kieker} is an APM tool that captures and analyzes execution traces from distributed software systems. It allows one to describe specific probes (data structures of the monitored information) using the \ac{IRL} \cite{jung2013model}. It also supports dynamic and adaptive monitoring by activating or deactivating probes during run time.
\subsection{iObserve}
\label{iObserve}
iObserve \cite{heinrich2017a} aims to increase the human comprehensibility of run-time changes by updating an architecture model accordingly. The authors defined specific monitoring records using \ac{IRL} and map the resulting measurements to the corresponding parts in the \ac{PCM} using the so-called \ac{RAC} \cite{heinrich2014}. iObserve detects changes concerning the deployment and the user behavior and updates the related parts of the \ac{PCM} to analyze performance and privacy aspects \cite{heinrich2016b}. However, iObserve does not support updating component behavior models (SEFFs) with \ac{PMPs}.
\section{Introduction}
\label{Introduction}
In many application domains, software is nowadays developed iteratively and incrementally, e.g., following agile practices.
This means that there usually is no dedicated architecture design phase, but architectural design decisions are made continuously throughout the development. For performance analysis, this means that there is not a single point in time at which the performance analysis is carried out, but a performance analysis is required whenever performance-critical architecture design decisions have to be made.
In such iterative and incremental development projects, developers typically use \ac{APM} \cite{apm,BeEiFeGrRhJaShHoViWaWi2019-ICPE-DevOpsSurvey} to assess the current performance of their software. However, \ac{APM} cannot predict the impact of design decisions or to answer scalability and sizing questions in a large and distributed environment, where the resources may not be accessible and a huge amount of load generators would be needed.
Software performance engineering \cite{SPEwoodside,conni} uses models to assess the performance of a software system. It supports the evaluation of various architecture, design, implementation and workload choices. Compared to \ac{APM}, such model-based performance prediction \cite{mbpp} allows to predict future alternatives before implementing them \cite{conni}.
In particular, \ac{AbPP} approaches~\cite{reussner2016a} reflect the architecture of the software system in a \ac{PM} in order to easily study design alternatives~\cite{martens2010c}.
The problem of \ac{AbPP} in iterative-incremental development is that modelling is a time-consuming process. In particular, calibrating \ac{PMPs} (e.g., \ac{RDs}, branch probabilities, and loop counts) is challenging, because the \ac{PMPs} may depend on impacting factors such as input data, properties of required resources or usage profile. Ignoring these so-called \emph{parametric dependencies}~\cite{becker2008a}) will lead to inaccurate performance predictions. Thus, keeping the \ac{PM} and in particular the parametrized \ac{PMPs} consistent with the iteratively and incrementally evolving source code over time requires repeated manual effort.
To address the high effort to create performance models when source code is available, researchers have suggested several approaches to extract \acp{PM} automatically.
However, most of these approaches \cite{BrHuKo2011-ASE-AutomExtraction,WaStKoKo2017-QUDOS-PMXBuilder, Brunnert13,langhammer2016automated,spinner2016reference} have three shortcomings: First, they require instrumentation and execution of the whole system under study to extract the \ac{PM}, which causes a high overhead and is not feasible at high frequency, e.g., after each source code commit. Second, they do not consider parametric dependencies.
Notable exceptions are the approach by Krogmann et al.~\cite{krogmann2009ck,krogmann2012ashort} or Grohmann et al.~\cite{GrEiElMaKiKo2019-MASCOTS-DependencyIdentification}, which also extract parametric dependencies (but monitor the whole system to calibrate the model without keeping the manual changes).
Third, they reconstruct the whole architectural PM from scratch. Thus, they cannot keep manual changes of the architecture model, which is problematic if the architecture model shall be used as an architecture knowledge base and is enhanced manually with e.g. the rationale of architecture decisions~\cite{kramer2012b}.
In this paper, we present our approach, called \ac{CIPM}, to incrementally extract and calibrate architecture-level \acp{PM} with parametric dependencies after each source code commit.
Our approach builds upon the Palladio approach~\cite{reussner2016a} for modelling and simulating architecture-level \acp{PM}. We have extended an incremental extraction of such architectural \acp{PM}~\cite{langhammer2017a} with incremental calibration that considers the parametric dependencies and is based on adaptive monitoring. The goal of the approach is to keep the \ac{PM} up-to-date automatically to allow \ac{AbPP}. An initial version of this idea has been presented in a workshop publication~\cite{Mazkatli2018Qudos} without evaluation.
The contributions of our paper are twofold:
\begin{enumerate}
\renewcommand{\labelenumi}{(\Alph{enumi})}
\item \textit{Incremental Dev-time calibration:} We propose a novel incremental calibration at \ac{Dev-time} that responds to source code changes by adaptive instrumentation of the changed parts of the code and uses the resulting measurements from performance tests or the production system to estimate the \ac{PMPs} incrementally.
For this purpose, we propose a novel incremental \ac{RDE} that is based on adaptive monitoring.
Our calibration uses statistical analysis to learn potential dependencies, e.g., regression analysis for resource demands and decision trees for the estimation of branch transitions.
\item \textit{Model-based DevOps pipeline:}
We extend agile DevOps practices \cite{brunnert2015a} to integrate the \ac{CIPM} activities and thus to reduce the effort for \ac{AbPP}.
To do so, we propose a model-based DevOps pipeline and implement a part of it.
In addition to the above-mentioned incremental \ac{Dev-time} calibration, we also include an existing approach for calibration at \ac{Ops-time} in the pipeline. This \ac{Ops-time} calibration responds to changes in deployment and usage profile and updates the respective parts of the \ac{PM}. We include also prototypical self-validation steps.
\end{enumerate}
We evaluate our contributions using two case studies \cite{cocome,teastore}.
The evaluation confirmed that the incremental calibration was able to detect parametric dependencies while significantly reducing the monitoring overhead. Moreover, we showed that the calibrated performance models are accurate by comparing the simulation results with monitored data.
The next section gives an overview of the foundations. \Cref{Running Example} introduces a motivating example. The overall CIPM process is presented in \Cref{CIPM}. \Cref{MbDevOps} describes how to embed CIPM in the model-based DevOps pipeline. The next three sections provide detail on the CIPM activities: \Cref{adaptive instrumentation} presents the adaptive instrumentation, \Cref{incremental calibration} describes the incremental calibration and \Cref{parametric dependencies} describes how we estimate the parametric dependencies. Our evaluations are discussed in \Cref{evaluation}. The related work follows in \Cref{related works}. The paper ends with our conclusions and future work
(\Cref{Conclusion}).
\section{Related Work}
\label{related works}
A number of approaches for constructing the architectural model based on static (e.g., \cite{becker2010reverse}), dynamic, or hybrid analysis exist.
Walter et al. \cite{WaStKoKo2017-QUDOS-PMXBuilder} propose a tool to extract an architectural PM as well as performance annotation based on analysing monitoring traces.
Similarly, other works \cite{Brunnert13,spinner2016reference,BrHuKo2011-ASE-AutomExtraction} extract PM based on dynamic analysis. Krogmann et al. \cite{krogmann2009ck,krogmann2012ashort} extract parametrized PCM based on hybrid analysis.
Langhammer et al. introduce two reverse engineering tools \cite [P.~140]{langhammer2017a} \cite{langhammer2016automated} that extract the behavior of the underlying source code based on hybrid analysis.
The above-mentioned approaches do not support the iterative extraction of the architectural PMs or the incremental calibration. If they were used in an iterative development, they would re-extract and recalibrate the whole models after each iteration. This would cause a high monitoring overhead. Moreover, this process ignores the potential manual modifications of the extracted PM that may be formerly applied.
In addition to the work of Krogmann et al. \cite{krogmann2009ck,krogmann2012ashort}, the works of Ackermann et al.~\cite{AcGrEiKo2018-DependencyModeling} and Curtois et al.~\cite{Courtois2000} also consider the characterization of parametric dependencies in performance models, while Grohmann et al.~\cite{GrEiElMaKiKo2019-MASCOTS-DependencyIdentification} focus on the identification of those from monitoring data. However, our work considers the parametric dependencies during the incremental calibration of the architectural PMs.
Model extraction approaches derive the resource demands either based on coarse-grained monitoring data \cite{spinner2015evaluating,LibReDE} or fine-grained data
\cite{Brosig2009Automated,willnecker2015comparing}. The latter approaches give a higher accuracy by estimating \ac{PMPs} but have a downside effect because of the overhead of instrumentation and the monitoring. Our approach reduces the overhead by the automatic adaptive instrumentation and monitoring.
Spinner et al. \cite{SpGrEiKo2019-JSS-ModelLearning} propose an agent-based approach to updated architectural performance models. In contrast to our work, they focus on model updates at run-time.
\ac{DPE} \cite{WaEiGrOkKo2018-ICPE-Tools-for-DPE-Tutorial} technically enables to answer concerns based on measurement-based performance evaluation and model-based performance predictions. However, existing solutions do not provide a technical integration into the build pipeline. Also, \ac{DPE} does not answer how to keep PMs updated.
\section{Evaluation}
\label{evaluation}
This section introduces the evaluation goals, setup, environment and the results of the evaluation:
\vspace{-0.3 cm}
\subsection{Evaluation Goals}
\label{goals}
The goal is to evaluate the following three aspects of the approach and answer the Evaluation-Questions (EQs):
\begin{itemize}[leftmargin=*]
\item \textbf{Accuracy of the incremental calibration}: \\
\textbf{EQ1.1:} How accurate are the incremental calibrated PMs?
\\
\textbf{EQ1.2:} Is the accuracy of the incremental calibrated PM stable over the incremental evolution?
\item \textbf{Accuracy of the parametric dependencies}: \\
\textbf{EQ2:} Does the estimation of parametric dependencies improve the accuracy of PMs?
\item \textbf{Performance of the calibration pipeline:} \\ \textbf{EQ3:} How long does the pipeline need to update PMs?
\item \textbf{Monitoring overhead:} \\\textbf{EQ4.1:} How much is the required monitoring overhead?
\\\textbf{EQ4.2:} How much can the adaptive instrumentation reduce the monitoring overhead?
\end{itemize}
\vspace{-0.2 cm}
\subsection{Experiment Setup}
\label{evaluation:setup}
The evaluation is structured as follows:\\
1. We monitor the application over a defined period of
time (e.g., 60 minutes) and perform a load test at the same time to artificially simulate user interactions. \\
2. We run the monitoring in combination with the load test two times. First, we apply a fine-grained monitoring, pass the results to the
calibration pipeline for adjusting the architecture model (training set). Second, we only observe service calls (coarse-grained monitoring). These data are used to estimate the accuracy of the calibrated model (validation set). This procedure prevents the monitoring overhead from falsifying the evaluation results. We also record the execution times (overhead) of the pipeline parts
3. We adjust the usage model according to the validation set and perform a fixed number of simulations of the updated architecture
model (100 repetitions). Repetitions are necessary, as the simulation samples from stochastic distributions may not be representative. Afterward, we compare the results of the
simulations with the validation set. Both are distributions,
therefore we use the \ac{KS-Test} \cite{KSTest}, the Wasserstein metric \cite{majewski_et_al:LIPIcs:2018:9327}, and conventional statistical measures to determine the
similarity.
The lower the \ac{KS-Test} value, the higher is the accuracy of the updated architecture model. In order to limit the dependence on a single metric, we also included the Wasserstein distance to confirm the results. The Wasserstein metric is distance measure that quantifies the effort needed to transfer one distribution into the other. If we want to compare the accuracy of two simulation distributions (using real monitoring as reference), this metric is well suited, but reports absolute numbers that are difficult to interpret. We therefore combined both metrics in our analysis.
\vspace{-0.3 cm}
\subsection{Evaluation Environments}
We evaluated our approach using two case studies: \ac{CoCoME}~\cite{cocome} and
TeaStore \cite{teastore}.
\ac{CoCoME} is a trading system which is designed for the use in supermarkets. It supports several processes like scanning products at a cash desk or processing sales using a credit card. We used a cloud-based implementation of \ac{CoCoME} where the enterprise server and
the database run in the cloud.
\begin{comment}
TeaStore is a micro-service based reference application
\cite{teastore}. It implements a web shop where you can buy
different kinds of tea. Besides, the TeaStore supports user
authentication and automatic recommendations. Amongst others, it is designed to
evaluate approaches in the field of performance modelling \cite{teastore}.
\end{comment}
Additionally, we evaluated our approach using the aforementioned running example of the TeaStore case study (cf. \Cref{Running Example}). Compared to the original, we modified the application so that the "train" service is executed when a user places an order. This causes the number of users to have a direct impact on the service response times.
Because \ac{CoCoME} and TeaStore are not implemented using the \vitruv platform, we instrument the changed parts of the source code manually. In future work, we will use approaches for importing existing source code into \vitruv \cite{leonhardt2015a}, which allows the automatic adaptive instrumentation.
\vspace{-0.2cm}
\subsection{Evaluation of the Incremental Calibration }
\label{Evaluation incremental Calibration}
This section evaluates the accuracy of our calibration for the following incremental evolution scenarios.
\subsubsection{BookSale Calibration}
The ``bookSale'' service of CoCoMe case study is responsible for processing a sale after a user submitted his payment. As input, this service receives a list of the purchased products including all information related to the purchase. The service consists of several internal and external actions and two loops.
Figure \ref{fig:seff:booksale} visualizes the structure of the calibrated "bookSale" service
of \ac{CoCoME}. For reasons of simplification, we focused on the structure and omitted certain details.
In this evaluation scenario, we supposed that the ``bookSale'' service is newly added.
According to our assumption, we instrumented this service and all services that are subsequently called by it.
We performed a load test which simulates user purchases. We used the resulting monitoring data to calibrate our architecture model.
\paragraph{Results} Table \ref{table:cdf} shows the quartiles for both the monitoring and the simulation results for a single example execution. Besides, a KS test value of \textbf{0.1321} and a Wasserstein distance of \textbf{13.7726} were obtained by comparing the monitoring and the simulation distributions. It is well visible that these distributions are very close to each for the first, second and third quartile. The discrepancy in the minimum and maximum values of the distributions arises because we limited the time spent for simulation. The KS tests average over all 100 simulation runs is approximately \textbf{0.1467} (min \textbf{0.1248}, max \textbf{0.1774}). These results confirm the PM accuracy (EQ1.1).
\vspace{-0.1cm}
\input{externals/cdf_cocome_table.tex}
\subsubsection{Train Calibration}
\label{Train Calibration}
\begin{comment}
Similar to \textit{"BookSale"} evaluation scenario, we assume that the first version of 'train' service of "Recommender" (described in \Cref{Running Example}) is newly added.
Therefore we instrument the statements of all SEFF elements shown in Fig. \ref{fig:seff:train} - (1), (2), (3) and (4). Then we monitored and calibrated them.
Due to lack of space and the similarity to the results of the parametric dependency scenario (Table \ref{table:train:kss}, \nth{2} column), we show the accuracy results of train in Sec. \ref{whatIf}\footnote{See \url{https://sdqweb.ipd.kit.edu/wiki/CIPM} for more detail on evaluation.}.
Similar to \textit{"BookSale"} evaluation scenario, we assume that the first version of 'train' service of "Recommender" (cf. \Cref{Running Example}) was newly added.
Therefore we instrument the statements of all SEFF elements shown in Fig. \ref{fig:seff:train} - (1), (2), (3) and (4). Then we monitored and calibrated them.
The resulting Ks Test value (0.18) and Wasserstein value (29.9) confirmed the accuracy of train, which is evaluated again in another scenario in Sec. \ref{whatIf} (see Table \ref{table:train:kss}, \nth{2} column).
\end{comment}
Similar to ”BookSale” evaluation scenario, we assumed that ’train’ service (cf. \Cref{Running Example}) was newly added. Therefore, we instrumented and monitored the statements of all SEFF elements shown in Fig. \ref{fig:seff:train} in two cases: (1) for the first implementation of ’train’ and (2) after adding an enumeration parameter that refers to the used implementation of 'train' to represent the four implementations in one SEFF (cf. Sec. \ref{whatIf}). Then we calibrated the SEFFs of these two cases. Due to lack of space we show only the accuracy results of the second case in (Table \ref{table:train:kss}, \nth{2} column)\footnote{See \url{https://sdqweb.ipd.kit.edu/wiki/CIPM} for more detail on evaluation.}.
\subsubsection{TrainForRecommender Calibration}
\label{TrainForRecommender}
To evaluate that the accuracy of PM is stable (EQ1.2), we simulated an evolution scenario that assumes that the developers change the implementation of the \textit{'train'} service to examine the four implementation alternatives (cf. \Cref{Running Example}) in four subsequent source code commits $c1$, ..., $c4$. According to our assumption, except of the "trainForRecommender" internal action, the behavior of \textit{'train'} (\Cref{fig:seff:train}) stays the same.
To evaluate this scenario, we instrumented only the "trainForRecommender" internal action for the first implementation of \textit{'train'} ($c1$).
We performed 15 minutes of fine-granular monitoring and used the resulting data to calibrate the model. Next, we monitored the 'train' service for another 15 minutes of coarse granular. The coarse granular monitoring data was used as reference for the estimation of the accuracy of PM. In the next step, we instrumented the "trainForRecommender" internal action for the second implementation of \textit{'train'} ($c2$) and performed a fine granular and a coarse granular monitoring for 15 minutes each. We repeated the same procedure for commits $c3$ and $c4$.
To simulate additional commits, we changed the train implementation by randomly replacing the strategy with another one. We repeated this procedure six more times to simulate six additional iterations and calculate the accuracy of the calibrated model for each iteration. In addition, the database filled up over time, which leads to increasing response times, increasing the difficulty for the calibration process
\vspace{-0.15 cm}
\paragraph{Results}
\input{externals/evolution/evolution_table.tex}
Table \ref{table:evol} lists the aggregated results.
The values are slightly higher than without applying changes to the application. However, it can be seen that the accuracy of the models is almost constant throughout the evolutionary steps. From this, we conclude that the automated calibration is able to handle changes of the observed services.
\vspace{-0.25cm}
\subsection{Evaluation of the Estimation of Parametric Dependencies}
\label{whatIf}
This scenario evaluates the recognition of the parametric dependencies and compares the accuracy of the AbPP using the parameterized model (i.e., calibrated with our approach) with the non-parameterized model (i.e., calibrated ignoring parametric dependencies). For this goal, we extended the SEFF shown in \Cref{fig:seff:train} and added an enumeration input parameter that refers to the recommender type and determines which implementation is used within the "trainForRecommender" internal action. This allowed us to represent all different implementations in one SEFF. This means the RD of "trainForRecommender" can be related to the recommender type. Then, we calibrated the SEFF and predicted the performance for increasing the load from 20 to 40 concurrent users.
First, we monitored the service fine granular (F1) and subsequently coarse granular (C1) by applying the load of 20 users. We \textbf{only} used the resulting monitoring data (F1) to calibrate a parameterized model using our approach and the non-parameterized model using the distributions of the observed values. Then, we compared the simulation results of both models with the monitoring data (C1). Then, we performed AbPP for the higher load (40 users). As baseline, we measured the performance using coarse granular monitoring and applying the load of 40 users (C2). Finally, we compared the AbPP results of both models with the monitoring data (C2).
\paragraph{Results}
The results confirmed that our calibration was able to detect the parametric dependencies. Hence, the RD of \textit{trainForRecommender} is related to the type of recommendation implementation and the number of elements.
Table~\ref{table:train:kss} shows the comparison of the KS tests and the Wasserstein distances between the parameterized and the non-parameterized model. It can be seen that both models are very accurate for the standard load of 20 users. However, the non-parametric model loses significantly more accuracy than the parameterized model, when the load changes to 40 users. This confirms that the estimation of the parametric dependencies improves the accuracy of PMs (EQ1.1, EQ2).
\vspace{-0.15 cm}
\input{externals/comparisons/teastore_kstest_table.tex}
\vspace{-0.2 cm}
\subsection{Evaluation of the MbDevOps Pipeline}
This section evaluates the performance of the implemented part of MbDevOps Pipeline (monitoring, incremental calibration and self-validation) and answers EQ3.
For that, we computed how long the individual pipeline parts took to complete. We used the three evolution scenarios described in \Cref{Evaluation incremental Calibration}. The results are shown in Table \ref{table:exec}.
\vspace{-0.15 cm}\begin{table}[h]
\centering
\caption{Execution times of the pipeline parts}
\begin{tabular}{|c|c|c|c|}
\hline
Measure & bookSale & train & trainForRecommender \\ \hline \hline
Record Count & 220706 & 294237 & 27085 \\ \hline \hline
Load Records & 4.664s & 5.142s & 0.774s \\ \hline
PM Calibration & 9.168s & 4.996s & 0.789s \\ \hline
UM Adjustment & 2.365s & 0.194s & 0.165s \\ \hline
Self-Validation & 1.729s & 1.440s & 1.327s \\ \hline \hline
\textbf{Total} & 17.926s & 11.772s & 3.055s \\ \hline
\end{tabular}
\label{table:exec}
\end{table}
\vspace{-0.15 cm}
We can see that the required time strongly depends on the number of monitoring records and the complexity of the observed service. Except for the self-validation, the execution time of all parts of the pipeline depends on the number of records (see Table \ref{table:perf}). In all considered cases, the execution of the whole pipeline took less than \textbf{20} seconds.
We observe that adaptive monitoring limits the number of monitoring records. Accordingly, the second iteration of TeaStore monitoring produces far fewer records, which results in significantly lower execution times.
To gain detailed insights about the performance of the pipeline, we also examined the behavior for an increasing number of monitoring records. For this purpose, we generated monitoring data using \ac{CoCoME} and the load test that we also used before. Thereafter, we executed the pipeline several times, with an increasing number of monitoring records as input. The results in Table \ref{table:perf} showed that the execution time of the pipeline scales linearly with the number of monitoring records in this case.
\input{externals/performance_details_table.tex}
\vspace{-0.25 cm}
\subsection{Evaluation of Monitoring Overhead}
To answer the EQ4.1 and EQ4.2, we measured the overhead caused by the monitoring for the incremental calibration scenarios in \Cref{Evaluation incremental Calibration}.
The average monitoring overhead for \textit{bookSale} was \textbf{1.31ms} (fine-grained), \textbf{252\boldmath$\mu$s} (coarse-grained) and for \textit{train} \textbf{1.81ms} (fine-grained), \textbf{88\boldmath$\mu$s} (coarse-grained). We note that the coarse-grained monitoring has a negligible influence. The fine-grained monitoring overhead scales with the complexity of the service. In our case, the overhead of around 1ms had no drastic impact on the performance, since the execution times of the services are significantly higher.
The evolution of \textit{TrainForRecommender} calibration scenario (cf. \Cref{TrainForRecommender}) shows very well that the adaptive monitoring helps to significantly reduce the monitoring overhead.
Comparing to \textit{train} (cf. \Cref{Train Calibration}) where all SEFF elements are instrumented, we saved \textbf{at least 75\%} of the monitoring overhead by instrumenting only the changed internal action.
This is reflected in the calibration times, which are considerably lower compared to the first iteration (see \autoref{table:exec}) (EQ4.2).
In general, there is a trade off between the monitoring overhead and the accuracy of the model. However, our approach tries to balance them by using the incremental calibration in combination with the adaptive monitoring in order to minimize calibration time and monitoring overhead.
|
\section{Introduction}
Black hole pair creation has been an interesting subject, in the past few decades, partly because it leads to a deeper understanding of the black hole entropy~\cite{garfinkle1991,garfinkle1994,hawking1995,booth1998}. In fact, it has been shown in~\cite{garfinkle1994} that the pair production amplitude contains a factor of $e^{S_{bh}}$, where $S_{bh}$ denotes the black hole entropy. This is consistent with the view that $e^{S_{bh}}$ measures the number of black hole (internal or surface) states.
Pair creation through quantum mechanical tunneling effect can also explain the production of black holes less massive than a few Solar masses. Due to fermionic degeneracy pressure, gravitational collapse cannot be responsible for the production of these small black holes. A recent upper bound on the mass of cold spherical neutron stars is obtained to be $2.3\, M_{\astrosun}$, obtained using the results of GW170817, where it has been shown that neutron stars more massive than this limit would gravitationally collapse into a black hole ~\cite{shibata2019}.
A possible mechanism to produce black hole pairs is through breaking of a cosmic string~\cite{hr} (see also~\cite{achucarro1995}). The string could also ``fray'' to create black hole pair~\cite{eardley1995}. When there is no background magnetic field, it has been shown in~\cite{ashoorioon2014} that the tension of the strut between the two black holes, $\mu_4$, must be smaller than the tension of the string, $\mu_3$, from each black hole to infinity. Given the observational constraint on the tension, $\mu\leq 2\times 10^{-7}$~\cite{seljak2006}, either case of breaking or fraying of the cosmic string to produce black holes of mass larger than the Planck mass is rare~\cite{hr,eardley1995}.
On the other hand, pair creation of black holes in a background magnetic field, with no string, has been widely studied in the literature~\cite{garfinkle1991,dowker1994,hawking1995}. Using the instanton action, the rate of this process is found in the semiclassical approximation. As we show explicitly in this paper, for observationally viable values of the intergalactic magnetic field, the production rate of black holes of mass larger than the Planck mass is negligibly small.
In this paper we investigate the pair production of magnetically charged black holes on a cosmic string in the presence of a background magnetic field. To find the rate of this process, we use the action, $I$, of instanton interpolating between two following states. The first state is a spacetime with magnetic field and a cosmic string of tension $\mu_3$. We refer to this state as background spacetime. In the second state, which we refer to as physical spacetime, we have two black holes, each connected to a string of tension $\mu_3$ that runs to infinity. There may also be a strut between these two black holes with tension $\mu_4$. This configuration is in the presence of a magnetic field.
For a general value of the background magnetic field, we use numerical analysis and we find that the instanton action decreases as the background magnetic field increases. Since the pair production rate is given by $e^{-I}$, we find that for larger values of the background magnetic field the pair production is more probable.
We find an explicit relation for the action $I$ in the weak field limit. We compare this action, with that of the pair creation of black holes in a background magnetic field and no string. The setup we focus on is theoretical, however, we investigate the cosmological implications with real parameters. Using the cosmologically viable values for the tension of the cosmic string and the background magnetic field, we show that the presence of the string would significantly decrease the value of the action and, hence, increase the production rate.
For the sake of comparison with the black hole pair production, we study magnetically charged monopole-antimonopole pair production on a cosmic string in a background magnetic field. Here the configuration of physical spacetime is like that of the black hole case with black holes replaced by magnetic monopoles. To obtain the action of monopole pair production we use the WKB approximation.
We also heuristically study how the presence of a positive cosmological constant affects the black hole pair creation rate in the presence of a background magnetic field and cosmic string. We find that the pair creation rate is higher in a de Sitter (dS) background. During the inflation, where the cosmological constant has its highest value, black hole pair production is the most probable.
The outline of the paper is as follows. In the next section we review the Ernst metric which describes a pair of black holes connected to semi-infinite cosmic strings, in a background magnetic field. We study the production rate of such pairs in Sec.~\ref{sec:rate} by using the instanton action. The weak field limit of this action is studied in Sec.~\ref{sec:wf} where we provide some examples to show how the presence of a cosmic string would decrease the action. In Sec.~\ref{sec:mp} we study the pair production of monopole-antimonopole on a cosmic string in a background magnetic field. A heuristic derivation of the action in a dS background is presented in Sec.~\ref{sec:ds}. We conclude our paper in Sec.~\ref{sec:con}. We use the units in which $G=c=\hbar=\epsilon_0=1$.
\section{Black hole pair on a cosmic string}\label{sec:ernst}
Before the pair creation, the spacetime consists of a cosmic string with tension $\mu_3$ in a magnetic field. We call this state, the background spacetime and it will be described by Melvin solution~\cite{melvin1964}
\begin{align}
ds^2&=\bar{\Gamma}^2\left[-dt^2+dz^2+d\rho^2\right]+\bar{\Gamma}^{-2}\rho^2d\bar{\phi}^2, \label{melvin}\\
\bar{\Gamma}&=1+\frac{1}{4}\hat{B}_{M}^2\rho^2, \nonumber
\end{align}
where $\hat{B}_{M}$ denotes the magnetic field, and the gauge potential is given by $2\bar{\Gamma}A_{\bar{\phi}}=\hat{B}_{M}\rho^2$. The cosmic string results in a deficit angle $\delta_3=8\pi\mu_3$ in the azimuthal coordinate, so $\Delta\bar{\phi}=2\pi-\delta_3$.
After the pair creation, the spacetime, which now we refer to as physical spacetime, will be given by the Ernst metric~\cite{ernst1976}
\begin{align}
ds^2&=r^2\Gamma^2\left[G(y)dt^2-G^{-1}(y)dy^2+G^{-1}(x)dx^2\right] \nonumber\\ &+\frac{r^2G(x)}{\Gamma^2}d\phi^2, \label{ernst}\\
A_{\phi}&=-\frac{2}{B\Gamma}\left(1+\frac{1}{2}Bqx\right)+k, \nonumber
\end{align}
where $r=A^{-1}(x-y)^{-1}$ represents the radial coordinate~\cite{kinnersley1970}, and
\begin{align}
\Gamma&=\left(1+\frac{1}{2}B q x\right)^2+\frac{1}{4}r^2B^2G(x),\\
G(\xi)&=(1+r_{-}A\xi)(1-\xi^2-r_{+}A\xi^3). \nonumber
\end{align}
The parameter $q^2=r_{-}r_{+}$ is related to the charge of the black hole and the mass of the black hole is given by $m=(r_{-}+r_{+})/2$. This solution describes two charged black holes accelerating from each other. The acceleration is given by the parameter $A$ with its inverse representing the typical distance between the two black holes at the time of their creation~\cite{dias2004}.
The physically interesting case is the one in which the function $G(\xi)$ has four real roots. One root is given by $\xi_1=-1/(r_{-}A)$. To have three other real roots we need to restrict the parameters so that $r_{+}A\leq\sqrt{4/27}$, which means that the black holes should be small and/or they should be far apart from each other. For $r_{+}A\ll 1$ one finds the remaining roots as follows
\begin{align}
\xi_2&=-\frac{1}{r_{+}A}+r_{+}A+\cdots,\label{roots}\\
\xi_3&=-1-\frac{r_{+}A}{2}+\cdots, \nonumber\\
\xi_4&=1-\frac{r_{+}A}{2}+\cdots. \nonumber
\end{align}
To obtain the correct signature of the metric (\ref{ernst}), we require $\xi_3\leq x \leq \xi_4$ and $-\infty<y\leq x$. The surfaces $y = \xi_1$, $\xi_2$ , $\xi_3$ correspond to black hole inner horizon, event horizon, and acceleration horizon, respectively. Also $x=\xi_3$ and $x=\xi_4$ are axes that point to spatial infinity and to the other black hole, respectively. The parameter $k$, in the gauge potential of Eq. (\ref{ernst}), will be chosen so as to confine the Dirac string singularities to the $x=\xi_4$ axis. It should also be mentioned that the parameter $B$ in the Ernst metric is related to the physical magnetic field $\hat{B}$ by~\cite{dowker1994}
\begin{equation}
\hat{B}=\frac{B\,G'(\xi_3)}{2L^{3/2}}, \label{physb}
\end{equation}
where $L=\Gamma(\xi_3)$. It could be easily shown that in the weak field limit $\hat{B}=B$ to the leading order of background magnetic field. Also, the physical magnetic charge of the black hole is defined by $\hat{q} = 1/(4\pi)\int F$, where $F$ denotes the Maxwell field~\cite{dowker1994}. Integration over a two sphere surrounding the black hole yields
\begin{equation}
\hat{q} = q \frac{\Delta\phi\left(\xi_4-\xi_3\right)}{4\pi L^{1/2}\left(1+\frac{1}{2}B q \xi_4\right)}, \label{eqn:charge}
\end{equation}
where $\Delta\phi$ is the period of the azimuthal coordinate.
Consider the angular part of the Ernst metric (\ref{ernst})
\begin{equation}
d\Phi^2=r^2\left[\frac{\Gamma^2}{G(x)}dx^2+\frac{G(x)}{\Gamma^2}d\phi^2\right].
\end{equation}
Using the change of variable $\theta=\int_{\xi_i}^{x}\Gamma G^{-1/2}(x)dx$, near the poles $x=\xi_i$ ($i=3,4$), we obtain
\begin{equation}
d\Phi^2=r^2\left[d\theta^2+\frac{G'^2(\xi_i)}{4\Gamma^4(\xi_i)}\theta^2d\phi^2\right].
\end{equation}
We are interested in a configuration in which each black hole is connected to a string with deficit angle $\delta_3$ that runs to infinity and a strut with deficit angle $\delta_4$ that goes to the other black hole. In such case, one can show that the following relation holds near the poles
\begin{equation}
\frac{|G'(\xi_i)|}{2\Gamma^2(\xi_i)}\Delta\phi=2\pi-\delta_i, \qquad i=3,4. \label{phiatpoles}
\end{equation}
Now we write the period of the azimuthal coordinate in terms of $\xi_3$
\begin{equation}
\Delta\phi=\frac{2L^2}{G'(\xi_3)}(2\pi-\delta_3)\,. \label{deltaphi}
\end{equation}
It is easy to show that $G'(\xi_3)>0$. By using Eqs. (\ref{phiatpoles}) and (\ref{deltaphi}), we find
\begin{equation}
\left|\frac{G'(\xi_4)}{G'(\xi_3)}\right|\frac{L^2}{\Gamma^2(\xi_4)}=\frac{2\pi-\delta_4}{2\pi-\delta_3}. \label{wbnewton}
\end{equation}
This relation would lead to Newton's law in the weak field limit $Bq\ll1$. In observationally important case we have $\delta\ll1$. We are also working in the limit which the size of the black hole is small compared to the typical distance between them (the so called small black hole limit). Then Eqs. (\ref{roots}) and (\ref{wbnewton}) would result in
\begin{equation}
mA=\left(\mu_3-\mu_4\right)+Bq. \label{newton}
\end{equation}
Now consider the first two terms of the Ernst metric (\ref{ernst}). By Wick rotating the time coordinate $t=i\tau$ and following a similar calculations as above, we find that the period of the imaginary time at the surfaces $y=\xi_i$ ($i=2,3$) is $\Delta\tau=4\pi/|G'(\xi_i)|$. For the case of extreme black hole, $\xi_1=\xi_2$, the surface $y=\xi_2$ is infinitely far away from any point in the Euclidean sector of the metric~\cite{hawking1995}. As a result, one does not worry about the periodicity of $\tau$ at this surface (In this case we take $\Delta\tau=4\pi/G'(\xi_3)$). For the non-extreme case we request that $y=\xi_2$ and $y=\xi_3$ have the same periodicity of $\tau$, or equivalently, have the same temperature $1/\beta$. So
\begin{equation}
\Delta\tau=\beta=\frac{4\pi}{G'(\xi_3)}=-\frac{4\pi}{G'(\xi_2)},
\end{equation}
which leads to the following relation for the non-extremal case
\begin{equation}
\xi_4-\xi_3=\xi_2-\xi_1. \label{nerelation}
\end{equation}
\section{Production rate of black hole pair on a cosmic string in the presence of a background magnetic field}\label{sec:rate}
In this section we are going to obtain the rate of black hole pair production on cosmic string in the presence of a background magnetic field. To do so, we use the action of the instanton that interpolates between background and physical spacetimes. Based on the reasoning of~\cite{hr,hawking1996}, this action is equivalent to the Euclidean action
\footnote{We note that in Eqs. (\ref{neaction}) and (\ref{eaction}) all the surface terms have been taken into account~\cite{hawking1996}. The discussion of~\cite{hawking1996} applies to all spacetimes that can be foliated by complete, nonintersecting spacelike surfaces. In derivation of this formulae one needs to keep in mind that, since the Lorentzian action diverges for spatially non-compact geometries, one needs a reference background. One then fixes a boundary near infinity and requires that the metric and matter fields of the physical spacetime induce the same fields on the boundary as those of the background spacetime induce. Also, it is assumed that the periodicity of time coordinate $\beta$ be the same in physical and background spacetimes.}
\begin{equation}
I=\beta H-\frac{1}{4}\Delta\mathcal{A}, \label{eaction}
\end{equation}
in the extremal black hole case, and
\begin{equation}
I=\beta H-\frac{1}{4}\left(\Delta\mathcal{A}+A_{bh}\right), \label{neaction}
\end{equation}
in the non-extremal case.
\footnote{The extra term in Eq. (\ref{neaction}) relative to Eq. (\ref{eaction}) shows that the production rate in non-extreme case is enhanced by $e^{S_{bh}}$, where $S_{bh}=A_{bh}/4$ is the entropy of the non-extremal black hole, as the entropy of an extremal black hole is zero~\cite{hawking1995}. See also~\cite{emparan1995:2} to find out why only the entropy of one of the black holes enter the equations.}
Here $\Delta\mathcal{A}$ is the difference between the area of the acceleration horizon in physical and background spacetimes. $A_{bh}$ is the area of the black hole, and the Hamiltonian $H$ is given by~\cite{hawking1996}
\begin{equation}
H=\int_{\Sigma}N\mathcal{H}-\frac{1}{8\pi}\int_{S_{\infty}^2}N\left(^2K-{}^2K_0\right), \label{hamilton}
\end{equation}
where $N$ is the lapse function and $\Sigma$ is the constant time hypersurface. $\mathcal{H}$ is the Hamiltonian constraint which vanishes on the solution. $S_{\infty}^2$ is the boundary of $\Sigma$ and $^2K$ and $^2K_0$ are the extrinsic curvature of this boundary in physical and background spacetimes, respectively. Note that $S_{\infty}^2$ has been chosen so as to match in physical and background spacetimes. This way, a calculation similar to that of~\cite{hawking1995} would result in the conclusion that the second integral in Eq. (\ref{hamilton}) also vanishes.
Now, we are going to find the area of the black hole horizon, $A_{bh}$, which appears in the action (\ref{neaction}) for non-extremal black holes. The event horizon is at $y=\xi_2$ in the Ernst metric (\ref{ernst}). We find, by using (\ref{deltaphi}),
\begin{align}
A_{bh}&=\int_{y=\xi_2}\sqrt{g_{xx}g_{\phi\phi}}dxd\phi \nonumber\\ &=\frac{2L^2}{A^2G'(\xi_3)}\left(2\pi-\delta_3\right)\frac{\xi_4-\xi_3}{\left(\xi_3-\xi_2\right)\left(\xi_4-\xi_2\right)}. \label{bharea}
\end{align}
Now, we proceed to find the area of the acceleration horizons. Recalling the definition of the radial coordinate, $r=A^{-1}(x-y)^{-1}$, and the restrictions on $x$ and $y$, we note that equalities $x = y = \xi_3$ show the spatial infinity. To find the area of the acceleration horizon in the Ernst metric, which is located at $y=\xi_3$, we integrate from $x=\xi_4$ to $x=\xi_3+\epsilon$, and will take the limit $\epsilon\rightarrow0$ at the end. We would find
\begin{align}
\mathcal{A}&=\int_{y=\xi_3}\sqrt{g_{xx}g_{\phi\phi}}dxd\phi\nonumber\\ &=\frac{2L^2}{A^2G'(\xi_3)}\left(2\pi-\delta_3\right)\left(\frac{1}{\epsilon}-\frac{1}{\xi_4-\xi_3}\right). \label{ernstacch}
\end{align}
The area of the acceleration horizon in the Melvin metric (\ref{melvin}) is
\begin{equation}
\bar{\mathcal{A}}=\int\rho d\rho d\bar{\phi}=\frac{2\pi-\delta_3}{2}\bar{\rho}^2. \label{melvinacch}
\end{equation}
By requiring that the proper length of the boundary and the integral of the gauge potential around the boundary be the same in Ernst and Melvin metrics, it has been found in~\cite{hawking1995} that
\begin{equation}
\bar{\rho}^2=\frac{4L^2}{A^2G'(\xi_3)\epsilon}\left(1+\frac{G''(\xi_3)}{4G'(\xi_3)}\epsilon\right)^2.
\end{equation}
Substituting $\bar{\rho}$ from the above equation into Eq. (\ref{melvinacch}), we obtain
\begin{equation}
\bar{\mathcal{A}}\simeq\frac{2L^2}{A^2G'(\xi_3)}\left(2\pi-\delta_3\right)\left(\frac{1}{\epsilon}+\frac{G''(\xi_3)}{2G'(\xi_3)}\right). \label{melvinacch2}
\end{equation}
To find the second term in the above equation we note that one could write the function $G(\xi)$ as
\begin{equation}
G(\xi)=-(r_-A)(r_+A)(\xi-\xi_1)(\xi-\xi_2)(\xi-\xi_3)(\xi-\xi_4).
\end{equation}
By using Eqs. (\ref{ernstacch}) and (\ref{melvinacch2}) we find that
\begin{align}
\Delta\mathcal{A}=\mathcal{A}-\bar{\mathcal{A}}&=-\frac{2L^2}{A^2G'(\xi_3)}\left(2\pi-\delta_3\right) \nonumber\\
&\times\left(\frac{1}{\xi_3-\xi_1}-\frac{1}{\xi_3-\xi_2}\right). \label{acch}
\end{align}
By substituting this equation along with Eq. (\ref{bharea}) into Eq. (\ref{neaction}), and using the non-extremality relation (\ref{nerelation}), we obtain
\begin{equation}
I=\frac{L^2}{A^2G'(\xi_3)}\left(\frac{2\pi-\delta_3}{\xi_3-\xi_1}\right). \label{action}
\end{equation}
The same equation would be found for the action of extremal black hole, by using Eqs. (\ref{eaction}) and (\ref{acch}) and the relation $\xi_2=\xi_1$.
Here we would like to study the action (\ref{action}) for the case in which $q = m$ (or equivalently $r_+ = r_-$). We have restricted the black holes to be small on the scale set by the acceleration. That is, we are considering the case in which $r_+A\ll1$. By using this condition, we could show from (\ref{physb}) that
\begin{equation}
\hat{B}\simeq\frac{8B}{(2-Bm)^{3}}. \label{physb:genb}
\end{equation}
We see that $\hat{B}$ goes from zero to infinity as $B$ changes from zero to $2/m$.
On the other hand, for the case $q = m$, Eq. (\ref{action}) could be written as
\begin{equation}
I\simeq\frac{m(B m-2)^4}{32A}\left(2\pi-\delta_3\right), \label{action:genb}
\end{equation}
where the parameter $A$ has to satisfy Eq. (\ref{wbnewton}). We have not been able to solve Eq. (\ref{wbnewton}) analytically, to find $A$ for a general value of $B$. However by using numerical methods, we have used Eqs. (\ref{physb:genb}), (\ref{action:genb}), and (\ref{wbnewton}) to plot the action as function of physical magnetic field $\hat{B}$ for the $q = m$ case in Fig.~\ref{fig:action}. It is obvious from this figure that the instanton action decreases as the physical magnetic field increases. Therefore, $e^{-I}\rightarrow1$ as $\hat{B}\rightarrow\infty$, which means that the black hole pair production rate increases by increasing the magnetic field.
\begin{figure}[htp]
\centering
\includegraphics[width=0.45\textwidth]{action.pdf}
\caption{The action as a function of the physical magnetic field for the $q = m$ case. We see that the action decreases by increasing the physical magnetic field $\hat{B}$. We have set $m\,(=q)=10$, $\mu_3 = 2\times 10^{-7}$, and $\mu_4 = 10^{-7}$.}
\label{fig:action}
\end{figure}
\section{Weak field limit of the action}\label{sec:wf}
We see from Fig.~\ref{fig:action} that the action for the pair production of black holes of masses larger than the Planck mass is of order unity, only when the physical magnetic field is about 10 or larger. This is much larger than the value of intragalactic magnetic field, which, nowadays, is less than about $10^{-63}$ in natural units~\cite{ashoorioon2005}. It motivated us to study the weak field limit of the action (\ref{action}).
In the weak field limit, the physical magnetic field (\ref{physb}) and charge (\ref{eqn:charge}) can be approximated by
\begin{align}
\hat{B}&\simeq B\left[1-\frac{1}{2}Bq-2\left(\mu_3-\mu_4\right)\right], \label{eqn:physb:emp}\\
\hat{q}&\simeq q\left(1-4\mu_3\right)\left[1+2\left(\mu_3-\mu_4\right)\right],
\end{align}
where we have used Eq. (\ref{deltaphi}). These expressions reduce to the ones derived in ~\cite{emparan1995} when $\mu_3=0$. Using these equations, along with the Newton's law (\ref{newton}) and the non-extremality relation (\ref{nerelation}), one can write all the parameters in terms of $\hat{B}$, $\hat{q}$, $\mu_3$, and $\mu_4$
\begin{align}
r_{\pm}&\simeq \hat{Q}\left[1-2\left(\mu_3-\mu_4\right)\pm\left(\hat{B}\hat{Q}+\mu_3-\mu_4\right)\right], \label{eqn:para:r}\\
q&\simeq\hat{Q}\left[1-2\left(\mu_3-\mu_4\right)\right], \label{eqn:para:q}\\
A&\simeq\hat{B}\left[1+\frac{1}{2}\hat{B}\hat{Q}+\frac{\mu_3-\mu_4}{\hat{B}\hat{Q}}\left(1+2\hat{B}\hat{Q}\right)\right], \label{eqn:para:A}
\end{align}
where $\hat{Q}=\hat{q}\left(1+4\mu_3\right)$. Again see similar expressions in ~\cite{emparan1995}, when $\mu_3=0$.
By substituting the parameters (\ref{eqn:para:r})-(\ref{eqn:para:A}) into Eq. (\ref{action}) we find the weak field approximation of the action as
\begin{equation}
I\simeq\pi\hat{Q}^2\left(1-4\mu_3\right)\left[\frac{1}{\hat{B}\hat{Q}}-\frac{1}{2}-\frac{\mu_3-\mu_4}{\hat{B}^2\hat{Q}^2}\right]. \label{eqn:act:emp}
\end{equation}
If we have no string, $\mu_3 = \mu_4 = 0$, one obtains the result of~\cite{dowker1994}. Hence, Eq. (\ref{eqn:act:emp}) is also in agreement with Schwinger result~\cite{schwinger1951}.
\footnote{Eq. \eqref{eqn:act:emp} agrees to the first three terms of the action obtained in~\cite{emparan1995}. However, if we continue our expansion \eqref{eqn:act:emp}, the fourth term would be $-\pi\hat{Q}\left(1-4\mu_3\right)\left(\mu_3-\mu_4\right)/\hat{B}$ which differs from the fourth term of~\cite{emparan1995} by a factor of 2. We only consider the first three terms of the action \eqref{eqn:act:emp} when we are comparing black hole pair production action to that of monopoles in the next section.}
Now we study the action in the case where $q=m$ and we are also in the weak field limit. In this case, the action (\ref{action}) can be written as
\begin{equation}
I\simeq \frac{\pi m (1-2Bm)}{A}\left(1-4\mu_3\right). \label{action:smallb}
\end{equation}
By using Eqs. (\ref{newton}) and (\ref{eqn:physb:emp}) one can rewrite the above equation as
\begin{equation}
I\simeq \frac{\pi m^2 \left[1-2\left(J+\mu_3-\mu_4\right)\right]\left(1-4\mu_3\right)}{J-1+\left[1+2\left(\mu_3-\mu_4\right)\right]\left(\mu_3-\mu_4\right)}, \label{ours}
\end{equation}
where the auxiliary function $J$ is
\begin{equation}
J=\sqrt{\left[1-2\left(\mu_3-\mu_4\right)\right]\left[1-2\left(\hat{B}m+\mu_3-\mu_4\right)\right]}. \nonumber
\end{equation}
Let us see how the presence of the cosmic string would change the rate of the production. Suppose that a pair of black holes, each of mass $m=10$, pop up in the background magnetic field $\hat{B}=10^{-63}$ which is the today's value of the intragalactic magnetic field. Substituting these values into Eq. (\ref{ours}) with $\mu_3 = \mu_4 = 0$, we find $I\simeq 3.141592\times 10^{64}$ for the action of this pair creation. Let us compare this with the case where $\hat{B}=10^{-63}$, $m=10$, and $\mu_3 = \mu_4 =2\times 10^{-7}$ i.e.~when the pair creation of black holes occurs on a cosmic string with uniform tension. In this case we find $I\simeq 3.141590 \times 10^{64}$, which is slightly smaller than when black holes pair is created from vacuum. This is due to the factor $1-4\mu_3$ in Eq. \eqref{ours}. This implies that the presence of the string, even if it does not fray or break, reduces the action, although for the cosmic strings with observationally relevant values of tension, the change in the action will be quite small.
Now let us assume that we have a string of tension $\mu_3=2\times 10^{-7}$ in the Melvin universe (recall that this is the upper limit on the tension of the cosmic string). Assume that the string frays and produces a pair of black holes with the same mass $m=10$. Let us take $\mu_4=10^{-7}$, for the tension of the strut between the two black holes. Also take the same value of the background magnetic field as above, $\hat{B}=10^{-63}$. Substituting these values into the action (\ref{ours}), we find $I\simeq 3.141590 \times 10^{9}$.
\footnote{We note that if we omit the factor $\left(1-4\mu_3\right)$ in Eq. (\ref{ours}), we would find $I\simeq 3.141592 \times 10^{9}$. Therefore, the contribution of this factor is very small compared to the changes of the action upon breaking/fraying of the cosmic string.}
This is smaller than the value of the action in the absence of the string, by 55 order of magnitude!
In the early universe, the background magnetic field can be much stronger. It is supposed to be of the order $\hat{B}=10^{-13}$ in natural units at the end of inflationary era \footnote{We have assumed that the observed intragalactic field, has been produced during inflation but its strength has reduced super-adiabatically, $B\propto a^{-2}$, since then.}. Using this value of the background magnetic field along with $\mu_3 = \mu_4 = 0$ and $m=10$, we find $I\simeq 3.141592 \times 10^{14}$, which is smaller than the corresponding value in today magnetic field strength by 50 orders of magnitude.
For the case of pair creation on the cosmic string in the early universe, we use $\mu_3=2\times 10^{-7}$, $\mu_4=10^{-7}$, $m=10$, and $\hat{B}=10^{-13}$. By substituting these values into Eq. (\ref{ours}), we find $I\simeq 3.141559 \times 10^{9}$. One notices that the value of the action for pair creation on a cosmic string at the end of the inflationary phase does not differ significantly from its current value, although without cosmic string, the action reduces significantly.
We find from Eq. (\ref{ours}) that for fixed values of the parameters, the action is minimized if the string breaks, i.e.~$\mu_4=0$. For $\hat{B}=10^{-63}$, $m=10$, $\mu_3=2\times 10^{-7}$, and $\mu_4=0$ we find $I\simeq 1.570795 \times 10^{9}$. On the other hand if the cosmic string does not fray/break, one finds the maximum value of the action for a given set of the parameters, even though we should stress that the obtained value in this case is slightly smaller than the value of the action when the pair creation of black holes takes place only in presence of background magnetic field without a cosmic string.
We also see from Fig.~\ref{fig:actionwf} that for fixed $\mu_3$, $\mu_4$, and $\hat{B}$, the action (\ref{ours}) would increase by increasing the mass of the black holes. One concludes that the probability of pair producing more massive black holes is smaller.
\begin{figure}[htp]
\centering
\includegraphics[width=0.45\textwidth]{actionwf.pdf}
\caption{The action for the $q=m$ case in the weak field limit as a function of black hole mass. We see that the action increases by increasing the mass of the black holes. We have taken $\mu_3=2\times 10^{-7}$, $\mu_4=10^{-7}$, and $\hat{B}=10^{-63}$.}
\label{fig:actionwf}
\end{figure}
\section{Monopole pair production on cosmic strings in a background magnetic field}\label{sec:mp}
Monopoles are hypothetical particles with quantized magnetic charge~\cite{dirac1931}. It has been shown that they can be produced in theories with certain non-Abelian gauge fields~\cite{t1974}. In this section we study the production rate of magnetically charged monopole-antimonopole pair on cosmic string in the presence of a background magnetic field.
In the absence of a background magnetic field, the semiclassical probability for a cosmic string to break producing a monopole-antimonopole pair has been obtained and their cosmological evolution has been studied~\cite{vilenkin1982}. The production rate of monopole pairs in a background magnetic field has also been studied in~\cite{affleck1982}. Recently, the rate of monopole pair production has been investigated in thermal baths~\cite{gould2017} and strong background magnetic fields~\cite{rajantie2019,ho2020}. Investigation of monopole pair production on cosmic string in a background magnetic field is missing in the literature.
Here we use WKB approximation, similar to that of~\cite{vilenkin1982}, to find the action of monopole pair production on cosmic string in an external magnetic field. We compare this result with the action of black hole pair production on a cosmic string in a background magnetic field.
Assume a cosmic string with tension $\mu_3$ on an axis $z$. Assume also a background magnetic field $\hat{B}$ parallel to this axis. Suppose the string breaks and produces a monopole and an antimonopole each connected to a semi-infinite string. Because of the symmetry, we only consider the monopole in the following analyze and include the antimonopole rule by multiplying the monopole action by 2, in the end.
The monopole has mass $m$ and magnetic charge $\hat{q}$. The energy of the monopole as a function of $z$ is $m+\left(\mu_3+\hat{B}\hat{q}\right)z$. We note that if the string frays and we have strut of tension $\mu_4$ between the pair, then the force on monopole due to the string-strut system is $\mu_3-\mu_4$ and the energy would be
\begin{equation}
E=m+\left(\mu_3-\mu_4+\hat{B}\hat{q}\right)z.
\end{equation}
The additive constant has been chosen to make the monopole bounce at $z=0$, see \cite{vilenkin1982}. The action for the pair production is then
\begin{equation}
I_{mon}\simeq 4 \int_{\frac{-m}{\mu_3-\mu_4+\hat{B}\hat{q}}}^{0}\sqrt{m^2-E^2}dz.
\end{equation}
The lower bound of the integral has been chosen so as to make $E\geq 0$ throughout the integration domain. Integrating we find
\begin{equation}
I_{mon}\simeq \pi m^2 \left(\frac{1}{\hat{B}\hat{q}}-\frac{\mu_3-\mu_4}{\hat{B}^2\hat{q}^2}\right). \label{eqn:mnac1}
\end{equation}
It has been suggested in~\cite{dowker1994} that, regarding the production rate, monopoles behave more like extremal black holes than non-extremal ones. This suggestion is based on the calculations of~\cite{garfinkle1994} which shows that the action of the instanton describing monopole pair production is greater than the action of non-extreme black holes by the black hole entropy. On the other hand, it was found in~\cite{dowker1994} that the instanton action for the pair production of extremal black holes is larger than the action for the case of non-extreme black holes again by the black hole entropy. For the extreme black holes we have $m=\hat{q}$. Thus if we regard monopoles as extreme black holes we can rewrite Eq. \eqref{eqn:mnac1} as
\begin{equation}
I_{mon}\simeq \pi \hat{q}^2 \left(\frac{1}{\hat{B}\hat{q}}-\frac{\mu_3-\mu_4}{\hat{B}^2\hat{q}^2}\right). \label{eqn:mnac2}
\end{equation}
Eq. \eqref{eqn:act:emp} is the action for pair production of (slightly) non-extremal black holes, and, in $\mu_3\ll 1$ limit, can be written as the following
\begin{equation}
I\simeq \pi \hat{q}^2 \left(\frac{1}{\hat{B}\hat{q}}-\frac{\mu_3-\mu_4}{\hat{B}^2\hat{q}^2}-\frac{1}{2}\right).
\end{equation}
Comparing this equation with Eq. \eqref{eqn:mnac2}, we find that the action obtained by WKB approximation is greater than the action of non-extremal black holes by half the Bekenstein-Hawking entropy $\pi\hat{q}^2/2=S_{bh}/2$. Similar discrepancies between the actions for non-extremal black hole pair creation, \cite{dowker1994}, the exact action for monopole pair creation, \cite{affleck1982}, and the WKB result for the monopole pair generation, all in a uniform magnetic background, is noticed. The difference is probably mainly due to the inadequacy of the WKB approach in obtaining the exact result. More on this and related issues remain to be studied.
\section{Heuristic derivation of the production rate in a de Sitter background}\label{sec:ds}
In this section we study black hole pair production on a cosmic string in the presence of a background magnetic field along with a positive cosmological constant $\Lambda$. The rate of black hole pair creation in dS background has been first studied in~\cite{mellor1989}, and further investigated in~\cite{mann1995}. It was shown that the action of the instanton that mediate between dS spacetime and dS spacetime with a pair of black holes, each of mass $m$, is $I = \pi m \sqrt{3/\Lambda}$. Later, in~\cite{dias2004:2}, a cosmic string was invoked in a dS background and the rate of black hole pair production has been studied.
To study the pair production of dS black holes in the presence of a background magnetic field, one needs a dS version of the Ernst metric. However such a solution does not exist~\cite{dias2003}. In fact, Ernst has used an Ehlers-Harrison type transformation~\cite{ehlers1959,harrison1968} to add the magnetic field to the C metric. But, applying this transformation to the cosmological C metric would not yield a new solution to the Einstein-Maxwell theory~\cite{dias2003}.
Nonetheless we can deal with this problem heuristically. One can think of a positive cosmological constant doing the same role as the cosmic string; in the sense that it provides the energy to materialize the pair of black holes and, also, provides the force to accelerate the black holes apart.
To find the acceleration caused by a positive cosmological constant, we recall that the Newtonian potential of dS spacetime is $\Lambda r^2/6$. The force per unit mass or the acceleration would then be $\Lambda r/3$, where $r$ is the dS radius given by $r=\sqrt{3/\Lambda}$. Therefore, in the weak field limit, we can modify the Newton's law (\ref{newton}) to be
\begin{equation}
mA=\tilde{\mu}+Bq,
\end{equation}
where $\tilde{\mu}=\mu_3-\mu_4+m\sqrt{\Lambda/3}$. Likewise, we can modify the actions (\ref{eqn:act:emp}) and (\ref{ours}) by replacing $\mu_3-\mu_4$ by $\tilde{\mu}$.
Consider, for instance the $q=m$ case in the weak field limit. The dS counterpart of the action (\ref{ours}) is
\begin{equation}
I\simeq \frac{\pi m^2 \left[1-2\left(\tilde{J}+\tilde{\mu}\right)\right]\left(1-4\mu_3\right)}{\tilde{J}-1+\left(1+2\tilde{\mu}\right)\tilde{\mu}}, \label{ours:ds}
\end{equation}
where
\begin{equation}
\tilde{J}=\sqrt{\left(1-2\tilde{\mu}\right) \left[1-2\left(\hat{B}m+\tilde{\mu}\right)\right]}. \nonumber
\end{equation}
In the limit $\Lambda\ll m^{-2}$, and $\hat{B},\, \mu_3,\, \mu_4 \rightarrow 0$, Eq. (\ref{ours:ds}) reduces to $I\simeq \pi m \sqrt{3/\Lambda}$, to leading order of $\Lambda$. This is consistent with the result of~\cite{mann1995} for the case of lukewarm black holes in which one has non-extreme black holes satisfying $q^2=m^2$~\cite{romans1992}.
We note here that if we modify the parameters (\ref{eqn:para:r})-(\ref{eqn:para:A}) to include the cosmological term and substitute them into Eq. (\ref{action}), then, the limit $\Lambda\ll m^{-2}$, and $\hat{B},\, \mu_3,\, \mu_4 \rightarrow 0$ would lead to $I\simeq \pi m \sqrt{3/\Lambda}$ as well. This is because in this limit we have $r_+\rightarrow r_-$, or equivalently $q\rightarrow m$. Therefore, again we are left with lukewarm black holes. This result shows that the lukewarm instanton is the only one available for non-extreme black holes, for which the conical singularities are eliminated at both $y=\xi_2$ and $y=\xi_3$ horizons in a dS background (see~\cite{mann1995} and~\cite{dias2004:2}).
Now let us study the action (\ref{ours:ds}) through some examples, considering the current bounds on the physical quantities. For $\hat{B}=10^{-63}$, $\mu_3=2\times 10^{-7}$, $\mu_4=10^{-7}$, $m=10$, and $\Lambda=10^{-122}$ we find $I\simeq 3.141590 \times 10^{9}$. On the other hand, if we have no string (i.e.~$\mu_3=\mu_4=0$) and keep the other parameters the same, we find $I\simeq 5.348755 \times 10^{62}$.
For the inflationary era, by taking $\hat{B}=0$, $\Lambda=10^{-12}$, $m=10$, and $\mu_3=2\times 10^{-7}$, and $\mu_4=10^{-7}$, we find $I\simeq 5.348751 \times 10^{7}$. For the no string case (i.e.~$\mu_3=\mu_4=0$) it is $I\simeq 5.441398 \times 10^{7}$, which is a bit larger. Although the value of the action for pair creation of black holes would not change significantly during inflation regardless of the presence of the cosmic string, the action is still considerably smaller in this era than the time it frays in presence of today's background magnetic field.
\section{Concluding remarks}\label{sec:con}
The rate of the black hole pair production (per unit time per unit length of the cosmic string) is given by $e^{-I}$. We see from Fig.~\ref{fig:action} that as the background magnetic field increases the black hole pair creation would be more probable. These black holes are accelerated away from each other after their production due to the force of the string and the background magnetic field.
We worked on a theoretical basis, however, we have presented our numerical examples with observationally viable values. Since the background magnetic field in the Universe is very small, we could safely use the weak field limit of the instanton action in cosmological applications. We have compared the pair creation of black holes in a background magnetic field (and no string) with that on a cosmic string in the presence of the background magnetic field. We have shown that the fraying/breaking of the cosmic string, even with the observational bound on its tension, would decrease the action significantly.
In~\cite{emparan1995}, the production rate of black hole pair, connected by a strut but with no string, was obtained in the Melvin universe. It was shown that this rate is suppressed relative to the case where no strut is present. Here we showed that if a cosmic string is present in the Melvin universe from the beginning, the pair production rate is enhanced.
In the early universe at the end of the inflationary era, the background magnetic field is expected to be much larger if they are produced during inflation. We find that the action of the pair production in no string case could be much smaller in that period. However if the pair production took place on a cosmic string, the value of the action at the end of the inflationary era and present day universe do not differ significantly.
We have studied the production rate of monopole-antimonopole pair as well. Here we have used WKB approximation to find the exponent of the rate. We have found that the rate of monopole pair production is suppressed relative to non-extremal black hole pair production by the factor $e^{S_{bh}/2}$. This is not consistent with the suggestion that monopoles behave like extremal black holes in the sense that their production is suppressed relative to non-extremal black hole production by the factor $e^{S_{bh}}$. There might be a problem with the WKB approach. There are various loop corrections with the same order as black hole entropy that have not been taken into account in WKB approximation~\cite{affleck1982,garfinkle1994}. This issue needs further investigations in future studies.
We have also presented a heuristic derivation of the production rate on cosmic string in the presence of a background magnetic field in dS spacetime. Like the cosmic string the presence of a positive cosmological constant decreases the action, hence increases the production rate. By using the action obtained in this way, we find that the pair creation is the most probable in the inflationary era. Contrary to the epoch after inflation, the black hole pair creation rate during inflation is nearly insensitive to whether it takes place on a cosmic string.
In our numerical example we have taken the mass of the black holes to be $m=10$ in Planck units. Such small black holes may tolerate the evaporation~\cite{macgibbon1987}. Besides, for the evaporation of the magnetically charged black holes, one needs copious magnetically charged monopoles to take away the charge of the black holes. This is not observationally acceptable as we do not see many (if any) monopoles around us. In GUT theories with magnetically charged monopoles with with mass $10^{16}~{\rm GeV}\lesssim \mu \lesssim 10^{19}~{\rm GeV}$, it has been shown that the black holes with mass $m\geq 150 \left(\frac{\mu}{10^{17}~{\rm GeV}}\right)^2$ can become stable roughly over the lifetime of the universe \cite{Hiscock:1983ag}. The temperature of the accelerating black holes is also somewhat subtle to define (see~\cite{appels2016,appels2017,appels2020}). More study on the thermodynamics of such black holes is needed.
\bibliographystyle{model1-num-names}
|
\section*{Results}
\subsection*{Fitting a Logistic Regression Model}
We used the training dataset with 351 patients without missing any values of the three important features (i.e., LDH, lymphocyte, and hs-CRP) to fit a step-wise logistic regression model with all the second-order interaction items. We used the coefficient of determination adjusted $R^2$ \cite{hastie2009elements}, indicating the portion of variance in the dependent variable (i.e., patient outcome) explained by the independent variables, to determine which interaction items to include. We obtained a model with the largest adjusted $R^2 = 0.797$ with all the three variables, i.e., LDH ($P = 0.000$), lymphocyte ($P = 0.000$), and hs-CRP ($P = 0.000$), and two interaction items, i.e., LDH:lymphocyte ($P = 0.081$) and LDH:hs-CRP ($P = 0.000$).
\subsection*{Prediction Performance}
Table 1 shows the fatality prediction performance in terms of area under the curve (AUC) \cite{walter2005partial} scores of logistic regression using 100-round five-fold cross-validation. It can be seen that the logistic regression models are able to accurately predict the fatality outcomes of COVID-19 patients with three important biomarkers. As the number of the features increases, the performance also improves before the last interaction item (i.e.,lymphocyte:hs-CRP) was added. This was consistent with the results of fitting a logistic regression model found above. The model with optimal performance had two (product) interaction items, i.e., LDH:lymphocyte and LDH:hs-CRP.
\begin{figure*}[hbt!]
\centering
\includegraphics[width=17.8cm]{Figures/PredictTimeAhead.png}
\caption{The performance of multi-day ahead forecasting of the proposed logistic regression model: Histogram of the maximum days ahead of the correctly predicted patients of (a) all the 485 patients and (c) 110 patients in the external test set; The sample size, f1-score, accuracy in each day and cumulative f1-score and cumulative accuracy over the days of (b) all the the samples and (d) the samples in the external test set. Note there were three negative value records excluded in the figure as they were obtained after the patient was discharged or died.}
\label{fig:frog}
\end{figure*}
\subsection*{An Explainable Logistic Regression Model}
Based on the results in Table 1, we developed an explainable logistic regression model with two interaction items to predict the fatality rate of COVID-19 patients. The modeling of prediction using logistic regression models is transparent and the model can produce fatality probabilities between 0 and 1 rather than a binary value. Using $Y=1$ and $Y=0$ to indicate death and survival, respectively, we can formulate the logistic regression model in the following way \cite{hastie2009elements}:
\begin{equation} \label{eu_eqn}
l\ =\ ln\frac{P( Y=1)}{P( Y=0)} =ln\frac{p}{1-p} =\mathbf{B^{T}X} \ =\sum\limits _{i} \beta _{i} x_{i},
\end{equation}
where $\beta_{0}$ is a constant and $\beta_{i}, i=1,..., 5$, are the coefficients of $x_{1}$ (LDH), $x_{2}$ (lymphocyte), $x_{3}$ (hs-CRP), $x_{4}$ (LDH:lymphocyte), and $x_{5}$ (LDH:hs-CRP). We ran 100 rounds of five-fold cross-validation with random search to identify the optimal values of the regularization term $C$ between 0.0001 and 1000 with two types of penalty, i.e., $l1$ and $l2$. This process resulted in 500 sets of the coefficients and we used the median values as the coefficients of the final model in a vector form: $\mathbf{B^{T}} = [-4.976, 1.440e-2, -3.053e-1, 4.378e-2, 4.766e-4, -6.748e-5]$. We also optimized the prediction performance of the model by adjusting the threshold of the death probability both for the external test data and multi-day ahead forecasting below. We found the optimal threshold was 0.8. In other words, when one's fatality probability was larger than 0.8, the model predicted that a patient would die and the model had the best performance.
The model in Eq.(1) with the identified threshold was then used to predict the outcomes of 110 patients in the external test set that was not used to build the model. Although the data set was rather unbalanced (13 deaths and 97 survivals), the performance of the proposed logistic regression model was promising with 96 true negatives (1 false negative) and 12 true positives (1 false positive). The accuracy, f1-score, and AUC were $98.18\%$, $92.38\%$ and 0.996, respectively.
\subsection*{Multi-Day Ahead Forecasting}
Since there were multiple records of the three biomarkers for each patient, the model was used to forecast patient outcomes multi-days in advance. The samples were obtained by examining the records within each day for each patient. A total number of 909 records was obtained for all the 485 patients with 251 records for the 110 patients in the external test set. For multi-day ahead forecasting, we aimed to obtain the maximum days in advance. For example, if there are $N_{i}$ records in total for patient $i$ and the model was able to predict $m, m<=M$ days ahead, the predicted outcomes and the ground truth of the following maximum $n_{i}, 1<=n_{i}<=N_{i}$ consecutive records need to be the same. Here $m$/$M$ are the days between the dates of the $n_{i}$-th/first records and the date of the final outcome, respectively. For all the records, the model was able to predict 11.30 days (maximum = 34.91) ahead on average with a cumulative f1-score of 93.76\% (see Fig.1 (a) and (b)) and a cumulative accuracy value of 93.92\%. For the 251 records in the external test set, the model was able to predict 12.76 (maximum = 33.15) days ahead on average with a cumulative f1-score of 95.73\% (see Fig.1 (c) and (d)) and a cumulative accuracy value of 96.47\%. Thus, the proposed model can potentially give doctors the time needed to treat the patient accordingly.
\section*{Discussion}
Built on \cite{yan2020interpretable}, we proposed an explainable, intuitive, and yet accurate prediction model using logistic regression by incorporating two interaction items among the three most important biomarkers, including LDH, lymphocyte (\%) and hs-CRP, which can be easily measured in hospitals. Our model used no extra input information, compared to the decision tree model in \cite{yan2020interpretable}. Unlike the binary decisions produced by the decision tree, the logistic regression model produced a probability of fatality for each patient, which is more consistent with human-friendly explanations of machine learning models \cite{molnar2020interpretable}. For example, one of the rules in the decision tree in \cite{yan2020interpretable} was that IF LDH $> 365 UI^{-1}$, THEN death. Such a binary prediction may be not very intuitive without telling the likelihood of death. As a value of $364 UI^{-1}$ might result in a significantly different fatality probability from a value of $ 64 UI^{-1}$.
Our model, on the other hand, always gave a probability of death and also identified a threshold probability at 0.8, above which the model predicted that the patient would die. Furthermore, our model also outperformed the decision tree model in terms of average maximum days to the outcome and the cumulative f1-score and accuracy in Fig. 1.
Such a model can offer the clinicians time to identify high-risk patients before they become critical. Thus, an appropriate treatment strategy for COVID-19 patients depending on their likelihood of death can be made using corresponding medical resources. This can potentially alleviate the shortages of critical medical resources in hospitals in the current situation.
However, the model was built on a relatively small sample size. More research is needed to further test and optimize the model, taking both explainability and prediction performance into account.
\section*{Materials and Methods}
\subsection*{Samples} The data was originally from \cite{yan2020interpretable}. The model construction was based on the data of 375 (174 died) patients collected between January 10, 2020 and February 18, 2020 from Tongji Hospitals, Wuhan, China. Of them, 24 of the patients had missing data in the three biomarkers and thus were excluded from analysis. The external test data set was collected from another 110 patients (13 died) between February 19, 2020 and February 24, 2020 from the same hospital. We reported the performance of the model using metrics, including AUC \cite{walter2005partial}, micro-ave f1-score \cite{zhou2011affect}, and accuracy \cite{zhou2020driver}.
\subsection*{Data Analysis}
Interaction items in logistic regression can potentially improve the performance of the model to a great extent \cite{levy2019don}. Hence, we first fitted a logistic regression model using the three most important biomarkers, i.e., LDH, lymphocyte (\%), and hs-CRP identified in \cite{yan2020interpretable} and further identified two interaction items that could be useful to improve the prediction model. Then, we added one item at a time to the logistic regression model with five-fold cross-validation for 100 rounds and verified the two identified interaction items did improve its prediction performance as shown in Table 1. In order to have a model with good generalizability, we used the median values of the coefficients produced from the 500 models when producing the results in Table 1. Finally, this model was used to predict the external test set and multi-day ahead forecasting.
\section*{Conclusions}
The proposed logistic regression model can effectively predict the outcomes of COVID-19 patients with fatality probabilities. The model is accurate, intuitive, and explainable with only three blood biomarkers and two of their interaction items as input, which can potentially help the doctors determine the best treatment route for COVID-19 patients with high risks and optimize the logistic planning in the medical systems around the world amid this COVID-19 pandemic.
\acknow{We thank the authors in \cite{yan2020interpretable} for
providing their data sets.}
\showacknow{}
|
\section{Introduction}
Control Barrier Functions (CBFs) have proven to be effective at enforcing safety in nonlinear systems. The goal of this paper is to introduce a new form of CBFs that are suitable for feedback systems where the controller is defined by a differential equation \cite{Isidori90}, i.e., integral control \cite{freeman1995robust,khalil2000universal,jiang2001robust}.
The motivation comes from a dynamically defined tracking controller that has displayed effective tracking convergence \cite{Wardi19}, but also can exhibit large overshoots in the input controls at early transient phases. We investigate a CBF-based approach to limit these overshoots, i.e., satisfy input bounds, while also satisfying safety constraints on the state. The results are presented in a more-general setting of pointwise constraints on input and state encoded by a novel form of CBF.
The general CBF framework (see \cite{Ames19} and references therein) requires solving an optimization problem at each time to enforce the CBF condition. If the dynamics of the state are control affine, this optimization problem is a convex quadratic program (QP), solvable in real time. Yet this real-time solvability hinges on the control affine nature of the dynamics. Moreover, when input constraints are also present, adding them to the QP can result in its infeasibility. This paper will address both of these existing limitations in the context of integral control.
This paper introduces a new class of CBFs: \emph{integral control barrier functions (I-CBFs)}. These barrier functions are developed in the context of systems with dynamically defined controllers, i.e., systems where the evolution of the state and input are described by an ordinary differential equation. As a result, I-CBFs are defined on both the state and input, allowing for the inclusion of the input in the safety conditions encoded by this function. A related formulation is presented in \cite{huang2019guaranteed}, where control-dependent CBFs are defined: these, unlike I-CBFs, are considered for inputs with bounded time-derivative and, importantly, the integration with nominal tracking controllers was not considered. In this paper, given CBFs and I-CBFs, we present a resulting controller that guarantees safety in state and input, while minimally modifying a nominal dynamically defined controller. Additionally, we explicitly explore systems with both state and input constraints, and demonstrate how these can be unified through the I-CBF framework.
\section{Background Material}
\label{sec:background}
This section summarizes established results regarding safety-critical control via control barrier functions (CBFs), and the tracking-control technique that will inspire the main result of this paper.
Herein, we consider a nonlinear control system defined by the differential equation:
\begin{equation}
\label{eqn:controlsys}
\dot{x}(t)=f(x(t),u(t)),
\end{equation}
where $t\geq 0$, the state is $x(t)\in \mathbb{R}^n$, the input is $u(t)\in \mathbb{R}^m$, and an initial state $x_{0}:=x(0)\in \mathbb{R}^n$ is given.
Assume that the function $f:\mathbb{R}^n\times \mathbb{R}^m\rightarrow \mathbb{R}^n$ is continuously differentiable,
and it
satisfies sufficient conditions for the existence of a unique solution to \eqref{eqn:controlsys} over all $t\geq 0$ for every bounded, piecewise-continuous control $u(t)$ and $x_{0}\in \mathbb{R}^n$.
\subsection{Control Barrier Functions}
\label{ssec:cbfs}
Control Barrier Functions (CBFs) ensure the forward invariance of a set $\S\subset\mathbb{R}^n$, in which case the system is considered \emph{safe}. The CBF framework was introduced in \cite{Ames14,Ames17} where it was applied to adaptive cruise control, and has since been applied to a variety of application domains, including: automotive safety \cite{jankovic2018robust,xu2015robustness}, robotics \cite{nguyen2016exponential,wang2016multi} and multi-robot systems \cite{Wang17,lindemann2018control}.
See \cite{Ames19} for a recent survey.
Note that alternative approaches to CBFs include reference governors \cite{gilbert2002nonlinear,nicotra2018explicit} and integral barrier Lyapunov functions \cite{tee2012control}; yet, the former is computationally more expensive, whereas the latter only applies to systems in strict-feedback form and does not explicitly consider input bounds.
Let $h:\mathbb{R}^n\rightarrow \mathbb{R}$ be a continuously-differentiable function
such that $0$ is a regular value. We will consider the set, $\S\subset\mathbb{R}^n$, given as the 0-superlevel set of $h$, i.e., $\S$ is defined by:
\begin{eqnarray}
\S & := & \{ x \in \mathbb{R}^n ~ : ~ h(x) \geq 0 \} , \nonumber\\
\partial \S & = & \{ x \in \mathbb{R}^n ~ : ~ h(x) = 0 \}, \nonumber\\
\mathrm{Int}(\S) & = & \{ x \in \mathbb{R}^n ~ : ~ h(x) > 0 \}. \nonumber
\end{eqnarray}
Consider a feedback control law applied to \eqref{eqn:controlsys} of the form:
\begin{equation}
\label{eqn:feedbackcont}
u(t)=k(x(t)),
\end{equation}
with $k:\mathbb{R}^n\rightarrow \mathbb{R}^m$ continuous. Under this controller, we say that $x(t)$ is {\it safe} if and only if
$x_0 \in \S$ implies that $x(t)\in \S$ for all $t \geq 0$, i.e., $\S$ is \emph{forward invariant}, i.e., $\S$ is safe.
The function $h$ is a {\it control barrier function (CBF)} if there exists a continuous extended class-${\cal K}$ function $\gamma:\mathbb{R}\rightarrow \mathbb{R}$ (monotone increasing, $\gamma(0)=0$) together with a feedback controller \eqref{eqn:feedbackcont} such that along every trajectory of the closed-loop system, the following CBF condition holds \cite{Ames17}:
\begin{equation}
\label{eqn:CBFcond}
\frac{d}{dt}h(x(t))+\gamma(h(x(t)))\geq 0.
\end{equation}
Additionally, if Eq. \eqref{eqn:CBFcond} is satisfied for all $t\geq 0$, then the set $\S$ is forward invariant and asymptotically stable \cite{Ames17}.
Now by \eqref{eqn:controlsys}, it follows that $\dot{h}$ has the following form:
\begin{equation}
\label{eqn:cbfcondcontrol}
\dot{h}(x(t),u(t)) =\frac{d}{dt}h(x(t))=\frac{\partial h}{\partial x}(x(t))f(x(t),u(t)).
\end{equation}
Therefore, the closed-loop system is safe if for every $t\geq 0$,
\begin{equation}
\frac{\partial h}{\partial x}(x(t))f(x(t),u(t))+\gamma(h(x(t)))\geq 0.
\end{equation}
If this inequality is not satisfied for a particular $t\geq 0$, then the control law can be modified as follows to guarantee the safety of $\S$. Given $x\in \mathbb{R}^n$, define the set
$K_{x}\subset \mathbb{R}^m$ by:
\begin{align}
K_{x}= & \left\{u\in \mathbb{R}^m : \frac{\partial h}{\partial x}(x)f(x,u)+\gamma(h(x))\geq 0\large \right\}.
\end{align}
Therefore, $u(t)$ given by:
\begin{equation}
\label{eqn:optimizationbasedcontrol}
u(t)\in \underset{u \in \mathbb{R}^{m}}{\operatorname{argmin}} \{||u-k(x(t))||^2~:~u\in K_{x(t)}\}
\end{equation}
ensures the safety of the closed-loop system while modifying the control law defined in \eqref{eqn:feedbackcont} in a minimally invasive fashion.
Finally, note that if the dynamics of the plant are control-affine, namely
$f(x,u)=f_{0}(x)+f_{1}(x)u$ for functions $f_{0}:\mathbb{R}^n\rightarrow \mathbb{R}^n$
and $f_{1}:\mathbb{R}^n\rightarrow \mathbb{R}^{n\times m}$, then $u(t)$, defined by \eqref{eqn:optimizationbasedcontrol}, can be computed by a quadratic program (QP). In particular, we obtain an example of the feedback controller $u = k_{\mathrm{s}}(x)$ that ``filters'' the controller in \eqref{eqn:feedbackcont} and renders the system safe:
\begin{align}
\label{eqn:CBFQP}
k_{\mathrm{s}}(x) = \underset{u \in \mathbb{R}^{m}}{\operatorname{argmin}} & ~ ||u-k(x)||^2 \\
\mathrm{s.t.} & ~ \frac{\partial h}{\partial x}(x) f_0(x) +
\frac{\partial h}{\partial x}(x) f_1(x) u \geq - \gamma(h(x)), \nonumber
\end{align}
where we suppressed the dependence on $t$.
This QP admits a real-time implementation (see \cite{Ames14}).
Importantly, this paper will allow for quadratic program representations of safety-critical CBF-based controllers even when the plant is not control affine.
\subsection{Tracking Control by a Newton-Raphson Flow}
In this paper, we consider a tracking-control technique based on a flow version of the Newton-Raphson method for solving algebraic equations \cite{Wardi19}. This provides us with a convenient expression of a feedback control law defined by an ordinary differential equation, as it is recalled in this section (a more-detailed discussions and analyses can be found in \cite{Wardi19}). It is worth noting, however, that the main results presented in this paper do not depend on this specific method employed. We consider a specific method to define an integral control law that achieves tracking for definiteness.
To explain the main idea behind the tracking controller utilized, consider first the simple case where the plant is a memoryless nonlinearity, input-output system of the form $y(t)=g(u(t))$,
where $u(t)\in \mathbb{R}^m$ is the input, $y(t)\in \mathbb{R}^m$ is the output, and $g:\mathbb{R}^m\rightarrow \mathbb{R}^m$ is a continuously-differentiable function.
Given a continuously-differentiable reference signal $r(t)\in \mathbb{R}^m$,
suppose that the objective is to design a control law such that $y(t)$ converges to $r(t)$ in a suitable sense as described below. Consider the dynamically defined controller:
\begin{equation}
\label{eqn:udotnomem}
\dot{u}(t)=\Big(\frac{\partial g}{\partial u}(u(t))\Big)^{-1}\big(r(t)-y(t)\big).
\end{equation}
This controller essentially implements a Newton-Raphson (NR) flow whose vector field at time $t$, $\dot{u}(t)$, is the direction defined by the classical NR method for
solving the time-dependent algebraic equation: $r(t)-g(u)=0$.
Next, suppose that the plant is dynamic, i.e., defined by the state equation \eqref{eqn:controlsys}, and consider the output equation $y(t)=\zeta(x(t))$,
where the function $\zeta:\mathbb{R}^n\rightarrow \mathbb{R}^m$ is continuously differentiable. Now $x(t)$, and hence $y(t)$, are functions of the initial condition $x_0$ and past control actions: $u(\tau),~\tau\in[0,t)$. Yet, instantaneously, $y(t)$ is not a function of $u(t)$.
Therefore the controller $u(t)$ cannot be defined by an equation like \eqref{eqn:udotnomem}.
However, given $t\geq 0$ and $T> 0$, $x(t+T)$ and hence $y(t+T)$ depend on $u(\tau),~\tau\in[t,t+T]$. This observation forms the basis for our tracking controller.
Given a time $t \geq 0$ and $T > 0$, fixing $u_t := u(t)$ as a constant over $[t,t + T]$ and (approximately) forward integrating \eqref{eqn:controlsys}, i.e., $\dot{\hat{x}}(\tau) = f(\hat{x}(\tau),u_t)$, over this interval with initial condition $\hat{x}(t)= x(t)$ results in prediction of the state $\hat{x}(t + T)$ and hence a prediction of the output:
\begin{equation}
\label{eqn:predictor}
\hat{y}(t+T)= \zeta(\hat{x}(t + T)) =: g(x(t),u(t))
\end{equation}
that, therefore, depends on $x(t)$ and $u(t)$.
In this case, a suitable extension of \eqref{eqn:udotnomem} has the following form:
\begin{equation}
\label{eqn:controleqn}
\dot{u}(t)=\alpha\Big(\frac{\partial g}{\partial u}(x(t),u(t)\Big)^{-1}\big(r(t)-\hat{y}(t+T)\big);
\end{equation}
where $\alpha > 0$ is a controller gain.
In \cite{Wardi19}, the following convergence result was established:
\begin{equation}
\underset{t\rightarrow\infty}{\operatorname{limsup}}
||r(t)-y(t)||<\eta_{1}+\eta_{2}/\alpha,
\end{equation}
where $\eta_{1}:={\rm limsup}||y(t)-\hat{y}(t)||$, and $\eta_{2}:={\rm limsup}||\dot{r}(t)||$, where limsup means as $t\rightarrow\infty$. Thus, increasing the controller gain, $\alpha$, can reduce the error due variations in $r(t)$, but cannot attenuate the effects of prediction errors.
\section{Integral Control Barrier Functions}
\label{sec:mainresult}
This section presents the main construction and results of the paper. Specifically, we introduce integral control barrier functions, and demonstrate that they can enforce safety for dynamically defined control laws. This, in essence, creates a new paradigm for safety-critical integral control.
Consider a dynamical system defined by Eq. \eqref{eqn:controlsys} with an initial condition $x_{0}:=x(0)\in \mathbb{R}^n$. In a departure from the type of controller considered in Section II.A, which is algebraic, we define a general feedback law by the ordinary differential equation:
\begin{equation}
\label{eqn:dyncontrolequation}
\dot{u}(t)=\phi(x(t),u(t),t),
\end{equation}
with an initial condition $u_{0}:=u(0)\in \mathbb{R}^m$.
We assume that the function
$\phi:\mathbb{R}^n\times \mathbb{R}^m\times \mathbb{R}\rightarrow \mathbb{R}^m$ is continuously differentiable. The closed-loop system is defined by
the state equation \eqref{eqn:controlsys} and the control equation \eqref{eqn:dyncontrolequation}.
We write these two equations jointly in the following way:
\begin{equation}
\label{eqn:stateeqn}
\left[\begin{array}{l}
\dot{x}(t)\\
\dot{u}(t)
\end{array}
\right]~=~
\left[
\begin{array}{l}
f(x(t),u(t))\\
\phi(x(t),u(t),t)
\end{array}
\right].
\end{equation}
Define $z(t):=(x(t)^{\top},u(t)^{\top})^{\top}\in \mathbb{R}^n\times \mathbb{R}^m$ with $z(t)$ the {\it augmented state}, which can be viewed as the state of the closed-loop system. Note that the state equation \eqref{eqn:stateeqn} has no external input.
Let $\S\subset \mathbb{R}^n\times \mathbb{R}^m$ be a set, the {\it safety set}, defined as the 0-superlevel set of a continuously differentiable function $h:\mathbb{R}^{n}\times \mathbb{R}^m\rightarrow \mathbb{R}$ with $0$ a regular value. In addition, suppose the existence of a continuous extended class-${\cal K}$ function $\gamma:\mathbb{R}\rightarrow \mathbb{R}$ such that the following is satisfied along every trajectory $z(t)$ of the closed-loop system:
\begin{equation}
\label{eqn:zCBFcond}
\frac{d}{dt}h(z(t))+\gamma(h(z(t)))\geq 0.
\end{equation}
Then, the set $\S$ is forward invariant and asymptotically stable.
Throughout the remainder of this paper, for the sake of simplicity of exposition, we will suppress the dependence on time---this can easily be inferred from context.
For notational simplicity, define the following vector valued $p(x,u) \in \mathbb{R}^m$ and scalar valued $d(x,u) \in \mathbb{R}$ functions:
For given $t\geq 0$, define
the scalar $d $ and vector $p \in \mathbb{R}^m$ by:
\begin{align}
\label{eqn:pxu}
p(x,u) := & \Big(\frac{\partial h}{\partial u}(x ,u )\Big)^{\top} \\
\label{eqn:pdxut}
d(x,u,t) := & -\Big(\frac{\partial h}{\partial x}(x ,u )f(x ,u ) +\frac{\partial h}{\partial u}(x ,u )\phi(x ,u ,t) \nonumber\\
& \qquad \qquad +\gamma(h(x ,u ))\Big),
\end{align}
Note that Eq. \eqref{eqn:zCBFcond} can thus be recast as $d(x,u,t) \leq 0$.
\newsec{Main Result.}
Traditional CBF methods cannot be directly applied to systems of the form given in \eqref{eqn:stateeqn} due to the dynamically defined control law, i.e., the control law $u$ is a result of integrating the augmented state equation in \eqref{eqn:stateeqn}.
Our approach is to add an auxiliary input to the state equation so that, if the inequality in \eqref{eqn:zCBFcond} is not satisfied for a particular $t\geq 0$, we are able to determine the minimal modification of the dynamically defined control law that will guarantee safety. To this end, we modify the augmented-state equation
\eqref{eqn:stateeqn} to include the input $v\in \mathbb{R}^m$ as follows:
\begin{equation}
\label{eqn:stateeqnv}
\left[\begin{array}{l}
\dot{x}\\
\dot{u}
\end{array}
\right]~=~
\left[
\begin{array}{l}
f(x,u)\\
\phi(x ,u ,t)+v
\end{array}
\right].
\end{equation}
For every $z:=(x^{\top},u^{\top})^{\top}\in \mathbb{R}^n\times \mathbb{R}^m$
and $t\geq 0$, define:
\begin{align}
\label{eqn:Szset}
K_{z,t}:= & \big\{v\in \mathbb{R}^m~
:~\frac{\partial h}{\partial x}(x,u)f(x,u) \\
&+\frac{\partial h}{\partial u}(x,u)\big(\phi(x,u,t)+v\big) ~ +\gamma(h(x,u))~\geq~0\big\} \nonumber\\
\label{eqn:Szsetshort}
=& \left\{ v\in \mathbb{R}^m~
:~ p(x,u)^{\top} v \geq d(x,u,t) \right\}.
\end{align}
Therefore, if $v(x,t) \in K_{z,t}$, it follows that Eq. \eqref{eqn:zCBFcond} is satisfied. Additionally, if the dynamic control law $\dot{u} = \phi(x,u,t)$ is inherently safe, then $d(x,u,t) \leq 0$ and $v = 0$ imply that the system is safe.
This leads to the formulation of I-CBFs,
mirroring the classic definition of control Lyapunov functions \cite{sontag1989universal}) and similar in spirit to control-dependent CBFs \cite{huang2019guaranteed}.
\begin{definition}
For the system \eqref{eqn:stateeqnv}, with corresponding safe set $\S \subset \mathbb{R}^n \times \mathbb{R}^m$ defined as the 0-superlevel set of a function $h : \mathbb{R}^n \times \mathbb{R}^m \to \mathbb{R}$ with $0$ a regular value: $\S = \{(x,u) \in \mathbb{R}^n \times \mathbb{R}^m ~ : ~ h(x,u) \geq 0\}$. Then $h$ is an \emph{integral control barrier function (I-CBF)} if for any $(x,u) \in \mathbb{R}^n \times \mathbb{R}^m$ and $t \geq 0$:
\begin{eqnarray}
\label{eqn:relativedegree}
p(x,u) = 0 \qquad \Rightarrow \qquad d(x,u,t) \leq 0.
\end{eqnarray}
\end{definition}
We now have the necessary constructions to state the main result of this paper.
\begin{theorem}
\label{thm:main}
Consider the control system $\dot{x} = f(x,u)$, with $x \in \mathbb{R}^n$ and $u \in \mathbb{R}^m$, and suppose that there is a corresponding dynamically defined controller: $\dot{u} = \phi(x,u,t)$. If the safe set $\S \subset \mathbb{R}^n \times \mathbb{R}^m$ is defined by an integral control barrier function, $h : \mathbb{R}^n \times \mathbb{R}^m \to \mathbb{R}$,
then modifying the dynamically defined controller to be of the form:
\begin{eqnarray}
\label{eqn:umodifiedopt}
\dot{u} = \phi(x,u,t) + v^*(x,u,t)
\end{eqnarray}
with $v^*$ the solution to the QP:
\begin{align}
\label{eqn:CBFQPv}
\qquad \qquad \quad v^*(x,u,t) = \underset{v \in \mathbb{R}^{m}}{\operatorname{argmin}} & ~ ||v||^2 \\
\mathrm{s.t.} & ~ p(x,u)^{\top} v \geq d(x,u,t) \nonumber
\end{align}
results in safety, i.e., the control system $\dot{x} = f(x,u)$ with the dynamically defined controller \eqref{eqn:umodifiedopt} results in $\S$ being forward invariant: if $(x(0),u(0)) \in \S$ then $(x(t),u(t)) \in \S$ for all $t \geq 0$.
\end{theorem}
\begin{proof}
We only need to verify that Eq. \eqref{eqn:zCBFcond} is satisfied for $z(t) = (x(t),u(t))$ the solution to \eqref{eqn:stateeqnv} with $v = v^*(x,u,t)$. Per \eqref{eqn:Szset} and \eqref{eqn:Szsetshort}, this is equivalent to:
\begin{eqnarray}
\label{eqn:proofmaineqn}
p(x,u)^{\top} v^*(x,u,t) \geq d(x,u,t).
\end{eqnarray}
This, in turn, will imply that $(x(t),u(t)) \in \S$ for all $t \geq 0$ if $(x(0),u(0)) \in \S$ per the main result of \cite{Ames19}.
Thus, to establish the result, we must show the solution to \eqref{eqn:CBFQPv} satisfies \eqref{eqn:proofmaineqn}. This follows from the fact that we can obtain an explicit solution to \eqref{eqn:CBFQPv}. In particular, the condition that $h$ is an integral control barrier function and, in particular, that it satisfies \eqref{eqn:relativedegree}, implies that the linear independent constraint qualification condition is satisfied \cite{boyd2004convex}. Therefore, using the KKT optimality conditions (see \cite{xu2015robustness}), the solution to \eqref{eqn:CBFQPv} is given by the \emph{min-norm} controller:
\begin{align}
\label{eqn:minnormcon}
v^*(x,u,t) = \left\{
\begin{array}{lcr}
\frac{d(x,u,t)}{\| p(x,u) \|^2} p(x,u) & \mathrm{if~} d(x,u,t) > 0 \\
0 & \mathrm{if~} d(x,u,t) \leq 0
\end{array}
\right.
\end{align}
This controller is well-defined because, as $h$ is a control barrier function and thus satisfies \eqref{eqn:relativedegree}, it follows that: $d(x,u,t) > 0$ implies that $p(x,u) \neq 0$. Additionally, from this explicit form, one can verify the Lipschitz continuity of this controller (assuming Lipschitz continuity of $p$ and $d$).
Finally, the explicit form of \eqref{eqn:minnormcon} makes it clear that it satisfies \eqref{eqn:proofmaineqn}. If $d(x,u,t) \leq 0$ than it is naturally satisfied with $v^*(x,u,t) = 0$. If $d(x,u,t) > 0$, then:
\begin{eqnarray}
\quad p(x,u)^{\top} v^*(x,u,t) & = & \underbrace{p(x,u)^{\top} p(x,u)}_{\| p(x,u) \|^2} \frac{d(x,u,t)}{\| p(x,u) \|^2} \nonumber\\
& = & d(x,u,t). \hspace{2.6cm} \hfill \qedhere \nonumber
\end{eqnarray}
\end{proof}
\begin{remark}
A natural consequence of this form of control barrier functions is, as its name suggests, its application to integral control. In the case of ``pure'' integral control, we can consider a control system \eqref{eqn:stateeqnv} together with $\phi(x,u,t) \equiv 0$, i.e., the dynamic extension via the addition of an integrator. In this case, the controller from Theorem \ref{thm:main} is just $u(t)~=~\int_{0}^t v^*(x(\tau),u(\tau),\tau) d \tau$.
\end{remark}
\begin{remark}
The controller given in Theorem \ref{thm:main} can be viewed in the following fashion. We began with a control system $\dot{x} = f(x,u)$ for which we synthesized a dynamic controller:
\begin{align*}
\dot{u} = \phi(x,u,t) + \left\{
\begin{array}{lcr}
\frac{d(x,u,t)}{\| p(x,u) \|^2} p(x,u) & \mathrm{if} & d(x,u,t) > 0 \\
0 & \mathrm{if} & d(x,u,t) \leq 0
\end{array}
\right.
\end{align*}
with $\phi(x,u,t)$ the ``feedforward'' integral controller that is modulated by the additional term to ensure safety through a minimal modification. The controller, $\phi$, for example, can be given by \eqref{eqn:controleqn}. The advantages of the integral instantiation of control barrier functions are: they allow CBFs to be applied to systems not in control affine form; the dynamic equation describing $u$ is integrated thus smoothing out the non-smooth nature of solutions to QPs; they can encode input bounds.
\end{remark}
\begin{remark}
\label{rmk:staticfeedback}
We can also consider the case when we have a nominal controller $u = k(x)$ as in \eqref{eqn:feedbackcont}. In this case, if we pick the following ``feedforward'' integral controller:
$$
\phi(x,u,t) = \frac{\partial k}{\partial x}(x) f(x,u) + \frac{\alpha}{2}\left(k(x,t)-u\right),
$$
for $\alpha > 0$, we provably get that $\| u - k(x) \| \to 0$ exponentially. This can, therefore, be coupled with Theorem \ref{thm:main} to achieve provable tracking of the desired controller subject to safety. Proving this is not within the scope of this paper, but will be the subject of future work---as it indicates the ability to safely track desired controllers even for non-affine control systems.
\end{remark}
\section{Applications and Extensions of I-CBFs}
In this section, we explore some extensions and applications of integral control barrier functions. In particular, we demonstrate that I-CBFs can encode both input bounds and state constraints. Importantly, we show that I-CBFs can simultaneously enforce both, thereby enabling input bounded provably safe integral control subject to feasibility.
\subsection{Application to Input Constrained Systems}
Consider the case when we have a system $\dot{x} = f(x,u)$ subject to input constraints: for $u_{\max} > 0$,
$$
\| u \|^2 \leq u_{\max} \qquad \Rightarrow \qquad h_u(u) := u_{\max} - u^{\top} u \geq 0.
$$
Naturally, this is a conservative method for enforcing input constraints due to its scalar instantiation. Yet, it allows us to view input bounds as safety relative to the set $\S_u := \{ (x,u) \in \mathbb{R}^n \times \mathbb{R}^m ~ : ~ h_u(u) \geq 0\}$.
In this case, \eqref{eqn:pxu} and \eqref{eqn:pdxut} become:
\begin{eqnarray}
\label{eqn:pxu2}
p_u(x,u) & = & - 2 u \nonumber\\
\label{eqn:pdxut2}
d_u(x,u,t) & = & 2 u^{\top} \phi(x ,u ,t) - \gamma_u(h_u(u )). \nonumber
\end{eqnarray}
Thus, following from Theorem \ref{thm:main}:
\begin{lemma}
\label{lem:torquebound}
For the system \eqref{eqn:stateeqnv}, $h_u$ is an integral control barrier function, i.e., $\S_u$ can be rendered safe (forward invariant), with the integral control law \eqref{eqn:umodifiedopt} where:
\begin{align}
\label{eqn:CBFQPvubound}
v^*(x,u,t) = \underset{v \in \mathbb{R}^{m}}{\operatorname{argmin}} & ~ ||v||^2 \\
\mathrm{s.t.} & ~ 2 u^{\top} v \geq - 2 u^{\top} \phi(x ,u ,t) - \gamma_u(h_u(u ))
\nonumber
\end{align}
\end{lemma}
\begin{proof}
We only need to verify \eqref{eqn:relativedegree}. If $p_u(x,u)= 2u = 0$, then $u = 0$, and therefore $d_u(x,u,t) = - \gamma_u(h_u(u))$. But $u = 0$ implies that $(x,u) \in \S_u$, and therefore $h_u(u) \geq 0$ or $\gamma_u(h_u(u)) \geq 0$, and so $d_u(x,u,t) \leq 0$ as desired.
\end{proof}
\subsection{Application to State Constrained Control Affine Systems}
\label{ssec:affine}
Returning to the systems that are often considered in the context of control barrier functions---affine control systems---we wish to understand the relationship between classical and integral control barrier functions for systems of this form.
Consider the variation of \eqref{eqn:stateeqnv} in control affine form:
\begin{eqnarray}
\label{eqn:stateeqnaffine}
\dot{x} & = & f_0(x) + f_1(x) (\mu + u) \\
\dot{u} & = & \phi(x,u,t) \nonumber
\end{eqnarray}
where here we added an auxiliary control law in the $x$ dynamics---the reason for this will become apparent soon.
For this system suppose we have a safe set defined in terms of state (and not input), i.e., by $h_x : \mathbb{R}^n \to \mathbb{R}$ with:
$$
\S_x := \{ (x,u) \in \mathbb{R}^n \times \mathbb{R}^m ~ : ~ h_x(x) \geq 0\}.
$$
Additionally, suppose that $h_x$ is a valid control barrier functions for the affine dynamics without the corresponding integrator: $\dot{h}_x(x,u) \geq -\gamma_x(h_x(x))$.
This can be quantified by a conditions analogous to \eqref{eqn:relativedegree}, specifically:
\begin{align}
\label{eqn:hxCBFcond}
\underbrace{\frac{\partial h_x}{\partial x}(x) f_1(x)}_{:= p_x(x)^{\top}} = 0
\Rightarrow
\underbrace{
\frac{\partial h_x}{\partial x}(x) f_0(x) + \gamma_x(h_x(x)) }_{ := -d_x(x)} \geq 0
\end{align}
The following lemma gives a controller that will guarantee the forward invariance of $\S_x$ under the assumption that $h_x$ is a valid CBF. It will do so through a combined ``traditional'' CBF controller together with an integral controller.
\begin{lemma}
\label{lem:statebound}
Given the system \eqref{eqn:stateeqnaffine} with $h_x : \mathbb{R}^n \to \mathbb{R}$ a control barrier function for $\dot{x} = f_0(x) + f_1(x) u$. Then the controller:
$$
k_{\mathrm{s}}(x,u,t) = \mu^*(x(t),u(t)) + \int_{0}^{t} \phi(x(\tau),u(\tau),\tau) d \tau
$$
where, for any nominal controller $u = k(x)$:
\begin{align}
\label{eqn:CBFQPint}
\mu^*(x,u) = \underset{\mu \in \mathbb{R}^{m}}{\operatorname{argmin}}& ~ ||\mu +u -k(x)||^2 \\
\mathrm{s.t.} ~ & ~
p_x(x)^{\top} \mu \geq d_x(x) - p_x(x)^{\top} u \nonumber
\end{align}
renders the set $\S_x$ forward invariant, i.e., safe.
\end{lemma}
\begin{proof}
By \eqref{eqn:hxCBFcond}, the QP \eqref{eqn:CBFQPint} is well-defined and has a solution. This follows from the fact that $u$ not being a decision variable anymore does not affect \eqref{eqn:hxCBFcond}, i.e., $p_x(x) = 0$ still implies that $d_x(x) \leq 0$ since the term $-p_x(x) u$ vanishes. Thus safety, i.e., forward invariance of $\S_x$, is guarenteed by the classic CBF result \cite{Ames17}.
\end{proof}
\subsection{Extension to Multiple Control Barrier Functions}
With the goal of now simultaneously enforcing state and input constraints in a holistic fashion,
we will begin by taking inspiration from Section \ref{ssec:affine}, wherein we must achieve similar results for systems with only integral control.
Returning to original state equations \eqref{eqn:stateeqnv}, but in affine form:
\begin{eqnarray}
\label{eqn:stateeqnaffinetwop}
\dot{x} & = & f_0(x) + f_1(x) u \\
\dot{u} & = & \phi(x,u,t) + v, \nonumber
\end{eqnarray}
consider state constraints encoded by $\S_x$, i.e., by $h_x(x) \geq 0$. The input $v$ no longer appears in $\dot{h}_x$, but, the forward invariance of the set $\S_x$ is quantified by the condition $\dot{h}_x(x,u) + \gamma_x(h_x(x)) \geq 0 $. Thus, in order to ensure the latter inequality is satisfied, we follow the approach in \cite{nguyen2016exponential} and let:
\begin{eqnarray}
\label{eqn:heeqn}
h_e(x,u) := \dot{h}_x(x,u) + \gamma_x(h_x(x)) = p_x(x)^{\top} u - d_x(x).
\end{eqnarray}
The result is a function: $h_e : \mathbb{R}^n \times \mathbb{R}^m \to \mathbb{R}$ with corresponding set: $\S_e = \{ (x,u) \in \mathbb{R}^n \times \mathbb{R}^m ~ : ~ h_e(x,u) \geq 0\}$.
Let $p_e(x,u)$ and $d_e(x,u,t)$ be the corresponding functions defined from $h_e$, for some $\gamma_e$, as in \eqref{eqn:pxu} and \eqref{eqn:pdxut}.
\begin{lemma}
\label{lem:cbftoicbf}
Given the system \eqref{eqn:stateeqnaffinetwop} with $h_x : \mathbb{R}^n \to \mathbb{R}$ a control barrier function with $p_x(x) \neq 0$, i.e., $h_x$ has relative degree 1, then $h_e(x,u) = \dot{h}_x(x,u) + \gamma_x(h(x))$ is an integral control barrier function. Additionally, for the I-CBF control law \eqref{eqn:CBFQPv}, using $p_e$ and $d_e$ in place of $p$ and $d$, if:
$$
\begin{array}{r}
x_0 \in \S_x \\
(x_0,u_0) \in \S_e
\end{array}
\quad \Rightarrow \quad
\begin{array}{rr}
x(t) \in \S_x & \quad \forall~ t \geq 0 \\
(x(t),u(t)) \in \S_e & \quad \forall~ t \geq 0
\end{array}
$$
That is, $\S_x$ is safe subject to appropriate initial conditions.
\end{lemma}
\begin{proof}
It is easy to verify that $p_e(x,u) = p_x(x)$, thus the relative degree condition transfers from $h$ to $h_e$. As a result, the controller in \eqref{eqn:CBFQPv} is well-defined, i.e., \eqref{eqn:relativedegree} is trivially satisfied. As a result, $\S_e$ is forward invariant, i.e., assuming an initial condition $(x_0,u_0) \in \S_e$ then $\dot{h}_e(x(t),u(t)) = \dot{h}_x(x(t),u(t)) + \gamma(h_x(x(t))) \geq 0$. Coupling this with the assumption that $x_0 \in \S_x$, i.e, that $h_x(x_0) \geq 0$, it follows that $\S_x$ is also forward invariant.
\end{proof}
\newsec{Combining State Constraints with Input Bounds.} With the I-CBF $h_e$, obtained from $h_x$ via Lemma \ref{lem:cbftoicbf}, we can synthesize controllers that
enforce state constraints with input bounds.
\begin{theorem}
\label{thm:maintwo}
Consider the system in \eqref{eqn:stateeqnaffinetwop}, together with corresponding CBF $h_x : \mathbb{R}^n \to \mathbb{R}$ and I-CBF $h_u : \mathbb{R}^m \to \mathbb{R}$, and safe sets $\S_x$ and $\S_u$ respectively, and let $h_e : \mathbb{R}^n \times \mathbb{R}^m$ be given as in \eqref{eqn:heeqn}. If the following QP is feasible:
\begin{align}
\label{eqn:CBFQPintfinal}
v^*(x,u,t)
= &\underset{v\in \mathbb{R}^{m}}{\operatorname{argmin}} ~ ||v||^2 \\
& \quad \mathrm{s.t.} \quad
\begin{bmatrix}
p_x(x)^{\top} \\
p_u(x,u)^{\top}
\end{bmatrix}
v \geq
\begin{bmatrix}
d_e(x,u,t) \\
d_u(x,u,t)
\end{bmatrix}
\nonumber
\end{align}
then the dynamically defined (integral) controller
$$\dot{u}~=~\phi(x,u,t) + v^*(x,u,t)$$
renders the set $\S = \S_x \cap \S_u$ forward invariant, i.e., safe, for appropriate initial conditions:
$$
\begin{array}{r}
x_0 \in \S_x \\
(x_0,u_0) \in \S_e \cap \S_u
\end{array}
\quad \Rightarrow \quad
\begin{array}{r}
(x(t),u(t)) \in \S_x \cap \S_u \\
\forall ~ t \geq 0.
\end{array}
$$
\end{theorem}
\begin{proof}
The constraints in \eqref{eqn:CBFQPintfinal} are simply reformulations of:
\begin{eqnarray}
\hspace{-0.1cm}
\begin{array}{l}
\dot{h}_e(x,u, v) \geq - \gamma_e(h_e(x,u)) \\
\dot{h}_u(x,u, v) \geq - \gamma_u(h_u(u))
\end{array}
\Leftrightarrow
\begin{array}{r}
p_x(x)^{\top} v \geq d_e(x,u,t) \\
p_u(x,u)^{\top} v \geq d_u(x,u,t)
\end{array} \nonumber
\end{eqnarray}
Therefore, assuming a feasible QP implies that $\S_e \cap \S_u$ is forward invariant by Theorem \ref{thm:main} and Lemma \ref{lem:torquebound}. The result follows by Lemma \ref{lem:cbftoicbf}.
\end{proof}
\begin{remark}
The requirement on the feasibility of the QP in Theorem \ref{thm:maintwo} encodes the fact that we are not assuming that $\S_x$ is a control invariant set. Without this assumption, we do not know if there are feasible control inputs rendering $\S_x$ invariant---this is encoded by the feasibility of the QP. Moreover, if the QP is feasible, conditions for the Lipschitz continuity of its solution can be found (see, e.g., \cite{Ames17}). This and the relaxation of the feasibility assumption will be the subject of future work.
\end{remark}
\section{Simulation Results}
To demonstrate the results of the paper, we will return to an early motivating example considered for CBFs \cite{Ames14}: \emph{adaptive cruise control (ACC).}
The dynamics are given by:
\begin{eqnarray}
\label{eqn:accdynamics}
\dot{x} =
\begin{bmatrix}
x_2 \\
- \frac{1}{m} F_r(x) \\
v_0 - x_2
\end{bmatrix}
+
\begin{bmatrix}
0 \\
\frac{1}{m} \\
0
\end{bmatrix} u
\end{eqnarray}
where $(x_1,x_2)$ are the position and velocity ($x_2 = \dot{x}_1$) of the vehicle, $m$ its mass, $x_3$ is the distance between the vehicle and the lead vehicle traveling at a velocity of $v_0$, and $F_r(x) = c_0 + c_1 x_2 + c_2 x_2^2$ is the empirical form of rolling resistance. The control objective for the system is to drive the car to a desired speed ($x_2 \to v_d$). This can be represented as an output: $y = \zeta(x) = x_2 - v_d$. To obtain the predictor as in \eqref{eqn:predictor}, we can forward integrate \eqref{eqn:accdynamics} with $c_2 = 0$, i.e., the linear approximation, resulting in:
\[
\hat{y}(t + T) =
{\scriptstyle
-c_{1}^{-1}\left(c_{0}-u(t)+ m v_d-c_{1} e^{-\frac{c_{1} T }{m}} \left(x_2(t) +\frac{c_{0}-u(t)+m v_d}{c_{1}}\right)\right),
}
\]
yielding the dynamically defined control law per \eqref{eqn:controleqn}:
\[
\dot{u}(t) = \alpha c_1 \left( e^{\frac{-c_1}{m} T} - 1 \right)^{-1} \hat{y}(t + T) =: \phi(x,u,t).
\]
The state safety constraints encode the ``half the speedometer rule'' which yields the CBF (since $p_x(x) = -1.8/m \neq 0$):
\[
h_x(x) = x_3 - 1.8 x_2 \geq 0 \qquad \mapsto \qquad
\S_x = \{h_x(x) \geq 0\} .
\]
The input constraint is given by the constraint that the wheel force is bounded by $|u| \leq m c_{a/d} g $ yielding an I-CBF:
\[
h_u(x) = (m c_{a/d} g)^2 - u^2 \geq 0
\quad \mapsto \quad
\S_u = \{h_u(u) \geq 0\} .
\]
\noindent where $c_{a/d}$ is the factor of $g$ for acceleration/deceleration. In both cases, we pick $\gamma_x(r) = \gamma_u(r) = \gamma r$ for $\gamma > 0$. From $h_x$ we get $h_e$ as in \eqref{eqn:heeqn}, and we pick $\gamma_e(r) = \frac{1}{2} \gamma r$.
\twocolumn[%
{\centering\includegraphics[width=\textwidth]{figures/final_fig4.pdf}
\captionof{figure}{Methods developed in the paper applied to adaptive cruise control: the goal is for a vehicle to reach a desired velocity ($v_d = 24 m/s$) while not colliding with the lead vehicle traveling slower (at $14 m/s$) and not exceeding a maximum wheel force (input bounds).}
\label{fig:finalfig}}
\vspace{0.5cm}
]
Utilizing the constructions leading to Theorem \ref{thm:maintwo} results in the forward invariance of $\S_x \cap \S_u$. This is illustrated in Fig. \ref{fig:finalfig}, where the parameters and initial condition where chosen to match \cite{Ames14} with $\alpha = 10$ and $\gamma = 1$ above.
Three different controllers are plotted in Fig. \ref{fig:finalfig}, including: only input constraints and thus an I-CBF, $h_u$, with the controller from Lemma \ref{lem:torquebound} (red), only state constraints and thus a CBF, $h_x$, with the controller from Lemma \ref{lem:statebound} (green), and both input and state constraints, thus enforcing both $h_u$ and $h_e$, via the controller from Theorem \ref{thm:maintwo} (blue). Therefore, via the presented unified controller, we are able to simultaneously satisfy the input and state constraints, and thus render the system safe for both constraints, i.e., render $\S_x \cap \S_u$ forward invariant. Importantly, when compared against past uses of this example in \cite{Ames19,Ames14} we can do so holistically.
\section{Conclusions}
This paper introduced integral control barrier functions (I-CBFs). By considering dynamically defined controllers, we were able to guarantee safety using I-CBFs defined in terms of both state and input.
This was applied in the context of dynamically defined tracking controllers for general nonlinear control systems (not necessarily control affine), wherein I-CBFs lead to minimal modification of these controllers via I-CBF based QPs (Theorem \ref{thm:main}).
We then considered the specific cases of input bounds (Lemma \ref{lem:torquebound}) and state constraints (Lemma \ref{lem:cbftoicbf}), wherein both can be independently enforced via I-CBFs.
Additionally, we proved the state constraints and input bounds can be simultaneously satisfied via the framework of I-CBFs assuming a feasible QP (Theorem \ref{thm:maintwo}). This gives a holistic method for provably enforcing both through the use of safety-critical integral controllers.
Future work will be devoted to expanding this theoretic basis and demonstrating this theory experimentally. From a theoretic perspective, understanding when the QP in Theorem \ref{thm:maintwo} is feasible is a rich problem that has important implications.
Additionally, we wish to improve the means in which nominal controllers can be enforced in conjunction with integral control barrier functions.
From an experimentally perspective, the ability of I-CBFs to bound inputs while achieving safety in state has important ramifications that we wish to explore on everything from walking robots to multi-robot systems to safey-critical autonomy.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\input{content.tex}
\subsection*{Philosophy}
Type theories are typically presented by systems of natural deduction rules describing the inductive structure of the theory. Models of the type theory will therefore have corresponding structure. The observation that motivates our approach may be summarised by the following thesis.
\begin{center}\em
Natural deduction rules are syntax for polynomials.
\end{center}
\noindent In this paper we give an exposition of this idea, describing a polynomial approach to the semantics of simple type theories. Concretely, we will show how the natural deduction rules presenting the algebraic structure of a simple type theory, which are described precisely by a class of arities, induce polynomial functors in presheaf categories whose algebras are exactly models of the type theory. In particular, the initial algebras are the syntactic models, whose terms are inductively generated from the rules. This provides a correspondence between type theoretic and categorical structure. To build intuition for the general setting, we will frame the classical example of the simply-typed \lc{}{} in this new perspective.
The relationship between polynomials and the algebraic structure of type theories was first proposed by \textcite{fiore2012discrete}, in the context of generalised polynomial functors between presheaf categories. Though our approach is similarly motivated, the setting is different: we consider traditional polynomial functors between slice categories. This is a setting that has been more widely studied \cite{abbott2003categories, gambino2013polynomial} and one we suggest also extends more readily to modelling dependent type theories: \citeauthor{awodey2018polynomial} \cite{awodey2018natural, awodey2018polynomial, newstead2018algebraic}, for instance, have also considered a relationship between polynomial pseudomonads and natural models of type theory~\cite{awodey2018natural}. Their setting, however, is entirely semantic, and one in which the significance of polynomials in the structure of natural deduction rules is not apparent.
In this framework, we consider two classes of models: models of
\emph{simply typed syntax}~(\Cref{sec:models-syntax}), and models of
\emph{simple type theories}~(\Cref{sec:models-theories}). Both classes of models have algebraic type structure and multisorted binding (\latinabbr{i.e.}{}~second-order) algebraic term structure, but simple type theories extend syntax in two ways: while \emph{syntax} here refers to those terms solely built inductively from natural deduction rules, \emph{type theories} additionally have an associated notion of (capture-avoiding) substitution: a variable in a term may be replaced by a term of the same type, taking care not to bind any free variables. Typically, a syntax gives rise to a type theory, as one can add a substitution operation that commutes with the operators of the syntax. For this reason, many models of universal algebra do not draw a distinction between syntax and type theory: for instance, Lawvere theories~\cite{lawvere1963functorial} have a built-in notion of substitution, given by composition of morphisms. However, it is useful to consider these two notions separately: substitution gives rise to rich structure that one can only observe by treating it explicitly, for example the substitution lemma~(\Cref{thm:substitution-lemma}) that is ubiquitous in treatments of type theory.
We also consider only type theories (and not syntax) to be equational, as modelling equations involves identifying terms that are syntactically distinct.
\subsection*{Contributions}
The main contributions of this paper are the following.
\begin{enumerate}
\item A new perspective on natural deduction rules, presenting natural deduction rules for formation, introduction and elimination, as the syntax for polynomials in presheaf categories.
\item A general definition of models of simply typed syntax and simple type theories.
\item Initiality theorems, giving a construction of the initial models of simply typed syntax and simple type theories.
\item A correspondence
between models of simple type theories and classifying multicategories, generalising the classical Lambek correspondence between the simply-typed \lc{}{} and cartesian-closed categories.
\end{enumerate}
This work provides a basis for our ongoing development of algebraic dependent type theory.
\subsection*{Organisation of the paper}
We build up the definition of a simple type theory in parts, presenting the syntax and semantics in conjunction.
\Cref{sec:types} describes the monosorted nonbinding algebraic structure of types, which is standard from universal algebra.
\Cref{sec:contexts} considers variable contexts and introduces models thereof.
\Cref{sec:terms} is the central contribution of the paper and explains how the multisorted binding algebraic structure on terms may be presented by syntax for polynomials corresponding to natural deduction rules.
\Cref{sec:models-syntax} defines categories of models of simply typed syntax and gives a construction of the initial model (\Cref{thm:initiality-syntax}).
\Cref{sec:substitution} introduces substitution structure on terms and establishes a substitution lemma (\Cref{thm:substitution-lemma}).
\Cref{sec:equations} describes equations on terms, which crucially relies on the substitution structure from the preceding section.
\Cref{sec:models-theories} defines categories of models of simple type theories, which extend syntax by having substitution and equational structure, and
leads to a construction
of the initial model (\Cref{thm:initiality-theory}).
\Cref{sec:classifying-multicategories} demonstrates how models of simple type theories induce structured cartesian multicategories, establishing a generalised Lambek correspondence (\Cref{thm:equivalence-multicategories} and \Cref{cor:lambek}).
\section{Simple types}
\label{sec:types}
We consider types with monosorted nonbinding algebraic structure \`a la universal algebra. The type constructors of the simply-typed \lc{}{} are examples of such algebraic structure; consider the following formation rules.
\[\begin{prooftree}
\Infer0[$\mathsf{Unit}$-\textsc{form}]{\mathsf{Unit}\ \mathsf{type}}
\end{prooftree}\]
\[\begin{prooftree}
\Hypo{A\ \mathsf{type}}
\Hypo{B\ \mathsf{type}}
\Infer2[$\mathsf{Prod}$-\textsc{form}]{\mathsf{Prod}(A, B)\ \mathsf{type}}
\end{prooftree}\]
\[\begin{prooftree}
\Hypo{A\ \mathsf{type}}
\Hypo{B\ \mathsf{type}}
\Infer2[$\mathsf{Fun}$-\textsc{form}]{\mathsf{Fun}(A, B)\ \mathsf{type}}
\end{prooftree}\]
These types may be modelled by a set $S$ of sorts with a function expressing the denotations of the type constructors. Base types are described by nullary type constructors, as in universal algebra.
\[1 + S^2 + S^2 \xrightarrow{[\denop{Unit}, \denop{Prod}, \denop{Fun}]} S\]
This structure is an algebra for the endofunctor on $\cat{Set}$ mapping $S \mapsto S^0 + S^2 + S^2$. This is an example of a \emph{polynomial functor} on $\cat{Set}$. Polynomial functors are a categorification of the notion of polynomial functions and similarly represent ``sums of products of variables''. Just as a polynomial function is presented by a list of coefficients, polynomial functors are presented by \emph{polynomials}, which are diagrams of the following shape.
\[I \xleftarrow{s} A \xrightarrow{f} B \xrightarrow{t} J\]
Such a polynomial in $\cat{Set}$ induces a polynomial functor $\cat{Set}/I \to \cat{Set}/J$, given by the following, where $B_j = t^{-1}(j)$ and $A_b = f^{-1}(b)$.
\[(X_i \mid i \in I) \mapsto (\Sigma_{b \in B_j} \Pi_{a \in A_b} X_{s(a)} \mid j \in J)\]
This is slightly more sophisticated than the traditional sum of products: in particular, we also have a notion of reindexing. Clear introductions to polynomial functors are given in \textcite{weber2011polynomials, gambino2013polynomial}.
Type constructors correspond generally to polynomials in $\cat{Set}$. Consider the $\mathsf{Prod}$ type constructor, for instance. It induces the following very simple polynomial.
\[1 \leftarrow {1 + 1}
\to 1 \to 1\]
Each summand in the second component corresponds to a premiss in the formation rule. Here, every morphism is trivial, which is a consequence of types being monosorted. We will see more illustrative examples later. The polynomial induces the polynomial functor $(-) \mapsto (-) \times (-)$, algebras for which are sets $S$ with a function $\denop{Prod}: S^2 \to S$ as intended.
Type operators (\latinabbr{i.e.}{} formation rules) are described generally in terms of arities.
\begin{notation}
Let $M: \cat{Set} \to \cat{Set}$ be the free monoid endofunctor. For any functor $F: \cat{Set} \to \cat{Set}$, define $F^{\star} \stackrel{\textup{def}}{=} M \circ F$.
\end{notation}
\begin{definition}
\label{def:arity}
We define $\mathsf{ar}_k: \cat{Set} \to \cat{Set}$, for $k \in \mathbb{N}$, inductively.
\vspace{-1.6ex}
\begin{align*}
\mathsf{ar}_0 & \stackrel{\textup{def}}{=} \Id{} \\
\mathsf{ar}_{k + 1} & \stackrel{\textup{def}}{=} {\mathsf{ar}_k}^{\star} \times \mathsf{ar}_k
\end{align*}
We call $\mathsf{ar}_k(S)$ the set of \emph{$S$-sorted $k$\textsuperscript{th}{}-order arities}.
\end{definition}
\begin{notation}
We denote by $A_1, \ldots, A_n \to A$ the $S$-sorted first-order arity
$\big((A_1, \ldots, A_n), A\big) \in \mathsf{ar}_1(S)$. We identify nullary arities with constants and omit the arrow ($\to$) when $n = 0$.
\end{notation}
\begin{notation}
We denote by $\underline{n}$ the set $\{1, \ldots, n\}$, for $n \in \mathbb{N}$. In particular, $\underline{0}$ is the empty set.
\end{notation}
First-order arities correspond to the operators of (multisorted) universal algebra \cite{birkhoff1970heterogeneous}, though in this setting we are solely concerned with monosorted operators. Specifically, our type operators are represented by $\{*\}$-sorted first-order arities, where $*$ is the unique kind.
In general, an $n$-ary type operator
\[\mathsf{O} : \underbrace{*, \ldots, *}_{n \in \mathbb{N}} \to *\]
corresponds to a type formation rule of the form
\[\begin{prooftree}
\Hypo{A_1\ \mathsf{type}}
\Hypo{\cdots}
\Hypo{A_n\ \mathsf{type}}
\Infer3[$\mathsf{O}$-\textsc{form}]{\mathsf{O}(A_1, \ldots, A_n)\ \mathsf{type}}
\end{prooftree}\]
where $A_1, \ldots, A_n$ are type metavariables, universally quantified over all types.
An $n$-ary type operator induces a polynomial in $\cat{Set}$
\[1 \leftarrow \underline{n} \to 1 \to 1\]
intuitively the following.
\[\scalebox{0.9}{
$\{*\} \leftarrow \{A_1 : *\} + \dots + \{A_n : *\} \to \{\mathsf{O}(A_1, \ldots, A_n) : *\} \to \{*\}$
}\]
An algebra for the induced polynomial functor is given explicitly by a set $S$ and a function $\denop{O}: S^n \to S$. We collect the arities into a single \emph{signature}, which completely describes the inductive structure of the types.
\begin{definition}
A \emph{type operator signature}, denoted $\opty$, is given by a list of $\{*\}$-sorted first-order arities.
\end{definition}
\begin{notation}
To aid readability, we will use the following informal notation throughout. The $\rhd$ symbol separates the type metavariables
from a type, term or equation involving them. For example, the notation \[A, B : * \rhd \mathsf{Prod}(A, B) : *\] specifies a $\{*\}$-sorted first-order arity $*, * \to *$.
\end{notation}
\begin{example}[Formation rules for the simply-typed \lc{}{}]
Let $I \in \cat{Set}$ be a finite set of base types.
\begin{align*}
& \rhd \mathsf{Base}_i : * & (i \in I) \\
& \rhd \mathsf{Unit} : * \\
A, B : * & \rhd \mathsf{Prod}(A, B) : * \\
A, B : * & \rhd \mathsf{Fun}(A, B) : *
\end{align*}
\end{example}
A type operator signature induces a polynomial (resp.{} polynomial functor), given by taking the coproduct of the polynomials (resp.{} polynomial functors) induced by its elements.
\begin{notation}
We will denote by $\opty$ both a type operator signature and the polynomial
functor $\opty: \cat{Set} \to \cat{Set}$ it induces.
\end{notation}
The polynomial functor $\opty$ induces a monad giving the closure of a set of type metavariables under the operators of the signature.
\begin{notation}
Given a type operator signature $\opty$, we denote by $\opty[*]$ the free $\opty$-algebra monad on $\cat{Set}$.
\end{notation}
The Eilenberg--Moore category of the monad $\opty[*]$ is isomorphic to the category of $\opty$-algebras.
\subsection{Equations on types}
\label{subsec:equations-on-types}
We permit types to be identified by means of equational laws. For any $m \in \mathbb{N}$, the set $\opty[*](\underline{m})$ may be considered syntactically as the set of types parameterised by $m$ type metavariables. Each element of $\underline{m}$ acts as a placeholder, which one can substitute for a concrete type,
by the freeness of $\opty[*](\underline{m})$ as in the following. A morphism $\mathbf A$ as below corresponds to a family of sorts $(A_i)_{1 \leq i \leq m} \in S^m$.
\begin{equation}
\label{eq:type-equation-substitution}
\begin{tikzcd}
{\opty(\opty[*](\underline{m}))} && {\opty(S)} \\
{\opty[*](\underline{m})} && {S} \\
& {\underline{m}}
\arrow["{\mathbf A}"', from=3-2, to=2-3]
\arrow["{\eta}", from=3-2, to=2-1]
\arrow["{\Psi_{\mathbf A}}"', from=2-1, to=2-3, dashed]
\arrow["{\denop{ty}^*}"', from=1-1, to=2-1]
\arrow["{\denop{ty}}", from=1-3, to=2-3]
\arrow["{\opty(\Psi_{\mathbf A})}", from=1-1, to=1-3]
\end{tikzcd}
\end{equation}
\begin{definition}
An \emph{$\opty$-type equation} is given by a pair
$\big( {m \in \mathbb{N}} , {(L,R) \in \opty[*](\underline{m})^2} \big)$, representing an equation between types
$L \jeq R$ parameterised by $m$ metavariables.
\end{definition}
An $\opty$-type equation induces a term monad identifying the terms in the $(L,R)$ pair \cite{fiore2009construction}, intuitively given by quotienting $\opty[*]$ by the equation.
\begin{definition}
An \emph{equational type signature}, typically denoted $\sigty$, is given by a type operator signature $\opty$ and a list $\eqty$ of $\opty$-type equations.
\end{definition}
\begin{definition}
Given an equational type signature $\sigty = (\opty, \eqty)$, a \emph{$\sigty$-algebra} is an $\opty$-algebra satisfying the equations of $\eqty$.
\end{definition}
\begin{notation}
Given an equational type signature $\sigty$, we denote by ${\Sigma^*_{\sf{ty}}}$ the associated term monad on $\cat{Set}$.
\end{notation}
The term monad associated to an equational type signature $\big(\opty, [\,]\big)$ is the free $\opty$-monad $\opty[*]$. For any list of \text{$\opty$-type} equations $\eqty$, there is a canonical quotient monad morphism $\opty[*] \twoheadrightarrow {\Sigma^*_{\sf{ty}}}$.
\begin{example}[Unityped $\lambda$-calculus]
In the unityped $\lambda$-calculus there is a single type constant $\mathsf{D} : *$ and a single type constructor $\mathsf{Fun} : *, * \to *$, where function types are identified with the base constant: $\rhd\ \mathsf{D} \jeq \mathsf{Fun}(\mathsf{D}, \mathsf{D})$.
\end{example}
\section{Contexts}
\label{sec:contexts}
Type theories have a notion of \emph{(variable) context}, explicitly quantifying the free variables that may appear in a term. Here, we take the contexts of simple type theories to be cartesian: intuitively, lists of typed variables, admitting exchange, weakening, and contraction. Cartesian context structures model the structure of such contexts.
\begin{definition}
Given an equational type signature $\sigty = (\opty, \eqty)$, a \emph{cartesian $\sigty$-typed context structure} for an algebra $\denop{ty}: {\Sigma^*_{\sf{ty}}}(S) \to S$ consists of
\begin{itemize}
\item a small category $\smcat C$, the \emph{category of contexts}, with a specified terminal object $\epsilon$, the \emph{empty context};
\item a functor $\tuple{-}: \disc S \to \smcat C$, embedding sorts as \emph{single-variable contexts}, where $\disc S$ denotes the discrete category on a set $S$;
\item for all $\Gamma \in \smcat C$ and $A \in S$, a specified product $\Gamma \times \tuple{A}$, \emph{context extension} of $\Gamma$ by a variable of sort $A$.
\end{itemize}
\end{definition}
\begin{notation}
We write $\freecart{S}$ for the free strict cartesian category on a set $S$, given concretely by the opposite of the comma category
$(\mathbb{F} \hookrightarrow \cat{Set}) \downarrow (S : \mathbbm{1} \to \cat{Set})$,
where $\mathbb{F}$ is the skeleton of the category of finite sets and functions.
\end{notation}
\begin{example}
Every algebraic theory \cite{adamek2010algebraic} (that is, a cartesian category) $\smcat C$ is an example of a cartesian $\Id{|\smcat C|}$-typed context structure (in fact, one closed under concatenation, rather than just extension).
\end{example}
\begin{definition}
\label{def:context-structure-hom}
A \emph{homomorphism of cartesian $\sigty$-typed context structures} from $(\smcat C, S) \to (\smcat C', S')$ consists of
\begin{itemize}
\item a functor $H : \smcat C \to \smcat C'$;
\item a ${\Sigma^*_{\sf{ty}}}$-algebra homomorphism $h: S \to S'$,
\end{itemize}
such that the following diagram commutes.
\[\begin{tikzcd}
{\smcat C \times \disc S} && {\smcat C' \times \disc{S'}} \\
{\smcat C \times \smcat C} && {\smcat C' \times \smcat C'} \\
{\smcat C} && {\smcat C'} \\
& {\mathbbm{1}}
\arrow["{H}"', from=3-1, to=3-3]
\arrow["{\epsilon}", from=4-2, to=3-1]
\arrow["{\epsilon'}"', from=4-2, to=3-3]
\arrow["{\times}"', from=2-1, to=3-1]
\arrow["{\id{} \times \tuple{-}}"', from=1-1, to=2-1]
\arrow["{H \times \disc h}", from=1-1, to=1-3]
\arrow["{\id{} \times \tuple{-}'}", from=1-3, to=2-3]
\arrow["{\times}", from=2-3, to=3-3]
\end{tikzcd}\]
\end{definition}
Cartesian $\sigty$-typed context structures and their homomorphisms form a category.
\begin{proposition}
\label{prop:initial-cartesian-typed-context-structure}
There is a left-adjoint free functor taking sets $S$ to the free cartesian \text{$\sigty$-typed} context structure on $S$, given by $\freecart{{\Sigma^*_{\sf{ty}}}(S)}$ with $\tuple{-}$ the canonical embedding.
\end{proposition}
In particular, the free cartesian $\sigty$-typed context structure on $\emptyset$ is the initial object.
\section{Terms}
\label{sec:terms}
We follow the tradition of abstract syntax, initiated in \textcite{fiore1999abstract}, of representing models of terms as presheaves over categories of contexts. In particular, for a cartesian $\sigty$-typed context structure $\smcat C$, we consider presheaves $T : \smcat C^{\textup{op}} \to \cat{Set}$ as sets of terms, indexed by their context. For each context $\Gamma \in \smcat C^{\textup{op}}$, $T(\Gamma)$ is to be regarded as the set of terms with variables in $\Gamma$; while a morphism ${\rho : \Gamma \to \Gamma'}$ in $\smcat C^{\textup{op}}$, representing a context renaming, induces a mapping $T(\rho) : T(\Gamma) \to T(\Gamma')$ between terms in different contexts.
\begin{notation}
We use the same symbol for a set $S$ (resp.{} function $h: S \to S'$) and any constant presheaf on $S$ (resp.{} any constant natural transformation on $h$).
\end{notation}
The set of sorts $S$ embeds into $\spsh C$ as a constant presheaf: intuitively a presheaf of types that do not depend on their context. In this light, a natural transformation $\tau : T \to S$ in $\spsh C$ is to be regarded as an assignment of types to terms that respects context renaming. The slice category $\spsh C/S$ is thus an appropriate setting for considering typed terms in context. (Note that we work in the fibred setting, rather than the equivalent indexed setting of \textcite{fiore2002semantic}.)
\begin{definition}
A \emph{typed term structure} for a cartesian \text{$\sigty$-typed} context structure $(\smcat C, S)$ is an object of $\spsh C/S$, concretely
\begin{itemize}
\item a presheaf $T$ in $\spsh C$, the \emph{terms};
\item a natural transformation $\tau: T \to S$, the \emph{assignment of a type} for each term.
\end{itemize}
The type of any term $t \in T(\Gamma)$ is therefore given by $\tau_\Gamma(t)$ (\latinabbr{cf.}{}~the view taken in \textcite{fiore2012discretetalk} and \citeauthor{awodey2018natural}'s natural models~\cite{awodey2018natural}).
\end{definition}
\begin{example}
\label{eg:presheaf-of-variables}
The \emph{presheaf of variables} for a cartesian \text{$\sigty$-typed} context structure $(\smcat C, S)$ forms a typed term structure $\nu : V \to S$ given by the following, where $\mathbf{y}$ denotes the Yoneda embedding.
\begin{align*}
V & \stackrel{\textup{def}}{=} \coprod_{A \in S} \mathbf{y} \tuple{A} & \nu(\tuple{A, \rho}) & \stackrel{\textup{def}}{=} A
\end{align*}
\end{example}
The presheaf of variables is so called because, for any context $\Gamma$, the set $V(\Gamma)$ is to be regarded as the variables in $\Gamma$. We note that, for all presheaves $X \in \spsh C$ and $A \in S$, one has $X^{V_A} \equiv X(- \times \tuple A)$, illustrating that exponentiation by $V_A$ is the same as context extension \cite{fiore1999abstract} (in turn demonstrating that context extension is polynomial).
\begin{proposition}
For all $n \in \mathbb{N}$, the morphism ${\nu^n\!:\!V^n \to S^n}$ is representable.
\end{proposition}
Any presheaf of terms may be restricted to just those with a specified type, by taking pullbacks, as in the following example.
\begin{example}
Given a typed term structure $\tau : T \to S$ and a sort $A \in S$, we denote by $T_A$ the presheaf consisting of terms in $T$ whose type is $A$, given by the fibre:
\[\begin{tikzcd}
{T_A} & {T} \\
{1} & {S}
\arrow["{\tau}", from=1-2, to=2-2] \arrow["{\iota_A}", from=1-1, to=1-2] \arrow[from=1-1, to=2-1] \arrow["{A}"', from=2-1, to=2-2] \arrow["\lrcorner"{very near start, rotate=0}, from=1-1, to=2-2, phantom]
\end{tikzcd}\]
It induces a typed term structure
given by the composite $T_A\to 1\xrightarrow{A} S$.
\end{example}
\subsection{Algebraic models of the simply-typed \lc{}{}}
\label{sec:algebraic-models-of-the-stlc}
Terms have two additional forms of structure that is not found in simple types: multisorting and binding. We walk through the illustrative algebraic term structure of the simply-typed \lc{}{} to give intuition before providing the general construction in \Cref{sec:algebraic-term-structure}. First, we will identify the structure we expect our models to have, before seeing how this structure arises from our models being algebras for a polynomial functor in \Cref{sec:stlc-polynomials}.
As with the algebraic structure for types, the algebraic structure for terms is presented by natural deduction rules (typically introduction or elimination rules), each rule corresponding to an operator on terms.
\paragraph{Products}
The introduction rule for $\mathsf{Prod}$ is given by the following.
\begin{equation}
\label{rule:prod-intro}
\begin{prooftree}
\Hypo{\tr{a}{A}}
\Hypo{\tr{b}{B}}
\Infer2[$\mathsf{Prod}$-\textsc{intro}]{\tr{\mathsf{pair}(a, b)}{\mathsf{Prod}(A, B)}}
\end{prooftree}
\end{equation}
Conceptually, the introduction rule allows one to take two terms of any two types $A$ and $B$ and form a new term, their pair, such that the type of the new term is the product $\denop{Prod}(A, B)$, given by the algebraic structure of the types. A typed term structure $\tau : T \to S$ therefore models \ref{poly:prod-intro} when equipped with a morphism $\denop{pair}$ such that the following diagram commutes.
\[\begin{tikzcd}
{T \times T} & {T} \\
{S \times S} & {S}
\arrow["{\tau}", from=1-2, to=2-2]
\arrow["{\tau \times \tau}"', from=1-1, to=2-1]
\arrow["{\denote{\mathsf{pair}}}", from=1-1, to=1-2, dashed]
\arrow["{\denote{\mathsf{Prod}}}"', from=2-1, to=2-2]
\end{tikzcd}\]
The elimination rules for products are given by the following.
\vspace{-1.6ex}
\begin{equation}
\label{rule:prod-elim1}
\begin{prooftree}
\Hypo{\tr{p}{\mathsf{Prod}(A, B)}}
\Infer1[$\mathsf{Prod}$-\textsc{elim}$_1$]{\tr{\mathsf{proj}_1(p)}{A}}
\end{prooftree}
\end{equation}
\begin{equation}
\label{rule:prod-elim2}
\begin{prooftree}
\Hypo{\tr{p}{\mathsf{Prod}(A, B)}}
\Infer1[$\mathsf{Prod}$-\textsc{elim}$_2$]{\tr{\mathsf{proj}_2(p)}{B}}
\end{prooftree}
\end{equation}
A typed term structure $\tau : T \to S$ models the first projection when equipped with a morphism $\denop{proj_1}$ such that the following left-hand square commutes, where $T_{\denop{Prod}}$ is given by the following right-hand square.
\[\begin{tikzcd}
{T_{\denote{\mathsf{Prod}}}} & {T} \\
{S \times S} & {S}
\arrow["{\tau}", from=1-2, to=2-2]
\arrow[from=1-1, to=2-1]
\arrow["{\denote{\mathsf{proj_1}}}", from=1-1, to=1-2, dashed]
\arrow["{\pi_1}"', from=2-1, to=2-2]
\end{tikzcd}
\qquad
\begin{tikzcd}
{T_{\denote{\mathsf{Prod}}}} & {T} \\
{S \times S} & {S}
\arrow["{\tau}", from=1-2, to=2-2] \arrow["{\iota}", from=1-1, to=1-2] \arrow[from=1-1, to=2-1] \arrow["{\denote{\mathsf{Prod}}}"', from=2-1, to=2-2] \arrow["\lrcorner"{very near start, rotate=0}, from=1-1, to=2-2, phantom]
\end{tikzcd}
\]
This condition is analogous to the one for the introduction rule, the primary difference being that it is only possible to project from terms that have type $\mathsf{Prod}(A, B)$ for some types $A$ and $B$. The situation for \ref{poly:prod-elim2} is analogous.
\paragraph{Units}
Compared to that for products, the algebraic structure for units is almost trivial. The introduction rule for $\mathsf{Unit}$ is given by the following.
\begin{equation}
\label{rule:unit-intro}
\begin{prooftree}
\Infer0[$\mathsf{Unit}$-\textsc{intro}]{\tr{\mathsf{u}}{\mathsf{Unit}}}
\end{prooftree}
\end{equation}
A typed term structure $\tau : T \to S$ for the $\mathsf{Unit}$ type should therefore single out a
term $\denop{u}$ with type $\tau(\denop{u}) = \denop{\mathsf{Unit}}$ (in any context). That is, we expect the following diagram to commute.
\[\begin{tikzcd}
{1} & {T} \\
& {S}
\arrow["{\tau}", from=1-2, to=2-2]
\arrow["{\denop{Unit}}"', from=1-1, to=2-2]
\arrow["{\denop{u}}", from=1-1, to=1-2, dashed]
\end{tikzcd}\]
\paragraph{$\lambda$-abstraction}
Having considered sorting structure, we now consider variable binding. The introduction rule for $\mathsf{Fun}$ is given by the following.
\begin{equation}
\label{rule:fun-intro}
\begin{prooftree}
\Hypo{\trec{\twt{a}{A}}{b}{B}}
\Infer1[$\mathsf{Fun}$-\textsc{intro}]{\tr{\mathsf{abs}(\bind{\twt{a}{A}}{b})}{\mathsf{Fun}(A, B)}}
\end{prooftree}
\end{equation}
The $\mathsf{abs}$ operator allows one to take a term in an extended context and form a term in the original context. A typed term structure $\tau : T \to S$ therefore models \ref{poly:fun-intro} when equipped, for every context $\Gamma$ and types $A, B \in S$, with a
mapping
$\denop{abs}_{\Gamma}^{A, B}$, natural in $\Gamma$, such that the following diagram on the left commutes.
\[\begin{tikzcd}
{T_B(\Gamma \times \tuple{A})} & {T(\Gamma)} \\
{1} & {S}
\arrow["{\tau_\Gamma}", from=1-2, to=2-2]
\arrow["{\denop{abs}_{\Gamma}^{A, B}}", from=1-1, to=1-2, dashed]
\arrow[from=1-1, to=2-1]
\arrow["{\denop{Fun}(A,B)}"', from=2-1, to=2-2]
\end{tikzcd}
\qquad
\begin{tikzcd}
{{T_B}^{V_A}} & {T} \\
{1} & {S}
\arrow["{\tau}", from=1-2, to=2-2]
\arrow["{\denop{abs}^{A, B}}", from=1-1, to=1-2, dashed]
\arrow[from=1-1, to=2-1]
\arrow["{\denop{Fun}(A,B)}"', from=2-1, to=2-2]
\end{tikzcd}\]
Through the relationship between context extension and exponentiation by
representables (\Cref{eg:presheaf-of-variables}),
this is equivalent to the above diagram on the right, for all $A, B \in S$.
Finally, quantifying over the types, this is further equivalent to the following formulation.
\vspace{-1.6ex}
\[\begin{tikzcd}
{\coprod_{A, B \in S} {T_B}^{V_A}} & {T} \\
{S \times S} & {S}
\arrow["{\tau}", from=1-2, to=2-2]
\arrow["{\denop{abs}}", from=1-1, to=1-2, dashed]
\arrow["{\pi}"', from=1-1, to=2-1]
\arrow["{\denop{Fun}}"', from=2-1, to=2-2]
\end{tikzcd}\]
The elimination rule is simpler.
\begin{equation}
\label{rule:fun-elim}
\begin{prooftree}
\Hypo{\tr{f}{\mathsf{Fun}(A, B)}}
\Hypo{\tr{a}{A}}
\Infer2[$\mathsf{Fun}$-\textsc{elim}]{\tr{\mathsf{app}(f, a)}{B}}
\end{prooftree}
\end{equation}
A typed term structure $\tau : T \to S$ models function application when equipped with a morphism $\denop{app}$ such that the following square commutes.
\[\begin{tikzcd}
{\coprod_{A, B \in S} T_{\denop{Fun}(A, B)} \times T_A} & {T} \\
{S \times S} & {S}
\arrow["{\pi_2}"', from=2-1, to=2-2]
\arrow["{\tau}", from=1-2, to=2-2]
\arrow["{\denop{app}}", from=1-1, to=1-2, dashed]
\arrow["{\pi}"', from=1-1, to=2-1]
\end{tikzcd}\]
\subsection{Polynomials for the simply-typed \lc{}{}}
\label{sec:stlc-polynomials}
We now show how the above structure for models of the simply-typed \lc{}{} is actually algebraic structure for a polynomial functor. While, so far, we have only dealt with polynomials in $\cat{Set}$, we recall that the concept makes sense for any presheaf category $\spsh C$.
For every morphism $f : A \to B$ in $\spsh C$, there is an adjoint triple $\dsum{f} \dashv \pb{f} \dashv \dprod{f}$ where $\dsum{f}$ is postcomposition by $f$ and $\pb{f} : \spsh C/B \to \spsh C/A$ is pullback along $f$. Every polynomial $I \leftarrow A \to B \to J$ induces a polynomial functor $\dsum{t} \dprod{f} \pb{s} : \spsh C/I \to \spsh C/J$.
An algebra for a functor $F : \spsh C/S \to \spsh C/S$ is a typed term structure $\tau : T \to S$ with a morphism $\varphi : \text{dom}(F(\tau)) \to T$ such that the following diagram commutes.
\[\begin{tikzcd}
{\text{dom}(F(\tau))} & {T} \\
& {S}
\arrow["{\tau}", from=1-2, to=2-2]
\arrow["{F(\tau)}"', from=1-1, to=2-2]
\arrow["{\varphi}", from=1-1, to=1-2, dashed]
\end{tikzcd}\]
We will write $F(T)$ to mean $\text{dom}(F(\tau))$ when unambiguous.
In particular, algebras for a polynomial functor,
$\varphi: \dsum{t} \dprod{f} \pb{s}(\tau : T \to S) \to (\tau : T \to S)$,
are illustrated by the following diagram.
\[\begin{tikzcd}
{T} & {\pb{s}(T)} & {\dprod{f} \pb{s}(T)} & {T} \\
{S} & {A} & {B} & {S}
\arrow["{f}"', from=2-2, to=2-3]
\arrow["{s}", from=2-2, to=2-1]
\arrow["{t}"', from=2-3, to=2-4]
\arrow["{\tau}"', from=1-1, to=2-1]
\arrow["{\tau}", from=1-4, to=2-4]
\arrow["{\dprod{f} \pb{s}(\tau)}" description, from=1-3, to=2-3]
\arrow["{\varphi}", from=1-3, to=1-4, dashed]
\arrow[from=1-2, to=1-1]
\arrow["{\pb{s}(\tau)}" description, from=1-2, to=2-2]
\arrow["\lrcorner"{very near start, rotate=-90}, from=1-2, to=2-1, phantom]
\end{tikzcd}\]
We will sometimes depict polynomials geometrically as in the following.
\vspace{-1.6ex}
\[\begin{tikzcd}[cramped, row sep=tiny, column sep=small]
& {A} & {B} \\
{I} &&& {J}
\arrow["{s}"', from=1-2, to=2-1] \arrow["{t}", from=1-3, to=2-4] \arrow["{f}", from=1-2, to=1-3]
\end{tikzcd}\]
We may then unambiguously omit a component morphism, which is taken to be the identity. The composition of two polynomials is also a polynomial \cite[Proposition~1.12]{gambino2013polynomial}, depicted graphically as in the following.
\[\begin{tikzcd}[cramped, row sep=tiny, column sep=small]
& {A} & {B} && {C} & {D} \\
{I} &&& {J} &&& {K}
\arrow
from=1-2, to=2-1]
\arrow
from=1-3, to=2-4]
\arrow
from=1-2, to=1-3]
\arrow
from=1-5, to=2-4]
\arrow
from=1-5, to=1-6]
\arrow
from=1-6, to=2-7]
\end{tikzcd}\]
\paragraph{Products}
The condition for \ruleref{prod-intro} exactly states that $\tau : T \to S$ is an algebra for the polynomial functor induced by the following polynomial in $\spsh C$.
\begin{equation}
\label{poly:prod-intro}
\tag{\hyperref[rule:prod-intro]{$\mathsf{Prod}$-\textsc{intro}}}
S \xleftarrow{[\pi_1, \pi_2]} S^2 + S^2 \xrightarrow{\nabla_2} S^2 \xrightarrow{\denote{\mathsf{Prod}}} S
\end{equation}
The structure of this polynomial may seem opaque at first; we will attempt to provide some intuition. The polynomial describes a (many-in, one-out) transformation between terms, respecting the type structure. The middle component $\nabla_2 : S^2 + S^2 \to S^2$ represents the type metavariables $A$ and $B$: each summand in the domain represents the metavariables in a premiss, while the codomain represents the metavariables in the conclusion. While some metavariables may not appear in every premiss, each premiss is implicitly parameterised by each type metavariable; the codiagonal ensures that the metavariables available to each premiss (and the conclusion) are the same (\latinabbr{i.e.}{} unified).
The leftmost component $S \leftarrow S^2 + S^2 : [\pi_1, \pi_2]$ describes the types of each premiss, given the metavariables. In this case, the types are simply projections: the left-hand side to $A$ and the right-hand side to $B$. The rightmost component $\denop{Prod} : S^2 \to S$ describes the type of the conclusion, given the metavariables: in this case, constructing the product of $A$ and $B$.
An algebra for the functor induced by this polynomial is calculated explicitly below, to demonstrate that it aligns with the structure we deduced earlier.
\[\begin{tikzcd}
{T} & {T \times S + S \times T} & {T^2} & {T} \\
{S} & {S^2 + S^2} & {S^2} & {S}
\arrow["{\denop{Prod}}"', from=2-3, to=2-4] \arrow["{[\pi_1, \pi_2]}", from=2-2, to=2-1] \arrow["{\nabla_2}"', from=2-2, to=2-3] \arrow["{\tau}"', from=1-1, to=2-1] \arrow["{[\pi_1, \pi_2]}"', from=1-2, to=1-1] \arrow["{\tau \times \id{} + \id{} \times \tau}" description, from=1-2, to=2-2] \arrow["{\tau^2 = \dprod{\nabla_2}(\tau \times \id{} + \id{} \times \tau)}" description, from=1-3, to=2-3] \arrow["{\tau}", from=1-4, to=2-4] \arrow["{\denop{pair}}", from=1-3, to=1-4, dashed] \arrow["\lrcorner"{very near start, rotate=-90}, from=1-2, to=2-1, phantom]
\end{tikzcd}\]
The polynomials for the projections are similarly described. For $\tau : T \to S$ to be a model of the product eliminators \ruleref{prod-elim1} and \ruleref{prod-elim2}, we require it to be an algebra for the following polynomials.
\begin{equation}
\label{poly:prod-elim1}
\tag{\hyperref[rule:prod-elim1]{$\mathsf{Prod}$-\textsc{elim}$_1$}}
S \xleftarrow{\denote{\mathsf{Prod}}} S^2 \xrightarrow{\nabla_1} S^2 \xrightarrow{\pi_1} S
\end{equation}
\begin{equation}
\label{poly:prod-elim2}
\tag{\hyperref[rule:prod-elim2]{$\mathsf{Prod}$-\textsc{elim}$_2$}}
S \xleftarrow{\denote{\mathsf{Prod}}} S^2 \xrightarrow{\nabla_1} S^2 \xrightarrow{\pi_2} S
\end{equation}
Given some examination, the structure of the polynomials is analogous to that of the introduction rule, with the first component selecting the type of the premiss, the codiagonal (in this case trivially) unifying the premisses, and the final component selecting the type of the conclusion.
\paragraph{Units}
For \ruleref{unit-intro}, the polynomial inducing the structure is similarly defined. For $\tau : T \to S$ to be a model of $\mathsf{u}$, we require it to be an algebra for the following polynomial.
\begin{equation}
\label{poly:unit-intro}
\tag{\hyperref[rule:unit-intro]{$\mathsf{Unit}$-\textsc{intro}}}
S \xleftarrow{!} 0 \xrightarrow{\nabla_0} 1 \xrightarrow{\denop{Unit}} S
\end{equation}
One may see that, as the introduction rule for $\mathsf{Unit}$ has no premisses and no type metavariables, this (trivially) fits the same pattern as with $\mathsf{Prod}$.
\paragraph{$\lambda$-abstraction}
To describe binding structure, we need more sophisticated polynomials. For $\tau : T \to S$ to be a model of \ruleref{fun-intro}, we require it to be an algebra for the following polynomial.
\begin{equation}
\label{poly:fun-intro}
\tag{\hyperref[rule:fun-intro]{$\mathsf{Fun}$-\textsc{intro}}}
S \xleftarrow{\pi_2} V \times S \xrightarrow{\nu \times \id{}} S^2 \xrightarrow{\denote{\mathsf{Fun}}} S
\end{equation}
Here, the first and last components are familiar from the previous examples.
The form of the middle component is new: metavariables involved in context extension, and therefore in variable binding, must be fibred over the presheaf of variables $V$. The typed term structure $\nu: V \to S$ of~\Cref{eg:presheaf-of-variables} forgets
the information associated to a variable apart from its type.
For $\tau : T \to S$ to be a model of \ruleref{fun-elim}, we require it to be an algebra for the following polynomial.
\begin{equation}
\label{poly:fun-elim}
\tag{\hyperref[rule:fun-elim]{$\mathsf{Fun}$-\textsc{elim}}}
S \xleftarrow{[\denote{\mathsf{Fun}}, \pi_1]} S^2 + S^2 \xrightarrow{\nabla_2} S^2 \xrightarrow{\pi_2} S
\end{equation}
Polynomials are closed under taking coproducts: for a typed term structure to be a model of the entire structure of the simply-typed \lc{}{}, therefore, we require it to be an algebra for the polynomial endofunctor induced by the coproduct of all the aforementioned polynomials.
\subsection{Algebraic term structure}
\label{sec:algebraic-term-structure}
We now give syntax for a general natural deduction rule for a term operator and the construction of the polynomial it induces. As with type operators, we have a notion of arity corresponding to term operators.
\begin{notation}
\label{not:second-order-arity}
We denote by \[(A^1_1, \ldots, A^1_{k_1})A_1, \ldots, (A^n_1, \ldots, A^n_{k_n})A_n \to (B_1, \ldots, B_k)B\] the $S$-sorted second-order arity (\Cref{def:arity})
\[\scriptstyle ((((A^1_1, \ldots, A^1_{k_1}), A_1), \ldots, ((A^n_1, \ldots, A^n_{k_n}), A_n)), ((B_1, \ldots, B_k), B)) \in \mathsf{ar}_2(S)\]
Second-order arities correspond to the operators of multisorted binding algebra \cite{fiore2010second}: such an arity represents an operator taking $n$ arguments, the $i$\textsuperscript{th}{} of which binds $k_i$ variables, which is parameterised by $k$ variables. We identify nullary arities with constants.
\end{notation}
Given an equational type signature $\sigty$ and $m \in \mathbb{N}$ type metavariables, we can represent term operators by \text{${\Sigma^*_{\sf{ty}}}(\underline{m})$-sorted} second-order arities. An $n$-ary term operator
\begin{equation}
\label{eq:second-order-oper}
\footnotesize \mathsf{o} : (A^1_1, \ldots, A^1_{k_1})A_1, \ldots, (A^n_1, \ldots, A^n_{k_n})A_n \to (B_1, \ldots, B_k)B
\end{equation}
corresponds to a rule as in \Cref{fig:term-operator}, universally quantified over all contexts $\Gamma$.
\begin{figure*}
\begin{mdframed}
\[\begin{prooftree}
\Hypo{\trec{\twt{x^1_1}{A^1_1}, \ldots, \twt{x^1_{k_1}}{A^1_{k_1}}}{t_1}{A_1}}
\Hypo{\cdots}
\Hypo{\trec{\twt{x^n_1}{A^n_1}, \ldots, \twt{x^n_{k_n}}{A^n_{k_n}}}{t_n}{A_n}}
\Infer3{\trec{\twt{y_1}{B_1}, \ldots, \twt{y_k}{B_k}}{\mathsf{o}[\twt{y_1}{B_1}, \ldots, \twt{y_k}{B_k}]\big((\twt{x^1_1}{A^1_1}, \ldots, \twt{x^1_{k_1}}{A^1_{k_1}}) t_1, \ldots, (\twt{x^n_1}{A^n_1}, \ldots, \twt{x^n_{k_n}}{A^n_{k_n}}) t_n\big)}{B}}
\end{prooftree}\]
\caption{Natural deduction rule for a term operator}
\label{fig:term-operator}
\end{mdframed}
\end{figure*}
\begin{definition}
\label{def:parameterised-term-operators}
We say that a term operator is \emph{parameterised} when $k \neq 0$.
\end{definition}
A term operator for an equational type signature $\sigty$, as in \eqref{eq:second-order-oper}, induces a polynomial in $\spsh C$ for any cartesian $\sigty$-typed context structure, given in \Cref{fig:term-polynomial}.
\begin{figure*}
\begin{mdframed}
\[\!\!\!\begin{tikzcd}[column sep=scriptsize, row sep=normal]
& & & {\coprod_{1 \leq i \leq n} S^m} & {S^m} && {V^k \times S^m}
\\
&
{\coprod_{1 \leq i \leq n} V^{k_i} \times S^m}
\arrow{r}[swap, yshift=-1ex]{\coprod_{1 \leq i \leq n} \nu^{k_i} \times \id{}}
&
{\coprod_{1 \leq i \leq n} S^{k_i} \times S^m}
&
&&
{S^k \times S^m}
&&
{S}
\\
{S}
%
\arrow["{[\denote{A_i} \circ \pi_2]_{1 \leq i \leq n}}"', from=2-2, to=3-1]
\arrow["{\coprod_{1 \leq i \leq n} \tuple{\tuple{\denote{A^i_j}}_{1 \leq j \leq k_i}, \id{}}}"', from=1-4, to=2-3]
\arrow["{\nabla_n}", from=1-4, to=1-5]
\arrow["{\tuple{\tuple{\denote{B_j}}_{1 \leq j \leq k}, \id{}}}"', from=1-5, to=2-6]
\arrow["{\nu^k \times \id{}}", from=1-7, to=2-6]
\arrow["{\denote{B}\circ\pi_2}", from=1-7, to=2-8]
\end{tikzcd}\]
\caption{Polynomial induced by a term operator}
\label{fig:term-polynomial}
\end{mdframed}
\end{figure*}
\begin{figure*}
\begin{mdframed}
\[
\begin{tikzcd}
{\prod_{1 \leq i \leq n} T_{\denote{A_i}(\mathbf C)}(\Gamma \times \tuple{\denote{A^i_j}(\mathbf C)}_{1 \leq j \leq k_i})} & {T(\Gamma \times \tuple{\denote{B_j}(\mathbf C)}_{1 \leq j \leq k})} \\
{1} & {S}
\arrow["{\tau_{\Gamma \times \tuple{\denote{B_j}(\mathbf C)}_{1 \leq j \leq k}
}}", from=1-2, to=2-2]
\arrow["{\denote{B}(\mathbf C)}"', from=2-1, to=2-2]
\arrow[from=1-1, to=2-1]
\arrow["\denote{\mathsf{o}}^\sharp_\Gamma", from=1-1, to=1-2]
\end{tikzcd}
\qquad\qquad
(\mathbf C \in S^m)
\]
%
\[
\text{Natural in the context $\Gamma$, where
$\tuple{D_1, \ldots, D_\ell}
\stackrel{\textup{def}}{=}
(\cdots (\epsilon \times \tuple{D_1}) \times \cdots) \times \tuple{D_\ell}.$}
\]
%
\caption{Algebra structure induced by a term operator}
\label{fig:term-algebra}
\end{mdframed}
\end{figure*}
\begin{definition}
A \emph{term operator signature}, denoted $\optm$, for an equational type signature $\sigty$ is given by a list of pairs of natural numbers $m \in \mathbb{N}$ and ${\Sigma^*_{\sf{ty}}}(\underline{m})$-sorted second-order arities.
\end{definition}
\begin{example}[Term operators for the simply-typed \lc{}{}]
\setlength{\abovedisplayskip}{-1ex}
\begin{align*}
&\rhd \mathsf{u} : \mathsf{Unit} \\
A, B : * &\rhd \mathsf{abs} : \bind{A}{B} \to \mathsf{Fun}(A, B) \\
A, B : * &\rhd \mathsf{app} : \mathsf{Fun}(A, B), A \to B \\
A, B : * &\rhd \mathsf{pair} : A, B \to \mathsf{Prod}(A, B) \\
A, B : * &\rhd \mathsf{proj}_1 : \mathsf{Prod}(A, B) \to A \\
A, B : * &\rhd \mathsf{proj}_2 : \mathsf{Prod}(A, B) \to B
\end{align*}
\end{example}
A term operator signature induces a polynomial (resp.{} polynomial functor), given by taking the coproduct of the polynomials (resp.{} polynomial functors) induced by its elements.
\begin{notation}
We will denote by $\optm$ both a term operator signature and the polynomial functor it induces.
\end{notation}
\begin{remark}
\label{remark:explicit-dependent-products}
To gain intuition for the polynomial algebraic structure, it is instructive to evaluate the polynomials oneself, starting with an arbitrary $\tau : T \to S$ and taking pullbacks, dependent products and postcomposing. Of these operations, pullbacks and postcomposing are straightforward.
We give the two relevant calculations for the dependent products explicitly.
\[\begin{array}{l}
\dprod{\nabla_n}(P \to \coprod_{1 \leq i \leq n} S^m)
\equiv
\big(\coprod_{\mathbf A \in S^m} \prod_{1 \leq i \leq n} P_{\tuple{i, \mathbf A}} \big) \to S^m
\\[10pt]
\dprod{\nu^k \times \id{}}(V^k \times P \xrightarrow{\id{}\times p} V^k \times S^m)
\\[5pt]
\quad \equiv
\big(
\coprod_{\mathbf A \in S^k, \mathbf B \in S^m}
{P_{\mathbf B}}
^{ \prod_{1 \leq i \leq k}
V_{A_i}
}
\big)
\to S^k \times S^m
\end{array}\]
\end{remark}
\begin{proposition}
In elementary terms, $\optm$-algebras for the polynomial functor as in~\Cref{fig:term-polynomial} are equivalently given by typed term structures $\tau : T \to S$ with a natural transformation $\denop{o}^\sharp$ such that the diagram in \Cref{fig:term-algebra} commutes.
\end{proposition}
\begin{proposition}
\label{lemma:sigtm-finitary}
For all term signatures, the endofunctor $\optm$ on $\spsh C/S$ is finitary.
\end{proposition}
Thus, $\optm$ induces a monad~\cite{fiore2009construction} describing the term structure, closed under the operators of the signature.
\begin{notation}
Given a term operator signature $\optm$, we denote by $\optm[*]$ the free $\optm$-algebra monad on $\spsh C/S$.
\end{notation}
The Eilenberg--Moore category of the monad $\optm[*]$ is isomorphic to the category of $\optm$-algebras.
\section{Models of simply typed syntax}
\label{sec:models-syntax}
We now give the definition of simply typed syntax, along with its models. Note that this is not yet a full notion of simple type theory, as we lack substitution and equations.
\begin{definition}
A \emph{simply typed syntax} consists of:
\begin{itemize}
\item a type operator signature $\opty$;
\item a term operator signature $\optm$ for $\opty$.
\end{itemize}
\end{definition}
\begin{definition}
A \emph{model} for a simply typed syntax consists of
\begin{itemize}
\item an $\opty$-algebra $\denop{ty} : \opty(S) \to S$;
\item a cartesian $\opty$-typed context structure $\smcat C$ for $S$;
\item an $\optm$-algebra
$\denop{tm}\!: \optm {(\tau\!:\!T\!\to\!S)} \to {(\tau\!:\!T\!\to\!S)}$.
\end{itemize}
\end{definition}
\begin{proposition}
\label{prop:simply-typed-cwfs}
$S$-sorted simply-typed categories with families~\cite{castellan2019categories} are equivalent to models of simply typed syntax for an empty type and term signature, such that the carriers of the $\opty$- and $\optm$-algebras are $S$ and $\nu : V \to S$ respectively.
\end{proposition}
To discuss the relationships between different models of a simply typed syntax, and to prove that the syntactic model is initial, we need a notion of homomorphism. This necessarily involves a compatibility condition between algebraic term structures.
\begin{figure*}
\begin{mdframed}
\[\begin{tikzcd}[row sep=huge]
{\textstyle
\prod_{1 \leq i \leq n}
T_{\denote{A_i}(\mathbf C)}
( \Gamma \times \tuple{ \denote{A^i_j}(\mathbf C) }_{1\leq j\leq k_i} )
}
&
{\displaystyle
T_{\denote{B}(\mathbf C)}
( \Gamma \times \tuple{ \denote{B_j}(\mathbf C) }_{1\leq j\leq k} )
}
\\
{\textstyle
\prod_{1 \leq i \leq n}
T'_{\denote{A_i}'(h(\mathbf C))}
( H(\Gamma) \times \tuple{ \denote{A^i_j}'(h(\mathbf C)) }_{1\leq j\leq k_i} )
}
&
{\textstyle
T'_{\denote{B}'
(h(\mathbf C))}(H(\Gamma) \times \tuple{ \denote{B_j}'( h(\mathbf C))}_{1\leq j\leq k} )
}
\arrow["{\denop{tm}^\sharp_\Gamma}", from=1-1, to=1-2]
\arrow["{(\denop{tm}')^\sharp_\Gamma}", from=2-1, to=2-2]
\arrow["{
\prod_{1 \leq i \leq n} \displaystyle
(f_{\denote{A_i}(\mathbf C)})_{(\Gamma \times \tuple{ \denote{A^i_j}(\mathbf C) }_{1\leq j\leq k_i} )}
}" description, from=1-1, to=2-1]
\arrow["{\displaystyle
(f_{\denote{B}(\mathbf C)})_{(\Gamma \times \tuple{ \denote{B_j}(\mathbf C) }_{1\leq i\leq k} )}
}" description, from=1-2, to=2-2]
\end{tikzcd}\]
\[
\mathbf C = \tuple{C_1,\ldots,C_m} \in S^m
\qquad\qquad
h(\mathbf C) \stackrel{\textup{def}}{=} \tuple{h(C_1), \ldots, h(C_m)}
\]
\caption{Elementary term algebra coherence}
\label{fig:elementary-term-algebra-coherence}
\end{mdframed}
\end{figure*}
Categorically, this is made somewhat difficult to express by the fact that two typed term structures for the same signature may be algebras for polynomial endofunctors on different presheaf categories, depending on their cartesian \text{$\opty$-typed} context structures. To reconcile them, we will make use of the following lemma.
\begin{lemma}
\label{lemma:poly-commutes-with-hom}
Let $(\smcat{C}, \tau : T \to S)$ and $(\smcat{C}', \tau' : T' \to S')$ be models, and let
$( H: \smcat C \to \smcat C' , h : S \to S' )$ be a cartesian $\opty$-typed context structure homomorphism between them (\Cref{def:context-structure-hom}).
Then there is a canonical natural transformation as follows.
\[\begin{tikzcd}
{\spsh{C'}/S'} & {\spsh{C}/S} \\
{\spsh{C'}/S'} & {\spsh{C}/S}
\arrow["{\optm}", from=1-2, to=2-2] \arrow["{\optm[']}"', from=1-1, to=2-1] \arrow["{\pb h \circ \precomp H}", from=1-1, to=1-2] \arrow["{\pb h \circ \precomp H}"', from=2-1, to=2-2] \arrow[Rightarrow, from=1-2, to=2-1, shorten <=2mm, shorten >=2mm]
\end{tikzcd}\]
\end{lemma}
\begin{definition}
A \emph{homomorphism of models} for a simply typed syntax, from a model $(\smcat{C}, {\tau: T \to S})$ to a model $(\smcat{C}', {\tau' : T' \to S'})$, consists of
\begin{itemize}
\item a cartesian $\opty$-typed context structure homomorphism
$( H: \smcat C \to \smcat C', h : S \to S' )$;
\item a natural transformation $f: T \to T'H$,
\end{itemize}
such that the following diagrams, term-type coherence (left) and term algebra coherence~(right), commute:
\[\begin{tikzcd}
{T} & {T'H} \\
{S} & {S'}
\arrow["{h}"', from=2-1, to=2-2] \arrow["{\tau}"', from=1-1, to=2-1] \arrow["{\tau'H}", from=1-2, to=2-2] \arrow["{f}", from=1-1, to=1-2]
\end{tikzcd}
\qquad
\begin{tikzcd}
{\optm(T)} && {\optm(\pb{h}(T'H))} \\
&& {\pb{h}(\optm['](T')H)} \\
{T} && {\pb{h}(T'H)}
\arrow["{\denote{\mathsf{tm}}}"', from=1-1, to=3-1] \arrow["{f_{\lrcorner}}"', from=3-1, to=3-3] \arrow["{\optm( f_{\lrcorner})}", from=1-1, to=1-3] \arrow["{\text{(\Cref{lemma:poly-commutes-with-hom})}}", from=1-3, to=2-3] \arrow["{\pb{h}(\denote{\mathsf{tm}}' H)}", from=2-3, to=3-3]
\end{tikzcd}\]
where $f_\lrcorner$ is the mediating morphism as in the following diagram.
\vspace{-1.6ex}
\[\begin{tikzcd}
{T} & {\pb{h}(T'H)} & {T'H} \\
& {S} & {S'}
\arrow["{f_{\lrcorner}}" description, from=1-1, to=1-2]
\arrow["{\tau' H}", from=1-3, to=2-3]
\arrow["{\tau}"', from=1-1, to=2-2]
\arrow["{h}"', from=2-2, to=2-3]
\arrow[from=1-2, to=1-3]
\arrow["{\pb{h}(\tau' H)}" description, from=1-2, to=2-2]
\arrow["\lrcorner"{very near start, rotate=0}, from=1-2, to=2-3, phantom]
\arrow["{f}", from=1-1, to=1-3, bend left]
\end{tikzcd}\]
\end{definition}
The term algebra coherence diagram expresses that $f_\lrcorner$ is an $\optm$-algebra homomorphism. This equivalently expresses that $f$ is a form of term algebra heteromorphism as in the following diagram.
\[\begin{tikzcd}[column sep=tiny]
{\optm(T)} \arrow{r}[yshift=1ex]{\optm(f_{\lrcorner})} & {\optm(\pb{h}(T'H))} \arrow{r}[yshift=1ex]{\text{(\Cref{lemma:poly-commutes-with-hom})}} & {\pb{h}(\optm['](T')H)} \arrow{r}[yshift=1ex]{}
& {\optm['](T')H} \\
{T} &&& {T'H}
\arrow["{\denote{\mathsf{tm}}}"', from=1-1, to=2-1]
\arrow["{\denote{\mathsf{tm}}'H}", from=1-4, to=2-4]
\arrow["{f}"', from=2-1, to=2-4]
\end{tikzcd}\]
In elementary terms, this corresponds to the coherence condition expressed in \Cref{fig:elementary-term-algebra-coherence}.
This resolves the compatibility difficulty described earlier.
\begin{example}
For any cartesian $\opty$-typed context structure homomorphism
$(H, h)$,
there is a canonical model homomorphism
$(H, h, v)$
for $v : V \to V'H$ given by the action of $H$:
\[
v_\Gamma\big(\, A \, , \, {\rho:\Gamma\to\tuple A} \,\big)
\stackrel{\textup{def}}{=}{}
\big(\, h(A) \, ,\, {H(\rho) : H(\Gamma) \to \tuple{h(A)}} \,\big)
\]
\end{example}
Models of simply typed syntax and their homomorphisms, for a simply typed syntax $O$, form a category $\smcat{S}_O$.
\begin{theorem}
\label{thm:initiality-syntax}
$\smcat{S}_O$ has an initial object.
\end{theorem}
\begin{proof}
Let $\denop{ty}: \opty(S) \to S$ be the initial $\opty$-algebra and let $\smcat C$ be the free cartesian $\opty$-typed context structure on $S$ as in \Cref{prop:initial-cartesian-typed-context-structure}.
The slice category $\spsh{C}/S$ is cocomplete and the polynomial endofunctor $\optm$ is finitary (\Cref{lemma:sigtm-finitary}). Thus, we have an initial
$\optm$-algebra $\denop{tm} : \optm(\tau\!:\!T \to S) \longrightarrow (\tau\!:\!T \to S)$.
Then $M \stackrel{\textup{def}}{=} (\smcat C, \denop{ty}, \tau\!:\!T \to S, \denop{tm})$ is a model for the signature $O$.
Let $(\smcat C', \denop{ty}', \tau'\!:\!T' \to S', \denop{tm}')$ be a model of simply typed syntax for the signature $O$. There is a unique $\opty$-homomorphism $h: S \to S'$, by initiality of $S$, and ${H: \smcat C \to \smcat C'}$ is uniquely determined by the freeness of $\smcat C$. Furthermore, there is a unique $\optm$-homomorphism ${f: T \to T'H}$ satisfying the coherence conditions by the initiality of $\tau : T \to S$. Finally, $(H, h, f)$ is a unique model homomorphism and $M$ is therefore initial.
\end{proof}
The initial object in $\smcat{S}_O$ is the \emph{syntactic model}.
Indeed, according to the viewpoint of initial-algebra semantics~\cite{thomas1976initial}, syntactic models are precisely initial ones, for
these have
a canonical compositional interpretation into all models and, as such, uniquely characterise any concrete syntactic construction up to isomorphism. Here, it is further possible to make the finitary semantic construction of \Cref{thm:initiality-syntax} explicit to demonstrate its coincidence with familiar syntactic constructions.
\section{Substitution}
\label{sec:substitution}
\newcommand{\mathsf{wk}}{\mathsf{wk}}
\newcommand{\mathsf{contr}}{\mathsf{contr}}
\newcommand{\mathsf{exch}}{\mathsf{exch}}
\newcommand{\mathsf{subst}}{\mathsf{subst}}
\newcommand{\mathsf{var}}{\mathsf{var}}
$\optm$-algebras represent a notion of terms with (sorted and binding) algebraic structure. However, there are still two important concepts that are missing: that of (capture-avoiding) substitution, and that of equations. Substitution must be described before defining equations on terms, as many equational laws (such as the $\beta$-equality of the
simply-typed \text{$\lambda$-calculus})
involve this meta-operation. Substitution is an important metatheoretic concept even besides this, and is necessary to define the multicategorical composition operation that will appear in some of the models of simple type theories (\Cref{def:multisubstitutional-model}).
To begin to talk about substitution, one must have a notion of \emph{variables as terms}, corresponding to the following structure.
\vspace{-1.6ex}
\[\begin{tikzcd}
{V} & {T} \\
& {S}
\arrow["{\nu}"', from=1-1, to=2-2]
\arrow["{\mathsf{var}}", from=1-1, to=1-2]
\arrow["{\tau}", from=1-2, to=2-2]
\end{tikzcd}\]
This structure is not a term operator:
it may instead be added by considering free $\optm$-algebras on the typed term structure of variables, $\nu : V \to S$.
Substitution is traditionally given in one of two forms: single-variable substitution (typically denoted $\subst{t}{u}{x}$) and multivariable substitution (in which terms must be given for every variable in context). When the category of contexts is freely generated, these notions are equivalent. In our more general setting single-variable substitution is the appropriate primitive notion.
One may present substitution as an operation given by the following rule.
\[\begin{prooftree}
\Hypo{\Gamma, x : A \vdash t : B}
\Hypo{\Gamma \vdash u : A}
\Infer2[$\mathsf{subst}$]{\Gamma \vdash \subst{t}{u}{x} : B}
\end{prooftree}\]
It corresponds to the polynomial below, according to the general description of \Cref{sec:algebraic-term-structure}.
\[S \xleftarrow{[\pi_2, \pi_1]} V \times S + S^2 \xrightarrow{[\nu \times \id{}, \id{}]} S^2 \xrightarrow{\pi_2} S\]
An algebra for the functor induced by this polynomial is given explicitly by a morphism $\mathsf{subst}$ in $\spsh C/S$ as in the diagram below.
\[\begin{tikzcd}
{T} & {V \times T + T \times S} & {\coprod_{A, B \in S} {T_B}^{V_A} \times T_A} & {T} \\
{S} & {V \times S + S^2} & {S^2} & {S}
\arrow["{[\nu \times \id{}, \id{}]}"', from=2-2, to=2-3]
\arrow["{[\pi_2, \pi_1]}", from=2-2, to=2-1]
\arrow["{\pi_2}"', from=2-3, to=2-4]
\arrow["{\tau}"', from=1-1, to=2-1]
\arrow["{\tau}", from=1-4, to=2-4]
\arrow[from=1-3, to=2-3]
\arrow["{\mathsf{subst}}", from=1-3, to=1-4, dashed]
\arrow["{[\pi_2, \pi_1]}"'{yshift=0.5ex}, from=1-2, to=1-1]
\arrow["\lrcorner"{very near start, rotate=-90}, from=1-2, to=2-1, phantom]
\arrow["{\tau \times \id{} + \tau \times \id{}}" description, from=1-2, to=2-2]
\end{tikzcd}\]
Here, for expository purposes, we shall equivalently consider the structure as given by a family of morphisms in $\spsh C$,
\[\mathsf{subst}_{A, B} : {T_B}^{V_A} \times T_A \to T_B \qquad (A,B \in S)\]
which is closer to the syntactic intuition.
The substitution operator must obey equational laws (\latinabbr{cf.}{} \cite[Definition~3.1]{fiore1999abstract} and \cite[Section~2.1]{fiore2014substitution}). This structure must be described semantically,
as it makes use of the implicit cartesian structure of the categories of contexts, which is not available syntactically.
Specifically, we require the following diagrams to commute. They correspond respectively to trivial substitution, left and right identities, and associativity.
\[\begin{tikzcd}
{T_A \times T_B} & {{T_A}^{V_B} \times T_B} \\
& {T_A}
\arrow["{\mathsf{wk} \times \id{}}", from=1-1, to=1-2]
\arrow["{\pi_1}"', from=1-1, to=2-2]
\arrow["{\mathsf{subst}_{B, A}}", from=1-2, to=2-2]
\end{tikzcd}
\enspace
\begin{tikzcd}
{1\times T_A} & {{T_A}^{V_A} \times T_A} \\
& {T_A}
\arrow["{\lambda(\mathsf{var}_A)\times\id{}}"{yshift=1ex}, from=1-1, to=1-2]
\arrow["{\mathsf{subst}_{A, A}}", from=1-2, to=2-2]
\arrow["\pi_2"', from=1-1, to=2-2, no head]
\end{tikzcd}\]
\[\begin{tikzcd}
{{T_B}^{V_A} \times V_A} & {{T_B}^{V_A} \times T_A} \\
& {T_B}
\arrow["{\id{} \times \mathsf{var}_A}"{yshift=1ex}, from=1-1, to=1-2]
\arrow["{\mathsf{subst}_{A, B}}", from=1-2, to=2-2]
\arrow["{\mathsf{contr}}"', from=1-1, to=2-2]
\end{tikzcd}\]
\[\begin{tikzcd}
{({T_C}^{V_B \times V_A} \times {T_B}^{V_A}) \times T_A} & {{T_C}^{V_A} \times T_A}
&
\\
{({T_C}^{V_B \times V_A} \times T_A) \times ({T_B}^{V_A} \times T_A)}
&
{T_C}
\\
{({T_C}^{V_A \times V_B} \times {T_A}^{V_B}) \times ({T_B}^{V_A} \times T_A)}
&
{{T_C}^{V_B} \times T_B} &
\arrow["{(\mathsf{exch} \times \mathsf{wk}) \times \id{}}", from=2-1, to=3-1]
\arrow["{\mathsf{subst}_{B, C}^{V_A} \times \id{}}", from=1-1, to=1-2]
\arrow["{\mathsf{subst}_{A, C}^{V_B} \times \mathsf{subst}_{A, B}}"', from=3-1, to=3-2]
\arrow["{\mathsf{subst}_{A, C}}", from=1-2, to=2-2]
\arrow["{\mathsf{subst}_{B, C}}"', from=3-2, to=2-2]
\arrow["{\text{str}}", from=1-1, to=2-1]
\end{tikzcd}
\]
The morphism $\mathsf{exch}$ is given by $T^\gamma$ where $\gamma : X \times Y \xrightarrow{\equiv} Y \times X$ is the cartesian symmetry; $\mathsf{wk}$ by $X^! : X \to X^Y$; and $\mathsf{contr}$ by the evaluation.
By the extension structure of cartesian $\sigty$-typed context structures, they respectively correspond to the admissible syntactic operations of exchange, weakening, and contraction. The map $\text{str}$ is the canonical strength of products and the map $\mathsf{subst}_{A,B}^P$ is the composite
\[
{T_B}^{V_A\times P}\times {T_A}^{P}
\xrightarrow{\equiv}
({T_B}^{V_A}\times T_A)^{P}
\xrightarrow{(\mathsf{subst}_{A,B})^{P}}
{T_B}^{P}
\]
Crucially, substitution must also commute with all the operators of the theory: for every \emph{unparameterised} operator $\mathsf{o}$
as in \Cref{fig:term-operator}, we require the following diagram to commute for all $\mathbf C \in S^m$ and
$D\in S$, where $E_i \stackrel{\textup{def}}{=} \prod_{1 \leq j \leq k_i} V_{\denote{A^i_j}(\mathbf C)}$.
\[\begin{tikzcd}
{\big(\prod_{1 \leq i \leq n} {T_{\denote{A_i}(\mathbf C)}}^{E_i}\big)^{V_D} \times T_D}
&
{{T_{\denote B(\mathbf C)}}^{V_D} \times T_D}
\\
{\prod_{1 \leq i \leq n}
{T_{\denote{A_i}(\mathbf C)}}^{V_D\times E_i} \times {T_D}^{E_i}}
\\
{\prod_{1 \leq i \leq n} {T_{\denote{A_i}(\mathbf C)}}^{E_i}}
&
{T_{{\denote B}(\mathbf C)}}
\arrow["{(
\denop{o}^\sharp)
^{V_D} \times \id{}}"{yshift=1ex}, from=1-1, to=1-2]
\arrow["{\mathsf{subst}_{D, \denote B(\mathbf C)}}" description, from=1-2, to=3-2]
\arrow["{\prod_{1 \leq i \leq n} \mathsf{subst}^{E_i}_{D,\denote{A_i}(\mathbf C)}}", from=2-1, to=3-1]
\arrow["{\denop{o}^\sharp}"', from=3-1, to=3-2]
\arrow["{\cong \circ (\cong \times \tuple{\mathsf{wk}_i}_{1 \leq i \leq n})}",
from=1-1, to=2-1]
\end{tikzcd}\]
Models of simply typed syntax with unparameterised term operators may be extended to incorporate variable and substitution structure; together with homomorphisms that preserve this structure, they form a category.
\begin{proposition}
For a term signature $\optm$ with unparameterised operators, models of simply typed syntax with variable and substitution structure over a fixed cartesian typed context structure admit free constructions, and thereby generate a free monad
which we denote by $\optm[\circledast]$.
\end{proposition}
\begin{proposition}
$\optm[\circledast]$-algebras for the free cartesian \text{$([\,], [\,])$-typed} context structure on a single sort (\Cref{prop:initial-cartesian-typed-context-structure}) are, equivalently, $\optm$-substitution algebras
\cite{fiore1999abstract}.
\end{proposition}
\begin{theorem}[Substitution lemma, \latinabbr{cf.}{}~\cite{fiore1999abstract,fiore2008second}]
\label{thm:substitution-lemma}
For a fixed $\sigty$-algebra and the free cartesian $\sigty$-typed context structure
thereon, provided that the signature $\optm$ contains only unparameterised operators, the free $\optm[*]$-algebra on $\nu : V \to S$ and the initial \text{$\optm[\circledast]$-algebra} are isomorphic.
\end{theorem}
From the syntactic viewpoint, this means that substitution is admissible: adding a substitution operator to a simply typed syntax leaves the associated terms unchanged, because a term involving substitution is always equal to one that does not involve substitution.
\begin{proposition}
\label{prop:admissible-parameterised-operators}
In the presence of weakening and exchange (present in the simple type theories we consider here) and substitution, parameterised term operators (\Cref{def:parameterised-term-operators}) are admissible.
\end{proposition}
\section{Equations on terms}
\label{sec:equations}
Equations on terms may now be treated, analogously to those on types, with the proviso that one must keep track of sorts and variable contexts. In particular, we are interested in terms parameterised by a number of type metavariables, and term metavariables in
extended contexts~\cite{hamana2004free, fiore2008second}.
\begin{figure*}
\begin{mdframed}
\[\begin{prooftree}
\Hypo{\trec{\twt{x^1_1}{A^1_1}, \ldots, \twt{x^1_{k_1}}{A^1_{k_1}}}{t_1}{A_1}}
\Hypo{\cdots}
\Hypo{\trec{\twt{x^n_1}{A^n_1}, \ldots, \twt{x^n_{k_n}}{A^n_{k_n}}}{t_n}{A_n}}
\Infer3{\trec{\twt{y_1}{B_1}, \ldots, \twt{y_k}{B_k}}{l \jeq r}{B}}
\end{prooftree}\]
\caption{Natural deduction rule for a term equation}
\label{fig:term-equation}
\end{mdframed}
\end{figure*}
\begin{notation}
For ${m\in\mathbb{N}}$, let $K_m$ denote the free $\sigty$-algebra ${\Sigma^*_{\sf{ty}}}{(\underline m)}$ on a set of type metavariables $\underline m$ and let $\smcat K_m$ denote the category of contexts $\freecart{K_m}$
of the free cartesian \text{$\sigty$-typed} context structure on a set of type metavariables $\underline{m}$ (\Cref{prop:initial-cartesian-typed-context-structure}).
\end{notation}
\begin{definition}
An \emph{$\optm$-term equation} is given by a triple
\begin{equation}
\label{eq:term-equation}
\big(\,
m \in \mathbb{N}
\, , \,
\big(\mathbf{A}_1,\ldots\mathbf{A}_n\to\mathbf{B}\big) \in \mathsf{ar}_2\big(\opty[*](\underline{m})\big)
\, , \,
( l , r )
\,\big)
\end{equation}
with $l, r : \prod_{1 \leq j \leq k} V_{B_j} \to \optm[\circledast]( p : P \to K_m )_{B}$ a parallel pair of morphisms in $\widehat{\smcat K_m}$ for
\begin{align*}
P & \stackrel{\textup{def}}{=} \coprod_{1 \leq i \leq n} \prod_{1 \leq j \leq k_i} V_{A^i_j} &
p \big( \tuple{ i , (\rho_1,\ldots,\rho_{k_i}) } \big) & \stackrel{\textup{def}}{=} A_i
\end{align*}
where $\mathbf{A}_i = (A^i_1, \ldots, A^i_{k_i})A_i$ and $\mathbf{B}=(B_1,\ldots,B_k)B$.
\end{definition}
The parallel pair equivalently corresponds to a pair of terms in $\optm[\circledast](P)_{B}(\tuple{B_1,\ldots,B_k})$
which may be syntactically presented as in \Cref{fig:term-equation}.
\begin{definition}
An \emph{equational term signature}, typically denoted $\sigtm$, is given by a term operator signature $\optm$ and a list $\eqtm$ of $\optm$-term equations.
\end{definition}
Fix an $\optm$-term equation as in~\eqref{eq:term-equation} and consider an \text{$\optm$-algebra} in $\spsh{\smcat C}/S$:
\begin{equation}
\label{eq:optm-algebra}
\denop{tm} : \optm(\tau\!:\!T \to S) \longrightarrow (\tau\!:\!T \to S)
\end{equation}
Every $\mathbf C \in S^m$ freely induces a homomorphism $(H, h) : (\smcat K_m, K_m) \to (\smcat C, S)$, and every morphism $\mathbf t$ in $\psh{\smcat K_m}/{K_m}$ as below
\begin{equation}
\label{eq:term-assignment}
\begin{tikzcd}[row sep=scriptsize, column sep=scriptsize]
P & & h^*(TH)
\\
& K_m &
\arrow["\displaystyle \mathbf t", from=1-1, to=1-3]
\arrow["\displaystyle p"', from=1-1, to=2-2]
\arrow["\displaystyle h^*(\tau H)", from=1-3, to=2-2]
\end{tikzcd}
\end{equation}
freely induces the following situation, analogously to \eqref{eq:type-equation-substitution}.
\[\begin{tikzcd}
{\optm(\optm[\circledast]( P ))}
&&
{\optm(\pb{h}(TH))}
\\
{\optm[\circledast]( P )}
&&
{\pb{h}(TH)}
\\
& P
&&
\arrow["{\eta}", from=3-2, to=2-1]
\arrow["{\psi_{\mathbf t}}"', from=2-1, to=2-3, dashed]
\arrow["{\denop{tm}^\oast}"', from=1-1, to=2-1]
\arrow["{\pb{h}(\denote{\mathsf{tm}} H)\ \circ\ \text{(\Cref{lemma:poly-commutes-with-hom})}}", from=1-3, to=2-3]
\arrow["{\optm(\psi_{\mathbf t})}", from=1-1, to=1-3]
\arrow["\mathbf t"', from=3-2, to=2-3]
\end{tikzcd}\]
\begin{definition}
An $\optm$-algebra as in \eqref{eq:optm-algebra} satisfies an \text{$\optm$-term} equation as in~\eqref{eq:term-equation} whenever, for all $\mathbf C$ and $\mathbf t$ as in the preceding discussion,
$\psi_{\mathbf t} \circ l = \psi_{\mathbf t} \circ r$.
\end{definition}
A morphism $\mathbf t$ as in~\eqref{eq:term-assignment} corresponds to a family
\[
t_i \in T_{\denote{A_i}(\mathbf C)}(\tuple{\denote{A^i_1}(\mathbf C),\ldots,\denote{A^i_{k_i}}(\mathbf C)})
\qquad
(1\leq i\leq n)
\]
As such, it provides a valuation for the term placeholders of the terms in the equation. Indeed, the evaluation of $\psi_{\mathbf t}$ at ${u \in \optm[\circledast](P)_{B}(\tuple{B_1,\ldots,B_k}})$ is the
term
resulting from a meta-substitution operation replacing the term placeholders in $u$ with the concrete terms
$(t_i)_{1 \leq i \leq n}$.
\begin{definition}
Given an equational term signature $\sigtm = (\optm, \eqtm)$, a \emph{$\sigtm$-algebra} is an $\optm$-algebra that satisfies the equations of $\eqtm$.
\end{definition}
Equational term signatures (like equational type signatures) are an entirely syntactic notion and correspond exactly to systems of natural deduction rules presenting a simple type theory. We give examples.
\begin{example}
\label{eg:equational-presentations}
Equational presentations in multisorted universal algebra \cite{birkhoff1970heterogeneous} are examples of equational term signatures, whose operators are nonbinding and unparameterised.
\end{example}
\begin{notation}
We will informally denote by ${t : {(x\!:\!A)}B}$ a term metavariable $t$ of type $B$ in contexts extended by a fresh variable $x$ of type $A$, reminiscent of the notation for second-order arities (\Cref{not:second-order-arity}). The types of bound variables in term operators, and of terms themselves, may be inferred, and are elided.
\end{notation}
\begin{example}[$\beta$/$\eta$ rules for the simply-typed \lc{}{}]
\begin{align*}
A, B : * \rhd t : (x\!:\!A)B, a : A
& \vdash
\mathsf{app}\big(\mathsf{abs}\big(\bind{
}{\subst{t}{z}{x}}\big), a\big)
\jeq
\subst{t}{a}{x}
\\[2.5pt]
A, B : * \rhd f : \mathsf{Fun}(A, B)
& \vdash
\mathsf{abs}\big(\bind{
}{\mathsf{app}(f, x)}\big)
\jeq
f
\end{align*}
\end{example}
{
\allowdisplaybreaks
\begin{example}[Computational $\lambda$-calculus \cite{moggi1991notions}]
\label{example:computational-lambda-calculus}
The following extends the simply-typed \lc{}{}.
\begin{align*}
& \rhd \mathsf{T} : * \to *
\\[.5em]
A : * & \rhd \mathsf{return} : A \to \mathsf{T}(A)
\\[.25em]
A, B : * & \rhd \mathsf{bind} : \mathsf{T}(A), \bind{A}{\mathsf{T}(B)} \to \mathsf{T}(B)
\\[.5em]
A, B : * & \rhd a : A, f : \bind{x\!:\!A}{\mathsf{T}(B)}
\\
& \!\!\!\!\!\!\!\!\!\!\!\!
\vdash
\mathsf{bind}\big(\mathsf{return}(a), \bind{
}{\subst f z x}\big)
\jeq
\subst{f}{a}{x}
\\[.25em]
A : * & \rhd m : \mathsf{T}(A)
\vdash
\mathsf{bind}\big(m, \bind{
}{\mathsf{return}(x)}\big)
\jeq
m
\\[.25em]
A, B : * & \rhd m : \mathsf{T}(A), f : \bind{x\!:\!A}{\mathsf{T}(B)}, g : \bind{y\!:\!B}{\mathsf{T}(C)}
\\ & \!\!\!\!\!\!\!\!\!\!\!\!
\vdash
\mathsf{bind}\big(m, \bind{
}{\mathsf{bind}\big(\subst{f}{a}{x}, \bind{b}{\subst g b y}\big)}\big)
\\
& \!\!\!\!\!
\jeq
\mathsf{bind}\big(\mathsf{bind}(m,(a)\subst f a x),(b)\subst g b y\big)
\end{align*}
\end{example}
}
Models of simply typed syntax with variable and substitution structure may be restricted to algebras for equational term signatures.
\begin{proposition}
\label{prop:msigtmsubst}
Algebras for equational term signatures $\sigtm$ with unparameterised operators over a fixed cartesian typed context structure admit free constructions, and thereby generate a free monad, which
we denote by ${\Sigma^\circledast_{\sf{tm}}}$.
\end{proposition}
The monad associated to an equational term signature $(\optm, [\,])$ is the free $\optm$-monad with variable and substitution structure $\optm[\circledast]$. For any list of $\optm$-term equations $\eqtm$, there is a canonical quotient monad morphism $\optm[\circledast] \twoheadrightarrow {\Sigma^\circledast_{\sf{tm}}}$.
\section{Models of simple type theories}
\label{sec:models-theories}
Simple type theories extend simply typed syntax by incorporating variable, substitution, and equational structure.
\begin{definition}
A \emph{simple type theory} consists of:
\begin{itemize}
\item an equational type signature $\sigty$;
\item an equational term signature $\sigtm$ for $\sigty$.
\end{itemize}
\end{definition}
\begin{definition}
A \emph{model} for a simple type theory consists of
\begin{itemize}
\item a ${\Sigma^*_{\sf{ty}}}$-algebra $\denop{ty} : {\Sigma^*_{\sf{ty}}}(S) \to S$;
\item a cartesian $\sigty$-typed context structure $\smcat C$ for $S$;
\item a ${\Sigma^\circledast_{\sf{tm}}}$-algebra
$\denop{tm} : {\Sigma^\circledast_{\sf{tm}}} (\tau\!:\!T \to S) \to (\tau\!:\!T \to S)$.
\end{itemize}
In particular, the type and term algebras both satisfy the specified equations.
\end{definition}
\begin{definition}
A \emph{homomorphism of models} for a simple type theory is a homomorphism $(H, h, f)$ for the underlying simply typed syntax such that $f$ preserves the variable structure and is a heteromorphism for the substitution structure.
\end{definition}
Models of simple type theories and their homomorphisms, for a simple type theory $\Sigma$, form a category $\smcat{S}_\Sigma$.
\begin{theorem}
\label{thm:initiality-theory}
$\smcat{S}_\Sigma$ has an initial object.
\end{theorem}
The initial object is the \emph{syntactic model}. It is given by a construction analogous to the one in \Cref{thm:initiality-syntax}, taking \Cref{prop:msigtmsubst} into account.
\section{Classifying multicategories}
\label{sec:classifying-multicategories}
The classes of models we have considered so far are very general. First, contexts must be closed under extension, but may not necessarily be lists of sorts. More importantly, substitution is not inherent in simply typed syntax, which allowed us to consider models with and without substitution: it is only by making this distinction that we are able to prove metatheoretic properties regarding substitution, such as in \Cref{thm:substitution-lemma}. However, one typically wishes to consider simple type theories that do have an associated notion of substitution, along with contexts that are lists. In this setting, we can reformulate the models to be more familiar to the models dealt with in categorical algebra (see \latinabbr{e.g.}{} \textcite{crole1993categories}).
\begin{definition}
\label{def:multisubstitutional-model}
A model of simple type theory is \emph{multisubstitutional} if
the embedding of the set of sorts in the category of contexts presents the latter as the strict cartesian completion of the former.
\end{definition}
Multisubstitutional models have list-like contexts and admit a multivariable substitution operation~\cite{fiore1999abstract} in addition to, and induced by, the single-variable substitution operation of \Cref{sec:substitution}. In fact, we have the following result.
\begin{theorem}
\label{thm:equivalence-multicategories}
Multisubstitutional models of simple type theories with empty type operator signatures are equivalent to cartesian multicategories with corresponding
structure.
\end{theorem}
We sketch the idea. There is an equivalence taking such a multisubstitutional model
$(\smcat C, \tau : T \to S)$ to a cartesian multicategory $\smcat M$, with object set $S$; multihoms
$\smcat M(A_1, \ldots, A_n; B) = T_B(\tuple{A_1,\ldots,A_n})$; identities
arising from
the variable structure; composition given by the multivariable substitution operation (or, equivalently, by iterated single-variable substitution); and cartesian multicategory structure given by the functorial action of the presheaf $T$ along the exchange, weakening, and contraction structure of
$\smcat C$. Model homomorphisms define cartesian multifunctors.
The algebraic structure on $T$ induces
structure on $\smcat M$, where each term operator induces a pair of functors (corresponding to the premisses and conclusion), natural transformations between which correspond to
interpretations of
the operator.
There are a variety of notions equivalent to cartesian multi\-categories, such as
many-sorted abstract clones and multi\-sorted Lawvere theories, giving corresponding versions of \Cref{thm:equivalence-multicategories} for each notion. Relevant for future work on polynomial models of dependent type theories is the relationship with categories with
families~\cite{dybjer1995internal}. We note that the relationship with simply-typed categories with families in \Cref{prop:simply-typed-cwfs} extends to incorporate operators and equations. However, care must be taken: in the context of categories with families, type theoretic structure is typically expressed through generalised algebraic theories, which permit operators that are not natural in a categorical sense; while, conversely, such unnatural operators are forbidden in the current framework.
Theorems~\ref{thm:initiality-theory} and~\ref{thm:equivalence-multicategories} provide a general systematic construction of the \emph{classifying cartesian multicategory} of any simple type
theory.
In the context of universal algebra, we have the following.
\begin{corollary}
The initial model of the simple type theory for an equational presentation in universal algebra is, equivalently, its abstract clone.
\end{corollary}
Beyond universal algebra, we have a kind of ``generalised Lambek correspondence'' between models of simple type theories and structured cartesian multicategories. When the simple type theory has finite products, the classifying cartesian multicategory is representable and hence equivalent to a cartesian category. In particular, we recover the classical Lambek correspondence.
\begin{corollary}[Lambek correspondence]
\label{cor:lambek}
The initial model of the simple type theory for the simply-typed \lc{}{} with a set of base types $B$ is, equivalently, the free cartesian-closed category on $B$.
\end{corollary}
|
\section{Introduction}
In this paper, we prove the quenched homogenization of the equation
\begin{equation}\label{intro_eq}-\nabla\cdot a(\nicefrac{x}{\varepsilon},\o)\nabla u^\varepsilon+\frac{1}{\varepsilon}b(\nicefrac{x}{\varepsilon},\o)\cdot \nabla u^\varepsilon=f\;\;\textrm{in}\;\;U\;\;\textrm{with}\;\;u^\varepsilon=g\;\;\textrm{on}\;\;\partial U,\end{equation}
for a uniformly elliptic matrix $a$ and a mean zero, divergence-free drift $b$. The coefficients are jointly-measurable, stationary, and ergodic random variables defined on some probability space $(\O,\mathcal F,\P)$. Stationarity asserts that the random environment is statistically homogenous: there exists a measure preserving transformation group $\{\tau_x\colon\O\rightarrow\O\}_{x\in\mathbb{R}^d}$ such that
\begin{equation}\label{intro_rvs}(a(x,\omega),b(x,\omega))=(A(\tau_x\omega),B(\tau_x\omega))\;\textrm{for random variables}\;A\colon\O\rightarrow\mathbb{R}^{d\times d}, B\colon\O\rightarrow\mathbb{R}^d.\end{equation}
The ergodicity is a qualitative form of mixing: for $g\in L^\infty(\O)$,
\begin{equation}\label{intro_ergodic}
g(\omega)=g(\tau_x\omega)\;\;\textrm{almost surely for every}\;\;x\in\mathbb{R}^d\;\;\textrm{if and only if}\;\;g\;\;\textrm{is almost surely constant.}
\end{equation}
In terms of the coefficients, we will assume that the matrix $A$ is bounded and uniformly elliptic: there exist $\lambda,\Lambda\in(0,\infty)$ such that, almost surely for every $\xi\in\mathbb{R}^d$,
\begin{equation}\label{intro_ue}\abs{A\xi}\leq \Lambda\abs{\xi}\;\;\textrm{and}\;\;A\xi\cdot \xi\geq \lambda\abs{\xi}^2.\end{equation}
We will assume that, for some $\d\in(0,1)$, the random drift $B$ admits a stationary $L^{d\vee(2+\d)}$-integrable stream matrix: there exists a skew-symmetric $S\in L^{d\vee(2+\d)}(\O;\mathbb{R}^{d\times d})$ which satisfies
\begin{equation}\label{intro_stream} \nabla\cdot S=B\;\;\textrm{for}\;\;(\nabla\cdot S)_i=\partial_kS_{ik},\end{equation}
fixed by the choice of gauge
\begin{equation}\label{intro_gauge}-\Delta S_{ij}=\partial_iB_j-\partial_j B_i.\end{equation}
We prove in Proposition~\ref{prop_stream} below that a stream matrix exists in $d\geq 3$ if $B$ is $L^d$-integrable and satisfies a finite range of dependence. A stationary stream matrix does not exist in general if $d=2$, and for this reason the homogenization of \eqref{intro_eq} in $d=2$ remains largely an open problem.
In the symmetric case, for sufficiently regular coefficients and a sufficiently regular domain, solutions of \eqref{intro_eq} are related by the Feynman-Kac formula to a rescaling of the stochastic differential equation
\begin{equation}\label{intro_sde}\dd X_t=\sigma\left(X_t,\omega\right)\dd B_t+\left(\nabla\cdot a^t\left(X_t,\omega\right)-b\left(X_t,\omega\right)\right)\dt,\end{equation}
for $\sigma\sigma^t=2a$. Precisely, for the exit time $\tau^\varepsilon$ from the dilated domain $\nicefrac{U}{\varepsilon}$,
\[u^\varepsilon(x)=E_{\frac{x}{\varepsilon},\omega}\left[g\left(\varepsilon X_{\tau^\varepsilon}\right)+\varepsilon^2\int_0^{\tau^\varepsilon}f\left(\varepsilon X_s\right)\ds\right].\]
The homogenization of \eqref{intro_eq} is therefore equivalent to characterizing the asymptotic behavior of the exit distributions and exit times of \eqref{intro_sde} from large domains. Furthermore, in the case $A=I$, equation \eqref{intro_sde} is the passive tracer model
\[\dd X_t=\sqrt{2}\dd B_t+b\left(X_t,\omega\right)\dt,\]
which is a simple approximation for the transport of a passive tracer particle in a turbulent, incompressible flow. This model has applications to hydrology, meteorological sciences, and oceanography, and we point the reader, for instance, to Csanady \cite{Csa1973}, Frish \cite{Fri1995}, and Monin and Yaglom \cite{MonYag2007,MonYag2007II} for more details.
The stream matrix allows equation \eqref{intro_eq} to be rewritten in the form
\begin{equation}\label{intro_transform}-\nabla\cdot \left(a^\varepsilon+s^\varepsilon\right)\cdot\nabla u^\varepsilon=f\;\;\textrm{in}\;\;U\;\;\textrm{with}\;\;u^\varepsilon=g\;\;\textrm{on}\;\;\partial U,\end{equation}
for $s^\varepsilon(x,\omega)=S(\tau_{\nicefrac{x}{\varepsilon}}\omega)$. The transformation \eqref{intro_transform} formally justifies the two-scale expansion familiar from the homogenization of divergence-form equations without drift. That is, for the standard orthonormal basis $\{e_i\}_{i\in\{1,\ldots,d\}}$ of $\mathbb{R}^d$, we expect the corrector $\phi_i$ to be defined by a stationary gradient $\nabla\phi_i$ that solves
\begin{equation}\label{intro_corrector}-\nabla\cdot \left(A+S\right)\left(\nabla\phi_i+e_i\right)=0,\end{equation}
and we expect the homogenized coefficient $\overline{a}\in \mathbb{R}^{d\times d}$ to be defined by
\[\overline{a}e_i=\mathbb{E}\left[(A+S)(\nabla\phi_i+e_i)\right].\]
This is the case if $S$ is bounded, for which the methods of Papanicolaou and Varadhan \cite{PapVar1981} and Osada \cite{Osa1983} yield readily that, for the solution
\begin{equation}\label{intro_homogenized}-\nabla\cdot \overline{a}\nabla v = f \;\;\textrm{in}\;\;U\;\;\textrm{with}\;\;v=g\;\;\textrm{on}\;\;\partial U,\end{equation}
we have weak $H^1$-convergence of $u^\varepsilon$ to $v$, and strong $H^1$-convergence of the two-scale expansion in the sense that almost surely
\[\lim_{\varepsilon\rightarrow 0}\norm{u^\varepsilon-(v+\varepsilon\phi^\varepsilon_i\partial_i v)}_{H^1(U)}=0,\]
for the rescaled correctors $\phi^\varepsilon_i=\phi_i(\nicefrac{x}{\varepsilon},\omega)$, where here and throughout the paper we use Einstein's summation convention over repeated indices.
The case of an unbounded stream matrix $S$ is fundamentally different. Proving the existence of a solution to \eqref{intro_corrector} is essentially straightforward, arguing by approximation and the skew-symmetry of $S$. Uniqueness is however not clear and was posed as an open problem in Avellaneda and Majda \cite{AveMaj1991}. The reason for this is that, while the equation defines $S\nabla\phi_i$ as an element of the dual for any solution $\nabla\phi_i$, it is not clear that this rule extends to a skew-symmetric operator on the solution space. Issues related to this fact explain the strong regularity assumptions used in Oelschl\"ager \cite{Oel1988} and form the technical core of the more recent works Kozma and T\'oth \cite{KozTot2017} and T\'oth \cite{Tot2018}. In this paper we take a different approach based on the methods of \cite{Oel1988}. In Proposition~\ref{prop_corrector} below, assuming the existence of a square integrable stream matrix, we prove that there exists a unique stationary gradient $\nabla\phi_i$ satisfying \eqref{intro_corrector}. The quenched uniqueness of correctors under the higher $L^{d\vee(2+\d)}$-integrability assumption then follows from the Liouville theorem and Proposition~\ref{prop_sublinear} below.
Our first result is the quenched homogenization of \eqref{intro_transform} under the assumptions of uniform ellipticity and the existence of a $L^{d\vee(2+\d)}$-integrable stream matrix:
\begin{equation}\label{steady} \textrm{Assume}\;\eqref{intro_rvs},\; \eqref{intro_ergodic},\; \eqref{intro_ue},\; \eqref{intro_stream},\; \textrm{and}\; \eqref{intro_gauge}.\end{equation}
This result provides a new approach to the results of \cite{Oel1988} and \cite{AveMaj1991}. In comparison to \cite{KozTot2017,Tot2018}, which respectively assume the existence of an $L^2$/$L^{2+\d}$-integrable stream matrix to prove an annealed/quenched central limit theorem for the analogous random walk on the whole space, the methods of this paper explain how the higher $L^{d\vee(2+\d)}$-integrability assumption arises naturally to prove the strong convergence of the two-scale expansion in $H^1$. This is done by introducing the homogenization flux correctors $\sigma_i$ satisfying
\begin{equation}\label{intro_flux}\nabla\cdot \sigma_i = (a+s)(\nabla\phi_i+e_i)-\overline{a}e_i.\end{equation}
The flux correction was used originally in the context of stochastic homogenization by Gloria, Neukamm, and Otto \cite{GloNeuOtt2020}, and allows the residuum of the two-scale expansion to be written in the form
%
\begin{equation}\label{intro_00000}-\nabla\cdot (a^\varepsilon+s^\varepsilon)\nabla (u^\varepsilon-v-\varepsilon\phi^\varepsilon_i\partial_i v) = \nabla\cdot\left[\left(\varepsilon\phi_i^\varepsilon(a^\varepsilon+s^\varepsilon)-\varepsilon\sigma_i^\varepsilon\right)\nabla(\partial_i v)\right].\end{equation}
%
The $L^{d\vee(2+\d)}$-integrability of $S$ is exactly the threshold which guarantees that the righthand side of \eqref{intro_00000} vanishes strongly in $L^2$ as $\varepsilon\rightarrow 0$, using the sublinearity of $\phi_i$ and $\sigma_i$ proven in Proposition~\ref{prop_sublinear} below. These methods apply without change to the elliptic and parabolic settings, and thereby establish an invariance principle on the whole space while also characterizing asymptotically the exit times and exit distributions of \eqref{intro_sde} from large domains.
\begin{thm*}[cf.\ Theorem~\ref{thm_ts} below] Assume \eqref{steady}. For some $\a\in(0,1)$ let $U\subseteq\mathbb{R}^d$ be a bounded $\C^{2,\a}$-domain, let $f\in\C^\a(U)$, and let $g\in \C^{2,\a}(\partial U)$. For each $\varepsilon\in(0,1)$ let $u^\varepsilon\in H^1(U)$ be the unique solution of \eqref{intro_eq} and let $v\in H^1(U)$ be the unique solution of \eqref{intro_homogenized}. Then, almost surely as $\varepsilon\rightarrow 0$,
\[\lim_{\varepsilon\rightarrow 0}\norm{u^\varepsilon-v-\varepsilon\phi^\varepsilon_i\partial_iv}_{H^1(U)}=0.\]
\end{thm*}
The second main result of this work is an almost sure large-scale $\a$-H\"older regularity estimate for whole space solutions $u\in H^1_{\textrm{loc}}(\mathbb{R}^d)$ of the equation
\begin{equation}\label{intro_whole_space}-\nabla\cdot(a+s)\nabla u=0\;\;\textrm{in}\;\;\mathbb{R}^d.\end{equation}
Following \cite{GloNeuOtt2020}, based on the equivalence of Morrey, Campanato, and H\"older spaces (cf.\ eg.\ Giaquinta and Martinazzi \cite{GiaMar2012}), we introduce a version of the large-scale $\a$-H\"older semi-norm defined with respect to the intrinsic $(a+s)$-harmonic coordinates $(x_i+\phi_i)$: the excess $\textrm{Exc}(u;R)$ is defined by
\begin{equation}\label{intro_excess}\textrm{Exc}(u;R)=\inf_{\xi\in\mathbb{R}^d}\left(R^{-2\a}\fint_{B_R}\abs{\nabla u-\xi-\nabla\phi_\xi}^2\right),\end{equation}
for $\phi_\xi=\xi_i\phi_i$. The following theorem proves that there exists an almost surely finite radius $R_0\in(0,\infty)$ after which point the solutions of \eqref{intro_whole_space} enter the regime of $\a$-H\"older regularity. The radius $R_0$ is quantified precisely by the sublinearity of the correctors in Proposition~\ref{prop_excess} below.
\begin{thm*}[cf.\ Proposition~\ref{prop_excess}, Theorem~\ref{thm_excess} below] Assume \eqref{steady}. On a subset of full probability, for every $\a\in(0,1)$ there exists a random radius $R_0\in(0,\infty)$ and a deterministic $c\in(0,\infty)$ depending on $\a$ such that, for every weak solution $u\in H^1_{\textrm{loc}}(\mathbb{R}^d)$ of \eqref{intro_whole_space}, for every $R_1<R_2\in(R_0,\infty)$,
\[R^{-2\a}_1\textrm{Exc}(u;R_1)\leq cR_2^{-2\a}\textrm{Exc}(u;R_2),\]
for the excess $\textrm{Exc}(u;R)$ defined in \eqref{intro_excess}. \end{thm*}
The final result of this work is a first-order Liouville theorem. In analogy with the classical first-order Liouville theorem, the $(a+s)$-harmonic coordinates $(x_i+\phi_i)$ are the linear functions in the random geometry of the space, and every subquadratic $(a+s)$-harmonic function is a corrector. The sublinearity is quantified with respect to the $L^{2_*}$-norm, for $2_*>2$ defined below, as opposed to the $L^2$-norm used in \cite{GloNeuOtt2020}. This stronger condition is necessary for our arguments due to the unboundedness of the stream matrix. In combination, the Liouville theorem and Proposition~\ref{prop_sublinear} below prove the quenched uniqueness of the homogenization correctors and thereby provide a strong answer to the original question of \cite{AveMaj1991} on the physical space.
\begin{thm*}[cf.\ Theorem~\ref{thm_lvl} below] Assume \eqref{steady}, let $q_d=d\vee(2+\d)$, and let $\nicefrac{1}{2_*}=\nicefrac{1}{2}-\nicefrac{1}{q_d}$. Then, on a subset of full probability, every weak solution $u\in H^1_{\textrm{loc}}(\mathbb{R}^d)$ of \eqref{intro_whole_space} that is strictly subquadratic in the sense that, for some $\a\in(0,1)$,
\[\lim_{R\rightarrow\infty}\frac{1}{R^{1+\a}}\left(\fint_{B_R}\abs{u}^{2_*}\right)^\frac{1}{2_*}=0,\]
satisfies $u=c+\xi\cdot x+\phi_\xi$ in $H^1_{\textrm{loc}}(\mathbb{R}^d)$ for some $c\in\mathbb{R}$ and $\xi\in\mathbb{R}^d$.
\end{thm*}
\subsection{The organization of the paper} In Section~\ref{sec_prob}, we describe how equations \eqref{intro_stream}, \eqref{intro_gauge}, and \eqref{intro_corrector} are lifted to the probability space: in Section~\ref{sec_hom_cor} we construct the homogenization correctors, in Section~\ref{sec_flux_cor} we construct the homogenization flux correctors, and in Section~\ref{sec_stream} we prove the existence of a stationary stream matrix. The proof of quenched homogenization is presented in Section~\ref{sec_stoch_hom}, where we also prove the well-posedness of \eqref{intro_eq} and the uniform ellipticity of the homogenized coefficient field. In Section~\ref{sec_lsr}, we first obtain an energy estimate for the homogenization error in Proposition~\ref{prop_hom_energy} and then prove the large-scale regularity estimate. We prove the Liouville theorem in Section~\ref{sec_lvl}, which is a consequence of the large-scale regularity estimate and a version of the Caccioppoli inequality adapted to the divergence-free setting.
\subsection{Overview of the literature} The foundational theory of homogenization for elliptic and parabolic equations with periodic coefficients can be found in the references Bensoussan, Lions, and Papanicolaou \cite{BenLioPap2011} and Jikov, Kozlov, and Ole\u{\i}nik \cite{JikKozOle1994}. The stochastic homogenization of divergence-form equations, and non-divergence-form equations without drift, was initiated by Papanicolaou and Varadhan \cite{PapVar1981,PapVar1982}, Osada \cite{Osa1983}, and Kozlov \cite{Koz1985}. In the absence of additional assumptions on the drift, the general question of stochastic homogenization for diffusion equations of the type
\begin{equation}\label{lit_eq}-\nabla\cdot a(\nicefrac{x}{\varepsilon},\o)\nabla u^\varepsilon + \frac{1}{\varepsilon}b(\nicefrac{x}{\varepsilon},\o)\cdot\nabla u^\varepsilon = f\;\;\textrm{in}\;\;U\;\;\textrm{with}\;\;u^\varepsilon =g \;\;\textrm{on}\;\;\partial U\end{equation}
remains open. The difficulty lies in constructing the invariant measure for the process from the point of view of the particle. Thus far, the construction of this measure has required additional assumptions on the drift, such as the case when $b=\nabla U$ is the gradient of a stationary field, which has been treated, for instance, by Olla in \cite{Oll1994}, and the case when $b$ is divergence-free, which will be discussed in detail below. The only other known results apply to a perturbative, strongly mixing, and isotropic regime in $d\geq 3$, which have been obtained in the discrete case by Bricmont and Kupiainen \cite{BriKup1991}, Bolthausen and Zeitouni \cite{BolZei2007}, Baur and Bolthausen \cite{BauBol2015}, and Baur \cite{Bau2016} and in the continuous case by Sznitman and Zeitouni \cite{SznZei2006} and the author \cite{Feh2017, Feh2017a,Feh2019} where \cite{Feh2017a} constructs the invariant measure. A general overview can be found in the reference \cite{Oll1994} and the book Komorowski, Landim, and Olla \cite{KomLanOll2012}.
The homogenization of \eqref{lit_eq} with divergence-free drift was initiated by \cite{Osa1983}, who considered the case of a bounded stream matrix, and Oelschl\"ager \cite{Oel1988}, who proved an annealed invariance principle and the annealed homogenization of equations like \eqref{intro_eq} on the whole space assuming the existence of an $L^2$-integrable, $\C^2$-smooth stream matrix. More recently, in the discrete case, T\'oth and Kozma \cite{KozTot2017} have proven an annealed invariance principle for the analogous discrete random walk under the so-called $\mathcal{H}_{-1}$-condition, which is equivalent to the existence of a stationary, $L^2$-integrable stream matrix. T\'oth \cite{Tot2018} then proved a quenched central limit theorem in this setting assuming the existence of an $L^{2+\d}$-integrable stream matrix, using an adaptation of Nash's moment bound. The higher $L^{d\vee(2+\d)}$-integrability assumption was introduced in Avellaneda and Majda \cite{AveMaj1991} to prove the quenched homogenization of the parabolic version of \eqref{intro_eq} on the whole space. In \cite{AveMaj1991} correctors are constructed by approximation, and therefore lack an intrinsic characterization. Related problems under more restrictive integrability assumptions have been considered by Fannjiang and Komorowski \cite{FanKom1997}, and time-dependent problems have been considered by Landim, Olla, and Yau \cite{LanOllYau1998}, Fannjiang and Komorowski \cite{FanKom1999,FanKom2002}, and Komorowski and Olla \cite{KomOll2001}. Komorowski and Olla \cite{KomOll2002} have provided a counterexample to the annealed homogenization of equations like \eqref{intro_eq} on the whole space for drifts that do not admit a square-integrable stream matrix.
The relationship between Schauder estimates and Liouville theorems for constant-coefficient elliptic equations without drift was shown by Simon \cite{Sim1997}. In the context of the periodic homogenization of divergence-form elliptic equations without drift Avellaneda and Lin \cite{AveLin1989} obtained a full hierarchy of Liouville theorems based on the large-scale regularity theory in H\"older- and $L^p$-spaces of the same authors Avellaneda and Lin \cite{AveLin1987,AveLin1987a}. Armstrong and Smart \cite{ArmSma2016} first adapted the approach of \cite{AveLin1987a} to the stochastic case and obtained a large-scale regularity theory for environments satisfying a finite range of dependence. Their methods are based on the variational characterization of solutions and quantify the convergence of certain sub- and super-additive energies. Armstrong and Mourrat \cite{ArmMou2016} extended the results of \cite{ArmSma2016} to more general mixing conditions and these works have given rise to a significant literature on the subject. A complete account of these developments can be found in the monograph Armstrong, Kuusi and Mourrat \cite{ArmKuuMou2019}, which includes applications to percolation clusters Armstrong and Dario \cite{ArmDar2018} and time-dependent environments Armstrong, Bordas, and Mourrat \cite{ArmBorMou2018}.
The results of this work are most closely related to those of Gloria, Neukamm and Otto \cite{GloNeuOtt2020}, who established a large-scale regularity theory and first-order Liouville principle for divergence-form equations without drift under the qualitative assumption of ergodicity. In particular, the homogenization flux-correction introduced in \cite{GloNeuOtt2020} is used essentially in the proof of every result in this work, and their introduction of an intrinsic excess decay is used to obtain the large-scale regularity estimate and Liouville theorem. Marahrens and Otto \cite{MarOtt2015} had previously obtained a Liouville theorem assuming a quantified form of ergodicity. Fischer and Otto \cite{FisOtt2016, FisOtt2017} extended the results of \cite{GloNeuOtt2020} to obtain a full hierarchy of Liouville theorems under a mild quantification of ergodicity. Degenerate environments were considered by Bella, the author, and Otto \cite{BelFehOtt2018} and time-dependent environments by Bella, Chiarini, and the author \cite{BelChiFeh2019}. The work \cite{GloNeuOtt2020} has similarly given rise to a substantial literature on the subject including, for instance, Gloria and Otto \cite{GloOtt2017} and Duerinckx, Gloria, and Otto \cite{DueGloOtt2020}.
\section{The extended homogenization corrector}\label{sec_prob}
In this section, we will describe how the equations \eqref{intro_stream}, \eqref{intro_gauge}, and \eqref{intro_corrector} are lifted to the probability space. Following \cite{PapVar1981}, the transformation group $\{\tau_x\}_{x\in\mathbb{R}^d}$ is used to define so-called horizontal derivatives $\{D_i\}_{i\in\{1,\ldots d\}}$: for each $i\in\{1,\ldots,d\}$,
\[\mathcal{D}(D_i)=\{f\in L^2(\O)\colon \lim_{h\rightarrow 0}\nicefrac{f(\tau_{he_i}\o)-f(\o)}{h}\;\textrm{exists strongly in}\;L^2(\O)\},\]
and $D_i\colon \mathcal{D}(D_i)\rightarrow L^2(\O)$ is defined by $D_if=\lim_{h\rightarrow 0}\nicefrac{f(\tau_{he_i}\o)-f(\o)}{h}.$
The $D_i$ are closed, densely defined operators on $L^2(\O)$. We define $\mathcal{H}^1(\O)=\cap_{i=1}^d\mathcal{D}(D_i)$ and we will write $\mathcal{H}^{-1}(\O)$ for the dual of $\mathcal{H}^1(\O)$. For $\phi\in \mathcal{H}^1(\O)$ we will write $D\phi=(D_1\phi,\ldots,D_d\phi)$ for the horizontal gradient.
A natural class of test functions can be constructed by convolution. For each $\psi\in\C^\infty_c(\mathbb{R}^d)$ and $f\in L^\infty(\O)$ we define $\psi_f\in L^\infty(\O)$ as the convolution
\[\psi_f(\o)=\int_{\mathbb{R}^d}f(\tau_x\o)\psi(x)\dx,\]
and we will write $\mathcal{D}(\O)$ for the space of all such functions. The space $\mathcal{D}(\O)$ is dense in $L^p(\O)$ for every $p\in[1,\infty)$. We will write $\mathcal{D}'(\O)$ for the dual of $\mathcal{D}(\O)$, and we will understand distributional inequalities in $\mathcal{D}'(\O)$ in the sense that, for $f\in L^1(\O)$,
\[D_if=0\;\;\textrm{if and only if}\;\;\mathbb{E}[fD_i\psi]=0\;\;\textrm{for every}\;\;\psi\in\mathcal{D}(\O).\]
For a vector field $V=(V_i)_{i\in\{1,\ldots,d\}}\in L^2(\O;\mathbb{R}^d)$ we define the distributional divergence $D\cdot V=D_iV_i$. The space of vector fields $L^2(\O;\mathbb{R}^d)$ then admits the following Helmoltz decomposition. The space of potential or curl-free fields on $\O$ is defined by
\[L^2_{\textrm{pot}}(\O)=\overline{\left\{D\psi\in L^2(\O;\mathbb{R}^d)\colon \psi\in \mathcal{H}^1(\O)\right\}}^{L^2(\O;\mathbb{R}^d)},\]
which is the $L^2(\O;\mathbb{R}^d)$-closure of the space of $\mathcal{H}^1$-gradients. The space of solenoidal or divergence-free fields is defined by
\[L^2_{\textrm{sol}}(\O)=\{V\in L^2(\O;\mathbb{R}^d)\colon D\cdot V=0\}.\]
The space $L^2(\O;\mathbb{R}^d)$ then admits the orthogonal decomposition
\[L^2(\O;\mathbb{R}^d)=L^2_{\textrm{pot}}(\O)\oplus L^2_{\textrm{sol}}(\O),\]
which can be deduced from Proposition~\ref{prop_flux} below. We will now use this framework to lift equations like \eqref{intro_stream}, \eqref{intro_gauge}, and \eqref{intro_corrector} to the probability space.
\subsection{The homogenization corrector}\label{sec_hom_cor} We will construct the homogenization corrector as a stationary gradient $\Phi_i$ in $L^2_{\textrm{pot}}(\O)$ satisfying
\begin{equation}\label{re_1}-D\cdot(A+S)(\Phi_i+e_i)=0\;\;\textrm{in}\;\;\mathcal{D}'(\O).\end{equation}
The solution is identified by approximation. We first prove that for every $\a\in(0,1)$ there exists a unique $\Phi_{i,\a}\in \mathcal{H}^1(\O)$ satisfying the equation
\begin{equation}\label{cor_approx_cor_eq}\a\Phi_{i,\a}-D\cdot (A+S)(D\Phi_{i,\a}+e_i)=0,\end{equation}
where here, in comparison to \eqref{re_1}, the proof of uniqueness is simpler and relies crucially on the stationarity of $\Phi_{i,\a}$ itself. We will then show that the $D\Phi_{i,\a}$ converge along the full sequence $\a\rightarrow 0$ in $L^2_{\textrm{pot}}(\O)$ to the unique solution of \eqref{re_1}.
The subsection is organized as follows. We will first present a general proof of sublinearity for the homogenization correctors in Proposition~\ref{prop_sublinear} below. We analyze \eqref{cor_approx_cor_eq} in Proposition~\ref{prop_approx_cor} below. Finally, in Proposition~\ref{prop_corrector} below, we prove that there exists a unique stationary gradient satisfying \eqref{re_1}. The proof of Proposition~\ref{prop_corrector} is strongly motivated by \cite[Lemma~3.27]{Oel1988} and extends \cite[Lemma~3.27]{Oel1988} to the case of a general $L^2$-integrable stream matrix. The proof of sublinearity is essentially well-known, but we include details here, in particular, to handle the less standard case $q=p_*$ below.
\begin{prop}\label{prop_sublinear} Assume \eqref{steady}. Let $p\in(1,\infty)$, let $F\in L^p(\O;\mathbb{R}^d)$ satisfy
\[D_iF_j=D_jF_i\;\;\textrm{for every}\;\;i,j\in\{1,\ldots,d\}\;\;\textrm{and}\;\;\mathbb{E}\left[F\right]=0,\]
and let $\phi\colon\mathbb{R}^d\times\O\rightarrow \mathbb{R}$ almost surely satisfy $\phi\in W^{1,p}_{\textrm{loc}}(\mathbb{R}^d)$ with $\nabla\phi(x,\omega)=F(\tau_x\omega)$. If $p<d$ and $\nicefrac{1}{p_*}=\nicefrac{1}{p}-\nicefrac{1}{d}$ we have almost surely that, for every $q\in[1,p_*]$,
\begin{equation}\label{sub_00000}\lim_{R\rightarrow\infty}\frac{1}{R}\left(\fint_{B_R}\abs{\phi}^q\right)^\frac{1}{q}=0.\end{equation}
If $p\geq d$ then \eqref{sub_00000} holds for every $q\in[1,\infty)$. If $p>d$ then $\lim_{R\rightarrow\infty}\nicefrac{\norm{\phi}_{L^\infty(B_R)}}{R}=0$.
\end{prop}
\begin{proof} We will first consider the case $p<d$ and $q=p_*$. After rescaling we observe that
\begin{equation}\label{sub_0000}\limsup_{R\rightarrow\infty}\frac{1}{R}\left(\fint_{B_R}\abs{\phi}^{p_*}\right)^\frac{1}{p_*}=\limsup_{\varepsilon\rightarrow 0}\left(\fint_{B_1}\abs{\phi^\varepsilon}^{p_*}\right)^\frac{1}{p_*},\end{equation}
for $\phi^\varepsilon(x)=\varepsilon\phi(\nicefrac{x}{\varepsilon})$. We will first prove that
\begin{equation}\label{sub_000}\limsup_{\varepsilon\rightarrow 0}\left(\fint_{B_1}\abs{\phi^\varepsilon-\fint_{B_1}\phi^\varepsilon}^{p_*}\right)^\frac{1}{p_*}=0,\end{equation}
and then show that \eqref{sub_000} implies \eqref{sub_00000}. For every $\d\in(0,1)$ let $\rho^\d$ be a standard convolution kernel of scale $\d$, and for every $\varepsilon,\d\in(0,1)$ let $\phi^{\varepsilon,\d}=\phi^\varepsilon*\rho^\d$. The triangle inequality proves that, for every $\varepsilon,\d\in(0,1)$,
\begin{align*}
\left(\fint_{B_1}\abs{\phi^\varepsilon-\fint_{B_1}\phi^\varepsilon}^{p_*}\right)^\frac{1}{p_*} & \leq \left(\fint_{B_1}\abs{\phi^{\varepsilon,\d}-\fint_{B_1}\phi^{\varepsilon,\d}}^{p_*}\right)^\frac{1}{p_*}
\\ \nonumber & \quad + \left(\fint_{B_1}\abs{\left(\phi^\varepsilon-\fint_{B_1}\phi^\varepsilon\right)-\left(\phi^{\varepsilon,\d}-\fint_{B_1}\phi^{\varepsilon,\d}\right)}^{p_*}\right)^\frac{1}{p_*}.
\end{align*}
The Sobolev inequality proves that there exists $c\in(0,\infty)$ such that, for each $\varepsilon,\delta\in(0,1)$,
\begin{align}\label{sub_0}
\left(\fint_{B_1}\abs{\phi^\varepsilon-\fint_{B_1}\phi^\varepsilon}^{p_*}\right)^\frac{1}{p_*} & \leq c\left(\left(\fint_{B_1}\abs{\nabla \phi^{\varepsilon,\d}}^p\right)^\frac{1}{p}+ \left(\fint_{B_1}\abs{\nabla\left(\phi^\varepsilon-\phi^{\varepsilon,\d}\right)}^{p}\right)^\frac{1}{p}\right).
\end{align}
For the first term on the righthand side of \eqref{sub_0}, since Jensen's inequality proves that, for every $x\in B_1$ and $\varepsilon,\d\in(0,1)$,
\[\abs{\nabla\phi^{\varepsilon,\d}(x,\omega)}^p=\abs{\int_{\mathbb{R}^d}\nabla\phi^\varepsilon(y)\rho^\delta(y-x)\dy}^p\leq\norm{\rho^\delta}_{L^\infty(\mathbb{R}^d)}\int_{B_2}\abs{\nabla\phi^\varepsilon}^p,\]
the ergodic theorem and $F\in L^p(\O;\mathbb{R}^d)$ prove that, almost surely for each $\delta\in(0,1)$,
\begin{equation}\label{sub_3}\sup_{\varepsilon\in(0,1)}\left(\sup_{x\in B_1}\abs{\nabla\phi^{\varepsilon,\d}(x,\omega)}\right)<\infty.\end{equation}
Since the ergodic theorem and $\mathbb{E}[F]=0$ prove almost surely that, as $\varepsilon\rightarrow 0$,
\[\nabla\phi^\varepsilon(x,\omega)\rightharpoonup 0\;\;\textrm{weakly in}\;\;L^p_{\textrm{loc}}(\mathbb{R}^d;\mathbb{R}^d),\]
we have, almost surely for every $\delta\in(0,1)$ and $x\in B_1$,
\begin{equation}\label{sub_1}\lim_{\varepsilon\rightarrow 0}\abs{\nabla \phi^{\varepsilon,\d}(x)}^p=\lim_{\varepsilon\rightarrow 0}\abs{\int_{\mathbb{R}^d}\nabla\phi^\varepsilon(y)\rho^\d(y-x)\dy}^p=0.\end{equation}
The dominated convergence theorem, \eqref{sub_3}, and \eqref{sub_1} then prove that, almost surely for every $\delta\in(0,1)$,
\begin{equation}\label{sub_04}\limsup_{\varepsilon\rightarrow 0}\left(\fint_{B_1}\abs{\nabla \phi^{\varepsilon,\d}}^p\right)^\frac{1}{p}=0.\end{equation}
For the second term on the righthand side of \eqref{sub_0}, the ergodic theorem proves almost surely for every $\delta\in(0,1)$ that
\begin{equation}\label{sub_4}\lim_{\varepsilon\rightarrow 0}\left(\fint_{B_1}\abs{\nabla\left(\phi^\varepsilon-\phi^{\varepsilon,\d}\right)}^{p}\right)^\frac{1}{p}=\mathbb{E}\left[\abs{F-F^\delta}^p\right]^\frac{1}{p},\end{equation}
for $F^\delta(\omega)=\int_{\mathbb{R}^d}F(\tau_y\omega)\rho^\delta(y)\dy$. Returning to \eqref{sub_0}, it follows from \eqref{sub_04} and \eqref{sub_4} that, for every $\d\in(0,1)$,
\[\limsup_{\varepsilon\rightarrow 0}\left(\fint_{B_1}\abs{\phi^\varepsilon-\fint_{B_1}\phi^\varepsilon}^{p_*}\right)^\frac{1}{p_*} \leq \mathbb{E}\left[\abs{F-F^\delta}^p\right]^\frac{1}{p}.\]
It follows from $F\in L^p(\O;\mathbb{R}^d)$ that $\lim_{\delta\rightarrow 0}\mathbb{E}\left[\abs{F-F^\delta}^p\right]^\frac{1}{p}=0$, which complete the proof of \eqref{sub_000}.
It remains to prove that \eqref{sub_000} implies \eqref{sub_00000}. The following argument appears in \cite[Lemma~2]{BelFehOtt2018}. Due to the equivalence \eqref{sub_0000}, almost surely for every $\delta\in(0,1)$ there exists $R_0\in(0,\infty)$ such that, for every $R\geq R_0$,
\[\left(\fint_{B_R}\abs{\phi-\fint_{B_R}\phi}^{p_*}\right)^\frac{1}{p_*}\leq R\delta.\]
By the triangle inequality, for every $R\in[R_0,2R_0]$, for $c\in(0,\infty)$ independent of $R$,
\begin{align*}
\abs{\fint_{B_R}\phi-\fint_{B_{R_0}}\phi} & \leq \left(\fint_{B_{R_0}}\abs{\phi-\fint_{B_R}\phi}^{p_*}\right)^\frac{1}{p_*}+ \left(\fint_{B_{R_0}}\abs{\phi-\fint_{B_{R_0}}\phi}^{p_*}\right)^\frac{1}{p_*}
\\ & \leq \left(\frac{R}{R_0}\right)^\frac{d}{p_*}R\delta+R_0\delta \leq \left(2^{(\nicefrac{d}{p_*}+1)}+1\right)R_0\delta=cR_0\delta.
\end{align*}
Therefore, for every $R\in[R_0,2R_0]$,
\[\abs{\frac{1}{R}\fint_{B_R}\phi}\leq \left(\frac{R_0}{R}\right)\abs{\frac{1}{R_0}\fint_{B_{R_0}}\phi}+c\left(\frac{R_0}{R}\right)\delta.\]
It then follows inductively that, for every $R\in[2^{k-1}R_0,2^kR_0]$,
\begin{align*}
\abs{\frac{1}{R}\fint_{B_R}\phi} & \leq \left(\frac{2^{k-1}R_0}{R}\right)\abs{\frac{1}{2^{k-1}R_0}\fint_{B_{2^{k-1}R_0}}\phi}+c\left(\frac{2^{k-1}R_0}{R}\right)\delta
\\ & \leq \abs{\frac{1}{R}\fint_{B_{R_0}}\phi} + c\left(\sum_{j=0}^\infty2^{-j}\right)\delta= \abs{\frac{1}{R}\fint_{B_{R_0}}\phi}+2c\delta.
\end{align*}
Since $\delta\in(0,1)$ was arbitrary, we have almost surely that
\begin{equation}\label{sub_5}\limsup_{R\rightarrow\infty} \abs{\frac{1}{R}\fint_{B_R}\phi}=0.\end{equation}
The triangle inequality, \eqref{sub_000}, and \eqref{sub_5} prove almost surely that
\[\limsup_{R\rightarrow\infty}\frac{1}{R}\left(\fint_{B_R}\abs{\phi}^{p_*}\right)^\frac{1}{p_*}\leq \limsup_{R\rightarrow\infty}\frac{1}{R}\left(\fint_{B_R}\abs{\phi-\int_{B_R}\phi}^{p_*}\right)^\frac{1}{p_*}+\limsup_{R\rightarrow\infty}\abs{\frac{1}{R}\fint_{B_R}\phi}=0,\]
which completes the proof for the case $p<d$ and $q=p_*$. The fact that \eqref{sub_00000} holds for the cases $p<d$ and $q\in[1,p^*)$ and $p\geq d$ and $q\in[1,\infty)$ is then a consequence of H\"older's inequality. If $p>d$, returning to \eqref{sub_000}, the Sobolev embedding theorem implies that the sequence $\{\phi^\varepsilon-\int_{B_1}\phi^\varepsilon\}_{\varepsilon\in(0,1)}$ is almost surely bounded in $\C^\alpha(B_1)$ for $\alpha=1-\nicefrac{d}{p}$. The Arzel\`a-Ascoli theorem, \eqref{sub_1}, and \eqref{sub_5} then prove almost surely that
\[\lim_{\varepsilon\rightarrow 0}\norm{\phi^\varepsilon}_{L^\infty(B_1)}=\lim_{R\rightarrow\infty}\frac{1}{R}\norm{\phi}_{L^\infty(B_R)}=0.\qedhere\]
\end{proof}
\begin{prop}\label{prop_approx_cor} Assume \eqref{steady} under the weaker assumption $S\in L^2(\O;\mathbb{R}^{d\times d})$. Let $F\in L^2(\O;\mathbb{R}^d)$ and $\alpha\in(0,1)$. Then there exists a unique $\Phi\in \mathcal{H}^1(\O)$ satisfying the equation
\[\a\Phi-D\cdot (A+S)D\Phi=-D\cdot F\;\;\textrm{in}\;\;D'(\O).\]
Furthermore, $\Phi$ satisfies the energy identity
\begin{equation}\label{cor_000}\mathbb{E}\left[\alpha \Phi^2+AD\Phi\cdot D\Phi\right]=\mathbb{E}\left[F\cdot D\Phi\right].\end{equation}
\end{prop}
\begin{proof} We will write $S=(S_{ij})_{i,j\in\{1,\ldots,d\}}\in L^2(\O;\mathbb{R}^{d\times d})$ and for every $n\in\mathbb{N}$ we define
\[S_n=((S_{ij}\wedge n)\vee(-n))_{i,j\in\{1,\ldots,d\}}.\]
The Lax-Milgram theorem proves that there exists a unique $\Phi_n\in \mathcal{H}^1(\O)$ which satisfies
\[\a\Phi_n-D\cdot (A+S_n)D\Phi_n = -D\cdot F\;\;\textrm{in}\;\;\mathcal{H}^{-1}(\O).\]
The boundedness and anti-symmetry of $S_n$, the uniform ellipticity of $A$, H\"older's inequality, and Young's inequality prove that there exists $c\in(0,\infty)$ such that, for each $n\in\mathbb{N}$,
\begin{equation}\label{cor_1}\mathbb{E}\left[\a\Phi_n^2+\abs{D\Phi_n}^2\right]\leq c \mathbb{E}\left[\abs{F}^2\right].\end{equation}
It follows from \eqref{cor_1} that there exists $\Phi\in \mathcal{H}^1(\O)$ such that, after passing to a subsequence, as $n\rightarrow\infty$,
\begin{equation}\label{cor_2}\Phi_n\rightharpoonup\Phi\;\;\textrm{weakly in}\;\;\mathcal{H}^1(\O).\end{equation}
Since $S_n\rightarrow S$ strongly in $L^2(\O,\mathbb{R}^{d\times d})$, it follows from \eqref{cor_2} and $\mathcal{D}(\O)\subseteq \mathcal{H}^1(\O)$ that $\Phi$ solves
\begin{equation}\label{cor_3}\a\Phi-D\cdot(A+S)D\Phi=-D\cdot F\;\;\textrm{in}\;\;\mathcal{D}'(\O).\end{equation}
Uniqueness is an immediate consequence of the linearity, the uniform ellipticity of $A$, and the energy estimate. Therefore, it remains only to prove the energy estimate \eqref{cor_000}. The skew-symmetry of $S$ and $D\cdot(D\cdot S)=0$ prove that, for every $\psi\in D(\O)$,
\begin{equation}\label{cor_4}\mathbb{E}\left[SD\Phi\cdot D\psi\right]=-\mathbb{E}\left[(D\cdot S)\cdot D\psi \Phi\right]=\mathbb{E}\left[\left((D\cdot S)\cdot D\Phi\right)\psi\right].\end{equation}
For each $n\in \mathbb{N}$ let $\Phi_n=(\Phi\wedge n)\vee(-n)$ and for every $\varepsilon\in(0,1)$ let $\rho^\varepsilon$ denote a standard convolution kernel of scale $\varepsilon\in(0,1)$. For every $n\in\mathbb{N}$ and $\varepsilon\in(0,1)$ let
\[\Phi_{n,\varepsilon}(\o)=\int_{\mathbb{R}^d}\Phi_n(\tau_x\o)\rho^\varepsilon(x)\dx.\]
It follows by definition that the $\Phi_{n,\varepsilon}$ are admissible test functions for \eqref{cor_3} and, after using the boundedness of $\Phi_n$ to pass to the limit $\varepsilon\rightarrow 0$, it follows from \eqref{cor_3} and \eqref{cor_4} that
\begin{equation}\label{cor_5}\mathbb{E}\left[\a\Phi\Phi_n+AD\Phi\cdot D\Phi_n\right]=-\mathbb{E}\left[(D\cdot S)D\Phi\Phi_n\right]+\mathbb{E}\left[F\cdot D\Phi_n\right].\end{equation}
Since the distributional equality $D\Phi_n=D\Phi\mathbf{1}_{\{\abs{\Phi}\leq n\}}$ proves the distributional equality $D\Phi\Phi_n=D\left(\Phi\Phi_n-\nicefrac{1}{2}\Phi_n^2\right)$, the boundedness of $\Phi_n$ and $D\cdot(D\cdot S)=0$ prove that
\begin{equation}\label{cor_6}\mathbb{E}\left[(D\cdot S)D\Phi\Phi_n\right]=\mathbb{E}\left[(D\cdot S)\cdot D\left(\Phi\Phi_n-\nicefrac{1}{2}\Phi_n^2\right)\right]=0.\end{equation}
It then follows from \eqref{cor_5}, \eqref{cor_6}, and $D\Phi_n=D\Phi\mathbf{1}_{\{\abs{\Phi}\leq n\}}$ that
\begin{equation}\label{cor_7}\mathbb{E}\left[\a\Phi\Phi_n+AD\Phi\cdot D\Phi\mathbf{1}_{\{\abs{\Phi}\leq n\}}\right]=\mathbb{E}\left[F\cdot D\Phi\textbf{1}_{\{\abs{\Phi}\leq n\}}\right].\end{equation}
The energy estimate \eqref{cor_000} then follows by the dominated convergence theorem, after passing to the limit $n\rightarrow\infty$ in \eqref{cor_7}. This completes the proof.\end{proof}
\begin{prop}\label{prop_corrector} Assume \eqref{steady} under the weaker assumption $S\in L^2(\O;\mathbb{R}^{d\times d})$. Let $F\in L^2(\O;\mathbb{R}^d)$. Then there exists a unique $\Phi\in L^2_{\textrm{pot}}(\O)$ which satisfies the equation
\begin{equation}\label{cor_0008}-D\cdot (A+S)\Phi=-D\cdot F\;\;\textrm{in}\;\;D'(\O).\end{equation}
Furthermore, $\Phi$ satisfies the energy identity
\begin{equation}\label{cor_008}\mathbb{E}\left[A\Phi\cdot\Phi\right]=\mathbb{E}\left[F\cdot \Phi\right].\end{equation}
\end{prop}
\begin{proof} For every $\a\in(0,1)$ let $\Phi_\a\in \mathcal{H}^1(\O)$ be the unique solution of
\begin{equation}\label{cor_08}\a\Phi_\a-D\cdot (A+S)D\Phi_\a=-D\cdot F\;\;\textrm{in}\;\;D'(\O).\end{equation}
It follows from \eqref{cor_000}, the uniform ellipticity, H\"older's inequality, and Young's inequality that, for some $c\in(0,\infty)$ independent of $\a\in(0,1)$,
\[\mathbb{E}\left[\a(\Phi_\a)^2+\abs{D\Phi_\a}^2\right]\leq c\mathbb{E}\left[\abs{F}^2\right].\]
Therefore, after passing to a subsequence $\a\rightarrow 0$, there exists $\Phi\in L^2_{\textrm{pot}}(\O)$ such that
\begin{equation}\label{cor_8}a\Phi_\a\rightarrow 0\;\;\textrm{strongly in}\;\;L^2(\O)\;\;\textrm{and}\;\;D\Phi_\a\rightharpoonup \Phi\;\;\textrm{weakly in}\;\;L^2_\textrm{pot}(\O).\end{equation}
It follows from \eqref{cor_08} and \eqref{cor_8} that
\begin{equation}\label{cor_9}-D\cdot (A+S)\Phi=-D\cdot F\;\;\textrm{in}\;\;D'(\O).\end{equation}
This completes the proof of existence.
We will now prove the energy identity \eqref{cor_008}. Since $\Phi\in L^2_\textrm{pot}(\O)$ is curl-free and satisfies \eqref{cor_9}, by integration let $\phi\colon\mathbb{R}^d\times\O\rightarrow \mathbb{R}$ be the unique function almost surely satisfying $\fint_{B_1}\phi = 0$, $\phi\in H^1_\textrm{loc}(\mathbb{R}^d)$ with $\nabla\phi(x,\o)=\Phi(\tau_x\o)$, and, for every $\psi\in\C^\infty_c(\mathbb{R}^d)$,
\[\int_{\mathbb{R}^d}(a+s)\nabla\phi\cdot\nabla\psi =\int_{\mathbb{R}^d}f\cdot \nabla\psi,\]
for $a(x,\omega)=A(\tau_x\o)$, $s(x,\o)=S(\tau_x\o)$, and $f(x,\o)=F(\tau_x\o)$. Since $\nabla\cdot(\nabla\cdot s)= 0$ almost surely on $\mathbb{R}^d$, a repetition of the argument from Proposition~\ref{prop_approx_cor} proves that, for every $\psi\in\C^\infty_c(\mathbb{R}^d)$,
\[\int_{\mathbb{R}^d}s\nabla\phi\cdot\nabla\psi = \int_{\mathbb{R}^d}(\nabla\cdot s)\cdot \nabla\phi \psi.\]
Therefore, almost surely for every $\psi\in\C^\infty_c(\mathbb{R}^d)$,
\begin{equation}\label{cor_11}\int_{\mathbb{R}^d}a\nabla\phi\cdot\nabla\psi+\left((\nabla\cdot s)\cdot\nabla\phi\right)\psi =\int_{\mathbb{R}^d}f\cdot \nabla\psi.\end{equation}
Let $\eta\colon\mathbb{R}^d\rightarrow[0,1]$ be a smooth function satisfying $\eta=1$ on $\overline{B}_1$ and $\eta=0$ on $\mathbb{R}^d\setminus B_2$, and for every $R\in(0,\infty)$ let $\eta_R(x)=\eta(\nicefrac{x}{R})$. For every $\varepsilon\in(0,1)$ let $\rho^\varepsilon\in\C^\infty_c(\mathbb{R}^d)$ be a standard convolution kernel of scale $\varepsilon\in(0,1)$. For every $n\in\mathbb{N}$ let $\phi_n=(\phi\wedge n)\vee(-n)$. Then for every $R\in(0,\infty)$, $\varepsilon\in(0,1)$, and $n\in\mathbb{N}$, the function $(\phi_n*\rho^\varepsilon)\eta_R$ is an admissible test function for \eqref{cor_11}. Using the boundedness of $\phi_n$ to pass to the limit $\varepsilon\rightarrow 0$, we have almost surely for every $n\in\mathbb{N}$ that
\begin{equation}\label{cor_12}\int_{\mathbb{R}^d}a\nabla\phi\cdot\nabla\phi_n\eta_R+a\nabla\phi\cdot \nabla\eta_R\phi_n +\left((\nabla\cdot s)\cdot\nabla\phi\right)\phi_n\eta_R =\int_{\mathbb{R}^d}f\cdot \nabla\phi_n\eta_R+f\cdot \nabla\eta_R\phi_n.\end{equation}
The distributional equality $\nabla\phi\phi_n=\nabla\left(\phi\phi_n-\nicefrac{1}{2}\phi_n^2\right)$, the fact that $(\nabla\cdot s)$ is divergence-free, and the skew-symmetry of $s$ prove that
\begin{equation}\label{cor_014}\int_{\mathbb{R}^d}\left((\nabla\cdot s)\cdot\nabla\phi\right)\phi_n\eta_R=-\int_{\mathbb{R}^d}\left((\nabla\cdot s)\cdot \nabla\eta_R\right)\left(\phi\phi_n-\nicefrac{1}{2}\phi_n^2\right)=\int_{\mathbb{R}^d}\left(s\nabla\phi\cdot \nabla \eta_R\right)\phi_n.\end{equation}
It then follows from \eqref{cor_12}, \eqref{cor_014}, and the distributional equality $\nabla\phi_n=\nabla\phi\mathbf{1}_{\{\abs{\phi}\leq n\}}$ that
\begin{align}\label{cor_14}
& \int_{\mathbb{R}^d}a\nabla\phi\cdot\nabla\phi\mathbf{1}_{\{\abs{\phi}\leq n\}}\eta_R-\int_{\mathbb{R}^d}f\cdot \nabla\phi\mathbf{1}_{\{\abs{\phi}\leq n\}}\eta_R
\\ \nonumber & =\int_{\mathbb{R}^d}f\cdot \nabla\eta_R\phi_n- a\nabla\phi\cdot \nabla\eta_R\phi_n -\left(s\nabla\phi\cdot \nabla \eta_R\right)\phi_n.
\end{align}
For each $R\in(0,\infty)$ let $c_R=\int_{\mathbb{R}^d}\eta_R$. It follows from the definition of $\eta_R$ that $\abs{B_R}\leq c_R\leq \abs{B_{2R}}$. We now make the choice $n=R$. It then follows from the definition of $\eta_R$, the definition of $c_R$, the definition of $\phi_n$, the uniform ellipticity, \eqref{cor_14}, and H\"older's inequality that, for some $c\in(0,\infty)$ independent of $R$,
\begin{align}\label{cor_15}
& \abs{c_R^{-1}\int_{\mathbb{R}^d}a\nabla\phi\cdot\nabla\phi\mathbf{1}_{\{\abs{\phi}\leq R\}}\eta_R-c_R^{-1}\int_{\mathbb{R}^d}f\cdot \nabla\phi\mathbf{1}_{\{\abs{\phi}\leq R\}}\eta_R}
\\ \nonumber & \leq c\left(\frac{1}{R}\left(\fint_{B_{2R}}\abs{\phi}^2\right)^\frac{1}{2}\left(\fint_{B_{2R}}\abs{f}^2+\abs{\nabla\phi}^2\right)^\frac{1}{2}+\frac{1}{R}\left(\fint_{B_{2R}}\abs{s}^2\right)^\frac{1}{2}\left(\fint_{B_{2R}}\abs{\nabla\phi}^2\phi_R^2\right)^\frac{1}{2}\right).
\end{align}
The difficulty in the proof is that, since $\phi$ is not itself stationary, it is not obvious for instance that
\begin{equation}\label{cor_16}\lim_{R\rightarrow\infty}c_R^{-1}\int_{\mathbb{R}^d}a\nabla\phi\cdot\nabla\phi\mathbf{1}_{\{\abs{\phi}\leq R\}}\eta_R =\mathbb{E}\left[a\Phi\cdot\Phi\right],\end{equation}
as is formally suggested by the ergodic theorem. We will prove \eqref{cor_16} using the sublinearity of $\phi$. For each $R\in(0,\infty)$ we write
\[c_R^{-1}\int_{\mathbb{R}^d}a\nabla\phi\cdot\nabla\phi\mathbf{1}_{\{\abs{\phi}\leq R\}}\eta_R= c_R^{-1}\int_{\mathbb{R}^d}a\nabla\phi\cdot\nabla\phi\eta_R-c_R^{-1}\int_{\mathbb{R}^d}a\nabla\phi\cdot\nabla\phi\mathbf{1}_{\{\abs{\phi}>R\}}\eta_R.\]
Since the ergodic theorem proves almost surely that
\[\mathbb{E}\left[A\Phi\cdot\Phi\right] =\lim_{R\rightarrow\infty}c_R^{-1}\int_{\mathbb{R}^d}a\nabla\phi\cdot\nabla\phi\eta_R,\]
it remains only to prove almost surely that
\begin{equation}\label{cor_18}\lim_{R\rightarrow\infty}c_R^{-1}\int_{\mathbb{R}^d}a\nabla\phi\cdot\nabla\phi\mathbf{1}_{\{\abs{\phi}>R\}}\eta_R=0.\end{equation}
Chebyshev's inequality and the definition of $\eta_R$ prove that, for $c\in(0,\infty)$ independent of $R\in(0,\infty)$,
\[c_R^{-1}\abs{\{\abs{\phi}> R\}\cap \Supp(\eta_R)}\leq c_R^{-1}\abs{\{\abs{\phi}> R\}\cap B_{2R}}\leq \frac{c}{R^2}\fint_{B_{2R}}\abs{\phi}^2,\]
from which we almost surely conclude using Proposition~\ref{prop_sublinear} that
\begin{equation}\label{cor_19}\limsup_{R\rightarrow\infty}c_R^{-1}\abs{\{\abs{\phi}> R\}\cap \Supp(\eta_R)}\leq \limsup_{R\rightarrow\infty}\frac{c}{R^2}\fint_{B_{2R}}\abs{\phi}^2=0.\end{equation}
We now exploit the stationarity of $\nabla\phi$. For each $R\in(0,\infty)$ and $K\in\mathbb{N}$, the uniform ellipticity and the definitions of $\eta_R$ and $c_R$ prove that, for some $c\in(0,\infty)$ independent of $R$ and $K$,
\begin{align}\label{cor_0000020}
\abs{c_R^{-1}\int_{\mathbb{R}^d}a\nabla\phi\cdot\nabla\phi\mathbf{1}_{\{\abs{\phi}>R\}}\eta_R} & \leq \abs{c_R^{-1}\int_{\mathbb{R}^d}a\nabla\phi\cdot\nabla\phi\mathbf{1}_{\{\abs{\phi}>R,\abs{\nabla\phi}\leq K\}}\eta_R}
\\ \nonumber & \quad +c\abs{\fint_{B_{2R}}\abs{\nabla\phi}^2\mathbf{1}_{\{\abs{\nabla\phi}>K\}}}.
\end{align}
After applying the dominated convergence theorem, the ergodic theorem, the stationarity of $\nabla\phi$, and \eqref{cor_19} to \eqref{cor_0000020}, we have almost surely for every $K\in\mathbb{N}$ that
\[\limsup_{R\rightarrow\infty}\abs{c_R^{-1}\int_{\mathbb{R}^d}a\nabla\phi\cdot\nabla\phi\mathbf{1}_{\{\abs{\phi}>R\}}\eta_R} \leq c\mathbb{E}\left[\abs{\Phi}^2\mathbf{1}_{\{\abs{\Phi}>K\}}\right].\]
Therefore, since $\Phi\in L^2_{\textrm{pot}}(\O)$, after passing to the limit $K\rightarrow\infty$ we conclude the proof of \eqref{cor_18} and therefore the proof of \eqref{cor_16}. The identical proof shows almost surely that
\begin{equation}\label{cor_020}\lim_{R\rightarrow\infty} c_R^{-1}\int_{\mathbb{R}^d}f\cdot\nabla\phi\mathbf{1}_{\{\abs{\phi}\leq R\}}=\mathbb{E}\left[F\cdot\Phi\right].\end{equation}
It remains to treat the two terms on the righthand side of \eqref{cor_15}. Proposition~\ref{prop_sublinear}, $F,\Phi\in L^2(\O;\mathbb{R}^d)$, and the ergodic theorem prove almost surely that
\begin{equation}\label{cor_20}\lim_{R\rightarrow\infty}\left[\frac{1}{R}\left(\fint_{B_{2R}}\abs{\phi}^2\right)^\frac{1}{2}\left(\fint_{B_{2R}}\abs{f}^2+\abs{\nabla\phi}^2\right)^\frac{1}{2}\right]=0.\end{equation}
For the final term on the righthand side of \eqref{cor_15}, it follows from the definition of $\phi_R$ and the triangle inequality that, for each $K\in\mathbb{N}$ and $R\in(0,\infty)$,
\begin{align*}
& \frac{1}{R}\left(\fint_{B_{2R}}\abs{s}^2\right)^\frac{1}{2}\left(\fint_{B_{2R}}\abs{\nabla\phi}^2\phi_R^2\right)^\frac{1}{2}
\\ & \leq \frac{1}{R}\left(\fint_{B_{2R}}\abs{s}^2\right)^\frac{1}{2}\left(\fint_{B_{2R}\cap\{\abs{\nabla\phi}\leq K\}}\abs{\nabla\phi}^2\phi_R^2\right)^\frac{1}{2}
\\ & \quad+ \frac{1}{R}\left(\fint_{B_{2R}}\abs{s}^2\right)^\frac{1}{2}\left(\fint_{B_{2R}\cap\{\abs{\nabla\phi}>K\}}\abs{\nabla\phi}^2\phi_R^2\right)^\frac{1}{2}
\\ & \leq \frac{K}{R}\left(\fint_{B_{2R}}\abs{s}^2\right)^\frac{1}{2}\left(\fint_{B_{2R}}\phi^2\right)^\frac{1}{2} +\left(\fint_{B_{2R}}\abs{s}^2\right)^\frac{1}{2}\left(\fint_{B_{2R}\cap\{\abs{\nabla\phi}>K\}}\abs{\nabla\phi}^2\right)^\frac{1}{2}.
\end{align*}
The ergodic theorem, the stationarity of $\nabla\phi$, $S\in L^2(\O;\mathbb{R}^{d\times d})$, and Proposition~\ref{prop_sublinear} then prove almost surely that, for some $c\in(0,\infty)$ independent of $K\in\mathbb{N}$,
\[\limsup_{R\rightarrow\infty}\frac{1}{R}\left(\fint_{B_{2R}}\abs{s}^2\right)^\frac{1}{2}\left(\fint_{B_{2R}}\abs{\nabla\phi}^2\phi_R^2\right)^\frac{1}{2}\leq c\mathbb{E}\left[\abs{\Phi}^2\mathbf{1}_{\{\abs{\Phi}>K\}}\right].\]
Since $\Phi\in L^2_{\textrm{pot}}(\O)$, after passing to the limit $K\rightarrow\infty$ we conclude almost surely that
\begin{equation}\label{cor_22}\limsup_{R\rightarrow\infty}\frac{1}{R}\left(\fint_{B_{2R}}\abs{s}^2\right)^\frac{1}{2}\left(\fint_{B_{2R}}\abs{\nabla\phi}^2\phi_R^2\right)^\frac{1}{2}=0.\end{equation}
In combination \eqref{cor_15}, \eqref{cor_16}, \eqref{cor_020}, \eqref{cor_20}, and \eqref{cor_22} prove that
\begin{equation}\label{cor_23}\mathbb{E}\left[A\Phi\cdot\Phi\right]=\mathbb{E}\left[F\cdot\Phi\right],\end{equation}
which complete the proof of the energy identity.
It remains only to prove uniqueness. Suppose that $\Phi_1, \Phi_2\in L^2_{\textrm{pot}}(\O)$ satisfy \eqref{cor_0008} and \eqref{cor_008}. Then by linearity the difference $\Phi_1-\Phi_2$ satisfies both \eqref{cor_0008} and \eqref{cor_008} with $F=0$. The uniform ellipticity and the energy identity \eqref{cor_23} prove that
\[\lambda \mathbb{E}\left[\abs{\Phi_1-\Phi_2}^2\right]\leq \mathbb{E}\left[A\cdot(\Phi_1-\Phi_2)\cdot(\Phi_1-\Phi_2)\right]=0,\]
which proves that $\Phi_1=\Phi_2$ in $L^2_\textrm{pot}(\O)$ and completes the proof. \end{proof}
\subsection{The homogenization flux corrector}\label{sec_flux_cor} We will now construct the skew-symmetric flux correctors $\sigma_i$ satisfying \eqref{intro_flux}. Let $p_d\in(1,2)$ denote the integrability exponent
\begin{equation}\label{pd_exponent}p_d=\frac{2d}{d+2}\;\;\textrm{if}\;\;d\geq 3\;\;\textrm{and}\;\;p_d=\frac{4+2\d}{4+\d}\;\;\textrm{if}\;\;d=2,\end{equation}
and for each $i\in\{1,\ldots,d\}$, using H\"older's inequality, let $Q_i\in L^{p_d}(\O;\mathbb{R}^d)$ be befined by
\[Q_i=(A+S)(\Phi_i+e_i),\]
for the corrector fields $\Phi_i\in L^2_{\textrm{pot}}(\O)$ satisfying \eqref{re_1} constructed in Proposition~\ref{prop_corrector}. We will identify the flux correctors $\sigma_i=(\sigma_{ijk})$ by their stationary gradients $\Sigma_{ijk}$ satisfying the equation
\[-D\cdot \Sigma_{ijk}=D_jQ_{ik}-D_k Q_{ij}\;\;\textrm{in}\;\;\mathcal{D}'(\O).\]
We construct the $\Sigma_{ijk}$ in Proposition~\ref{prop_flux} below. In Proposition~\ref{prop_flux_corrector} below, we prove that the resulting skew-symmetric matrices $\sigma_i$ defined on $\mathbb{R}^d$ by integration almost surely satisfy $\nabla\cdot\sigma_i= q_i$ for $q_i(x,\o)=Q_i(\tau_x\o)-\mathbb{E}[Q_i]$. The proof of existence and uniqueness for the flux correctors is an extension of \cite[Lemma~1]{BelFehOtt2018}.
\begin{prop}\label{prop_flux} Assume \eqref{steady}. Let $p\in(1,\infty)$ and let $F\in L^p(\O;\mathbb{R}^d)$. Then there exists a unique weak solution $\Phi\in L^p(\O;\mathbb{R}^d)$ of the equation
\begin{equation}\label{pf_1}-D\cdot \Phi= -D\cdot F\;\;\textrm{in}\;\;\mathcal{D}'(\O),\end{equation}
with $\mathbb{E}[\Phi]=0$ and such that, for every $i,j\in\{1,2,\ldots,d\}$,
\begin{equation}\label{pf_2}D_i\Phi_j=D_j\Phi_i\;\;\textrm{in}\;\;\mathcal{D}'(\O).\end{equation}
\end{prop}
\begin{proof} Let $p\in(1,\infty)$. We will first consider a smooth righthand side $F=(F_1,\ldots,F_d)\in \mathcal{D}(\O)^d$. The Lax-Milgram theorem proves that, for every $\alpha\in(0,1)$, there exists a unique solution $\Phi_\a\in L^2_{\textrm{pot}}(\O)$ of the equation
\[\a\Phi_\a-D\cdot D\Phi_\a= -D\cdot F\;\;\textrm{in}\;\;\mathcal{D}'(\O).\]
Since $\Phi_\a\in L^2_{\textrm{pot}}(\O)$ is mean zero and curl-free, define by integration $\phi_\a\colon\mathbb{R}^d\times\O\rightarrow\mathbb{R}$ which almost surely satisfies $\phi_\a\in H^1_{\textrm{loc}}(\mathbb{R}^d)$, $\nabla\phi_\a(x,\o)=\Phi_\a(\tau_x\o)$, and
\[a\phi_\a-\Delta\phi_\a=-\nabla\cdot f\;\;\textrm{in}\;\;\mathbb{R}^d,\]
for $f(x,\omega)=F(\tau_x\o)$. Due to the $\C^1$-boundedness of $f$, it follows from the Feynman-Kac formula that
\[\phi_\a(x,\o)=\int_0^\infty\int_{\mathbb{R}^d}e^{-\a s}(4\pi s)^{-\nicefrac{d}{2}}\exp(-\nicefrac{\abs{x-y}^2}{4s})\left(-\nabla_y\cdot f(y,\omega)\right) \dy\ds,\]
from which a direct computation proves almost surely that, for $c\in(0,\infty)$ independent of $\a\in(0,1)$,
\begin{equation}\label{cor_24}\norm{\nabla\phi_\a}_{L^\infty(\mathbb{R}^d;\mathbb{R}^d)}\leq \frac{c}{\a}.\end{equation}
Therefore, almost surely,
\begin{align*}
\nabla\phi_\a(x,\omega) & = \int_0^\infty\int_{\mathbb{R}^d} e^{-\a s}(4\pi s)^{-\nicefrac{d}{2}}\nabla_x\nabla_y\left(\exp(-\nicefrac{\abs{x-y}^2}{4s})\right)f(y,\omega) \dy\ds
\\ & = \int_{\mathbb{R}^d}\nabla_x\nabla_y K_\a(x,y)f(y,\omega) \dy,
\end{align*}
for $K_\a(x,y)=\int_0^\infty e^{-\a s}(4\pi s)^{-\nicefrac{d}{2}}\exp\left(-\nicefrac{\abs{x-y}^2}{4s}\right)\ds$.
A direct computation proves that, for $c\in(0,\infty)$ independent of $\a\in(0,1)$,
\[\abs{\nabla_xK_\a(x,y)}+\abs{\nabla_yK_\a(x,y)}\leq c\abs{x-y}^{1-d}e^{-\sqrt{\a}\abs{x-y}},\]
and, for $c\in(0,\infty)$ independent of $\a\in(0,1)$,
\begin{equation}\label{cor_25}
\abs{\nabla_x\nabla_y K_\a(x,y)} \leq c\abs{x-y}^{-d}e^{-\sqrt{\a}\abs{x-y}}.
\end{equation}
Therefore, $\nabla_x\nabla_y K_\a(x,y)$ defines a Calderon-Zygmund kernel (cf.\ eg.\ Stein \cite{Ste1970}). Let $\eta\colon\mathbb{R}^d\rightarrow [0,1]$ be a smooth function satisfying $\eta = 1$ on $\overline{B}_1$ and $\eta=0$ on $\mathbb{R}^d\setminus B_2$, and for every $R\in(0,\infty)$ let $\eta_R(x)=\eta(\nicefrac{x}{R})$. For each $R\in(0,\infty)$ let
\[\nabla\phi_{\a,R}(x,\omega)= \int_{\mathbb{R}^d}\nabla_x\nabla_y K_\a(x,y)\eta_R(y)f(y,\omega)\dy.\]
It follows almost surely from \eqref{cor_25} that, for constants $c_1,c_2\in(0,\infty)$ independent of $\a\in(0,1)$,
\begin{equation}\label{cor_26}\sup_{x\in B_{\nicefrac{R}{2}}}\abs{\nabla\phi_\a(x,\omega)-\nabla\phi_{\a,R}(x,\omega)}\leq c_1\norm{F}_{L^\infty(\O)}(\sqrt{\a}R)^{-1}\exp(-c_2\sqrt{\a}R).\end{equation}
It follows from \eqref{cor_25} and the Calderon-Zygmund estimate (cf.\ eg.\ \cite{Ste1970}) that there exists $c\in(0,\infty)$ depending on $p$ and $d$ such that
\begin{equation}\label{cor_27}\int_{\mathbb{R}^d}\abs{\nabla\phi_{\a,R}}^p\leq c\int_{\mathbb{R}^d}\abs{\eta_Rf}^p.\end{equation}
In combination \eqref{cor_26}, \eqref{cor_27}, and the definition of $\eta_R$ prove almost surely that, for every $R\in(0,\infty)$, for $c\in(0,\infty)$ depending on $p$ and $d$ but independent of $R\in(0,\infty)$,
\[\fint_{B_{\frac{R}{2}}}\abs{\nabla\phi_\a}^p\leq c\fint_{B_{2R}}\abs{f}^p+c_1\norm{F}_{L^\infty(\O)}(\sqrt{\a}R)^{-1}\exp(-c_2\sqrt{\a}R).\]
Therefore, after passing to the limit $R\rightarrow\infty$, the ergodic theorem and \eqref{cor_24} prove that, for $c\in(0,\infty)$ depending on $p$ and $d$,
\begin{equation}\label{cor_270}\mathbb{E}\left[\abs{D\Phi_\a}^p\right]\leq c\mathbb{E}\left[\abs{F}^p\right].\end{equation}
Then, after passing to a subsequence $\alpha\rightarrow 0$, the weak lower-semicontinuity of the Sobolev norm proves that there exists $\Phi\in L^2_{\textrm{pot}}(\O)$ satisfying, for $c\in(0,\infty)$ depending on $p$ and $d$,
\[-D\cdot \Phi = -D\cdot F\;\;\textrm{in}\;\;D'(\O)\;\;\textrm{with}\;\;\mathbb{E}\left[\abs{\Phi}^p\right]\leq c\mathbb{E}\left[\abs{F}^p\right].\]
The proof of existence for $F\in L^p(\O;\mathbb{R}^d)$ then follows from the density of $\mathcal{D}(\O)$ in $L^p(\O)$, the definition of $L^2_{\textrm{pot}}(\O)$, and the weak lower-semicontinuity of the Sobolev norm. It remains to prove uniqueness.
By linearity, it suffices to prove that the only $\Phi\in L^p(\O;\mathbb{R}^d)$ satisfying \eqref{pf_1} and \eqref{pf_2} with $F=0$ is $\Phi=0$. Since $\Phi$ is mean zero and curl-free let $\phi\colon\mathbb{R}^d\times\O\rightarrow \mathbb{R}$ be the unique function that almost surely satisfies $\fint_{B_1}\phi = 0$, that $\phi\in W^{1,p}_{\textrm{loc}}(\mathbb{R}^d)$ with $\nabla\phi(x,\o)=\Phi(\tau_x\o)$, and that $\phi$ is a weak solution of $-\Delta\phi = 0$ on $\mathbb{R}^d$. For every $\varepsilon\in(0,1)$ let $\rho^\varepsilon\in\C^\infty_c(\mathbb{R}^d)$ be a standard convolution kernel of scale $\varepsilon$ and let $\phi^\varepsilon=u*\rho^\varepsilon$. Then $\phi^\varepsilon$ is almost surely harmonic on $\mathbb{R}^d$ and the Feynman-Kac formula proves that there exists $c\in(0,\infty)$ such that, for every $\varepsilon\in(0,1)$ and $t\in(0,\infty)$,
\begin{align}\label{cor_29}
\abs{\nabla \phi^\varepsilon(0)} & =\abs{\int_{\mathbb{R}^d}\phi^\varepsilon(y)(4\pi t)^{-\nicefrac{d}{2}}\frac{y}{2}\exp\left(-\nicefrac{\abs{y}^2}{4t}\right)\dy}
\\ \nonumber & \leq c\left(\int_{\mathbb{R}^d}\frac{\phi^\varepsilon(\sqrt{t}y)}{\sqrt{t}}\abs{y}\exp\left(-\nicefrac{\abs{y}^2}{4}\right)\dy\right).
\end{align}
For each $R\in(0,\infty)$ there exists $c\in(0,\infty)$ independent of $R$ such that
\begin{equation}\label{cor_30}
\abs{\nabla \phi^\varepsilon(0)}\leq c\left(R^d\fint_{B_R} \frac{\phi^\varepsilon(\sqrt{t}y)}{\sqrt{t}}\dy+\int_R^\infty\left(\frac{\phi^\varepsilon(0)}{\sqrt{t}}+\fint_{B_r}\abs{\nabla \phi^\varepsilon(\sqrt{t} y)}\dy\right)r^{2d}e^{-\frac{\abs{r}^2}{4}}\dr\right).
\end{equation}
Proposition~\ref{prop_sublinear}, the ergodic theorem, and $\Phi\in L^p(\O;\mathbb{R}^d)$ prove almost surely for some $c\in(0,\infty)$ that, after passing to the limit $t\rightarrow\infty$,
\[\abs{\nabla \phi^\varepsilon(0)}\leq c\mathbb{E}\left[\abs{\Phi^\varepsilon}\right]\int_R^\infty r^{2d}e^{-\frac{r^2}{4}}\dr,\]
for $\Phi^\varepsilon(\o)=\int_{\mathbb{R}^d}\Phi(\tau_x\o)\rho^\varepsilon(x)\dx$. After passing to the limit $R\rightarrow \infty$, we conclude almost surely that $\abs{\nabla \phi^\varepsilon(0)}=0$ and therefore by stationarity that $\Phi^\varepsilon=0$. After passing to the limit $\varepsilon\rightarrow 0$, we conclude that $\Phi=0$. This completes the proof. \end{proof}
\begin{prop}\label{prop_flux_corrector} Assume \eqref{steady}. Let $p_d\in(1,\infty)$ be defined in \eqref{pd_exponent}. For every $i,j,k\in\{1,\ldots,d\}$ let $\Sigma_{ijk}\in L^{p_d}(\O;\mathbb{R}^d)$ be the unique solution of
\[-D\cdot \Sigma_{ijk}=D_j Q_{ik}-D_k Q_{ij}\;\;\textrm{in}\;\;D'(\O),\]
defined in Proposition~\ref{prop_flux} and let $\sigma_{ijk}\colon\mathbb{R}^d\times\O\rightarrow\mathbb{R}$ be the unique function that almost surely satisfies $\fint_{B_1}\sigma_{ijk}=0$, $\sigma_{ijk}\in W^{1,p_d}_{\textrm{loc}}(\mathbb{R}^d)$ with $\nabla\sigma(x,\o)=\Sigma_{ijk}(\tau_x\o)$, and, for every $\psi\in\C^\infty_c(\mathbb{R}^d)$,
\[\int_{\mathbb{R}^d}\nabla\sigma_{ijk}\cdot\nabla\psi = \int_{\mathbb{R}^d}\partial_k\psi q_{ij}-\partial_j\psi q_{ik},\]
for $q_i(x,\o)=Q_i(\tau_x\o)-\mathbb{E}[Q_i]$. Then for every $i\in\{1,\ldots,d\}$ the matrix $\sigma_i=(\sigma_{ijk})$ is skew-symmetric and almost surely satisfies
\begin{equation}\label{pfc_1}\nabla\cdot\sigma_i = q_i \;\;\textrm{in}\;\;\mathbb{R}^d\;\;\textrm{for}\;\;(\nabla\cdot \sigma_i)_j=\partial_k\sigma_{ijk}.\end{equation}
\end{prop}
\begin{proof} Let $i,j,k\in\{1,\ldots,d\}$. It follows from the uniqueness of Proposition~\ref{prop_flux} that $\Sigma_{ijk}=-\Sigma_{ikj}$ and therefore it follows from the definition of $\sigma_{ijk}$ that $\sigma_{ijk}=-\sigma_{ikj}$. This proves that $\sigma_i$ is skew-symmetric. It remains only to prove the equality \eqref{pfc_1}. This will follow from the distributional equality, for every $j\in\{1,\ldots,d\}$,
\[\Delta\left((\nabla\cdot \sigma_i)_j-q_{ij}\right)=0.\]
Indeed, using the equation satisfied by the $\sigma_{ijk}$ and the fact that $q_i$ is divergence-free, for each $j\in\{1,\ldots,d\}$ we have as distributions that
\begin{equation}\label{pfc_2} \Delta\left((\nabla\cdot \sigma_i)_j-q_{ij}\right) = \partial_s\partial_s\left(\partial_k\sigma_{ijk}-q_{ij}\right) =\Delta q_{ij}-\partial_k\partial_jq_{ik}-\Delta q_{ij} = -\partial_j(\nabla\cdot q_{ik}) = 0. \end{equation}
Equation \eqref{pfc_2} proves that, for standard convolution kernels $\rho^\varepsilon\in\C^\infty_c(\mathbb{R}^d)$ of scale $\varepsilon\in(0,1)$, for every $j\in\{1,\ldots,d\}$ and $\varepsilon\in(0,1)$,
\[\Delta\left[\left((\nabla\cdot \sigma_i)_j-q_{ij}\right)*\rho^\varepsilon\right]=0\;\;\textrm{in}\;\;\mathbb{R}^d.\]
A repetition of the arguments leading to \eqref{cor_29} and \eqref{cor_30} in the proof of Proposition~\ref{prop_flux} proves that, for each $j\in\{1,\ldots,d\}$ there exists $c^\varepsilon_j\in L^\infty(\O)$ such that almost surely
\[[\left((\nabla\cdot \sigma_i)_j-q_{ij}\right)*\rho^\varepsilon](x,\o)=c_j^\varepsilon(\o)\;\;\textrm{for every}\;\;x\in\mathbb{R}^d.\]
Since the gradient fields $\Sigma_{ijk}$ are mean zero, the stationarity of the gradient, the stationarity of the flux, and the definition of the $q_i$ prove almost surely with the ergodic theorem that, for every $j\in\{1,\ldots,d\}$,
\[0=\lim_{R\rightarrow\infty}\fint_{B_R}\left((\nabla\cdot \sigma_i)_j-q_{ij}\right)*\rho^\varepsilon=c^\varepsilon_j(\o).\]
After passing to the limit $\varepsilon\rightarrow 0$, we have almost surely that
\[\nabla\cdot\sigma_i = q_i\;\;\textrm{in}\;\;\mathbb{R}^d.\qedhere\]
\end{proof}
\subsection{The stream matrix}\label{sec_stream} In Proposition~\ref{prop_stream} below, we will prove that every mean zero, divergence-free, $L^p$-integrable vector field $B$ satisfying a finite-range of dependence admits an $L^p$-integrable stream matrix provided $p\in[2,\infty)$ and the dimension $d\geq 3$. We assume a finite range of dependence for simplicity: that is, there exists $R\in(0,\infty)$ such that for subsets $A_1,A_2\subseteq\mathbb{R}^d$ the sigma algebras
\[\sigma(B(\tau_x\o)\colon x\in A_1)\;\;\textrm{and}\;\;\sigma(B(\tau_x\o)\colon x\in A_2)\;\;\textrm{are independent whenever}\;\;\dd(A_1,A_2)\geq R.\]
In the case $p=2$, for instance, the same proof yields the existence of a stationary stream matrix provided the spatial correlations of $B$ decay faster than a square.
\begin{prop}\label{prop_stream} Assume~\eqref{steady}. Let $d\in[3,4,\ldots)$, let $p\in[2,\infty)$, and let $B\in L^p(\O;\mathbb{R}^d)$ satisfy $\mathbb{E}[B]=0$, $D\cdot B=0$, and, for some $R\in(0,\infty)$, for every $A_1,A_2\subseteq\mathbb{R}^d$,
\begin{equation}\label{ps_1}\sigma(B(\tau_x\o)\colon x\in A_1)\;\;\textrm{and}\;\;\sigma(B(\tau_x\o)\colon x\in A_2)\;\;\textrm{are independent whenever}\;\;\dd(A_1,A_2)\geq R.\end{equation}
Then there exists skew-symmetric matrix $S=(S_{jk})_{j,k\in\{1,\ldots,d\}}\in L^p(\O;\mathbb{R}^{d\times d})$ that satisfies
\[D\cdot S = B\;\;\textrm{in}\;\;L^p(\O;\mathbb{R}^d).\]
\end{prop}
\begin{proof} Let $\mathcal F_B$ denote the sigma algebra generated by $B$. It follows from \eqref{ps_1} that every $\mathcal F_B$-measurable random variable satisfies a finite range of dependence. Let $X=(X_i)_{i\in\{1,\ldots,d\}}\in L^\infty(\O;\mathbb{R}^d)$ be $\mathcal F_B$-measurable and for every $\a\in(0,1)$ let $S_\a\in\mathcal{H}^1(\O)$ denote the unique Lax-Milgram solution of the equation
\[\a S_\a-D\cdot DS_\a=D\cdot X.\]
Due to the boundedness of $X$, we have the representation
\begin{align}\label{ps_2}
S_\a(\o) & =\int_0^\infty \int_{\mathbb{R}^d}(4\pi s)^{-\frac{d}{2}}e^{-\a s-\frac{\abs{x}^2}{4s}}\frac{x}{2s}\cdot X(\tau_x\o)\dx\ds
\\ \nonumber & =(4\pi)^{-\frac{d}{2}}\int_0^\infty \int_{\mathbb{R}^d}\abs{x}^{1-d}s^{-(\frac{d}{2}+1)}e^{-\a s\abs{x}^2-\frac{1}{4s}}\frac{x}{2\abs{x}}\cdot X(\tau_x\o)\dx\ds.
\end{align}
Let $q\in\{2,4,6,\ldots\}$ be a nonzero even integer and let $\mathcal{I}_q$ denote the collection of partitions of $\{1,2,\ldots,q\}$ of the form
\[\mathcal{I}_q=\{\beta=\left(\beta_1,\ldots,\beta_{N(\beta)}\right)\colon \beta_j\in\{2,3,\ldots\}\;\forall\;j\in\{1,\ldots,N(\beta)\},\;\textrm{and}\;\sum_{j=1}^{N(\beta)}\beta_j=q\},\]
which are exactly the partitions of $\{1,2,\ldots,q\}$ that contain no singletons. We define for every $\beta\in\mathcal{I}_q$ the integral
\[I_\beta = \prod_{j=1}^{N(\beta)}\int_{\mathbb{R}^d}\int_{B_{Rq}(x_{\beta_j})^{\beta_j-1}}\prod_{k=\beta_{j-1}+1}^{\beta_j}\abs{x_k}^{1-d}\dx_{\beta_{j-1}+1}\ldots\dx_{\beta_j}\]
and observe from the assumption $d\geq 3$ and the fact that $\beta_j\geq 2$ for every $j\in\{1,\ldots,N(\beta)\}$ that, for some $c\in(0,\infty)$ independent of $\a\in(0,1)$, for every $\beta\in\mathcal{I}_q$,
\[I_\beta \leq c \prod_{j=1}^{N(\beta)}\int_{\mathbb{R}^d}\left(1\wedge \abs{x_{\beta_j}}^{1-d}\right)^{\beta_j}\dx_{\beta_j}\leq c\left(\int_0^\infty (1\wedge r^{(1-d)})\dr\right)^{N(\beta)}<\infty.\]
It then follows from the $\mathcal F_B$-measurability of $X$, \eqref{ps_1}, the fact that transformation group preserves the measure, H\"older's inequality, and an explicit calculation based on \eqref{ps_2} that, for some $c\in(0,\infty)$ independent of $\a\in(0,1)$,
\begin{equation}\label{ps_3}\mathbb{E}\left[S_\a^q\right]\leq c\mathbb{E}\left[\abs{X}^q\right]\sum_{\beta\in\mathcal{I}_q}I_\beta\leq c\mathbb{E}\left[\abs{X}^q\right].\end{equation}
Since it follows from \eqref{cor_270} that, for some $c\in(0,\infty)$ independent of $\a\in(0,1)$,
\begin{equation}\label{ps_4} \mathbb{E}\left[\abs{DS_\a}^q\right]\leq c\mathbb{E}\left[\abs{X}^q\right],\end{equation}
it follows after passing to a subsequence $\a\rightarrow 0$ that there exists $S\in L^q(\O)\cap \mathcal{H}^1(\O)$ with $DS\in L^q(\O;\mathbb{R}^d)$ such that
\[S_\a\rightharpoonup S\;\;\textrm{weakly in}\;\;L^q(\O)\;\;\textrm{and}\;\;DS_\a\rightharpoonup DS\;\;\textrm{weakly in}\;\;L^q(\O;\mathbb{R}^d).\]
It follows from Proposition~\ref{prop_flux}, \eqref{ps_3}, \eqref{ps_4}, and the weak lower-semicontinuity of the Sobolev norm that $DS\in L^q(\O;\mathbb{R}^d)$ is the unique curl-free, mean zero solution of
\begin{equation}\label{ps_6}-D\cdot DS = D\cdot X\;\;\textrm{in}\;\;D'(\O),\end{equation}
and that, for some $c\in(0,\infty)$ depending on $q\in\{2,4,6,\ldots\}$ but independent of $X$,
\begin{equation}\label{ps_7}\mathbb{E}\left[\abs{S}^q+\abs{DS}^q\right]\leq c\mathbb{E}\left[\abs{X}^q\right].\end{equation}
The density of bounded functions in $L^q(\O)$ for every $q\in\{2,4,6,\ldots\}$ proves that, for every $\mathcal F_B$-measurable $X\in L^q(\O;\mathbb{R}^d)$ there exists a unique $S\in L^q(\O)\cap \mathcal{H}^1(\O)$ with $DS\in L^q(\O;\mathbb{R}^d)$ that satisfies \eqref{ps_6} and \eqref{ps_7}. Finally, since $q\in\{2,4,6,\ldots\}$ was arbitrary, it follows from the Riesz-Thorin interpolation theorem applied to the spaces $L^p(\O,\mathcal F_B)$ for $p\in[2,\infty)$ that for every $\mathcal F_B$-measurable $X\in L^p(\O;\mathbb{R}^d)$ there exists a unique $S\in L^p(\O)\cap\mathcal{H}^1(\O)$ with $DS\in L^p(\O;\mathbb{R}^d)$ satisfying \eqref{ps_6} and \eqref{ps_7}.
Now let $B=(B_i)_{i\in\{1,\ldots,d\}}\in L^p(\O;\mathbb{R}^d)$ for some $p\in[2,\infty)$ be mean zero and divergence-free in the sense that $\mathbb{E}[B]=0$ and $D\cdot B=0$, and let $B$ satisfy a finite range of dependence. For every $j,k\in\{1,\ldots,d\}$ let $S_{jk}\in L^p(\O)\cap\mathcal{H}^1(\O)$ be the unique solution of
\[-D\cdot DS_{jk}=D_jB_k-D_kB_j.\]
The uniqueness proves that $S_{jk}=-S_{kj}$ for every $j,k\in\{1,\ldots,d\}$ and it follows from Proposition~\ref{prop_flux_corrector} and $\mathbb{E}[B]=0$ that for $S=(S_{jk})_{j,k\in\{1,\ldots,d\}}$ we have $D\cdot S=B$ in $L^p(\O;\mathbb{R}^d)$. This completes the proof. \end{proof}
\section{Quenched stochastic homogenization}\label{sec_stoch_hom}
In this section, we will prove the quenched stochastic homogenization of the equation
\begin{equation}\label{hom_2}-\nabla\cdot(a^\varepsilon+s^\varepsilon)\nabla u^\varepsilon=f\;\;\textrm{in}\;\;U\;\;\textrm{with}\;\;u^\varepsilon=g\;\;\textrm{on}\;\;\partial U.\end{equation}
The proof is based on estimating the energy of the two-scale expansion
\[w^\varepsilon=u^\varepsilon-v-\varepsilon\phi^\varepsilon_i\partial_i v,\]
where, for the gradient fields $\Phi_i\in L^2_{\textrm{pot}}(\O)$ constructed in Proposition~\ref{prop_corrector}, the physical correctors $\phi_i$ are the unique functions that almost surely satisfy $\fint_{B_1}\phi_i=0$, $\phi_i\in H^1_{\textrm{loc}}(\mathbb{R}^d)$ with $\nabla\phi_i(x,\o)=\Phi_i(\tau_x\o)$, and, for every $\psi\in\C^\infty_c(\mathbb{R}^d)$,
\begin{equation}\label{hom_03}\int_{\mathbb{R}^d}(a+s)(\nabla\phi_i+e_i)\cdot\nabla\psi=0,\end{equation}
and where $\phi_i^\varepsilon(x,\o)=\phi(\nicefrac{x}{\varepsilon},\o)$. The limit $v\in H^1(U)$ solves the homogenized equation
\begin{equation}\label{hom_3}-\nabla\cdot\overline{a}\nabla v = f\;\;\textrm{in}\;\;U\;\;\textrm{with}\;\; u=f\;\;\textrm{on}\;\;\partial U,\end{equation}
for the homogenized coefficient field $\overline{a}\in\mathbb{R}^{d\times d}$ defined for each $i\in\{1,\ldots,d\}$ by
\begin{equation}\label{hom_0003}\overline{a}e_i=\mathbb{E}\left[(A+S)(\Phi_i+e_i)\right].\end{equation}
Motivated by the analogous computation in \cite{GloNeuOtt2020}, after introducing the flux correctors $\sigma_i$ we will prove that, up to boundary terms,
\[-\nabla\cdot (a^\varepsilon+s^\varepsilon)\nabla w^\varepsilon=\nabla\cdot\left[(\varepsilon\phi^\varepsilon_i(a^\varepsilon+s^\varepsilon)-\varepsilon\sigma^\varepsilon_i)\nabla(\partial_i v)\right].\]
The strong convergence of $\nabla w^\varepsilon$ to zero in the $\varepsilon\rightarrow 0$ limit then follows formally from the $L^{d\vee(2+\d)}$-integrability of the stream matrix, Proposition~\ref{prop_sublinear}, and the regularity of $\overline{a}$-harmonic functions.
This section is organized as follows. We prove the well-posedness of \eqref{hom_2} in Proposition~\ref{prop_hom_wp} below. We prove that $\overline{a}$ is uniform elliptic in Proposition~\ref{prop_hom_ue} below, a fact which relies on the energy identity \eqref{cor_008}. Finally, we prove the quenched homogenization of \eqref{hom_2} in Theorem~\ref{thm_ts} below.
\begin{prop}\label{prop_hom_wp} Let $U\subseteq\mathbb{R}^d$ be a bounded $\C^{2,\a}$-domain for some $\a\in(0,1)$, let $a\in L^\infty(U;\mathbb{R}^{d\times d})$ be uniformly elliptic, and let $s=(s_{jk})\in H^1(U;\mathbb{R}^{d\times d})$ be skew-symmetric. Then for every $f_1\in L^2(U)$, $f_2\in L^2(U;\mathbb{R}^d)$, and $g\in W^{1,\infty}(\partial U)$ there exists a unique weak solution $u\in H^1(U)$ of the equation
\begin{equation}\label{hom_5}-\nabla\cdot (a+s)\nabla u = f_1+\nabla\cdot f_2\;\;\textrm{in}\;\;U\;\;\textrm{with}\;\; u = g\;\;\textrm{on}\;\;\partial U.\end{equation}
\end{prop}
\begin{proof} The regularity of the domain $U$ and the tubular neighborhood theorem prove that there exists a globally Lipschitz continuous function $\overline{g}\colon\mathbb{R}^d\rightarrow \mathbb{R}$ such that $\overline{g}|_{\partial U}=g$. Then by considering $\tilde{u}=u-\overline{g}$ it follows that $u\in H^1(U)$ solves \eqref{hom_5} if and only if $\tilde{u}\in H^1_0(U)$ solves
\[-\nabla\cdot (a+s)\nabla \tilde{u} = f_1+\nabla\cdot \tilde{f}_2\;\;\textrm{in}\;\;U\;\;\textrm{with}\;\; u = 0\;\;\textrm{on}\;\;\partial U,\]
for $\tilde{f}_2=f_2+a\nabla\overline{g}+s\nabla\overline{g}\in L^2(U;\mathbb{R}^d)$. It is therefore sufficient to consider the case $g=0$.
Let $f_1\in L^2(U)$ and $f_2\in L^2(U;\mathbb{R}^d)$ and for each $n\in\mathbb{N}$ let $s_n=(s^n_{jk})$ be defined by $s^n_{jk}=(s_{jk}\wedge n)\vee(-n)$. The Lax-Milgram theorem proves for every $n\in\mathbb{N}$ that there exists a unique solution $u_n\in H^1_0(U)$ of the equation
\begin{equation}\label{hom_7}-\nabla\cdot (a+s_n)\nabla u_n = f_1+\nabla\cdot f_2\;\;\textrm{in}\;\;U\;\;\textrm{with}\;\; u = 0\;\;\textrm{on}\;\;\partial U,\end{equation}
which due to the skew-symmetry of $s_n$, the uniform ellipticity, and the Poincar\'e inequality satisfies the energy inequality, for some $c\in(0,\infty)$ independent of $n$,
\begin{equation}\label{hom_8}\int_U\abs{\nabla u_n}^2\leq c\int_U\abs{f_1}^2+\abs{f_2}^2.\end{equation}
Therefore, after passing to a subsequence $n\rightarrow\infty$, it follows from \eqref{hom_7}, \eqref{hom_8}, and the strong convergence of $s_n$ to $s$ in $L^2(U;\mathbb{R}^{d\times d})$ that there exists $u\in H^1_0(U)$ such that $u_n\rightharpoonup u$ weakly in $H^1_0(U)$ and such that, for every $\psi\in\C^\infty_c(U)$,
\[\int_U(a+s)\nabla u\cdot\nabla\psi = \int_U f_1\psi -f_2\cdot \nabla \psi.\]
It remains to prove the uniqueness of $u$. By linearity, it suffices to prove that the only $u\in H^1_0(U)$ that solves \eqref{hom_9} with $f_1=0$ and $f_2=0$ is $u=0$. Since $s\in H^1(U;\mathbb{R}^{d\times d})$ is skew symmetric and since $\nabla\cdot s$ is divergence-free, we have, for every $\psi\in\C^\infty_c(U)$,
\[\int_Us\nabla u\nabla\psi = -\int_U (\nabla\cdot s)\cdot\nabla\psi u = \int_U(\nabla\cdot s)\cdot\nabla u\psi.\]
Therefore, for every $\psi\in\C^\infty_c(U)$,
\begin{equation}\label{hom_9}\int_U a\nabla u\cdot\nabla\psi + \int_U(\nabla\cdot s)\nabla u\psi = 0.\end{equation}
It follows as in the proof of Proposition~\ref{prop_corrector} that for each $n\in\mathbb{N}$ the function $u_n=(u\wedge n)\vee(-n)$ is an admissible test function for \eqref{hom_9}. The distributional equalities $\nabla u_n=\nabla u\mathbf{1}_{\{\abs{u}\leq n\}}$ and $\nabla u u_n = \nabla(uu_n-\nicefrac{1}{2}u_n^2)$ and the fact that $\nabla\cdot s$ is divergence-free then prove, for each $n\in\mathbb{N}$,
\[\int_Ua\nabla u\cdot\nabla u\mathbf{1}_{\{\abs{u}\leq n\}}=0.\]
After passing to the limit $n\rightarrow\infty$, we conclude using the uniform ellipticity and the monotone convergence theorem that $\nabla u=0$ and therefore that $u=0$. This completes the proof. \end{proof}
\begin{prop}\label{prop_hom_ue} Assume \eqref{steady}. Let $\overline{a}\in\mathbb{R}^{d\times d}$ be defined by \eqref{hom_0003}. Then, for every $\xi\in\mathbb{R}^d$,
\[\abs{\overline{a}\xi}\leq 2 \left(\Lambda+\mathbb{E}\left[\abs{S}^2\right]^\frac{1}{2}\right)\left(\sum_{i=1}^d\mathbb{E}[\abs{\Phi_i+e_i}^2]\right)^\frac{1}{2}\abs{\xi}\;\;\textrm{and}\;\;\overline{a}\xi\cdot\xi\geq\lambda\abs{\xi}^2.\]
\end{prop}
\begin{proof} The uniform ellipticity, H\"older's inequality, the linearity, and the definition of $\overline{a}$ prove that, for every $\xi\in\mathbb{R}^d$,
\begin{align*}
\abs{\overline{a}\xi} =\abs{\xi_i\mathbb{E}\left[(A+S)(\Phi_i+e_i)\right]} & \leq 2\left(\Lambda+\mathbb{E}\left[\abs{S}^2\right]^\frac{1}{2}\right)\abs{\xi_i}\mathbb{E}\left[\abs{\Phi_i+e_i}^2\right]^\frac{1}{2}
\\ & \leq 2 \left(\Lambda+\mathbb{E}\left[\abs{S}^2\right]^\frac{1}{2}\right)\left(\sum_{i=1}^d\mathbb{E}[\abs{\Phi_i+e_i}^2]\right)^\frac{1}{2}\abs{\xi}.
\end{align*}
Similarly it follows by definition that
\[\overline{a}\xi\cdot \xi = \xi_i^2\mathbb{E}\left[(A+S)(\Phi_i+e_i)\cdot e_i\right],\]
and it follow from the skew-symmetry of $S$, the energy identity \eqref{cor_008}, the fact that $-D\cdot (A+S)\Phi_i=0$ in $L^2_{\textrm{pot}}(\O)$, the uniform ellipticity, Jensen's inequality, and $\mathbb{E}[\Phi_i]=0$ that
\begin{align*}
\overline{a}\xi\cdot \xi & = \xi_i^2\mathbb{E}\left[(A+S)(\Phi_i+e_i)\cdot (\Phi_i+e_i)\right] =\xi_i^2\mathbb{E}\left[A(\Phi_i+e_i)\cdot (\Phi_i+e_i)\right],
\\ & \geq \lambda \xi_i^2\mathbb{E}\left[\abs{\Phi_i+e_i}^2\right] \geq \lambda\xi^2_i\abs{\mathbb{E}[\Phi_i+e_i]}^2=\lambda\abs{\xi}^2, &
\end{align*}
which completes the proof. \end{proof}
\begin{thm}\label{thm_ts} Assume \eqref{steady}. Let $\a\in(0,1)$, let $U\subseteq\mathbb{R}^d$ be a bounded $\C^{2,\a}$-domain, let $f\in\C^\a(U)$, and let $g\in\C^{2,\a}(\partial U)$. For every $\varepsilon\in(0,1)$ let $u^\varepsilon\in H^1(U)$ be the unique solution of \eqref{hom_2} and let $v\in H^1(U)$ be the unique solution of \eqref{hom_3}. Then, almost surely as $\varepsilon\rightarrow 0$,
\[\lim_{\varepsilon\rightarrow 0}\norm{u^\varepsilon-v-\varepsilon\phi^\varepsilon_i\partial_iv}_{H^1(U)}=0.\]
\end{thm}
\begin{proof} We will essentially study the equation satisfied by the homogenization error
\begin{equation}\label{sh_29}w^\varepsilon=u^\varepsilon-v-\varepsilon\phi^\varepsilon_i\partial_iv,\end{equation}
after introducing a cutoff to ensure that $w^\varepsilon$ vanishes along the boundary. Using the fact that $U\subseteq\mathbb{R}^d$ is a bounded $\C^{2,\a}$-domain, for each $\rho\in(0,1)$ we define $\eta_\rho\colon\overline{U}\rightarrow[0,1]$ to be a smooth cutoff function satisfying $\eta_\rho(x)=1$ if $\dd(x,\partial U)\geq 2\rho$, $\eta_\rho(x)=0$ if $\dd(x,\partial U)<\rho$, and $\abs{\nabla\eta_\rho(x)}\leq \nicefrac{c}{\rho}$ for some $c\in(0,\infty)$ independent of $\rho\in(0,1)$. For each $\varepsilon,\rho\in(0,1)$ we define
\[w^{\varepsilon,\rho}=u^\varepsilon-v-\varepsilon\phi^\varepsilon_i\eta_\rho\partial_iv\;\;\textrm{in}\;\;H^1_0(U).\]
It follows by definition that
\[\nabla w^{\varepsilon,\rho}=\nabla u^\varepsilon-\nabla v-\eta_\rho\partial_iv\nabla\phi^\varepsilon_i-\varepsilon\phi^\varepsilon_i\nabla(\eta_\rho\partial_i v).\]
Distributionally, using the equation satisfied by $u^\varepsilon$,
\[-\nabla\cdot (a^\varepsilon+s^\varepsilon)\nabla w^{\varepsilon,\rho} = f+\nabla\cdot (a^\varepsilon+s^\varepsilon)\nabla v+\nabla\cdot(a^\varepsilon+s^\varepsilon)\left(\eta_\rho\partial_iv\nabla\phi^\varepsilon_i +\varepsilon\phi^\varepsilon\nabla(\eta_\rho\partial_iv)\right),\]
and, using the equation satisfied by $v$,
\begin{align}\label{sh_30} -\nabla\cdot (a^\varepsilon+s^\varepsilon)\nabla w^{\varepsilon,\rho} & = \nabla\cdot\left[ (1-\eta_\rho)\left((a^\varepsilon+s^\varepsilon)-\overline{a}\right)\nabla v\right]
\\ \nonumber & \quad +\nabla \cdot\left[\left((a^\varepsilon+s^\varepsilon)(\nabla\phi_i+e_i)-\overline{a}e_i\right)\eta_\rho\partial_iv\right]
\\ \nonumber & \quad + \nabla\cdot\left[(a^\varepsilon+s^\varepsilon)\varepsilon\phi^\varepsilon_i\nabla(\eta_\rho\partial_iv)\right].
\end{align}
The second term on the righthand side of \eqref{sh_30} is defined for each $i\in\{1,\ldots,d\}$ by $q^\varepsilon_i(x,\o)=Q_i(\tau_{\nicefrac{x}{\varepsilon}}\o)-\mathbb{E}\left[Q_i\right]$ for the flux $Q_i$ defined by
\[Q_i=(A+S)(\Phi_i+e_i)\;\;\textrm{in}\;\;L^{p_d}(\O;\mathbb{R}^d),\]
for $p_d$ defined in \eqref{pd_exponent}. The $q^\varepsilon_i$ do not vanish in a strong sense as $\varepsilon\rightarrow 0$, and it is for this reason that we introduce the flux correctors defined in Proposition~\ref{prop_flux_corrector}. For each $i\in\{1,\ldots,d\}$ let $\sigma_i=(\sigma_{ijk})\in W^{1,p_d}_{\textrm{loc}}(\mathbb{R}^d;\mathbb{R}^{d\times d})$ be as in Proposition~\ref{prop_flux_corrector} and let $\sigma_i^\varepsilon(x,\o)=\sigma_i(\nicefrac{x}{\varepsilon},\o)$. Then, for every $\psi\in \C^\infty_c(U)$,
\begin{align*}
\int_{\mathbb{R}^d}q^\varepsilon_i\eta_\rho\partial_i v\cdot\nabla\psi & = \int_{\mathbb{R}^d}(\eta_\rho\partial_i v)q^\varepsilon_{ij}\partial_j\psi=\int_{\mathbb{R}^d}(\eta_\rho\partial_iv)\partial_k(\varepsilon\sigma^\varepsilon_{ijk})\partial_j\psi
\\ & =-\int_{\mathbb{R}^d}\varepsilon\sigma^\varepsilon_{ijk}\partial_k(\eta_\rho\partial_i v)\partial_j\psi,
\end{align*}
where the final inequality relies on the skew-symmetry. So, as distributions on $\mathbb{R}^d$,
\[\nabla\cdot\left[q^\varepsilon_i\eta_\rho\partial_i v\right]=-\nabla\cdot\left[\varepsilon\sigma^\varepsilon_i\nabla(\eta_\rho\partial_i v)\right].\]
Returning to \eqref{sh_30}, we conclude that
\begin{align}\label{sh_32}
& -\nabla\cdot (a^\varepsilon+s^\varepsilon)\nabla w^{\varepsilon,\rho}
\\ & \nonumber = \nabla\cdot\left[ (1-\eta_\rho)\left((a^\varepsilon+s^\varepsilon)-\overline{a}\right)\nabla v\right] + \nabla\cdot\left[\left(\varepsilon\phi^\varepsilon_i(a^\varepsilon+s^\varepsilon)-\varepsilon\sigma^\varepsilon_i\right)\nabla(\eta_\rho\partial_iv)\right].
\end{align}
The uniform ellipticity, H\"older's inequality, Young's inequality, and the definition of $\eta_\rho$ prove that, for some $c\in(0,\infty)$ independent of $\varepsilon,\rho\in(0,1)$, for $q_d=d\vee(2+\d)$ and $\nicefrac{1}{2_*}=\nicefrac{1}{2}-\nicefrac{1}{q_d}$,
\begin{align*}\int_U\abs{\nabla w^{\varepsilon,\rho}}^2 & \leq c\norm{\nabla v}_{L^\infty(U;\mathbb{R}^d)}^2\left(\int_U(1-\eta_\rho)^2\left(\abs{a^\varepsilon}^2+\abs{s^\varepsilon}^2\right)\right)
\\ \nonumber & \quad +c\norm{\nabla (\eta_\rho\partial_iv)}_{L^\infty(U;\mathbb{R}^d)}^2\left(\int_U \abs{a^\varepsilon}^{q_d}+\abs{s^\varepsilon}^{q_d}\right)^\frac{2}{q_d}\left(\int_U\abs{\varepsilon\phi^\varepsilon_i}^{2_*}\right)^\frac{2}{2_*}
\\ \nonumber & \quad +c\norm{\nabla (\eta_\rho\partial_iv)}^2_{L^\infty(U;\mathbb{R}^d)}\left(\int_U\abs{\varepsilon\sigma^\varepsilon_i}^2\right).
\end{align*}
The regularity of the domain and Schauder estimates (cf.\ eg.\ Gilbarg and Trudinger \cite[Chapter~6]{GilTru2001}) prove that, for some $c\in(0,\infty)$ depending on $U$,
\begin{equation}\label{sh_34}\norm{v}_{\C^{2,\alpha}(U)}\leq c\left(\norm{f}_{\C^\a(U)}+\norm{g}_{\C^{2,a}(\partial U)}\right).\end{equation}
It follows almost surely from Proposition~\ref{prop_sublinear}, \eqref{sh_34}, $\Phi_i\in L^2_{\textrm{pot}}(\O)$, $\Sigma_{ijk}\in L^{p_d}(\O;\mathbb{R}^d)$, $S\in L^{q_d}(\O;\mathbb{R}^{d\times d})$, the uniform ellipticity, the ergodic theorem, and the definition of $\eta_\rho$ that, for each $\rho\in(0,1)$, for $c\in(0,\infty)$ depending on $U$ but independent of $\rho\in(0,1)$,
\begin{equation}\label{sh_35}\limsup_{\varepsilon\rightarrow 0}\int_U\abs{\nabla w^{\varepsilon,\rho}}^2\leq c\rho\norm{\nabla v}^2_{L^\infty(U;\mathbb{R}^d)}\mathbb{E}\left[\abs{A}^2+\abs{S}^2\right].\end{equation}
Then for each $\varepsilon\in(0,1)$ let $w^\varepsilon\in H^1(U)$ be defined by \eqref{sh_29} and for every $\rho\in(0,1)$ observe that
\[\nabla w^\varepsilon=\nabla w^{\varepsilon,\rho}+\nabla \phi^\varepsilon_i(1-\eta_\rho)\partial_iv+\varepsilon\phi^\varepsilon_i\nabla\left((1-\eta_\rho)\partial_i v)\right).\]
It follows from \eqref{sh_35}, the triangle inequality, and Young's inequality that, for $c\in(0,\infty)$ independent of $\varepsilon,\rho\in(0,1)$,
\begin{align*}
& \int_U\abs{\nabla w^\varepsilon}^2
\\ & \leq c\left(\int_U\abs{\nabla w^{\varepsilon,\rho}}^2+\norm{\partial_i v}^2_{L^\infty}\int_U(1-\eta_\rho)^2\abs{\nabla\phi^\varepsilon_i}^2+\norm{\nabla((1-\eta_\rho)\partial_iv)}^2_{L^\infty}\int_U\abs{\varepsilon\phi^\varepsilon_i}^2 \right).
\end{align*}
Proposition~\ref{prop_sublinear}, \eqref{sh_34}, the definition of $\eta_\rho$, and the ergodic theorem therefore prove almost surely that for every $\rho\in(0,1)$, for $c\in(0,\infty)$ depending on $U$ but independent of $\rho$,
\[\limsup_{\varepsilon\rightarrow 0} \int_U\abs{\nabla w^\varepsilon}^2\leq c\rho\left(\norm{\nabla v}^2_{L^\infty(U;\mathbb{R}^d)}E\left[\abs{A}^2+\abs{S}^2\right]+\norm{\partial_i v}^2_{L^\infty(U)}\mathbb{E}\left[\abs{\Phi_i}^2\right]\right).\]
Passing to the limit $\rho\rightarrow 0$, we conclude that, almost surely as $\varepsilon\rightarrow 0$,
\begin{equation}\label{sh_36}\nabla w^\varepsilon\rightarrow 0\;\;\textrm{strongly in}\;\;L^2(U;\mathbb{R}^d).\end{equation}
Finally, since Proposition~\ref{prop_sublinear} and $\Phi_i\in L^2_{\textrm{pot}}(\O)$ prove that, almost surely as $\varepsilon\rightarrow 0$,
\[\varepsilon\phi^\varepsilon_i\partial_i v\rightarrow 0\;\;\textrm{strongly in}\;\;L^2(U),\]
it follows from the uniform boundedness of the $u^\varepsilon-v$ in $H^1_0(U)$, the Sobolev embedding theorem, and \eqref{sh_36} that, almost surely as $\varepsilon\rightarrow 0$,
\begin{equation}\label{sh_38} w^\varepsilon\rightarrow 0\;\;\textrm{strongly in}\;\;L^2(U).\end{equation}
In combination \eqref{sh_36} and \eqref{sh_38} complete the proof. \end{proof}
\section{The large-scale regularity estimate}\label{sec_lsr}
In this section, motivated by the methods of \cite{GloNeuOtt2020}, we will establish an almost sure intrinsic large-scale $\C^{1,\a}$-regularity estimate for solutions $u\in H^1_{\textrm{loc}}(\mathbb{R}^d)$ of
\begin{equation}\label{ls_1}-\nabla\cdot(a+s)\nabla u=0\;\;\textrm{in}\;\;\mathbb{R}^d.\end{equation}
In analogy with the the characterization of H\"older spaces by Morrey and Campanato, for each $\a\in(0,1)$ and $R\in(0,\infty)$ we define the excess $\textrm{Exc}(u;R)$ to be the large-scale $\C^{1,\a}$-Campanato semi-norm with respect to the intrinsic $(a+s)$-harmonic coordinates $(x_i+\phi_i)$:
\[\textrm{Exc}(u;R) = \inf_{\xi\in\mathbb{R}^d}\frac{1}{R^{2\a}}\fint_{B_R}\abs{\nabla u-\xi-\nabla\phi_\xi}^2.\]
Formally the homogenization of \eqref{ls_1} in $H^1(U)$ and the ergodic theorem imply that the excess is well-controlled for large radii $R$ by the regularity of an $\overline{a}$-harmonic function and the energy of the random gradient fields $\Phi_i$. The arguments of this section make this precise.
The section is organized as follows. In Propositions~\ref{prop_ue_int} and \ref{prop_ue_est} below we recall some standard results from constant-coefficient elliptic regularity theory. We estimate the energy of the two-scale expansion in Proposition~\ref{prop_hom_energy} below. We then prove the large-scale H\"older estimate and excess decay in Proposition~\ref{prop_excess} and Theorem~\ref{thm_excess} below. The proof of excess decay is most closely related to the methods of \cite{GloNeuOtt2020} in the uniformly elliptic setting, and shares aspects of the work \cite{BelFehOtt2018} in the degenerate elliptic setting. Here, in analogy with the degenerate setting, the regularity estimate comes into effect after controlling both the sublinearity of the correctors and the large-scale averages of the unbounded stream matrix. In this way Propositions~\ref{prop_hom_energy} and \ref{prop_excess} are wholly analytic and essentially deterministic, taking as input only this large-scale behavior. Theorem~\ref{thm_excess} combines these statements with the probabilistic input of Proposition~\ref{prop_sublinear} and the ergodic theorem to obtain the complete statement.
\begin{remark} In this section, we will write $a\lesssim b$ if $a\leq cb$ for a constant $c$ depending only on the dimension and ellipticity constants.\end{remark}
\begin{prop}\label{prop_ue_int} Let $\overline{a}\in\mathbb{R}^{d\times d}$ be uniformly elliptic and let $v\in H^1_{\textrm{loc}}(\mathbb{R}^d)$ be a weak solution of the equation
\begin{equation}\label{ue_1}-\nabla\cdot \overline{a}\nabla v = 0\;\;\textrm{in}\;\;B_1.\end{equation}
Then, for each $r_1< r_2\in (0,1)$ and $c\in\mathbb{R}$,
\[\int_{B_{r_1}}\abs{\nabla v}^2\lesssim \frac{1}{(r_2-r_1)^2}\int_{B_{r_2}}(v-c)^2.\]
And, for every $\rho\in(0,1)$,
\[\sup_{B_{(1-\rho)}}\left(\abs{\nabla^2 v}^2+\frac{1}{\rho^2}\abs{\nabla v}^2\right)\lesssim \frac{1}{\rho^{2(d+1)}}\int_{B_1}\abs{\nabla v}^2.\]
\end{prop}
\begin{proof} Let $r_1<r_2\in(0,1)$ and let $\eta\colon \mathbb{R}^d \rightarrow[0,1]$ be a smooth cutoff function satisfying $\eta=1$ on $\overline{B}_{r_1}$ and $\eta=0$ on $\mathbb{R}^d\setminus B_{r_2}$ with $\abs{\nabla\eta}\leq \nicefrac{2}{(r_2-r_1)}$. After testing \eqref{ue_1} with $\eta^2(v-c)$,
\[\int_{B_1}\left(\overline{a}\nabla v\cdot\nabla v\right)\eta^2 = -2\int_{B_1}\left(\overline{a}\nabla v\cdot \nabla\eta_r\right)(v-c)\eta.\]
The uniform ellipticity, H\"older's inequality, and Young's inequality prove using the definition of $\eta$ that
\begin{equation}\label{ue_2}\int_{B_{r_1}}\abs{\nabla v}^2\lesssim \frac{1}{(r_2-r_1)^2}\int_{B_{r_2}}(v-c)^2.\end{equation}
Let $K\in\mathbb{N}$ and $\rho\in(0,1)$. Since for every multi-index $\alpha=(\alpha_1,\ldots,\alpha_d)\in\mathbb{N}_0^d$ the partial derivative $\partial^{\a_1}_1\ldots\partial^{\a_d}_d v$ satisfies \eqref{ue_1}, a repeated application of \eqref{ue_2} on the subintervals of length $\nicefrac{\rho}{2K}$ proves that, for every $k\in\{1,\ldots,K\}$,
\[\int_{B_{(1-\rho)}}\abs{\nabla^kv}^2\lesssim \frac{4K^2}{\rho^2}\int_{B_{\left(1-\rho+\frac{\rho}{2K}\right)}}\abs{\nabla^{k-1}v}^2\lesssim\ldots\lesssim \frac{(4K)^{k-1}}{\rho^{2(k-1)}}\int_{B_{\left(1-\rho+\frac{(k-1)\rho}{2K}\right)}}\abs{\nabla v}^2.\]
After choosing $K=d+2$ the Sobolev embedding theorem proves that
\[\sup_{B_{(1-\rho)}}\left(\abs{\nabla^2 v}^2+\frac{1}{\rho^2}\abs{\nabla v}^2\right)\lesssim \frac{1}{\rho^{2(d+1)}}\int_{B_1}\abs{\nabla v}^2.\qedhere\] \end{proof}
\begin{prop}\label{prop_ue_est} Let $\overline{a}\in\mathbb{R}^{d\times d}$ be uniformly elliptic, let $\psi\in\C^\infty(B_1)$, and let $v\in H^1(B_1)$ be a weak solution of the equation
\[-\nabla\cdot \overline{a}\nabla v = 0\;\;\textrm{in}\;\;B_1\;\;\textrm{with}\;\;v=\psi\;\;\textrm{on}\;\;\partial B_1.\]
Then, for every $p\in[2,\infty)$ there exists $c_1,c_2\in(0,\infty)$ depending on $p$ such that
\[\norm{\nabla v}_{L^p(B_1)}\leq c_1 \norm{\nabla^{\textrm{tan}}\psi}_{L^p(\partial B_1)}\leq c_2 \norm{\nabla^{\textrm{tan}}\psi}_{L^\infty(B_1)},\]
where $\nabla^{\textrm{tan}}$ denotes the tangential derivative on $\partial B_1$.
\end{prop}
\begin{proof} We may assume without loss of generality that $\fint_{\partial B_1}\psi = 0$ since subtracting a constant does not change the gradient. Let $\eta\colon\mathbb{R} \rightarrow[0,1]$ be a smooth function satisfying $\eta=1$ on $[\nicefrac{3}{4},\infty)$ and $\eta=0$ on $(-\infty,\nicefrac{1}{4}]$. Then $\overline{\psi}(x)=\psi(\nicefrac{x}{\abs{x}})\eta(\abs{x})$ is a smooth extension of $\psi$ into $B_1$ that satisfies, using the fact that $\psi$ has average zero on $\partial B_R$,
\[\norm{\nabla\overline{\psi}}_{L^p(B_1)}\lesssim \norm{\nabla^{\textrm{tan}}\psi}_{L^p(\partial B_1)}\lesssim \norm{\nabla^{\textrm{tan}}\psi}_{L^\infty(\partial B_1)}.\]
It then follows from \cite[Theorem~7.1]{GiaMar2012} and $p\geq 2$ that, for some $c_1,c_2\in(0,\infty)$ depending on $p$,
\[\norm{\nabla v}_{L^p(B_1)}\leq c_1 \norm{\nabla^{\textrm{tan}}\psi}_{L^p(\partial B_1)}\leq c_2\norm{\nabla^{\textrm{tan}}\psi}_{L^\infty(\partial B_1)}.\qedhere\]\end{proof}
\begin{prop}\label{prop_hom_energy} Assume \eqref{steady}. Let $R\in(0,\infty)$ and let $u\in H^1(B_R)$ be a distributional solution of
\[-\nabla\cdot (a+s)\nabla u=0\;\;\textrm{in}\;\;B_R.\]
Then there exists $c\in(0,\infty)$ so that for every $\varepsilon\in(0,1)$ there exists an $\overline{a}$-harmonic function $v^\varepsilon\in H^1(B_{\nicefrac{1}{2}})$ such that, for every $\rho\in(0,\nicefrac{1}{4})$, for $q_d=d\vee(2+\d)$ and $\nicefrac{1}{2_*}=\nicefrac{1}{2}-\nicefrac{1}{q_d}$,
\begin{align*}
& \fint_{B_{\nicefrac{R}{4}}}\abs{\nabla\left( u - v^\varepsilon-\phi_i\partial_iv^\varepsilon\right)}^2
\\ \nonumber & \leq c \left(\varepsilon+\varepsilon^{1-\frac{d-1}{q_d}}\left(\fint_{B_R}\abs{s}^{q_d}\right)^\frac{1}{q_d}\right)\fint_{B_R}\abs{\nabla u}^2
\\ \nonumber & \quad + c\varepsilon^{-(d-1)}\rho^\frac{1}{2_*}\left(1+\left(\fint_{B_R}\abs{s}^{q_d}\right)^\frac{2}{q_d}\right)\fint_{B_R}\abs{\nabla u}^2
\\ \nonumber & \quad + c\rho^{-2(d+1)}R^{-2}\left[\left(1+\left(\fint_{B_R}\abs{s}^{q_d}\right)^\frac{2}{q_d}\right)\left(\fint_{B_R}\abs{\phi_i}^{2_*}\right)^\frac{2}{2_*}+\left(\fint_{B_R}\abs{\sigma_i}^2\right)\right]\fint_{B_R}\abs{\nabla u}^2.
\end{align*}
\end{prop}
\begin{proof} We will first consider the case $R=1$ and obtain the general result by scaling. Let $u\in H^1_{\textrm{loc}}(B_1)$ be an arbitrary distributional solution of the equation
\[-\nabla\cdot (a+s)\cdot \nabla u=0\;\;\textrm{in}\;\; B_1.\]
We will first prove that for every $\varepsilon\in(0,1)$ there exists an $\overline{a}$-harmonic function $v^\varepsilon\in H^1(B_{\nicefrac{1}{2}})$ such that the homogenization error $w^\varepsilon=u-v^\varepsilon-\phi_i\partial_iv^\varepsilon$ satisfies, for every $\rho\in(0,\nicefrac{1}{4})$,
\begin{align}\label{lsr_0}
& \int_{B_{\nicefrac{1}{4}}}\abs{\nabla w^\varepsilon}^2 \lesssim \left(\varepsilon+\varepsilon^{1-\frac{d-1}{q_d}}\left(\int_{B_1}\abs{s}^{q_d}\right)^\frac{1}{q_d}\right)\int_{B_1}\abs{\nabla u}^2
\\ \nonumber & \quad + \varepsilon^{-(d-1)}\rho^\frac{1}{2_*}\left(1+\left(\int_{B_r}\abs{s}^{q_d}\right)^\frac{2}{q_d}\right)\int_{B_1}\abs{\nabla u}^2
\\ \nonumber & \quad + \rho^{-2(d+1)}\left[\left(1+\left(\int_{B_r}\abs{s}^{q_d}\right)^\frac{2}{q_d}\right)\left(\int_{B_r}\abs{\phi_i}^{2_*}\right)^\frac{2}{2_*}+\left(\int_{B_r}\abs{\sigma_i}^2\right)\right]\int_{B_1}\abs{\nabla u}^2,
\end{align}
for the correctors $\phi_i$ defined in \eqref{hom_03}. Using Fubini's theorem, fix $r\in(\nicefrac{1}{2},\nicefrac{3}{4})$ such that
\begin{equation}\label{lsr_1}\int_{\partial B_r}\abs{\nabla u}^2\leq 4\int_{B_1}\abs{\nabla u}^2\;\;\textrm{and}\;\;\int_{\partial B_r}\abs{s}^{q_d}\leq 4\int_{B_1}\abs{s}^{q_d},\end{equation}
and for every $\varepsilon\in(0,1)$ let $u^\varepsilon$ denote a standard convolution of scale $\varepsilon$ of $u$ on $\partial B_r$. For each $\varepsilon\in(0,1)$ let $v^\varepsilon \in H^1(B_r)$ solve
\begin{equation}\label{lsr_2}-\nabla\cdot \overline{a}\nabla v^\varepsilon=0\;\;\textrm{in}\;\;B_r\;\;\textrm{with}\;\;v^\varepsilon=u^\varepsilon\;\;\textrm{on}\;\;\partial B_r.\end{equation}
It then follows from Dirichlet-to-Neumann estimates Fabes, Jodeit and Rivi\`ere \cite[Theorem~2.4]{FabJodRiv1978} and Stein \cite[Chapter~7]{Ste1993}, \eqref{lsr_1}, and the fact that the convolution preserves the $L^2$-norm that
\[\int_{\partial B_r}\abs{\nu\cdot \nabla v^\varepsilon} \lesssim \int_{\partial B_r}\abs{\nabla^{\textrm{tan}}v^\varepsilon}^2 = \int_{\partial B_r}\abs{\nabla^{\textrm{tan}}u^\varepsilon}^2\leq \int_{\partial B_r}\abs{\nabla u}^2\lesssim \int_{B_1}\abs{\nabla u}^2,\]
for the outward unit normal $\nu$ to $\partial B_r$. Finally, for each $\rho\in(0,\nicefrac{1}{4})$ let $\eta_\rho\colon\mathbb{R}^d\rightarrow[0,1]$ be a smooth function satisfying $\eta_\rho=1$ on $\overline{B}_{1-\rho}$, satisfying $\eta_\rho = 0$ on $\mathbb{R}^d\setminus B_{1-\nicefrac{\rho}{2}}$, and satisfying $\abs{\eta_\rho(x)}\leq \nicefrac{c}{\rho}$ for some $c\in(0,\infty)$ independent of $\rho\in(0,\nicefrac{1}{4})$. The first step will be to estimate the energy of the homogenization error $w^{\varepsilon,\rho}\in H^1_0(B_r)$ defined by
\[w^{\varepsilon,\rho}=u-v^\varepsilon-\phi_i\eta_\rho\partial_i v^\varepsilon.\]
A repetition of the derivation leading to \eqref{sh_32} proves that
\[-\nabla\cdot(a+s)\nabla w^{\varepsilon,\rho} = \nabla\cdot\left[(1-\eta_\rho)((a+s)-\overline{a})\nabla v^\varepsilon\right]+\nabla\cdot\left[\left(\phi_i(a+s)-\sigma_i\right)\nabla(\eta_\rho\partial_iv^\varepsilon)\right]\]
in $B_r$ with boundary condition $w^{\varepsilon,\rho}=u-u^\varepsilon$ on $\partial B_r$, for the flux correctors $\sigma_i$ defined in Proposition~\ref{prop_flux_corrector}. It follows from H\"older's inequality, Young's inequality, the triangle inequality, the uniform ellpticity, the definition of $\eta_\rho$, and a repetition of the argument leading to \eqref{hom_9} that
\begin{align} \label{lsr_5} & \int_{B_r}\abs{\nabla w^{\varepsilon,\rho}}^2 \lesssim \abs{\int_{\partial B_r} (u-u^\varepsilon)\nu\cdot \left((a+s)\nabla u-\overline{a}\nabla v\right)}
\\ \nonumber & +\rho^{\frac{1}{2_*}}\left(1+\left(\int_{B_r}\abs{s}^{q_d}\right)^\frac{2}{q_d}\right)\left(\int_{B_r}\abs{\nabla v^\varepsilon}^{2\cdot 2_*}\right)^\frac{1}{2_*}
\\ \nonumber &+ \sup_{B_{(1-\nicefrac{\rho}{2})}}(\abs{\nabla(\partial_i v^\varepsilon)}^2+\frac{1}{\rho^2}\abs{\partial_iv^\varepsilon}^2)[(1+(\int_{B_r}\abs{s}^{q_d})^\frac{2}{q_d})(\int_{B_r}\abs{\phi_i}^{2_*})^\frac{2}{2_*}+(\int_{B_r}\abs{\sigma_i}^2)].
\end{align}
H\"older's inequality proves that, for the first term on the righthand side of \eqref{lsr_5},
\begin{align}\label{lsr_006}
& \abs{\int_{\partial B_r} (u-u^\varepsilon)\nu\cdot \left((a+s)\nabla u-\overline{a}\nabla v\right)}
\\ \nonumber & \lesssim \left(\left(\int_{\partial B_r}\abs{\nabla u}^2\right)^\frac{1}{2}+\left(\int_{\partial B_r}\abs{\nabla v}^2\right)^\frac{1}{2}\right)\left(\int_{\partial B_r}\abs{u-u^\varepsilon}^{2}\right)^\frac{1}{2}
\\ \nonumber & \quad + \left(\int_{\partial B_r}\abs{s}^{q_d}\right)^\frac{1}{q_d}\left(\int_{\partial B_r}\abs{\nabla u}^2\right)^\frac{1}{2}\left(\int_{\partial B_r}\abs{u-u^\varepsilon}^{2_*}\right)^\frac{1}{2_*}.
\end{align}
Since for each $p\in[1,\infty)$ we have the convolution estimate
\[\left(\int_{\partial B_r}\abs{u-u^\varepsilon}^p\right)^\frac{1}{p}\lesssim \varepsilon\left(\int_{\partial B_r}\abs{\nabla^{\textrm{tan}} u}^p\right)^\frac{1}{p},\]
it follows from \eqref{lsr_1} that the first term on the righthand side of \eqref{lsr_006} is bounded by
\begin{equation}\label{lsr_00006}\left(\left(\int_{\partial B_r}\abs{\nabla u}^2\right)^\frac{1}{2}+\left(\int_{\partial B_r}\abs{\nabla v}^2\right)^\frac{1}{2}\right)\left(\int_{\partial B_r}\abs{u-u^\varepsilon}^{2}\right)^\frac{1}{2}\lesssim \varepsilon\int_{\partial B_r}\abs{\nabla u}^2 \lesssim \varepsilon\int_{B_1}\abs{\nabla u}^2.\end{equation}
For the second term on the righthand side of \eqref{lsr_006}, since the definition of the convolution proves that $\fint_{B_r}(u-u^\varepsilon)=0$, it follows from the Sobolev inequality, the fact that the convolution does not increase $L^p$-norms for $p\in[1,\infty)$, and the triangle inequality that
\begin{equation}\label{lsr_6}\left(\int_{\partial B_r}\abs{u-u^\varepsilon}^{2_*}\right)^\frac{1}{2_*}\lesssim \left(\int_{\partial B_r}\abs{\nabla^{\textrm{tan}} (u-u^\varepsilon)}^q\right)^\frac{1}{q}\lesssim \left(\int_{\partial B_r}\abs{\nabla u}^q\right)^\frac{1}{q},\end{equation}
for $q\in(1,2)$ defined by $\frac{1}{q}=\frac{1}{2_*}+\frac{1}{d-1}$. Interpolating between the the convolution estimate with $p=2_*$ and \eqref{lsr_6} proves with \eqref{lsr_1} that
\begin{equation}\label{lsr_8} \left(\int_{\partial B_r}\abs{u-u^\varepsilon}^{2_*}\right)^\frac{1}{2_*}\lesssim \varepsilon^{1-\frac{d-1}{q_d}}\left(\int_{\partial B_r}\abs{\nabla u}^2\right)^\frac{1}{2}\lesssim \varepsilon^{1-\frac{d-1}{q_d}}\left(\int_{B_r}\abs{\nabla u}^2\right)^\frac{1}{2}.\end{equation}
In combination \eqref{lsr_1} and \eqref{lsr_8} prove that the second term on the righthand side of \eqref{lsr_006} satisfies
\[\left(\int_{\partial B_r}\abs{s}^{q_d}\right)^\frac{1}{q_d}\left(\int_{\partial B_r}\abs{\nabla u}^2\right)^\frac{1}{2}\left(\int_{\partial B_r}\abs{u-u^\varepsilon}^{2_*}\right)^\frac{1}{2_*}\lesssim \varepsilon^{1-\frac{d-1}{q_d}}\left(\int_{B_1}\abs{s}^{q_d}\right)^\frac{1}{q_d}\int_{B_1}\abs{\nabla u}^2.\]
Returning to \eqref{lsr_006}, it follows from \eqref{lsr_00006} that
\begin{equation}\label{lsr_10} \abs{\int_{\partial B_r} (u-u^\varepsilon)\nu\cdot \left((a+s)\nabla u-\overline{a}\nabla v\right)}\lesssim \left(\varepsilon+\varepsilon^{1-\frac{d-1}{q_d}}\left(\int_{B_1}\abs{s}^{q_d}\right)^\frac{1}{q_d}\right)\int_{B_1}\abs{\nabla u}^2.\end{equation}
For the second term on the righthand side of \eqref{lsr_5}, since it follows from H\"older's inequality, the definition of the convolution kernel, and \eqref{lsr_1} that
\begin{equation}\label{lsr_11}\sup_{\partial B_r}\abs{\nabla^{\textrm{tan}} u^\varepsilon}\lesssim \varepsilon^{-\frac{d-1}{2}}\left(\int_{\partial B_r}\abs{\nabla^{\textrm{tan}} u}^2\right)^\frac{1}{2}\lesssim \varepsilon^{-\frac{d-1}{2}}\left(\int_{B_1}\abs{\nabla u}^2\right)^\frac{1}{2},\end{equation}
it follows from Proposition~\ref{prop_ue_int} and \eqref{lsr_11} that
\begin{equation}\label{lsr_12}\left(\int_{B_r}\abs{\nabla v^\varepsilon}^{2\cdot 2_*}\right)^\frac{1}{2_*}\lesssim \varepsilon^{-(d-1)}\int_{B_1}\abs{\nabla u}^2.\end{equation}
Therefore, using \eqref{lsr_12}, the second term on the righthand side of \eqref{lsr_5} is bounded by
\begin{align}\label{lsr_14}& \rho^{\frac{1}{2_*}}\left(1+\left(\int_{B_r}\abs{s}^{q_d}\right)^\frac{2}{q_d}\right)\left(\int_{B_r}\abs{\nabla v^\varepsilon}^{2\cdot 2_*}\right)^\frac{1}{2_*}
\\ \nonumber & \lesssim \varepsilon^{-(d-1)}\rho^\frac{1}{2_*}\left(1+\left(\int_{B_r}\abs{s}^{q_d}\right)^\frac{2}{q_d}\right)\int_{B_1}\abs{\nabla u}^2.\end{align}
For the final term of \eqref{lsr_5}, it follows from Proposition~\ref{prop_ue_int} that
\begin{align}\label{lsr_15}
& \sup_{B_{(1-\nicefrac{\rho}{2})}}(\abs{\nabla(\partial_i v^\varepsilon)}^2+\frac{1}{\rho^2}\abs{\partial_iv^\varepsilon}^2)[(1+(\int_{B_r}\abs{s}^{q_d})^\frac{2}{q_d})(\int_{B_r}\abs{\phi_i}^{2_*})^\frac{2}{2_*}+(\int_{B_r}\abs{\sigma_i}^2)]
\\ \nonumber & \lesssim \rho^{-2(d+1)}\left[\left(1+\left(\int_{B_r}\abs{s}^{q_d}\right)^\frac{2}{q_d}\right)\left(\int_{B_r}\abs{\phi_i}^{2_*}\right)^\frac{2}{2_*}+\left(\int_{B_r}\abs{\sigma_i}^2\right)\right]\int_{B_1}\abs{\nabla u}^2.
\end{align}
For every $\varepsilon\in(0,1)$ let $w^\varepsilon=u-v^\varepsilon-\phi_i\partial_i v^\varepsilon$. It follows from $\rho\in(0,\nicefrac{1}{4})$ and the definition of $\eta_\rho$ that $w^\varepsilon=w^{\varepsilon,\rho}$ in $B_{\nicefrac{1}{4}}$ and therefore it follows from \eqref{lsr_5}, \eqref{lsr_10}, \eqref{lsr_14}, and \eqref{lsr_15} that, for each $\rho\in(0,\nicefrac{1}{4})$,
\begin{align}\label{lsr_17}
& \int_{B_{\nicefrac{r}{4}}}\abs{\nabla w^\varepsilon}^2 \lesssim \left(\varepsilon+\varepsilon^{1-\frac{d-1}{q_d}}\left(\int_{B_1}\abs{s}^{q_d}\right)^\frac{1}{q_d}\right)\int_{B_1}\abs{\nabla u}^2
\\ \nonumber & \quad + \varepsilon^{-(d-1)}\rho^\frac{1}{2_*}\left(1+\left(\int_{B_1}\abs{s}^{q_d}\right)^\frac{2}{q_d}\right)\int_{B_1}\abs{\nabla u}^2
\\ \nonumber & \quad + \rho^{-2(d+1)}\left[\left(1+\left(\int_{B_1}\abs{s}^{q_d}\right)^\frac{2}{q_d}\right)\left(\int_{B_1}\abs{\phi_i}^{2_*}\right)^\frac{2}{2_*}+\left(\int_{B_1}\abs{\sigma_i}^2\right)\right]\int_{B_1}\abs{\nabla u}^2,
\end{align}
which completes the proof of \eqref{lsr_0} with $v^\varepsilon\in H^1(B_{\nicefrac{1}{2}})$ defined by \eqref{lsr_2}. It then follows by scaling that, for each $R\in(0,\infty)$, for any $u\in H^1(B_R)$ that is a weak solution of
\[-\nabla\cdot (a+s)\cdot \nabla u=0\;\;\textrm{in}\;\; B_R,\]
there exists for every $\varepsilon\in(0,1)$ an $\overline{a}$-harmonic function $v^\varepsilon\in H^1(B_{\nicefrac{R}{2}})$ such that the homogenization error $w=u-v^\varepsilon-\phi_i\partial_i v^\varepsilon$ satisfies, for every $\rho\in(0,\nicefrac{1}{4})$, for some $\overline{c}\in(0,\infty)$ independent of $R$, $\varepsilon$, and $\rho$,
\begin{align}\label{lsr_18}
& \fint_{B_{\nicefrac{R}{4}}}\abs{\nabla w^\varepsilon}^2 \leq\overline{c} \left(\varepsilon+\varepsilon^{1-\frac{d-1}{q_d}}\left(\fint_{B_R}\abs{s}^{q_d}\right)^\frac{1}{q_d}\right)\fint_{B_R}\abs{\nabla u}^2
\\ \nonumber & \quad + \overline{c}\varepsilon^{-(d-1)}\rho^\frac{1}{2_*}\left(1+\left(\fint_{B_R}\abs{s}^{q_d}\right)^\frac{2}{q_d}\right)\fint_{B_R}\abs{\nabla u}^2
\\ \nonumber & \quad + \overline{c}\rho^{-2(d+1)}R^{-2}\left[\left(1+\left(\fint_{B_R}\abs{s}^{q_d}\right)^\frac{2}{q_d}\right)\left(\fint_{B_R}\abs{\phi_i}^{2_*}\right)^\frac{2}{2_*}+\left(\fint_{B_R}\abs{\sigma_i}^2\right)\right]\fint_{B_R}\abs{\nabla u}^2.
\end{align}
The proof follows by considering the rescalings $u^R(x)=R^{-1}u(Rx)$, $\phi^R_i(x)=R^{-1}\phi_i(Rx)$, and $\sigma_i^R(x)=R^{-1}\sigma_i(x)$ and by repeating the argument leading to \eqref{lsr_17} to obtain an $\overline{a}$-harmonic function $\tilde{v}^\varepsilon$ in $H^1(B_{\nicefrac{1}{2}})$ such that the homogenization error $w^{R,\varepsilon}(x)=u^R-v^\varepsilon-\phi_i^R\partial_i\tilde{v}^\varepsilon$ satisfies \eqref{lsr_17} with $\phi_i^R$, $\sigma_i^R$, and $s^R(x)=s(Rx)$. We then define $v^\varepsilon(x)=R\tilde{v}^\varepsilon(\nicefrac{x}{R})$ and obtain \eqref{lsr_18} from \eqref{lsr_17} after rescaling. This completes the proof. \end{proof}
\begin{prop}\label{prop_excess} Assume \eqref{steady} and let $q_d=d\vee(2+\d)$. For every $\a\in(0,1)$ there exist $c_\a,C_\a\in(0,\infty)$ such that if for any $R_1<R_2\in(0,\infty)$ we have, for every $R\in[R_1,R_2]$ and $i\in\{1,\ldots,d\}$,
\[\frac{1}{R}\left(\fint_{B_R}\abs{\phi_i}^{2_*}\right)^\frac{1}{2_*}+\frac{1}{R}\left(\fint_{B_R}\abs{\sigma_i}^2\right)^\frac{1}{2}\leq \nicefrac{1}{C_\a},\]
and we have, for every $R\in[R_1,R_2]$,
\[\left(\fint_{B_R}\abs{s}^{q_d}\right)^\frac{1}{q_d}\leq\mathbb{E}\left[\abs{S}^{q_d}\right]^\frac{1}{q_d}+1\;\;\textrm{and}\;\;\sum_{i=1}^d\fint_{B_R}\abs{\Phi_i+e_i}^2\leq \sum_{i=1}^d\mathbb{E}\left[\abs{\Phi_i+e_i}^2\right]+1,\]
then every weak solution $u\in H^1_{\textrm{loc}}(\mathbb{R}^d)$ of the equation
\[-\nabla\cdot (a+s)\nabla u = 0 \;\;\textrm{in}\;\;\mathbb{R}^d,\]
satisfies
\[R_1^{-2\a}\textrm{Exc}(u;R_1)\leq c_\a R_2^{-2\a}\textrm{Exc}(u;R_2).\]
\end{prop}
\begin{proof} Let $u\in H^1_{\textrm{loc}}(\mathbb{R}^d)$ be a distributional solution of $-\nabla\cdot (a+s)\nabla u = 0$ in $\mathbb{R}^d$, let $R\in(0,\infty)$, and for each $\varepsilon\in(0,1)$ let $w^\varepsilon$ be defined in \eqref{lsr_18} on $B_{\nicefrac{R}{2}}$. We will first show that, for each $\delta_0\in(0,1)$, there exists $C_1\in(1,\infty)$ depending on $\delta_0$ such that if
\[\left(\fint_{B_R}\abs{s}^{q_d}\right)^\frac{1}{q_d}\leq \mathbb{E}\left[\abs{S}^{q_d}\right]^\frac{1}{q_d}+1,\]
and such that if
\[\sum_{i=1}^d\left(\frac{1}{R^2}\left(\fint_{B_R}\abs{\phi_i}^{2_*}\right)^\frac{2}{2_*}+\frac{1}{R^2}\fint_{B_R}\abs{\sigma_i}^2\right)\leq \frac{1}{C_1},\]
then there exists a deterministic $\varepsilon_0\in(0,1)$ depending on $\delta_0$ such that
\begin{equation}\label{lsr_19}\fint_{B_{\nicefrac{R}{4}}}\abs{\nabla w^{\varepsilon_0}}^2\leq \delta_0\fint_{B_R}\abs{\nabla u}^2,\end{equation}
for $w^{\varepsilon_0}$ defined in \eqref{lsr_18}. First fix $\varepsilon_0\in(0,1)$ such that
\begin{equation}\label{lsr_20}\overline{c}\left(\varepsilon_0+\varepsilon_0^{1-\frac{d-1}{q_d}}\left(\mathbb{E}\left[\abs{S}^{q_d}\right]^{\frac{2}{q_d}}+1\right)\right)<\nicefrac{\delta_0}{3}.\end{equation}
Then fix $\rho_0\in(0,\nicefrac{1}{4})$ such that
\begin{equation}\label{lsr_21}\overline{c}\left(\varepsilon_0^{-(d-1)}\rho_0^\frac{1}{2_*}\left(2+\mathbb{E}\left[\abs{S}^{q_d}\right]^\frac{2}{q_d}\right)\right)<\nicefrac{\delta_0}{3}.\end{equation}
Finally fix $C_1\in(1,\infty)$ satisfying
\begin{equation}\label{lsr_22}\overline{c}\rho_0^{-2(d+1)}\left(2+\mathbb{E}\left[\abs{S}^{q_d}\right]^\frac{2}{q_d}\right)< \nicefrac{C_1\delta_0}{3}.\end{equation}
The claim \eqref{lsr_19} then follows from Proposition~\ref{prop_hom_energy}, \eqref{lsr_20}, \eqref{lsr_21}, and \eqref{lsr_22}.
We will now prove that there exists $\theta_0\in(0,\nicefrac{1}{4})$ and $C_2\in(1,\infty)$ such that for any $R\in(0,\infty)$ satisfying, for every $r\in[\theta_0R,R]$,
\begin{equation} \label{lsr_103}
\sum_{i=1}^d\frac{1}{r}\left(\fint_{B_r}\abs{\phi_i}^{2_*}\right)^\frac{1}{2_*}+\frac{1}{r}\left(\fint_{B_{r}}\abs{\sigma_i}^2\right)^\frac{1}{2}\leq \nicefrac{1}{C_2}, \end{equation}
and, for every $r\in[\theta_0R,R]$,
\begin{equation}\label{lsr_104}\left(\fint_{B_{r}}\abs{s}^{q_d}\right)^\frac{1}{q_d}\leq\mathbb{E}\left[\abs{S}^{q_d}\right]^\frac{1}{q_d}+1\;\;\textrm{and}\;\;\sum_{i=1}^d\fint_{B_{r}}\abs{\Phi_i+e_i}^2\leq \sum_{i=1}^d\mathbb{E}\left[\abs{\Phi_i+e_i}^2\right]+1,\end{equation}
we have the exact inequality
\[(\theta_0R)^{-2\a}\textrm{Exc}(u;\theta_0R)\leq R^{-2\a}\textrm{Exc}(u;R).\]
For each $\varepsilon\in(0,1)$ let $\xi^\varepsilon=\nabla v^{\varepsilon}(0)\in\mathbb{R}^d$ for $v^{\varepsilon}$ defined in \eqref{lsr_18} and observe that
\[u-\xi^\varepsilon-\nabla\phi_{\xi^\varepsilon} = \nabla w^{\varepsilon}+(\nabla v^{\varepsilon}-\nabla v^{\varepsilon}(0))+\nabla \phi_i(\partial_iv^{\varepsilon}-\partial_iv^{\varepsilon}(0))+\phi_i\nabla(\partial_iv^{\varepsilon})\;\;\textrm{on}\;\;B_{\nicefrac{R}{2}},\]
for $w^\varepsilon$ defined in \eqref{lsr_18}. For every $r\in(0,\nicefrac{R}{4})$ the triangle inequality, Young's inequality, and the mean value theorem prove that
\begin{align*}
& \int_{B_r}\abs{u-\xi^\varepsilon-\nabla\phi_{\xi^\varepsilon}}^2
\\ & \lesssim \int_{B_{\nicefrac{R}{4}}}\abs{\nabla w^{\varepsilon}}^2+r^2\sup_{B_r}\abs{\nabla\partial_i v^{\varepsilon}}^2\int_{B_r}\abs{e_i+\nabla \phi_i}^2+\sup_{B_r}\abs{\nabla \partial_iv^{\varepsilon}}^2\int_{B_r}\abs{\phi_i}^2.
\end{align*}
Proposition~\ref{prop_ue_int} proves that
\[\int_{B_r}\abs{u-\xi^\varepsilon-\nabla\phi_{\xi^\varepsilon}}^2\lesssim \int_{B_{\nicefrac{R}{4}}}\abs{\nabla w^{\varepsilon}}^2+\left(\frac{r^2}{R^2}\int_{B_r}\abs{e_i+\nabla \phi_i}^2+\frac{1}{R^2}\int_{B_r}\abs{\phi_i}^2\right)\fint_{B_R}\abs{\nabla u}^2.\]
After dividing by $r^d$, for some $\overline{c}\in(0,\infty)$ independent of $\varepsilon$, $r$, and $R$,
\begin{align*}
& \fint_{B_r}\abs{u-\xi^\varepsilon-\nabla\phi_{\xi^\varepsilon}}^2
\\ & \leq\overline{c}\left(\frac{R^d}{r^d}\fint_{B_{\nicefrac{R}{4}}}\abs{\nabla w^{\varepsilon}}^2+ \left(\frac{r^2}{R^2}\left(\fint_{B_r}\abs{e_i+\nabla \phi_i}^2+\frac{1}{R^2}\fint_{B_r}\abs{\phi_i}^2\right)\right)\fint_{B_R}\abs{\nabla u}^2\right).
\end{align*}
Since $\a\in(0,1)$ fix $\theta_0\in(0,\nicefrac{1}{4})$ such that
\begin{equation}\label{lsr_37}\overline{c}\theta_0^2\left(\sum_{i=1}^d\mathbb{E}\left[\abs{\Phi_i+e_i}^2\right]+2\right)\leq \nicefrac{\theta_0^{2\a}}{2},\end{equation}
then fix $\delta_0\in(0,1)$ such that
\begin{equation}\label{lsr_38}\overline{c}\theta_0^{-d}\delta_0\leq \nicefrac{\theta_0^{2\a}}{2},\end{equation}
and let $C_2\in(1,\infty)$ and $\varepsilon_0\in(0,1)$ satisfy the conclusion of \eqref{lsr_19} for this $\delta_0$. The definition of the excess, $C_2\in(1,\infty)$, \eqref{lsr_19}, \eqref{lsr_103}, and \eqref{lsr_104} then prove after choosing $r=\theta_0R$ that
\begin{align}\label{lsr_36}
\textrm{Exc}(u;\theta_0 R) & \leq \fint_{B_{\theta_0 R}}\abs{u-\xi^{\varepsilon_0}-\nabla\phi_{\xi^{\varepsilon_0}}}^2
\\ & \leq\overline{c} \left(\delta_0\theta_0^{-d}+\theta_0^2\left(\sum_{i=1}^d\mathbb{E}\left[\abs{\Phi_i+e_i}^2\right]+2\right)\right)\fint_{B_R}\abs{\nabla u}^2.\nonumber
\end{align}
In combination \eqref{lsr_37}, \eqref{lsr_38}, and \eqref{lsr_36} prove that
\begin{equation}\label{lsr_40}\textrm{Exc}(u;\theta_0R)\leq \theta_0^{2\alpha}\fint_{B_R}\abs{\nabla u}^2.\end{equation}
We now observe that for every $\xi\in\mathbb{R}^d$ the function $u-\xi\cdot x-\phi_\xi\in H^1_{\textrm{loc}}(\mathbb{R}^d)$ solves
\begin{equation}\label{lsr_41}-\nabla\cdot(a+s)\nabla (u-\xi\cdot x-\phi_\xi)=0\;\;\textrm{in}\;\;\mathbb{R}^d,\end{equation}
and by definition of the excess and linearity we have
\[\textrm{Exc}(u;\theta_0 R)=\textrm{Exc}(u-\xi\cdot x-\phi_\xi;\theta_0 R)\;\;\textrm{for every}\;\;\xi\in\mathbb{R}^d.\]
Therefore, since $u\in H^1_{\textrm{loc}}(\mathbb{R}^d)$ solving \eqref{lsr_41} was arbitrary, we have from \eqref{lsr_40} and the definition of the excess that
\begin{align}\label{lsr_42}\textrm{Exc}(u;\theta_0R) & =\inf_{\xi\in\mathbb{R}^d}\textrm{Exc}(u-\xi\cdot x-\phi_\xi;\theta_0R)
\\ \nonumber & \leq \theta_0^{2\alpha}\inf_{\xi\in\mathbb{R}^d}\left(\fint_{B_R}\abs{\nabla u-\xi-\nabla\phi_\xi}^2\right)=\theta_0^{2\a}\textrm{Exc}(u;R).\end{align}
We will now use the exact inequality \eqref{lsr_42} to conclude. For $R_1\leq R_2\in(0,\infty)$ suppose that \eqref{lsr_103} and \eqref{lsr_104} are satisfied for every $r\in[R_1,R_2]$. We will prove that there exists $c\in(0,\infty)$ depending on $\a$ but independent of $R_1,R_2\in(0,\infty)$ such that
\[\textrm{Exc}(u;R_1)\leq c\left(\nicefrac{R_1}{R_2}\right)^{2\a}\textrm{Exc}(u;R_2).\]
For $\theta_0\in(0,\nicefrac{1}{4})$ defined in \eqref{lsr_38}, if $\nicefrac{R_1}{R_2}\geq\theta_0$ then by definition of the excess
\[\textrm{Exc}(u;R_1)\leq \left(\nicefrac{R_2}{R_1}\right)^d\textrm{Exc}(u;R_2)\leq\theta_0^{-d}\textrm{Exc}(u;R_2)\leq \theta_0^{-(d+2\a)}\left(\nicefrac{R_1}{R_2}\right)^{2\a}\textrm{Exc}(u;R_2).\]
If $\nicefrac{R_1}{R_2}<\theta_0$ let $n\in\mathbb{N}$ be the unique positive integer satisfying $\theta_0^{n}\leq\nicefrac{R_1}{R_2}<\theta_0^{n-1}$ and observe by induction, the previous step, and \eqref{lsr_42} that
\begin{align*}
\textrm{Exc}(u;R_1)& \leq \theta_0^{-(d+2\a)}\textrm{Exc}(u;\theta_0^{n-1}R_2) \leq \theta_0^{-(d+2\a)}\theta_0^{2\a}\textrm{Exc}(u;\theta_0^{n-2}R_2)
\\ & \leq \theta_0^{-(d+2\a)}(\theta_0^{n-1})^{2\a}\textrm{Exc}(u;R_2)\leq \theta_0^{-(d+4\a)}(\nicefrac{R_1}{R_2})^{2\a}\textrm{Exc}(u;R_2),
\end{align*}
which completes the proof.
\end{proof}
\begin{thm}\label{thm_excess} Assume~\eqref{steady}. On a subset of full probability, for every $\a\in(0,1)$ there exists a deterministic $c\in(0,\infty)$ and a random radius $R_0\in(0,\infty)$ such that, for every weak solution $u\in H^1_{\textrm{loc}}(\mathbb{R}^d)$ of the equation
\[-\nabla\cdot (a+s)\nabla u = 0 \;\;\textrm{in}\;\;\mathbb{R}^d,\]
for every $R_1<R_2\in(R_0,\infty)$,
\[R_1^{-2\a}\textrm{Exc}(u;R_1)\leq cR_2^{-2\a}\textrm{Exc}(u;R_2).\]
\end{thm}
\begin{proof} The proof is an immediate consequence of the ergodic theorem, Proposition~\ref{prop_sublinear}, and Proposition~\ref{prop_excess}.\end{proof}
\section{The Liouville theorem}\label{sec_lvl} In this section, we will prove the first-order Liouville theorem for subquadratic solutions $u\in H^1_{\textrm{loc}}(\mathbb{R}^d)$ of the equation
\[-\nabla\cdot(a+s)\nabla u = 0\;\;\textrm{in}\;\;\mathbb{R}^d.\]
That is, in analogy with the Liouville theorem for harmonic functions on Euclidean space, the space of subquadratic $(a+s)$-harmonic functions is spanned by the $(a+s)$-harmonic coordinates. The section is organized as follows. We prove in Proposition~\ref{prop_lvl_est} below a version of the Caccioppoli inequality adapted to the divergence-free setting. We then prove the Liouville theorem in Theorem~\ref{thm_lvl} below, which is a consequence of the large-scale regularity estimate of Theorem~\ref{thm_excess} and the Caccioppoli inequality. These methods are motivated by the analogous results in \cite{BelFehOtt2018,GloNeuOtt2020} from the elliptic setting.
\begin{prop}\label{prop_lvl_est} Assume \eqref{steady}. Let $q_d=d\vee(2+\d)$, let $\nicefrac{1}{2_*}=\nicefrac{1}{2}-\nicefrac{1}{q_d}$, and let $u\in H^1_{\textrm{loc}}(\mathbb{R}^d)$ be a weak solution of
\begin{equation}\label{lvl_1}-\nabla\cdot(a+s)\nabla u=0\;\;\textrm{in}\;\;\mathbb{R}^d.\end{equation}
Then, for every $R\in(0,\infty)$, for some $c\in(0,\infty)$ independent of $R$,
\[\fint_{B_R}\abs{\nabla u}^2\leq \frac{c}{R^2}\left(\fint_{B_{2R}}\abs{u}^2+\left(\fint_{B_{2R}}\abs{s}^{q_d}\right)^{\frac{2}{q_d}}\left(\fint_{B_{2R}}\abs{u}^{2_*}\right)^\frac{2}{2_*}\right).\]
\end{prop}
\begin{proof} Let $\eta\colon\mathbb{R}^d\rightarrow\mathbb{R}$ be a smooth cutoff function satisfying $\eta=1$ on $\overline{B}_1$, satisfying $\eta=0$ on $\mathbb{R}^d\setminus B_2$, and for each $R\in(0,\infty)$ define $\eta_R(x)=\eta(\nicefrac{x}{R})$. A repetition of the argument leading to \eqref{hom_9} proves that, after testing \eqref{lvl_1} with $\eta_R^2 u_n$ for $u_n=(u\wedge n)\vee(-n)$ and passing to the limit $n\rightarrow\infty$,
\[\int_{B_R}a\nabla u\cdot\nabla u \eta^2_R = -2\int_{B_R}a\nabla u\cdot\nabla\eta_R u\eta_R -2\int_{B_R}s\nabla u\cdot \nabla\eta_R u\eta_R.\]
It then follows by definition of $\eta_R$, the uniform ellipticity, H\"older's inequality, and Young's inequality that, for some $c\in(0,\infty)$ independent of $R$,
\[\fint_{B_R}\abs{\nabla u}^2\leq \frac{c}{R^2}\left(\fint_{B_{2R}}\abs{u}^2+\left(\fint_{B_{2R}}\abs{s}^{q_d}\right)^{\frac{2}{q_d}}\left(\fint_{B_{2R}}\abs{u}^{2_*}\right)^\frac{2}{2_*}\right).\qedhere\]
\end{proof}
\begin{thm}\label{thm_lvl} Assume \eqref{steady}. Let $q_d=d\vee(2+\d)$ and $\nicefrac{1}{2_*}=\nicefrac{1}{2}-\nicefrac{1}{q_d}$. Then almost surely every weak solution $u\in H^1_{\textrm{loc}}(\mathbb{R}^d)$ of the equation
\begin{equation}\label{lvl_20}-\nabla\cdot(a+s)\nabla u=0\;\;\textrm{in}\;\;\mathbb{R}^d,\end{equation}
that is strictly subquadratic in the sense that, for some $\a\in(0,1)$,
\begin{equation}\label{lvl_021}\lim_{R\rightarrow\infty}\frac{1}{R^{1+\a}}\left(\fint_{B_R}\abs{u}^{2_*}\right)^\frac{1}{2_*}=0,\end{equation}
satisfies $u=c+\xi\cdot x+\phi_\xi$ on $\mathbb{R}^d$ for some $c\in(0,\infty)$ and $\xi\in\mathbb{R}^d$.
\end{thm}
\begin{proof} By the ergodic theorem and $\mathbb{E}[\Phi_i]=0$ for each $i\in\{1,\ldots,d\}$ let $\O_1\subseteq\O$ be the subset of full probability satisfying, for every $\xi\in\mathbb{R}^d$,
\begin{equation}\label{lvl_19}\lim_{R\rightarrow\infty}\fint_{B_R}\abs{\nabla\phi_\xi+\xi}^2=\mathbb{E}\left[\abs{\Phi_\xi}^2\right]+\abs{\xi}^2\geq \abs{\xi}^2\;\;\textrm{and}\;\;\lim_{R\rightarrow\infty}\fint_{B_R}\abs{s}^{q_d}=\mathbb{E}\left[\abs{S}^{q_d}\right].\end{equation}
Let $\O_2\subseteq\O$ be the subset of full probability satisfying the conclusion Theorem~\ref{thm_excess}, and let $\O_3=\O_1\cap\O_2$. For $\o\in\O_3$ let $R_0\in(0,\infty)$ be such that every weak solution $u\in H^1_{\textrm{loc}}(\mathbb{R}^d)$ of \eqref{lvl_20} satisfies, for every $R_1<R_2\in(R_0,\infty)$, for a deterministic $c\in(0,\infty)$ depending on $\a$,
\begin{equation}\label{lvl_21}R_1^{-2\a}\textrm{Exc}(u;R_1)\leq c R_2^{-2\a}\textrm{Exc}(u;R_2),\end{equation}
and such that, for every $R\in(R_0,\infty)$ and $\xi\in\mathbb{R}^d$,
\begin{equation}\label{level_22}\fint_{B_R}\abs{\nabla\phi_\xi+\xi}^2\geq\nicefrac{\abs{\xi}^2}{2}.\end{equation}
The definition of the excess, $u\in H^1_{\textrm{loc}}(\mathbb{R}^d)$, and \eqref{level_22} prove that, for every $R\in(R_0,\infty)$,
\begin{equation}\label{lvl_23}\textrm{Exc}(u;R)=\inf_{\xi\in\mathbb{R}^d}\left(R^{-2\a}\fint_{B_R}{\nabla u-\xi-\nabla\phi_\xi}^2\right)=\min_{\xi\in\mathbb{R}^d}\left(R^{-2\a}\fint_{B_R}{\nabla u-\xi-\nabla\phi_\xi}^2\right).\end{equation}
Fix $R_1\in(R_0,\infty)$. We have by definition of the excess, Proposition~\ref{prop_lvl_est}, and \eqref{lvl_21} that, for every $R\in(R_1,\infty)$,
\begin{align}
R_1^{-2\a}\textrm{Exc}(u;R_1) & \leq c R^{-2\a}\fint_{B_R}\abs{\nabla u}^2
\\ \nonumber & \leq c R^{-2(1+\a)}\left(\fint_{B_{2R}}\abs{u}^2+\left(\fint_{B_{2R}}\abs{s}^{q_d}\right)^{\frac{2}{q_d}}\left(\fint_{B_{2R}}\abs{u}^{2_*}\right)^\frac{2}{2_*}\right).
\end{align}
H\"older's inequality, the ergodic theorem, $2_*\in(2,\infty)$, \eqref{lvl_021}, and \eqref{lvl_19} prove almost surely that
\begin{align*}
& R_1^{-2\a}\textrm{Exc}(u;R_1)
\\ & \leq c \limsup_{R\rightarrow\infty}R^{-2(1+\a)}\left(\fint_{B_{2R}}\abs{u}^2+\left(\fint_{B_{2R}}\abs{s}^{q_d}\right)^{\frac{2}{q_d}}\left(\fint_{B_{2R}}\abs{u}^{2_*}\right)^\frac{2}{2_*}\right)=0.
\end{align*}
It then follows from \eqref{lvl_23} that there exists $\xi_{R_1}\in\mathbb{R}^d$ such that
\[\nabla u-\xi_{R_1}-\nabla\phi_{\xi_{R_1}}=0\;\;\textrm{in}\;\;L^2(B_{R_1};\mathbb{R}^d),\]
and therefore there exists $c_{R_1}\in\mathbb{R}$ such that
\[u=c_{R_1}+\xi_{R_1}\cdot x+\phi_{\xi_{R_1}}\;\;\textrm{in}\;\;H^1(B_{R_1}).\]
Since the linearity and \eqref{level_22} prove that $c_{R_1}=c_{R_2}$ and $\xi_{R_1}=\xi_{R_2}$ whenever $R_1\leq R_2\in(R_0,\infty)$, there exists $\xi\in\mathbb{R}^d$ and $c\in\mathbb{R}$ such that
\[u=c+\xi\cdot x+\phi_\xi\;\;\textrm{in}\;\;H^1_{\textrm{loc}}(\mathbb{R}^d).\qedhere\]
\end{proof}
\bibliographystyle{plain}
|
\section{Details of the calculations for Si}
First, we compute the self-consistent potential and Kohn-Sham states on a
$12\times12\times12$ Monkhorst-Pack {\bf{k}}-point grid using DFT and lattice-dynamical properties with
DFPT\cite{baroni2001phonons} on a $3\times3\times3$ {\bf{q}}-point grid,
as implemented in the Quantum Espresso distribution\cite{giannozzi2009quantum}
using the Perdew-Burke-Ernzerhoff exchange-correlation functional\cite{perdew1996generalized}.
We used a full-relativistic norm-conserving
optimized Vanderbilt pseudopotential\cite{scherpelz2016implementation}.
The unit cell consists of Si in the diamond structure with
an experimental lattice parameter of $5.43~\si{\angstrom}$.
The {\it{e-ph}} matrix elements are first computed
on coarse grids, then they are determined in the significantly finer grids using
both Wannier-Fourier (W-F) interpolation only, through EPW code
and our dual interpolation method, Turbo-EPW (T-EPW).
Maximally localized Wannier functions\cite{marzari1997maximally} for the
wannierization procedure are obtained
from Wannier90.\cite{mostofi2008wannier90} Thus, Bloch-to-Wannier
rotation matrices and then Wannier-to-Bloch diagonalizer matrices are used
to interpolate {\it{el-ph}} matrix elements.
\section{Convergence analysis}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figS1.png}
\vspace*{6mm}\caption{\label{figureS1}
Convergence analysis of the scattering rate of Si at 300K
due to el-ph coupling in the Fan-Migdal approximation as a
function of the electron energy. The calculation has been
performed on different {\bf{q}} meshes, namely, $(20)^3$ (green crosses),
$(60)^3$ (blue squares), $(100)^3$ (red dots) {\bf{q}}-meshes in the
first W-F interpolation, while keeping fixed the number of irreducible
{\bf{k}} points in the second plane-waves interpolation,
leading to $(100)^3$ {\bf{k}}-mesh.}
\end{figure}
Fig.~\ref{figureS1} shows how the scattering rate aproaches
convergence by increasing the mesh size of W-F interpolation,
from $(20)^3$ to $(100)^{3}$ ${\bf{q}}$ points,
while keeping the number of irreducible points
fixed at $5216$ ${\bf{k}}$ points.
The second interpolation by star functions leads
into a converged grid with $(100)^3$ ${\bf{k}}$ points.
Fig.~\ref{figureS2} shows the difference in
scattering rates calculated by different approaches, namely,
different parameters in the second plane-waves interpolation (PWI),
over equivalent meshes.
The analysis shows that the accompanying error due to the increase in $M$
decreases by enlarging the number of irreducible points, $n_{{\bf{\bar{k}}}}$ (see main text).
Moreover, the difference between scattering rates calculated over $(100)^3$ ${\bf{q}}$/ $1661$ ${\bf{k}}$
points expanded by using $M = 5$ to reach $(100)^3$ ${\bf{k}}$-mesh and
data from $(100)^3$ ${\bf{q}}$/ $256$ ${\bf{k}}$ points expanded by using $M = 30$
to reach $(100)^3$ ${\bf{k}}$-mesh,
is within $\approx \pm 6\%$.
For the remaining, the difference between the data points is about
the same, within $\approx \pm 3\%$.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figS2.png}
\vspace*{6mm}\caption{\label{figureS2}
Difference between scattering rates computed by using
different parameters in the second plane-waves interpolation,
over equivalent meshes. Green crosses correspond to the difference
between scattering rates calculated over $(100)^3$ {\bf{q}}/ $1661$ {\bf{k}}
points expanded by using $M = 5$ to reach $(100)^3$ {\bf{k}}-mesh and
data from $(100)^3$ {\bf{q}}/ $256$ {\bf{k}} points expanded by using $M = 30$
to reach $(100)^3$ {\bf{k}}-mesh. Blue squares correspond to the
difference between scattering rates calculated over $(100)^3$ {\bf{q}}/
$5216$ {\bf{k}} points expanded by using $M = 10$ to reach $(180)^3$ {\bf{k}}-mesh
and data from $(100)^3$ {\bf{q}}/ $1661$ {\bf{k}} points expanded by
using $M = 30$ to reach $(180)^3$ {\bf{k}}-mesh. Red dots
correspond to the difference between scattering rates calculated over
$(100)^3$ {\bf{q}}/ $5216$ {\bf{k}} points expanded by using $M = 8$
to reach $(160)^3$ {\bf{k}}-mesh and data from $(100)^3$
{\bf{q}}/ $1661$ {\bf{k}} points expanded by using $M = 25$ to reach
$(160)^3$ {\bf{k}}-mesh.}
\end{figure}
\section{Calculation of electron self-energy and thermoelectric properties}
The expression for the imaginary part of electronic self-energy due to
{\it{el-ph}} coupling in the
Fan-Migdal approximation can be derived from quantum field theory\cite{giustino2017electron}
and it is expressed as
\begin{equation}
\label{imag}
\begin{split}
\Sigma^{\prime\prime}_{n,{\bf{k}}}(\omega,T)=\pi \sum_{m,\beta} \int_{BZ} \frac{d{\bf{q}}}{\Omega_{BZ}} |g_{mn,\beta}({\bf{k},{\bf{q}}})|^2 \\
\times \Bigg[\left[n_{{\bf{q}}\beta}(T) + f_{m{\bf{k}}+{\bf{q}}}\right]\delta(\omega - (\epsilon_{m{\bf{k}}+{\bf{q}}} - \epsilon_F) + \omega_{{\bf{q}}\beta}) \\
+ [n_{{\bf{q}}\beta}(T) + 1 - f_{m{\bf{k}}+{\bf{q}}}]\delta(\omega - (\epsilon_{m{\bf{k}}+{\bf{q}}} - \epsilon_F) - \omega_{{\bf{q}}\beta})\Bigg]~,
\end{split}
\end{equation}
where $n_{{\bf{q}}\beta}(T)$ and $f_{m{\bf{k}}+{\bf{q}}}$ are the Bose-Einstein and the Fermi-Dirac distributions,
$\Omega_{BZ}$ is the BZ volume, $m$ and $n$ are the corresponding electronic states, while $\beta$ represents the
phonon branch, $\epsilon_{m{\bf{k}}+{\bf{q}}}$ are the electronic eigenenergies of the state $m{\bf{k}}+{\bf{q}}$ and
$\omega_{{\bf{q}}\beta}$ are the corresponding eigenfrequencies with wavevector ${\bf{q}}$ and phonon branch $\beta$.
Basically, from the first W-F interpolation we can get
$f({\bf{\bar{k}}}) = \Sigma^{\prime\prime}_{n,{\bf{\bar{k}}}}(\omega,T)$,
over the irreducible points,
which will be interpolated throughout the whole BZ by star functions,
resulting in $\Sigma^{\prime\prime}_{n,{\bf{k^{\prime}}}}(\omega,T)$ over denser
${\bf{k^{\prime}}}$ grids.
$\Sigma^{\prime\prime}$ is directly related to the scattering rate, that is, inversely
proportional to the relaxation time
\begin{equation}
\label{tau}
\frac{1}{\tau_{n,{\bf{k}}}} = 2\Sigma^{\prime\prime}_{n,{\bf{k}}}(\omega = 0,T)~,
\end{equation}
which enters in kinetic transport equations.
Indeed, the kinetic coefficient tensors can be expressed through
\begin{equation}
\label{Lambda}
\Lambda^{(\alpha)}(\mu;T)= e{^2}\int\Xi(\epsilon,\mu,T)(\epsilon - \mu)^{\alpha}\left(-\frac{\partial f^{(0)}(\mu;\epsilon,T)}{\partial \epsilon}\right)d\epsilon~,
\end{equation}
where $\mu$ is the chemical potential and $\Xi(\epsilon,\mu,T)$
is the transport distribution kernel given by
\begin{equation}
\label{Xi}
\Xi(\epsilon,\mu,T) = \int \sum_n{{\bf{v}}_{n,{\bf{k}}}\otimes{\bf{v}}_{n,{\bf{k}}}{\tau}_{n,{\bf{k}}}}(\mu,T)\delta(\epsilon - \epsilon_{n,{\bf{k}}})\frac{d{\bf{k}}}{8\pi^3}~,
\end{equation}
with ${\bf{v}}_{n,{\bf{k}}}$ being the electron velocity.
From both experimental conditions of zero temperature gradient ($\nabla T = 0$)
and zero electric current, the kinetic coefficient tensors
can be identified with the electrical conductivity tensor, $\sigma = \Lambda^{(0)}$,
the Seebeck coefficient tensor, $S = (eT)^{-1}\Lambda^{(1)}/\Lambda^{(0)}$,
and the charge carrier contribution to thermal conductivity tensor,
$\kappa_e = (e^2T)^{-1} \left({\Lambda^{(1)}\cdot{\Lambda^{(0)}}^{-1}}\cdot{\Lambda^{(1)}} - \Lambda^{(2)}\right)$.
Consequently, once we have $\tau_{n,{\bf{k}}}$ from Eq.~(\ref{tau}),
we can compute the electrical conductivity, the Seebeck coefficient, Lorenz function and
the charge carrier contribution to thermal conductivity from the solution of Eq.~(\ref{Xi}) and~Eq.~(\ref{Lambda}).
Note that the both bandstructure and phonon dispersion have also been interpolated by the method
presented in the main text over the same grid as electron self-energy.
Indeed, we have implemented these equations on top of the BoltzTraP code,\cite{madsen2006boltztrap}
from which we can obtain these transport properties
directly from first principles.
\section{Scattering by ionized impurities}
For the calculation of thermoelectric transport properties of
$n$- and $p$-type Si polycrystals, we have also considered the scattering
by ionized impurities.
Such scattering has been treated theoretically by
Brooks and Herring (B-H)\cite{brooks1955theory,chattopadhyay1981electron}
by considering a screened Coulomb potential, the
Born approximation for the evaluation of transition probabilities
and neglecting perturbation effects of the
impurities on the electron energy levels and wave functions.
In the B-H theory the electron is scattered independently
by dilute concentrations of ionized centers randomly distributed in the semiconductor.
The per-unit-time transition probability for the scattering of charge carriers
by ionized impurities can be given in the plane-wave approximation as
\begin{multline}
\label{U}
W({\bf{k}}|{\bf{{k}^{\prime}}}) = \frac{2\pi}{\hbar}\frac{N_i}{V} \\
\left\vert{\int U({\bf{r}}) \exp{\left[i({\bf{k}}-{\bf{{k}^{\prime}}};{\bf{r}})\right]}d{\bf{r}}}\right\vert^2 \delta(\epsilon({\bf{{k}^{\prime}}})-\epsilon({\bf{k}}))~,
\end{multline}
where $U({\bf{r}})$ is the scattering potential and $N_i$ is the ionized impurity concentration.
A long-range Coulomb field,
$U(\mat{r}) = e\phi(r) = \pm e^2/\zeta r$, with potential $\phi$ at a point $r$ of the crystal
is created by the presence of positive (donor) or negative (acceptor)
impurity ions, within a medium with dieletric constant $\zeta$.
The straightforward application of this field in Eq.~(\ref{U}) leads
into a logarithmic divergence, and hence, a screened Coulomb potential has to be considered.
From the B-H theory the potential can be expressed in a more rigorous form as
$\phi(r) = \pm e/{\zeta r} \left(\exp{\left(-{r}/{r_0}\right)}\right)$,
where $r_0$ is the radius of ion field screening defined by
\begin{equation}
\label{r0}
r{_0}^{-2}({\bf{k}}) = \frac{4\pi{e^2}}{\zeta{_0}}\int{-\frac{\partial f_0}{\partial \epsilon({\bf{k}})}}g(\epsilon)d\epsilon~,
\end{equation}
where $f^{(0)}(\epsilon)$ is the equilibrium electron
distribution function, $\zeta_0$ is the static dielectric constant,
and $g(\epsilon)$ is the density of states,
calculated numerically on an energy grid
with spacing $d\epsilon$ sampled over $N_k$ ${\bf{k}}$-points
\begin{equation}
\label{dos2}
g(\epsilon) = \int \sum_n \delta(\epsilon - \epsilon_{n,{\bf{k}}}) \frac{d{\bf{k}}}{8\pi^3} = \frac{1}{\Omega N_k}\sum_{n,{\bf{k}}} \frac{\delta(\epsilon - \epsilon_{n,{\bf{k}}})}{d\epsilon}~,
\end{equation}
where $\Omega$ is the volume of the unit cell.
Within the relaxation time approximation for the
Boltzmann transport equations, the relaxation time for the scattering of the
charge carriers by ionized impurities can be expressed as
\begin{equation}
\label{tau22}
\tau_{imp}({\bf{k}}) = \frac{\hbar\zeta{_0}{^2}}{2{\pi}{e^4}{N_i}F_{imp}({\bf{k}})}{\bf{k}}^2 \left\vert\frac{\partial \epsilon({\bf{k}})}{\partial {\bf{k}}}\right\vert
\end{equation}
where
\begin{equation}
F_{imp}({\bf{k}}) = ln(1+\eta) - \frac{\eta}{1+\eta}~,
\end{equation}
is the screening function with $\eta = (2{\bf{k}}r_0)^2$.
Here, we interpolated $\left\vert\frac{\partial \epsilon({\bf{k}})}{\partial {\bf{k}}}\right\vert$
within Eq.~(\ref{tau22}) by using the plane-waves interpolation (see the main text) and
used Mathiessen's rule to consider both the scattering by ionized impurities and phonons.
\section{Calculation of phonon-assisted optical absorption}
To calculate the phonon-assisted absorption coefficient,
we use the Fermi's golden rule expression\cite{parravicini1975g,noffsinger2012}:
\begin{eqnarray}
\label{imag}
\alpha(\omega) = 2 \frac{4\pi^2 e^2}{\omega c n_r(\omega)}\frac{1}{\Omega} \frac{1}{N_{{\bf{k}}} N_{{\bf{q}}}} \sum_{\beta i j {\bf{k}} {\bf{q}}}|{\mat{\lambda}} \cdot ({\bf{S_1}} + {\bf{S_2}})|^2 \\
\times P \delta(\epsilon_{j,{\bf{k}}+{\bf{q}}} - \epsilon_{i,{\bf{k}}} - \hbar\omega - \pm \hbar\omega_{\beta {\bf{q}}}) \nonumber ~,
\end{eqnarray}
with
\begin{equation}
{\bf{S_1}}({\bf{k}} {\bf{q}}) = \sum_m \frac{{\bf{v}}_{im}({\bf{k}}) g_{m j,\beta}({\bf{k}},{\bf{q}})}{\epsilon_{m,{\bf{k}}} - \epsilon_{i {\bf{k}}} - \hbar\omega + i\Gamma_{m,{\bf{k}}}}~,
\end{equation}
\begin{equation}
{\bf{S_2}}({\bf{k}} {\bf{q}}) = \sum_m \frac{g_{i m,\beta}({\bf{k}},{\bf{q}}) {\bf{v}}_{mj}({\bf{k}} + {\bf{q}})}{\epsilon_{m,{\bf{k}} + {\bf{q}}} - \epsilon_{i {\bf{k}}} \pm \hbar\omega_{\beta {\bf{q}}} + i\Gamma_{m,{\bf{k}}+{\bf{q}}}} ~,
\end{equation}
and
\begin{equation}
P = (n_{\beta {\bf{q}}} + \frac{1}{2} \pm \frac{1}{2})(f_{i {\bf{k}}} - f_{j, {\bf{k}}+{\bf{q}}})~.
\end{equation}
In these equations, $\omega$ is the photon frequency,
$c$ is the speed of light, $n_r$ is the refractive index
(for silicon we used $n_r = 3.4$), and $\mat{\lambda}$ is
the photon polarization. ${\bf{S_1}}$ and ${\bf{S_2}}$ are
the two possible ways for the indirect absorption process, while
$P$ is related to the carrier and phonon statistics.
For calculations of Si, the DFT band gap and all
conduction bands have been shifted up by $0.7$\si{\eV}
to simulate experimental gap.
Our calculations have been carried out over $(60)^3${\bf{q}}/$(40)^3${\bf{k}}
meshes. It took $\approx 36$ minutes to perform the calculation by using
our dual interpolation method on $8$
CPU cores at the Odyssey cluster (Harvard University).
\end{appendices}
\setcounter{secnumdepth}{0}
\section{Acknowledgments}
The authors thank the Harvard FAS Research Computing facility and the Brazilian
CCJDR-IFGW-UNICAMP for computational resources.
A.S.C. and A.A. gratefully acknowledge financial support from the
Brazilian agency FAPESP under Grants No.2015/26434-2,
No.2016/23891-6, No.2017/26105-4, No.2018/01274-0 and No.2019/26088-8.
A.S.C. also acknowledges the kind hospitality of SEAS-Harvard University.
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
In this paper, all groups are finite. Given a group $G$, the commuting graph of $G$, denoted by $\Gamma(G)$, is defined as the graph whose vertex set is the noncentral elements of $G$ and two vertices are adjacent if and only if they commute in $G$. Giudici and Pope in \cite{GiuPo} state that commuting graphs were first studied by Brauer and Fowler in \cite{BrFo} in relation to the classification of simple groups. They go on to say that commuting graphs were first studied in their own right by Segev and Seitz in \cite{SeSe} in terms of the classical simple groups. In fact, much of the research regarding the commuting graph is related to simple groups. We are not going to try describe all of this research, but note that it culminates in the work of Solomon and Woldar in \cite{SoWo} were they complete the final step in proving that if $S$ is a simple group and $X$ is any group where $\Gamma (S) \cong \Gamma (X)$, then $S \cong X$.
In the course of this research, Iranmanesh and Jafarzedah conjecture in \cite{IraJaf} that there is a universal bound on the diameter of commuting graphs. Giudici and Pope in \cite{GiuPo} study the diameters of commuting graphs for a number of families of group. Surprisingly, Giudici and Parker produced in \cite{GiuPark} a family of $2$-groups of nilpotence class $2$ where there is no bound on the diameter of the commuting graphs. On the other hand, in the seminal 2013 paper \cite{parkerSoluble}, Parker proves that the commuting graph of a solvable group $G$ with a trivial center is disconnected if and only if $G$ is a Frobenius or $2$-Frobenius group. We will state the definition of $2$-Frobenius groups in Section 2. In addition, when the commuting graph of $G$ is connected, he proves that it has diameter at most $8$. He also provides an example of a solvable group where this bound is met. Parker and Morgan remove the solvability hypothesis on $G$ in \cite{parkerTrivialCenter} by proving for the commuting graph of any group with a trivial center that all the connected components have diameter at most $10$.
In this paper, we look to extend Parker's and Parker and Morgan's results. In particular, we would like to extend the class of groups where we can determine which commuting graphs are disconnected and determine further classes of groups where we can show that diameters of the commuting graph are bounded by the constants $8$ and $10$. In particular, we show that we can replace the hypothesis that $\centerof{G} = 1$ with the hypothesis that $G' \cap \centerof{G} = 1$.
\begin{mainthm}
Let $G$ be a group and suppose that $G' \cap \centerof{G} = 1$.
\begin{enumerate}
\item $\Gamma (G)$ is connected if and only if $\Gamma (G/\centerof{G})$ is connected.
\item Every connected component of $\Gamma (G)$ has diameter at most $10$.
\item If $G$ is solvable and $\Gamma (G)$ is connected, then $\Gamma (G)$ has diameter at most $8$.
\item If $G$ is solvable, then $\Gamma (G)$ is disconnected if and only if $G/Z$ is either a Frobenius group or a $2$-Frobenius group.
\end{enumerate}
\end{mainthm}
In fact, we will see that we can relax the hypothesis that $G' \cap \centerof{G} = 1$ even further. In particular, we show that it suffices to assume that $C (G) \cap \centerof{G} = \{ 1 \}$ where $C (G) = \{ [x,y] \mid x,y \in G \}$. I.e., we only need the set of commutators, not the whole commutator subgroup. (We note that we use $1$ to denote the trivial subgroup. Since $C(G)$ is not a subgroup, it is not appropriate to view $C(G) \cap \centerof{G}$ as a subgroup, so we use $\{ 1 \}$ to denote the set consisting only of the identity.) We will present examples of groups $G$ where $C (G) \cap \centerof{G} = \{ 1 \}$, but $G' \cap Z (G) > 1$, so this replacement does actually improve the result.
Following the literature, we say that a group $G$ is an {\it $A$-group} if every Sylow subgroup of $G$ is abelian. We show that if $G$ is an $A$-group then $G' \cap \centerof{G} = 1$. In particular, this shows that the results of the Main Theorem apply to $A$-groups; however, we would not be surprised if one could show that the diameter bounds can be lowered for $A$-groups, especially solvable $A$-groups, but we have not investigated this at this time.
One other consequence of $G' \cap \centerof{G} = 1$ is that $\centerof{G/\centerof{G}} = 1$, and so it makes sense to ask whether we can determine which groups satisfying $\centerof{G/\centerof{G}} = 1$ have a disconnected commuting graph and if we can bound the diameter of the commuting graphs of groups where the commuting graph is connected. At this point, we can find one class of groups where the commuting graph is disconnected as seen in the following theorem, but we will present examples of other groups satisfying this condition where the commuting graph is disconnected.
\begin{theorem}
If $G$ is a group where $G/\centerof{G}$ is either a Frobenius or a $2$-Frobenius group, then $\Gamma (G)$ is disconnected.
\end{theorem}
This research was conducted during a summer REU in 2020 at Kent State University with the funding of NSF Grant DMS-1653002.
We thank the NSF and Professor Soprunova for their support.
\section{Results}
Let $G$ be a finite group and $\Gamma(G)$ the commuting graph of $G$. When $x = y$ or $x$ and $y$ are adjacent in $\Gamma(G)$, we write $x \sim y$. In particular, writing $x \sim y$ emphasizes that $x,y \in G \setminus \centerof{G}$. Viewing $x,y$ as vertices of $\Gamma (G)$, we use $d(x,y)$ to denote the distance between $x$ and $y$. (I.e., the number of edges in the shortest path between $x$ and $y$.) For the remainder of the paper, we set $Z = \centerof{G}$, the center of the group, and let $Z_2 = \ithcenter{2}{G}$, the preimage of $\centerof{G/\centerof{G}}$. We also set $C = C(G)$. Note that $G'$ is the group generated by $C$, but $C$ itself is not always a group.
The following lemma addresses the relationship between adjacent elements in $\Gamma(G)$ and adjacent elements in $\Gamma(G/Z)$.
\begin{lemma} \label{three}
Let $G$ be a group and fix $x, y\in G \setminus Z_2$. If $x \sim y$ in $\Gamma(G)$, then $xZ \sim yZ$ in $\Gamma(G/Z)$. If $C \cap Z = \{ 1 \} $ and $xZ \sim yZ$ in $\Gamma(G/Z)$, then $x \sim y$ in $\Gamma(G)$.
\end{lemma}
\begin{proof}
Suppose that $x \sim y$ in $\Gamma(G)$. Then, we have that $xy=yx$, and hence, $xZ yZ = yZ xZ$ in $G/Z$. Thus, $xZ \sim yZ$.
Now, suppose $C \cap Z = \{ 1 \}$ and $xZ \sim yZ$. This implies that $[x,y]\in Z$. Also, $[x,y] \in C$, and since $C \cap Z = \{ 1 \}$, we obtain $[x,y] = 1$. Therefore, $xy = yx$ and so $x \sim y$.
\end{proof}
We now consider the relationship between $\Gamma (G)$ and $\Gamma (G/Z)$ when $Z_2 = Z$ and in particular, when $C \cap Z = \{ 1 \}$.
\begin{lemma} \label {four}
Let $G$ be a group and suppose that $Z_2 = Z$.
\begin{enumerate}
\item If $\Gamma (G)$ is connected, then $\Gamma (G/Z)$ is connected and the diameter of $\Gamma (G/Z)$ is less than or equal to the diameter of $\Gamma (G)$.
\item If $C \cap Z = \{ 1 \}$ and $\Gamma (G/Z)$ is connected, then $\Gamma (G)$ is connected and has the same diameter as $\Gamma (G/Z)$.
\item If $C \cap Z = \{ 1 \}$ and $\Gamma (G/Z)$ is disconnected, then there is a one-to-one correspondence between the connected components of $\Gamma (G/Z)$ and of $\Gamma (G)$ that preserves diameter except in the case that a connected component of $\Gamma (G/Z)$ consists of a single coset and $Z > 1$; in this case, the corresponding component of $\Gamma (G)$ will have diameter $1$.
\end{enumerate}
\end{lemma}
\begin{proof}
Suppose $x, y \in G \setminus Z$. Since $\Gamma (G)$ is connected, we can find $x = x_0, x_1, \dots, x_n = y \in G \setminus Z$ so that $x_i \sim x_{i+1}$ for $i = 0, \dots, n -1$. By Lemma \ref{three}, we have $x_i Z \sim x_{i+1}Z$. It follows that $xZ$ and $yZ$ are connected by a path of length of $n$. It follows that $d(xZ,yZ) \le d(x,y)$. We deduce that $\Gamma (G/Z)$ is connected and its diameter is at most the diameter of $\Gamma (G)$.
We now add the assumption that $C \cap Z = \{ 1 \}$. From Lemma \ref{three}, we see that $x \sim y$ in $\Gamma (G)$ if and only if $xZ \sim yZ$ in $\Gamma (G/Z)$. It follows that $x$ and $y$ lie in the same connected component of $\Gamma (G)$ if and only if $xZ$ and $yZ$ lie in the same connected component of $\Gamma (G/Z)$. In the case where they are in the same connected component, then $d (x,y) = d(xZ, yZ)$; except when $xZ = yZ$ and $x \ne y$, in this case $d(x,y) = 1$ and $d (xZ,yZ) = 0$. Notice that if $xZ \ne yZ$, then $x \ne y$, so a connected component of $\Gamma (G/Z)$ with more than one element will correspond to a connected component of $\Gamma (G)$ with more than one element and they will have the same diameter. Finally, if $\{ xZ \}$ is a connected component of $\Gamma (G/Z)$, then $\{ xz \mid z \in Z\}$ will be the corresponding connected component of $\Gamma (G)$, so the component in $\Gamma (G/Z)$ has diameter $0$, but the corresponding component in $\Gamma (G)$ will have diameter $1$.
\end{proof}
Following the literature, a group $G$ is a {\it Frobenius group} if it contains a proper, nontrivial subgroup $H$ so that $H \cap H^g = 1$ for all $g \in G \setminus H$. The subgroup $H$ is called a {\it Frobenius complement} for $H$. Frobenius proved that $N = (G \setminus \cup_{g \in G} H^g) \cup \{ 1 \}$ is a normal Hall subgroup of $G$ that satisfies $G = HN$ and $H \cap N = 1$; the subgroup $N$ is called the {\it Frobenius kernel} of $G$. We say $G$ is a {\it $2$-Frobenius group} if there exist normal subgroups $K \le L$ so that $L$ and $G/K$ are Frobenius groups with Frobenius kernels $K$ and $L/K$ respectively.
We now show that when $G/Z$ is either a Frobenius group or a $2$-Frobenius group, then $\Gamma (G)$ is disconnected.
\begin{corollary}
If $G$ is a group so that $G/Z$ is either a Frobenius group or a $2$-Frobenius group, then $\Gamma (G)$ is disconnected.
\end{corollary}
\begin{proof}
We know that the center of a Frobenius group or a $2$-Frobenius group is trivial, so $Z_2 (G) = Z$. Now, Parker has shown that $\Gamma (G/Z)$ is disconnected. We apply the contrapositive of Lemma \ref{four}(1) to see that $\Gamma (G)$ is disconnected.
\end{proof}
We next show that the condition that $C \cap Z = \{ 1 \}$ implies that $Z = Z_2$.
For $x \in G$, set $D_G (x) = \{ d \in G \mid [d, x] \in Z \}$. Thus, $D_G(x)/Z = C_{G/Z} (Zx)$.
\begin{lemma}\label{five}
If $G$ is a group such that $C \cap Z = \{ 1 \}$, then $C_G(x) = D_G(x)$ for each
$x \in G$. In particular, $Z = Z_2$.
\end{lemma}
\begin{proof}
Let $x \in G$. Of course, $C_G(x) \leq D_G(x)$. If $d \in D_G(x)$, then $[d,x] \in
C \cap Z = \{ 1 \}$ and so $d \in C_G(x)$. Thus, $D_G(x) \le C_G(x)$, and we conclude that
$C_G (x) = D_G (x)$. Finally, observe that $Z = \cap_{x \in G} C_G(x) = \cap_{x \in G}
D_G(x) = Z_2$, as wanted.
\end{proof}
We now obtain Parker and Morgan's and Parker's result for $G$ when $C \cap Z = \{ 1 \}$.
\begin{corollary} \label{six}
Let $G$ be a group and suppose that $C \cap Z = \{ 1 \}$.
\begin{enumerate}
\item $\Gamma (G)$ is connected if and only if $\Gamma (G/Z)$ is connected.
\item Every connected component of $\Gamma (G)$ has diameter at most $10$.
\item If $G$ is solvable and $\Gamma (G)$ is connected, then $\Gamma (G)$ has diameter at most $8$.
\item If $G$ is solvable, then $\Gamma (G)$ is disconnected if and only if $G/Z$ is either a Frobenius group or a $2$-Frobenius group.
\end{enumerate}
\end{corollary}
\begin{proof}
Conclusion (1) is an immediate consequence of Lemma \ref{four} (1) and (2) combined with Lemma \ref{five}. Since $Z(G/Z) = 1$, Parker's results apply to $G/Z$, and applying Parker's results in $G/Z$ with Lemma \ref{four} (2) and (3) yields (2), (3), and (4).
\end{proof}
The following result has appeared in the literature. See, for example, Corollary 4.5 in \cite{Bro} or Theorem 4.1 in \cite{Tau }.
\begin{lemma} \label{agroup}
Let $G$ be an $A$-group. Then $G' \cap Z = 1$.
\end{lemma}
Using Lemma \ref{agroup}, we see that if $G$ is an $A$-group, then $G$ satisfies the hypothesis of Corollary \ref{six}, and thus, $G$ satisfies the conclusions of Corollary \ref{six}.
We close by presenting some examples using the small groups library \cite{small} that can be accessed by the computer algebra systems GAP \cite{gap} or Magma \cite{magma} to illustrate various points. We first present examples of groups $G$ where $C \cap Z = \{ 1 \}$ but $G' \cap Z > 1$. Take $G$ to be one of SmallGroup(768,1083474), SmallGroup(768,1083475), or SmallGroup(768,1083476).
We next present examples of groups $G$ where $\Gamma (G/Z)$ is connected but $\Gamma (G)$ is not connected. Taking $G$ to be one of the Small Groups (72,22), (72,23), (120,11), (144,125), and (288,565) yield examples of groups $G$ with this property. We actually have many more examples in this category, but we have just pulled a few examples at random from the list. We include the graphs $\Gamma (G)$ and $\Gamma (G/Z)$ for $G = {\rm SmallGroup} (72,22)$ in Figures \ref{g7222} and \ref{g7222qc}.
\begin{figure}[h]
\centering
\begin{minipage}{0.8\textwidth}
\centering
\includegraphics[width=0.7\linewidth]{EX-72-22-Graph}
\caption{\texttt{$\Gamma (G)$ for $G = $ SmallGroup(72,22)}}
\label{g7222}
\end{minipage}
\bigskip
\begin{minipage}{0.8\textwidth}
\centering
\includegraphics[width=0.7\linewidth]{EX-72-22-Quotient}
\caption{\texttt{$\Gamma (G/Z)$ for $G = $ SmallGroup(72,22)}}
\label{g7222qc}
\end{minipage}
\end{figure}
The groups that we have considered all satisfy the condition that $\centerof{G/Z} = 1$. It makes sense to ask what can be said about groups that satisfy $\centerof{G/Z} = 1$ and do not satisfy $C \cap Z = \{ 1 \}$. Notice that if $G$ is a such group, then there will definitely exist elements $x,y \in G \setminus Z$ so that $xZ \sim yZ$ in $\Gamma (G/Z)$ but $x \not\sim y$ in $\Gamma (G)$; so the results of Lemma \ref{four} (2) and (3) are not guaranteed to hold. We will present some examples to show in fact conclusion (3) does not hold.
First, we claim that it is not hard to see that if $G$ is a Frobenius group with Frobenius kernel $N$, then $\Gamma (G)$ has $1 + |N|$ connected components and if $G$ is a $2$-Frobenius group with normal subgroups $K \le L$ as in the definition above, then $\Gamma (G)$ has $1 + |K|$ connected components. (Although neither of these computations are difficult, the second computation is done explicitly at the end of Section 3 of \cite{CoLe}.) Also, Thompson's celebrated theorem shows that a Frobenius kernel is always nilpotent and it known that a Frobenius complement always has a nontrivial center. Since the nonidentity elements of these subgroups form the connected components of $\Gamma (G)$ when $G$ is a Frobenius group, it follows that each connected component has diameter at most $2$.
When $G$ is a $2$-Frobenius group, there are $|K|$ connected components in $\Gamma (G)$ that each consist of the nonidentity elements of a Frobenius complement of $L$. Since these Frobenius complements are all cyclic, the corresponding connected components in $\Gamma (G)$ are complete graphs. The remaining nonidentity elements of $G$ form a single connected component in $\Gamma (G)$. We know that every element of prime order in $G \setminus L$ centralizes some nonidentity element in $K$. (See Lemma 3.8 of \cite{CoLe}.) It follows that every element outside of $K$ has distance at most $2$ to a nonidentity element of $K$, and since $K$ is nilpotent, we deduce that this remaining connected component of $\Gamma (G)$ has diameter at most $6$. (We would not be surprised if one could actually get a tighter bound for the diameter of this component.)
Take $P$ to be an extra special group of order $p^3$ and exponent $p$ for an odd prime $p$. Let $q$ be a prime divisor of $p - 1$. It is well-known that $P$ has an automorphism $\sigma$ of order $q$ that centralizes $\centerof{P}$. Let $C = \langle \sigma \rangle$ so that $C$ acts on $P$ via automorphisms, and let $G$ be the resulting semi-direct product. Observe that $\centerof{P} = Z$ and $G/Z$ is a Frobenius group of order $p^2q$. Also, we have $Z = C (P) \subseteq C (G)$, so we do not have $C \cap Z = \{ 1 \}$. We see that $\Gamma (G/Z)$ is disconnected and has $1 + p^2$ connected components. It follows that $\Gamma (G)$ is disconnected. It is not difficult to see that the $p^2$ connected components of $\Gamma (G/Z)$ that correspond to the Frobenius complements in $G/Z$ will correspond to $p^2$ connected components in $\Gamma (G)$.
On the other hand, since $Z (P) = Z (G)$, we see that the remaining connected component of $\Gamma (G/Z)$ consists of the nontrivial cosets in $P/Z (P)$. On the other hand, in the graph of $\Gamma (G)$, the elements in $P \setminus Z$ will have the same connected components as the graph $\Gamma (P)$. It is not difficult to see that every noncentral element of $P$ has a centralizer that is abelian of order $p^2$. This implies that $P$ is what is sometimes called a CA-group in the literature. In particular, it is not difficult to see that the elements in $P \setminus Z$ split into $p + 1$ different connected components each having $p^2 - p$ elements. In particular, $\Gamma (G)$ has $1 + p + p^2$ different connected components. This shows that the correspondence in Lemma \ref{four} (3) does not hold when we do not have $C \cap Z = \{ 1 \}$.
We next present ${\rm GL}_2 (3)$. In this case, $G/Z \cong S_4$ is a $2$-Frobenius group so that $\Gamma (G/Z)$ has $4$ connected components that are complete graphs and one connected component that has diameter $3$. On the other hand, $\Gamma (G)$ has $13$ connected components that are all complete graphs. We include these graphs as Figures \ref{gl23} and \ref{gl23qc}.
\begin{figure}[h]
\centering
\begin{minipage}{0.8\textwidth}
\centering
\includegraphics[width=0.5\linewidth]{GL23Graph}
\caption{\texttt{$\Gamma (G)$ for $G = \rm{GL}_2 (3)$ }}
\label{gl23}
\end{minipage}
\end{figure}
\bigskip
\begin{figure}
\begin{minipage}{0.8\textwidth}
\centering
\includegraphics[width=0.4\linewidth]{S4Graph}
\caption{\texttt{$\Gamma (G/Z)$ for $G = \rm{GL}_2 (3)$ }}
\label{gl23qc}
\end{minipage}
\end{figure}
\eject
The last examples we present are examples where $\Gamma (G)$ and $\Gamma (G/Z)$ are both connected and the diameters are different. The groups are SmallGroups (400,125), (400,126), and (400,127). In all three cases, we have that $\Gamma (G)$ has diameter $5$ and $\Gamma (G/Z)$ has diameter $3$.
|
\section{Introduction}
Last year on May $24$ the SpaceX launched the first set of $60$ Starlink satellites and up to now the total number is approximately $400$ aiming to reach at least $12000$ \citep{starlink} at the end of a decade program announced by Elon Musk. Such a huge number of satellites, distributed over almost the whole surface of the Earth might be considered as the first prototype of a possible megastructure around the Earth, which in principal, might be visible from the cosmos. Similarly, one may search for techno-signatures of alien civilizations.
The discovery of the Tabby's star \citep{kic846} and "Oumuamua" \citep{oum1} have provoked the revival of the search for extraterrestrial intelligence (SETI). The idea to search for techno-signatures of advanced alien societies have been proposed by \cite{dyson}. Assuming that a civilization is advanced enough to build a megastructure around a host star to consume its whole energy, Dyson has concluded that such a huge (having the length-scale of the order of one AU) spherical construction - Dyson sphere (DS) - should be visible in the infrared (IR) spectrum. Civilizations harnessing the host star's total energy belong to the Type-II societies according to the classification by \cite{kardashev}. Type-I civilization is harnessing the total energy coming from the sun to the Earth. Our society is consuming less than the mentioned energy, therefore, an index, $K = log_{10}\left(P\right)/10-0.6$, introduced by \cite{shkl} for Earthlings is $0.7$, where $P$ denotes the average harnessed power in Watts. In the framework of the same classification Type-III is the alien high tech society which is able to use the total energy of the host galaxy.
It is clear that detection of Type-II and Type-III civilizations is much easier than Type-I because of the much higher total consumed energies. Therefore, a special interest in the Dysonian SETI projects deserve Type-II,III techno-signatures and a series of papers are dedicated to identification of DS candidates \citep{timofeev,carrigan,gaia}. Dyson's original idea has been extended to hot DSs \citep{paper3,paper4} and the megastructures around pulsars \citep{paper1,paper2}.
Despite high radiation intensity of Type-II,III technologies compared to Type-I techno-signatures, the latter still can be considered seriously in the SETI context. In particular, \cite{warming} studied effects of global warming as detectable biomarkers in Earth-like societies.
Our civilization consumes approximately $1.5\times 10^{20}$ ergs s$^{-1}$ \citep{warming}, which is less than for Type-I society, $1.7\times 10^{24}$ ergs s$^{-1}$. If one assumes $1\%$ of an average growth rate of industry and the subsequent energy consumption, one can straightforwardly show that our civilization might reach Type-I in $\sim1000$yrs. It is quite probable that in $1000$yrs the level of technology will differ from ours, likewise ours is different from the one of middle ages. Therefore, one can assume that Type-I alien society is able to cloak their planet by a sphere-like (or ring-like) structure to harness the total energy emitted from their host star toward the planet.
In this paper we consider the possible observational characteristics of a planetary megastructure partially or completely covering an Earth-like planet located in the habitable zone.
The paper is organized in the following way: in Sec. 2, we introduce main theory and study the techno-signatures of Type-I megaconstructions, obtaining major results and in Sec. 3 we outline them.
\section[]{Discussion and results}
At first we consider the question: energetically how feasible is launching material to cover an area comparable to the Earth's surface. As an example we examine the construction designed by means of Graphene as up to now the strongest material, our civilization is able to produce. Then, for the total mass of the megastructure we obtain
$$M\simeq 4\kappa\pi R_p^2\rho\Delta r\simeq$$
\begin{equation}
\label{mass}
\simeq 2\kappa\times 10^{17}\times\frac{\rho}{0.4g/cm^3}\times\frac{\Delta r}{1mm} \; g,
\end{equation}
where $\kappa\leq 1$ is a dimensionless geometric factor, $R_p$ is the Earth-like planet's radius ($R_{\oplus}\simeq 6400\;km$) density, $\rho$, is normalised by the Graphene density and it is assumed that since the Graphene is $10$ times stronger than steel the effective thickness of the construction, $\Delta r$, is normalised by $1$mm value. As it clear from here, the total mass is less than the total storage of Earth's Carbon, $1.85\times 10^{24}$ by seven orders of magnitude and thus one can find enough material to build a construction. Here we assume that the altitude is much less than $R_{\oplus}$. One can straightforwardly check that at geostationary orbit ($R\simeq 36000$km) the required mass becomes enormous compared to the above considered.
Another issue one should address, is to understand time scale required for launching such a huge mass of material. Roughly speaking, the civilization can start launching material from the times when its index was equal to ours, $0.7$ (See the introduction). Then, as we have already mentioned, to reach the Type-I level, it needs $1000$ yrs. The process will be feasible if the annual rate of growth of mass launching, $\mu$, is a small parameter. If one assumes that the first year's launched mass equals $M_0$, then in $1000$ yrs the total mass will be given by
\begin{equation}
\label{launch}
M=M_0\left(1+\mu\right)^{1000}.
\end{equation}
After taking into account Eq. (\ref{mass}) and the fact that up to date there are $895$ satellites each with the mass $260$ kg \citep{starlink}, one can straightforwardly show that $\mu\simeq 0.021$.
Generally speaking, the megastructure might be constructed by advanced Type-I civilization from extraterrestrial resources without launch costs \citep{haliki}, but let us consider the worst case: the Solar energy is stored and then utilized to launch material on the orbit.
In this case energy required to launch material on the altitude $H\sim 500$km is as follows
\begin{equation}
\label{energy}
E\simeq \frac{GMM_pH}{R_{\oplus}^2}\simeq 5.7\kappa\times 10^{27}\; ergs,
\end{equation}
where $M_p = M_E $ is Earth-like planet's mass and $M_E\simeq 6\times 10^{27}$ g is the Earth's mass.
On the other hand, power from the host star toward the planet is given by
\begin{equation}
\label{power}
P\simeq \frac{L}{4}\times\left(\frac{R_p}{r}\right)^2\simeq 1.7\kappa\times 10^{24}\; ergs\;s^{-1},
\end{equation}
where $L\simeq 3.8\times 10^{33}$ergs s$^{-1}$ is the solar-type star's luminosity and $r\simeq 1.5\times 10^{13}$ cm is the radius of the habitable zone. Whatever the propulsion mechanism (electromagnetic launch or some other mechanisms), it is evident from Eq. (\ref{power}) that the energy required to launch a thin shell around a planet can be extracted in approximately ten hours with $10\%$ of efficiency of energy conversion. If the megastructure is a web of many satellites orbiting the planet, then the kinetic energy should be added to the aforementioned value, leading to ten times more total energy, which, energetically is quite feasible for Type-I alien societies.
For a uniformly distributed spherical mass such a construction will not require any energy to stabilise it. Unlike this scenario, if instead of the sphere one uses a ring, the planet will be characterised by the out of plane stability, whereas for in-plane displacements the dynamics is unstable \citep{paper1}. In this work we have discussed that such rings may have sense only if the power required to maintain stability is small compared to the total received power, leading to the following condition for maximum displacement from equilibrium
\begin{equation}
\label{ksi}
\xi<<\frac{0.37}{R_p}\left(\frac{P}{M}\right)^{1/2}\left(\frac{2R_p^3}{GM_p}\right)^{3/4}\simeq 0.07,
\end{equation}
where $\xi\equiv d/R_p$ and $d$ denotes the displacement of the ring from the equilibrium position. We would like to emphasise that such a precision is not difficult to achieve: in the Lunar laser ranging experiment\footnote{Data is avalable from the Paris Observatory Lunar Analysis Center: http://polac.obspm.fr/llrdatae.htm}
the distance is measured with the precision of the order of $10^{-10}$.
By assuming that the whole emission incident on the megastructure is absorbed, one can easily show that the temperature is given by
\begin{equation}
\label{temp}
T\simeq\left(\frac{P}{16\sigma r^2}\right)^{1/4}\simeq 280 K,
\end{equation}
leading by means of the Wien's law to the emission peak in the IR spectral band with the following wavelength
\begin{equation}
\label{lambda}
\lambda = \frac{b}{T}\simeq 10 \mu m,
\end{equation}
where $\sigma$ denotes the Stefan-Boltzmann constant and $b\simeq 2898\mu$mK.
The average value of the flux can be estimated as follows
\begin{equation}
\label{flux}
F\simeq\frac{P}{4\pi D^2}\simeq 1.6\times10^{-17}\times\left(\frac{100 \;ly}{D}\right)^2 ergs\;s^{-1} cm^{-2},
\end{equation}
where the distance to the extrasolar system, $D$, is normalised by $100$ light years. The Very Large Telescope Interferometer (VLTI) could detect such fluxes. In particular, the limiting flux value of the VLTI instruments\footnote{https://www.eso.org/sci/facilities/paranal/telescopes/vlti} in $1$ hour is of the order of $2.3\times 10^{-18} ergs\;s^{-1} cm^{-2}$. Therefore, potentially one can monitor the spherical volume with radius of the order of $R_0 = 260\; ly$. In the Solar neighbourhood the stellar number density, $n$, is of the order of $0.138$ star pc$^{-3}$ \citep{dens}, which means that the total number of stars could be
\begin{equation}
\label{numb}
N\simeq\frac{4\pi R_0^3}{3}\; n\simeq 3\times 10^5,
\end{equation}
with approximately $1000$ G-type (Solar-type) stars. Although the flux method cannot distinguish emission from the planet and the megastructure, it allows to find objects in the habitable zone. For identifying the candidates of megaconstruction one can use the spatial resolution of the telescopes.
The VLTI telescope has the maximum angular resolution, $\theta_m$, of the order of $0.001$ mas (milliarcsecond - please see the technical characteristics of the VLTI instruments. By taking into account this value, one can obtain the maximum resolving distance
\begin{equation}
\label{dist}
D_m = \frac{2R}{\theta_m}\simeq 280\;ly.
\end{equation}
From the aforementioned two critical values, it is clear that planetary-scale megastructures inside $260\; ly$ might be detectable.
It is worth noting that a monolithic DS will be less stable compared to concentric rotating rings. In this case very interesting observational features might arise. By means of the non-relativistic Doppler effect, the observed wavelength is given by \citep{carroll}
\begin{equation}
\label{dopp1}
\lambda \simeq \lambda_s\left(1+v\cos\theta/c\right),
\end{equation}
where by $\lambda_s$ we represent the original wavelength emitted by the ring, $\theta$ is an angle of velocity direction measured in an observer's frame of reference, $v$ denotes the corresponding velocity and $c$ is the sped of light. Without going into geometric details, to make the orders of magnitude, we consider diametrically opposite sides of the ring relative to an observer moving with velocities $\upsilon_p \pm v_{_I}$, where $\upsilon_p$ denotes the orbital velocity of the planet around a host star and $v_{_I}=\sqrt{GM_p/R_p}$ is the orbital speed of the ring around a planet. Then, for the diametrically opposite sides of the ring from Eq. (\ref{dopp1}) one obtains
\begin{equation}
\label{dopp2}
\frac{\lambda_s}{\Delta\lambda}\simeq\frac{c}{2\upsilon_{_I}}\simeq 1.9\times 10^4\times \left(
\frac{R_p}{R_{\oplus}}\times \frac{M_{\oplus}}{M_p} \right)^{1/2}.
\end{equation}
In this regard, one should note that up to now the VLTI has the highest resolving power $RP\equiv\lambda/\Delta\lambda = 25000$ for $\lambda \approx 10 \mu$m (which corresponds to the blackbody radiation with $T=300$K) \footnote{www.eso.org/public/teles-instr/paranal-observatory/vlt/} and since $RP>\lambda_s/\Delta\lambda$, such orbital motions of the ring might be detected by the spectrographs of the VLTI instruments. Similarly, if instead of a solid mega-construction one uses a web of satellites, the same method might be used. It is worth noting that the observational pattern might be characterized by absorption lines of the matter the megastructure is made of. Therefore, in principle, the emission fingerprint might carry information about the matter, but this is not the scope of the current paper and we do not consider it.
One has to note that the SpaceX starlink satellites use radio communication. Therefore, a reasonable question might appear: is it possible to detect their radio signals by means of the China's FAST telescope? By taking into account that its system temperature is $T_{sys}\simeq 20$K and the illuminated aperture area, $A\simeq 7108$ cm$^2$ \citep{fast}, one can conclude that the minimum spectral flux density, $kT_{sys}/A$, which can be distinguished from noise is of the order of $6\times 10^{-24}$ ergs s$^{-1}$cm$^{-2}$Hz$^{-1}$. On the other hand, by assuming that only $1\%$ of the incident Solar energy is used for interstellar communication, one can find that if one uses the Hydrogen atom's frequency, 1420 Hz, the FAST telescope can detect such isotropic sources up to the distances of the order of $160$ pc.
Making use of simple geometry we have performed our calculations for solar-type stars with Earth-like planets and it is clear that the similar estimates can be straightforwardly performed for Type-I civilizations living in the systems with different parameters.
\section{Conclusion}
Extrapolating the idea of SpaceX's Starlink constellation we assume that an alien society with index $K = 0.7$ (our civilization) will reach Type-I in $1000$ years, which is enough to build a planetary megastructure for collecting the required material.
Launching has been analysed from the point of view energy costs and it has been shown that energy required to construct a megastructure is small compared to the total energy received from the star.
We have shown that the construction will be visible in the IR spectrum, which might be detected by VLTI instruments up to the distance $260$ light years, with $\sim 10^3$ Solar-type stars.
We have also emphasised that the spectral variability method might be an efficient tool to detect either orbital rotation of the solid megastructure or internal motions of small satellites.
It has been estimated the possibility to detect the radio emission operating on the Hydrogen atom?s frequency. It has been found that for reasonable parameters FAST can detect such radio sources from relatively large distances.
\section*{Acknowledgments}
The research was supported by the Shota Rustaveli National Science Foundation grant (NFR17-587).
|
\section{Introduction}
In 5G and beyond, the number of wireless devices, and their rate requirements, will increase by orders of magnitude \cite{paperG6}. To cope with such demands, different methods are proposed to improve the capacity and spectral efficiency. Here, one of the promising techniques is network densification, i.e., the deployment of many base stations (BSs) of different types such that there are more resource blocks per unit area. The BSs, however, need to be connected to the operators' core network via a transport network, the problem which becomes challenging
as the number of the BSs/users increases. Particularly, the increase of backhaul traffic may lead to backhaul congestion which, in turn, leads to end-to-end latency increment. This is the main motivation for wireless caching schemes reducing the backhaul load.
Caching is defined as storing popular reusable information at intermediate nodes to reduce the backhauling load. Such a technique is of interest in delay- and/or backhaul-constrained applications such as D2D, V2X and integrated access and backhaul.
Not every type of information is cacheable, for example, interactive applications such as gaming and voice calls. However, most of the network traffic today, including trending tweets, breaking news and video, is cacheable. Here, particular attention is paid to video. This is because, according to \cite{video2022}, by 2022, $79\%$ of the world's mobile traffic will be video. For instance, Netflix and YouTube alone account for almost half of peak downstream traffic in USA \cite{youtube}. Also, video is typically long and pre-recorded, which makes planning, prediction and segmentation easy.
Specially, as demonstrated in Fig. 1, video has high variation of the daily traffic profile, and increases the backhaul peak rate during the night significantly. This is important because the wireless network is designed based on the peak traffic. Then, with a variant daily traffic profile, the network will be underutilized most of the time, which is not economically viable. Thus, predicting the videos of interest in the high-traffic (HT) periods and caching them in the access points close to the devices during the low-traffic (LT) periods will lead to considerable cost reduction. For these reasons, caching is currently used by different content providers where, for instance, caches serve approximately $60\%-80\%$ (resp. up to $90\%$) of the Netflix \cite{Netflix2} (resp. Facebook \cite{facebook}) content requests.
\begin{figure}
\vspace{-2mm}
\centering
\includegraphics[width=0.6\columnwidth]{Figtraffic-eps-converted-to.pdf}\vspace{-2mm}
\caption{Application traffic daily profile in western Europe \cite[Fig. A3.3]{ssitu}.}\label{figure111}
\vspace{-4mm}
\end{figure}
The initial (un-coded) caching schemes where based on distributing the same information between the cache nodes, and minimizing the cache miss probability defined as the probability of the event that the device's requested file is not in the cache node. Here, depending on the amount of the information available about the sequence of the future requested files, different, e.g., Belady, highest-popularity-first, least-frequently-used and least-recently-used, algorithms have been proposed \cite{7080842}. Minimizing the cache miss probability improves the average system performance. On the other hand, the fundamental work of M. A. Maddah-ali and U. Niesen \cite{6763007,6807823} exploited the multicasting opportunity of cache networks and network coding concept to introduce coded-cashing, minimizing the worst-case backhaul peak rate in HT period. Also, following \cite{6763007,6807823}, there have been several works on the performance analysis of coded-caching systems, e.g., \cite{7857805,8674819,cairecodedcaching}. In these works, it is mainly concentrated on proper partitioning and distribution of the sub-packets between the cache nodes while the wireless channel between the serve and the cache nodes has been rarely studied, e.g. \cite{9014575}.
In this letter, we study the effect of adaptive data transmission and different data decoding/buffering schemes on the performance of wireless coded-caching networks. We present adaptive rate and power allocation schemes between the sub-packets such that the network successful transmission probability (STP) is maximized. Here, STP is defined as the probability of the event that all cache nodes can decode their intended signals correctly. Moreover, we investigate the effect of different decoding and buffering schemes on the network STP. We concentrate on the worst-case peak backhaul traffic cases where the cache nodes request for different signals during the HT periods. As we show, the performance of coded-caching networks is considerably affected by the decoding scheme as well as rate/power allocation.
\section{System Model}
\begin{figure}
\vspace{-2mm}
\centering
\includegraphics[width=0.6\columnwidth]{xFigchache2.jpg}\vspace{-2mm}
\caption{Coded-caching concept. Subplot A: Placements phase (at 2 AM) where the cache nodes are filled in with different sub-packets. Subplot B: Delivery phase (at 9 PM) where a single subpacket is braodcasted to the cache nodes. }\label{figure222}
\vspace{-5mm}
\end{figure}
As illustrated in Fig. 2, consider the simplest case of coded-caching networks where a server connects to two cache nodes $\text{C}_1$ and $\text{C}_2$. However, it is straightforward to extend the results to the cases with different numbers of cache nodes. In general, coded-caching has two, namely, placement and delivery, phases. During the placement phase, performed in the LT period (say, at 2 AM), the server divides the packets, for instance, packets $X$ and $\tilde X$, to sub-packets $X_1, X_2, \tilde X_1, \tilde X_2$ with $X=[X_1, X_2]$, $\tilde X=[\tilde X_1 \tilde X_2],$ and fills in the caches with different sub-packets having no knowledge, or probably a rough estimation, of the cache nodes' data requests during the HT period (see Fig. 2a). Assuming the packets $X$ and $\tilde X$ to be of length $2L$, the sub-packets $X_1, X_2, \tilde X_1, \tilde X_2$ are of length $L$. In this way, during the LT period, the server sends separate signals $[X_1 \tilde X_1]$ and $[X_2 \tilde X_2]$, each of length $2L,$ to $\text{C}_1$ and $\text{C}_2$, respectively.
During the HT period (say, 9 PM), the server serves the cache nodes based on their instantaneous data requests. Let $\bigoplus$ be the superposition operator. Also, considering the worst-case scenario in terms of backhaul traffic, assume that the caches $\text{C}_1$ and $\text{C}_2$ request for different packets $X$ and $\tilde X$, respectively. Then, as demonstrated in Fig. 2b, the server broadcasts a single sub-packet $X_2\bigoplus \tilde X_1$ of length $L$. Also, using the accumulated signals, each cache node may use different methods to decode its message of interest (see Section III). In this way, as shown in \cite{6763007,6807823}, coded-caching reduces the peak HT backhaul traffic by $50\%$ because, unlike uncoded caching, only a single sub-packet is broadcasted at, say, 9 PM. Finally, it is interesting to note that the presented coded-caching approach is a specific combination of the orthogonal multiple access (OMA) and NOMA (N:non) schemes used in different time slots with proper packet partitioning and signal decoding at the receivers.
Let us denote the server-cache $i$ channel coefficient by $h_i, i=1,2,$ and define the channel gains as $g_i=|h_i|^2, i=1,2.$ We consider Rayleigh-fading conditions with channel probability density functions (PDFs) $f_{g_i}(u)=\lambda_ie^{-\lambda_iu}, i=1,2,$ where $\lambda_i, i=1,2,$ depends on the long-term channel quality. Then, the signals received by $\text{C}_1$ and $\text{C}_2$ during the LT period are given by
\begin{align}\label{eq:eqLT}
\left\{\begin{matrix}
\left[Y_{\text{C}_1}^\text{LT}(t)\, \tilde Y_{\text{C}_1}^\text{LT}(t)\right]=\sqrt{P}h_1^\text{LT}\left[X_1(t) \tilde X_1(t)\right]+\left[Z_{\text{C}_1,1}(t)\, Z_{\text{C}_1,2}(t)\right], t=1,\ldots,L
\\
\left[Y_{\text{C}_2}^\text{LT}(t)\, \tilde Y_{\text{C}_2}^\text{LT}(t)\right]=
\sqrt{P}h_2^\text{LT}\left[X_2(t) \tilde X_2(t)\right]+\left[Z_{\text{C}_2,1}(t)\, Z_{\text{C}_2,2}(t)\right], t=1,\ldots,L
\end{matrix}\right.
\end{align}
while, at the HT period, the received signals are \begin{align}\label{eq:eqHT}
\left\{\begin{matrix}
Y_{C_1}^\text{HT}(t)=\sqrt{P}h_{1}^\text{HT}S(t)+Z_{\text{C}_1,1}(t), t=1,\ldots,L
\\
Y_{C_2}^\text{HT}(t)=\sqrt{P}h_{2}^\text{HT}S(t)+Z_{\text{C}_2,1}(t), t=1,\ldots,L
\\
S(t)=\alpha X_{2}(t)+\sqrt{1-\alpha^2}\tilde X_{1}(t).\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,
\end{matrix}\right.
\end{align}
Here, $h_i^\text{LT}$ and $h_i^\text{HT}, i=1,2,$ represent the channel realizations of $h_i$ at LT and HT periods, respectively. Then, $Z_{\text{C}_i,j},i,j=1,2,$ denotes the unit-variance additive Gaussian noise, and $P$ is the server maximum transmit power. Also, $X_i, \tilde X_i, i=1,2,$ are the unit-variance signals of the sub-packets, $S(t)$ represents the unit-variance superimposed signals of $X_2$ and $\tilde X_1$, and $\alpha\in [0,1]$ gives the power partitioning between these signals.
\vspace{-4mm}
\section{Analytical results}
Let us denote the number of information bits in the packets $X$ and $\tilde X$ by $K$ and $\tilde K$, respectively. Also, the information nats are divided between the sub-packets according to
\begin{align}\label{eq:eqbitpart}
\left\{\begin{matrix}
K=K_1+K_2
\\
\tilde K=\tilde K_1+\tilde K_2,
\end{matrix}\right.
\end{align}
which, defining the code rates $R=\frac{K}{2L}$ and $\tilde R=\frac{\tilde K}{2L},$ leads to \begin{align}\label{eq:eqratepart}
\left\{\begin{matrix}
2R=R_1+R_2
\\
2\tilde R=\tilde R_1+\tilde R_2.
\end{matrix}\right.
\end{align}
Our goal is to design proper rate allocation between sub-packets, i.e., $R_i,\tilde R_i, i=1,2,$ and power split parameter $\alpha$ in (\ref{eq:eqHT}) such that the network STP is maximized. STP is defined as
\begin{align}\label{eq:eqSTPdef}
\Pr(\text{ST})=\frac{1}{2}\left(\Pr\left(\text{C}_1 \,\text{successful}\right)+\Pr\left(\text{C}_2 \,\text{successful}\right)\right),
\end{align}
i.e., the probability of the event that both cache nodes can decode their intended packets correctly. Depending on the buffering and the coding schemes of the cache nodes, the network may lead to different STPs as follows.
\vspace{-3mm}
\subsection{Joint Decoding at HT Period using Successive Interference Cancellation}
In one approach, referred to as Method 1 in the following, the cache nodes buffer the signals received in the LT period, and use both maximum ratio combining (MRC) and successive interference cancellation (SIC) for message decoding at HT periods. Let us concentrate on cache node $\text{C}_1$. Receiving $Y_{\text{C}_1}^\text{HT}$ in (\ref{eq:eqHT}) during the HT period and with $\tilde Y_{\text{C}_1}^\text{LT}$ in (\ref{eq:eqLT}) already buffered, $\text{C}_1$ first uses MRC to decode $\tilde X_1(t)$. Then, with a successful decoding of $\tilde X_1(t)$, node $\text{C}_1$ uses the SIC concept to remove $\tilde X_1(t)$ from $Y_{\text{C}_1}^\text{HT}(t)$, leading to an interference-free signal
\vspace{-2mm}
\begin{align}\label{eq:Eqsic1}
\mathcal{Y}_{\text{C}_1}^\text{HT}=\sqrt{P}h_1^\text{HT}\alpha X_2(t)+Z_{\text{C}_1,1}(t).
\end{align}
Finally, the decoder generates the concatenated signal $[Y_{\text{C}_1}^\text{LT}(t) \,\mathcal{Y}_{\text{C}_1}^\text{HT}(t)]$, with $Y_{\text{C}_1}^\text{LT}(t)$ and $\mathcal{Y}_{\text{C}_1}^\text{HT}(t)$ given in (\ref{eq:eqLT}) and (\ref{eq:Eqsic1}), respectively, and decodes the packet $X$ in \emph{one-shot.} In this way, and considering the same procedure in $\text{C}_2$ by using (\ref{eq:eqLT})-(\ref{eq:eqHT}), STP in (\ref{eq:eqSTPdef}) is given by
\begin{align}\label{eq:eqSTPdef1}
&\Pr(\text{ST})=\frac{1}{2}(\eta_1\gamma_1+\eta_2\gamma_2),\nonumber\\&
\eta_1=\Pr\left(\log\left(1+\frac{(1-\alpha^2)Pg_1^\text{HT}}{1+\alpha^2Pg_1^\text{HT}}+Pg_1^\text{LT}\right)\ge \tilde R_1\right),
\nonumber\\&\gamma_1=\Pr\left(\log\left(1+Pg_1^\text{LT}\right)+\log\left(1+\alpha^2Pg_1^\text{HT}\right)\ge 2R\right)\nonumber\\&
\eta_2=\Pr\left(\log\left(1+\frac{\alpha^2Pg_2^\text{HT}}{1+(1-\alpha^2)Pg_2^\text{HT}}+Pg_2^\text{LT}\right)\ge R_2\right),\nonumber\\
&\gamma_2=\Pr\left(\log\left(1+Pg_2^\text{LT}\right)+\log\left(1+(1-\alpha^2)Pg_2^\text{HT}\right)\ge 2\tilde R\right).
\end{align}
Here, $\eta_1$ (resp. $\eta_2$) is the probability of successful decoding of $\tilde X_1(t)$ (resp. $X_2(t)$) at $\text{C}_1$ (resp. $\text{C}_2$) using MRC. Then, $\gamma_1$ and $\gamma_2$ give the probability that, removing the interference from the received signal in HT period, the caches can decode their intended signals correctly. Note that in $\gamma_1$ and $\gamma_2$ we have used the results on the maximum achievable rates of parallel Gaussian channels.
Considering Rayleigh-fading conditions, we have
\begin{align}\label{eq:eqeta11}
&\eta_1=1-\Pr\left(\frac{(1-\alpha^2)Pg_1^\text{HT}}{1+\alpha^2Pg_1^\text{HT}}+Pg_1^\text{LT}\le e^{\tilde R_1}-1\right)=
\nonumber\\&
1-\int_0^{\frac{e^{\tilde R_1}-1}{P}}f_{g_1}(y)\Pr\left(\frac{(1-\alpha^2)Pg_1^\text{HT}}{1+\alpha^2Pg_1^\text{HT}} \le {e^{\tilde R_1}-1}-Py \right)\text{d}y
\nonumber\\&
=e^{-\frac{\lambda_1\left({e^{\tilde R_1}-1}\right)}{P}}+
\int_{\frac{\alpha^2 e^{\tilde R_1}-1}{\alpha^2P}}^{\frac{e^{\tilde R_1}-1}{P}}e^{\lambda_1\left(y+\frac{{e^{\tilde R_1}-1}-Py}{\alpha^2P-(1-\alpha^2)P\left({e^{\tilde R_1}-1}-Py\right)}\right)}\text{d}y,
\end{align}
which can be calculated numerically. Also, following the same procedure, we have
\begin{align}\label{eq:Eqeta22}
&\eta_2=e^{-\frac{\lambda_2\left(e^{R_2}-1\right)}{P}}
+\int_{\frac{\left(1-\alpha^2\right)e^{R_2}-1}{\left(1-\alpha^2\right)P}}^{\frac{e^{R_2}-1}{P}}{\lambda_2e^{-\lambda_2\left(y+\frac{e^{R_2}-1-Py}{\alpha^2P-\left(1-\alpha^2\right)P\left(e^{R_2}-1-Py\right)}\right)}}\text{d}y.
\end{align}
The terms $\gamma_i,i=1,2,$ on the other hand, do not have closed-form or easy-to-deal integration expressions. Thus, we use the Jensen's inequality $\frac{1}{n}\sum_{i=1}^n\log(1+x_i)\le \log\left(1+\frac{1}{n}\sum_{i=1}^nx_i\right)$ \cite[Eq. (30)]{7445896} and $f_{g_i}(u)=\lambda_ie^{-\lambda_iu}, i=1,2,$ to rephrase $\gamma_i,i=1,2,$ as
\vspace{-2mm}
\begin{align}\label{eq:eqgamma1}
\gamma_1&\le 1-\Pr\left(g_1^\text{LT}+\alpha^2g_1^\text{HT}\le \frac{2}{P}\left(e^R-1\right)\right) \nonumber\\&
=1-\int_0^{\frac{2\left(e^R-1\right)}{P}}{f_{g_1}(x)\Pr\left(g_1^\text{HT}\le\frac{\frac{2\left(e^R-1\right)}{P}-x}{\alpha^2}\right)}\text{d}x\nonumber\\&
=e^{\frac{-2\lambda_1\theta}{P}}+\frac{\alpha^2 e^{\frac{-2\lambda_1\theta}{P\alpha^2}}}{\alpha^2-1}\left(1-e^{\frac{-2\lambda_1\left(\alpha^2-1\right)\theta}{P\alpha^2}}\right),
\end{align}
\vspace{-2mm}
\begin{align}\label{eq:eqgamma2}
\gamma_2&\le e^{\frac{-2\lambda_2\theta}{P}}+\frac{\left(\alpha^2-1\right) e^{\frac{-2\lambda_2\theta}{P\left(1-\alpha^2\right)}}}{\alpha^2}\left(1-e^{\frac{-2\lambda_2\alpha^2\theta}{P\left(\alpha^2-1\right)}}\right),
\end{align}
where (\ref{eq:eqgamma2}) follows the same procedure as in (\ref{eq:eqgamma1}). In this way, the optimal rate/power allocation maximizing STP is given by
\begin{align}\label{eq:eqoptprob}
&\max \frac{1}{2}\{\eta_1\gamma_1+\eta_2\gamma_2\}\nonumber\\
&\text{s.t. } \,\alpha\in[0,1], R_1+R_2=2R, \,\tilde R_1+\tilde R_2=2\tilde R,
\end{align}
which can be effectively solved by, e.g., exhaustive search or the machine-learning based scheme of \cite{8520925}.
\vspace{-2mm}
\subsection{Joint Decoding at HT Period without SIC}
Implementation of MRC and SIC, to decode and remove the interference, increases the decoding complexity/delay. Also, SIC suffers from error propagation problem, e.g., \cite{nomaharq}. For these reasons, in Method 2, each cache node decodes its intended packets in one-shot by considering the interference as an additive noise. Here, (\ref{eq:eqSTPdef1}) is rephrased as
\begin{align}\label{eq:eqSTPdef3}
&\Pr(\text{ST})=\frac{1}{2}(\bar \gamma_1+\bar \gamma_2),\nonumber\\&
\bar\gamma_1=\Pr\left(\log\left(1+Pg_1^\text{LT}\right)+\log\left(1+\frac{\alpha^2Pg_1^\text{HT}}{1+(1-\alpha^2)Pg_1^\text{HT}}\right)\ge 2R\right)\nonumber\\&
\bar\gamma_2=\Pr\left(\log\left(1+Pg_2^\text{LT}\right)+\log\left(1+\frac{(1-\alpha^2)Pg_2^\text{HT}}{1+\alpha^2Pg_2^\text{HT}}\right)\ge 2\tilde R\right),
\end{align}
where, using the Jensen's inequality and the same procedure as in (\ref{eq:eqgamma1}), we have
\begin{align}\label{eq:eqjensen2}
&\bar\gamma_1\le e^{-\frac{2\lambda_1\left(e^R-1\right)}{P}}+\int_{\frac{2\left(e^R-1\right)-\frac{\alpha^2}{1-\alpha^2}}{P}}^{\frac{2\left(e^R-1\right)}{P}}\lambda_1e^{-\lambda_1\left(x+\frac{2\left(e^R-1\right)-Px}{\alpha^2P+(1-\alpha^2)P(Px-2\left(e^R-1\right))}\right)}\text{d}x,\nonumber\\&
\bar\gamma_2\le e^{-\frac{2\lambda_2\left(e^{\tilde R}-1\right)}{P}}
+\int_{\frac{2\left(e^{\tilde R}-1\right)-\frac{1-\alpha^2}{\alpha^2}}{P}}^{\frac{2\left(e^{\tilde R}-1\right)}{P}}\lambda_2e^{-\lambda_2\left(x+\frac{2\left(e^{\tilde R}-1\right)-Px}{(1-\alpha^2)P+\alpha^2P(Px-2\left(e^{\tilde R}-1\right))}\right)}\text{d}x,
\end{align}
which can be calculated numerically.
Finally, note that, replacing (\ref{eq:eqSTPdef3}) into (\ref{eq:eqoptprob}), the optimal performance of the cache nodes in Method 2 is independent of the rate split between the sub-packets. This, although Method 1 gives the best performance in terms of the worst-case peak traffic,
may give an advantage to Method 2, compared to Method 1. This is because in Method 1 the rate split is performed by considering the worst-case condition with the cache nodes requesting for different signals during HT period. However, if the caches request for the same signals during HT period, the rate split scheme of Method 1 is not necessarily optimal. As opposed, in Method 2, the rate split is independent of the caches requested signals in HT periods.
\subsection{Separate Decoding using SIC}
In Methods 1-2, one needs to follow the coding schemes of incremental redundancy hybrid automatic repeat request (HARQ)-based protocols or Raptor codes, e.g., \cite{6164088}, where the message is decoded by concatenating different sub-packets. Alternatively, in Method 3, we consider the case where, while MRC and SIC are used to decode and remove the interference signal, respectively, each cache node decodes its sub-packets of interest separately. That is, considering $\text{C}_1,$ $X_1$ (resp. $X_2$) is decoded during the LT (resp. HT) period. In this case, the STP (\ref{eq:eqSTPdef1}) is changed to
\vspace{-1mm}
\begin{align}\label{eq:eqSTPdef2}
&\Pr\left(\text{ST}\right)=\frac{1}{2}\left(\eta_1\breve{\gamma}_{11}\breve{\gamma}_{12}+\eta_2\breve{\gamma}_{21}\breve{\gamma}_{22}\right),\nonumber\\&
\breve{\gamma}_{11}=\Pr\left(\log\left(1+Pg_1^\text{LT}\right)\ge R_1\right)=e^{-\frac{\lambda_1\left(e^{R_1}-1\right)}{P}},\nonumber\\&
\breve{\gamma}_{12}=\Pr\left(\log\left(1+\alpha^2Pg_1^\text{HT}\right)\ge R_2\right)=e^{-\frac{\lambda_1\left(e^{R_2}-1\right)}{\alpha^2P}}
\nonumber\\&
\breve{\gamma}_{21}=\Pr\left(\log\left(1+(1-\alpha^2)Pg_2^\text{HT}\right)\ge \tilde R_1\right)=e^{-\frac{\lambda_2\left(e^{\tilde R_1}-1\right)}{(1-\alpha^2)P}},\nonumber\\&
\breve{\gamma}_{22}=\Pr\left(\log\left(1+Pg_2^\text{LT}\right)\ge \tilde R_2\right)=e^{-\frac{\lambda_2\left(e^{\tilde R_2}-1\right)}{P}},
\end{align}
with $\eta_i,i=1,2,$ given in (\ref{eq:eqSTPdef1}), and (\ref{eq:eqoptprob}) is adapted correspondingly. In (\ref{eq:eqSTPdef2}), $\breve{\gamma}_{11}$ is the probability that $\text{C}_1$ decodes $X_1$ during the LT period. Also, $\breve{\gamma}_{12}$ gives the probability that, after decoding and removing $\tilde X_1$, the cache node $\text{C}_1$ correctly decodes $X_2$ in the HT period. Also, the same arguments hold for $\breve{\gamma}_{2i},i=1,2.$
Note that, although Method 1 maximizes the achievable rate/STP, Method 3 has a number of advantages including:
\begin{itemize}
\item \textbf{Low decoding complexity}: Because, as opposed to Methods 1-2 decoding long codewords of length $2L$, Method 3 is based on decoding sub-packets of length $L$.
\item \textbf{Efficient HARQ-based transmissions}: In Methods 1-2, all packets are decoded during the HT periods and, in case of decoding failure, the message is retransmitted at that period. Such HARQ-based retransmissions increase the backhauling load at HT period. As opposed, in Method 3, the decoding of the first received sub-packets and all their required HARQ-based retransmissions are performed during the LT period, which reduces the backhauling cost of HARQ.
\end{itemize}
Finally, depending on the considered method, the buffering scheme of the caches during the LT period may change. Particularly, in Methods 1-2 the caches buffer the signals received during LT period without decoding. In Method 3, however, the caches buffer the sub-packets successfully decoded during LT period.
\vspace{-1mm}
\subsection{Separate Decoding without SIC}
To further reduce the complexity of Method 3, one can consider the case where, while decoding the sub-packets separately, the cache nodes consider the interference as an additive noise (Method 4). In this case, where the sub-packets are decoded in different LT and HT periods without SIC, the STP is given by
\begin{align}\label{eq:eqSTPdef4}
&\Pr\left(\text{ST}\right)=\frac{1}{2}(\breve{\gamma}_{11}\hat{\gamma}_{12}+\hat{\gamma}_{21}\breve{\gamma}_{22}),\nonumber\\&
\hat{\gamma}_{12}=\Pr\left(\log\left(1+\frac{\alpha^2Pg_1^\text{HT}}{1+(1-\alpha^2)Pg_1^\text{HT}}\right)\ge R_2\right)\nonumber\\&\,\,\,\,\,\,\,\,=\left\{\begin{matrix}
0\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\, & \text{if } R_2\ge -\log\left(1-\alpha^2 \right )\\
e^{-\lambda_1\frac{\left(e^{R_2}-1 \right )}{\alpha^2P-(1-\alpha^2)P\left(e^{R_2}-1 \right )}} & \text{otherwise}
\end{matrix}\right.
\nonumber\\&
\hat{\gamma}_{21}=\Pr\left(\log\left(1+\frac{(1-\alpha^2)Pg_2^\text{HT}}{1+\alpha^2Pg_2^\text{HT}}\right)\ge \tilde R_1\right)\nonumber\\&=\left\{\begin{matrix}
0\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\, & \text{if } \tilde R_1\ge -2\log\alpha\\
e^{-\lambda_2\frac{\left(e^{\tilde R_1}-1 \right )}{(1-\alpha^2)P-\alpha^2P\left(e^{\tilde R_1}-1 \right )}} & \text{otherwise}
\end{matrix}\right.
\end{align}
with $\breve{\gamma}_{11}$ and $\breve{\gamma}_{22}$ given in (\ref{eq:eqSTPdef2}). Also, in (\ref{eq:eqSTPdef4}) we use Rayleigh channel PDFs and some manipulations to derive the probabilities. For further comparisons between Methods 1-4, see Section IV.
\section{Simulation Results}
The simulation results are presented for the cases with $\lambda_1=1$ and $\lambda_2=0.1$, i.e., with $10$ dB difference between the channel gains of the server-cache links, and we define the transmission signal-to-noise ratio (SNR) as $10\log_{10} P$, considering the additive noises to be unit-variance. Note that we have evaluated the results for different parameter settings, and they show the same qualitative conclusions as those presented in the following. In Figs. 3-4, the results are obtained by optimizing the rate and power allocation. Here, both exhaustive search and the genetic-algorithm based scheme of \cite{8520925} have been used which have ended up in the same results, indicating the accuracy of the optimization process. In Fig. 5, we study the effect of rate/power allocation.
\begin{figure}
\vspace{-3mm}
\centering
\includegraphics[width=0.6\columnwidth]{FigCodedcaching1-eps-converted-to.pdf}\\\vspace{-3mm}
\caption{Comparison between the STP of Methods 1-4, $\lambda_1=1,$ $\lambda_2=0.1,$ $R=\tilde R=1$ npcu. The results are obtained by optimal rate and power allocation.}
\label{fig:fig_ivd9}
\vspace{-2mm}
\end{figure}
Considering $R=\tilde R=1$ nats-per-channel-use (npcu), Fig. 3 compares the performance of Methods 1-4. Also, the figure verifies the tightness of the Jensen's inequality-based approximation results of (\ref{eq:eqgamma1})-(\ref{eq:eqgamma2}). Then, Fig. 4 shows the STP versus the data rates $R=\tilde R$ for the cases with different decoding/buffering methods and transmission SNRs.
Considering Methods 1 and 3, with joint and separate decoding on the sub-packets, respectively, Fig. 5 studies the effect of optimal rate and power allocation on the network STP. Particularly, the figure compares the optimal results obtained by (\ref{eq:eqoptprob}) with the cases using uniform power allocation, i.e., with $\alpha=\frac{\sqrt{2}}{2}$ in (\ref{eq:eqHT}), and/or equal rate split, i.e., $R_i=\tilde R_i, i=1,2$. According to the figures, the following conclusions can be drawn:
\begin{figure}
\vspace{-3mm}
\centering
\includegraphics[width=0.6\columnwidth]{FigCodedcachingrate-eps-converted-to.pdf}\\\vspace{-2mm}
\caption{Achievable STP versus the data rate for different methods. The results are obtained by optimal rate and power allocation and $\lambda_1=1, \lambda_2=0.1.$}
\label{fig:fig_ivd10}
\vspace{-0mm}
\end{figure}
\begin{figure}
\vspace{-3mm}
\centering
\includegraphics[width=0.6\columnwidth]{plottingcomparison-eps-converted-to.pdf}\\\vspace{-2mm}
\caption{On the effect of optimal rate and power allocation in Methods 1 and 3, $R=\tilde R=1$ npcu and $\lambda_1=1, \lambda_2=0.1$.}
\label{fig:fig_ivd11}
\vspace{-0mm}
\end{figure}
\begin{itemize}
\item The approximation results of (\ref{eq:eqgamma1})-(\ref{eq:eqgamma2}) properly approximate the probabilities $\gamma_i, i=1,2$ (Fig. 3. Also, the same point is observed for $\bar \gamma_i,i=1,2,$ in (\ref{eq:eqjensen2}) although not shown in the figure). Thus, the approximations can be well utilized for the performance evaluation of Methods 1-2, i.e., in the cases with joint decoding of the sub-packets.
\item Compared to the cases with separate decoding of sub-packets, i.e., Methods 3-4, considerable STP improvement is observed by joint decoding of the sub-packet, i.e., Methods 1-2 (Figs. 3-4). However, as explained in Section III.C, the STP increment of Methods 1-2 is at the cost of decoding complexity and possible HARQ-based retransmissions at HT periods. On the other hand, for both cases with joint and separate decoding of the sub-packets, using SIC-based interference cancellation leads to marginal performance improvement at low rates while its effect increases with the data rate (Fig. 4).
Finally, as the data rate increases, the performance gap between Methods 2 and 3 decreases, i.e., one can reach the same STP as in the cases with joint sub-packet decoding of interference-affected signals by separate sub-packets decoding if the interference signals are removed using SIC. (Fig. 4).
\item For all parameter settings, Method 1 leads to the highest STP, compared to Methods 2-4, at the cost of decoding delay/possible retransmissions at HT periods (Figs. 3-4). For instance, with the parameter settings of Fig. 4 and data rate 1.5 npcu, the implementation of Method 1 with transmit SNR 10 dB results in the same STP, $80\%$, as in the cases with Method 4 and SNR 15 dB, i.e., using advanced decoding methods leads to 5 dB gain in SNR (Fig. 4).
\item For both cases with and without interference cancellation (Methods 1 and 3), optimal rate allocation leads to considerable STP increment (Fig. 5. The same conclusion is observed in Methods 2 and 4, although not presented in the figure). Also, the relative performance gain of optimal rate split increases in the cases with interference cancellation. Finally, considering the interference as additive noise, optimal power allocation between sub-packets during HT period increases the STP. However, with interference cancellation and joint decoding of sub-packets, the effect of optimal power allocation between HT period sub-packets is marginal (Fig. 5).
\end{itemize}
\section{Conclusions}
This paper studied the performance coded-caching networks in the cases with adaptive rate/power allocation and different decoding/buffering schemes. As we showed, joint decoding of the sub-packets at HT periods leads to considerable performance improvement of coded-caching setups. Also, for different decoding schemes, optimal rate split between the sub-packets increases the STP considerably while optimal power allocation between the sub-packets of HT period only improves the STP if SIC-based receiver is not implemented and the sub-packets are decoded separately.
\vspace{-2mm}
\bibliographystyle{IEEEtran}
\section{Introduction}
In 5G and beyond, the number of wireless devices, and their rate requirements, will increase by orders of magnitude \cite{paperG6}. To cope with such demands, different methods are proposed to improve the capacity and spectral efficiency. Here, one of the promising techniques is network densification, i.e., the deployment of many base stations (BSs) of different types such that there are more resource blocks per unit area. The BSs, however, need to be connected to the operators' core network via a transport network, the problem which becomes challenging
as the number of the BSs/users increases. Particularly, the increase of backhaul traffic may lead to backhaul congestion which, in turn, leads to end-to-end latency increment. This is the main motivation for wireless caching schemes reducing the backhaul load.
Caching is defined as storing popular reusable information at intermediate nodes to reduce the backhauling load. Such a technique is of interest in delay- and/or backhaul-constrained applications such as D2D, V2X and integrated access and backhaul.
Not every type of information is cacheable, for example, interactive applications such as gaming and voice calls. However, most of the network traffic today, including trending tweets, breaking news and video, is cacheable. Here, particular attention is paid to video. This is because, according to \cite{video2022}, by 2022, $79\%$ of the world's mobile traffic will be video. For instance, Netflix and YouTube alone account for almost half of peak downstream traffic in USA \cite{youtube}. Also, video is typically long and pre-recorded, which makes planning, prediction and segmentation easy.
Specially, as demonstrated in Fig. 1, video has high variation of the daily traffic profile, and increases the backhaul peak rate during the night significantly. This is important because the wireless network is designed based on the peak traffic. Then, with a variant daily traffic profile, the network will be underutilized most of the time, which is not economically viable. Thus, predicting the videos of interest in the high-traffic (HT) periods and caching them in the access points close to the devices during the low-traffic (LT) periods will lead to considerable cost reduction. For these reasons, caching is currently used by different content providers where, for instance, caches serve approximately $60\%-80\%$ (resp. up to $90\%$) of the Netflix \cite{Netflix2} (resp. Facebook \cite{facebook}) content requests.
\begin{figure}
\vspace{-2mm}
\centering
\includegraphics[width=0.6\columnwidth]{Figtraffic-eps-converted-to.pdf}\vspace{-2mm}
\caption{Application traffic daily profile in western Europe \cite[Fig. A3.3]{ssitu}.}\label{figure111}
\vspace{-4mm}
\end{figure}
The initial (un-coded) caching schemes where based on distributing the same information between the cache nodes, and minimizing the cache miss probability defined as the probability of the event that the device's requested file is not in the cache node. Here, depending on the amount of the information available about the sequence of the future requested files, different, e.g., Belady, highest-popularity-first, least-frequently-used and least-recently-used, algorithms have been proposed \cite{7080842}. Minimizing the cache miss probability improves the average system performance. On the other hand, the fundamental work of M. A. Maddah-ali and U. Niesen \cite{6763007,6807823} exploited the multicasting opportunity of cache networks and network coding concept to introduce coded-cashing, minimizing the worst-case backhaul peak rate in HT period. Also, following \cite{6763007,6807823}, there have been several works on the performance analysis of coded-caching systems, e.g., \cite{7857805,8674819,cairecodedcaching}. In these works, it is mainly concentrated on proper partitioning and distribution of the sub-packets between the cache nodes while the wireless channel between the serve and the cache nodes has been rarely studied, e.g. \cite{9014575}.
In this letter, we study the effect of adaptive data transmission and different data decoding/buffering schemes on the performance of wireless coded-caching networks. We present adaptive rate and power allocation schemes between the sub-packets such that the network successful transmission probability (STP) is maximized. Here, STP is defined as the probability of the event that all cache nodes can decode their intended signals correctly. Moreover, we investigate the effect of different decoding and buffering schemes on the network STP. We concentrate on the worst-case peak backhaul traffic cases where the cache nodes request for different signals during the HT periods. As we show, the performance of coded-caching networks is considerably affected by the decoding scheme as well as rate/power allocation.
\section{System Model}
\begin{figure}
\vspace{-2mm}
\centering
\includegraphics[width=0.6\columnwidth]{xFigchache2.jpg}\vspace{-2mm}
\caption{Coded-caching concept. Subplot A: Placements phase (at 2 AM) where the cache nodes are filled in with different sub-packets. Subplot B: Delivery phase (at 9 PM) where a single subpacket is braodcasted to the cache nodes. }\label{figure222}
\vspace{-5mm}
\end{figure}
As illustrated in Fig. 2, consider the simplest case of coded-caching networks where a server connects to two cache nodes $\text{C}_1$ and $\text{C}_2$. However, it is straightforward to extend the results to the cases with different numbers of cache nodes. In general, coded-caching has two, namely, placement and delivery, phases. During the placement phase, performed in the LT period (say, at 2 AM), the server divides the packets, for instance, packets $X$ and $\tilde X$, to sub-packets $X_1, X_2, \tilde X_1, \tilde X_2$ with $X=[X_1, X_2]$, $\tilde X=[\tilde X_1 \tilde X_2],$ and fills in the caches with different sub-packets having no knowledge, or probably a rough estimation, of the cache nodes' data requests during the HT period (see Fig. 2a). Assuming the packets $X$ and $\tilde X$ to be of length $2L$, the sub-packets $X_1, X_2, \tilde X_1, \tilde X_2$ are of length $L$. In this way, during the LT period, the server sends separate signals $[X_1 \tilde X_1]$ and $[X_2 \tilde X_2]$, each of length $2L,$ to $\text{C}_1$ and $\text{C}_2$, respectively.
During the HT period (say, 9 PM), the server serves the cache nodes based on their instantaneous data requests. Let $\bigoplus$ be the superposition operator. Also, considering the worst-case scenario in terms of backhaul traffic, assume that the caches $\text{C}_1$ and $\text{C}_2$ request for different packets $X$ and $\tilde X$, respectively. Then, as demonstrated in Fig. 2b, the server broadcasts a single sub-packet $X_2\bigoplus \tilde X_1$ of length $L$. Also, using the accumulated signals, each cache node may use different methods to decode its message of interest (see Section III). In this way, as shown in \cite{6763007,6807823}, coded-caching reduces the peak HT backhaul traffic by $50\%$ because, unlike uncoded caching, only a single sub-packet is broadcasted at, say, 9 PM. Finally, it is interesting to note that the presented coded-caching approach is a specific combination of the orthogonal multiple access (OMA) and NOMA (N:non) schemes used in different time slots with proper packet partitioning and signal decoding at the receivers.
Let us denote the server-cache $i$ channel coefficient by $h_i, i=1,2,$ and define the channel gains as $g_i=|h_i|^2, i=1,2.$ We consider Rayleigh-fading conditions with channel probability density functions (PDFs) $f_{g_i}(u)=\lambda_ie^{-\lambda_iu}, i=1,2,$ where $\lambda_i, i=1,2,$ depends on the long-term channel quality. Then, the signals received by $\text{C}_1$ and $\text{C}_2$ during the LT period are given by
\begin{align}\label{eq:eqLT}
\left\{\begin{matrix}
\left[Y_{\text{C}_1}^\text{LT}(t)\, \tilde Y_{\text{C}_1}^\text{LT}(t)\right]=\sqrt{P}h_1^\text{LT}\left[X_1(t) \tilde X_1(t)\right]+\left[Z_{\text{C}_1,1}(t)\, Z_{\text{C}_1,2}(t)\right], t=1,\ldots,L
\\
\left[Y_{\text{C}_2}^\text{LT}(t)\, \tilde Y_{\text{C}_2}^\text{LT}(t)\right]=
\sqrt{P}h_2^\text{LT}\left[X_2(t) \tilde X_2(t)\right]+\left[Z_{\text{C}_2,1}(t)\, Z_{\text{C}_2,2}(t)\right], t=1,\ldots,L
\end{matrix}\right.
\end{align}
while, at the HT period, the received signals are \begin{align}\label{eq:eqHT}
\left\{\begin{matrix}
Y_{C_1}^\text{HT}(t)=\sqrt{P}h_{1}^\text{HT}S(t)+Z_{\text{C}_1,1}(t), t=1,\ldots,L
\\
Y_{C_2}^\text{HT}(t)=\sqrt{P}h_{2}^\text{HT}S(t)+Z_{\text{C}_2,1}(t), t=1,\ldots,L
\\
S(t)=\alpha X_{2}(t)+\sqrt{1-\alpha^2}\tilde X_{1}(t).\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,
\end{matrix}\right.
\end{align}
Here, $h_i^\text{LT}$ and $h_i^\text{HT}, i=1,2,$ represent the channel realizations of $h_i$ at LT and HT periods, respectively. Then, $Z_{\text{C}_i,j},i,j=1,2,$ denotes the unit-variance additive Gaussian noise, and $P$ is the server maximum transmit power. Also, $X_i, \tilde X_i, i=1,2,$ are the unit-variance signals of the sub-packets, $S(t)$ represents the unit-variance superimposed signals of $X_2$ and $\tilde X_1$, and $\alpha\in [0,1]$ gives the power partitioning between these signals.
\vspace{-4mm}
\section{Analytical results}
Let us denote the number of information bits in the packets $X$ and $\tilde X$ by $K$ and $\tilde K$, respectively. Also, the information nats are divided between the sub-packets according to
\begin{align}\label{eq:eqbitpart}
\left\{\begin{matrix}
K=K_1+K_2
\\
\tilde K=\tilde K_1+\tilde K_2,
\end{matrix}\right.
\end{align}
which, defining the code rates $R=\frac{K}{2L}$ and $\tilde R=\frac{\tilde K}{2L},$ leads to \begin{align}\label{eq:eqratepart}
\left\{\begin{matrix}
2R=R_1+R_2
\\
2\tilde R=\tilde R_1+\tilde R_2.
\end{matrix}\right.
\end{align}
Our goal is to design proper rate allocation between sub-packets, i.e., $R_i,\tilde R_i, i=1,2,$ and power split parameter $\alpha$ in (\ref{eq:eqHT}) such that the network STP is maximized. STP is defined as
\begin{align}\label{eq:eqSTPdef}
\Pr(\text{ST})=\frac{1}{2}\left(\Pr\left(\text{C}_1 \,\text{successful}\right)+\Pr\left(\text{C}_2 \,\text{successful}\right)\right),
\end{align}
i.e., the probability of the event that both cache nodes can decode their intended packets correctly. Depending on the buffering and the coding schemes of the cache nodes, the network may lead to different STPs as follows.
\vspace{-3mm}
\subsection{Joint Decoding at HT Period using Successive Interference Cancellation}
In one approach, referred to as Method 1 in the following, the cache nodes buffer the signals received in the LT period, and use both maximum ratio combining (MRC) and successive interference cancellation (SIC) for message decoding at HT periods. Let us concentrate on cache node $\text{C}_1$. Receiving $Y_{\text{C}_1}^\text{HT}$ in (\ref{eq:eqHT}) during the HT period and with $\tilde Y_{\text{C}_1}^\text{LT}$ in (\ref{eq:eqLT}) already buffered, $\text{C}_1$ first uses MRC to decode $\tilde X_1(t)$. Then, with a successful decoding of $\tilde X_1(t)$, node $\text{C}_1$ uses the SIC concept to remove $\tilde X_1(t)$ from $Y_{\text{C}_1}^\text{HT}(t)$, leading to an interference-free signal
\vspace{-2mm}
\begin{align}\label{eq:Eqsic1}
\mathcal{Y}_{\text{C}_1}^\text{HT}=\sqrt{P}h_1^\text{HT}\alpha X_2(t)+Z_{\text{C}_1,1}(t).
\end{align}
Finally, the decoder generates the concatenated signal $[Y_{\text{C}_1}^\text{LT}(t) \,\mathcal{Y}_{\text{C}_1}^\text{HT}(t)]$, with $Y_{\text{C}_1}^\text{LT}(t)$ and $\mathcal{Y}_{\text{C}_1}^\text{HT}(t)$ given in (\ref{eq:eqLT}) and (\ref{eq:Eqsic1}), respectively, and decodes the packet $X$ in \emph{one-shot.} In this way, and considering the same procedure in $\text{C}_2$ by using (\ref{eq:eqLT})-(\ref{eq:eqHT}), STP in (\ref{eq:eqSTPdef}) is given by
\begin{align}\label{eq:eqSTPdef1}
&\Pr(\text{ST})=\frac{1}{2}(\eta_1\gamma_1+\eta_2\gamma_2),\nonumber\\&
\eta_1=\Pr\left(\log\left(1+\frac{(1-\alpha^2)Pg_1^\text{HT}}{1+\alpha^2Pg_1^\text{HT}}+Pg_1^\text{LT}\right)\ge \tilde R_1\right),
\nonumber\\&\gamma_1=\Pr\left(\log\left(1+Pg_1^\text{LT}\right)+\log\left(1+\alpha^2Pg_1^\text{HT}\right)\ge 2R\right)\nonumber\\&
\eta_2=\Pr\left(\log\left(1+\frac{\alpha^2Pg_2^\text{HT}}{1+(1-\alpha^2)Pg_2^\text{HT}}+Pg_2^\text{LT}\right)\ge R_2\right),\nonumber\\
&\gamma_2=\Pr\left(\log\left(1+Pg_2^\text{LT}\right)+\log\left(1+(1-\alpha^2)Pg_2^\text{HT}\right)\ge 2\tilde R\right).
\end{align}
Here, $\eta_1$ (resp. $\eta_2$) is the probability of successful decoding of $\tilde X_1(t)$ (resp. $X_2(t)$) at $\text{C}_1$ (resp. $\text{C}_2$) using MRC. Then, $\gamma_1$ and $\gamma_2$ give the probability that, removing the interference from the received signal in HT period, the caches can decode their intended signals correctly. Note that in $\gamma_1$ and $\gamma_2$ we have used the results on the maximum achievable rates of parallel Gaussian channels.
Considering Rayleigh-fading conditions, we have
\begin{align}\label{eq:eqeta11}
&\eta_1=1-\Pr\left(\frac{(1-\alpha^2)Pg_1^\text{HT}}{1+\alpha^2Pg_1^\text{HT}}+Pg_1^\text{LT}\le e^{\tilde R_1}-1\right)=
\nonumber\\&
1-\int_0^{\frac{e^{\tilde R_1}-1}{P}}f_{g_1}(y)\Pr\left(\frac{(1-\alpha^2)Pg_1^\text{HT}}{1+\alpha^2Pg_1^\text{HT}} \le {e^{\tilde R_1}-1}-Py \right)\text{d}y
\nonumber\\&
=e^{-\frac{\lambda_1\left({e^{\tilde R_1}-1}\right)}{P}}+
\int_{\frac{\alpha^2 e^{\tilde R_1}-1}{\alpha^2P}}^{\frac{e^{\tilde R_1}-1}{P}}e^{\lambda_1\left(y+\frac{{e^{\tilde R_1}-1}-Py}{\alpha^2P-(1-\alpha^2)P\left({e^{\tilde R_1}-1}-Py\right)}\right)}\text{d}y,
\end{align}
which can be calculated numerically. Also, following the same procedure, we have
\begin{align}\label{eq:Eqeta22}
&\eta_2=e^{-\frac{\lambda_2\left(e^{R_2}-1\right)}{P}}
+\int_{\frac{\left(1-\alpha^2\right)e^{R_2}-1}{\left(1-\alpha^2\right)P}}^{\frac{e^{R_2}-1}{P}}{\lambda_2e^{-\lambda_2\left(y+\frac{e^{R_2}-1-Py}{\alpha^2P-\left(1-\alpha^2\right)P\left(e^{R_2}-1-Py\right)}\right)}}\text{d}y.
\end{align}
The terms $\gamma_i,i=1,2,$ on the other hand, do not have closed-form or easy-to-deal integration expressions. Thus, we use the Jensen's inequality $\frac{1}{n}\sum_{i=1}^n\log(1+x_i)\le \log\left(1+\frac{1}{n}\sum_{i=1}^nx_i\right)$ \cite[Eq. (30)]{7445896} and $f_{g_i}(u)=\lambda_ie^{-\lambda_iu}, i=1,2,$ to rephrase $\gamma_i,i=1,2,$ as
\vspace{-2mm}
\begin{align}\label{eq:eqgamma1}
\gamma_1&\le 1-\Pr\left(g_1^\text{LT}+\alpha^2g_1^\text{HT}\le \frac{2}{P}\left(e^R-1\right)\right) \nonumber\\&
=1-\int_0^{\frac{2\left(e^R-1\right)}{P}}{f_{g_1}(x)\Pr\left(g_1^\text{HT}\le\frac{\frac{2\left(e^R-1\right)}{P}-x}{\alpha^2}\right)}\text{d}x\nonumber\\&
=e^{\frac{-2\lambda_1\theta}{P}}+\frac{\alpha^2 e^{\frac{-2\lambda_1\theta}{P\alpha^2}}}{\alpha^2-1}\left(1-e^{\frac{-2\lambda_1\left(\alpha^2-1\right)\theta}{P\alpha^2}}\right),
\end{align}
\vspace{-2mm}
\begin{align}\label{eq:eqgamma2}
\gamma_2&\le e^{\frac{-2\lambda_2\theta}{P}}+\frac{\left(\alpha^2-1\right) e^{\frac{-2\lambda_2\theta}{P\left(1-\alpha^2\right)}}}{\alpha^2}\left(1-e^{\frac{-2\lambda_2\alpha^2\theta}{P\left(\alpha^2-1\right)}}\right),
\end{align}
where (\ref{eq:eqgamma2}) follows the same procedure as in (\ref{eq:eqgamma1}). In this way, the optimal rate/power allocation maximizing STP is given by
\begin{align}\label{eq:eqoptprob}
&\max \frac{1}{2}\{\eta_1\gamma_1+\eta_2\gamma_2\}\nonumber\\
&\text{s.t. } \,\alpha\in[0,1], R_1+R_2=2R, \,\tilde R_1+\tilde R_2=2\tilde R,
\end{align}
which can be effectively solved by, e.g., exhaustive search or the machine-learning based scheme of \cite{8520925}.
\vspace{-2mm}
\subsection{Joint Decoding at HT Period without SIC}
Implementation of MRC and SIC, to decode and remove the interference, increases the decoding complexity/delay. Also, SIC suffers from error propagation problem, e.g., \cite{nomaharq}. For these reasons, in Method 2, each cache node decodes its intended packets in one-shot by considering the interference as an additive noise. Here, (\ref{eq:eqSTPdef1}) is rephrased as
\begin{align}\label{eq:eqSTPdef3}
&\Pr(\text{ST})=\frac{1}{2}(\bar \gamma_1+\bar \gamma_2),\nonumber\\&
\bar\gamma_1=\Pr\left(\log\left(1+Pg_1^\text{LT}\right)+\log\left(1+\frac{\alpha^2Pg_1^\text{HT}}{1+(1-\alpha^2)Pg_1^\text{HT}}\right)\ge 2R\right)\nonumber\\&
\bar\gamma_2=\Pr\left(\log\left(1+Pg_2^\text{LT}\right)+\log\left(1+\frac{(1-\alpha^2)Pg_2^\text{HT}}{1+\alpha^2Pg_2^\text{HT}}\right)\ge 2\tilde R\right),
\end{align}
where, using the Jensen's inequality and the same procedure as in (\ref{eq:eqgamma1}), we have
\begin{align}\label{eq:eqjensen2}
&\bar\gamma_1\le e^{-\frac{2\lambda_1\left(e^R-1\right)}{P}}+\int_{\frac{2\left(e^R-1\right)-\frac{\alpha^2}{1-\alpha^2}}{P}}^{\frac{2\left(e^R-1\right)}{P}}\lambda_1e^{-\lambda_1\left(x+\frac{2\left(e^R-1\right)-Px}{\alpha^2P+(1-\alpha^2)P(Px-2\left(e^R-1\right))}\right)}\text{d}x,\nonumber\\&
\bar\gamma_2\le e^{-\frac{2\lambda_2\left(e^{\tilde R}-1\right)}{P}}
+\int_{\frac{2\left(e^{\tilde R}-1\right)-\frac{1-\alpha^2}{\alpha^2}}{P}}^{\frac{2\left(e^{\tilde R}-1\right)}{P}}\lambda_2e^{-\lambda_2\left(x+\frac{2\left(e^{\tilde R}-1\right)-Px}{(1-\alpha^2)P+\alpha^2P(Px-2\left(e^{\tilde R}-1\right))}\right)}\text{d}x,
\end{align}
which can be calculated numerically.
Finally, note that, replacing (\ref{eq:eqSTPdef3}) into (\ref{eq:eqoptprob}), the optimal performance of the cache nodes in Method 2 is independent of the rate split between the sub-packets. This, although Method 1 gives the best performance in terms of the worst-case peak traffic,
may give an advantage to Method 2, compared to Method 1. This is because in Method 1 the rate split is performed by considering the worst-case condition with the cache nodes requesting for different signals during HT period. However, if the caches request for the same signals during HT period, the rate split scheme of Method 1 is not necessarily optimal. As opposed, in Method 2, the rate split is independent of the caches requested signals in HT periods.
\subsection{Separate Decoding using SIC}
In Methods 1-2, one needs to follow the coding schemes of incremental redundancy hybrid automatic repeat request (HARQ)-based protocols or Raptor codes, e.g., \cite{6164088}, where the message is decoded by concatenating different sub-packets. Alternatively, in Method 3, we consider the case where, while MRC and SIC are used to decode and remove the interference signal, respectively, each cache node decodes its sub-packets of interest separately. That is, considering $\text{C}_1,$ $X_1$ (resp. $X_2$) is decoded during the LT (resp. HT) period. In this case, the STP (\ref{eq:eqSTPdef1}) is changed to
\vspace{-1mm}
\begin{align}\label{eq:eqSTPdef2}
&\Pr\left(\text{ST}\right)=\frac{1}{2}\left(\eta_1\breve{\gamma}_{11}\breve{\gamma}_{12}+\eta_2\breve{\gamma}_{21}\breve{\gamma}_{22}\right),\nonumber\\&
\breve{\gamma}_{11}=\Pr\left(\log\left(1+Pg_1^\text{LT}\right)\ge R_1\right)=e^{-\frac{\lambda_1\left(e^{R_1}-1\right)}{P}},\nonumber\\&
\breve{\gamma}_{12}=\Pr\left(\log\left(1+\alpha^2Pg_1^\text{HT}\right)\ge R_2\right)=e^{-\frac{\lambda_1\left(e^{R_2}-1\right)}{\alpha^2P}}
\nonumber\\&
\breve{\gamma}_{21}=\Pr\left(\log\left(1+(1-\alpha^2)Pg_2^\text{HT}\right)\ge \tilde R_1\right)=e^{-\frac{\lambda_2\left(e^{\tilde R_1}-1\right)}{(1-\alpha^2)P}},\nonumber\\&
\breve{\gamma}_{22}=\Pr\left(\log\left(1+Pg_2^\text{LT}\right)\ge \tilde R_2\right)=e^{-\frac{\lambda_2\left(e^{\tilde R_2}-1\right)}{P}},
\end{align}
with $\eta_i,i=1,2,$ given in (\ref{eq:eqSTPdef1}), and (\ref{eq:eqoptprob}) is adapted correspondingly. In (\ref{eq:eqSTPdef2}), $\breve{\gamma}_{11}$ is the probability that $\text{C}_1$ decodes $X_1$ during the LT period. Also, $\breve{\gamma}_{12}$ gives the probability that, after decoding and removing $\tilde X_1$, the cache node $\text{C}_1$ correctly decodes $X_2$ in the HT period. Also, the same arguments hold for $\breve{\gamma}_{2i},i=1,2.$
Note that, although Method 1 maximizes the achievable rate/STP, Method 3 has a number of advantages including:
\begin{itemize}
\item \textbf{Low decoding complexity}: Because, as opposed to Methods 1-2 decoding long codewords of length $2L$, Method 3 is based on decoding sub-packets of length $L$.
\item \textbf{Efficient HARQ-based transmissions}: In Methods 1-2, all packets are decoded during the HT periods and, in case of decoding failure, the message is retransmitted at that period. Such HARQ-based retransmissions increase the backhauling load at HT period. As opposed, in Method 3, the decoding of the first received sub-packets and all their required HARQ-based retransmissions are performed during the LT period, which reduces the backhauling cost of HARQ.
\end{itemize}
Finally, depending on the considered method, the buffering scheme of the caches during the LT period may change. Particularly, in Methods 1-2 the caches buffer the signals received during LT period without decoding. In Method 3, however, the caches buffer the sub-packets successfully decoded during LT period.
\vspace{-1mm}
\subsection{Separate Decoding without SIC}
To further reduce the complexity of Method 3, one can consider the case where, while decoding the sub-packets separately, the cache nodes consider the interference as an additive noise (Method 4). In this case, where the sub-packets are decoded in different LT and HT periods without SIC, the STP is given by
\begin{align}\label{eq:eqSTPdef4}
&\Pr\left(\text{ST}\right)=\frac{1}{2}(\breve{\gamma}_{11}\hat{\gamma}_{12}+\hat{\gamma}_{21}\breve{\gamma}_{22}),\nonumber\\&
\hat{\gamma}_{12}=\Pr\left(\log\left(1+\frac{\alpha^2Pg_1^\text{HT}}{1+(1-\alpha^2)Pg_1^\text{HT}}\right)\ge R_2\right)\nonumber\\&\,\,\,\,\,\,\,\,=\left\{\begin{matrix}
0\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\, & \text{if } R_2\ge -\log\left(1-\alpha^2 \right )\\
e^{-\lambda_1\frac{\left(e^{R_2}-1 \right )}{\alpha^2P-(1-\alpha^2)P\left(e^{R_2}-1 \right )}} & \text{otherwise}
\end{matrix}\right.
\nonumber\\&
\hat{\gamma}_{21}=\Pr\left(\log\left(1+\frac{(1-\alpha^2)Pg_2^\text{HT}}{1+\alpha^2Pg_2^\text{HT}}\right)\ge \tilde R_1\right)\nonumber\\&=\left\{\begin{matrix}
0\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\, & \text{if } \tilde R_1\ge -2\log\alpha\\
e^{-\lambda_2\frac{\left(e^{\tilde R_1}-1 \right )}{(1-\alpha^2)P-\alpha^2P\left(e^{\tilde R_1}-1 \right )}} & \text{otherwise}
\end{matrix}\right.
\end{align}
with $\breve{\gamma}_{11}$ and $\breve{\gamma}_{22}$ given in (\ref{eq:eqSTPdef2}). Also, in (\ref{eq:eqSTPdef4}) we use Rayleigh channel PDFs and some manipulations to derive the probabilities. For further comparisons between Methods 1-4, see Section IV.
\section{Simulation Results}
The simulation results are presented for the cases with $\lambda_1=1$ and $\lambda_2=0.1$, i.e., with $10$ dB difference between the channel gains of the server-cache links, and we define the transmission signal-to-noise ratio (SNR) as $10\log_{10} P$, considering the additive noises to be unit-variance. Note that we have evaluated the results for different parameter settings, and they show the same qualitative conclusions as those presented in the following. In Figs. 3-4, the results are obtained by optimizing the rate and power allocation. Here, both exhaustive search and the genetic-algorithm based scheme of \cite{8520925} have been used which have ended up in the same results, indicating the accuracy of the optimization process. In Fig. 5, we study the effect of rate/power allocation.
\begin{figure}
\vspace{-3mm}
\centering
\includegraphics[width=0.6\columnwidth]{FigCodedcaching1-eps-converted-to.pdf}\\\vspace{-3mm}
\caption{Comparison between the STP of Methods 1-4, $\lambda_1=1,$ $\lambda_2=0.1,$ $R=\tilde R=1$ npcu. The results are obtained by optimal rate and power allocation.}
\label{fig:fig_ivd9}
\vspace{-2mm}
\end{figure}
Considering $R=\tilde R=1$ nats-per-channel-use (npcu), Fig. 3 compares the performance of Methods 1-4. Also, the figure verifies the tightness of the Jensen's inequality-based approximation results of (\ref{eq:eqgamma1})-(\ref{eq:eqgamma2}). Then, Fig. 4 shows the STP versus the data rates $R=\tilde R$ for the cases with different decoding/buffering methods and transmission SNRs.
Considering Methods 1 and 3, with joint and separate decoding on the sub-packets, respectively, Fig. 5 studies the effect of optimal rate and power allocation on the network STP. Particularly, the figure compares the optimal results obtained by (\ref{eq:eqoptprob}) with the cases using uniform power allocation, i.e., with $\alpha=\frac{\sqrt{2}}{2}$ in (\ref{eq:eqHT}), and/or equal rate split, i.e., $R_i=\tilde R_i, i=1,2$. According to the figures, the following conclusions can be drawn:
\begin{figure}
\vspace{-3mm}
\centering
\includegraphics[width=0.6\columnwidth]{FigCodedcachingrate-eps-converted-to.pdf}\\\vspace{-2mm}
\caption{Achievable STP versus the data rate for different methods. The results are obtained by optimal rate and power allocation and $\lambda_1=1, \lambda_2=0.1.$}
\label{fig:fig_ivd10}
\vspace{-0mm}
\end{figure}
\begin{figure}
\vspace{-3mm}
\centering
\includegraphics[width=0.6\columnwidth]{plottingcomparison-eps-converted-to.pdf}\\\vspace{-2mm}
\caption{On the effect of optimal rate and power allocation in Methods 1 and 3, $R=\tilde R=1$ npcu and $\lambda_1=1, \lambda_2=0.1$.}
\label{fig:fig_ivd11}
\vspace{-0mm}
\end{figure}
\begin{itemize}
\item The approximation results of (\ref{eq:eqgamma1})-(\ref{eq:eqgamma2}) properly approximate the probabilities $\gamma_i, i=1,2$ (Fig. 3. Also, the same point is observed for $\bar \gamma_i,i=1,2,$ in (\ref{eq:eqjensen2}) although not shown in the figure). Thus, the approximations can be well utilized for the performance evaluation of Methods 1-2, i.e., in the cases with joint decoding of the sub-packets.
\item Compared to the cases with separate decoding of sub-packets, i.e., Methods 3-4, considerable STP improvement is observed by joint decoding of the sub-packet, i.e., Methods 1-2 (Figs. 3-4). However, as explained in Section III.C, the STP increment of Methods 1-2 is at the cost of decoding complexity and possible HARQ-based retransmissions at HT periods. On the other hand, for both cases with joint and separate decoding of the sub-packets, using SIC-based interference cancellation leads to marginal performance improvement at low rates while its effect increases with the data rate (Fig. 4).
Finally, as the data rate increases, the performance gap between Methods 2 and 3 decreases, i.e., one can reach the same STP as in the cases with joint sub-packet decoding of interference-affected signals by separate sub-packets decoding if the interference signals are removed using SIC. (Fig. 4).
\item For all parameter settings, Method 1 leads to the highest STP, compared to Methods 2-4, at the cost of decoding delay/possible retransmissions at HT periods (Figs. 3-4). For instance, with the parameter settings of Fig. 4 and data rate 1.5 npcu, the implementation of Method 1 with transmit SNR 10 dB results in the same STP, $80\%$, as in the cases with Method 4 and SNR 15 dB, i.e., using advanced decoding methods leads to 5 dB gain in SNR (Fig. 4).
\item For both cases with and without interference cancellation (Methods 1 and 3), optimal rate allocation leads to considerable STP increment (Fig. 5. The same conclusion is observed in Methods 2 and 4, although not presented in the figure). Also, the relative performance gain of optimal rate split increases in the cases with interference cancellation. Finally, considering the interference as additive noise, optimal power allocation between sub-packets during HT period increases the STP. However, with interference cancellation and joint decoding of sub-packets, the effect of optimal power allocation between HT period sub-packets is marginal (Fig. 5).
\end{itemize}
\section{Conclusions}
This paper studied the performance coded-caching networks in the cases with adaptive rate/power allocation and different decoding/buffering schemes. As we showed, joint decoding of the sub-packets at HT periods leads to considerable performance improvement of coded-caching setups. Also, for different decoding schemes, optimal rate split between the sub-packets increases the STP considerably while optimal power allocation between the sub-packets of HT period only improves the STP if SIC-based receiver is not implemented and the sub-packets are decoded separately.
\vspace{-2mm}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Let $S$ be a scheme. Let $\mathfrak{G}$ be an
$S$--group scheme. It is natural to ask
whether $\mathfrak{G}$ is linear; that is, there exists a group monomorphism
$\mathfrak{G} \to \GL(\mathcal E)$ where
$\mathcal E$ is a locally free $\mathcal{O}_S$-module of finite rank. In particular, $\mathfrak{G}$ admits a faithful representation on $\mathcal E$.
This holds for affine algebraic groups over a field
\cite[II, \S 2.3.3]{DG}.
In the case $S$ is locally noetherian and $G$ is of multiplicative type
of constant type and of finite type, Grothendieck has shown that
$G$ is linear if and if $G$ is isotrivial, i.e. $G$ is split by a finite \'etale extension of $S$ \cite[XI.4.6]{SGA3}. In particular there exist non linear tori of rank $2$ over the local ring (at a node) of a nodal
algebraic curve ({\it ibid}, X.1.6).
Firstly we extend that criterion over an arbitrary base by using
Azumaya and separable $\mathcal{O}_S$--algebras (Theorem \ref{thm_torus}).
Secondly we deal with the case $\mathfrak{G}$ reductive; that
is, $\mathfrak{G}$ is smooth affine with reductive (connected) geometric fibers.
In this case a faithful representation is necessarily a closed immersion \cite[XVI.1.5]{SGA3}.
Positive results on the linearity question are due to
M. Raynaud \cite[VI$_B$]{SGA3} and R. Thomason \cite[3.1]{T}
which is essentially the implication $(i) \Longrightarrow (ii)$
in the theorem below.
We can restrict our attention to the case
when $\mathfrak{G}$ is of constant type (recall that the type is a locally
constant function on $S$);
this implies that there exists a Chevalley $\mathbb{Z}$--group scheme $G$
such that $\mathfrak{G}$ is locally isomorphic to $G_S$ for the \'etale topology
\cite[XXII.2.3, 2.5]{SGA3}.
A short version of our main result is the following.
\begin{stheorem}\label{thm_main}
The following are equivalent:
\smallskip
(i) The radical torus $\mathrm{rad}(\mathfrak{G})$ is isotrivial;
\smallskip
(ii) $\mathfrak{G}$ is linear.
\smallskip
\noindent Furthermore if $S$ is affine, the above are equivalent to
\smallskip
(ii') there exists a
closed immersion $i: \mathfrak{G} \to \GL_n$ with $n \geq 1$
which is a homomorphism.
\end{stheorem}
We recall that $\mathrm{rad}(\mathfrak{G})$ is the maximal central subtorus of $\mathfrak{G}$ \cite[XXIV.4.3.6]{SGA3} and that (i)
means that $\mathrm{rad}(\mathfrak{G})$ splits
after passing to a finite \'etale cover $S'\to S$.
In the noetherian setting, a variant of the implication $(i) \Longrightarrow (ii)$
has been shown by Margaux who furthermore provided an $\text{\rm{Aut}}(\mathfrak{G})$-equivariant representation \cite{M}.
Note that condition (i) depends only on the quasi-split form of
$\mathfrak{G}$ and also that it is always satisfied
if $\mathfrak{G}$ is semisimple or if $\mathrm{rad}(\mathfrak{G})$ is of rank one.
Furthermore, if $S$ is a semilocal scheme,
Demazure's characterization of isotrivial group schemes \cite[XXIV.3.5]{SGA3}
permits us to deduce that
$\mathfrak{G}$ is isotrivial if and only if $\mathfrak{G}$ is linear,
see section \ref{section_semilocal}.
Finally for $S=\mathop{\rm Spec}\nolimits(R)$ with $R$ noetherian, we complete Thomason's approach
by showing that linearity for $\mathfrak{G}$ is equivalent to the
resolution property (Th. \ref{thm_re}).
\medskip
\smallskip
\noindent{\bf Acknowledgements.} I thank Vladimir Chernousov, Laurent Moret-Bailly, Erhard Neher, Arturo Pianzola,
and Anastasia Stavrova for their valuable suggestions.
I thank the referee for a simplification of the proof of Proposition \ref{prop_torus}.
\section{Definitions and basic facts}
\subsection{Notation}
We use mainly the terminology and notation of Grothendieck-Dieudonn\'e \cite[\S 9.4 and 9.6]{EGA1}
which agrees with that of Demazure-Grothendieck used in \cite[Exp. I.4]{SGA3}.
\smallskip
\noindent (a) Let $S$ be a scheme and let $\mathcal E$ be a quasi-coherent sheaf over $S$.
For each morphism $f:T \to S$,
we denote by $\mathcal E_{T}=f^*(\mathcal E)$ the inverse image of $\mathcal E$
by the morphism $f$.
We denote by $\mathbf{V}(\mathcal E)$ the affine $S$--scheme defined by
$\mathbf{V}(\mathcal E)=\mathop{\rm Spec}\nolimits\bigl( \mathrm{Sym}^\bullet(\mathcal E)\bigr)$;
it represents the $S$--functor $Y \mapsto \mathop{\rm Hom}\nolimits_{\mathcal{O}_Y}(\mathcal E_{Y}, \mathcal{O}_Y)$
\cite[9.4.9]{EGA1}.
\smallskip
\noindent (b) We assume now that $\mathcal E$ is locally free of finite rank and denote by $\mathcal E^\vee$ its dual.
In this case the affine $S$--scheme $\mathbf{V}(\mathcal E)$ is of finite presentation
(ibid, 9.4.11); also
the $S$--functor $Y \mapsto H^0(Y, \mathcal E_{Y})=
\mathop{\rm Hom}\nolimits_{\mathcal{O}_Y}(\mathcal{O}_Y, \mathcal E_{Y} )$
is representable by the affine $S$--scheme $\mathbf{V}(\mathcal E^\vee)$
which is also denoted by $\mathbf{W}(\mathcal E)$ \cite[I.4.6]{SGA3}.
The above applies to the locally free quasi-coherent sheaf
${\mathcal E}nd(\mathcal E) = \mathcal E^\vee \otimes_{\mathcal{O}_S} \mathcal E$
over $S$ so that we can consider
the affine $S$--scheme $\mathbf{V}\bigl({\mathcal E}nd(\mathcal E)\bigr)$
which is an $S$--functor in associative and unital algebras
\cite[9.6.2]{EGA1}.
Now we consider the $S$--functor $Y \mapsto \text{\rm{Aut}}_{\mathcal{O}_Y}(\mathcal E_{Y})$.
It is representable by an open $S$--subscheme of $\mathbf{V}\bigl({\mathcal E}nd(\mathcal E)\bigr)$
which is denoted by $\GL(\mathcal E)$ ({\it loc. cit.}, 9.6.4).
\smallskip
\noindent (c) If $\mathcal B$ is a locally free $\mathcal{O}_S$--algebra (unital, associative) of finite rank,
we recall that the functor of invertible elements of $\mathcal B$ is representable
by an affine $S$-group scheme which is denoted by $\GL_1(\mathcal B)$ \cite[2.4.2.1]{CF}.
For separable and Azumaya algebras, we refer to \cite{KO}.
Note that in \cite[\S 2.5.1]{CF}, separable algebras are supposed furthermore to be locally free of finite rank.
If $\mathcal B$ is a separable $\mathcal{O}_S$--algebra which is a
locally free $\mathcal{O}_S$--algebra of finite rank,
then $\GL_1(\mathcal B)$ it is a reductive $S$--group scheme \cite[3.1.0.50]{CF}.
\smallskip
\noindent (d) We use the theory and terminology of tori and multiplicative group schemes of \cite{SGA3}; see also Oesterl\'e's survey \cite{O}.
\subsection{Finite \'etale covers}
The next lemma is a consequence of the equivalence of categories describing
finite \'etale $\mathcal{O}_S$--algebra of rank $N$
\cite[\S 2.5.2]{CF}; it admits a simple direct proof.
\begin{slemma} \label{lem_brian} Let $N$ be a positive integer and let
$\mathcal C$ be a finite \'etale $\mathcal{O}_S$--algebra of rank $N$. Then
there exists a finite \'etale cover $T$ of $S$
of degree $N!$ such that
$\mathcal C \otimes_{\mathcal{O}_S} {\mathcal{O}_T} \buildrel\sim\over\lgr (\mathcal{O}_T)^N $.
\end{slemma}
\begin{proof}
We proceed by induction on $N$, the case $N=1$ being obvious.
We put $S'= \mathop{\rm Spec}\nolimits(\mathcal{O}_C)$,
this is a finite \'etale cover of $S$ of degree $N$.
Since the diagonal map $S' \to S' \times_S S'$ is closed and
open \cite[$_4$.17.4.2]{EGA4}
there exists a decomposition $\mathcal C \otimes_{\mathcal{O}_S} \mathcal{O}_S'=\mathcal{O}_{S'} \times \mathcal C'$ where $\mathcal C'$ is a finite \'etale $\mathcal{O}_{S'}$--algebra of rank $N-1$.
Applying the induction process to $\mathcal C'$ provides
a finite \'etale cover $T$ of $S'$
of degree $(N-1)!$ such that
$\mathcal C' \otimes_{\mathcal{O}_{S'}} {\mathcal{O}_T} \cong (\mathcal{O}_T)^{N-1}$.
Thus $\mathcal C \otimes_{\mathcal{O}_S} \mathcal{O}_T=\mathcal{O}_{T} \times (\mathcal{O}_T)^{N-1}$
and $T$ is a finite \'etale cover of $S$ of degree $N!= N \times (N-1)!$.
\end{proof}
\subsection{Isotriviality} \cite[XXIV.4]{SGA3}
Let $\mathcal H$ be a fppf $S$--sheaf in groups and let
$\mathcal X$ be a $\mathcal H$--torsor. We say that $\mathcal X$ is {\it isotrivial}
if there exists a finite \'etale cover $S'$ of $S$
which trivializes $\mathcal X$; that is, satisfying $\mathcal X(S') \not = \emptyset$.
The notion of locally isotrivial (with respect to the Zariski topology) is then clear and
there is also the following variant of {\it semilocally isotrivial}.
We say that $\mathcal X$ is {\it semilocally isotrivial}
if for each subset $\{s_1,\dots, s_n\}$ of points of $S$
contained in an affine open subset of $S$, there exists
an open subscheme $U$ of $S$ containing $s_1, \dots, s_n$
such that $\mathcal X \times_S U$ is isotrivial over $U$.
\medskip
A reductive $S$--group scheme $\mathfrak{G}$ is {\it isotrivial} if it is split by
a finite \'etale cover $S'$ of $S$.
An isotrivial reductive $S$--group scheme $\mathfrak{G}$ is necessarily of constant type.
If $\mathfrak{G}$ is of constant type with underlying Chevalley group scheme $G$,
$\mathfrak{G}$ is isotrivial if and only if
the $\text{\rm{Aut}}(G)$--torsor $\mathrm{Isom}( G_S, \mathfrak{G})$ is isotrivial.
\subsection{Rank one tori}\label{subsec-rank1}
The simplest case is that of $G=\mathbb{G}_{m,S}$, the split
$S$--torus of rank $1$.
The $S$--functor $S' \mapsto \mathop{\rm Hom}\nolimits_{S'-gp}( \mathbb{G}_{m,S'}, \mathbb{G}_{m,S'})$
is representable by the constant $S$--group scheme $\mathbb{Z}_S$
\cite[VIII.1.5]{SGA3}. It follows that
the $S$--functor $S' \mapsto \mathop{\rm Isom}\nolimits_{S'-gp}( \mathbb{G}_{m,S'}, \mathbb{G}_{m,S'})$
is representable by the constant $S$--group scheme $(\mathbb{Z}/2\mathbb{Z})_S
=\text{\rm{Aut}}_{S-gp}(\mathbb{Z}_S)$.
On the other hand, $(\mathbb{Z}/2\mathbb{Z})_S$ is the automorphism group
of the split \'etale cover $S \sqcup S \to S$ of degree $2$.
By definition an $S$--torus of rank one is a form
of $\mathbb{G}_m$ for the fpqc topology;
in the other hand, a degree $2$ \'etale cover
is a form of $S \sqcup S$ for the finite \'etale topology
(see for example Lemma \ref{lem_brian}) and a fortiori
for the fpqc topology.
According to the faithfully flat descent technique (e.g. \cite[XXIV.1.17]{SGA3}),
there is then an equivalence of categories between
the groupoid of rank one tori over $S$ (resp.\ the groupoid of degree $2$ \'etale covers of $S$) and
the groupoid of $(\mathbb{Z}/2\mathbb{Z})_S$-torsors.
More precisely one associates to an $S$-torus $T$ of rank $1$
(resp.\ to an \'etale cover $E$ of degree $2$) the
$(\mathbb{Z}/2\mathbb{Z})_S$-torsor $\mathrm{Isom}_{gr}( \mathbb{G}_m, T)$
(resp. $\mathrm{Isom}( S \sqcup S , E)$).
The inverse map is given by twisting the split object
by a given $(\mathbb{Z}/2\mathbb{Z})_S$-torsor.
Let $T$ be an $S$--torus of rank one and let $S'\to T$ be its
associated \'etale cover of degree $2$.
Then $T$ is splits after base change to $S'$ and so is
isotrivial. Furthermore we claim that $T$ is isomorphic to the quotient
$Q=R_{S'/S}(\mathbb{G}_{m,S'})/ \mathbb{G}_m$ where $R_{S'/S}(\mathbb{G}_{m,S'})$
stands for the Weil restriction.
We consider the $(\mathbb{Z}/2\mathbb{Z})_S$--equivariant exact sequences
$$
1 \to \mathbb{G}_{m,S} \xrightarrow{\Delta} \mathbb{G}_{m,S} \times \mathbb{G}_{m,S}
\xrightarrow{\Pi} \mathbb{G}_{m,S} \to 1
$$
where $\Pi(x,y)=x \, y^{-1}$ and the $\mathbb{Z}/2\mathbb{Z}$-action on $\mathbb{G}_{m,S} \times \mathbb{G}_{m,S}$ is by $(x,y) \mapsto (y,x)$.
The $\mathbb{Z}/2\mathbb{Z}$-action on the last factor is then by $x \mapsto x^{-1}$.
Twisting this sequence by the $(\mathbb{Z}/2\mathbb{Z})_S$-torsor $\mathrm{Isom}( S \sqcup S , S')$ yields the desired exact sequence
$$
1 \to \mathbb{G}_m \to R_{S'/S}(\mathbb{G}_{m,S'}) \to Q \to 1
$$
where the identification of the second term is left to the reader
(it is similar to that of the proof of \cite[XXIX.3.13]{SGA3}).
\medskip
\subsection{Linear representations}
Let $\mathfrak{G}$ be an $S$--group scheme. We say that $\mathfrak{G}$ is {\it linear} if there exists
a locally free $\mathcal{O}_S$-module $\mathcal E$ is of finite rank
and a group homomorphism $\mathfrak{G} \to \GL(\mathcal E)$ which is a monomorphism.
The notion of locally linear $S$--group scheme is then clear and
there is also the following variant of {\it semilocally linear}.
We say that $\mathfrak{G}$ is {\it semilocally linear}
if for each subset $\{ s_1,\dots, s_n\}$ of points of $S$
contained in an affine open subset of $S$, there exists
an open subscheme $U$ of $S$ containing $s_1, \dots, s_n$
such that $\mathfrak{G} \times_S U$ is linear over $U$.
\begin{slemma}\label{lem_affine}
Assume that $S=\mathop{\rm Spec}\nolimits(R)$ and let $\mathcal E$ be a locally free $\mathcal{O}_S$--module of finite rank.
Then $\GL(\mathcal E)$ embeds as a closed $S$--subgroup scheme in $\GL_n$ for some $n \geq 1$.
\end{slemma}
\begin{proof} Since the rank of $\mathcal E$ is a locally constant function \cite[Ch. \, 0, 5.4.1]{EGA1},
we can assume that $\mathcal E$ is locally free of constant of rank $r$.
Then $E=H^0(R, \mathcal E)$ is a locally free $R$--module of rank $r$,
so is finitely generated projective \cite[Tag 00NX]{St}. It follows that there exists an integer
$n \geq 1$ and a decomposition $R^n =E \oplus E'$.
The homomorphism $\GL(\mathcal E) \to \GL_n$ is a closed immersion.
\end{proof}
\begin{slemma}\label{lem_weil}
Let $\mathfrak{G}$ be an $S$--group scheme
and let $S'$ be a finite locally free cover of $S$.
Then $\mathfrak{G}$ is linear if and only if $\mathfrak{G} \times_S S'$
is linear.
\end{slemma}
\begin{proof} We denote by $p: S' \to S$ the structure map.
If $\mathfrak{G}$ is linear, then $\mathfrak{G} \times_S S'$ is linear.
Conversely we assume that there exists a monomorphism
$i: \mathfrak{G} \times_S S' \to \GL(\mathcal E')$ where
$\mathcal E'$ is a locally free $\mathcal{O}_{S'}$--module of finite rank. We put $\mathcal E= p_*( \mathcal E')$, this is
a locally free $\mathcal{O}_S$--module of finite rank.
We consider the sequence of $S$--functors in $S$--groups
$$
\mathfrak{G} \to R_{S'/S}( \mathfrak{G} \times_S S') \xrightarrow{R_{S'/S}(i)} R_{S'/S}( \GL(\mathcal E') ) \to \GL( \mathcal E)
$$
where $R_{S'/S}$ stands for the Weil restriction
and the first map is the diagonal map which is a monomorphism.
Since the Weil restriction for $S'/S$ transforms monomorphisms into monomorphisms,
the map $R_{S'/S}(i)$ is also a monomorphism and
so is the last map since
$R_{S'/S}( \GL(\mathcal E') )(T) \subset \GL( \mathcal E)(T)$
corresponds to automorphisms of $\mathcal E \otimes_{\mathcal{O}_S} \mathcal{O}_T$
which are $\mathcal{O}_{S'} \otimes_{\mathcal{O}_S} \mathcal{O}_T$-linear.
Since all maps are monomorphisms, we conclude that $\mathfrak{G}$ is linear.
\end{proof}
\section{Tori and group of multiplicative type}\label{section_torus}
\subsection{Maximal tori of linear groups}
Let $\mathcal A$ be an Azumaya $\mathcal{O}_S$--algebra.
We consider the reductive $S$--group scheme $\GL_1(\mathcal A)$.
Let $\mathcal B \subset \mathcal A$ be a separable $\mathcal{O}_S$--subalgebra of $\mathcal A$
which is a locally free $\mathcal{O}_S$--module of finite rank and
which is locally a direct summand of $\mathcal A$ as $\mathcal{O}_S$--module.
We get a monomorphism of reductive $S$-group schemes
$\GL_1(\mathcal B) \to \GL_1(\mathcal A)$.
In particular, if $\mathcal B$ is commutative, $\mathcal B$ is a finite
\'etale $\mathcal{O}_S$--algebra and $\GL_1(\mathcal B)$ is a torus.
We come now to Grothendieck's definition of maximal \'etale subalgebras.
\begin{sdefinition} \cite[D\'ef. 5.6]{G}.
We say that a finite \'etale $\mathcal{O}_S$--subalgebra $\mathcal C \subset \mathcal A$ is
{\it maximal} if $\mathcal C$ is locally a direct summand of $\mathcal A$ as $\mathcal{O}_S$--module
and if the rank of $\mathcal C \otimes_{\mathcal{O}_S} \kappa(s)$ is the degree of $\mathcal A_s \otimes_{\mathcal{O}_S} \kappa(s)$
for each $s \in S$.
\end{sdefinition}
If $\mathcal C \subset \mathcal A$ is maximal finite \'etale $\mathcal{O}_S$--subalgebra of
$\mathcal A$, then the torus $\GL_1(\mathcal C)$ is a maximal torus of $\GL_1(\mathcal A)$
since it is the case on geometric fibers.
According to \cite[\S 7.5]{G}, all maximal $S$--tori of $\GL_1(\mathcal A)$ occur in that manner;
this is part (3) of the following enlarged statement.
\begin{sproposition} \label{prop_torus}
Let $\mathfrak{S}$ be a subgroup scheme of multiplicative type of $\GL_1(\mathcal A)$
and put $\mathcal B=\mathcal A^\mathfrak{S}$, the centralizer subalgebra of $\mathfrak{S}$.
\smallskip
\noindent (1) $\mathcal B$ is a separable $\mathcal{O}_S$--algebra which is locally a direct summand of $\mathcal A$ as $\mathcal{O}_S$--module.
\smallskip
\noindent (2) Let $\mathcal C$ be the center of $\mathcal B$; this is a
finite \'etale $\mathcal{O}_S$--algebra of positive rank which is locally a direct summand of $\mathcal B$ (and $\mathcal A$)
as $\mathcal{O}_S$--module.
We have the closed immersions
\[
\mathfrak{S} \, \subset \, \GL_1(\mathcal C) \, \subset \, \GL_1(\mathcal A).
\]
\smallskip
\noindent (3) If $\mathfrak{S}$ is a maximal torus of $\GL_1(\mathcal A)$,
then $\mathfrak{S}= \GL_1(\mathcal C)$ and $\mathcal C$ is a maximal finite \'etale $\mathcal{O}_S$--subalgebra of
$\mathcal A$.
\smallskip
\noindent (4) If $\mathfrak{S}$ is of constant type, then $\mathfrak{S}$ is isotrivial.
\end{sproposition}
\begin{proof}
According to \cite[IX.2.5]{SGA3}, the map $\mathfrak{S} \to \GL_1(\mathcal A)$ is a closed
immersion so that $\mathfrak{S}$ is of finite type over $S$.
As a preliminary observation we notice that
(1), (2), (3) are local for the \'etale topology.
We can assume that $S=\mathop{\rm Spec}\nolimits(R)$, $\mathcal A=\mathop{\rm M}\nolimits_n(R)$ and that $\mathfrak{S}=D(M)$ for
$M$ an abelian group. Note that $M$ is finitely generated since $\mathfrak{S}$ is
of finite type ({\it ibid}, VII.2.1.b).
\smallskip
\noindent (1) We consider the $M$--grading
$R^n= \bigoplus_{m \in M} R^n_m$.
The $R$--modules $(R^n_m)_{m \in M}$ are finitely generated projective
so locally free of finite rank.
There is a finite subset $M' \subset M$ such that
$R^n= \bigoplus_{m \in M'} R^n_m$.
Then $\mathcal B= \prod_{m \in M'} \mathop{\rm End}\nolimits_{R}( R^n_m)$
and each $\mathop{\rm End}\nolimits_{R}(R^n_m)$ is a separable $R$-algebra which is locally free of finite rank
\cite[III, example 2.8]{KO}.
Since a product of separable algebras is
a separable algebra ({\it ibid}, III, proposition 1.7),
it follows that $\mathcal B$ is a separable $R$-algebra.
Furthermore $\mathcal B$ is locally a direct summand of $\mathop{\rm M}\nolimits_n(R)$
as $R$-module.
\smallskip
\noindent (2) Let $\mathcal C$ be the center of $\mathcal B$.
We have $R$--monomorphisms of groups $\mathfrak{S} \, \subset \, \GL_1(\mathcal C) \, \subset \, \GL_1(\mathcal A)$.
\smallskip
\noindent (3) We assume that $\mathfrak{S}$ is a maximal torus of $\GL_1(\mathcal A)$
and want to establish that $\mathfrak{S}= \GL_1(\mathcal C)$. So $\GL_1(\mathcal C)$ is an $R$--torus containing $\goth T$ and since maximality
holds also in the naive sense \cite[XII.1.4]{SGA3},
we conclude that $\goth T=\GL_1(\mathcal C)$.
\smallskip
\noindent (4)
We come back to the initial setting (i.e. without localizing).
We want to show that $\mathfrak{S}$ is isotrivial.
According to \cite[ch.0, 5.4.1]{EGA1}, for each integer $l \geq 0$,
$S_l= {\bigl\{ s \in S \, \mid \, \mathop{\rm rank}\nolimits( \mathcal C_{\kappa(s)}) =l \bigr\}}$ is an open subset of $S$
so that we have a decomposition in clopen subschemes $S=\coprod\limits_{l \geq 0} S_l$.
Without loss of generality we can assume that $\mathcal C$ is locally free of rank $l$.
Lemma \ref{lem_brian} provides a finite \'etale cover $S'$ of $S$
such that $\mathcal C \otimes_{\mathcal{O}_S} \mathcal{O}_{S'} \cong (\mathcal{O}_{S'})^l$.
Hence $\GL_1(\mathcal C) \times_S S \cong (\mathbb{G}_m)^l \times_S S'$.
It follows that $\mathfrak{S} \times_S S'$ is a subgroup $S'$--scheme of $(\mathbb{G}_m)^l \times_S S'$
of multiplicative type.
According to \cite[IX.2.11.(i)]{SGA3} there exists a partition in clopen subsets
$S'= \sqcup_{i \in I} S'_i$ such that each $\mathfrak{S} \times_S S'_i$ is diagonalizable.
Since $\mathfrak{S} \times_S S'$ is of constant type we conclude that $\mathfrak{S} \times_S S'$
is diagonalizable.
\end{proof}
\subsection{Characterization of isotrivial groups of multiplicative type}
\begin{stheorem}\label{thm_torus} Let $\mathfrak{S}$ be an $S$-group scheme of multiplicative type of finite type and of
constant type. Then the following are equivalent:
\smallskip
(i) $\mathfrak{S}$ is isotrivial;
\smallskip
(ii) $\mathfrak{S}$ is linear;
\smallskip
(iii) $\mathfrak{S}$ is an $S$--subgroup scheme of an $S$--group scheme $\GL_1(\mathcal A)$
where $\mathcal A$ is an Azumaya $\mathcal{O}_S$--algebra.
\end{stheorem}
\begin{proof}
$(i) \Longrightarrow (ii)$. This follows from Lemma \ref{lem_weil}.
\smallskip
\noindent $(ii) \Longrightarrow (iii)$. By definition we have a monomorphism
$\mathfrak{S} \to \GL(\mathcal E)$ where $\mathcal E$ is a locally free $\mathcal{O}_{S}$--module of finite rank.
Since $\mathop{\rm End}\nolimits_{\mathcal{O}_S}(\mathcal E)$ is an Azumaya $\mathcal{O}_S$--algebra, we get (iii).
\smallskip
\noindent $(iii) \Longrightarrow (i)$. This follows of Proposition \ref{prop_torus}.(4).
\end{proof}
\begin{sexamples} \label{ex_groth} {\rm (a) Grothendieck constructed
a scheme $S$ and an $S$--torus $\goth T$ which is locally trivial of rank $2$
but which is not isotrivial \cite[\S X.1.6]{SGA3} (e.g. $S$ consists of
two copies of the projective line over a field pinched at $0$ and $\infty$).
Theorem \ref{thm_torus} shows that such an $S$--torus is not linear.
\newline
(b) Also there exists a local ring $R$ and an $R$--torus $\goth T$ of rank $2$ which is not isotrivial \cite[\S 1.6]{SGA3};
the ring $R$ can be taken as the local ring of an algebraic curve at a double point.
Theorem \ref{thm_torus} shows that such an $R$--torus is not linear.
}
\end{sexamples}
\section{Reductive case}
For stating the complete version of our main result,
we need more notation.
As in the introduction, $\mathfrak{G}$ is a reductive $S$--group
scheme of constant type and $G$ is the underlying Chevalley
$\mathbb{Z}$--group scheme.
We denote by $\text{\rm{Aut}}(G)$ the automorphism group scheme of $G$
and we have an exact sequence of $\mathbb{Z}$--group schemes \cite[th. XXIV.1.3]{SGA3}
$$
1 \to G_{ad} \to \text{\rm{Aut}}(G) \to \text{\rm{Out}}(G) \to 1.
$$
We remind the reader of the representability of the fppf sheaf
$\underline{\mathrm{Isom}}(G_S, \mathfrak{G})$ by
a $\text{\rm{Aut}}(G)_S$--torsor $\mathrm{Isom}(G_S, \mathfrak{G})$
defined in \cite[XXIV.1.8]{SGA3}. The contracted product
$$\mathrm{Isomext}(G_S, \mathfrak{G}) := \mathrm{Isom}(G_S, \mathfrak{G})
\wedge^{\text{\rm{Aut}}(G)_S} \text{\rm{Out}}(G)_S$$ is an $\text{\rm{Out}}(G)_S$--torsor ({\it ibid}, 1.10)
which encodes the isomorphism class of the quasi-split form of $\mathfrak{G}$.
\begin{stheorem}\label{thm_main_complete}
The following are equivalent:
\smallskip
(i) The torus $\mathrm{rad}(\mathfrak{G})$ is isotrivial;
\smallskip
(ii) the $\text{\rm{Out}}(G)_S$--torsor $\mathop{\rm Isomext}\nolimits(G_S, \mathfrak{G})$ is isotrivial;
\smallskip
(iii) $\mathfrak{G}$ is linear;
\smallskip
(iv) $\mathrm{rad}(\mathfrak{G})$ is linear.
\smallskip
\noindent Furthermore if $S$ is affine, we can take a faithful linear representation in some $\GL_n$ for $(iii)$ and $(iv)$.
\end{stheorem}
\begin{proof}$(i) \Longrightarrow (ii)$. We assume that $\mathrm{rad}(\mathfrak{G})$
is isotrivial and want to show that the $\text{\rm{Out}}(G)_S$--torsor $\mathcal F=\mathop{\rm Isomext}\nolimits(G_S, \mathfrak{G})$ is isotrivial.
In other words, we want to show that there exists a finite \'etale cover $S'$ of
$S$ such that $\mathfrak{G} \times_S S'$ is an inner form of $G$.
Without loss of generality we can assume that $\mathrm{rad}(\mathfrak{G})$ is a split torus.
We quote now \cite[XXIV.2.16]{SGA3} for the Chevalley group $G$ over $\mathbb{Z}$
which introduces the $\mathbb{Z}$-group scheme
$$H= \ker\bigl( \text{\rm{Aut}}(G) \to \text{\rm{Aut}}(\mathrm{rad}(G) \bigr);$$
furthermore there is an equivalence of categories
between the category of \break $H$-torsors over $S$
and the category of pairs $(\goth M, \phi)$ where $\goth M$ is an $S$--form
of $G$ and $\phi: \mathrm{rad}(G)_S \buildrel\sim\over\lgr \mathrm{rad}(\goth M)$.
Since $\mathrm{rad}(\mathfrak{G})$ is split, we choose an isomorphism $\phi: \mathrm{rad}(G)_S \buildrel\sim\over\lgr \mathrm{rad}(\mathfrak{G})$
and consider an $H$-torsor $\mathfrak{P}$ mapping to an object isomorphic to
$(\mathfrak{G},\phi)$. Furthermore the quoted reference provides an exact sequence
of $\mathbb{Z}$--group schemes
$$
1 \to G_{ad} \to H \xrightarrow{p} F \to 1
$$
where $F$ is finite \'etale over $\mathbb{Z}$, so is constant.
We denote by $S'= \goth M \wedge^{H_S} F_S$ the contracted product of $\goth M$ and $F_S$ with respect to $H_S$;
this is an $F$--torsor over $S$, hence is a finite \'etale cover of $S$.
It follows that $\mathfrak{P} \times_S S'$ admits a reduction to
a $G_{ad,S'}$--torsor $\mathfrak{Q}'$. Since the map $G_{ad} \to H \to \text{\rm{Aut}}(G)$
is the canonical map, we conclude that $\mathfrak{G}_{S'} \cong \, ^{\mathfrak{Q}'}\!\!G$ is an inner form of $G$.
\smallskip
\noindent $(ii) \Longrightarrow (iii)$. Our assumption is that there
exists a finite \'etale cover $S'/S$ which splits
the $\text{\rm{Out}}(G)$--torsor $\mathrm{Isomext}(G_S, \mathfrak{G})$.
Lemma \ref{lem_weil} permits us to replace $S$ by $S'$, so we can assume
that the $\text{\rm{Out}}(G)$--torsor $\mathrm{Isomext}(G_S, \mathfrak{G})$ is trivial;
that is, $\mathfrak{G}$ is an inner form of $G$.
There exists a $G_{ad}$--torsor $\mathfrak{Q}$ over $S$
such that $\mathfrak{G} \cong \, {^\mathfrak{Q} \! G}$.
Since $G \mathbin{{>}\!{\triangleleft}} G_{ad}$ is defined over $\mathbb{Z}$, it admits
a faithful representation $\rho: G \mathbin{{>}\!{\triangleleft}} G_{ad} \to \GL_n$ over $\mathbb{Z}$ \cite[\S 1.4.5]{BT}.
The map $\rho$ is then $G_{ad}$--equivariant and can be twisted
by the $G_{ad}$--torsor $\mathfrak{Q}$. We obtain a faithful representation
${^\mathfrak{Q} \! G} \mathbin{{>}\!{\triangleleft}} {^\mathfrak{Q} \! G_{ad}} \to {^\mathfrak{Q} \! \GL_n}= \GL(\mathcal E)$ where
$\mathcal E$ is the locally free $\mathcal{O}_S$--module of rank $n$
which is the twist of $(\mathcal{O}_S)^n$ by the $\GL_n$--torsor
$\mathfrak{Q} \wedge^{G_{ad}} \GL_n$. Thus ${^\mathfrak{Q} \! G}$ is linear.
\smallskip
\noindent $(iii) \Longrightarrow (iv)$. Obvious.
\smallskip
\noindent $(iv) \Longrightarrow (i)$.
Since $\mathrm{rad}(\mathfrak{G})$ is a form of $\mathrm{rad}(G)$, it is of constant rank
and Theorem \ref{thm_torus} shows that $\mathrm{rad}(G)$ is isotrivial.
\smallskip
Finally the refinement for $S$ affine follows from Lemma \ref{lem_affine}.
\end{proof}
\begin{scorollary}\label{cor_main}
Under the assumptions of $\mathfrak{G}$, let $\mathfrak{G}^{qs}$ be the quasi-split form of
$\mathfrak{G}$.
Then $\mathfrak{G}$ is linear if and only if $\mathfrak{G}^{qs}$ is
linear.
\end{scorollary}
The next corollary slightly generalizes a result by Thomason \cite[cor. 3.2]{T}.
\begin{scorollary}\label{cor_main2} Assume that either
\smallskip
\noindent (i) $S$ is locally noetherian and geometrically unibranch
(e.g. normal);
\smallskip
\noindent (ii) $\mathrm{rad}(\mathfrak{G})$ is of rank $\leq 1$ (in particular if $G$ is semisimple).
\smallskip
\noindent Then $\mathfrak{G}$ is linear.
\end{scorollary}
\begin{proof}
In case (i), the torus $\mathrm{rad}(\mathfrak{G})$ is isotrivial \cite[X.5.16]{SGA3}.
In case (ii), we have $\mathrm{rad}(G)=1$ or $\mathbb{G}_m$ (since $G$ is split), so
that $\mathrm{rad}(\mathfrak{G})$ is
split by a quadratic \'etale cover of $S$ (\S \ref{subsec-rank1}), hence is isotrivial.
Hence Theorem \ref{thm_main} implies that $\mathfrak{G}$ is linear.
\end{proof}
The next corollary extends
Demazure's characterization of locally isotrivial
reductive group schemes \cite[XXIV.4.1.5]{SGA3}.
\newpage
\begin{scorollary}\label{cor_main_complete}
The following are equivalent:
\smallskip
(i) $\mathfrak{G}$ is locally (resp.\ semilocally) isotrivial;
\smallskip
(ii) The torus $\mathrm{rad}(\mathfrak{G})$ is locally (resp.\ semilocally) isotrivial;
\smallskip
(iii) the $\text{\rm{Out}}(G)_S$--torsor $\mathop{\rm Isomext}\nolimits(G_S, \mathfrak{G})$ is locally (resp.\ semilocally) isotrivial;
\smallskip
(iv) $\mathfrak{G}$ is locally (resp.\ semilocally) linear;
\smallskip
(v) $\mathrm{rad}(\mathfrak{G})$ is locally (resp.\ semilocally) linear.
\end{scorollary}
\begin{proof}
In view of Theorem \ref{thm_main_complete}, it remains
to establish the equivalence $(i) \Longleftrightarrow (ii)$.
Since this is precisely the quoted result \cite[XXIV.3.5]{SGA3}, the proof is complete.
\end{proof}
\begin{scorollary}\label{cor_main_complete2}
Let $\goth H$ be a reductive $S$--subgroup scheme of $\mathfrak{G}$.
If $\mathfrak{G}$ is locally (resp.\ semilocally) isotrivial, then
$\goth H$ is locally (resp.\ semilocally) isotrivial.
\end{scorollary}
\begin{proof}
Corollary \ref{cor_main_complete} shows that $\mathfrak{G}$ is
locally (resp.\ semilocally) linear and so is $\goth H$.
Therefore $\goth H$ is locally (resp.\ semilocally) isotrivial.
\end{proof}
\smallskip
\section{The semilocal case} \label{section_semilocal}
We assume that $S=\mathop{\rm Spec}\nolimits(R)$ where $R$ is a semilocal ring
and continue to assume that the reductive $S$--group scheme
$\mathfrak{G}$ is of constant type.
We remind the reader that $\mathfrak{G}$ admits a maximal torus
(Grothendieck, \cite[XIV.3.20 and footnote]{SGA3}).
\begin{scorollary} \label{cor_semilocal} Let $\goth T$ be a maximal torus of $\mathfrak{G}$.
The following are equivalent:
\smallskip
(i) $\mathfrak{G}$ is isotrivial;
\smallskip
(ii) The torus $\mathrm{rad}(\mathfrak{G})$ is isotrivial;
\smallskip
(iii) the $\text{\rm{Out}}(G)_S$--torsor $\mathop{\rm Isomext}\nolimits(G_S, \mathfrak{G})$ is isotrivial;
\smallskip
(iv) $\mathfrak{G}$ is linear;
\smallskip
(v) $\mathrm{rad}(\mathfrak{G})$ is linear;
\smallskip
(vi) $\goth T$ is linear;
\smallskip
(vii) $\goth T$ is isotrivial.
\end{scorollary}
\begin{proof}
From Corollary \ref{cor_main_complete}, we have the equivalences
$(i) \Longleftrightarrow (ii) \Longleftrightarrow (iii) \Longleftrightarrow (iv)
\Longleftrightarrow (v)$. On the other hand,
the equivalence $(vi) \Longleftrightarrow (vii)$ holds
according to Theorem \ref{thm_torus}. Now we observe that
the implications $(iv) \Longrightarrow (vi)$ and
$(vi) \Longrightarrow (v)$ are obvious so the proof is complete.
\end{proof}
\section{Equivariant resolution property}
\begin{sdefinition} Let $\mathfrak{G}$ be a flat group scheme over $S$ acting on a locally noetherian $S$--scheme
$\goth X$. One says that $(\mathfrak{G}, S, \goth X)$ has the
resolution property ( $(RE)$ for short) if for every coherent $\mathfrak{G}$-module $\mathcal F$ on $\goth X$, there
is a locally free coherent $\mathfrak{G}$-module ( i.e. a $\mathfrak{G}$-vector bundle $\mathcal E$) and a
$\mathfrak{G}$-equivariant epimorphism $\mathcal E \to \mathcal F \to 0$.
\end{sdefinition}
We strengthen Thomason's results.
\begin{stheorem} \label{thm_re} Let $\mathfrak{G}$ be a reductive $S$--group scheme.
We assume that $S$ is separated noetherian and that
$(1,S,S)$ satisfies the resolution property, e.g. $S$ is affine or regular or
admits an ample family of line bundles.
Then the following are equivalent:
\smallskip
(i) $\mathfrak{G}$ is linear;
\smallskip
(ii) $\mathrm{rad}(\mathfrak{G})$ is isotrivial;
\smallskip
(iii) $\mathfrak{G}$ satisfies $(RE)$.
\end{stheorem}
\begin{proof}
$(i) \Longleftrightarrow (ii)$. This is a special case of Theorem \ref{thm_main_complete}.
\smallskip
\noindent $(ii) \Longrightarrow (iii)$. This is Thomason's result \cite[Theorem 2.18]{T}.
\smallskip
\noindent $(iii) \Longrightarrow (i)$. This is Thomason's result \cite[Theorem 3.1]{T}, see also \cite[VI$_B$.13.5]{SGA3}.
\end{proof}
\begin{sremark}{\rm Example \ref{ex_groth}.(b) is an example of
a local noetherian ring $R$ and of a rank two non-isotrivial torus $\goth T$.
Theorem \ref{thm_re} shows that $\goth T$ does not satisfy (RE).
This answers a question of Thomason \cite[\S 2.3]{T}.
}
\end{sremark}
\smallskip
|
\section{Introduction}
\label{introduction}
The speed of sound, $c_s$, is a fundamental property of any substance. In fluids, it is the velocity of a longitudinal compression wave propagating through the medium, and its square is computed as the ratio of a change in the pressure, $P$, corresponding to a change in the energy density, $\mathcal{E}$. Therefore, it is directly related to the thermodynamic properties of the system, including its equation of state (EOS).
In dense nuclear matter, $c_s$ is of particular interest to neutron star research: its behavior as a function of baryon number density, $n_B$, influences the mass-radius relationship and, consequently, the maximum possible mass of neutron stars \cite{Ozel:2016oaf}. Current neutron star data suggest that $c_s$ rises significantly for $n_B$ larger than the nuclear saturation density, $n_0$, and that it perhaps exceeds
$c_s\sim1/\sqrt{3}$ at densities as low as few times that of normal nuclear matter. This possibility was first suggested in \cite{Bedaque:2014sqa}, followed by other studies, e.g.\ \cite{Tews:2018kmu,McLerran:2018hbz,Fujimoto:2019hxv}.
Presently, heavy-ion collisions are the only means of studying dense nuclear matter in a
laboratory. Experiments probing nuclear matter at high $n_B$, such as the Beam
Energy Scan (BES) program at RHIC, put special significance on the search for the QCD critical point (CP). Here, $c_s$ also conveys relevant information: it displays a local minimum at a crossover transition, whereas it vanishes at the CP and on the associated spinodal lines. Indeed, lattice QCD (LQCD) shows that at vanishing baryon chemical potential, $\mu_B=0$, a minimum in $c_s$ occurs at temperature $T_0=156.5\pm1.5$ MeV \cite{Bazavov:2018mes} (see also \cite{Borsanyi:2020fev}), corresponding to a crossover transition between hadron gas and quark-gluon plasma (QGP).
To date, a few attempts have been made to evaluate $c_s$ from heavy-ion collision data. In \cite{Gardim:2019xjs}, $c_s$ is estimated in ultra-relativistic collisions, where $\mu_B\approx n_B\approx0$, based on the proportionality of entropy density, $s$, and temperature, $T$, to charged particle multiplicity and mean transverse momentum, respectively. The estimated value agrees with LQCD results. At finite $\mu_B$, the Landau model and hybrid UrQMD simulations were utilized in \cite{Steinheimer:2012bp} to reproduce the widths of the negatively charged pion rapidity distribution. That study purports to locate a minimum in $c_s$ within the collision energy range $\sqrt{s_{NN}}=4\txt{-}9\ \txt{GeV}$.
In this work we suggest a novel approach to exploring the behavior of $c_s$ by utilizing cumulants of the baryon number distribution. The sensitivity of the cumulants to the EOS near the CP \cite{Asakawa:2009aj, Stephanov:2011pb}, making them central observables pursued in the BES, follows directly from their sensitivity to derivatives of the pressure with respect to $\mu_B$. The key observation in this paper is that besides the vicinity of the CP, cumulants provide rich information about the EOS at all points of the phase diagram, and in particular they allow a measurement of $c_s$ in matter created in heavy-ion collisions.
\section{Cumulants and the speed of sound}
\label{relationship_cumulants_cT2}
Cumulants of net baryon number $\kappa_j$ are defined as $\kappa_j=VT^{j-1}\left(\frac{d^jP}{d\mu_B^j}\right)_T$, where $V$ is the volume. Expressed in terms of derivatives with respect to $n_B$, the first three cumulants are given by
\begin{eqnarray}
&& \kappa_1 = V n_B ~, \label{cumulant_1} \hspace{8mm} \kappa_2 = \frac{VTn_B}{\left( \frac{dP}{dn_B} \right)_T}~, \label{cumulant_2} \nonumber \\
&& \kappa_3= \frac{VT^2n_B}{\left( \frac{dP}{dn_B} \right)_T^2} \left[ 1 - \frac{n_B}{\left( \frac{dP}{dn_B} \right)_T} \left(\frac{d^2P}{dn_B^2}\right)_T \right] \label{cumulant_3} ~.
\end{eqnarray}
Importantly, cumulants are related to moments of the baryon number distribution. In particular, for $j\leq3$, $\kappa_j\equiv\big\langle\big(N_B-\big\langle N_B\big\rangle\big)^j\big\rangle$.
The definition of $c_s$ requires specifying which properties of the system are
considered constant during the propagation of the compression wave. One often uses the speed of
sound at constant entropy $S$ per net baryon number $N_B$, $c_{\sigma}^2\equiv\left(
\frac{dP}{d\mathcal E}\right)_{\sigma}$, where $\sigma=S/N_B$. Similarly, the speed of sound at constant temperature is $c_T^2\equiv\left(\frac{dP}{d\mathcal E}\right)_{T}$. These variants have specific regions of applicability: For example, the propagation of sound in air is governed by adiabatic
compression, so that using $c_{\sigma}^2$ is appropriate. On the other hand, when
there is a temperature reservoir (e.g.\ in porous media) or when the cooling timescale is fast compared with the sound wave period (as is the case e.g.\ for an interstellar medium subject to radiative cooling), $c_T^2$ is applicable.
Explicitly,
$c_{\sigma}^2$ and $c_T^2$
can be written as
\begin{eqnarray}
c_{\sigma}^2 = \frac{ \Big( \frac{d P}{dn_B} \Big)_{T} \Big(\frac{d s}{d T} \Big)_{n_B} + \Big( \frac{ d P}{dT} \Big)_{n_B} \bigg[ \frac{s}{n_B} - \Big(\frac{d s}{d n_B}\Big)_T \bigg] }{ \Big(\frac{sT}{n_B} + \mu_B \Big) \Big( \frac{d s}{d T} \Big)_{n_B} }
\label{speed_isentropic}
\end{eqnarray}
and
\begin{eqnarray}
c_T^2 = \frac{\Big( \frac{dP}{dn_B} \Big)_T}{ T \Big(\frac{d s}{d n_B} \Big)_T + \mu_B }~.
\label{speed_isothermal}
\end{eqnarray}
In the limit $T\to0$, the above expressions both lead to
\begin{eqnarray}
c^2\Big|_{T=0} = \frac{1}{\mu_B} \bigg( \frac{dP}{dn_B} \bigg)_T ~.
\label{cT2_approx}
\end{eqnarray}
Consequently, for $\frac{\mu_B}{T}\gg1$ the values of $c_{\sigma}^2$ and $c_T^2$ should largely coincide. Moreover, Eq.\ (\ref{speed_isothermal}) can be transformed to express $c_T^2$ as a function of the cumulants, Eq.\ (\ref{cumulant_3}),
\begin{eqnarray}
c_T^{2} = \left[\bigg(\parr{\log \kappa_1}{\log T}\bigg)_{\mu_B} + \frac{\mu_B}{T} \frac{\kappa_2}{\kappa_1} \right]^{-1}~.
\label{cT2_as_function_of_cumulants}
\end{eqnarray}
The first term in Eq.\ (\ref{cT2_as_function_of_cumulants}) is challenging to estimate from experimental data, however, it can be shown to be negligible for a degenerate Fermi gas, $\frac{\mu_B}{T}\gg1$, where it constitutes an order $\left(T/\mu_B\right)^2$ correction; then
\begin{eqnarray}
c_T^2 \approx \frac{T \kappa_1}{\mu_B \kappa_2}~.
\label{magic_equation_1}
\end{eqnarray}
We note that Eq.\ (\ref{magic_equation_1}) provides an upper limit to the value of $c_T^2$ as long as $\Big(\parr{\log \kappa_1}{\log T}\Big)_{\mu_B} > 0$.
Using Eq.\ (\ref{speed_isothermal}), one can also calculate the logarithmic derivative of $c_T^2$,
\begin{eqnarray}
\hspace{-1mm}\bigg(\frac{d \ln c_T^2}{d \ln n_B} \bigg)_T
= \frac{n_B \Big( \frac{d^2P}{dn_B^2} \Big)_T}{ \Big( \frac{dP}{dn_B} \Big)_T} - \frac{ \Big( \frac{dP}{dn_B} \Big)_T + Tn_B \Big( \frac{d^2s}{dn_B^2} \Big)_T }{\mu_B + T \Big( \frac{ds}{dn_B} \big)_T} ~.
\end{eqnarray}
It is again possible to rewrite the above equation in terms of the cumulants,
\begin{eqnarray}
\bigg(\frac{d \ln c_T^2}{d \ln n_B} \bigg)_T + c_T^2
= 1 - \frac{\kappa_3 \kappa_1}{\kappa_2^2} - c_T^2 \bigg(\frac{d \ln \big(\frac{\kappa_2}{T}\big)}{d \ln T}\bigg)_{n_B} ~,
\end{eqnarray}
and neglecting the last term on the right-hand side yields
\begin{eqnarray}
\left(\frac{d \ln c_T^2}{d \ln n_B} \right)_T + c_T^2 \approx 1 - \frac{\kappa_3 \kappa_1}{\kappa_2^2} ~.
\label{magic_equation_2}
\end{eqnarray}
This approximation is again valid for $\frac{\mu_B}{T} \gg 1$, and the correction due to the neglected term is likewise of order $\left(T/\mu_B\right)^2$.
We note that in the opposite limit, $\mu_B \to 0$, Eq.\ (\ref{cT2_as_function_of_cumulants}) reveals a similarly simple form, $c_T^{2} = \left( \frac{d \ln \kappa_2}{d \ln T} \right)^{-1}_{\mu_B=0}$, suggesting that $c_T^2$ can be estimated in ultrarelativistic heavy-ion collisions, provided measurements of $\kappa_2$ are available at different temperatures. It might be possible to achieve this with data from a combination of centralities, energies, collision species, or rapidity ranges. In this work, however, we are interested in utilizing Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}) applied to collisions at medium and low energies.
\begin{figure*}[t]
\includegraphics[width = 0.99\textwidth]{model_comparison_combined.jpeg}
%
\caption{(Color online) Model study of regions of applicability of Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}). The left (right) panels show results obtained in the VDF (Walecka) model. The upper and lower panels show quantities entering Eq.\ (\ref{magic_equation_1}) and Eq.\ (\ref{magic_equation_2}), respectively. Results at $T=50,100,150,200\ \txt{MeV}$ are given by blue and green solid lines, dark and light purple long-dashed lines, red and pink short-dashed lines, and orange and brown dash-dotted lines, respectively. For each $T$, the thickest lines correspond to the exact results and the medium-thick lines correspond to the approximations, given by the right-hand sides of Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}). Additionally, on upper panels the thinnest lines correspond to $c_{\sigma}^2$. Upper panels: for both models, Eq.\ (\ref{magic_equation_1}) is valid for $T\lesssim100\ \txt{MeV}$ and $\mu_B\gtrsim600\ \txt{MeV}$. Lower panels: for both models, Eq.\ (\ref{magic_equation_2}) is valid for $\mu_B\gtrsim200\ \txt{MeV}$; the exception is the Walecka model at $T=200\ \txt{MeV}$, where a phase transition to an almost massless gas of nucleons dramatically decreases the applicability of both Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}).
}
\label{tests_of_formulas}
\end{figure*}
\section{Validation}
We are interested in finding the limitations of the low-temperature approximation used to derive Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}), and for this we use effective models. Anticipating applying our formulas in regions of the phase diagram described by hadronic degrees of freedom, we choose two models of dense nuclear matter: the vector density functional (VDF) model with two phase transitions \cite{Sorensen:2020ygf} and the Walecka model \cite{Walecka:1974qa}. The VDF model utilizes interactions of vector type, while the Walecka model employs both vector- and scalar-type interactions. Both models describe the nuclear liquid-gas phase transition, while the VDF model additionally describes a conjectured high-density, high-temperature phase transition modeling the QGP phase transition. In this work, the QGP-like phase transition is chosen to exhibit a CP at $T_c=100\ \txt{MeV}$ and $n_c=3\ n_0$, with the $T=0$ boundaries of the spinodal region in the $T$-$n_B$ plane given by $n_{B, \txt{left spinodal} } (T=0) \equiv \eta_L=2.5\ n_0$ and $n_{B, \txt{right spinodal}} (T=0) \equiv\eta_R=3.32\ n_0$, where $n_0=0.160\ \txt{fm}^{-3}$; this choice is arbitrary and serves as a plausible example.
We plot both sides of Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}) as functions of $\mu_B$ at a series of
temperatures in Fig.\ \ref{tests_of_formulas}. We note that the explored temperature range reaches beyond the region of the phase diagram where hadronic models plausibly describe matter created in heavy-ion collisions, nevertheless, it is instructive to test our approximations in this regime.
We utilize natural units in which the speed of light in vacuum is $c=1$. We note that in the VDF model, $c_s$ quickly becomes acausal for $\mu_B$ above the QGP-like phase transition. It is an expected behavior in models utilizing interactions dependent on high powers of $n_B$ \cite{Zeldovich:1962emp}, and while not ideal, it does not affect the current analysis.
On all panels in Fig.\ \ref{tests_of_formulas}, the exact model calculations show expected
features as functions of $\mu_B$. In the upper left panel, showing both $c_{T}^2$ and $c_{\sigma}^2$ in the VDF model, at small $\mu_B$ we see a softening of the EOS due to the influence of the nuclear CP, followed by an increase at densities of the order of $n_0$, then a dive in $c_s^2$ caused by the QGP-like phase transition, and finally a steep rise for high values of $\mu_B$. In the upper right panel, showing $c_T^2$ and $c_{\sigma}^2$ in the Walecka model, we similarly observe a soft EOS at small $\mu_B$, while the value of $c_s^2$ goes asymptotically to 1 for large $\mu_B$. Additionally, for $T=200\ \txt{MeV}$, the Walecka model shows effects due to a phase transition in the nucleon-antinucleon plasma, occurring around $T\approx190\ \txt{MeV}$ and $n_B=0$; above this transition, the model describes an almost non-interacting gas of nearly massless nucleons \cite{Theis:1984qc}. The behavior of the curves in the lower panels, showing $c_T^2+\big(\frac{d\ln c_T^2}{d\ln n_B}\big)_T$, can be directly traced to the behavior of the curves in the upper panels. In particular, for the VDF model we observe strong divergences due to the softening of the EOS in the QGP-like phase transition region.
Comparing the exact results with the approximations, we see that while Eq.\ (\ref{magic_equation_1}) is valid for $T\lesssim100\ \txt{MeV}$ and $\mu_B\gtrsim600\ \txt{MeV}$, it behaves poorly, both qualitatively and quantitatively, for $T$ and $\mu_B$ corresponding to regions of the phase diagram probed by moderately- to highly-energetic heavy-ion collisions (upper panels). On the other hand, the approximation introduced in Eq.\ (\ref{magic_equation_2}) is qualitatively valid for most of the probed $T$ and $\mu_B$, with the exception of regions characterized by $\mu_B\lesssim200\ \txt{MeV}$ (lower panels).
\begin{figure}[t]
\includegraphics[width = 0.99\columnwidth]{STAR+HADES_and_VDF_plots_combined_v2.jpeg}\\
%
\caption{(Color online) Comparison of the right-hand sides of Eq.\ (\ref{magic_equation_1}) (upper panel) and Eq.\ (\ref{magic_equation_2}) (lower panel) for experimental data (red triangles), ideal gas at the freeze-out (small grey circles), the VDF model at the freeze-out (light green stars), and the VDF model at a set of points chosen to reproduce the data (dark purple stars); exact results, that is left-hand sides of Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}), are shown for the two cases considered in the VDF model (green and purple circles). The data points for the matched VDF results (shown only for collisions at low energies, where using the model is justified) are chosen to reproduce experimental values of $1-\kappa_3\kappa_1/\kappa_2^2$ (see Fig.\ \ref{diagram}). We note that at $\sqrt{s}=2.4\ \txt{GeV}$, matching the value of $1-\kappa_3 \kappa_1/\kappa_2^2$ exactly is possible, but would place the matched point close to the nuclear liquid-gas CP, which we find unlikely.
}
\label{STAR_HADES_plots}
\end{figure}
\begin{figure}[t]
%
\includegraphics[width = 0.99\columnwidth]{Cumulants_diagram_1minusk3k1overk2k2_with_legend.jpeg}
\caption{(Color online) Contour plot of $1-\kappa_3\kappa_1/\kappa_2^2$ in the VDF model. Yellow and black lines correspond to the spinodal and coexistence lines, respectively; white contours signifiy regions where $1-\kappa_3\kappa_1/\kappa_2^2=1\pm0.03$. Light green stars denote experimentally measured freeze-out parameters $(T_{\txt{fo}},\mu_{\txt{fo}})$, while dark purple stars denote points where $1-\kappa_3\kappa_1/\kappa_2^2$, taken along lines informed by average phase diagram trajectories for STAR collision energies \cite{Shen:2020jwv}, matches the experimentally measured values for a given collision energy. The softening of the EOS, leading to negative values of $\big(\frac{d\ln c_T^2}{d\ln n_B}\big)_T$, occurs in two regions of the phase diagram, corresponding to the ordinary nuclear matter phase transition and to the conjectured QGP-like phase transition.
}
\label{diagram}
\end{figure}
\section{Experimental data and interpretation}
We consider cumulants of net proton number and chemical freeze-out parameters, $(T_{\txt{fo}},\mu_{\txt{fo}})$, in collisions at 0-5\% centrality, determined by the STAR \cite{Abdallah:2021fzj} and HADES \cite{Adamczewski-Musch:2020slf, HADES_MLorentz_talk} experiments, and we use them to plot Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}) (red triangles, upper and lower panel in Fig.\ \ref{STAR_HADES_plots}, respectively) against $\mu_B$.
Based on the previous section, we trust the results presented in the upper panel of Fig.\ \ref{STAR_HADES_plots}, approximating $c_T^2$, only for the lowest collision energy, $\sqrt{s}=2.4\ \txt{GeV}$ from the HADES experiment. Here, $c_T^2$ as obtained from Eq.\ (\ref{magic_equation_1}) is small: less than half of the ideal gas value. At the same time, the value of $1-\kappa_3\kappa_1/\kappa_2^2$ (shown in the lower panel in Fig.\ \ref{STAR_HADES_plots}), which we assume is dominated by $\big(\frac{d\ln c_T^2}{d\ln n_B}\big)_T$, drops with decreasing collision energy to reach a minimum at the lowest STAR point, $\sqrt{s}=7.7\ \txt{GeV}$, and then steeply rises for the HADES point. This could mean that in $\sqrt{s} = 7.7\ \txt{GeV}$ collisions, $c_T^2$ is approximately constant as a function of $n_B$, while in $\sqrt{s} = 2.4\ \txt{GeV}$ collisions the matter is characterized by a small $c^2_T$ which nevertheless has a large slope as a function of $n_B$.
To further understand this behavior, we study the dependence of $1-\kappa_3\kappa_1/\kappa_2^2$ on $\mu_{B}$ and $T$ within the VDF model, shown in Fig.\ \ref{diagram}. Comparing model results with experiment requires choosing at which $T$ and $\mu_B$ to take values of $1-\kappa_3 \kappa_1/\kappa_2^2$. A natural choice is to use values at $(T_{\txt{fo}},\mu_{\txt{fo}})$ (light green stars), but as shown in the lower panel of Fig.\ \ref{STAR_HADES_plots}, these values do not lead to an agreement with experimental data, with the biggest discrepancy for the HADES point. (We note that $(T_{\txt{fo}},\mu_{\txt{fo}})$ are established with hadron interactions neglected, and the degree to which this affects our results may vary across the phase diagram.) However, critical fluctuations exhibit a large relaxation time \cite{Berdnikov:1999ph,Stephanov:2017ghc,Du:2020bxp}, and their measured values could be affected by stages of the collision preceding the freeze-out. With this insight and taking guidance from average phase diagram trajectories of hybrid simulations of heavy-ion collisions \cite{Shen:2020jwv}, we consider $1-\kappa_3\kappa_1/\kappa_2^2$ at slightly earlier stages of the evolution. In this way we obtain values of $1-\kappa_3\kappa_1/\kappa_2^2$ (dark purple stars) that reproduce the experimental results for a given collision energy (lower panel in Fig.\ \ref{STAR_HADES_plots}); here the exception is the HADES point, for which we prioritized choosing a point in a reasonable vicinity of the measured freeze-out over obtaining a value equal to the experimental data. Comparing with the exact model results for $c_T^2$ and $c_T^2 + \big(\frac{d\ln c_T^2}{d\ln n_B}\big)_T$, also displayed in Fig.\ \ref{STAR_HADES_plots}, as well as with the upper left panel of Fig.\ \ref{tests_of_formulas}, we can confirm that at the point reproducing the experimental value of $1-\kappa_3\kappa_1/\kappa_2^2$ for the lowest STAR energy, $c_T^2$ is nearly constant as a function of $\mu_B$ (thick short-dashed line, $T=150\ \txt{MeV}$, at $\mu_B\approx410\ \txt{MeV}$ in Fig.\ \ref{tests_of_formulas}), while at the point reproducing the result for the HADES energy, $c_T^2$ increases sharply with $\mu_B$ (thick solid line, $T=50\ \txt{MeV}$, at $\mu_B\approx850\ \txt{MeV}$ in Fig.\ \ref{tests_of_formulas}).
Naturally, the choice of $T$ and $\mu_B$ at which we compare model calculations with STAR and HADES cumulant data is driven by the wish to match the experimental results, and it serves mainly to show that baryon number cumulants measured in heavy-ion collisions can be connected to the speed of sound in hot and dense nuclear matter. Whether values of higher order cumulants are indeed significantly affected by stages of the evolution preceding the freeze-out needs to be further investigated (for recent developments, see \cite{Nahrgang:2018afz,Jiang:2017sni}). Moreover, while experiments measure proton number cumulants, the VDF model provides baryon number cumulants, putting more strain on our interpretation. Baryon number conservation should likewise be important \cite{Bzdak:2012an,Vovchenko:2020gne}. Finally, our model results may not be applicable in regions of the phase diagram where quarks and gluons become increasingly relevant.
Nonetheless, hadronic models are well-justified for describing low-energy collisions whose evolution is dominated by the hadronic stage. The comparison between the experimental data and the VDF model suggests that collisions at the lowest STAR and HADES energies may be probing regions of the phase diagram affected most by the nuclear liquid-gas phase transition, and the corresponding cumulants of the baryon number may be telling us more about hadronic physics than the QCD CP. If this is true, it may be worthwhile to study the cumulants at even lower collision energies, starting from $0.1\ \txt{GeV}$ projectile kinetic energy, and obtain the speed of sound around the nuclear liquid-gas CP. Conversely, at higher energies it could be possible to utilize collisions at different centralities and different rapidity windows to estimate the neglected terms in Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}), and obtain a stronger estimate for the speed of sound in the respective regions of the phase diagram.
\section{Summary and conclusions}
In this work we utilize cumulants of the baryon number distribution to estimate the isothermal speed of sound squared and its logarithmic derivative with respect to the baryon number density. This result provides a new method for obtaining information about fundamental properties of nuclear matter studied in heavy-ion collisions, with consequences for both the search for the QCD CP and neutron star studies. While the approximations and the model comparison we considered apply to experiments at low energies, the approach itself can be used at any collision energy provided that measurements of cumulants of baryon number as well as their temperature dependence are available. Further studies of effects due to dynamics, in particular utilizing state-of-the-art simulations, will be absolutely essential in determining the extent to which the proposed method provides a reliable extraction of sound velocities and their derivatives.
{\hspace{10mm}}
\section{Acknowledgements}
A.S.\ thanks Chun Shen for providing average phase diagram trajectories of hybrid heavy-ion collision simulations at RHIC BES energies.
A.S.\ and V.K.\ received support through the U.S. Department of Energy,
Office of Science, Office of Nuclear Physics, under contract number
DE-AC02-05CH11231231 and received support within the framework of the
Beam Energy Scan Theory (BEST) Topical Collaboration. D.O.\ and L.M.\ were supported by the U.S.\ DOE under Grant No.\ DE-FG02-00ER4113.
\section{Introduction}
\label{introduction}
The speed of sound, $c_s$, is a fundamental property of any substance. In fluids, it is the velocity of a longitudinal compression wave propagating through the medium, and its square is computed as the ratio of a change in the pressure, $P$, corresponding to a change in the energy density, $\mathcal{E}$. Therefore, it is directly related to the thermodynamic properties of the system, including its equation of state (EOS).
In dense nuclear matter, $c_s$ is of particular interest to neutron star research: its behavior as a function of baryon number density, $n_B$, influences the mass-radius relationship and, consequently, the maximum possible mass of neutron stars \cite{Ozel:2016oaf}. Current neutron star data suggest that $c_s$ rises significantly for an $n_B$ larger than the nuclear saturation density, $n_0$, and that it perhaps exceeds
$c_s\sim1/\sqrt{3}$ at densities as low as a few times that of normal nuclear matter. This possibility was first suggested in \cite{Bedaque:2014sqa}, followed by other studies, e.g.,\ \cite{Tews:2018kmu,McLerran:2018hbz,Fujimoto:2019hxv}.
Presently, heavy-ion collisions are the only means of studying dense nuclear matter in a
laboratory. Experiments probing nuclear matter at high $n_B$, such as the Beam
Energy Scan program at the Relativistic Heavy Ion Collider (RHIC), put special significance on the search for the QCD critical point (CP). Here, $c_s$ also conveys relevant information: it displays a local minimum at a crossover transition, whereas it vanishes at the CP and on the associated spinodal lines. Indeed, lattice QCD shows that at vanishing baryon chemical potential, $\mu_B=0$, a minimum in $c_s$ occurs at temperature $T_0=156.5\pm1.5$ MeV \cite{Bazavov:2018mes} (see also \cite{Borsanyi:2020fev}), corresponding to a crossover transition between hadron gas and quark-gluon plasma (QGP).
To date, a few attempts have been made to evaluate $c_s$ from heavy-ion collision data. In \cite{Gardim:2019xjs}, $c_s$ is estimated in ultrarelativistic collisions, where $\mu_B\approx n_B\approx0$, based on the proportionality of the entropy density, $s$, and the temperature, $T$, to the charged particle multiplicity and mean transverse momentum, respectively. The estimated value agrees with lattice QCD results. At finite $\mu_B$, the Landau model as well as hybrid hydrodynamics and hadronic transport simulations were used in \cite{Steinheimer:2012bp} to reproduce the widths of the negatively charged pion rapidity distribution. That study purports to locate a minimum in $c_s$ within the collision energy range $\sqrt{s_{NN}}=4\txt{-}9\ \txt{GeV}$.
In this Letter, we suggest a novel approach to exploring the behavior of $c_s$ by using cumulants of the baryon number distribution. The sensitivity of the cumulants to the EOS near the CP \cite{Asakawa:2009aj, Stephanov:2011pb}, which makes them central observables pursued in the Beam Energy Scan, follows directly from their sensitivity to derivatives of the pressure with respect to $\mu_B$. The key observation in this Letter is that, besides the vicinity of the CP, cumulants provide rich information about the EOS at all points of the phase diagram, and in particular they allow a measurement of $c_s$ in matter created in heavy-ion collisions.
\section{Cumulants and the speed of sound}
\label{relationship_cumulants_cT2}
Cumulants of net baryon number $\kappa_j$ are defined as $\kappa_j=VT^{j-1}\left(d^jP/d\mu_B^j\right)_T$, where $V$ is the volume. Expressed in terms of derivatives with respect to $n_B$, the first three cumulants are given by
\begin{eqnarray}
&& \kappa_1 = V n_B ~, \label{cumulant_1} \hspace{8mm} \kappa_2 = \frac{VTn_B}{\left( \frac{dP}{dn_B} \right)_T}~, \label{cumulant_2} \nonumber \\
&& \kappa_3= \frac{VT^2n_B}{\left( \frac{dP}{dn_B} \right)_T^2} \left[ 1 - \frac{n_B}{\left( \frac{dP}{dn_B} \right)_T} \left(\frac{d^2P}{dn_B^2}\right)_T \right] \label{cumulant_3} ~.
\end{eqnarray}
Importantly, cumulants are related to moments of the baryon number distribution. In particular, for $j\leq3$, $\kappa_j\equiv\big\langle\big(N_B-\big\langle N_B\big\rangle\big)^j\big\rangle$.
The definition of $c_s$ requires specifying which properties of the system are
considered constant during the propagation of the compression wave. One often uses the speed of
sound at constant entropy $S$ per net baryon number $N_B$, $c_{\sigma}^2\equiv\left(
dP/d\mathcal E\right)_{\sigma}$, where $\sigma=S/N_B$. Similarly, the speed of sound at constant temperature is $c_T^2\equiv\left(dP/d\mathcal E\right)_{T}$. These variants have specific regions of applicability. For example, the propagation of sound in air is governed by adiabatic
compression, so that using $c_{\sigma}^2$ is appropriate. On the other hand, when
there is a temperature reservoir (e.g.,\ in porous media) or when the cooling timescale is fast compared with the sound wave period (as is the case, e.g.,\ for an interstellar medium subject to radiative cooling), $c_T^2$ is applicable.
Explicitly,
$c_{\sigma}^2$ and $c_T^2$
can be written as
\begin{eqnarray}
c_{\sigma}^2 = \frac{ \Big( \frac{d P}{dn_B} \Big)_{T} \Big(\frac{d s}{d T} \Big)_{n_B} + \Big( \frac{ d P}{dT} \Big)_{n_B} \bigg[ \frac{s}{n_B} - \Big(\frac{d s}{d n_B}\Big)_T \bigg] }{ \Big(\frac{sT}{n_B} + \mu_B \Big) \Big( \frac{d s}{d T} \Big)_{n_B} }
\label{speed_isentropic}
\end{eqnarray}
and
\begin{eqnarray}
c_T^2 = \frac{\Big( \frac{dP}{dn_B} \Big)_T}{ T \Big(\frac{d s}{d n_B} \Big)_T + \mu_B }~.
\label{speed_isothermal}
\end{eqnarray}
In the limit $T\to0$, the above expressions both lead to
\begin{eqnarray}
c^2\Big|_{T=0} = \frac{1}{\mu_B} \bigg( \frac{dP}{dn_B} \bigg)_T ~.
\label{cT2_approx}
\end{eqnarray}
Consequently, for $(\mu_B/T)\gg1$, the values of $c_{\sigma}^2$ and $c_T^2$ should largely coincide. Moreover, Eq.\ (\ref{speed_isothermal}) can be transformed to express $c_T^2$ as a function of the Eq.\ (\ref{cumulant_3}) cumulants,
\begin{eqnarray}
c_T^{2} = \left[\bigg(\parr{\log \kappa_1}{\log T}\bigg)_{\mu_B} + \frac{\mu_B}{T} \frac{\kappa_2}{\kappa_1} \right]^{-1}~.
\label{cT2_as_function_of_cumulants}
\end{eqnarray}
The first term in Eq.\ (\ref{cT2_as_function_of_cumulants}) is challenging to estimate from experimental data, however, it can be shown to be negligible for a degenerate Fermi gas, $(\mu_B/T)\gg1$, where it constitutes an order $\left(T/\mu_B\right)^2$ correction; then
\begin{eqnarray}
c_T^2 \approx \frac{T \kappa_1}{\mu_B \kappa_2}~.
\label{magic_equation_1}
\end{eqnarray}
We note that Eq.\ (\ref{magic_equation_1}) provides an upper limit to the value of $c_T^2$ as long as $(\partial \log \kappa_1/\partial \log T)_{\mu_B} > 0$.
Using Eq.\ (\ref{speed_isothermal}), one can also calculate the logarithmic derivative of $c_T^2$,
\begin{eqnarray}
\hspace{-1mm}\bigg(\frac{d \ln c_T^2}{d \ln n_B} \bigg)_T
= \frac{n_B \Big( \frac{d^2P}{dn_B^2} \Big)_T}{ \Big( \frac{dP}{dn_B} \Big)_T} - \frac{ \Big( \frac{dP}{dn_B} \Big)_T + Tn_B \Big( \frac{d^2s}{dn_B^2} \Big)_T }{\mu_B + T \Big( \frac{ds}{dn_B} \big)_T} ~.
\end{eqnarray}
It is again possible to rewrite the above equation in terms of the cumulants,
\begin{eqnarray}
\bigg(\frac{d \ln c_T^2}{d \ln n_B} \bigg)_T + c_T^2
= 1 - \frac{\kappa_3 \kappa_1}{\kappa_2^2} - c_T^2 \bigg(\frac{d \ln (\kappa_2/T)}{d \ln T}\bigg)_{n_B} ~,
\end{eqnarray}
and neglecting the last term on the right-hand side yields
\begin{eqnarray}
\left(\frac{d \ln c_T^2}{d \ln n_B} \right)_T + c_T^2 \approx 1 - \frac{\kappa_3 \kappa_1}{\kappa_2^2} ~.
\label{magic_equation_2}
\end{eqnarray}
This approximation is again valid for $(\mu_B/T) \gg 1$, and the correction due to the neglected term is likewise of order $\left(T/\mu_B\right)^2$.
We note that in the opposite limit, $\mu_B \to 0$, Eq.\ (\ref{cT2_as_function_of_cumulants}) reveals a similarly simple form, $c_T^{2} = \left( d \ln \kappa_2/d \ln T \right)^{-1}_{\mu_B=0}$, suggesting that $c_T^2$ can be estimated in ultrarelativistic heavy-ion collisions, provided measurements of $\kappa_2$ are available at different temperatures. It might be possible to achieve this with data from a combination of centralities, energies, collision species, or rapidity ranges. In this work, however, we are interested in utilizing Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}) applied to collisions at medium and low energies.
\begin{figure*}[t]
\includegraphics[width = 0.99\textwidth]{model_comparison_combined.jpg}
%
\caption{Model study of regions of applicability of Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}). The left (right) panels show results obtained in the VDF (Walecka) model. The upper and lower panels show quantities entering Eq.\ (\ref{magic_equation_1}) and Eq.\ (\ref{magic_equation_2}), respectively. Results at $T=50,100,150,200\ \txt{MeV}$ are given by blue and green solid lines, dark and light purple long-dashed lines, red and pink short-dashed lines, and orange and brown dash-dotted lines, respectively. For each $T$, the thickest lines correspond to the exact results and the medium-thick lines correspond to the approximations, given by the right-hand sides of Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}). Additionally, in upper panels the thinnest lines correspond to $c_{\sigma}^2$. Upper panels: for both models, Eq.\ (\ref{magic_equation_1}) is valid for $T\lesssim100\ \txt{MeV}$ and $\mu_B\gtrsim600\ \txt{MeV}$. Lower panels: for both models, Eq.\ (\ref{magic_equation_2}) is valid for $\mu_B\gtrsim200\ \txt{MeV}$; the exception is the Walecka model at $T=200\ \txt{MeV}$, where a phase transition to an almost massless gas of nucleons dramatically decreases the applicability of both Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}).
}
\label{tests_of_formulas}
\end{figure*}
\section{Validation}
We are interested in finding the limitations of the low-temperature approximation used to derive Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}), and for this we use effective models. Anticipating applying our formulas in regions of the phase diagram described by hadronic degrees of freedom, we choose two models of dense nuclear matter: the vector density functional (VDF) model with two phase transitions \cite{Sorensen:2020ygf} and the Walecka model \cite{Walecka:1974qa}. The VDF model utilizes interactions of the vector type, while the Walecka model employs both vector- and scalar-type interactions. Both models describe the nuclear liquid-gas phase transition, while the VDF model additionally describes a conjectured high-density, high-temperature phase transition modeling the QGP phase transition. In this work, the QGP-like phase transition is chosen to exhibit a CP at $T_c=100\ \txt{MeV}$ and $n_c=3n_0$, with the $T=0$ boundaries of the spinodal region in the $T$-$n_B$ plane given by $n_{B, \txt{left spinodal} } (T=0) \equiv \eta_L=2.5 n_0$ and $n_{B, \txt{right spinodal}} (T=0) \equiv\eta_R=3.32 n_0$, where $n_0=0.160\ \txt{fm}^{-3}$; this choice is arbitrary and serves as a plausible example.
We plot both sides of Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}) as functions of $\mu_B$ at a series of
temperatures in Fig.\ \ref{tests_of_formulas}. We note that the explored temperature range reaches beyond the region of the phase diagram where hadronic models plausibly describe matter created in heavy-ion collisions; nevertheless, it is instructive to test our approximations in this regime.
We use natural units in which the speed of light in vacuum is $c=1$. We note that in the VDF model, $c_s$ quickly becomes acausal for $\mu_B$ above the QGP-like phase transition. It is an expected behavior in models using interactions dependent on high powers of $n_B$ \cite{Zeldovich:1962emp}, and while not ideal, it does not affect the current analysis.
In all panels in Fig.\ \ref{tests_of_formulas}, the exact model calculations show expected
features as functions of $\mu_B$. In the upper left panel, showing both $c_{T}^2$ and $c_{\sigma}^2$ in the VDF model, at small $\mu_B$ we see a softening of the EOS due to the influence of the nuclear CP, followed by an increase at densities of the order of $n_0$, then a dive in $c_s^2$ caused by the QGP-like phase transition, and finally a steep rise for high values of $\mu_B$. In the upper right panel, showing $c_T^2$ and $c_{\sigma}^2$ in the Walecka model, we similarly observe a soft EOS at small $\mu_B$, while the value of $c_s^2$ goes asymptotically to 1 for large $\mu_B$. Additionally, for $T=200\ \txt{MeV}$, the Walecka model shows effects due to a phase transition in the nucleon-antinucleon plasma, occurring around $T\approx190\ \txt{MeV}$ and $n_B=0$; above this transition, the model describes an almost noninteracting gas of nearly massless nucleons \cite{Theis:1984qc}. The behavior of the curves in the lower panels, showing $c_T^2+\big(\frac{d\ln c_T^2}{d\ln n_B}\big)_T$, can be directly traced to the behavior of the curves in the upper panels. In particular, for the VDF model we observe strong divergences due to the softening of the EOS in the QGP-like phase transition region.
Comparing the exact results to the approximations, we see that, while Eq.\ (\ref{magic_equation_1}) is valid for $T\lesssim100\ \txt{MeV}$ and $\mu_B\gtrsim600\ \txt{MeV}$, it behaves poorly, both qualitatively and quantitatively, for $T$ and $\mu_B$ corresponding to regions of the phase diagram probed by moderately to highly energetic heavy-ion collisions (upper panels). On the other hand, the approximation introduced in Eq.\ (\ref{magic_equation_2}) is qualitatively valid for most of the probed $T$ and $\mu_B$, with the exception of regions characterized by $\mu_B\lesssim200\ \txt{MeV}$ (lower panels).
\begin{figure}[t]
\includegraphics[width = 0.99\columnwidth]{STAR+HADES_and_VDF_plots_combined_v2.jpg}\\
%
\caption{Comparison of the right-hand sides of Eq.\ (\ref{magic_equation_1}) (upper panel) and Eq.\ (\ref{magic_equation_2}) (lower panel) for experimental data (red triangles), ideal gas at the freeze-out (small gray circles), the VDF model at the freeze-out (light green stars), and the VDF model at a set of points chosen to reproduce the data (dark purple stars); exact results, that is, the left-hand sides of Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}), are shown for the two cases considered in the VDF model (green and purple circles). The data points for the matched VDF results (shown only for collisions at low energies, where using the model is justified) are chosen to reproduce experimental values of $1-\kappa_3\kappa_1/\kappa_2^2$ (see Fig.\ \ref{diagram}). We note that at $\sqrt{s}=2.4\ \txt{GeV}$, matching the value of $1-\kappa_3 \kappa_1/\kappa_2^2$ exactly is possible, but would place the matched point close to the nuclear liquid-gas CP, which we find unlikely.
}
\label{STAR_HADES_plots}
\end{figure}
\begin{figure}[t]
%
\includegraphics[width = 0.99\columnwidth]{Cumulants_diagram_1minusk3k1overk2k2_with_legend.jpg}
\caption{Contour plot of $1-\kappa_3\kappa_1/\kappa_2^2$ in the VDF model. Yellow and black lines correspond to the spinodal and coexistence lines, respectively; white contours signify regions where $1-\kappa_3\kappa_1/\kappa_2^2=1\pm0.03$. Light green stars denote experimentally measured freeze-out parameters $(T_{\txt{fo}},\mu_{\txt{fo}})$, while dark purple stars denote points where $1-\kappa_3\kappa_1/\kappa_2^2$, taken along lines informed by average phase diagram trajectories for STAR collision energies \cite{Shen:2020jwv}, matches the experimentally measured values for a given collision energy. The softening of the EOS, leading to negative values of $\big(\frac{d\ln c_T^2}{d\ln n_B}\big)_T$, occurs in two regions of the phase diagram, corresponding to the ordinary nuclear matter phase transition and to the conjectured QGP-like phase transition.
}
\label{diagram}
\end{figure}
\section{Experimental data and interpretation}
We consider cumulants of the net proton number and chemical freeze-out parameters, $(T_{\txt{fo}},\mu_{\txt{fo}})$, in collisions at 0-5\% centrality, determined by the solenoidal tracker at RHIC (STAR) \cite{Abdallah:2021fzj} and high acceptance dielectron spectrometer (HADES) \cite{Adamczewski-Musch:2020slf, HADES_MLorentz_talk} experiments, and we use them to plot Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}) (red triangles, upper and lower panel in Fig.\ \ref{STAR_HADES_plots}, respectively) against $\mu_B$.
Based on the previous section, we trust the results presented in the upper panel of Fig.\ \ref{STAR_HADES_plots}, approximating $c_T^2$, only for the lowest collision energy, $\sqrt{s}=2.4\ \txt{GeV}$ from the HADES experiment. Here, $c_T^2$ as obtained from Eq.\ (\ref{magic_equation_1}) is small: less than half of the ideal gas value. At the same time, the value of $1-\kappa_3\kappa_1/\kappa_2^2$ (shown in the lower panel in Fig.\ \ref{STAR_HADES_plots}), which we assume is dominated by $\big(\frac{d\ln c_T^2}{d\ln n_B}\big)_T$, drops with decreasing collision energy to reach a minimum at the lowest STAR point, $\sqrt{s}=7.7\ \txt{GeV}$, and then steeply rises for the HADES point. This could mean that in $\sqrt{s} = 7.7\ \txt{GeV}$ collisions, $c_T^2$ is approximately constant as a function of $n_B$, while in $\sqrt{s} = 2.4\ \txt{GeV}$ collisions the matter is characterized by a small $c^2_T$ which nevertheless has a large slope as a function of $n_B$.
To further understand this behavior, we study the dependence of $1-\kappa_3\kappa_1/\kappa_2^2$ on $\mu_{B}$ and $T$ within the VDF model, shown in Fig.\ \ref{diagram}. Comparing model results to experiment requires choosing at which $T$ and $\mu_B$ to take values of $1-\kappa_3 \kappa_1/\kappa_2^2$. A natural choice is to use values at $(T_{\txt{fo}},\mu_{\txt{fo}})$ (light green stars), but as shown in the lower panel of Fig.\ \ref{STAR_HADES_plots}, these values do not lead to an agreement with experimental data, with the biggest discrepancy for the HADES point. (We note that $(T_{\txt{fo}},\mu_{\txt{fo}})$ are established with hadron interactions neglected, and the degree to which this affects our results may vary across the phase diagram.) However, critical fluctuations exhibit a large relaxation time \cite{Berdnikov:1999ph,Stephanov:2017ghc,Du:2020bxp}, and their measured values could be affected by stages of the collision preceding the freeze-out. With this insight and taking guidance from average phase diagram trajectories of hybrid simulations of heavy-ion collisions \cite{Shen:2020jwv}, we consider $1-\kappa_3\kappa_1/\kappa_2^2$ at slightly earlier stages of the evolution. In this way we obtain values of $1-\kappa_3\kappa_1/\kappa_2^2$ (dark purple stars) that reproduce the experimental results for a given collision energy (lower panel in Fig.\ \ref{STAR_HADES_plots}); here the exception is the HADES point, for which we prioritized choosing a point in a reasonable vicinity of the measured freeze-out over obtaining a value equal to the experimental data. Comparing to the exact model results for $c_T^2$ and $c_T^2 + \big(\frac{d\ln c_T^2}{d\ln n_B}\big)_T$, also displayed in Fig.\ \ref{STAR_HADES_plots}, as well as to the upper left panel of Fig.\ \ref{tests_of_formulas}, we can confirm that at the point reproducing the experimental value of $1-\kappa_3\kappa_1/\kappa_2^2$ for the lowest STAR energy, $c_T^2$ is nearly constant as a function of $\mu_B$ (thick short-dashed line, $T=150\ \txt{MeV}$, at $\mu_B\approx410\ \txt{MeV}$ in Fig.\ \ref{tests_of_formulas}), while at the point reproducing the result for the HADES energy, $c_T^2$ increases sharply with $\mu_B$ (thick solid line, $T=50\ \txt{MeV}$, at $\mu_B\approx850\ \txt{MeV}$ in Fig.\ \ref{tests_of_formulas}).
Naturally, the choice of $T$ and $\mu_B$ at which we compare model calculations with STAR and HADES cumulant data is driven by the wish to match the experimental results, and it serves mainly to show that baryon number cumulants measured in heavy-ion collisions can be connected to the speed of sound in hot and dense nuclear matter. Whether values of higher order cumulants are indeed significantly affected by stages of the evolution preceding the freeze-out needs to be further investigated (for recent developments, see \cite{Nahrgang:2018afz,Jiang:2017sni}). Moreover, while experiments measure proton number cumulants, the VDF model provides baryon number cumulants, putting more strain on our interpretation. Baryon number conservation should likewise be important \cite{Bzdak:2012an,Vovchenko:2020gne}. Finally, our model results may not be applicable in regions of the phase diagram where quarks and gluons become increasingly relevant.
Nonetheless, hadronic models are well-justified for describing low-energy collisions whose evolution is dominated by the hadronic stage. The comparison between the experimental data and the VDF model suggests that collisions at the lowest STAR and HADES energies may be probing regions of the phase diagram where the cumulants of the baryon number tell us more about hadronic physics than the QCD CP. In particular, the change in the sign of $\kappa_3$, predicted to take place in the vicinity of a critical point \cite{Asakawa:2009aj} and apparent in the HADES data (see lower panel of Fig.\ \ref{STAR_HADES_plots}), may mark the region of the phase diagram affected by the nuclear liquid-gas phase transition. If this is true, it may be worthwhile to study the cumulants at even lower collision energies, starting from $0.1\ \txt{GeV}$ projectile kinetic energy, and obtain the speed of sound around the nuclear liquid-gas CP. Conversely, at higher energies it could be possible to use collisions at different centralities and different rapidity windows to estimate the neglected terms in Eqs.\ (\ref{magic_equation_1}) and (\ref{magic_equation_2}), and obtain a stronger estimate for the speed of sound in the respective regions of the phase diagram.
\section{Summary and conclusions}
In this work, we use cumulants of the baryon number distribution to estimate the isothermal speed of sound squared and its logarithmic derivative with respect to the baryon number density. This result provides a new method for obtaining information about fundamental properties of nuclear matter studied in heavy-ion collisions, with consequences for both the search for the QCD CP and neutron star studies. While the approximations and the model comparison we considered apply to experiments at low energies, the approach itself can be used at any collision energy provided that measurements of cumulants of baryon number distribution as well as their temperature dependence are available. Further studies of effects due to dynamics, in particular using state-of-the-art simulations, will be absolutely essential in determining the extent to which the proposed method provides a reliable extraction of sound velocities and their derivatives.
{\hspace{10mm}}
\section{Acknowledgements}
A.S.\ thanks Chun Shen for providing average phase diagram trajectories of hybrid heavy-ion collision simulations at RHIC Beam Energy Scan energies.
A.S.\ and V.K.\ received support through the U.S. Department of Energy,
Office of Science, Office of Nuclear Physics, under contract no.\
DE-AC02-05CH11231231 and received support within the framework of the
Beam Energy Scan Theory (BEST) Topical Collaboration. D.O.\ and L.M.\ were supported by the U.S.\ DOE under Grant No.\ DE-FG02-00ER4113.
|
\section{Introduction}
Food fraud has become a issue on a global scale for producers, consumers, governments and other involved actors \cite{Galvez2018}. In fact although it is not a new problem - fraud and adulteration of food were widespread during ancient times so much so Roman laws included edicts against the commercial frauds or adulteration - unfortunately appears rising in the last years and is a threat to public health and the economy.
According to \citet{spink2011defining} \emph{``Food fraud is a collective term used to encompass the deliberate and intentional substitution, addition, tampering, or misrepresentation of food, food ingredients, or food packaging; or false or misleading statements made about a product, for economic gain.''}
The globalization of the markets reduces the food traceability and so facilitates food fraud. Consequently, food safety can be increased through higher traceability \cite{Galvez2018}.
In order to identify and address food frauds worldwide, it is essential to trace and authenticate the food supply chain to understand provenance and to ensure food quality in compliance with international standards and national legislation.
In the literature several definitions of traceability and of traceability system can be found \cite{Olsen2013}, \cite{Bosona2013}, \cite{Galli2015}, \cite{Karlsen2013}, \cite{Dabbene2014}.
Food traceability is \emph{``the ability to follow the movement of a feed or food through specified stage(s) of production, processing and
distribution''} \cite{ISO22005_2007} as defined by the International Organization for Standardization in ISO 22005:2007 - a specific standard for traceability in the food and feed chain. In addiction, a traceability system is based on product labelling and contains information, both quantitative and qualitative, about the final product and its provenance.
In Europe, food legislation is particularly strict and the implementation of traceability systems are mandatory but are unable to insure consumers against fraud. For this reason innovative methods for traceability systems based on product identification are needed.
Deterring food fraud requires interdisciplinary research combining food science, food law, supply chain management, and other fields such as informatics, mathematics and statistics \cite{DANEZIS2016123}.
In this context, a distributed ledger technology such as Blockchain (BC) provides a full and immutable audit trail of transactions data for all stages of the food supply chain allowing for transparency and verifiable and immutable records in the form of digital certificates.
Immutability of the data enables the technology to be considered for regulated industries such as agri-food.
As reported by \citet{Galvez2018} in his research paper most blockchain systems for traceability management have been developed since 2015.
Numerous research papers, \citet{Demestichas2020}, \citet{costa2013review}, \citet{tian2016agri}, \citet{Antonucci2019} just to cite a few, have shown that using blockchain can advantageously help to achieve traceability by irreversibly and immutably storing data.
Although many studies are promising that the application of blockchain technologies to the agri-food supply chain can ensure the food traceability and some companies have launched pilot or proof of concept projects to manage their supply chains with blockchain technology, certain limits remain to be considered and addressed. Most of the published works concerning the application of blockchain technologies to the supply chain management have reported no detailed information about the technical implementation and there are still few uses to the state of practice (See Section \ref{sec:relatedWork}).
For this reason we investigated the potential applications of blockchain technologies in the agri-food sector both from some large companies that have declared their intention to develop pilot or proof of concept projects and several research papers in literature about the possibility of applying such technologies in agri-food domain. Our investigations indicated us a main direction of research: devising new methods for agri-food supply chain implementation processes based on blockchain technology. In our opinion research in this topic area should investigate how blockchain-based processes for implementing the agri-food supply chain traceability can be efficiently specified and deployed. Among others, formal methods to facilitate and make more efficient the development of applications on blockchain technologies for the management of the agri-food supply chain. Researchers and developers could benefit from a general-purpose approach for agri-food supply-chain management.
In light of these considerations and to better understand this technology and generate new user-friendly implementations, the primary purpose of this paper is to propose a blockchain-based methodology to facilitate and make more efficient the development of blockchain applications for the agri-food supply chain management, by building configurable blocks to be assembled together, so as not to start from scratch every time. Finally, we evaluated the proposed approach using a specific case study: a traceability system for the olive oil supply chain.
\begin{comment}
The main contributions of this work are summarized as follows:
\begin{itemize}
\item To investigate and define the roles of major actors involved in the system;
\item To determine the entities that emerge and recur in this type of system;
\item To determine the relationships between these entities and the events of interest for the traceability system (events that need to be made permanent for the traceability);
\item Starting from the aforementioned points, to propose configurable and modular building blocks for agri-food supply chain management systems which can be represented on Blockchain.
\item To evaluate the proposed approach using a specific case study: a traceability system for the olive oil supply chain.
\end{itemize}
\end{comment}
This methodology can help to compose Smart Contracts (SCs) specific in an easy way, so as to be able to reuse the code and modules and automate the process to make it fast, keeping its safety.
As far as we know, this is the first attempt to develop a semi-automatic configurable system that supports the entire class of supply chains for agri-food industrial domain.
The rest of the paper is organized as follows. Section \ref{sec:relatedWork} deals with the background and related work. Section \ref{sec:Methodology} presents the proposed methodology, followed by Section \ref{sec:Case study} including a case study on a traceability system for the olive oil supply chain to evaluate the proposed methodology. Section \ref{sec:Conclusions} makes a brief conclusion for this paper.
\section{Background and Related Work} \label{sec:relatedWork}
In this section, we first introduce an overview on decentralized application and smart contracts. Then, we briefly introduce some background knowledge of related work which leads the readers to better understand the issue.
A decentralized application, or DApp, is a computer application that runs on a distributed system, that is on a network of nodes, with no node acting as supervisor. In the context of blockchain peer-to-peer network, a DApp is stored and executed on the blockchain, in order to be decentralized, transparent, deterministic and redundant. A DApp is developed by writing Smart Contracts (SCs), which are small script programs running on every node of the blockchain and may have a User Interface (UI) that allows users and devices to interact with them.
SCs are immutable - no one can tamper with the code of the contract - and distributed, because of their storage inside the blockchain
Smart contracts have been developed and used in several fields \cite{zheng2018blockchain}; here we focus on agri-food industrial supply chains.
Based on various definitions of traceability systems (\cite{Olsen2013}, \cite{Bosona2013}, \cite{Galli2015}, \cite{Karlsen2013}, \cite{Dabbene2014}, \cite{ISO22005_2007}, just to cite a few), the main purposes of a agri-food traceability system are:
\begin{itemize}
\item To document in a transparent and irreversible way any event relevant to production;
\item To allow authorities, laboratories and certified experts to asseverate the production, giving proof of their identity and their certifications;
\item To integrate manual registrations and automatic registrations, made by Internet of Things (IoT) devices that are increasingly widespread;
\item To keep track of the quantities produced, so that these cannot be increased by introducing products of non-certified origin;
\item To give evidence of all stages of production to the authorities responsible for verifying the specifications;
\item To allow retailers and end consumers to learn about the history of the products purchased, from the field to the purchased product, using an app.
\end{itemize}
Traditional approaches for the agri-food traceability systems are usually inefficient. Blockchain is a promising technology that is tamper-proof and decentralized and self-executing and self-verifying smart contracts are able to conduct transactions between mutually untrusted parties. In this context scholarly literature on the adoption of the blockchain technology in specific traceability systems for the agri-food supply chain is beginning to emerge.
In particular, since 2018, a lot of research efforts have been made on the use of blockchain technology for traceability systems. \cite{Demestichas2020}
Since 2018, Tripoli and Schmidhuber highlighted the potential of distributed systems to transform the agri-food industry \cite{tripoli2018emerging}. \citet{tribis2018supply} searched for relevant papers about the adoption of blockchain technology for generic supply chain management, covering the literature from all years to 2018, and found out that most of the papers were focused on the use of BC for traceability, but only one out of 40 papers dealt with agricultural field. \citet{chang2020blockchain} claim that by 2023 the global blockchain supply chain market will grow to \$ 3,314.6 million, recording an increase in the annual growth rate of 87\% .
Various papers presented real blockchain solutions for supply chain management, which proved to be successful. Among others, AgriDigital \footnote{https://www.agridigital.io/}, an Australian system for managing grain supply chain released in 2017, counts at the end of 2020, more than 7000 users and a transaction value of \$ 3,793 million.
\citet{caro2018blockchain} developed AgriBlockIoT, a fully decentralized, blockchain-based traceability solution for agri-food supply chain management, able to seamless integrate IoT devices, using and comparing both Ethereum and Hyperledger Sawtooth blockchains. \citet{tian2016agri} studied and developed an agri-food supply chain traceability system for China based on RFID (Radio-Frequency Identification) and blockchain technology, to guarantee food safety. \citet{baralla2019ensure} proposed a generic agri-food supply chain traceability system based on blockchain technology implementing the "from-farm-to-fork" (F2F) model currently used in the European Union, which can integrate current traceability rules and processes, using Hyperledger Sawtooth, and following an agile approach.
\citet{wang2019smart} propose a product traceability system based on blockchain technology, in which all product registration and transfer histories are perpetually recorded by using smart contracts and an event response mechanism was designed to verify the identities of both parties of the transaction and guarantee the validity of the transaction.
\citet{Yu2020a} propose a monitoring framework that combines smart contracts and evaluation models for the automatic evaluation of the quality of fruit juice samples. Smart contracts are executed to record production data on a blockchain and can decide whether the production process can be resumed or not. The feasibility of the system has been evaluated by implementing a prototype version of the quality monitoring system for flat peach juice production based on the Ethereum platform and executed in the Remix IDE.
Many studies in the literature have focused on highlighting the benefits and value derived by the blockchain implementation in agri-food supply chain domain.
In particular, \citet{kamble2020modeling} conducted a analysis in order to model a traceability system based on blockchain technologies in the agriculture supply chain. They identified thirteen enablers that encourage the blockchain adoption in agriculture supply chain, such as anonymity and privacy, immutability, smart contracts, secured and shared database, traceability, transparency, and others. Then they established hierarchical levels and relationships between the involved actors in the supply chain through interpretive structural modeling (ISM) and decision-making trial and evaluation laboratory (DEMATEL) methodologies. The enablers were identified from existing literature and validated by experts from the field of agrobased supply chain and technology.
Moreover the authors conducted an interesting literature review revealing that BC technology offers various benefits leading to improvements in the sustainable performance of the agriculture supply chains.
Based on previous research papers (\cite{tian2016agri}, \cite{tian2017supply}, \cite{kamble2020modeling}, \cite{wang2019understanding}, \cite{Li2018}), just to cite a few, by using a blockchain the traceability system guarantees:
\begin{itemize}
\item Data integrity and provenance of documents and records on blockchain;
\item Immutability and transparency of data recorded on the blockchain;
\item Respect for the quantities of the products involved (grapes, wine, bottles), based on the annual production of the land and the yield in the various stages of processing. This is achieved with the system of tokens, which are associated with the various products and cannot be altered as they are managed on the blockchain;
\item Buyers take charge of their products, with complete traceability of the supply chain;
\item Ability to retrace the entire supply chain, simply by accessing the blockchain and public servers with documents, starting from the QR code shown on the final product.
\end{itemize}
To the best of our knowledge, and by comparing our work to the others that deal with the use of DApps and blockchain to certify the origin of food and prevent food fraud, we found out that this is the first paper that proposes the use of a self-generating system.
\section{Methodology} \label{sec:Methodology}
We studied the agri-food production domain, and we found that virtually all agri-food supply chain systems start from one or more primary sources (soil, herd, beehives, lake, sea, and so on.), produce a primitive resource (harvest, milk, raw honey, fishes, etc.), then this product is transformed possibly several times, until the final product; events relevant to the process can occur in each of the phases.
This problem is too complex to be tackled top-down, therefore we adopted a bottom-up approach. Using analysis techniques and object-oriented design, we performed the analysis of the supply chain domain in order to:
\begin{itemize}
\item Determine the actors, concepts and entities that emerge and recur in this type of systems;
\item Determine the relationships between these entities;
\item Determine the events of interest for tracking and for the generation of new products (transformations, splitting, merging);
\item Design configurable and modular SCs/building blocks, representing the identified entities and relationships, which can then be represented on the blockchain;
\item Semi-automatically map the user interface needed to interact with the system, describing the entities, events and transformations.
\end{itemize}
In this section we present our approach and we show how the traceability system is then generated automatically, saving time and costs. The entire software process was carried on following an agile approach such as ABCDE (Agile BlockChain Dapp Engineering) \cite{marchesi2020abcde}.
\subsection{Actors}
We define the key actors of an agri-food system. Each actor is identified by a unique address able to send transactions to the blockchain. The actor owns the private key associated with the address, thus being the only person able to send messages from that address.
\begin{itemize}
\item \textbf{Administrator/Owner}: Administers the system, managing the addresses of the other actors and their permissions.
\item \textbf{Producer}: Produces the raw materials that are the inputs of the supply chain.
\item \textbf{Supplier}: Supplies materials, services or devices needed for producing the target good.
\item \textbf{Transformer}: Transforms raw material, or already transformed material, into intermediate goods or into the target good.
\item \textbf{Wholesaler}: Buys the target good in large quantities, and sell it to other wholesalers or retailers.
\item \textbf{Retailer}: Buys the target good in order to sell it to End Customers.
\item \textbf{End Customer}: Buys the target good from Retailers. S/he is usually not provided with an address.
\item \textbf{Certification Authority}: Public or private authority in charge of controlling and certifying a given production.
\item \textbf{Professional}: A person with a given degree and experience, qualified to certify the goodness of amounts and documents.
\item \textbf{Analysis Lab}: A laboratory able to perform physical, chemical and/or biological analysis of given materials, products or goods.
\item \textbf{Warehouse}: Receives, stores and sends goods.
\item \textbf{Device}: An IoT device connected to the Web, able to send transactions with measurements relevant to the supply chain.
\end{itemize}
\subsection{Entities}
The main entities involved in an agri-food supply chain are:
\begin{itemize}
\item \textbf{Address Catalog}: for each address, the identity and the role(s) of the owner are specified; the catalog is managed by the Administrator/Owner of the system.
\item \textbf{Producer}: a farmer or a firm producing or transforming agri-food products. Its representative(s) are identified by blockchain addresses.
\item \textbf{Productive Resource}: it represents something that produces the main raw agri-food products. Typically, it is a field (orchard, vineyard, wheat field, vegetable garden, greenhouse, olive grove, etc.), a group of animals (flock, herd, poultry, etc.), a set of beehives. It is owned by a Producer. The system can hold information and documents on it, and can register events related to its cultivation or farming.
\item \textbf{Product}: it represents a production lot of something that comes from a Productive Resource, or from the transformation of other products. For instance, the grapes are produced from the vineyard, the must from the pressed grapes, the wine from the must. It is linked to a Producer which is in charge of its processing. The system can hold information and documents on it, and can register events related to its processing.
\item \textbf{Token}: a given quantity (a number) created and assigned to a given address. The token represents the ownership of a specific amount of material, good or asset. It can be split, and transferred to other address(es). Since the token, once created, cannot be increased, it guarantees that only the original material/good/asset is managed in the system. Many kinds of tokens can be managed by the system.
\item \textbf{Notarization} of documents assessing the process (treatments, harvest, chemical analysis, quantity produced in subsequent steps, etc.). This is a known pattern [10]. The original document must be kept on a server, and available to download to authorized users. The notarization must include:
\begin{itemize}
\item hash of the document;
\item registration date (always available as date and time of the transaction);
\item link to the document in the server, or information on how to access it;
\item possible metadata.
\end{itemize}
\end{itemize}
\subsection{Main Events}
There are two types of events:
\begin{enumerate}
\item \textbf{Transformation events}, which create a product starting from one or more resources, transform one or more products into one or more others, or divide the product into several sub-products that are the same, but of lower quantities;
\item \textbf{Documentation events}, which associate data related to the production process to a product (or resource), but do not transform it and do not create other products.
\end{enumerate}
The main events managed by the system are:
\begin{itemize}
\item \textbf{Asseveration}: a given entity stored on the blockchain (hash of a document, data, etc.) is certified by a transaction sent from an address of a person/body able to certify. Also the correctness of product and token creation or transformation can be asseverated.
\item \textbf{Creation of tokens} associated to some product of the process, at a given date. The creation can be provided of further data about the physical product associated to the tokens, and even of the notarization of documents attesting the truthfulness of the creation.
\item \textbf{Product Merging}: the act of merging products of the same kind, invalidating them and producing products of the same kind. Different lots of the same material can be merged, producing one or more new lots. The tokens associated to the products must be burned, and new tokens associated to the new products are created, preserving the overall number of tokens.
\item \textbf{Product Splitting}: the act of splitting one product into two or more of the same kind, invalidating the original and producing products of the same kind. Also in this case the tokens associated to the original product must be burned, and new tokens associated to the new products are created, preserving the overall number of tokens.
\item \textbf{Product Transformation}: the act of merging one or more products into one or more products of different kinds, invalidating the original one(s) and producing products of one or more other kinds. For instance, grapes can be transformed to must, used to produce wine, but also to marc, used to produce grape pomace brandy. Also in this case the tokens associated to the original products must be burned, and new tokens associated to the new products are created, preserving the overall balance of quantities. The transformation is often associated to asseveration events.
\item \textbf{Data Registration}: a record holding specific data is stored in the blockchain by a given address, guaranteeing the date and the actor who stored it. For instance, fertilization, pesticide treatments and pruning are recorded as events linked to a field (Productive Resource).
\item \textbf{Notarization Event}: the data concerning a Notarization (see previous section) are registered by a specific address, ensuring date and inalterability of the documentation, and signature of the registrant.
\item \textbf{Certification of data or of a notarization}: a third party certifies, with a transaction coming from its address, the correctness of a data registration, or of a notarized document.
\item \textbf{Unlocking}: one or more transactions from given addresses are needed to unlock a process, that is to register another event (typically, a transformation event).
\item \textbf{Payment}: a payment (in cryptocurrency) is made available to a given address.
\end{itemize}
\subsection{Smart Contracts}
After identifying the actors, concepts and events related to the process, the system SCs are created.
Depending on the specific use case to be implemented, the
developer can design either dynamic, static, or oracle driven
smart contracts.
At the moment in our system the smart contract is a static file regardless of the use case (olive or wine or honey supply chain use case, just to cite a few) listing all the relevant fields for each concept, as described in the following section. The SCs are automatically generated from the JSON file. We will have a SC for each producer (\texttt{Producer}), and a SC for each resource (\texttt{ProductiveResource}) and for each product being processed or transformed (\texttt{AgriProduct}), which takes into account the registrations (\texttt{AgriEvent}) on it. Since a \texttt{ProductiveResource} and an \texttt{AgriProduct} share several data and operations, both inherit from \texttt{AbstractResource} abstract SC.
The \texttt{Producer} contains a mapping with links to all productive resources and products owned by, or related to it. Products are generated or transformed starting from one or more productive resources, or from one or more products. Each product tracks its origin or origins through an array of addresses of the upstream resources or products. An \texttt{AbstractResource} holds a list of events \texttt{AgriEvent}, which in turn contain a list of data related to the event (all encoded in the string of bytes named "parameters"), to flexibly attribute data to the events, as per the above approach for the GUI.
A QR code on the final product allows to find the related \texttt{AgriProduct}. From here you can find all the events in (reverse) registration order and, for each event, all relevant data, including links to documents and their possible hash digest. If the \texttt{AgriProduct} has one or more origins, by going to these further origins (\texttt{AgriProduct} or \texttt{ProductiveResource}), all upstream products/resources can be found, with the related events, and so on. At the end of each product, you can also access its \texttt{Producer}.
Backward navigation must be thoroughly designed, to make it easier for multiple origins. If necessary, a list of generic data could also be added to the \texttt{AbstractResource}, encoded in a string of bytes, as in the "parameters" field of \texttt{AgriEvent}. The relationship between product and upstream products/resources is navigable in both directions. An \texttt{AgriProduct} contains the array "origins", with the addresses of the contracts of type \texttt{ProductiveResource} or \texttt{AgriProduct} that created it (the \texttt{ProductiveResource} has no origins). This is a list created during the creation of the \texttt{AgriProduct}, and cannot be modified. A generic product/resource also contains a “produced” array with the addresses of the generated \texttt{AgriProduct}, for Creation, Transformation, Division or Contribution. This array can be updated, typically by appending the address of a newly created \texttt{AgriProduct}.
It is recommended to refer to some best practices given by researchers or communities also to avoid vulnerabilities.
In order to optimize the code, we use OpenZeppelin \cite{OpenZeppelin} - a toolkit to develop, compile, update, deploy and interact with Smart Contracts. We also systematically applied gas saving patterns \cite{marchesi2018gas}.
Figure \ref{fig:UML} shows the UML class diagram representing the Smart Contracts used in our system.
\begin{figure}[h]
\centering\includegraphics[width=1\linewidth]{images/UML_Class_Diagram.png}
\caption{UML class diagram representing the Smart Contracts used in our system.}\label{fig:UML}
\end{figure}
\subsection{Data Types}
All product and events are characterized by data which are the same in every agri-food supply chain, such as type, name, unique identifier, owner. However, they may hold many more data, specific of the particular production, and production process.
Since our goal is to develop a general-purpose system, able to be configured for every agri-food production process, we use a flexible data representation. Besides the above quoted common data, each element can be provided of a list of data, each represented by a string. The string representing a data includes three substrings giving the name of the data, its type, chosen among a set of common types, and its value. The types allowed by the system are: int, float, string, text (multiline string), enum, link, hashlink (link to an hashed file, stored off-chain), upload (it is possible to upload a local file or a photo to a server, keeping its link on the blockchain), hashupload (same of upload but the file is firstly hashed, and its hash digest is stored in the blockchain).
For instance, the harvest event could have associated the list of parameters that describe how many kilograms were collected (type int), the grape variety (type string) and even a photo taken during the harvest (type hashupload).
\subsection{User Interface}
The entities and the events of an agri-food supply chain, as defined in the sections above, are standard, and are able to cover virtually every production process. Consequently, also the applications enabling their input, editing (when allowed), and retrieving, are standard, and their user interface (UI) can be automatically generated starting from the description of the system.
More precisely, once the events of a specific production process are defined, with their data, constraints and authorizations, it is possible to automatically generate an app (a HTML5 Web page provided of Javascript code) able to create and edit the event. Clearly, the style and the appearance of the UI can be customized, but the data input, with all proper checks, do not require further programmer's intervention.
Also the navigation among the events, the products and the productive resources can be automatically programmed, starting from a QR code written on the final product. This navigation does not require a blockchain address, and can be performed by every customer.
Figures \ref{fig:HomePageOperator} and and \ref{fig:NavigazioneEventi} show some possible screenshots of the application, automatically generated.
\begin{figure}[h]
\centering\includegraphics[width=1\linewidth]{images/homepage_operator.png}
\caption{Homepage of an operator, menu and selection of the recordable events.}\label{fig:HomePageOperator}
\end{figure}
\begin{figure}[h]
\centering\includegraphics[width=1\linewidth]{images/EventHistoryNavigation.png}
\caption{Recording an event and navigating a product's event history.}\label{fig:NavigazioneEventi}
\end{figure}
\section{The Case Study: a blockchain traceability system for the olive oil supply chain} \label{sec:Case study}
The object of the case study is a blockchain traceability system to trace the olive oil production, from the agricultural treatments, to the olive harvest, to the oil mill, to the bottled oil, until the sale of bottles to a wholesaler and eventually to shops and supermarkets.
The choice is not accidental because olive oil and in particular Extra-Virgin olive oil remains today one of the most fraudulent food products.
The architecture of the system includes a cloud space as off-chain data repository. In this case, it has been used a decentralized technology, with Ipfs, an App to send, register and read events (written in Javascript using Angular/Ionic framework), the use of a real blockchain, such as Ethereum testnet, with app to send transactions or read from blockchain (Javascript, web3.js, and so on) and finally Smart Contracts written in Solidity.
A well-known issue with DApps developed on Ethereum-like blockchains is the cost of gas, which is specially related to the amount of Storage memory used. Solidity provides different types of memories. Those of interest to our application are:
\begin{itemize}
\item Storage: it is the persistent memory, and the most expensive;
\item Event-log: it stores on the blockchain data related to the Events raised by SCs (these Events are the syntactic constructs of Solidity). These data are cheap, but can be accessed only by external applications.
\end{itemize}
The information related to actors, entities and events may be too big. To optimize gas consumption, we save on Storage only simple data, keeping the large documents off-chain and storing here only their hash code and a link to retrieve them. The data related to the agri-food production events are stored in the Event-log, because they are not needed by the SCs of the system, but are only read by users' apps.
On average, this information storage choice reduces the gas consumption to 1/5, compared to using only the Storage.
Through a GUI, the system allows system actors to record events related to food production, and then a user to navigate through these events over time.
All the concepts of the system are cataloged in a specific JSON file, with the relevant associated fields.
To make the system usable even for users with a low level of experience, these JSON files can be created through an easy-to-use back-office application.
In particular, we have a JSON files for:
\begin{itemize}
\item The list of actors. For each actor id and name fields are specified.
\item The list of the companies involved in the process (producer, processor or retailer). For each company name, resources (the list of ids of resources and products managed by that company), authorized actors (the list of ids of the actors authorized to record events for products related to that company) are specified.
\item For each type of productive resource or agro-industrial product managed by the production process, the fields id, type (R for resource or P for product), name, authorized actors (the list of ids of the actors authorized to record events for that product) are specified. Further fields can be automatically enhanced by the system when creating specific products: description, quantity (a positive integer describing the quantity of the product or the size of the productive resource), unit of measure (mandatory if quantity is used).
\item The list of relevant events.
Events are always associated with a product or a resource. For each event we specify id, name, products/resources (the list of products/resources ids that can be associated with that event), authorized actors (the list of ids of the actors authorized to record that event), type (D for documentation, T for transformation), generated products (the list of product ids that can be generated by that event), parameters (list of names and allowed types for the parameters associated to that event).
\item The list of activities that allows to customize the application template, based on the company and on the users.
\end{itemize}
Each of these files contains a key that determine the management of its content by the application.
\begin{figure}[h]
\centering\includegraphics[width=1\linewidth]{images/OliveOilProduction.png}
\caption{Olive oil production flow.}\label{fig:OliveOilProduction}
\end{figure}
Figure \ref{fig:OliveOilProduction} shows the simplified olive oil production flow. The agro-industrial products are denoted as rectangles, while the associated events as "documents" in the flow chart notation (rectangles with wavy bottom side), possibly multiple.
Transformation events (denoted in the figure as ET), always single, are associated near the right side of products or resources, and cause them to be transformed into other products, or divided into batches. In the shown example, the olives come from two olive groves (1 and 2). Once harvested with a single event (Harvest) they are divided into two lots (Olive 1 and Olive 2), immediately before milling, which takes place separately for the two lots. Consequently, they have no other associated description events.
The flow of agro-industrial products, from left to right, can be represented by a tree: each product can have more "fathers", and more "sons". Consequently, the "navigation" backwards, starting from the last product, is not unique and must be managed in the user interface.
\section{conclusions and future works} \label{sec:Conclusions}
Traceability systems are considered important to ensure the safety of a food product and prevent food fraud in the food supply chain.
System based on Blockchain and smart contracts for monitoring, integrated with the Internet of Things, allow a traceability system to occur without third party which usually is responsible for controlling the agri-food supply chain, transparency of data and indication of the origin of the products, as well as allowing the end customer to control the characteristics of the product purchased.
Specifically the advantages of the management system of an agri-food supply chain via blockchain, generated in a semi-automatic way, are:
\begin{itemize}
\item The consumer can be sure of the origin, the production process and the quality of the product purchased.
\item The Authority in charge of production control is facilitated in the controls, and can reduce on-site inspections.
\item The manufacturer can certify in a simple and non-falsifiable way all the steps of a production.
\item Development times and costs are reduced, while maintaining a high level of security and trust.
\item Blockchain system can also manage contractual transactions and payments, in this case, the system could also be extended to payments and the transfer of ownership of a product could also be marked by a cryptocurrency transfer.
\end{itemize}
To the best of our knowledge, this is the first attempt to automatically develop custom DApps for agri-food supply chain, by building configurable SCs to be assembled together and we truly believe in the innovation of the proposed method. Moreover, it was actually used to develop different real systems, thus confirming its efficiency.
Note that our flexible approach may lead to high gas cost in creating and updating SCs on Ethereum blockchain. A data represented as a string is way more costly than a native data, both in terms of storage and of computations needed to code and decode it. For this reason, we advise to use a permissioned blockchain publicly accessible for reading as a backbone. Future works may include the creation of a DSL (Domain Specific Language), the generalization to generic and non-ethereum-like blockchains, and also the generalization to other kind of supply chains.
\bibstyle{IEEEtran}
|
\section{Introduction}
This work is on the interface between Computer Science in the area of quantum computing, which is based particularly on a diagrammatic formulation of ZX calculus\cite{CD}, and the theory of Hopf algebras or quantum groups as understood in algebra and mathematical physics. The preliminaries in Section~\ref{secquaZX} start with a concise but self-contained reworking of the central algebraic structure, namely that of a Frobenius Hopf algebra or `interacting pair of Hopf algebras', but without assuming that our algebras are commutative or coalgebras cocommutative or our Frobenius forms symmetric. Here, Corollary~\ref{corHH*} revisits a recent result in \cite{CoD} that every finite-dimensional Hopf algebra $H$ `amplifies' to such a pair, but with a more direct treatment that extends to the braided case and identifies the other member of the pair as $H^{*op}$. We keep track of braid crossings in the diagrammatic proofs as preparation for the braided theory later, and we compute details for the example of $u_q(sl_2)$ at $q$ a primitive $n$th root of unity. The simplest case is $n=2$ which, in our conventions from \cite{AziMa}, is noncommutative and noncocommutative even for $q=-1$ and given in detail.
Section~\ref{secstar} introduces $*$-structures on Frobenius and Frobenius-Hopf algebras. The nicest assumptions are where $H$ but require unimodularity. For $u_q(sl_2)$ we show how this gets modified, with its natural $*$ operation making it a {\rm flip}-$*$ Hopf algebra. Phases and an abstract formulation of a `Hadamard gate' are also discussed, the latter in three types. Here, Type 1 is the more obvious one but Type 2 is equivalent to $H$ self-dual as a Hopf algebra and Type 3 to $H$ anti-self dual, which are therefore more natural from a Hopf algebra point of view. $u_{-1}(sl_2)$ and the reduced Taft algebra $u_q(b_+)\subset u_q(sl_2)$ provide examples.
The main new result of the paper is in Section~\ref{secbraZX}, where we construct interacting Hopf algebras and ZX calculus in a braided tensor category. The general result in the braided case that every braided Hopf algebra with suitable integrals amplifies to a braided Frobenius-Hopf algebra (Corollary~\ref{corbraHH*}) is read off from the diagrammatic proofs in Section~\ref{secquaZX}, so that the main additional work is the construction of examples. We do this (Proposition~\ref{proptransH}) by a process of transmutation~\cite{Ma:bra,Ma:tra} whereby every quasitriangular Hopf algebra in the sense of Drinfeld\cite{Dri,Ma} has a braided version. We include detailed computations for $b_q(sl_2)$ obtained by transmutation from $u_q(sl_2)$. Our Type 2 `braided hadamard gate' $\hbox{{$\mathfrak h$}}$ here is essentially the braided Fourier transform for this class of braided Hopf algebras, shown in \cite{LyuMa} along with the ribbon structure to obey the modular identities needed for topological invariants.
We conclude this introduction with a little of the background from Computer Science. In quantum computing, any unitary operator can be regarded as a `quantum gate', but the useful content is to identify a specific collection of such `gates' from which others can be built or approximated to arbitrary accuracy. To this end, let $V=\mathbb{C}^2$ for a singe `qubit' system and consider unitaries on the $m$-qubit system $V^{\mathop{{\otimes}} m}$. Then some useful gates are the Hadamard gate
\[ \hbox{{$\mathfrak h$}}: V\to V,\quad \hbox{{$\mathfrak h$}}={1\over\sqrt 2}\begin{pmatrix}1 & 1\\ 1 & -1\end{pmatrix} \]
and two flavours of maps $\Delta_{\color{green}\bullet}:V\mathop{{\otimes}} V\to V, \mu_{\color{red}\bullet}:V\to V\mathop{{\otimes}} V$ (traditionally depicted with red and green nodes) which, up to normalisation, come from the product and coproduct of the group Hopf algebra $\mathbb{C}\mathbb{Z}_2$ and its dual Hopf algebra\cite{CD}. One can further include versions of the nodes modified by phases. The merit of this particular collection of `gates' is that the Hopf algebra structures allow one to simplify and rewrite compositions according to certain identities, leading to an efficient framework for working with them. Moreover, composition of these linear maps can conveniently be expressed as `string diagrams' where the copies of $V$ are written side by side and maps applied as for a flow chart or `quantum circuit' flowing down the page. In this case, the core identities between the `gates' appear more abstractly at the diagram level as a graphical ZX-calculus\cite{CD} and the role of $\mathbb{C} \mathbb{Z}_2$ is to realise this concretely. The role of $\mathbb{C}\mathbb{Z}_2$ and its dual can then be played by another `interacting pair' of Hopf algebras\cite{DD} namely $H$ and an associated Hopf algebra $\tilde H$ with green product $\mu_{\color{green}\bullet}$ and red coproduct $\Delta_{\color{red}\bullet}$.
In these string diagrams as used in ZX calculus, there is no significance to the over or under crossings of strings. On the other hand, the idea of doing algebraic operations at a string diagram level has already been familiar in algebra since the early 1990s as the theory of `braided groups' or Hopf algebras in braided categories \cite{Ma:bra,Ma:bg,Ma:tra,Ma:cro,Ma:alg,Ma}. Here the `wiring diagram' approach to algebraic identities was essential to keep track of under- and over- crossings, as these correspond to typically different operations. In the simplest `anyonic' example, the braiding is a power a parameter $q$ and the inverse crossing at the same point would be require $q^{-1}$. The theory was needed as a way to understand deeper aspects of $q$-deformation quantum groups\cite{Dri}. This body of work therefore begs the question as to whether ZX-calculus has a useful version at the level of braided tensor categories. Here, we show that at least for the core algebraic structures, namely the notion of a pair interacting Hopf algebras and Hadamard gate, this is indeed the case. The paper concludes with some remarks about further directions.
\subsection*{Acknowledgements} I thank the team at CQC for helpful discussions and references, notably B. Coecke, A. Cowtan, R. Duncan, A. Kissinger, Q. Wang. I also thank T. Fritz for a correction.
\section{Construction of noncommutative interacting Hopf algebras}\label{secquaZX}
This section is based heavily on \cite{CD,DD,CoD} and does not claim to be particularly new, but we are careful to ensure that where possible we do not assume that our (co)algebras $A$ are (co)commutative or that our bilinear forms are symmetric. Section~\ref{secpre}, in particular, is strictly preliminary and included only to be self-contained for readers from a more algebraic background. Section~\ref{secFHopf} is a reworking of \cite{CoD} but we go further with new more direct proofs and in which we keep track of braid crossings as preparation for Section~\ref{secbraZX}. We also strip out the formal categorical setting and focus on the content at an algebraic level.
We will write our maps composed flowing downward and use juxtaposition of objects to denote $\mathop{{\otimes}}$. Braidings are represented by braid crossings\cite{JoyStr} through which other morphisms can be pulled as part of the functoriality of the braiding. An introduction to doing actual algebra and Hopf algebra at this level can be found in \cite{Ma:alg,Ma,Ma:pri}. We will give main formulae in both conventional and diagrammatic form. However, we will denote coproducts by $\Delta$ as usual in Hopf algebra theory, not by $\delta$ as favoured in the Computer Science literature. Adjunction of morphisms is done by rotating anticlockwise in the plane, which it seems is the opposite of the more common convention in Computer Science. Purely diagrammatic proofs work in any braided category, but for examples built from Hopf algebras we will assume we are in the $k$-linear case, where $k$ is a field.
\subsection{Preliminaries on F-algebras}\label{secpre}
We recall that a unital algebra $A$ over a field $k$ is {\em Frobenius} if there is a bilinear form $(\ ,\ ):A\mathop{{\otimes}} A\to k$ which is nondegenerate and such that $(ab,c)=(a,bc)$ for all $a,b,c\in A$. The latter says that the Frobenius bilinear form descends to a map $A\mathop{{\otimes}}_A A\to k$ or equivalently a map $A\to k$. Nondegenerate as a linear map has its usually meaning which we write explicitly as existence of a {\em metric} $g=g^1\mathop{{\otimes}} g^2\in A\mathop{{\otimes}} A$ (sum of terms understood) such that for all $a\in A$,
\[ (a,g^1)g^2=a=g^1(g^2,a),\quad \includegraphics[scale=0.6]{dual.pdf}\]
where we also write the diagrammatic form with $(\ ,\ )=\cup$ and $g=\cap$. This makes $A$ a rigid object in the category of vector spaces and requires that $A$ is finite-dimensional.
\begin{lemma}\label{gcentral} In a Frobenius algebra, $g$ commutes with the algebra product.\end{lemma}
{\noindent {\bfseries Proof:}\quad } $g^1(g^2a,b)=g^1(g^2,ab)=ab=a g^1(g^2,b)$ for all $a,b\in A$, using the Frobenius and inverse properties. This implies $g^1(g^2a,g^{1'})\mathop{{\otimes}} g^{2'}=ag^1(g^2,g^{1'})\mathop{{\otimes}} g^{2'}$ or $g^1\mathop{{\otimes}} (g^2a,g^{1'})g^{2'}=ag^1\mathop{{\otimes}} (g^2,g^{1'}) g^{2'}$ and hence $ g^1\mathop{{\otimes}} g^2 a=a g^1 \mathop{{\otimes}} g^2$ for all $a\in A$. \endproof
In particular, the element $\mu(g)$, where we apply the product $\mu:A\mathop{{\otimes}} A\to A$ of $A$ to $g$, is in the centre, $\mu(g)\in Z(A)$. Next, recall that reversing the axioms of a unital algebra gives the notion of a counital coalgebra $(A,\Delta,{\epsilon})$ where the coproduct $\Delta:A\to A\mathop{{\otimes}} A$ is coassociative and the counit ${\epsilon}:A\to k$ obeys $({\epsilon}\mathop{{\otimes}}\mathrm{id})\Delta=\mathrm{id}=(\mathrm{id}\mathop{{\otimes}}{\epsilon})\Delta$. These have a standard diagrammatic representation in terms of tree diagrams and pruning identities\cite{Ma:alg}. The parallel notion in Computer Science to a Frobenius algebra is at first sight different and we will therefore denote it differently, as an F-algebra.
\begin{definition}cf\cite{CD}\label{defFalg} An F-algebra is a pair consisting of a unital algebra $(A,\mu,1)$ and counital coalgebra $(A,\Delta,{\epsilon})$ where $\mu=\includegraphics{prodfrag}$ and $\Delta=\includegraphics{deltafrag}$ are such that
\[ (\mathrm{id}\mathop{{\otimes}}\mu)(\Delta\mathop{{\otimes}}\mathrm{id})=\Delta\mu=(\mu\mathop{{\otimes}}\mathrm{id})(\mathrm{id}\mathop{{\otimes}}\Delta),\quad \includegraphics[scale=0.5]{Falg.pdf}\]
An $F$-algebra is called {\em special} if $\mu\Delta=\mathrm{id}$ or \includegraphics[scale=0.3]{special.pdf} (or quasispecial if a nonzero multiple of this).
\end{definition}
\begin{lemma}\label{lemFalg}The following are equivalent for a unital algebra $A$
(1) A is the algebra part of an F-algebra.
(2) A is a Frobenius algebra.
\noindent The F-algebra is special if and only if $\mu(g)=1$.
\end{lemma}
{\noindent {\bfseries Proof:}\quad } (i) If we are given an F-algebra then let $(\ ,\ )=\cup={\epsilon}\mu$ and $g=\cap=\Delta(1)$. The first half of the F-algebra axiom with $1$ on the top left leg and ${\epsilon}$ on the bottom right leg, recovers the first half of the axiom for a nondegenerate bilinear pairing. Similarly for the other side. Figure~\ref{figFalg} part (a) then checks that we have a Frobenius algebra with respect to this pairing and that the categorical adjoint of the product with respect to the pairing is the coproduct, and vice versa. Note that these categorical adjoints are the opposite of the usual vector space adjoints so as to avoid an unnecessary transposition, as for braided-Hopf algebras in \cite{Ma:alg}. (ii) Conversely, suppose that $A$ is a Frobenius algebra with bilinear $(\ ,\ )=\cup$ and inverse $g=\cap$. It is convenient to define $\Delta$ in terms of the bilinear and $\mu$ as shown in Figure~\ref{figFalg}, and ${\epsilon}=(\ ,1)$ as adjoint to 1. This implies that ${\epsilon}\mu=(\ ,\ )$, $\Delta(1)=g$ and $1=(\mathrm{id}\mathop{{\otimes}}{\epsilon})(g)$. In the figure, we also recover $\mu$ in terms of $\Delta$ by the up-side-down formula and check the main identities for an F-algebra. Once these are established, we go back in the last line of part (b) and check that $\Delta$ is coassociative. From an identity in part (a) we also see that $\Delta,\mu$ are adjoint in the categorical sense. It is clear that this construction in inverse to the one in part (i), proving the equivalence. (iii) Part (c) of the figure makes it clear that the special property corresponds to $\mu(g)$ being a left and right identity for the product $\mu$, i.e. 1. (The same proof up-side-down means it is also equivalent to $(\ ,\ )\Delta={\epsilon}$.)
\endproof
\begin{figure}
\[ \includegraphics[scale=1]{Falgpf.pdf}\]
\caption{\label{figFalg} (a)-(b) Proof that F-algebras and Frobenius algebras are equivalent and (c) the special property as $\mu(g)=1$.}
\end{figure}
The significance of the F-algebra point of view is that all the identities for associativity, unity, coassociativity, counity, and the identities in Definition~\ref{defFalg} take the form that all different ways to compose $\Delta,\mu, 1, {\epsilon}$ to give a connected planar graph with fixed $m$ legs in and $n$ legs out are equal. Here $1$ is regarded as a node with no legs in and one leg out ($m=0,n=1$) and ${\epsilon}$ is regarded as a node with one leg in and no legs out ($m=1,n=0$). This leads iteratively to:
\begin{corollary}\cite{MaRie} cf\cite{CD,Q} For a special F-algebra, all compositions of $\Delta,\mu, 1, {\epsilon}$ corresponding to a connected planar graph with $m$ legs in and $n$ legs out can be obtained in the standard form of an iterated $m-1$-fold product and an iterated $n-1$-fold coproduct. By associativity and coassociativity, we can depict these as a single node with $m$-legs going in at the top and $n$ legs out at the bottom (a `spider'), in our case keeping the order of legs.
\end{corollary}
Composing spiders along a contiguous subset of legs necessarily gives the spider with the remaining legs, which is also clear using coassociativity and associativity to isolate the branches to be contracted and using the special property iteratively. The empty morphism is the element 1 of the field, which must therefore equal ${\epsilon}(1)$ as the spider with $m=0,n=0$. (This is not really a connected graph but we formally include it.) The same will apply up to an overall scale factor in the quasispecial case. The commutative and cocommutative case of primary interest in Computer Science (and where we do not need to restrict to planar graphs, i.e., can allow crossings) is the celebrated spider theorem in \cite{CD}. The noncommutative case but with the Frobenius form symmetric occurs somewhat implicitly in \cite{LauPf}, and is also studied in \cite{Q}. A short proof in the fully general case is in \cite{MaRie}.
\subsection{F-Hopf algebras}\label{secFHopf}
We recall that an algebra and coalgebra $H$ form a bialgebra if $\Delta,{\epsilon}$ are unital algebra homs (where $H\mathop{{\otimes}} H$ has the tensor product algebra structure) and a Hopf algebra if there is additionally a map $S:H\to H$ such $\mu(S\mathop{{\otimes}}\mathrm{id})\Delta=1{\epsilon}=\mu(\mathrm{id}\mathop{{\otimes}} S)\Delta$. See \cite{Ma} for more details.
A single F-algebra can never be a bialgebra other than the trivial case where $A=k$ is the field (which we exclude). This is because $(h,h')={\epsilon}(hh')$ for all $h,h'\in A$ and $g=\Delta(1)=1\mathop{{\otimes}} 1$ need to be inverses to each other, which needs $1{\epsilon}$ to be the identity map on $A$.
\begin{remark}\label{remunnorm}\rm In the Computer Science literature, a `bialgebra' often means what we will call an {\em unnormalised bialgebra} (or `scaled bialgebra'\cite{CD}), where we assume an algebra and a coalgebra and only that ${\epsilon}(1)$ is invertible and
\[1\mathop{{\otimes}} 1= {\epsilon}(1)\Delta(1),\quad {\epsilon}\mathop{{\otimes}}{\epsilon}={\epsilon}(1){\epsilon}\mu,\quad \Delta\mu={\epsilon}(1)\mu_{A\mathop{{\otimes}} A}(\Delta\mathop{{\otimes}}\Delta)\]
\[ \includegraphics{bialg.pdf}\]
There is no significance to the braid crossing at the moment (the operation so far is just the usual flip map). The unit and counit are represented by a univalent node as before.
In this case, it is easy to see that $\lambda\Delta,\lambda^{-1}{\epsilon}$ form a usual bialgebra. An unnormalised Hopf algebra is an unnormalised bialgebra and an antipode $S$ obeying the same condition as for bialgebras. In this case $\lambda^{-2}S$ makes the associated bialgebra into a Hopf algebra in the usual sense. Hence it suffices to work in the usual normalised case. There is also no significance so far to colouring of the algebra and coalgebra nodes of a bialgebra other than to remind that they cannot belong to the same F-algebra.\end{remark}
\begin{definition}\cite{DD, CoD} An F-bialgebra is a pair of F-algebras on the same vector space such that for each F-algebra, the algebra of one and the coalgebra of the other forms a bialgebra. It is called an F-Hopf algebra if these bialgebras are Hopf algebras. It is called special if both F-algebras are special.
\end{definition}
Here is where, because there are now two F-algebras, it is useful colour one red and the other green (in a black and white, printout, the green will appear lighter). Then the definition
says that they fit together e.g. as in Remark~\ref{remunnorm}, except that we stay in the usual (normalised) case. Similarly with the colours swapped. If we are given one, say the red product/green coproduct Hopf algebra, as primary then we refer to the other as the associated Hopf algebra in the pair.
\begin{proposition}cf\cite{DD}\label{propFant} If an $F$-bialgebra obeys the three conditions
\[ \includegraphics[scale=0.9]{Fant.pdf}\]
then it is an $F$-Hopf algebra with antipode for the red product/green coproduct bialgebra given by the expressions in the third condition. The latter holds if the Frobenius forms are symmetric, in which case the antipode for the red product bialgebra and the green product bialgebra respectively can also be written as
\[ \includegraphics[scale=.8]{S.pdf}.\]
\end{proposition}
{\noindent {\bfseries Proof:}\quad } The proof that we have an antipode for the red product bialgebra is shown in Figure~\ref{figFant} (a) and is similar to \cite{DD} but with a different $S$ as we do not assume here that our algebra is commutative. The left-right reflections of the first two stated conditions also hold by applying the unit or counit to the third stated condition. Next, applying the snake identities for duals to the third condition in the statement gives two equivalent versions in succession as shown in (b). One similarly has that the colour-reversal of the first two stated conditions hold. The colour-reversal of (a) is then the proof for the other Hopf algebra.
Part (c) shows that the third stated condition holds if the red Frobenius form and the green metric are symmetric (with the latter equivalent to the green Frobenius form symmetric) and in this case it provides the second way to write the antipode for the red product bialgebra. Similarly for the green product bialgebra . \endproof
\begin{figure}
\[ \includegraphics[scale=1]{Fantpf.pdf}\]
\caption{\label{figFant} (a) Proof of Proposition~\ref{propFant} that we obtain a Hopf algebra. (b) Three equivalent forms of the third condition. (c) Proof that the third condition holds in the symmetric case.}
\end{figure}
The first two conditions here say that the categorical dual of the green unit with respect to the red Frobenius form is the green counit, and the green Frobenius dual of the red counit is the red unit. This result is for motivation and in what follows we only take away the second special form of the antipode, without the symmetry or other assumptions. In fact, we will assume $S$ to be the inverse of the one at the end of Proposition~\ref{propFant}, i.e. the stated form on the Hopf algebra with opposite coproduct.
In the following, we will start with one Hopf algebra $H$ and denote by $H^\medstar$ the categorical dual defined by nested adjoints (in the style of Figure~\ref{figFalg}(a), but now for the duality between $H$ and its dual space). Hence $H^\medstar$ is the op-algebra and op-coalgebra to the usual $H^*$, but the two are isomorphic as Hopf algebras via the antipode. We will assume that $H,H^\medstar$ as algebras are Frobenius. The latter means that $H$ is a {\em Frobenius coalgebra} in the sense of a coalgebra and an invertible metric $g$ such that
\[ (\Delta\mathop{{\otimes}}\mathrm{id})g=(\mathrm{id}\mathop{{\otimes}}\Delta)g,\]
equivalent to Definition~\ref{defFalg} by Lemma~\ref{lemFalg} read up side down.
\begin{proposition}\label{propFhopf} Let $H=(\mu_{\color{red}\bullet},1_{\color{red}\bullet}, \Delta_{\color{green}\bullet},{\epsilon}_{\color{green}\bullet})$ be a Hopf algebra such that the algebra and coalgebra are Frobenius and the antipode $S$ is \includegraphics[scale=0.7]{S3.pdf}.
Then we have an F-Hopf algebra where the associated Hopf algebra $\tilde H=(\mu_{\color{green}\bullet},1_{\color{green}\bullet}, \Delta_{\color{red}\bullet},{\epsilon}_{\color{red}\bullet})$ has antipode \includegraphics[scale=0.7]{S4.pdf}. Moreover, $\tilde H{\cong} H^{\medstar op}$ as Hopf algebras.
\end{proposition}
{\noindent {\bfseries Proof:}\quad } Given $H$ with red product and green coproduct we use the Frobenius structure on $H$ to define a dual red coproduct also on $H$. We use that the coalgebra is Frobenius to define a dual green product also on $H$. We then show in Figure~\ref{figFhopf} (a) that these fit together to form a bialgebra $\tilde H$, recognising $S^{-1}$ where $S$ denotes assumed antipode for the red product and green coproduct and using the (braided)-anti-algebra property of antipodes\cite{Ma} to take it through the product and cancel against the red metric, turning it green. In the second line {\em we start off with the wrong braid crossing}, which is of no significance at the moment, recognise $S$ and use the same (braided)-anti-algebra property to move them through the product. This then cancels against the red metric to turn it green and give the same expression as in the result of the first line after rearrangement. That the red counit (red-adjoint to the red unit) provides the counit of the bialgebra, etc are immediate and left to the reader. In part (b), we verify the proposed antipode $\tilde S$ for the green product/red coproduct Hopf algebra in terms of the antipode property for $S$. In (c) we check that $H^\medstar$ computed using (say) the red duality gives the product and antipode of $\tilde H^{op}$.
\endproof
\begin{figure}
\[ \includegraphics[scale=1]{Fhopfpf.pdf}\]
\caption{\label{figFhopf} Proof of Proposition~\ref{propFhopf}. Part (a) checks the bialgebra property and (b) the antipode properties.}
\end{figure}
This gives a canonical F-Hopf algebra amplified under certain conditions from a single Hopf algebra. It provides a slightly different route to the following result in \cite{CD}, which also applies more generally for a Hopf algebra in a symmetric monoidal category such that integrals exist and give a nondegenerate bilinear form as for the vector space case on which we focus.
\begin{corollary}\label{corHH*}\cite{CD} Every finite-dimensional Hopf algebra $H$ and its dual $H^{\medstar op}$ form an $F$-Hopf algebra. This is quasispecial, with $\mu_{\color{red}\bullet}(g_{\color{red}\bullet})\ne 0$ if and only if $H$ is semisimple and $\mu_{\color{green}\bullet}(g_{\color{green}\bullet})\ne0$ if and only if $H$ is cosemisimple (these being equivalent in characteristic 0).
\end{corollary}
{\noindent {\bfseries Proof:}\quad } The algebra of any finite-dimensional Hopf algebra $H$ is necessarily Frobenius\cite{Par} with
\[ (h,h')_{\color{red}\bullet}=\int hh',\quad g_{\color{red}\bullet}=(\mathrm{id}\mathop{{\otimes}} S) \Delta\Lambda,\quad\mu_{\color{red}\bullet}(g_{\color{red}\bullet})={\epsilon}(\Lambda)\]
\[ (\int\mathop{{\otimes}} \mathrm{id})\Delta= 1\int,\quad \int\Lambda=1,\quad h\Lambda =\Lambda{\epsilon}(h)\]
for all $h,h'\in H$. The (red) product of $H$ is omitted and $\Delta$ is its (green) coproduct. The map $\int: H\to k$ here is a `right-invariant integral' on $H$ and is unique up to scale. The element $\Lambda$ is a left integral in $H$ and corresponds to a left-invariant integral on $H^*$. Both are unique up to scale and can be normalised as stated. We use the compact notation $\Delta h=h_1\mathop{{\otimes}} h_2$ (with more numbers for iterated coproducts and sum of such terms understood; alternatively this can be done with strings, given later). We check the metric inversion identities,
\begin{align*} (\int h\Lambda_1)S\Lambda_2&=(\int h_1\Lambda_1) (S(h_2 \Lambda_2))h_3=(\int h_1\Lambda)S(1)h_2={\epsilon}(h_1)h_2\int\Lambda=h\\
\Lambda_1\int( (S\Lambda_2) Sh)&=(Sh_1)h_2\Lambda_1\int S(h_3\Lambda_2)=(Sh)\Lambda_1 \int S\Lambda_2\\
&=(Sh)S^{-1}(S\Lambda)_2\int (S\Lambda)_1=(Sh)S^{-1}(1)\int S\Lambda=Sh \int S\Lambda\end{align*}
for all $h\in H$. Here, computing the following middle expression two ways by the results just found,
\[ \Lambda_1'\mathop{{\otimes}} S\Lambda_2'=\Lambda'_1\int (S\Lambda'_2)\Lambda_1\mathop{{\otimes}} S\Lambda_2=\Lambda_1 \int S\Lambda\mathop{{\otimes}} S\Lambda_2\]
tells us that $\int S\Lambda=1$, as required. Here $\Lambda'$ is another copy of $\Lambda$. These calculations can also be done with diagrams, which we defer to Section~\ref{secbraZX}. This gives the red Frobenius form, i.e. on $H$ as an algebra. We have $\mu_{\color{red}\bullet}(g_{\color{red}\bullet})=\Lambda_1S\Lambda_2=1{\epsilon}(\Lambda)$, so this is quasispecial. A property of integrals is that ${\epsilon}(\Lambda)\ne 0$ if and only if the algebra is semisimple.
Moreover, the same data $\int,\Lambda$ can be viewed as $\int\in H^\medstar$ and $\Lambda:H^\medstar\to k$ and make this similarly Frobenius, i.e. the coalgebra of $H$ Frobenius. We have explicitly
\[ (h,h')_{\color{green}\bullet}=\int (Sh)h'=(Sh,h')_{\color{red}\bullet},\quad g_{\color{green}\bullet}=\Delta\Lambda=(\mathrm{id}\mathop{{\otimes}} S^{-1})g_{\color{red}\bullet}\]
from which it is clear that these are also inverse to each other, since the previous pair were, and that $S$ has the form needed in Proposition~\ref{propFhopf}. We have $\mu_{\color{green}\bullet}(g_{\color{green}\bullet})=\Lambda\int 1$, so as a Frobenius algebra, this too is quasispecial with nonzero value if and only if $H^\medstar$ is semisimple. The latter in characteristic 0, is equivalent to $H$ semisimple and to $S^2=\mathrm{id}$ due to results of Larson and Radford, see \cite[Thm~3.14]{Sch} for an exposition.
\endproof
For an example, we recall\cite{Ma} that if $X$ is a group (not necessarily Abelian) then its associated group Hopf algebra $kX$ has basis given by elements of $X$ with the product extended linearly and $\Delta x=x\mathop{{\otimes}} x, {\epsilon}(x)=1, Sx=x^{-1}$ for all $x\in X$. Elements, as here, where the coproduct is diagonal are called {\em group-like}\cite{Ma} (or set-like in the computer science literature). If $X$ is finite then the associated function Hopf algebra $k(X)$ has basis $\{\delta_x\}$ of $\delta$-functions with $\delta_x\delta_y=\delta_{x,y}\delta_x$, $\Delta\delta_x=\sum_{yz=x}\delta_y\mathop{{\otimes}}\delta_z$, ${\epsilon}\delta_x=\delta_{x,e}$ and $S\delta_x=\delta_{x^{-1}}$ for all $x\in X$. Here $e\in X$ is the group identity.
\begin{example}\label{exCX}\rm If $X$ is a finite group then $H=kX$ with product $x{\color{red}\bullet}y=xy$ and coproduct $\Delta_{\color{green}\bullet}x=x\mathop{{\otimes}} x$ amplifies to an $F$-Hopf algebra by Proposition~\ref{corHH*}. The integral is $\int x=\delta_{x,e}$ and the integral element is $\Lambda=\sum_{x\in X}x$. The Frobenius form on $H$ as an algebra, its inverse, and the associated coproduct are
\[ (x,y)_{\color{red}\bullet}=\delta_{x^{-1},y},\quad g_{\color{red}\bullet}=\sum_{x\in X} x\mathop{{\otimes}} x^{-1},\quad \Delta_{\color{red}\bullet}x=\sum_{y,z\in X\, |\, yz=x}y\mathop{{\otimes}} z,\quad{\epsilon}_{\color{red}\bullet}x=\delta_{e,x}\]
and $\mu_{\color{red}\bullet}(g_{\color{red}\bullet})=|X|$, which is quasi-special. The Frobenius form on $H$ as a coalgebra, its inverse and the associated product are
\[(x,y)_{\color{green}\bullet}=\delta_{x,y},\quad g_{\color{green}\bullet}=\sum_{x\in X} x\mathop{{\otimes}} x,\quad x_{\color{green}\bullet}y= \delta_{x,y}x,\quad 1_{\color{green}\bullet}=\sum_{x\in X}x \]
from which we see that the green algebra/red coalgebra has the structure of $H^*$ on matching $x$ to $\delta_x$ there. We also see that the green F-algebra is special.
\end{example}
The work \cite{CoD} considers the simplest noncommutative noncocommutative example, the Taft algebra. Going beyond this, we will now look at the reduced Drinfeld-Jimbo quantum group $u_q(su_2)$ at $q$ a root of unity. There are different conventions for this and we use essentially the ones in the recent work\cite{AziMa} which is noncommutative and noncocommutative even when $q=-1$, when understood correctly.
\begin{example}\label{exsl2}\rm Let $q$ be a primitive $n$-th root of unity and let $H=u_q(sl_2)$ as generated by $E,F,K$ and
\[ E^n=F^n=0,\quad K^n=1,\quad KEK^{-1}=q^{-1}E,\quad KFK^{-1}=qF,\quad [E,F]={K-K^{-1}\over q-q^{-1}}\]
\[ \Delta K=K\mathop{{\otimes}} K,\quad \Delta E=1\mathop{{\otimes}} E+E\mathop{{\otimes}} K,\quad \Delta F=F\mathop{{\otimes}} 1+ K^{-1}\mathop{{\otimes}} F,\quad{\epsilon} K=1,\quad {\epsilon} E={\epsilon} F=0\]
with antipode (and the quasitriangular structure, needed later),
\[ SK=K^{-1},\quad SE=-EK^{-1},\quad SF=-KF,\quad \hbox{{$\mathcal R$}}={1\over n}\sum_{r,a,b=0}^{n-1}{(-1)^r(q-q^{-1})^rq^{-ab}\over [r]_{q^{-1}}!}F^rK^a\mathop{{\otimes}} E^r K^b\]
where $[m]_q:=(1-q^m)/(1-q)$ denotes a `q-integer'. When $q=-1$ we omit the $q-q^{-1}$ factors in the relations and in $\hbox{{$\mathcal R$}}$. The right integral on $H$ and left integral in $H$ are
\[ \int K^iF^j E^k=\delta_{1,i}\delta_{n-1,j}\delta_{n-1,k},\quad \Lambda=\Lambda_KF^{n-1}E^{n-1};\quad \Lambda_K=\sum_{r=0}^{n-1}K^r\]
so that $\int\Lambda=1$. One also has that $S\Lambda=\Lambda$. We compute the associated $\tilde H$ product and coproduct
\[ h{\color{green}\bullet}h'=h_1\int ((Sh_2)h'),\quad \Delta_{\color{red}\bullet}h=\Lambda_1\mathop{{\otimes}} (S\Lambda_2)h\]
for all $h,h'\in u_q(sl_2)$ can then be computed and must be isomorphic to $c_q[SL_2]^{op}$ for the appropriate reduced quantum group dual to $u_q(sl_2)$. We have $\mu_{\color{red}\bullet}(g_{\color{red}\bullet})=\mu_{\color{green}\bullet}(g_{\color{green}\bullet})=0$.
We compute these products for the $n=2$ case where $q=-1$ and\cite{AziMa}
\[K^2=1,\quad E^2=F^2=0,\quad [E,F]=0,\quad \{E,K\}=\{F,K\}=0,\]
\[\Delta K=K\mathop{{\otimes}} K, \quad \Delta E=E\mathop{{\otimes}} K+1\mathop{{\otimes}} E,\quad \Delta F=F\mathop{{\otimes}} 1+K\mathop{{\otimes}} F\]
\[ {\epsilon} K=1,\quad {\epsilon} E={\epsilon} F=0,\quad SK=K,\quad SE=KE,\quad SF=-KF\]
\[\Lambda=(1+K)EF,\quad \hbox{{$\mathcal R$}}=(1\mathop{{\otimes}} 1-F\mathop{{\otimes}} E)\hbox{{$\mathcal R$}}_K,\quad \hbox{{$\mathcal R$}}_K={1\over 2}(1\mathop{{\otimes}} 1+K\mathop{{\otimes}} 1+1\mathop{{\otimes}} K-K\mathop{{\otimes}} K).\]
This is the initial Hopf algebra with red algebra and green coalgebra understood.
From these we work out the values of $(\mathrm{id}\mathop{{\otimes}} S)\Delta$ on our monomial basis. In particular,
\[ g_{\color{red}\bullet}=\Lambda_1\mathop{{\otimes}} S\Lambda_2=F\mathop{{\otimes}} KE-KE\mathop{{\otimes}} F+ EF\mathop{{\otimes}} K+K\mathop{{\otimes}} EF+E\mathop{{\otimes}} KF-KF\mathop{{\otimes}} E+KEF\mathop{{\otimes}} 1+1\mathop{{\otimes}} KEF\]
We let
\[ x=(1+K)E,\quad y=(1+K)F,\quad t=(K-1)EF\]
and find for the associated Hopf algebra
\[ \Delta_{\color{red}\bullet} t=t\mathop{{\otimes}} t,\quad \Delta_{\color{red}\bullet} x=x\mathop{{\otimes}} t+\Lambda\mathop{{\otimes}} x,\quad \Delta_{\color{red}\bullet} y=y\mathop{{\otimes}} t+\Lambda\mathop{{\otimes}} y\]
\[ x{\color{green}\bullet}x=y{\color{green}\bullet}y=0,\quad t{\color{green}\bullet}t=\Lambda\]
\[ x{\color{green}\bullet}y=-y{\color{green}\bullet}x=1+K,\quad x{\color{green}\bullet}t=-t{\color{green}\bullet}x=(K-1)E,\quad y{\color{green}\bullet}t=-t{\color{green}\bullet}y=(1-K)F.\]
Note that $1_{\color{green}\bullet}=\Lambda$ is the unit element. In fact, $u_q(sl_2)$ at $q=-1$ is both self-dual and anti-self dual, and indeed the associated Hopf algebra here is isomorphic to $u_q(sl_2)$ if we identify
\[ 1\mapsto 1_{\color{green}\bullet},\quad K\mapsto t,\quad E\mapsto x,\quad F\mapsto y{\color{green}\bullet}t.\]
\end{example}
\section{ $*$-algebra and Hadamard forms} \label{secstar}
Next, we recall the usual notion of a Hopf $*$-algebra. This means a Hopf algebra over $\mathbb{C}$ where the algebra is a $*$-algebra (i.e., it has an antilinear anti-algebra involution $*$) and $\Delta,{\epsilon}$ commute with $*$ (which on $\mathbb{C}$ is complex conjugation). In this case, one has that $(S*)^2=\mathrm{id}$. For a group algebra $\mathbb{C} X$, we have $x^*=x^{-1}$ and for a function Hopf algebra $\mathbb{C}(X)$ we have $f^*(x)=\overline{f(x)}$ where the bar denotes complex conjugation of the value. This is the standard notion but will not be the most relevant. Instead, there is another which we called a {\rm flip-Hopf} $*$-algebra in \cite{Ma, BegMa}, where
\[ \Delta * =\dagger\Delta,\quad \overline{\ }\circ{\epsilon}={\epsilon}\circ *,\quad S*=*S\]
with the `hermitian conjugation'
\[ \dagger:={\rm flip}(*\mathop{{\otimes}} *)\]
on any tensor product space over $\mathbb{C}$. The categorical picture behind these ideas is that of a {\rm bar} category\cite{BegMa:bar,BegMa} which will be particularly useful in the braided case. For any flip-Hopf $*$-algebra, we define the associated antilinear Hopf algebra automorphism
\[ \theta:= S*,\]
which squares to $S^2$.
\subsection{Unimodular setting}\label{secuni}
We start with the simplest setting, which is adequate for finite group examples and simpler quantum groups.
\begin{definition}\label{starform} A Frobenius $*$-algebra is a Frobenius algebra $A$ over $\mathbb{C}$ which is a $*$-algebra and for which $\overline{(a,b)}=(b^*,a^*)$, or equivalently $g^\dagger=g$ for the metric.
\end{definition}
The metric property here is similar to the notion of a real metric in noncommutative geometry \cite{BegMa} and the proof that this is equivalent to the property of (\ ,\ ) is parallel to the proof there. In this case, it makes sense to define a sesquilinear bilinear form on $H$ by
\[ \<h|h'\>=(h^*,h')\]
which, under our assumption makes $A$ into a Hilbert space (upon completion in the infinite-dimensional case). In the group algebra Example~\ref{exCX} we have $\<x|y\>=\delta_{x,y}$ for the (red) inner product there, i.e. the group basis elements become an orthonormal basis.
\begin{lemma}\label{starfrob} A $*$-algebra $A$ is Frobenius if and only if the corresponding F-algebra obeys \rm
\[ \Delta * =\dagger \Delta,\quad \overline{\ }\circ{\epsilon}={\epsilon}\circ *.\]
\end{lemma}
{\noindent {\bfseries Proof:}\quad } Here $\Delta a=ga$ and we use Lemma~\ref{gcentral} to deduce that $\Delta (a^*)= g a^*=a^*g=\dagger(g^\dagger a)=\dagger(ga) =\dagger\Delta a$. We used that the metric obeys $g^\dagger=g$. We also have ${\epsilon}(a^*)=(1,a^*)=\overline{(a,1)}=\overline{{\epsilon}(a.1)}=\overline{{\epsilon}(a)}$. \endproof
We equally define a Frobenius $*$-coalgebra as a Frobenius coalgebra over $\mathbb{C}$ where the coalgebra obeys the conditions in Lemma~\ref{starfrob} and the Frobenius structure obeys the conditions in Definition~\ref{starform}, in which case the associated algebra is a $*$-algebra. In either case, we say that the associated F-algebra is an F-$*$-algebra.
\begin{corollary}\label{corbasicHH*}(Unimodular case.) If $H$ is a flip-Hopf $*$-algebra and $\int *=\overline{\ }\circ\int, \Lambda^*=\Lambda$ in Corollary~\ref{corHH*} then both associated F-algebras are F-$*$-algebras and the associated Hopf algebra in the induced F-Hopf algebra is also a flip-Hopf $*$-algebra.
\end{corollary}
{\noindent {\bfseries Proof:}\quad } Here $(h'{}^*,h^*)_{\color{red}\bullet}=\int(hh')^*=\overline{\int(h,h')_{\color{red}\bullet}}$ in Corollary~\ref{corHH*} under our assumption on $\int$. Similarly, $\dagger g_{\color{green}\bullet}=\dagger\Delta\Lambda=\Delta(\Lambda^*)=g_{\color{green}\bullet}$ for the Frobenius coalgebra in Corollary~\ref{corHH*}. Hence the red algebra and the green coalgebra are both Frobenius as *-(co)algebras. The associated red coproduct therefore obeys the condition in Lemma~\ref{starfrob} and the green product is a *-algebra. Hence the Hopf algebra that these form by Corollary~\ref{corHH*} is a flip-Hopf $*$-algebra. \endproof
We say in this case that we have an example of an F-Hopf $*$-algebra.
\begin{example}\rm \label{exCXstar} Let $H=\mathbb{C} X$ for a finite group $X$ with $x^*=x^{-1}$. This is a Hopf $*$-algebra but we regard it as a flip-Hopf $*$-algebra since it is cocommutative. We have $\int x^*=\delta_{e,x^{-1}}=\delta_{e,x}=\int x=\overline\int x$ and $\Lambda^*=\sum_{x\in X}x^{-1}=\Lambda$ as required. The induced red coproduct in Example~\ref{exCX} obeys $\Delta_{\color{red}\bullet}(x^*)=\sum_{yz=x^{-1}}y\mathop{{\otimes}} z=\dagger\sum_{yz=x^{-1}}z^{-1}\mathop{{\otimes}} y^{-1}=\dagger\Delta_{\color{red}\bullet} x$ as it had to by Corollary~\ref{corbasicHH*}. Likewise the induced green product obeys $y^*{\color{green}\bullet}x^*=y^{-1}{\color{green}\bullet}x^{-1}=\delta_{x,y}x^{-1}=(x{\color{green}\bullet}y)^*$, so this is a $*$-algebra as it had to be for an F-Hopf $*$-algebra.\end{example}
Note that $\mathbb{C}(X)$ with its usual $*$-algebra structure is not a flip-Hopf *-algebra when $X$ is nonAbelian. For that, we should take $\delta_x^*=\delta_{x^{-1}}$. This then matches up with the red coproduct and green product just found.
\begin{example}\label{exflipsl2}\rm Let $H=u_q(sl_2)$ at $q=-1$ as in Example~\ref{exsl2}. This forms a flip-Hopf $*$-algebra with
\[ K^*=K,\quad E^*=F,\quad F^*=E;\quad \theta(K)=K,\quad \theta(E)=-KF,\quad \theta(F)=KE.\]
We also have $\int S= \int$, $S\Lambda=\Lambda$ (the Hopf algebra is unimodular in both aspects) and $\int h^*=\overline{\int h}$,\ $\Lambda^*=\Lambda$ as required. Using the same underlying $*$-operation, we have
\[ t^*=FE(K-1)=t,\quad x^*=F(1+K)=(1-K)F=y{\color{green}\bullet}t,\quad y^*=E(1+K)=(1-K)E=t\circ x\]
which indeed makes the green product a $*$-algebra also. Similarly, for example
\begin{align*}\Delta_{\color{red}\bullet}x^*&=\Delta_{\color{red}\bullet}(y{\color{green}\bullet}t)=(\Delta_{\color{red}\bullet}y){\color{green}\bullet}(\Delta_{\color{red}\bullet}t)=(y\mathop{{\otimes}} t+\Lambda\mathop{{\otimes}} y){\color{green}\bullet}(t\mathop{{\otimes}} t)=y{\color{green}\bullet}t\mathop{{\otimes}}\Lambda+t\mathop{{\otimes}} y{\color{green}\bullet}t\\
&=x^*\mathop{{\otimes}}\Lambda+t\mathop{{\otimes}} x^*=\dagger\Delta_{\color{red}\bullet} x \end{align*}
as it must for a flip-Hopf $*$-algebra. In fact, this is isomorphic as a flip-Hopf $*$-algebra to $u_q(sl_2)$ by the map in Example~\ref{exsl2}.
\end{example}
On the other hand, the assumptions we made are not typical because in general $\int*$ is an (antilinear) left invariant integral and can not be equated to $\overline{\ }\circ\int$, and $\Lambda^*$ is a right integral element and can not be equated to $\Lambda$. This is because a typical quantum group will not be unimodular in the sense that the left and right integrals will not be proportional. Note that the Frobenius form, when it exists, is not unique. Hence it is also possible that we could stay in the setting of Definition~\ref{starform} and Lemma~\ref{starfrob}, which seems reasonable from a noncommutative geometry point of view, but with the Frobenius form and its inverse not given by integrals. However, we do not have any general results about this more general possibility. One could also explore a version of the theory adapted to ordinary Hopf $*$-algebras, which is more relevant to usual quantum geometry and applicable to $U_q(su_2)$ and $\mathbb{C}_q[SU_2]$ but not to the reduced finite-dimensional versions at $q$ a root of unity.
\subsection{General Hopf algebra case based on integrals} \label{secgenstar}
The less special assumptions we will take for the properties of the integrals on a flip-Hopf $*$-algebra under $*$ are
\begin{equation}\label{intstar} \int h^*= \overline{\int S h},\quad \Lambda^*=S\Lambda\end{equation}
These have to hold up to scale as each space of integrals is 1-dimensional, so it is not unreasonable to assume them as part of the normalisations. This does, however, imply that $\int\circ S^2=\int$ and $S^2\Lambda=\Lambda$
\begin{proposition}\label{propflipHH*} Let $H$ be a flip-Hopf $*$-algebra with integrals in Corollary~\ref{corHH*} obeying the condition (\ref{intstar}) and suppose that the antipodes on $H$ and on associated Hopf algebra of the induced F-Hopf algebra square to the same. Then the associated Hopf algebra is a flip-Hopf $*$-algebra with the same associated antilinear automorphism. \end{proposition}
{\noindent {\bfseries Proof:}\quad } The issue here is that $\int,\Lambda$ are not assumed to obey the stricter reality properties in Corollary~\ref{corbasicHH*}, with the result that the associated green product is not a $*$-algebra with the same $*$ as $H$, and likewise the red coproduct does not skew-commute with $*$. We will resolve this by introducing a new green $*$-operation $\color{green}*$ making this into a flip-Hopf $*$-algebra. First note that the first of (\ref{intstar}) says that $\overline{\int h}=\int\theta(h)$. Hence
\begin{align*}\theta(h {\color{green}\bullet}h')& = \theta(h_1{})\overline{(Sh_2)h'}= \theta(h_1{})\int \theta(Sh_2)\theta(h')=\theta(h{})_1\int (S\theta(h)_2)\theta(h')=\theta(h){\color{green}\bullet}\theta(h').\end{align*}
Next, the second of (\ref{intstar}) implies that
\begin{align*} g^\dagger&=(S\Lambda_2)^*\mathop{{\otimes}}\Lambda_1{}*=S(\Lambda_2{}^*)\mathop{{\otimes}}\Lambda_1{}*=S\Lambda^*{}_1\mathop{{\otimes}}\Lambda^*{}_2\\
&=S(S\Lambda)_1\mathop{{\otimes}} (S\Lambda)_2=S^2\Lambda_2\mathop{{\otimes}} S\Lambda_1=(S\mathop{{\otimes}} S){\rm flip}(g)\end{align*}
which is equivalent to $(\theta\mathop{{\otimes}}\theta)(g)=g$. Consequently,
\[ (\theta\mathop{{\otimes}}\theta)\Delta_{\color{red}\bullet}h=(\theta\mathop{{\otimes}}\theta)(gh)=g\theta(h)=\Delta_{\color{red}\bullet}\theta(h)\]
This $\theta$ is {\em also} a bialgebra and hence Hopf algebra automorphism on the associated green product/red coproduct Hopf algebra. If we let $\tilde S$ denote the antipode for the associated green algebra/red coalgebra $\tilde H$ then $\tilde S\theta=\theta \tilde S$. We now see that
${\color{green}*}:=\tilde S^{-1}\theta$ is an antilinear anti-algebra and anti-coalgebra map and commutes with $\tilde S$. Hence
\[ {\color{green}*}^2=\tilde S^{-1}\theta \tilde S^{-1}\theta=\theta^2 S^{-2}_{\color{green}\bullet}=S^2 S^{-2}_{\color{green}\bullet}=\mathrm{id}\]
if and only if $S^2=S^2_{\color{green}\bullet}$. Under this additional hypothesis, we see that ${\color{green}*}$ makes the associated Hopf algebra into a flip-Hopf $*$-algebra.
\endproof
In this case, we say that we have a general F-Hopf $*$-algebra, where we have a green $*$-operation for the green product, not necessarily the same as the initial one for the initial (red) product. The notion of a Frobenius $*$-algebra also gets modified. We saw that now the natural property is that $(\theta\mathop{{\otimes}}\theta)(g_{\color{red}\bullet})=g_{\color{red}\bullet}$ and one similarly has
\[\overline{(h,h')_{\color{red}\bullet}}=\overline{\int hh'}=\int\theta(h)\theta(h')=(\theta(h),\theta(h'))_{\color{red}\bullet}\]
under our assumptions. The same on the green side, noting that our definition of ${\color{green}*}$ is such that the green $\theta$ automorphism is the same as the initial one on $H$. This therefore departs from the more natural setting from the Frobenius $*$-algebra point of view in the preceding section.
\begin{example}\rm We make $H=u_q(sl_2)$ in Example~\ref{exsl2} for $q$ a primitive $n$-th root of unity into a flip-Hopf $*$-algebra\cite{Ma} with
\[ K^*=K^{-1},\quad E^*=F,\quad F^*=E;\quad \hbox{{$\mathcal R$}}^\dagger=\hbox{{$\mathcal R$}}^{-1}.\]
The right integral on $H$ and left integral in $H$ obey the conditions (\ref{intstar}) (in fact, one has that $\Lambda^*=\Lambda=S\Lambda$ is unimodular, but not $\int$). These are straightforward calculations which we omit, focusing on the additional condition in Proposition~\ref{propflipHH*}. In lieu of a formal proof, we do a detailed calculation to illustrate the method. Thus, from Corollary~\ref{corHH*},
\[ \tilde Sh=\Lambda_1\int \Lambda_2 h=\sum_j(\mathrm{id}\mathop{{\otimes}}\int)\Delta (K^j F^{n-1}E^{n-1})h\]
which on a monomial element $h$ picks out one term of the coproduct such that the powers of $F,E$ in the right hand factor add up with the powers in $h$ to $n-1$ (else the integral will kill the term). For example $h=K^a F$ means the only relevant terms are
\[ \Delta E^{n-1}= (E\mathop{{\otimes}} K+1\mathop{{\otimes}} E)^{n-1}=1\mathop{{\otimes}} E^{n-1}+...\]
\[ \Delta F^{n-1}=(F\mathop{{\otimes}} 1+K^{-1}\mathop{{\otimes}} F)^{n-1}=...+[n-1]_{q^{-1}} FK^{-(n-2)}\mathop{{\otimes}} F^{n-2}+...\]
on expanding in powers of $E,F$ in the right factors and using the commutation relations of $u_q(sl_2)$ to collect terms (an instance of the $q$-binomial formula\cite{Ma}). Hence,
\begin{align*} \tilde S(K^aF)&=\sum_j(\mathrm{id}\mathop{{\otimes}}\int)\left( (K^j\mathop{{\otimes}} K^j)([n-1]_{q^{-1}} FK^{-(n-2)}\mathop{{\otimes}} F^{n-2})(1\mathop{{\otimes}} E^{n-1} K^a F)\right)\\
&=\sum_j [n-1]_{q^{-1}}K^j F K^{2}\int K^j F^{n-2} E^{n-1}K^aF\\
&=\sum_j [n-1]_{q^{-1}} q^{a-2} K^{2+j}F\int K^{a+j} F^{n-1} E^{n-1}\\ &=\sum_j [n-1]_{q^{-1}} q^{a-2} K^{2+j}F\delta_{1,a+j}=[n-1]_{q^{-1}} q^{a-2} K^{3-a}F=-q^{a-1}K^{3-a}F.
\end{align*}
We commuted $E,F$ inside the integral as their commutator has lower degree in $E,F$. Hence $S^2_{\color{green}\bullet}(K^aF)=q^{a-1+ (3-a)-1}K^{3-(3-a)}F=q K^aF$. Meanwhile, $S(K^aF)=-KFK^{-a}=-q^a K^{1-a}F$ so that $S^2(K^aF)=q^{a+1-a}K^{1-(1-a)}= qK^aF$ also, so $S^2_{\color{green}\bullet}=S^2$ on these elements. The proof for the general case is rather more complicated, but follows the same method using the $q$-binomial formula and extracting the relevant terms for the coproduct. Thus, the conditions of Proposition~\ref{propflipHH*} apply and we have an F-Hopf $*$-algebra. The associated Hopf algebra in this example is presumably isomorphic to a reduced version of the flip-Hopf $*$-algebra $\mathbb{C}_q[SU_2^{fl}]$ in \cite[Ex.~2.113]{BegMa}. \end{example}
\subsection{Phases} Another important ingredient for ZX calculus is the notion of phases. These are defined in \cite{CD,DD} as map $A\to A$ on a Frobenius algebra $A$ such that applied to either leg of the product gives the same result as applying the map after the product. It is shown in the commutative setting that these take the form of left multiplication by elements of $A$. This map can also be applied to coproducts leading to the decoration of any spider by this operation applied to any leg.
In our setting with $A$ a possibly noncommutative Frobenius $*$-algebra, we let $Z(A)$ be the centre of $A$ and
\[ P(A)=\{\alpha\in Z(A)\ |\ \alpha^*\alpha=\alpha\alpha^*=1\}\]
i.e. of invertible elements such that $\alpha^*=\alpha^{-1}$. This forms an Abelian group (the `group of phases'). The corresponding operation is $L_\alpha(a)=\alpha a$ as before, which clearly can be applied to any leg of a composite product.
\begin{lemma} If $A$ if an F-$*$-algebra then the action of $\alpha\in P(A)$ can be taken through the coproduct to either leg.
\end{lemma}
{\noindent {\bfseries Proof:}\quad } This is immediate using the relevant form of the coproduct. For example, $(\mathrm{id}\mathop{{\otimes}} L_\alpha)\Delta a=g^1\mathop{{\otimes}} \alpha g^2 a=g^1\mathop{{\otimes}} g^2\alpha a=\Delta(L_\alpha a)$ and $(L_\alpha\mathop{{\otimes}} \mathrm{id})\Delta a=\alpha g^1\mathop{{\otimes}} g^2 a=g^1\mathop{{\otimes}} g^2\alpha a=\Delta(L_\alpha a)$ using centrality of $g$ (or using the other form $\Delta a=ag$.) \endproof
It follows, as in \cite{CD,DD}, that we get the same result applying $L_\alpha$ to any leg of a spider, and hence that we can denote such an $\alpha$ in a neutral way inside the node of the spider. When combining spiders, it is also clear that phases multiply using the group structure of $P(A)$ inherited from the product of $A$. In \cite{CD,DD}, there is also a discussion of group-like (`set-like') elements and classical values which we omit.
\subsection{Hadamard gates and self-duality}
Another ingredient of $ZX$ calculus is the Hadamard gate, which for the standard 1 qubit realisation using $\mathbb{C}\mathbb{Z}_2$ in the introduction, is clearly just the Fourier transform. This observation was used in \cite{GogZen} to propose Fourier theory on internal groups in their categorical setting of ZX calculus. Here we give a different take on the role of this gate in our algebraic setting.
Note first that Fourier transform works canonically on any finite-dimensional Hopf algebra\cite{Ma,BegMa}. If $H$ is a Hopf algebra with right-invariant integral as above, then the Hopf algebra Fourier transform is an invertible linear map
\begin{equation}\label{fou} \hbox{{$\mathcal F$}}: H\to H^\medstar,\quad \hbox{{$\mathcal F$}}(h)=\sum_a f^a(\int e_a h),\end{equation}
where $\{e_a\}$ is a basis of $H$ and $\{f_a\}$ a dual basis. We have seen that in the F-Hopf algebra case, the associated (green product, red coproduct) Hopf algebra is isomorphic to $H^\medstar$ as a vector space and indeed if we use the metric $g_{\color{red}\bullet}$ to refer $\hbox{{$\mathcal F$}}$ back to a map $H\to H$ then it is just the identity map. Thus, the canonical Fourier transform itself is not the content of the Hadamard gate.
In fact, Fourier transform as encountered on $\mathbb{R}^n$ or on $\mathbb{Z}_n$ in terms of the group algebra is an operator $H\to H$ after the canonical Fourier transform is combined with the fact that these groups are self-dual.
We propose this self-duality of some Hopf algebras as an approach to a Hadamard gate, but different from another point of view that it should map the green spiders to red spiders, which seems important for ZX calculus. We start with the latter.
\begin{definition}\label{defhad} A Type 1 {\em Hadamard form} on an $F$-bialgebra $H$ is a non-degenerate bilinear form $\Theta$ on $H$ such that (a)-(b) of
\[ \includegraphics{had.pdf}\]
hold for adjunction with respect to $\Theta$. $\Theta$ is called {\em quasi-Hadamard} if these hold up to constant scale factors. \end{definition}
We also show the definition of a corresponding `Hadamard gate' $\hbox{{$\mathfrak h$}}:H\to H$, which we will use throughout the section (also for the other types below). In the case where $(\ ,\ )_{\color{red}\bullet}=\int\circ\mu_{\color{red}\bullet}$,
\begin{equation}\label{hadfou} \hbox{{$\mathfrak h$}}^{-1}(h)=\Theta^1(\Theta^2,h )_{\color{red}\bullet}=\Theta^1\int \Theta^2 h\end{equation}
is the Hopf algebra Fourier transform (\ref{fou}) with the associated `metric' $\Theta=\Theta^1\mathop{{\otimes}}\Theta^2\in H\mathop{{\otimes}} H$ in the role of `exponential' used to map $H^\medstar\to H$ by evaluation against its second leg (i.e., used to provide the required duality).
\begin{lemma} Let $H$ be an F-bialgebra or F-Hopf algebra and $\Theta$ a bilinear form on $H$. The following are equivalent.
(1) $\Theta$ is a Type 1 Hadamard form.
(2) $\hbox{{$\mathfrak h$}}$ is an isomorphism from the green F-algebra to the red F-algebra,
\[\mu_{\color{green}\bullet}=\hbox{{$\mathfrak h$}}^{-1}\circ\mu_{\color{red}\bullet}\circ(\hbox{{$\mathfrak h$}}\mathop{{\otimes}}\hbox{{$\mathfrak h$}}),\quad \Delta_{\color{green}\bullet}=(\hbox{{$\mathfrak h$}}^{-1}\mathop{{\otimes}}\hbox{{$\mathfrak h$}}^{-1})\circ\Delta_{\color{red}\bullet}\circ\hbox{{$\mathfrak h$}},\quad 1_{\color{green}\bullet}=\hbox{{$\mathfrak h$}}^{-1}(1_{\color{red}\bullet}),\quad {\epsilon}_{\color{green}\bullet}={\epsilon}_{\color{red}\bullet}\circ\hbox{{$\mathfrak h$}}\]
\end{lemma}
{\noindent {\bfseries Proof:}\quad } The content of the Hadamard form assumption is shown explicitly in Figure~\ref{fighadpf}. The proof is then straightforward but necessary to be sure that all relevant conventions match up, as shown in parts (a)-(b) of the figure. In (a) we cover the adjunction of the red algebra to obtain the green coalgebra, and in (b) the same for the algebra. In each case, we start with (1) the adjunction as a `rotation' using the bilinear form $\Theta$ and its inverse $\Theta\in H\mathop{{\otimes}} H$. We then eliminate the latter to give the conditions for a Hadamard form on $H$. We then write our conditions in terms of $\hbox{{$\mathfrak h$}}$ by inserting the red Frobenius form as needed and then recognise the dualisation of the red product and coproduct from Figure~\ref{figFalg} to obtain (2). \endproof
\begin{figure}
\[ \includegraphics[scale=1.1]{hadpf.pdf}\]
\caption{\label{fighadpf} Definition (a) and (b) of a Hadamard form on a Hopf algebra and equivalent properties of $\hbox{{$\mathfrak h$}}$.}
\end{figure}
\begin{example}\rm\label{exhZn} $H=\mathbb{C}\mathbb{Z}_n$ as an example of Example~\ref{exCX} has a quasi-Hadamard form as follows. We write the group $X=\mathbb{Z}_n$ additively and to avoid confusion we let $|i\>$ explicitly denote the basis vector of $\mathbb{C} X$ spanned by $i\in \mathbb{Z}_n$ (a standard notation in the Computer Science literature). We let $q=e^{2\pi\imath\over n}$ be a primitive $n$-th root of unity and define
\[ \Theta(i,j)=q^{ij}\]
extended as a bilinear form on $H$. This is invertible and provides a Hadamard form. For example, on the part (a) side, $\Theta(i+j,k)=\Theta(i,k)\Theta(j,k)$ as required since $\Delta_{\color{green}\bullet}(|k\>)=|k\>\mathop{{\otimes}}|k\>$. On the part (b) side, we have using $\Delta_{\color{red}\bullet}$,
\[\sum_{r+s=i}\Theta(s,j)\Theta(r,k)= \sum_rq^{(i-r)j}q^{rk}=\sum_rq^{r(k-j)}q^{ij}=n\delta_{j,k}q^{ij}=n\Theta(i,j{\color{green}\bullet}k)\]
where see the factor $n$ which makes this only true up to a constant factor. This is related to the fact that the red Frobenius algebra is quasi-special, not special. The associated Hadamard gate is
\[ \hbox{{$\mathfrak h$}}(|i\>)=g^1_{\color{red}\bullet}\Theta(g^2_{\color{red}\bullet},|i\>)=\sum_j|j\> q^{-ji}\]
which is the usual $\mathbb{Z}_n$ (inverse) Fourier transform. Note for general $n$ that $\hbox{{$\mathfrak h$}}^2(|i\>)=n|-i\>$, so $\hbox{{$\mathfrak h$}}$ can be normalised to be order $4$. For $n=2$ only, one can normalise by $1/\sqrt{2}$ to obtain $\hbox{{$\mathfrak h$}}=\hbox{{$\mathfrak h$}}^{-1}$, with both (a) and (b) conditions sharing the scale factor discrepancy. \end{example}
Although this works better from the point of view of mapping red spiders to green spiders, it is not very natural from a Hopf algebra point to map both the algebra and coalgebra in the same direction. This suggests two natural variants.
\begin{definition}\label{defself} Let $H$ be an $F$-Hopf algebra with antipode of the form in Proposition~\ref{propFhopf} and $\Theta$ a nondegenerate bilinear form on $H$. We call it a Type 2 (resp. Type 3) Hadamard form if it obeys (a) in Definition~\ref{defhad} and (b') (resp. (b'')), where
\[ \includegraphics{hadvar.pdf}\]
and $\tilde\Theta=\Theta\circ(S\mathop{{\otimes}}\mathrm{id})$.
\end{definition}
Recall that when making a Hopf algebra $H$ into an $F$-Hopf algebra, we take it as the red product/green coproduct one and refer the green product/red coproduct one as the associated Hopf algebra $\tilde H$.
\begin{lemma}\label{lemsd} Let $H$ be a Hopf algebra extended to an F-Hopf algebra by Proposition~\ref{propFhopf} and $\theta:H\mathop{{\otimes}} H\to k$. The following are equivalent
(1) $\Theta$ is a Type 2 (resp Type 3) Hadamard form.
(2) $h\mapsto \Theta(h, )$ is an isomorphism $H{\cong} H^\medstar$ (resp. $H{\cong} H^{\medstar op}$) of Hopf algebras.
(3) $\hbox{{$\mathfrak h$}}$ is and isomorphism $H\cong \tilde H^{op}$ (resp. $H{\cong} \tilde H$) of Hopf algebras, i.e.
\[{\rm Type\ 2}:\quad \mu^{op}_{\color{green}\bullet}=\hbox{{$\mathfrak h$}}\circ\mu_{\color{red}\bullet}\circ(\hbox{{$\mathfrak h$}}^{-1}\mathop{{\otimes}}\hbox{{$\mathfrak h$}}^{-1});\quad {\rm Type\ 3}:\quad \mu_{\color{green}\bullet}=\hbox{{$\mathfrak h$}}\circ\mu_{\color{red}\bullet}\circ(\hbox{{$\mathfrak h$}}^{-1}\mathop{{\otimes}}\hbox{{$\mathfrak h$}}^{-1})\]
for the two cases and
\[ \Delta_{\color{green}\bullet}=(\hbox{{$\mathfrak h$}}^{-1}\mathop{{\otimes}}\hbox{{$\mathfrak h$}}^{-1})\circ\Delta_{\color{red}\bullet}\circ\hbox{{$\mathfrak h$}},\quad 1_{\color{green}\bullet}=\hbox{{$\mathfrak h$}}(1_{\color{red}\bullet}),\quad {\epsilon}_{\color{green}\bullet}={\epsilon}_{\color{red}\bullet}\circ\hbox{{$\mathfrak h$}}.\]
\end{lemma}
{\noindent {\bfseries Proof:}\quad } For Type 2, part (b') in Figure~\ref{fighadvarpf} shows the modified adjunction condition in (1) and converts it to a condition on $\Theta$. This version of conditions (a) and (b') explicitly is
\[ \Theta(hh',h'')=\Theta(h,h''_2 )\Theta(h',h''_1),\quad \Theta(h,h' h'')=\Theta(h_2,h')\Theta(h_1,h''),\quad \]
as well as $\Theta(1,h)={\epsilon}(h)=\Theta(h,1)$, for all $h,h',h''\in H$. This is (2). In the figure, we then replace $\Theta$ by $\hbox{{$\mathfrak h$}}$ and recognise the antipode $\tilde S$ of the associated Hopf algebra $\tilde H$. We then use the (braided) anti-algebra homomorphism property of antipodes to obtain the variant in (3). Similarly, for Type 3, part (b'') of Figure~\ref{fighadvarpf} shows the modified adjunction condition in (1) and converts it to a condition on $\tilde\Theta$. Then the (a) and (b'') conditions on the bilinear form say respectively that
\[ \Theta(hh',h'')=\Theta(h,h''_2 )\Theta(h',h''_1),\quad \Theta(h,h' h'')=\Theta(h_1,h')\Theta(h_2,h''),\]
as well as $\Theta(1,h)={\epsilon}(h)=\Theta(h,1)$, for all $h,h',h''\in H$. This is (2). In the figure, we then insert the assumed specific form of $S$ and recognise the green-dual of the green product of $H$ as needed for (3). \endproof
\begin{figure}
\[ \includegraphics[scale=1]{hadvarpf.pdf}\]
\caption{\label{fighadvarpf} Variants (b') and (b'') of a Hadamard form as Hopf algebra self-duality and anti-self-duality.}
\end{figure}
Note that $H^\medstar{\cong} H^*$ as Hopf algebras (via the antipode), so Type 2 is equivalent to saying that $H$ is self-dual in the usual sense, while Type 3 is equivalent to saying that $H$ is anti-self-dual.
\begin{example}\rm $H=\mathbb{C} Z_n$ is a self-dual (or anti-self dual) Hopf algebra as the underlying finite Abelian group is self-dual. Here the condition (a) for $\Theta(i,j)=q^{ij}$ is as in Example~\ref{exhZn} but instead of condition (b), we now have (b') which holds for the same reason as (a), given that the form is symmetric. Thus $\Theta$ is an exact (not quasi) Type 2 Hadamard form leading more conventionally to $\hbox{{$\mathfrak h$}}$ as Fourier transform. The actual $n=2$ Hadamard gate still changes the normalisation to have $\hbox{{$\mathfrak h$}}=\hbox{{$\mathfrak h$}}^{-1}$. \end{example}
\begin{example}\rm Let $H=u_q(b_+)$ be the Taft algebra with $q$ a primitive $n$-th root of unity. This was presented as an F-Hopf algebra in \cite{CoD}, and $u_q(b_+)$ is, moreover, known to be self-dual, see \cite[Ex~7.2.9]{Ma} at least for generic $q$. In our conventions, it is the sub-Hopf algebra of $u_q(sl_2)$ in Example~\ref{exsl2} generated by $K,F$ (say) and has
\[ \int K^iF^j=\delta_{i,0}\delta_{n-1,j},\quad \Lambda=\Lambda_KF^{n-1}.\]
We compute this for $n=3$. Then $\Lambda=(1+K+K^2)F^2$ and
\begin{align*} g_{\color{red}\bullet}&=F^2\mathop{{\otimes}} 1+q^2K^2F\mathop{{\otimes}} KF+q^2K\mathop{{\otimes}} K^2F^2\\
&\quad +KF^2\mathop{{\otimes}} K^2+F\mathop{{\otimes}} F+qK^2\mathop{{\otimes}} KF^2\\
&\quad +K^2F^2\mathop{{\otimes}} K+qKF\mathop{{\otimes}} K^2F+1\mathop{{\otimes}} F^2\end{align*}
\[ (K^iF^m){\color{green}\bullet}(K^jF^k)=\delta_{[i-j],2-k}\left[{m\atop 2-k}\right]_q(-1)^kq^{\delta_{k,1}}K^jF^{m+k-2}\]
treated as 0 unless $m+k\ge 2$. Here $i,j$ are mod 3 and $[i-j]$ denotes the value in the range 0,1,2 and the $q$-binomials are defined using $q$-integers. We used the $q$-binomial formula to compute $\Delta(K^iF^m)$ and pick off the term with the right power of $F$ for the integral to not vanish. We let
\[ t=(1+qK +q^2K^2)F^2,\quad x=(1+K+K^2)F\]
then we find
\[ t{\color{green}\bullet}t{\color{green}\bullet}t=\Lambda,\quad x{\color{green}\bullet}t=(1+qK+q^2K^2)F,\quad t{\color{green}\bullet}x=qx{\color{green}\bullet}t,\quad x{\color{green}\bullet}x{\color{green}\bullet}x=0,\]
where $\Lambda=1_{\color{green}\bullet}$. The red coproduct is given by multiplication by $g_{\color{red}\bullet}$ but this is central, so we can conveniently compute it as
\[\Delta_{\color{red}\bullet}(K^iF^j)=K^ig_{\color{red}\bullet}F^j;\quad \Delta_{\color{red}\bullet} t=t\mathop{{\otimes}} t,\quad \Delta_{\color{red}\bullet}x=x\mathop{{\otimes}} t^{-1}+ \Lambda\mathop{{\otimes}} x.\]
Thus, $H$ is isomorphic to the opposite of the associated Hopf algebra by $K\mapsto t^{-1}$ and $F\mapsto t{\color{green}\bullet} x$, so we have a Type 2 Hadamard form. We can now read off some of the values of $\hbox{{$\mathfrak h$}}$ and deduce the rest from the algebra homomorphism property. Thus, on the vector space of $H$,
\begin{align*} \hbox{{$\mathfrak h$}}(K^i)&=\delta_{-i}F^2,\quad
\hbox{{$\mathfrak h$}}(K^iF)=q\delta_{1-i} F,\quad \hbox{{$\mathfrak h$}}(K^iF^2)=\delta_{2-i},\end{align*}
where we define
\[ \delta_i=\sum_j q^{ij} K^j.\]
We see some similarities to the $\mathbb{Z}_n$ Fourier transform. \end{example}
\begin{example}\rm Let $H=u_{-1}(sl_2)$ as studied in Example~\ref{exsl2}. This is both self dual and anti-self dual as a Hopf algebra. Here, we already gave an isomorphism from $H$ to the associated Hopf algebra by specifying it on the generators, and deduce the rest now as
\[ \hbox{{$\mathfrak h$}}(1)=(1+K)EF,\quad \hbox{{$\mathfrak h$}}(K)=(K-1)EF,\quad\hbox{{$\mathfrak h$}}(E)=(1+K)E,\quad\hbox{{$\mathfrak h$}}(F)=(1-K)F\]
\[ \hbox{{$\mathfrak h$}}(KE)=(K-1)E,\quad \hbox{{$\mathfrak h$}}(KF)=(1+K)F,\quad \hbox{{$\mathfrak h$}}(FE)=-K,\quad \hbox{{$\mathfrak h$}}(KFE)=1+K\]
as the corresponding Type 3 Hadamard gate.
\end{example}
The last example here as well as $\mathbb{C} Z_n$ are flip-Hopf *-algebras, so it is reasonable to ask that $\Theta$ and $\hbox{{$\mathfrak h$}}$ be compatible with the $*$-algebra structure when they exist. We do not have a general analysis for this, but for $\mathbb{C} \mathbb{Z}_n$ we have $\Theta(i,j)=q^{ij}$ as a matrix is unitary up to a constant factor $n$ in our orthonormal basis defined from the $*$-structure. It is not, however, the case in this example that $\hbox{{$\mathfrak h$}}$ as an operator is a $*$-algebra map, but that is the case in the $u_{-1}(sl_2)$ example.
To close this section, we note that for a more categorical view on Frobenius $*$-algebras and F-Hopf $*$-algebras, one cannot work in the category of vector spaces alone since $*$ is not a morphism of vector spaces. An approach that is useful in noncommutative geometry is the notion of a bar-category\cite{BegMa:bar}\cite[Chap.~2.8]{BegMa}, in which a monoidal category is equipped with a functor `bar' that conjugates objects. Here every complex vector space $V$ has a conjugate $\bar V$ which is the same abelian group but $z\in \mathbb{C}$ acts by its conjugate. Then for a $*$-algebra $A$, $*:A\to \bar A$ is a linear map. There are related notions in the Computer Science literature.
\section{Braided interacting Hopf algebras}\label{secbraZX}
Braided ZX calculus was considered in \cite{BPW}, where it was shown that the bialgebra rewrite rules as stated there force the category to be symmetric. Here we revisit this from the point of view of building up the required algebraic axioms or `rewrite rules' by understanding the correct construction of a braided interacting pair of Hopf algebras.
We recall that a braided category is first of all a monoidal category, so there is a functor for $\mathop{{\otimes}}$ of objects and morphisms, with unit object $\underline 1$ and associated maps including an natural equivalence $(\ \mathop{{\otimes}}\ )\mathop{{\otimes}} \to \mathop{{\otimes}}(\ \mathop{{\otimes}}\ )$ (the associator). The Mac Lane coherence theorem for monoidal categories says that we can ignore brackets and insert the associator as needed. In addition, there is a natural equivalence $\Psi: \mathop{{\otimes}}\to \mathop{{\otimes}}^{op}$ (the braiding). The coherence theorem for braided categories was in \cite{JoyStr}, replacing symmetric group by the braid group in Mac Lane's coherence theorem for symmetric monoidal categories. This underlies the diagrammatatic notation for working in braided categories, with $\Psi=\includegraphics{braid.pdf}$ a braid crossing, which in turn underlies the theory of braided-Hopf algebras\cite{Ma:bra,Ma:tra,Ma:cro,Ma:alg}. The axioms for these are as in Section~\ref{secFHopf}, but now caring about over- and under- crossings. The most important thing we need from the theory of braided-Hopf algebras is a result\cite{Ma:pri,Ma:alg} that
\begin{equation}\label{antialg} \includegraphics{antant.pdf}\end{equation}
where we have inserted the factor for an unnormalised Hopf algebra. Here $S$ denotes the antipode of a braided Hopf algebra $H$ with red product and green coproduct, not necessarily of any special form. We proceed in the normalised case of a usual braided Hopf algebra.
We already used this {\em braided anti-homomorphism} property in the proofs in the preceding section, with the braided theory of the present section in mind.
The definition of an F-Hopf algebra still makes sense declaring $H$ to be braided, but we saw in the proof of Proposition~\ref{propFhopf} {\em one spot} where we needed a wrong braid crossing. Thus, what we actually proved in parts (a)-(b) of Figure~\ref{figFhopf} was the following.
\begin{corollary} Let $H$ be a Hopf algebra in a braided category $\hbox{{$\mathcal C$}}$ with red algebra and green coalgebra Frobenius and with antipode of the form in Proposition~\ref{propFhopf}. Then the associated green algebra and red coalgebra form a Hopf algebra $\tilde H$ in the braided category with reversed braid crossings (or a reverse-braided Hopf algebra in $\hbox{{$\mathcal C$}}$).
\end{corollary}
In fact, the opposite algebra to a Hopf algebra with reversed braiding is a braided Hopf algebra in the original category\cite{Ma:pri}. Equivalently, the categorical dual $H^\medstar$ of $H$ remains in $\hbox{{$\mathcal C$}}$\cite{Ma:alg,Ma:pri} and part (c) of Figure~\ref{figFhopf} says that if we compute this using the red duality then we indeed land on the opposite product. Hence, we also proved the following.
\begin{corollary}\label{propbraFhopf} If a braided Hopf algebra $H$ is Frobenius as an algebra and coalgebra and its antipode $S$ has the special form in Proposition~\ref{propFhopf} then
\[ \includegraphics[scale=0.85]{specialtildeH.pdf}\]
are a pair of braided Hopf algebras on the same object $H$ while the red and green maps are a pair of F-algebras. The second braided Hopf algebra is isomorphic to $H^\medstar$. \end{corollary}
Either point of view gives a canonical braided F-Hopf algebra amplified under our assumptions from a single braided Hopf algebra. Note that, as with ordinary F-Hopf algebras, it is not clear that there are many examples beyond these, so this is more of a construction rather than a general concept of interacting Hopf algebras. We let $\underline 1$ be the unit object of the braided category.
\begin{corollary}\label{corbraHH*} Let $H$ be a braided-Hopf algebra with invertible antipode equipped with a morphism $\Lambda:\underline 1\to H$ which is a left integral and a morphism $\int:H\to \underline 1$ which is a right integral, such that $\int\Lambda$ is the identity on $\underline 1$. Then Proposition~\ref{propbraFhopf} applies in the braided category and we have a braided F-Hopf algebra.
\end{corollary}
{\noindent {\bfseries Proof:}\quad } We follow the definitions and steps in the proof of Corollary~\ref{corHH*} but now as a morphism in a braided category. Thus, we set
\[ (\ ,\ )_{\color{red}\bullet}=\int\circ\mu_{\color{red}\bullet},\quad g_{\color{red}\bullet}=(\mathrm{id}\mathop{{\otimes}} S)\Delta\Lambda;\quad (\ ,\ )_{\color{green}\bullet}=(\ ,\ )_{\color{red}\bullet}(S\mathop{{\otimes}}\mathrm{id}),\quad g_{\color{green}\bullet}=(\mathrm{id}\mathop{{\otimes}} S^{-1})g_{\color{red}\bullet}.\]
The proof for the red Frobenius form is then shown in Figure~\ref{figbraint}. Part (a) uses the braided-antimultiplicativity property for the 3rd equality and the integral property of $\int$ for the 6th. At the end, we need that $\Lambda$ or $\int$ are morphisms to obtain the identity map. Part (b) uses braided-antimultiplicativity for the 1st equality, the morphism and integral property of $\Lambda$ for the 6th and the morphism and integral property of $\int$ for the 8th. These results together imply in (c) that $\int S\Lambda$ is the identity on $\underline 1$ under either morphism assumption. It is immediate that we have the duality properties for the green form and that this makes the coalgebra Frobenius, and that $S$ then has the form required in Proposition~\ref{propbraFhopf}.\endproof
\begin{figure}
\[ \includegraphics[scale=1]{braint.pdf}\]
\caption{\label{figbraint} Construction of Frobenius form from integrals $\Lambda,\int$ on a braided Hopf algebra.}
\end{figure}
In fact, integrals on braided-Hopf algebras are not typically bosonic (i.e. of trivial braiding), so this involves a certain complication already familiar in the literature on braided Fourier transform\cite{BesLyu}. This remark already applies to the {\em braided line} $\mathbb{C}[x]/(x^n)$ with degree $|x|=-1$ in the category of $\mathbb{Z}_n$-graded spaces with primitive coalgebra and antipode
\[ \Delta x^m=\sum_{r=0}^m\left[{m\atop r}\right]_q x^r\mathop{{\otimes}} x^{m-r},\quad S x^m= q^{m(m-1)\over 2}(-x)^m,\quad {\epsilon} x^m=\delta_{m,0}\]
extended with braiding $\Psi(x^i\mathop{{\otimes}} x^j)=q^{ij}x^j\mathop{{\otimes}} x^i$. This has an integral $\int: H\to K$ where $K=k$ regarded as a 1-dimensional object of grade -1 and given by $\int x^m=\delta_{n-1,m}$ and an integral element $\Lambda=x^{n-1}$ also of grade -1. As a result, the corollary does not apply on this simplest example, i.e.
\[ (x^a,x^b)_{\color{red}\bullet}=\delta_{n-1,a+b},\quad g_{\color{red}\bullet}=\sum_a x^a\mathop{{\otimes}} x^{n-1-a},\quad \mu_{\color{red}\bullet}(g_{\color{red}\bullet})=n x^{n-1}\]
where $g_{\color{red}\bullet}$ provides the inverse to $(\ ,\ )_{\color{red}\bullet}$ but is not of the expected form $(\mathrm{id}\mathop{{\otimes}} S)\Delta x^{n-1}$. Likewise,
\[ (x^a,x^b)_{\color{green}\bullet}={\delta_{n-1,a+b}\over \left[{n-1\atop a}\right]_q},\quad g_{\color{green}\bullet}=\Delta x^{n-1}=\sum_{a=0}^{n-1}\left[{n-1\atop a}\right]_q x^a\mathop{{\otimes}} x^{n-1-a}\]
where $(\ ,\ )_{\color{green}\bullet}$ provides the correct inverse to $g_{\color{green}\bullet}$, but is not of the expected form $(S(\ ),\ )_{\color{red}\bullet}$. As a result, $S$ is not of the required form for the braided reading of Proposition~\ref{propbraFhopf}.
\subsection{Self-dual braided interacting pairs by transmutation} \label{sectrans}
On the other hand, all quasitriangular Hopf algebras, such as $u_q(sl_2)$, have a self-dual braided Hopf algebra version via a process of {\em transmutation}\cite{Ma:bra,Ma:tra,Ma,Ma:pri}. We show that these provide examples for the braided theory. They also underly TQFT invariants and hence should be relevant to fault tolerant TQFT based quantum computing.
Let $(H,\hbox{{$\mathcal R$}})$ be an ordinary quasitriangular Hopf algebra\cite{Dri,Ma}. Its transmutation $\underline H$ has the same algebra and counit but a different coproduct and antipode\cite{Ma:bra,Ma:tra}
\begin{equation}\label{trans} \underline{\Delta}h=h_1S\hbox{{$\mathcal R$}}^2\mathop{{\otimes}} \hbox{{$\mathcal R$}}^1{\triangleright} h_2,
\quad \underline S h=\hbox{{$\mathcal R$}}^2 S(\hbox{{$\mathcal R$}}^1{\triangleright} h)\end{equation}
forming a braided-Hopf algebra in the category of $H$-modules as an object by the adjoint action
\begin{equation}\label{adj} h{\triangleright} h'= h_1 h' S h_2.\end{equation}
Throughout this section, we will use underlines to separate the braided Hopf algebra structures from the original ordinary ones. We will also be interested in the {\em factorisable} case which, by definition, means that the quantum Killing form $\hbox{{$\mathcal Q$}}:=\hbox{{$\mathcal R$}}_{21}\hbox{{$\mathcal R$}}$ is invertible when viewed as a linear map $H^*\to H$ by evaluation against its first factor. Here, if $S\Lambda=\Lambda$ then $h{\triangleright} \Lambda={\epsilon}(h)\Lambda$, so we can regard it as a morphism from the trivial object to $\underline{H}$. The following is a reworking of arguments in \cite{LyuMa}.
\begin{proposition}\label{proptransH} Let $H$ be a factorisable quasitriangular Hopf algebra with $S\Lambda=\Lambda$. Then Corollary~\ref{corbraHH*} applies to $\underline H$ and provides a braided F-Hopf algebra in the braided category of $H$-modules.
\end{proposition}
{\noindent {\bfseries Proof:}\quad } If $S\Lambda=\Lambda$ then $h{\triangleright} \Lambda={\epsilon}(h)\Lambda$, so we can regard it as a morphism from the trivial object to $\underline{H}$. Next, $H^*$ is coquasitriangular and a dual transmutation construction keeps the same coalgebra but modifies the algebra, making it into a braided-Hopf algebra, which can be viewed in the same braided category by the quantum coadjoint action. In this case, $\hbox{{$\mathcal Q$}}$ as a map becomes an isomorphism of braided-Hopf algebras, see \cite{Ma,Ma:pri} for more details. Finally, the right integral $\int$ on $H^*$ is also a right integral on the braided version and therefore transfers under the braided Hopf algebra self-duality isomorphism to a right integral $\underline\int$ on $\underline H$. By arguments \cite[Lem.~3.5]{LyuMa}, it is ad-invariant and hence a morphism. Therefore, we meet the conditions of Corollary~\ref{corbraHH*}. \endproof
Rather than more formal details, will now show how this all works in the simplest quantum group example.
\begin{example}\rm Let $q$ be a primitive $n$-th root of unity {\em with $n$ odd} and $H=u_q(sl_2)$ as in Example~\ref{exsl2}. In this case $q$ has a square root $q^{n+1\over 2}$ and $u_q(sl_2)$ in our conventions (denoted by gothic in \cite{AziMa}) is isomorphic to $u_{q^{1\over 2}}(sl_2)$ in usual conventions as shown there. Likewise for our conventions for the reduced $c_q[SL_2]$ as its Hopf algebra dual. The braided Hopf algebra $\underline{H}=b_q(sl_2)$ has the same algebra as $u_q(sl_2)$ and hence $\Lambda=\delta_0(K)F^nE^n$ there is also a left integral in $H$, where we use the notation $\delta_i:=\sum_i q^iK^i$. We have $S\Lambda=\Lambda$, so this is invariant under the adjoint action, which is equivalent to $\Lambda\in Z(H)$. The latter can also be checked from the commutation relations. The coalgebra of $b_q(sl_2)$ is obtained from that of $u_q(sl_2)$ via (\ref{trans}) using the adjoint action and the quasitriangular structure.
Also when $n$ is odd, $u_q(sl_2)$ is factorisable and we have a self-duality braided-Hopf algebra isomorphism $b_q(sl_2){\cong} b_q[SL_2]$, where $b_q[SL_2]$ is obtained by a dual transmutation construction from. It is built on the same coalgebra and viewing the matrix of generators of $c_q[SL_2]$ in $b_q[SL_2]$, where we denote them $\alpha,\beta,\gamma,\delta$, these have the relations, coproduct and antipode
\[ \alpha^n=1,\quad\beta^n=\gamma^n=0;\quad\quad\beta\alpha=q\alpha\beta,\quad \gamma\alpha=q^{-1}\alpha\gamma,\quad\delta\alpha=\alpha\delta, \quad \alpha\delta-q\gamma\beta=1\]
\[ [\beta,\gamma]=(1-q^{-1})\alpha(\delta-\alpha),\quad [\gamma,\delta]=(1-q^{-1})\gamma\alpha,\quad [\delta,\beta]=(1-q^{-1})\alpha\beta, \]
\[\underline{\Delta}\begin{pmatrix}\alpha&\beta\\ \gamma & \delta\end{pmatrix}=\begin{pmatrix}\alpha&\beta\\ \gamma & \delta\end{pmatrix}\mathop{{\otimes}} \begin{pmatrix}\alpha&\beta\\ \gamma & \delta\end{pmatrix},\quad \underline{S}\begin{pmatrix}\alpha&\beta\\ \gamma & \delta\end{pmatrix}=\begin{pmatrix}q\delta+(1-q)\alpha&-q\beta\\ -q\gamma & \alpha\end{pmatrix}\]
and $\underline{\epsilon}(\alpha)=\underline{\epsilon}(\delta)=1, \underline{\epsilon}(\beta)=\underline{\epsilon}(\gamma)=0$. We replaced $q^2$ in \cite{Ma} by $q$ to match our conventions. The braiding $\Psi$ on the matrix of generators is given explicitly in \cite[Ex~10.3.3]{Ma}, upon making the same change from $q^2$ to $q$. Here
\[ \Psi(\alpha\mathop{{\otimes}}\alpha)=\alpha\mathop{{\otimes}}\alpha+(1-q)\beta\mathop{{\otimes}}\gamma,\quad \Psi(\alpha\mathop{{\otimes}}\beta)=\beta\mathop{{\otimes}}\alpha,\quad\Psi(\alpha\mathop{{\otimes}}\gamma)=\gamma\mathop{{\otimes}}\alpha+(1-q)(\delta-\alpha)\mathop{{\otimes}}\gamma\]
\[ \Psi(\beta\mathop{{\otimes}}\alpha)=\alpha\mathop{{\otimes}}\beta+(1-q)\beta\mathop{{\otimes}}(\delta-\alpha),\quad\Psi(\beta\mathop{{\otimes}}\beta)=q\beta\mathop{{\otimes}}\beta,\]\[\quad \Psi(\beta\mathop{{\otimes}}\gamma)=q^{-1}\gamma\mathop{{\otimes}}\beta+(1+q)(1-q^{-1})^2\beta\mathop{{\otimes}}\gamma-(1-q^{-1})(\delta-\alpha)\mathop{{\otimes}}(\delta-\alpha)\]
\[ \Psi(\gamma\mathop{{\otimes}}\alpha)=\alpha\mathop{{\otimes}}\gamma,\quad\Psi(\gamma\mathop{{\otimes}}\beta)=q^{-1}\beta\mathop{{\otimes}}\gamma,\quad\Psi(\gamma\mathop{{\otimes}}\gamma)=q\gamma\mathop{{\otimes}}\gamma\]
with the braiding on products determined by iterated $\Psi$ as the algebra product is a morphism and hence passes through braid crossings. For example, $\Psi(\gamma\mathop{{\otimes}}\delta)=\delta\mathop{{\otimes}}\gamma$ and $\Psi(\delta\mathop{{\otimes}}\beta)=\beta\mathop{{\otimes}}\delta$ can be deduced.
The self-duality isomorphism to $b_q(sl_2)$ is likewise given in our conventions by
\[ \begin{pmatrix}\alpha&\beta\\ \gamma & \delta\end{pmatrix}\mapsto \begin{pmatrix}K& (1-q^{-1})E\\
(q^{-2}-1)KF & K^{-1}+(1-q)(1-q^{-2})FE\end{pmatrix}.\]
Moreover, as $b_q[SL_2]$ has the same coalgebra, it gets the right-invariant integral of $c_q[SL_2]$. Hence under the isomorphism, $b_q(sl_2)$ gets a right-invariant integral $\underline\int$ at least for odd $n$. Also note that the element $c$ defined as the image under the isomorphism of $\alpha+q\delta$ (a multiple of the $q$-trace), namely
\[ c=K+qK^{-1}+(1-q)(q-q^{-1})FE\]
is central and hence invariant. Thus, $1,c,c^2,\Lambda$ are bosonic (they have trivial braiding with everything) and we also necessarily have $\underline Sc=Sc=c$ and $\underline S\Lambda= S\Lambda=\Lambda$ using the general formula for $\underline S$.
Finally, we start with $\int$ on $c_q[SL_2]$ as in \cite{AziMa}, which when converted to the matrix generators $a,b,c$ (the fourth matrix generator $d$ is determined in the root of unity case) amounts to
\[ \int a^ib^jc^k=\delta_{i,0}\delta_{n-1,j}\delta_{n-1,k}.\]
We take take this $b_q[SL_2]$ also as it has the same vector space and map it over to $b_q(sl_2)$ under the isomorphism of braided Hopf algebras. Corollary~\ref{corbraHH*} then applies and we have a Frobenius structures for the algebra and coalgebra to give a braided F-Hopf algebra.
In what follows, we focus on $n=3$, so $q$ is a primitive cube root of unity and $1+q+q^2=0$. Using the transmutation formulae \cite[Eqn (7.37)]{Ma} extended to degree 4 gives $c^2b^2=q c\bullet c\bullet b\bullet b=q\gamma^2\beta^2$ where the modified product is denoted $\bullet$ and we then view the elements in $b_q[SL_2]$. Thus, the braided integral is zero except on $\gamma^2\beta^2$ in the monomial basis for $\gamma,\beta,\alpha$. Normalised so that $\underline\int \Lambda=1$, we have after the isomorphism with $b_q(sl_2)$,
\[ \underline\int K^iF^jE^j=\delta_{2,i}\delta_{2,j}\delta_{2,k}\]
as the braided integral. It is easy to see that this is invariant in a Hopf algebra sense under the quantum adjoint action and hence a morphism. Also note that this is not the same as the integral on $u_q(sl_2)$ as the coproduct is different, but similar.
We will now do some checks of the construction. From the matrix form of braided coproduct of $b_q[SL_2]$ and the stated isomorphism, we can read off some coproducts and antipodes
\begin{align*}
\underline\Delta K&=K\mathop{{\otimes}} K+q^{-1}(q-1)^2E\mathop{{\otimes}} KF,\quad \underline S K=c-qK,\\
\underline\Delta E&=E\mathop{{\otimes}} q^{-1}(c-K)+K\mathop{{\otimes}} E,\qquad \underline S E=-qE,\\
\underline\Delta(KF)&=KF\mathop{{\otimes}} K+q^{-1}(c-K)\mathop{{\otimes}} KF,\quad\underline S(KF)=-qKF\end{align*}
where $\delta$ maps under the isomorphism to $q^{-1}(c-K)$.
The second of these is relatively easy to compute directly using the first four of
\begin{gather*}K{\triangleright} K=K,\quad K{\triangleright}(KF)=q KF,\quad \\ F{\triangleright} K=(q^{-1}-1)KF,\quad F{\triangleright}(KF)=0,\quad F{\triangleright} K^2=(q-1)K^2F\end{gather*}
and $\hbox{{$\mathcal R$}}$ from Example~\ref{exsl2}, as higher order terms of $\hbox{{$\mathcal R$}}$ do not contribute. One can also check that the algebras are isomorphic. Some examples of the braiding can likewise be read off as
\begin{align*} \Psi(K\mathop{{\otimes}} E)&=E\mathop{{\otimes}} K,\quad \Psi(E\mathop{{\otimes}} E)=qE\mathop{{\otimes}} E,\quad \Psi(KF\mathop{{\otimes}} K)=K\mathop{{\otimes}} KF,\\
\Psi(KF\mathop{{\otimes}} E)&=q^{-1}E\mathop{{\otimes}} KF,\quad \Psi(KF\mathop{{\otimes}} KF)=qKF\mathop{{\otimes}} KF,\\
\Psi(E\mathop{{\otimes}} K)&=K\mathop{{\otimes}} E+ (1-q)E\mathop{{\otimes}} (q^{-1}c+qK),\\
\Psi(K\mathop{{\otimes}} KF)&=KF\mathop{{\otimes}} K+(1-q)(q^{-1}c+qK)\mathop{{\otimes}} KF\\
\Psi(K\mathop{{\otimes}} K)&=K\mathop{{\otimes}} K-q^{-1}(q-1)^3E\mathop{{\otimes}} KF,\\
\Psi(E\mathop{{\otimes}} KF)&=q^{-1}KF\mathop{{\otimes}} E-(q-1)^2E\mathop{{\otimes}} KF-{1\over q-1}(q^{-1}c+qK)\mathop{{\otimes}} (q^{-1}c+qK)
\end{align*}
which allows us to compute more coproducts and antipodes using the braided-homomorphism and anti-homomorphism properties respectively. For example,
\begin{align*}\underline S(E^2)&=\cdot\Psi((-qE)\mathop{{\otimes}}(-qE))=E^2\\
\underline S (K^2F^2)&=q\cdot\Psi( (-q KF)\mathop{{\otimes}}(-qKF))=qKFKF=K^2F^2\\
\underline S(KE)&=\cdot\Psi((c-qK)\mathop{{\otimes}}(-qE))=qE(qK-c)=(K-qc)E\\
\underline S (K^2)&=\cdot\Psi((c-qK)\mathop{{\otimes}}(c-qK))= c^2+q^2(K^2+(1-q)(1-q^{-1})(q^{-2}-1)EKF)- 2 q c K\\
&=c^2+ c K + K^2 + q(q-1)=:s(c,K)\\
\underline S(K^2F)&=q\underline S((KF)K)=q\cdot\Psi((-qKF)\mathop{{\otimes}}(c-qK))=-q^2(c-qK)KF=(K-q^2c)KF\\
\underline S(KE^2)&=\cdot\Psi((c-qK)\mathop{{\otimes}} E^2)=E^2(c-qK)=(c-K)E^2\\
\underline S(F^2)&=\underline S((KFKF)K)=\cdot\Psi(KFKF\mathop{{\otimes}}(c-qK))=(c-qK)KFKF=(q^{-1}cK^2-1)F^2\\
\underline S F&=q^{-1}\underline S(KFK^2)=q^{-1}\cdot\Psi((-qKF)\mathop{{\otimes}} s(c,K))=- s(c,K)KF\\%\end{align*}\begin{align*}
\underline S(KFE)&=\underline S((KF)E)=\cdot \Psi((-qKF)\mathop{{\otimes}}(-qE))=q EKF=q^2 KEF\\
\underline S(F^2E^2)&=q^2\underline S(K^2F^2KE^2)=\cdot(KFKF\mathop{{\otimes}} (c-K)E^2)=q^{-1}(c-K)E^2KFKF=(cK^2-1)E^2F^2\\
\underline S(K^2E)&=\cdot\Psi(s(c,K)\mathop{{\otimes}} (-qE))=-q Es(c,K)\\
\underline S(K^2E^2)&=\cdot\Psi(s(c,K)\mathop{{\otimes}} E^2)=E^2s(c,K)\\
\underline S(FE^2)&=q^{-1}\underline S(KF(K^2E^2))=-\cdot\Psi(KF\mathop{{\otimes}} E^2 s(c,K))=-q^{-2}E^2 s(c,K) KF\\
\underline S(KF^2E)&=q^{-1}\underline S(KFKF(K^2 E))=-\cdot\Psi(KFKF\mathop{{\otimes}} Es(c,K))=-q^{-2}E s(c,K)KFKF\\
&=-E s(c,K)K^2F^2.
\end{align*}
We now let $(\ ,\ )_{\color{red}\bullet}=\underline\int\mu_{\color{red}\bullet}$ and $g_{\color{red}\bullet}=(\mathrm{id}\mathop{{\otimes}}\underline S)\underline\Delta \Lambda$. By construction, these will be inverse to each other and
as a spot check (to make sure there was no confusion in our conventions) we demonstrate this on a sample element, $F$, say, i.e. we check that
$(F,g^1_{\color{red}\bullet})_{\color{red}\bullet} g^2_{\color{red}\bullet}=F$. The same method will apply on other elements similarly. We start with the unbraided coproduct
\begin{align*} \Delta\Lambda&=(1\mathop{{\otimes}} 1+K\mathop{{\otimes}} K+K^2\mathop{{\otimes}} K^2)(F^2\mathop{{\otimes}} 1+K\mathop{{\otimes}} F^2-q^2K^2F\mathop{{\otimes}} F)\\
&\qquad\qquad\qquad (1\mathop{{\otimes}} E^2+E^2\mathop{{\otimes}} K^2-q^2E\mathop{{\otimes}} KE)\\
&=(1\mathop{{\otimes}} 1+K\mathop{{\otimes}} K+K^2\mathop{{\otimes}} K^2)\big(F^2\mathop{{\otimes}} E^2+F^2E^2\mathop{{\otimes}} K^2-q^2F^2E\mathop{{\otimes}} KE\\
&\qquad\qquad\qquad+K\mathop{{\otimes}} F^2E^2+q^2KE^2\mathop{{\otimes}} K^2F^2-KE\mathop{{\otimes}} KF^2E\\ &\qquad\qquad\qquad -q^2K^2F\mathop{{\otimes}} FE^2-K^2FE^2\mathop{{\otimes}} K^2F+K^2FE\mathop{{\otimes}} KFE\big)\end{align*}
and we now compute
\begin{align*}\Big( \underline\int F&\Lambda_1 S \hbox{{$\mathcal R$}}^2 \Big) \underline S(\hbox{{$\mathcal R$}}^1{\triangleright} \Lambda_2)=\left({1\over 3}\sum_{a,b,r}\int {(-1)^r(q-q^{-1})^r\over[r]_{q^{-1}}!} q^{-ab} F\Lambda_1 S(E^r K^b)\right) \underline S((F^r K^a){\triangleright} \Lambda_2)\\
&=\left(\sum_{r}\int {(q-q^{-1})^r\over[r]_{q^{-1}}!} F\Lambda_1 K^2q^{-{r(r+1)\over 2}}E^r\right) \underline S(F^r{\triangleright} \Lambda_2)\\
&=\left(\int F(-FE^2)K^2\right)\underline S F+q^{-1}(q-q^{-1})\left(\int F(FE)K^2E\right)\underline S F{\triangleright}(K^2FE)\\
&\qquad+{(q-q^{-1})^2q^{-3}\over1+q^{-1}}\left(\int F( -q^2F)K^2E^2\right)\underline S F^2{\triangleright}(KFE^2)\\
& = \underline S\big( -F+q(1-q)F{\triangleright}(K^2FE)+(1-q)^2F^2{\triangleright}(KFE^2)\big)\\
&=\underline S\big( -F+(1-q)^{-1}F{\triangleright}(c K^2-1-qK)+F^2{\triangleright}((q^2 cK-q^2 K^2-1)E)\big)\\
&=\underline S\big( -F+(1-q)^{-1}F{\triangleright}(c K^2-q K-q c^2 K+q^2K)\big)\\
&=\underline S( -F+ (q-1-c^2-cK)KF)=F\end{align*}
as required. Here, in order to have a nonzero integral, $\Lambda_1$ must involve $F E^{2-r}$, which means $\Lambda_2$ has $F E^r$ and hence $K^a{\triangleright} \Lambda_2=q^{(1-r)a}\Lambda_2$. We then do the sum over $a$ which absorbs the 1/3 and sets $b=1-r$, which is our second expression on using $S E^r=(-1)^r q^{-{r(r+1)\over 2}}K^{-r}E^r$. For the example of $F$ that we are testing on, the relevant terms in the 2nd displayed factor of $\Delta\Lambda$ are one of $ -K^2FE^2, K^2FE,-q^2K^2F$ according to $r=0,1,2$. In each case the contributing from the prefactor in $\Delta\Lambda$ must be $K\mathop{{\otimes}} K$ in order for the total power of $K$ in the integrand to be 2. This gives the 3rd expression. Evaluating the integrals gives the 4th expression and we then replace $FE$ in terms of $c$ for the 5th. Next,
\begin{align*}F{\triangleright} &((cK-K^2-q)E)=F(cK-K^2-q)E+K^{-1}(cK-K^2-q)E(-KF)\\
&=(cq^{-1}-qK^2-q)FE-q(cK-K^2-q)EF={1\over q-1}((c^2-q)K-q c-1)\end{align*}
which leads to the 6th expression when the further $F{\triangleright}$ is combined with the existing one. We then compute $F{\triangleright}$ noting that $c$ is invariant and $F{\triangleright} K, F{\triangleright} K^2$ were already given above. We then compute $\underline S$, noting that $c$ is unchanged and passes through it, and using $\underline S F, \underline S(KF)$ and $\underline S(K^2F)$ already found above. In this way, we obtain $F$ as expected.
Replacing $F$ by a different monomial will pick of 3 different terms for $\Lambda_1$, etc, i.e. we can proceed similarly. As in the proof of Corollary~\ref{corbraHH*}, the green Frobenius structures follow from the red ones via $\underline S$ and its inverse, to complete the construction. Moreover, our arguments apply for general odd $n$. \end{example}
Proposition~\ref{proptransH} also applies in principle to other reduced `Lusztig kernel' quantum groups $u_q(\hbox{{$\mathfrak g$}})$, although the precise version of these needed and at which roots of unity is not fully understood, see \cite{AziMa}. Returning to the general construction of $\underline H$ in the factorisable case, as these are self-dual, we also have a Type 2 Hadamard form in the sense of Lemma~\ref{lemsd}.
\begin{proposition} Let $(H,\hbox{{$\mathcal R$}})$ be a factorisable quasitriangular Hopf algebra. Then there is a Type 2 Hadamard form $\underline H\mathop{{\otimes}} \underline H\to \underline 1$ according to the braided version of Lemma~\ref{lemsd}, inverse to the metric element
\[ \Theta=(S\mathop{{\otimes}}\mathrm{id})\hbox{{$\mathcal Q$}}\in \underline H\mathop{{\otimes}}\underline H.\]
\end{proposition}
{\noindent {\bfseries Proof:}\quad } This is a version of the self-duality isomorphism between the the transmutation of $H$ and the transmutation of $H^*$, which was used in the construction above. It is known \cite{LyuMa,Ma} that $(S\mathop{{\otimes}}\mathrm{id})\hbox{{$\mathcal Q$}}$ is quantum ad-invariant and hence a morphism from $\underline 1$. That this is invertible when viewed as a map $\underline H^\medstar\to\underline H$ is essentially the assumption of being factorisable. The diagrams in Figure~\ref{fighadpf}(a) and Figure~\ref{fighadvarpf} part (b') when written in terms of the metric element $\Theta$ (rather than the bilinear form itself) amount to
\[ (\underline\Delta\mathop{{\otimes}}\mathrm{id})\Theta=(\mathrm{id}\mathop{{\otimes}}\mu)\Theta_{23}\Theta_{13},\quad (\mathrm{id}\mathop{{\otimes}}\underline\Delta)\Theta=(\mu\mathop{{\otimes}}\mathrm{id})\Theta_{12}\Theta_{23},\quad(\mathrm{id}\mathop{{\otimes}}{\epsilon})\Theta=({\epsilon}\mathop{{\otimes}}\mathrm{id})\Theta=1\]
The counit parts are immediate from $({\epsilon}\mathop{{\otimes}}\mathrm{id})\hbox{{$\mathcal R$}}=(\mathrm{id}\mathop{{\otimes}}\mathrm{id})\hbox{{$\mathcal R$}}=1$ while
\begin{align*} (\underline\Delta&\mathop{{\otimes}}\mathrm{id})\Theta=S(\hbox{{$\mathcal R$}}^{iv 2}\hbox{{$\mathcal R$}}^2{}_2\hbox{{$\mathcal R$}}^{'1}{}_2)\mathop{{\otimes}} \hbox{{$\mathcal R$}}^{iv 1}{}_1S(\hbox{{$\mathcal R$}}^{iv 1}{}_2 \hbox{{$\mathcal R$}}^2{}_1\hbox{{$\mathcal R$}}^{'1}{}_1)\mathop{{\otimes}}\hbox{{$\mathcal R$}}^1\hbox{{$\mathcal R$}}^{'2}\\
&=S(\hbox{{$\mathcal R$}}^{iv2}\hbox{{$\mathcal R$}}^{v2}\hbox{{$\mathcal R$}}^{''2}\hbox{{$\mathcal R$}}^{'''1})\mathop{{\otimes}} \hbox{{$\mathcal R$}}^{iv 1} S(\hbox{{$\mathcal R$}}^{v 1} \hbox{{$\mathcal R$}}^2\hbox{{$\mathcal R$}}^{'1})\mathop{{\otimes}}\hbox{{$\mathcal R$}}^{''1}\hbox{{$\mathcal R$}}^1\hbox{{$\mathcal R$}}^{'2}\hbox{{$\mathcal R$}}^{'''2}\\
&=S(\hbox{{$\mathcal R$}}^{iv2}\hbox{{$\mathcal R$}}^{''2}\hbox{{$\mathcal R$}}^{v2}\hbox{{$\mathcal R$}}^{'''1})\mathop{{\otimes}} \hbox{{$\mathcal R$}}^{iv 1} S( \hbox{{$\mathcal R$}}^2\hbox{{$\mathcal R$}}^{v 1}\hbox{{$\mathcal R$}}^{'1})\mathop{{\otimes}}\hbox{{$\mathcal R$}}^1\hbox{{$\mathcal R$}}^{''1}\hbox{{$\mathcal R$}}^{'2}\hbox{{$\mathcal R$}}^{'''2}\\
&=S(\hbox{{$\mathcal R$}}^{iv2}\hbox{{$\mathcal R$}}^{v2}\hbox{{$\mathcal R$}}^{''2}\hbox{{$\mathcal R$}}^{'''1})\mathop{{\otimes}} \hbox{{$\mathcal R$}}^{iv 1} S( \hbox{{$\mathcal R$}}^2\hbox{{$\mathcal R$}}^{'1}\hbox{{$\mathcal R$}}^{v 1})\mathop{{\otimes}}\hbox{{$\mathcal R$}}^1\hbox{{$\mathcal R$}}^{'2}\hbox{{$\mathcal R$}}^{''1}\hbox{{$\mathcal R$}}^{'''2}\\
&=S(\hbox{{$\mathcal R$}}^{''2}\hbox{{$\mathcal R$}}^{'''1})\mathop{{\otimes}} S( \hbox{{$\mathcal R$}}^2\hbox{{$\mathcal R$}}^{'1})\mathop{{\otimes}}\hbox{{$\mathcal R$}}^1\hbox{{$\mathcal R$}}^{'2}\hbox{{$\mathcal R$}}^{''1}\hbox{{$\mathcal R$}}^{'''2}=\Theta_{23}\Theta_{13}\\
\end{align*}
where $\hbox{{$\mathcal R$}}=\hbox{{$\mathcal R$}}^1\mathop{{\otimes}}\hbox{{$\mathcal R$}}^2$ (summation understood) and the primed/roman superfixes denote independent copies. We used (\ref{trans}) for $\underline\Delta$ and properties of the antipode for the first equality, the quasitriangularity properties $(\Delta\mathop{{\otimes}}\mathrm{id})\hbox{{$\mathcal R$}}=\hbox{{$\mathcal R$}}_{13}\hbox{{$\mathcal R$}}_{23}$ and $(\mathrm{id}\mathop{{\otimes}}\Delta)\hbox{{$\mathcal R$}}=\hbox{{$\mathcal R$}}_{13}\hbox{{$\mathcal R$}}_{12}$ for the 2nd, then the braid relations for $\hbox{{$\mathcal R$}}$ for the 3rd and 4th. For the 5th we use $(S\mathop{{\otimes}}\mathrm{id})\hbox{{$\mathcal R$}}=\hbox{{$\mathcal R$}}^{-1}$ applied to $\hbox{{$\mathcal R$}}^{v}$ and cancelled with $\hbox{{$\mathcal R$}}^{iv}$. Similarly on the other side,
\begin{align*} (\mathrm{id}\mathop{{\otimes}}&\underline\Delta)\Theta=S(\hbox{{$\mathcal R$}}^2\hbox{{$\mathcal R$}}^{'1})\mathop{{\otimes}} \hbox{{$\mathcal R$}}^1{}_1\hbox{{$\mathcal R$}}^{'2}{}_1S\hbox{{$\mathcal R$}}^{iv2}\mathop{{\otimes}}\hbox{{$\mathcal R$}}^{iv1}{}_1(\hbox{{$\mathcal R$}}^1{}_2\hbox{{$\mathcal R$}}^{'2}{}_2)S\hbox{{$\mathcal R$}}^{iv}{}_2\\
&=S(\hbox{{$\mathcal R$}}^2\hbox{{$\mathcal R$}}^{''2}\hbox{{$\mathcal R$}}^{'''1}\hbox{{$\mathcal R$}}^{'1})\mathop{{\otimes}}\hbox{{$\mathcal R$}}^1\hbox{{$\mathcal R$}}^{'2}S(\hbox{{$\mathcal R$}}^{iv2}\hbox{{$\mathcal R$}}^{v2})\mathop{{\otimes}} \hbox{{$\mathcal R$}}^{iv1}\hbox{{$\mathcal R$}}^{''1}\hbox{{$\mathcal R$}}^{'''2}S\hbox{{$\mathcal R$}}^{v1}\\
&=S(\hbox{{$\mathcal R$}}^2\hbox{{$\mathcal R$}}^{''2}\hbox{{$\mathcal R$}}^{'''1}\hbox{{$\mathcal R$}}^{'1})\mathop{{\otimes}}\hbox{{$\mathcal R$}}^1\hbox{{$\mathcal R$}}^{'2}\hbox{{$\mathcal R$}}^{v2}S(\hbox{{$\mathcal R$}}^{iv2})\mathop{{\otimes}} \hbox{{$\mathcal R$}}^{iv1}\hbox{{$\mathcal R$}}^{''1}\hbox{{$\mathcal R$}}^{'''2}\hbox{{$\mathcal R$}}^{v1}\\
&=S(\hbox{{$\mathcal R$}}^2\hbox{{$\mathcal R$}}^{''2}\hbox{{$\mathcal R$}}^{'1}\hbox{{$\mathcal R$}}^{'''1})\mathop{{\otimes}}\hbox{{$\mathcal R$}}^1\hbox{{$\mathcal R$}}^{v2}\hbox{{$\mathcal R$}}^{'2}S(\hbox{{$\mathcal R$}}^{iv2})\mathop{{\otimes}} \hbox{{$\mathcal R$}}^{iv1}\hbox{{$\mathcal R$}}^{''1}\hbox{{$\mathcal R$}}^{v1}\hbox{{$\mathcal R$}}^{'''2}\\
&=S(\hbox{{$\mathcal R$}}^2\hbox{{$\mathcal R$}}^{'1}\hbox{{$\mathcal R$}}^{''2}\hbox{{$\mathcal R$}}^{'''1})\mathop{{\otimes}}\hbox{{$\mathcal R$}}^1\hbox{{$\mathcal R$}}^{'2}\hbox{{$\mathcal R$}}^{v2}S(\hbox{{$\mathcal R$}}^{iv2})\mathop{{\otimes}} \hbox{{$\mathcal R$}}^{iv1}\hbox{{$\mathcal R$}}^{v1}\hbox{{$\mathcal R$}}^{''1}\hbox{{$\mathcal R$}}^{'''2}\\
&=S(\hbox{{$\mathcal R$}}^{''2}\hbox{{$\mathcal R$}}^{'''1})S(\hbox{{$\mathcal R$}}^2\hbox{{$\mathcal R$}}^{'1})\mathop{{\otimes}}\hbox{{$\mathcal R$}}^1\hbox{{$\mathcal R$}}^{'2}\mathop{{\otimes}}\hbox{{$\mathcal R$}}^{''1}\hbox{{$\mathcal R$}}^{'''2}=(\mu\mathop{{\otimes}}\mathrm{id})\Theta_{12}\Theta_{23}
\end{align*}
where for the 3rd equality we used that $(S\mathop{{\otimes}} S)(\hbox{{$\mathcal R$}})=\hbox{{$\mathcal R$}}$. The other steps are similar to the previous case. \endproof
For the associated Hadamard gate we recover $\hbox{{$\mathfrak h$}}^{-1}={\hbox{{$\mathcal S$}}}$ the braided Fourier transform $\underline H\to \underline H$ in \cite{LyuMa}, where it was shown in the case of a ribbon Hopf algebra, such as $u_q(sl_2)$, that ${\hbox{{$\mathcal S$}}}$ along with left multiplication $\hbox{{$\mathcal T$}}$ by the ribbon element, obey the modular identity
\begin{equation}\label{mod} ({\hbox{{$\mathcal S$}}}\hbox{{$\mathcal T$}})^3=\lambda{\hbox{{$\mathcal S$}}}^2\end{equation}
for some constant $\lambda$. This is key to the construction of topological invariants from quantum groups. One also has ${\hbox{{$\mathcal S$}}}^2=\underline S^{-1}$ and hence
\begin{equation}\label{modh2} \mathfrak h^2=\underline S\end{equation}
is the braided antipode. Finally, in the case where $H$ is a quasitriangular flip-Hopf $*$-algebra with $\hbox{{$\mathcal R$}}^\dagger=\hbox{{$\mathcal R$}}^{-1}$, which is the case for $u_q(sl_2)$ in Example~\ref{exsl2}, the category of $H$-modules becomes a braided bar category, see \cite[Prop 3.7, Thm~3.6]{BegMa:bar} for details. In the ribbon case, this is a strong one (meaning that the natural equivalence between applying the bar functor twice and the identity respects tensor products). It is then proven \cite[Thm~5.7]{BegMa:bar} that $\underline H$ becomes a braided $*$-Hopf algebra in the category with a modified operation
\begin{equation}\label{transtar}\underline{*}:\underline H\to \overline{\underline H},\quad h^{\underline *}=(S^2\hbox{{$\mathcal R$}}^1) h^* \hbox{{$\mathcal R$}}^2v;\quad v=\hbox{{$\mathcal R$}}^1 S\hbox{{$\mathcal R$}}^2,\end{equation}
where the element $v\in H$ implements $S^{-2}$ by conjugation as part of Drinfeld theory\cite{Dri} and the output of $\underline{*}$ is viewed in the conjugate object built on the same vector space. Note that in this bar category of $H$-modules, the natural equivalence $\Upsilon:\overline{ V\mathop{{\otimes}} W}\to \overline W\mathop{{\otimes}}\overline{V}$ is, like the braiding, governed by $\hbox{{$\mathcal R$}}$ and is not the usual flip map as for vector spaces, hence $\underline{*}$ is not a usual $*$-involution. We do however, have $\underline{*}\circ\underline S=\underline S^{-1}\circ\underline{*}$, see \cite[Prop.~5.2]{BegMa:bar}. We limit ourselves to one new observation.
\begin{lemma} For $H$ a quasitriangular flip-Hopf $*$-algebra with $\hbox{{$\mathcal R$}}^\dagger=\hbox{{$\mathcal R$}}^{-1}$, we have
\[ \underline S\circ \underline{*}= S\circ *\]
i.e. the underlying antilinear automorphism $\theta=S\circ*$ on $H$ is unchanged by transmutation.
\end{lemma}
{\noindent {\bfseries Proof:}\quad } From the definitions,
\begin{align*}\underline S(h^{\underline*})&=\hbox{{$\mathcal R$}}^2S(\hbox{{$\mathcal R$}}^1{}_1(S^2\hbox{{$\mathcal R$}}^{''1})h^*\hbox{{$\mathcal R$}}^{''2}vS\hbox{{$\mathcal R$}}^1{}_2=\hbox{{$\mathcal R$}}^2\hbox{{$\mathcal R$}}^{'2}S(\hbox{{$\mathcal R$}}^1 (S^2\hbox{{$\mathcal R$}}^{''1})h^*\hbox{{$\mathcal R$}}^{''2}v S\hbox{{$\mathcal R$}}^{'1})\\
&=\hbox{{$\mathcal R$}}^2(\hbox{{$\mathcal R$}}^{'2}S^2\hbox{{$\mathcal R$}}^{'1})(Sv)\hbox{{$\mathcal R$}}^{''2}(Sh^*)S(\hbox{{$\mathcal R$}}^1S\hbox{{$\mathcal R$}}^{''1})=\hbox{{$\mathcal R$}}^2\hbox{{$\mathcal R$}}^{''2}(Sh^*)S(\hbox{{$\mathcal R$}}^1S\hbox{{$\mathcal R$}}^{''1})=Sh^*\end{align*}
where $\hbox{{$\mathcal R$}}^2S\hbox{{$\mathcal R$}}^1=u^{-1}$ in Drinfeld theory and $u=S v$, see \cite[Prop.~2.1.8]{Ma}.\endproof
This suggests that a braided version of Section~\ref{secstar} should be possible, even though significantly more complicated due to working in a nontrivial bar category.
\section{Concluding remarks}
An open question at the algebra level is whether the notion of F-Hopf algebra or its braided version go truly beyond those obtained by amplifying a single Hopf algebra as in Propositions~\ref{propFhopf},~\ref{propbraFhopf}. In principle, the red and green Frobenius structures need not be closely related, with the result that the two Hopf algebras need not be op-dual to each other, but no examples are known even in the unbraided case. We also saw that the simplest version of the $*$-algebra theory in Section~\ref{secuni}, where $(\ ,\ )$ combines with a $*$-structure to give a Hilbert space, does not apply to non-unimodular quantum groups such as $u_q(sl_2)$ with $q$ a primitive odd root of unity. This means that the natural sesquilinear form defined by $\int h^*h'$ does not make $H$ into a Hilbert space since $\overline{\int h^*h'}=\int S( h'{}^*h)$ is not necessarily $\int h'{}^*h$. We saw at the end of Section~\ref{sectrans} that these problems are even more pronounced in the braided case where the natural transmuted $\underline{*}$ on $\underline{H}$, while it makes the latter into a $*$-Hopf algebra in the relevant braided bar category, is not a $*$-algebra structure in the usual sense. This topic of $*$-structures merits further investigation, perhaps guided by is its precise role in quantum computing, which is currently unclear. In particular, we have not developed a general theory for when the `gates' in question are unitary, which would be needed in applications. The situation is rather better for a Drinfeld double $H=D(G)$ of a finite group $G$, which is unimodular.
Another feature of the paper was an abstract notion of `Hadamard gate'. Its inverse appears (\ref{hadfou}) as Hopf algebra Fourier transform composed evaluated against an element $\Theta\in H\mathop{{\otimes}} H$ give an operator. We identified three types, with Type 1 connecting red and green Frobenius structures, which fits well with their usual role in ZX calculus, and Type 2 (resp. Type 3) corresponding to Hopf algebra (anti) self duality of $H$, which fits better with the algebraic picture (where the map on coproducts would naturally go the other way to the map on products related by duality). The practicality of this in quantum computing would need to be looked at further. The theory also applies in the braided case if we have integrals. Here, braided Fourier transform is also of interest in the noncommutative geometry of quantum groups\cite{Ma:hod}, but one has to deal with the fact that the integral is often not a morphism to the trivial object, even on some very basic examples such as the braided line. It would be interesting to adapt the approach to the braided line in \cite{Ma:hod} so as to still obtain a braided F-Hopf algebra, which we saw did not otherwise work. We saw that there is no problem, however, for the braided theory in the case of braided Hopf algebras obtained by transmutation from a factorisable quantum group such as $u_q(sl_2)$ at odd roots of unity. It is also striking that the braided self-duality/Type 2 Hadamard gate that applies canonically in this case obeys\cite{LyuMa} the modular identities in (\ref{mod}), making it central to one approach to quantum group topological invariants.
Finally, while the present constructions are interesting from an algebraic point of view, it remains to be seen how useful they are in quantum computing. The general idea is that while standard ZX-calculus is well adapted to conventional quantum computers, these suffer in practice from noise problems resulting in a current focus is on the development of `topologically fault tolerant' methods. The prototype of these is the Kitaev model\cite{Kit} for which the underlying algebraic structure is a Drinfeld double $D(G)$. This can be generalised to the quantum double of a finite-dimensional Hopf $C^*$-algebras\cite{BMCA} and beyond\cite{CowMa}. As noted in \cite{Meu}, such models are closer in spirit to the Turaev-Viro invariants of 3-manifolds, which have underlying them the quantum double of $u_q(sl_2)$, while the related Jones/Reshetkhin-Turaev knot invariant come from $u_q(su_2)$. The fact that modular identities (\ref{mod}) for the Turaev-Viro invariant have a natural role as Hadamard gates loosely supports the possibility of a braided ZX-calculus approach to topological quantum computing. This could be explored further. Here, Fibonnaci anyons of interest in fault tolerant quantum computing are related to $u_q(sl_2)$ at a 5th root of unity\cite{fib}.
Another setting in which all of these ideas could potentially be explored concretely is `digital quantum computing' as proposed in \cite{MaPac}. This work classifies Hopf algebras over $\mathbb{F}_2$ in low dimension, including which of them are quasitriangular. One could then search for digital Hadamard forms as well look at digital braided versions.
|
\section{Introduction}\label{sec:introduction}
Modern technologies such as Internet of Things and cloud computing have resulted in the collection and storage of a huge amount of data such as images and videos. With the help of such huge amount of data, deep supervised learning \cite{7780459,7298594} has achieved great success, provided that these data are labelled. Unfortunately, labelling such huge datasets is extremely laborious, and in many cases intractable. As a result, many image datasets are not fully utilized due to the lack of labels. In addition to the huge volumes, image data are usually characterized by a high dimensionality and multi-modal structure. Thus, the performance of traditional clustering approaches \cite{BEZDEK1984191,DBLP:conf/iccv/ComaniciuM99,Heller2005BayesianHC,Williams1999AMA,10.1007/978-3-642-33718-5_31} seriously deteriorates on such data \cite{DBLP:conf/iccv/ChangWMXP17,Steinbach2004}. By contrast, deep unsupervised methods have demonstrated superiority and scalability in handling vision data \cite{NIPS2012_c399862d}, including representation learning \cite{caron2020unsupervised,pmlr-v119-chen20j,He_2020_CVPR} and clustering \cite{8237888,DBLP:conf/eccv/HanPPKC20,Ji_2019_ICCV,Ren2020DeepDI,10.1007/978-3-030-58607-2_16}.
Representation learning techniques are gaining popularity, since they can generate discriminative features by considering spatial properties, object shapes and photometric information without the requiring of human annotations. Recent studies \cite{caron2020unsupervised,pmlr-v119-chen20j,He_2020_CVPR,2018arXiv180703748V} have dramatically enhanced the learning capacity and thus minimized the performance gap between supervised and unsupervised learning tasks. This is achieved by proposing self-training objectives for representation learning, also called self-supervised methods.
In particular, contrastive learning strategies have widely been applied in self-supervised methods \cite{ caron2020unsupervised,pmlr-v119-chen20j,pmlr-v9-gutmann10a,2018arXiv180703748V}. They aim to increase the concordance of positive (similar) features of augmented views and decrease the discordance in negative (dissimilar) instances. Alternatively, feature comparisons are replaced in grouping techniques \cite{asano2020self, caron2018deep,caron2020unsupervised} by identifying sub-classes and assigning pseudo-labels to the relevant representations based on their similarities. These pseudo-labelling annotations are either derived through traditional clustering methods \cite{caron2018deep} (i.e., extracted features using a convolutional encoder are annotated through K-means clustering) or through an optimal transport plan \cite{asano2020self,caron2020unsupervised} that can ensure consistency and balance population.
\par
Despite the remarkable success of grouping based self-supervised tasks, the pseudo-labels they generate have so far been evaluated for learning discriminative features only. Additionally, the deep clustering methods have not yet fully exploited the recently developed self-supervised algorithms. Some existing approaches \cite{Hu2017,Ji_2019_ICCV} implement a convolutional framework that introduces mutual information (MI) in the clustering objective during training. The application of MI has demonstrated to be beneficial; however, the performance of these methods degrades in tackling more challenging datasets. Most recently, self-supervised mechanisms were introduced to further improve clustering results on challenging datasets \cite{DBLP:conf/eccv/HanPPKC20,vangansbeke2020scan}, by firstly learning valuable properties of the training instances and then performing the clustering. Although the performance is enhanced, they often require multiple training phases whose clustering performance relies on the previous training phase and the effectiveness of each individual stage. Furthermore, many individual training phases require an additional set of augmentation operations and/or different objective functions. As a result, the final clustering accuracy heavily relies on an increasing number of hyper-parameters. Finally, multiple training phases often increase the computational complexity.
This study aims to address the issues of multiple-phase clustering methods by extending the functionality of Swapping Assignments between multiple Views (SwAV) \cite{caron2020unsupervised}, a recently proposed grouping based self-supervised learning strategy that performs discriminative feature learning. Therefore, we propose an online single-phase clustering framework by integrating a deep classifier net into the SwAV framework to simultaneously learn the representations and assign the desired annotations. Specifically, the proposed clustering method exploits the semantic structure obtained through a discrete probabilistic distribution that is derived by comparing the latent representations with a set of trainable prototypes. This is accomplished by maximizing the mutual dependency between the discrete representation of the integrated classifier and the discrete probabilistic distribution derived by the prototypes. To further enhance the clustering performance, we adopt the multi-crop strategy as presented in \cite{caron2020unsupervised} and optimize the mutual information across multi-crop views. We call the proposed framework Information Maximization Clustering by Swapping Assignments between multiple Views (IMC-SwAV). Our empirical studies show that the proposed framework is highly competitive and outperforms a wide range of state-of-the-art methods. Additionally, it shows robustness to the number of prototypes and the number of clusters. Finally, we demonstrate that other grouping-based self-supervised learning methods such as SeLa \cite{asano2020self} can also be adopted, although the proposed single-phase clustering method is based on an extended SwAV.
We highlight our contributions as follows:
\begin{enumerate}
\item A single-phase training framework is proposed by extending a feature learning strategy to simultaneously generated features and assign labels. This way, the proposed algorithm avoids extra hyper-parameters, and does not require additional transformation functions, nor extra training phases.
\item A modified clustering objective based on mutual information that maximizes the dependency between an over-clustering discrete probabilistic distribution and the classifier's discrete representation. A multi-crop view strategy is adopted in optimizing the classifier to further enhance the classifier's performance.
\item We demonstrate that our clustering strategy is not limited to a specific self-supervised learning strategy and can in principle be extended to any other grouping based self-supervised learning strategies within a single-phase training framework.
\end{enumerate}
The rest of the paper is organized as follows. In Section II, related work on clustering and representation learning are briefly reviewed. The proposed method, including the overall representation learning framework and a modified objective function for joint training on the basis of mutual information maximization is presented in Section III. Comparative experiments and ablation studies, along with the parameter settings and performance metrics, are described in Section IV. A summary of the proposed algorithm and future work are provided in Section V.
\section{Related work}
\textbf{Embedding strategies} have emerged as an effective means to eliminate the necessity of the desired annotations in terms of disentangled feature learning. Previous approaches employ traditional frameworks, such as autoencoders \cite{HintonSalakhutdinov2006b,Vincent2008} and deep belief networks \cite{10.1162/neco.2006.18.7.1527}. Generative models including variational autoencoders \cite{DBLP:conf/aistats/KhemakhemKMH20} and generative adversarial networks \cite{10.5555/2969033.2969125,DBLP:journals/corr/RadfordMC15} also became popular. This type of frameworks rely on two separate models, one generating training instances from a latent space, and the other mapping the training samples into a latent domain. \par
\textbf{Self-supervised learning}, on the other hand, obtains visual representations using a single unit model. A range of training objectives have been proposed in literature: 1) each training sample is treated as a unique class \cite{DBLP:journals/pami/DosovitskiyFSRB16}; 2) an image instance is partitioned in several patches and the encoder either attempts to solve the jigsaw puzzle \cite{10.1007/978-3-319-46466-4_5}, or 3) predicts the code of next image's patch \cite{2018arXiv180703748V}. Contrastive learning addresses feature extraction by comparing positive pairs (augmented views of similar instances). Many implementations \cite{caron2020unsupervised,pmlr-v119-chen20j,He_2020_CVPR,2018arXiv180703748V, DBLP:journals/ijcv/ZhaoXWWTL20} adopt variations of noise contrastive estimation (NCE) \cite{pmlr-v9-gutmann10a} loss function, that achieves a good representation by comparing positive pairs versus a large number of negative instances. To deal with the large negative sampling, momentum contrast (MoCo) \cite{He_2020_CVPR} has been suggested that compares representations between a momentum and a simple encoder, or in \cite{DBLP:journals/ijcv/ZhaoXWWTL20} through a memory bank for storing past representations. As another option, the grouping based strategies \cite{asano2020self,caron2018deep,caron2020unsupervised} assign the representations to a numerous surrogate sub-classes. A method presented in \cite{caron2018deep} employs a traditional K-means to cluster representations, whose performance, however, depends on network initialization, and an additional computational time is required for computing the centroids in each iteration. Some recently proposed algorithms \cite{asano2020self,caron2020unsupervised} take advantage of the efficient optimal transport plan, and introduce a self-labelling mechanism as target distribution which is computed via the Shinkhorn-Knopp \cite{NIPS2013_af21d0c9} algorithm. Nevertheless, these methods aim to extract features without producing cluster annotations.
\begin{figure*}[t]
\centering
\includegraphics[width=1.\textwidth]{figs/diagram.png}
\caption{A diagram presents the framework's structure and a training instance $x_i$, transformed twice through $T()$. Here $E_\psi$ denotes the encoder model, and the comparable prototypes as $\textbf{C}$. $A_\omega$ indicates the introduced classification model implemented on top of the embedding output (diagram is designed via PlotNeuralNet \cite{haris_iqbal_2018_2526396}).}
\label{fig:framework}
\end{figure*}
\par
\textbf{Image clustering} techniques group the corresponding representation basis to the defined priorities. Two main strategies are considered from different perspectives. A first group of studies \cite{DBLP:conf/iccv/ChangWMXP17,Ji_2019_ICCV, Yang2016JointUL,zhao2020deep} apply a defined clustering loss function to identify patterns within the training samples. A method called deep adaptive clustering (DAC) \cite{8237888} implements a convolutional net, and measures the cosine similarities of the generated features. The model is gradually trained to assign the most similar features to the same group index. Ji \textit{et al.} propose invariant information clustering (IIC) \cite{Ji_2019_ICCV} which includes the MI as a training objective to maximize the dependency between the categorical outputs of numerous augmented views. A similar algorithm is reported in \cite{Hu2017}, which suggests a variation of MI as the training objective for learning discrete representations regularized through virtual adversarial training. Nevertheless, the performance of the aforementioned single-phase training methods deteriorate on more challenging datasets. A second group of most recent methods is based on multiple sequential training phases. State-of-the-art performance has reported in \cite{vangansbeke2020scan}, where the model is initialized in the first phase by applying a contrastive learning objective. In the second phase, the $k$ nearest features of each instance are measured and considered to belong to the same group, then the model is trained accordingly. Due to the mismatch prediction of the $k$ nearest features, the method implements a third round of training, where the model learns from the most confident predictions. Likewise, a two-stage clustering is incorporated in \cite{eccv_han}, where the encoder parameters are firstly initialized through a self-supervised strategy. This pre-training phase significantly improves the performance by a large margin of the later clustering objective in the second stage. However, having multiple training stages will no doubt increase the computation time and introduce dependency on previous phases. In addition, each training phase usually involves its specific hyper-parameters, transformation functions, optimizers as well as objective functions. By contrast, our proposed strategy is based on a single-phase training method, nevertheless, it demonstrates competitive performance comparable to the most recent multiple-phases strategies.
\section{The Proposed Method}
\label{sec:method}
Assume there is an unlabelled set of images denoted as $X = \{x_i\}_{i=1}^{n}$, which holds a relation with a finite set of classes $Y= \{y_i \in \mathbb{N}, 0 < y_i \leq k\}$, where $\textit{k}$ is a given hyper-parameter equal to the number of classes. The goal of this work is to instantiate a convolutional model and map the described relation as $f_{\theta}: X \rightarrow Y$, and $\theta$ indicates the model's parameters. Figure \ref{fig:framework} presents the overall framework of the proposed IMC-SwAV, consisting of three main components: 1) a ConvNet encoder ($E_\psi$) that projects the training instances onto a latent space ($Z$); 2) the trainable prototype vectors $\textbf{C}$ that are compared with the projected features to derive the computed distribution; and 3) an introduced classifier ($A_\omega$), which maps the generated features to the corresponding discrete representation.
In the following, we begin with a brief description of a representation learning strategy based on an online self-labelling assignment method and its application to the optimization of an encoder model. We then present a joint training objective, which is a modified form of mutual information, with the aim to maximise the mutual dependency of the computed distribution obtained by the prototypes and the classifier's predictions.
\subsection{Unsupervised Representation Learning}
The encoder model adopted in this work $E_\psi: X \rightarrow Z$, where $\psi$ denotes the encoder's trainable parameters, aims to learn the important semantic information of the given set of data without supervision, while ignoring less valuable semantics such as background or noise. Hence, motivated by the recent state-of-the-art achievement of SwAV \cite{caron2020unsupervised}, we employ the SwAV contrastive learning strategy to impose the consistency between representations ($Z$) obtained from augmented views of the same instances by comparing them to a set of prototype vectors $\textbf{C}$. These prototypes are evaluated only for uniformly mapping the obtained representations by exchanging their predicted probabilistic distributions, respectively. This is achieved by minimizing the \textit{swapped} self-labelling training objective \cite{caron2020unsupervised}:
\begin{equation}
\begin{split}
L_{\psi, C}^{\textit{swap}}(z^{(1)},z^{(2)}) &= \ell(z^{(1)}, q^{(2)}) + \ell(z^{(2)}, q^{(1)}) \\
\text{where \ \ \ } \ell(z^{(j)}, q^{(l)}) &= -\frac{1}{m} \sum_{i=1}^m q_i^{(j)}log(u_{i}^{(l)})
\label{eq:swav objective}
\end{split}
\end{equation}
where $z_{i}^{(t)}=E_{\psi}(T(x_{i}))$ denotes a latent representation derived by the encoder ($E_\psi$) of the $i\text{-th}$ training instance, respectively, $T \in \{f_1,f_2,..,f_n\}$ is a collection of stochastic transformation functions applied on the original instance beforehand to obtain a transformed view $x_i^{(t)}=T(x_i)$, where $t$ is the augmented view index. For example, representations obtained by two alternative views of the same instance ($x_i$) are indicated as $z_i^{(1)}$ and $z_i^{(2)}$ in Equation \ref{eq:swav objective}, as illustrated in Fig. \ref{fig:framework}. $q_i^{(1)}$ and $q_i^{(2)}$ are the swapped self-labelling target distributions of the two transformed views of the $i$-th training instance, and $u_i^{(1)}$ and $u_i^{(2)}$ are the computed probabilistic distributions, respectively, also called ``codes", determined by comparing the corresponding representations with the trainable prototype vectors $\textbf{C}$ as follows:
\begin{equation}
u^{(t)}_i = \frac{\text{exp}(\frac{1}{\tau}\text{sim}(z_{i}^{(t)},c_j))}{\sum_{j=1}^{k'}\text{exp}(\frac{1}{\tau}\text{sim}(z_{i}^{(t)},c_j))}
\label{eq:u definition}
\end{equation}
where $\tau$ is a temperature parameter that controls the smoothness of the probabilistic Softmax output, and $k'$ indicates the number of prototypes. Lastly, we define $\text{sim}(z_{i},c_j)=z_{i}^T c_j/\|z_{i}\|\|c_j\|$ as the cosine similarity distance of both $\ell_2$ normalized vectors, and $c_j$ represents the $j\text{-th}$ prototype vector. \par
To minimize the cross-entropy terms of the \textit{swapped} training objective (Equation \ref{eq:swav objective}), the computation of the exchanged pair distributions ($q^{(t)}_i$) for each augmented view is required. The self-labelling target distributions ($q^{(t)}_i$) prevent the assignment of trivial solutions, and enforce a uniform mapping to the prototypes. Ultimately, we accommodate the online computation as introduced in \cite{caron2020unsupervised} to derive the required distribution. Given a collection of feature vectors $Z=[z_1,z_2,...,z_m]$, with $m$ being the size of the training mini-batch, and the prototype vectors $C=[c_1,c_2,...,c_{k'}]$, we aim to define a target distribution $Q=[q_1,q_2,...,q_m]$ to maximize the correlation between the generated feature vectors and the trainable prototypes by satisfying the aforementioned conditions as:
\begin{equation}
\mymax\limits_{\bold{Q} \in \mathcal{Q}} \ \text{Tr}(\bold{Q^{T}C^{T}Z}) + \epsilon H(\bold{Q})
\label{eq:max objective}
\end{equation}
\begin{equation}
\mathcal{Q} = \{\bold{Q} \in \mathbb{R}_{+}^{k'\times m}| \bold{Q}1_M = \frac{1}{K'}1_{K'}, \bold{Q}^T1_{K'}=\frac{1}{M}1_M\}
\label{eq:q definition}
\end{equation}
where $\epsilon$ is a scaling parameter that regularizes the entropy term $H(\bold{Q})$ by controlling the mapping of the distribution. In practice, the required target distribution can be efficiently optimized by iteratively employing the Sinkhorn-Knopp \cite{NIPS2013_af21d0c9} algorithm. Hence, the constraints of a uniform mapping and homogeneity are encouraged within the training mini-batch. The reader is referred to \cite{caron2020unsupervised} and \cite{NIPS2013_af21d0c9,peyre2020computational}, respectively, for a more detailed description of the SwAV and Sinkhorn-Knopp algorithms.
\subsection{Joint Cluster Representations}
Thus far, the described encoder generates features without producing the desired annotation or discrete properties to the given $k$ classes. Instead, the training set is mapped in domain $U=\{U \in \mathbb{R}^{m \times k'}_+| \sum_{j=1}^{k'} u_{i,j}=1 \}$. Our goal is to define a parametric classifier that assigns the generated representations into a given number of classes $\textit{k}$ such as $A_\omega : Z \rightarrow Y $, where $\omega$ are parameters of classifier $A$.\par
\begin{figure}[t]
\centering
\includegraphics[width=0.4\textwidth]{figs/jsd_pairwise.png}
\caption{A scatter illustration of JSD pairwise distances between two probabilistic distributions $U^{(1)}$ and $U^{(2)}$ generated by a pre-trained encoder of the same 32 image instances, where different transformations applied in each instance.
}
\label{fig:scatter similarities}
\end{figure}
To this end, the encoder is optimized by assigning pairs of the same instances to their corresponding prototypes, while ensuring different instances, that are distinct for each prototype. We evaluate this argumentation by computing two distributions of a mini-batch denoted as $U^{(1)}$ and $U^{(2)}$ using a pre-trained encoder. Each distribution includes the same 32 instances from the training set, where each instance is alternated with the application of the transformation functions $T()$. We measure the pairwise similarity of the two probabilistic distributions by applying the Jensen–Shannon divergence (JSD) \cite{1365067}, a symmetric bounded distance measurement for evaluating two probability distributions. As presented in Fig. \ref{fig:scatter similarities}, the diagonal positions express the lower divergences with the mean value $\approx 0.06$ as these transformed views are from the same original instances. On the other hand, the larger divergences are observed in the instances of different indices (off diagonals) with the mean value $\approx 0.63$. Note that the upper bound of JSD is $log(2) \approx 0.69$ and zero the lower bound. \par
Complementary to the above observation, we assume that the transformed views of an instance produce similar distribution in domain $U$. We can also assume that the probabilistic output ($U$) in Eq. \ref{eq:u definition} and the corresponding classifier's outputs $Y = A_\omega(Z)$ holds a relationship since both are obtained from representations $Z$, hence defining a training objective that \textit{maximizes the mutual dependency} between these two distributions with respect to parameters $\omega$. Then classifier $A_\omega$ will map semantically similar instances in domain $U$ to the same cluster in domain $Y$. Note that the distributions of $Y$ and $U$ are a product of the Softmax output; hence, both belong to the
probability simplex. Inspired by \cite{Ji_2019_ICCV}, we propose a measurement describing the amount of dependency between the two probabilistic outputs, which is the direct application of mutual information (MI) \cite{Cover:1991:EIT:129837}:
\begin{equation}
\begin{split}
I(U;Y) &= - H(Y|U) + H(Y) \\
&= \sum_{(u,y)} p(u,y) log \left (\frac{p(u,y)}{p(u)} \right ) -\sum_{y} p(y) log (p(y))
\label{eq:mi}
\end{split}
\end{equation}
The MI expression is the relative entropy for measuring the divergence of the joint probability $p(u,y)$ and the product of two marginals. We modify Equation \ref{eq:mi} to define our clustering objective function with respect to parameters $\omega$ of the classifier. In this work, we employ a deep net as the classifier $A_{\omega}$ on top of the embedding layer ($Z$) to jointly perform the training. We invert the corresponding expression into a minimization loss function, and optimizing the model using the stochastic gradient descent:
\begin{equation}
\begin{split}
L^{\textit{cluster}}_\omega (Y,U) &= \frac{1}{V^2}\sum_{i}^V \sum_{j}^V \left [ H(P_{Y^{j}}|P_{U^i}) - \beta H(P_{Y^{j}}) \right] \\
&= \frac{1}{V^2}\sum_{i}^V \sum_{j}^V \left[ \sum_{Y^j, U^i} P_{Y^j U^i } log \left( \frac{P_{U^i}}{P_{Y^{j} U^{i}}} \right) \right . \\ &+\left . \beta \sum_{Y} P_{Y^j} log(P_{Y^j})\right] ~.
\label{eq:mi v2}
\end{split}
\end{equation}
\begin{lstinputlisting}[language=Python,label={lst: psedocode},float=bp, caption=Pseudo-code written in Python presents the clustering objective based on Pytorch library \cite{NEURIPS2019_9015}.]{sup_includes/psedocode.py}
\end{lstinputlisting}
where $P_{Y^j U^i} = \{\frac{1}{m} {Y^j}^{\!\!\top} U^i \in \mathbb{R}_+^{k \times k'}| Y^j \in \mathbb{R}_+^{m \times k}, U^i \in \mathbb{R}_+^{m \times k'} \}$, $U^i$ denotes the computed codes, $Y_j$ the classifier's prediction, $P_{U^i}= \{\frac{1}{m} \sum^m_{l=1} u^i_{{l,k'}} \in \mathbb{R}^{k'}_{+}\}$ and $P_{Y^j}= \{\frac{1}{m} \sum^m_{l=1} y^j_{{l,k}} \in \mathbb{R}^{k}_{+}\}$ the two marginal terms. Here, $m$ indicates the number of elements in the mini-batch and indices $i$ and $j$ denote the corresponding transformed views of the mini-batch, as illustrated in Fig. \ref{fig:framework}, where $V$ indicates the total number of transformed views obtained from instances in the mini-batch. The scalar $\beta$ denotes an introduced weight parameter of the entropy term $H(P_{Y^j})$, that encourages the classifier to uniformly assign the predicted class indices and thus prevent degeneracy. The conditional entropy term is minimized by increasing the classifier's prediction confidence. In Listing \ref{lst: psedocode}, we provide the pseudo-code of the clustering objective function as described in Eq. \ref{eq:mi v2}, which returns a quantitative measurement of the dependency between the classifier's predictions ($Y^j$) and the distribution ($U^i$). The pseudo-code is written in Python and includes commands extended in the Pytorch library \cite{NEURIPS2019_9015}. The final training objective with respect to all parameters of the overall framework ${\{\psi,\omega,\textbf{C}\}} \in \theta$ is defined as follows:
\begin{equation}
\begin{split}
\textit{L}_\theta = \textit{L}_{\psi,C}^{\textit{swap}} + L^{\textit{cluster}}_\omega ~.
\label{eq:full objective}
\end{split}
\end{equation}
The internal computation of the IMC-SwAV is presented in Algorithm \ref{alg:framework} and the full training process, including all steps for the two transformed views, are given in Algorithm 2.
\SetKwInOut{Initialize}{Initialize}
\SetKwInOut{Generate}{Generate}
\begin{algorithm}[t]
\SetKwFunction{FMain}{Framework}
\SetKwProg{Fn}{Function}{:}{}
\Fn{\FMain{$B$}}{
\hspace{0.1mm} $B^{t} = T(B)$ apply transformation functions\\
\hspace{0.1mm} $Z^{t}=E_\psi(B^{t})$ generate the representations\\
\hspace{0.1mm} $U^{t}= \frac{\text{exp}(\frac{1}{\tau}\text{sim}(z_{i}^{t},c_j))}{\sum_{j=1}^{k'}\text{exp}(\frac{1}{\tau}\text{sim}(z_{i}^{t},c_j))}$ compute the distribution \\
\hspace{0.1mm} $Y^{t} = A_\omega(Z^{t})$ classifier's output \\
\hspace{0.1mm} $Q^{t} = $ derived by $U^{t}$ through Shinkhorn-Knopp \cite{NIPS2013_af21d0c9}\\
\Return $U^{t}$ , $Q^{t}$ , $Y^{t}$
}
\textbf{End Function}
\caption{Framework method}
\label{alg:framework}
\end{algorithm}
\begin{algorithm}[t]
\KwData{$X=\{x_{i}\}_{i=1}^{n}$}
\Initialize{$\{\psi, \omega,\textbf{C} \} \in \theta$}
\While{Convergence condition not satisfied}{
\hspace{0.7mm}{ \textbullet \hspace{0.1mm} Sample minibatch $B=\{x_{1},...,x_{m}\}$} \\
\hspace{1.mm}{ \Comment{First Views} }\\
\hspace{0.7mm}{ \textbullet \hspace{0.1mm} $U^{(1)}$ , $Q^{(1)}$ , $Y^{(1)} =$ \FMain{$B$} }\\
\hspace{1.mm}{ \Comment{Second Views} }\\
\hspace{0.7mm}{ \textbullet \hspace{0.1mm} $U^{(2)}$ , $Q^{(2)}$ , $Y^{(2)} =$ \FMain{$B$}} \\ \\
\hspace{1.mm}{ \Comment{Swap loss} }
$$
\textit{L}_{\psi, C}^{\textit{swap}} = \frac{1}{m} \sum_{i=1}^m \left ( -q_i^{(2)}log(u_{i}^{(1)}) - q_i^{(1)}log(u_{i}^{(2)}) \right )
$$ \\
\hspace{1.mm}{ \Comment{Clustering loss} }\\
$$L^{\textit{cluster}}_\omega (U,Y) = \frac{1}{2^2} \sum_{i=1}^{2} \sum_{j=1}^{2} \left [ H(Y^j|U^i) - \beta H(U^i) \right] $$
\textbullet \hspace{0.1mm} Jointly update the framework's parameters to minimize the combined training objective $\textit{L}_\theta$: \\
\begin{align*}
\textit{L}_\theta = \textit{L}_{\psi,C}^{\textit{swap}} + L^{\textit{cluster}}_\omega
\end{align*}
}
\caption{The proposed joint training algorithm.}
\label{alg:Process}
\end{algorithm}
In the following, we briefly discuss the reason for using mutual information in the objective function. Recall that the convergence of the SwAV \cite{caron2020unsupervised} framework is achieved by gradually dividing the full dataset into $k'$ partitions, where $k'$ is the number of defined prototypes, which is usually large. Thus, on average $n/k'$ instances are assigned to each partition. By maximizing the dependency between the distribution $U=\{U \in \mathbb{R}^{m \times k'}_+| \sum_{j=1}^{k'} u_{i,j}=1 \}$ and the classifier's prediction $Y$, the classifier converges by identifying partitions that hold similar semantics in $U$ and grouping them to the same cluster in $Y$. The marginal entropy in Equation \ref{eq:mi v2} regularizes the classifier to ensure that each cluster is approximately assigned $n/k$ elements. On the other hand, the classifier increases its confidence on its predictions by minimizing the conditional entropy. Despite its simplicity, the experimental results presented in Section \ref{sec:results} indicate that the proposed clustering objective can effectively leverage the derived distribution $U$ to achieve high performance. \par
\subsection{Multi-Crop Strategy}
In this work, we adopt the same multi-crop views training strategy as introduced in \cite{caron2020unsupervised} for a full exploitation of the proposed framework. Each training instance is transformed into a set of augmented views. The first two main transformed views are cropped in a negligible lower resolution of the original image, we later called this transformation as high resolution views. The additional mini-cropped views cover only a small part of the image with their resolution to be approximately the half of the original image size to reduce the convolutional operations and thus the time complexity. We named this transformation as low resolution views. During the training, $Q$ target distributions are computed only for the high resolution views, and used across all augmented views. We extend this process also to the proposed clustering objective (Eq. \ref{eq:mi v2}). Here, we compute pairwise the MI quantity based on all views; hence, if a mini-batch is transformed into two high and two low resolution views, so $V=4$, resulting in 16 combinations in total. It is found that the multi-crop strategy, as demonstrated in our empirical studies, effectively enhances the model's prediction capability.
\section{Experimental Studies}
\label{sec:results}
We evaluate the proposed clustering method, IMC-SwAV, on four challenging colour image datasets: 1) CIFAR-10, which contains 10 equally populated classes; 2) CIFAR-100 containing 100 classes with all elements being uniform distributed. These 100 classes are also grouped into 20 super-classes with each super class consisting of five classes. For convenience, we refer to the 20 super-classes as CIFAR-100/20; 3) STL10, which contains labels only for the 13000 images and the remaining unlabeled images are from various classes. The encoder is trained across the labeled and unlabeled instances, where the classifier is trained and evaluated only for the labeled instances; and 4) Tiny-ImageNet/200, which is a subset of ImageNet containing 200 classes downsampled to a lower resolution. Table \ref{Dataset description} presents the details of each set, including the number of training and validation elements, the number of clusters, the image resolution, and the multi-crop ranges during the training. By ``2x28+4x18", we mean two high resolution views of crop size $28$, and four low resolution views of crop size $18$.
The proposed algorithm is compared with a range of state-of-the-art visual clustering approaches in terms of unsupervised learning metrics. We further demonstrate the effectiveness of the proposed single-phase method on datasets with up to 200 classes. Ablation studies are also performed to evaluate the effects of different hyper-parameters to examine the role of the most important components of the proposed algorithm and the sensitivity to the key parameters.
\subsection{Metrics for Unsupervised Learning}
Three quantitative metrics for supervised learning are adopted to measure the performance of IMC-SwAV: 1) Accuracy (ACC); 2) Normalized mutual information (NMI) \cite{vinh2009information}; and 3) Adjusted rand index (ARI). The ACC reports the accuracy by finding the best mapping between the predicted labels and the ground truth labels. The mapping matrix is found using the Hungarian algorithm \cite{doi:10.1002/nav.3800020109}. NMI measures the mutual information between the two distributions (the model's prediction and the ground truth), which is scaled between zero and one. ARI measures the similarity between the two particular distributions by comparing all possible pairs and measuring those assigned in the same cluster and those in different ones.
\begin{table}[!]
\caption{Descriptions of the Datasets}
\resizebox{\columnwidth}{!}{
\begin{tabular}{lcccccccccccc}
\hline
\textbf{Name} & \begin{tabular}{@{}c@{}}\textbf{Train.} \\ \textbf{No.}\end{tabular} & \begin{tabular}{@{}c@{}}\textbf{Val.} \\ \textbf{No.}\end{tabular} & \begin{tabular}{@{}c@{}}\textbf{Classes} \\ \textbf{($\textit{k}$)}\end{tabular}& \textbf{Res.} & \begin{tabular}{@{}c@{}}\textbf{Multi-}\\ \textbf{\textbf{crops}}\end{tabular} \\ \hline
\textbf{CIFAR-10} & 50000 & 10000 & 10 & 32x32&
\begin{tabular}{@{}c@{}} 2x28+\\4x18 \end{tabular} \\
\begin{tabular}{@{}l@{}}\textbf{CIFAR-100}\\ \textbf{\textbf{CIFAR-100/20}}\end{tabular}& 50000& 10000 & \begin{tabular}{@{}c@{}} 100\\20\end{tabular} & 32x32 & \begin{tabular}{@{}c@{}} 2x28+\\4x18 \end{tabular} \\
\textbf{STL10} & 105000 & 8000 & 10 &
96x96 & \begin{tabular}{@{}c@{}} 2x76+\\4x52 \end{tabular} \\
\textbf{Tiny-ImageNet} & 100000 & 10000 & 200 &
64x64 & \begin{tabular}{@{}c@{}} 2x56+\\4x36 \end{tabular} \\\hline
\end{tabular}
}
\label{Dataset description}
\end{table}
\begin{table*}[t]
\caption{Comparative results on the three benchmarks. The top three methods in terms of the best results highlighted. Note that our method is evaluated for $15$ independent runs across all datasets, and the average and best results are reported.}
\footnotesize
\begin{center}
\setlength\tabcolsep{1.5pt}
\begin{tabularx}{\textwidth}{l| Y Y Y|Y Y Y|Y Y Y|Y Y Y}
\hline
\multirow{2}{*}{\textbf{Method/Dataset}}& \multicolumn{3}{c}{\textbf{CIFAR-10}} & \multicolumn{3}{c}{\textbf{CIFAR-100/20}} & \multicolumn{3}{c}{\textbf{STL10}} & \multicolumn{3}{c}{\textbf{Tiny-ImageNet/200}}\\
\cline{2-13}
& ACC & NMI & ARI & ACC & NMI & ARI & ACC & NMI & ARI & ACC & NMI & ARI \\ \hline
K-means & 22.9 & 8.7 & 4.9 & 13.0 & 8.4 & 2.8 & 19.2 & 12.5 & 6.1 & 2.5 & 6.5 & 0.5 \\
AE \cite{NIPS2006_3048} & 31.4 & 23.9 & 16.9 & 16.5 &10.0 & 4.8 & 30.3&25.0 &16.1 & 4.1 & 13.1 & 0.7 \\
VAE \cite{Kingma2014} & 29.1 & 24.5 & 16.7 & 15.2 & 10.8 & 4.0 & 28.2 & 20.0 & 14.6 & 3.6 & 11.3 & 0.6\\
DCGAN \cite{DBLP:journals/corr/RadfordMC15} & 31.5 & 26.5 & 17.6 & 15.1& 12.0&4.5 &29.8 &21.0 &13.9 & 4.1 & 13.5 & 0.7\\
DEC \cite{pmlr-v48-xieb16} & 30.1 & 25.7 & 16.1 & 18.5& 13.6 & 5.0 & 35.9 & 27.6 & 18.6 & - & - & -\\
JULE \cite{Yang2016JointUL} & 27.2 & 19.2 & 13.8 & 13.7 & 10.3 & 3.3 & 27.7 & 18.2 & 16.4 & 3.3 & 10.2 & 0.6\\
ADC \cite{DBLP:conf/dagm/HausserPGAC18} & 32.5 & - & - & 16.0 & - & - & 53.0 & - & - & - & - & -\\
DAC \cite{DBLP:conf/iccv/ChangWMXP17} & 52.2 & 39.6 & 30.6 & 23.8 & 18.5 & 8.8 & 47.0 & 36.6 & 25.7 & 6.6 & 19.0 & 1.7 \\
IMSAT-DCGAN \cite{ntelemis2020image} & 70.0 & - & - & 32.4 & - & - & 58.7 & - & - & - & - & -\\
DDC \cite{2019arXiv190501681C} & 52.4 & 42.4 & 32.9 & - & - & - & 48.9 & 37.1 & 26.7 & - & - & -\\
DCCM \cite{DCCM} & 62.3 & 49.6 & 40.8 & 32.7 & 28.5 & 17.3 & 48.2 & 37.6 & 26.2 & \textbf{10.8} & 22.4 & 3.8\\
IIC \cite{Ji_2019_ICCV} & 61.7 & 51.1 & 41.1 & 25.7 & 22.5 & 11.7 & 59.6 & 49.6 & 39.7 & - & - & -\\
DCCS \cite{zhao2020deep} & 65.6 & 56.9 & 46.9 & - & - & - & 53.6 & 49.0 & 36.2 & - & - & -\\
PICA \cite{huang2020pica} & 69.6 & 59.1 & 51.2 & 33.7 & 31.0 & 17.1 & 71.3 & 61.1 & 53.1 & 9.8 & \textbf{27.7} & \textbf{4.0} \\
DRC \cite{zhong2020deep} & 72.7 & 62.1 & 54.7 & 36.7 & 35.6 & 20.8 & 74.7 & 64.4 & 56.9 & - & - & -\\
EmbedUL \cite{DBLP:conf/eccv/HanPPKC20} & \textbf{81.0} & - & - & 35.3 & - & - & 66.5 & - & - & - & - & -\\
CC \cite{li2020contrastive} & 79.0 & \textbf{70.5} & \textbf{63.7} & \textbf{42.9} & \textbf{43.1} & \textbf{26.6} & \textbf{85.0} & \textbf{76.4} & \textbf{72.6} & \textbf{14.0} & \textbf{34.0} & \textbf{7.1}\\
SCAN \cite{vangansbeke2020scan} (Best) & \textbf{88.3} & \textbf{79.7} & \textbf{77.2} & \textbf{50.7 }& \textbf{48.6} & \textbf{33.3} & \textbf{80.9} & \textbf{69.8} & \textbf{64.6} & - & - & -\\
\hline
Supervised & 92.8 & - & - & 76.1 & - & - & 89.2 & - & - & 45.4 & - & -\\
SwAV\cite{caron2020unsupervised} + K-means & 78.4 & 67.5 & 61.3 & 40.1 & 47.0 & 10.6 & 74.9 & 70.5 & 54.0 & 18.2 & 46.8 & 5.2 \\
IMC-SwAV (Avg$\pm$) & \textbf{89.1}$\pm$0.5 & \textbf{81.1}$\pm$0.7 & \textbf{79.0}$\pm$1.0 & \textbf{49.0}$\pm$1.8 & \textbf{50.3}$\pm$1.2 & \textbf{33.7}$\pm$1.3 & \textbf{83.1}$\pm$1.0 & \textbf{72.9}$\pm$0.9 & \textbf{68.5}$\pm$1.4 & \textbf{27.9}$\pm$0.3 & \textbf{48.5}$\pm$2.0 & \textbf{14.3}$\pm$2.1 \\
IMC-SwAV (Best) & \textbf{89.7} & \textbf{81.8} & \textbf{80.0} & \textbf{51.9} & \textbf{52.7} & \textbf{36.1} & \textbf{85.3} & \textbf{74.7} & \textbf{71.6} & \textbf{28.2} & \textbf{52.6} & \textbf{14.6} \\
\hline
\end{tabularx}
\label{tab:Unsupervised Clustering}
\end{center}
\end{table*}
\subsection{Experimental Settings}
\subsubsection{Architecture Implementation}
For fair comparisons, we adopt the same settings across all datasets as those given in \cite{caron2020unsupervised,pmlr-v119-chen20j}. Specifically, the proposed framework implements an encoder based on ResNet18 \cite{he2015deep} architecture. The classifier employs a multilayer perceptron net containing two hidden layers on top of the embedding layer. Similar to \cite{caron2020unsupervised,pmlr-v119-chen20j}, we use a projection head to reduce the dimension of the embedding layer prior to the comparison with the prototypes. For the overall framework, the Adam optimizer \cite{DBLP:journals/corr/KingmaB14} is adopted for training. The learning rate is set to $5 \times 10^{-4}$ with a warmup schedule in the first $500$ training iterations. We use a decay learning rate of $0.4$ in epochs $[150, 300, 400]$ and a total of $500$ epochs is run. The mini-batch size is set to $256$ across all datasets except Tiny-ImageNet/200, for which the batch size was set $512$ due to the large number of classes in this dataset. A $\ell_2$ weight decay regularizer is used with a rate of $1\times10^{-5}$. We apply an additional two-sided regularizer, in a similar form to that reported in \cite{NEURIPS2019_ddf35421,NEURIPS2020_f3ada80d}, to the logit outputs of the classifier's representation. Specifically, we penalize any absolute value higher than five prior to the Softmax activation function, hence preventing the classifier from making predictions with a high confidence level in the early stage of the training phase as:
\begin{equation}
L^{\textit{cluster}}_\omega (U,Y) + \alpha \cdot \frac{1}{m} \sum_i^m \sum_j^k \text{max}(|x_{i,j}|-5,0)
\end{equation}
where $x$ is the logit outputs of the classifier, $\alpha$ is a weight parameter set to $1\times10^{-2}$ across all experiments, $m$ is the batch size, and $k$ is the number of clusters.
\subsubsection{Hyper-Parameters Selection}
The settings below are applied across all experiments. The temperature scalar $\tau$ in Softmax smoothness is set to $0.1$ as recommended in \cite{caron2020unsupervised}. Similarly, to prevent degeneracy, we set the hyperparameter $\epsilon$ of the weighted entropy of Sinkhorn-Knopp to $0.05$ and optimize it for three iterations. The implemented prototype vectors are set to $k'= 1000$ (except in the ablation studies). In the clustering objective, we set the weight factor of the marginal entropy $\beta=4$ for all experiments.
\subsubsection{Transformation Functions}
In this work, we follow the same transformation/augmentation scheme as presented in \cite{pmlr-v119-chen20j} across all experiments.
Each instance is horizontally flipped at a probability of $0.5$. All images are modified with color jittering at a rate: brightness $0.4$, contrast $0.4$, saturation $0.4$, and hue $0.2$. Color jitter is applied at a probability of $0.8$. A probability of $0.25$ is also used for a gray-scale instance transformation. The re-sizing rates are set to the range between $0.2$ and $1.0$, between $0.08$ and $0.4$, respectively, for the two main views (of high resolution) and four smaller views (of low resolution). The aspect ratio in both cases is set to $(3/4, 4/3)$. The corresponding crop sizes are given in Table \ref{Dataset description}.
\subsection{Comparative Results}
Table \ref{tab:Unsupervised Clustering} presents the results over $15$ independent runs. To demonstrate the stability of the proposed IMC-SwAV, in each run the model's parameters are randomly initialized. The same process is followed across all datasets. We present the mean, standard deviation (STD), and the best result. In contrast to the previous clustering studies in which all instances are used for training and validation, we use a similar validation process as that in \cite{10.1007/978-3-030-58607-2_16}, where the validation results are derived from the validation data that has not been seen by the model. During the training, all components of the framework are jointly trained with training instances only. We list the results obtained by a broad range of state-of-the-art unsupervised learning algorithms for comparison. It should be stressed that the ground truth is used only for computing the relevant metrics. To further demonstrate the effectiveness of our approach, we freeze the encoder parameters and train a single layer net on top of the embedding layer in a supervised manner. We present these supervised learning result to further demonstrate the capability of the proposed unsupervised clustering approach. Finally, we compare the clustering results of K-means by using the same features extracted by the SwAV method. We also report the best results of "SwAV + K-means'' over 20 initializations of K-means' centroids. \par
As presented in Table \ref{tab:Unsupervised Clustering}, the average performance of IMC-SwAV outperforms all state-of-the-art methods on CIFAR-10, CIFAR-100/20, and Tiny-ImageNet and produces competitive results on STL10, while the best results of our method are the best among all algorithms under comparison. Note that only the best results of all algorithms under comparison are reported in Table \ref{tab:Unsupervised Clustering}. IMC-SwAV performs clustering based on the online mode in the \textit{swap} training strategy without the requirement of a pre-trained model or the implementation of multi-phase strategies, thus providing an additional advantage in terms of simplicity and training time. We observe a large margin between the clustering performance of IMC-SwAV and SwAV + K-means. Since both methods are trained on the same features, this performance improvement can be attributed to the defined clustering objective and the jointly training proposed in this work. Furthermore, the increased number of parameters of the classifier has a low computational impact during the training, with the training time on a Nvidia Quadro RTX6000 GPU being increased by $1.08\text{x}$ only.
\begin{figure}[t]
\centering
\includegraphics[width=0.4 \textwidth]{figs/confusion_matrix.png}
\caption{The above confusion matrix showing the predictions and ground truth made by the proposed model on CIFAR-10 validation set.
}
\label{fig:confusion matrix}
\end{figure}
In the following, we further demonstrate our model performance by evaluating the individual classes and instances.
\subsubsection{Individual class performance}
To examine the performance of individual class accuracies on CIFAR-10, a confusion matrix with the predictions made by IMC-SwAV on the unseen validation set is presented in Fig. \ref{fig:confusion matrix}. Note that in contrast to the previous methods, none of the implemented components are trained on the validation set. Each class consists of $1000$ instances. In the figure, $x$-axis and $y$-axis indicates the model's predictions and the ground truth, respectively. From these results, it can be seen that IMC-SwAV shows a high accuracy on the majority of the classes. Mismatch inaccurate predictions are mainly observed between classes `Cat' and `Deer'.
\begin{figure}[t]
\centering
\includegraphics[width=0.45 \textwidth]{figs/gradcam.png}
\caption{This illustration is an interpretation of visual activation heatmap of accurate (positive - green frame) and inaccurate (negative - red frame) prediction made by our model on STL10.
}
\label{fig:gradcam}
\end{figure}
\subsubsection{Visual interpretation}
We visualize the predictions made by IMC-SwAV on STL10 dataset in Fig. \ref{fig:gradcam}. The predictions are visualized via Grad-Cam \cite{8237336}. IMC-SwAV is able to find specific visual elements to achieve the successful predictions such as airplane wings or horse's body shape. On the other hand, from the visual heatmap layers of the negative predictions, we see that the model fails to recognize specific object elements.
\begin{figure}[t]
\centering
\includegraphics[width=0.45 \textwidth]{figs/sample_size.png}
\caption{Performance of IMC-SwAV made on CIFAR-10 for a limited number of training elements. x-axis presents the number of training samples (by thousand) per class.
}
\label{fig:sample size}
\end{figure}
\begin{table}[t]
\caption{CIFAR100 - Evaluation on 100 Classes}
\resizebox{\columnwidth}{!}{
\begin{tabular}{lcccc}
\hline
\textbf{} & \textbf{Top-1 ACC} & \textbf{Top-5 ACC} & \textbf{NMI} & \textbf{ARI} \\ \hline
IMC-SwAV & 45.1 & 67.5 & 60.8 & 30.7 \\
SwAV + K-means & 30.2& 48.6 & 56.25 & 12.3 \\
\hline
\end{tabular}
}
\label{tab: cifar100}
\end{table}
\subsubsection{Influence of sample size}
We vary the number of the training instances per class by an increasing interval of $500$ per experiment within the range of $[1000,5000]$, where 5000 is the maximum number of samples in each CIFAR-10 class (recall that CIFAR-10's classes are equally spread with each class containing $5000$ element). The model is trained on the selected subset only, and the performance of validation set is illustrated in Fig. \ref{fig:sample size}.
As expected, the model's precision is gradually decreasing when fewer training samples are considered. Notably, IMC-SwAV's performance remains stable when the sample size is larger than $2500$.
\subsubsection{Effectiveness to a large number of clusters}
To further validate our method on a dataset with a large number of clusters, we conduct an experiment on the 100 sub-class of CIFAR-100 to evaluate the clustering performance of IMC-SwAV. The features extracted by IMC-SwAV are directly compared with those obtained by the K-means algorithm. The metrics of Top-1 and Top-5 accuracy, as well as the rest unsupervised metrics are listed in Table \ref{tab: cifar100}.
These results demonstrate that IMC-SwAV performs well on this highly demanding clustering task. The proposed algorithm achieves a large margin of $15\%$ on Top-1 ACC and $19\%$ on Top-5 ACC.
\subsection{Ablation Studies}
In this subsection, we experiment with a variety of hyper-parameters to evaluate the effectiveness of the proposed IMC-SwAV. All parameters remain unchanged, as per the main settings, except for those to be studied. \par
\begin{figure}[t]
\centering
\includegraphics[width=0.45 \textwidth]{figs/prototypes.png}
\caption{These diagrams plot the performance metrics on CIFAR-10 over different numbers of prototypes. From top to low: ACC, NMI and ARI. All figures report the number of prototypes in x-axis.
}
\label{fig:prototypes illustration}
\end{figure}
\begin{table}[t]
\caption{IMC-SwAV with different ResNet architectures indicated by the method suffix}
\setlength\tabcolsep{3.5pt}
\begin{tabularx}{250.pt}{l|YYY|YYY|YYY}
\hline
\multirow{2}{*}{\begin{tabular}{@{}c@{}}\end{tabular}}& \multicolumn{3}{c}{\textbf{CIFAR-10}} & \multicolumn{3}{c}{\textbf{CIFAR-100/20}} & \multicolumn{3}{c}{\textbf{STL10}} \\
\cline{2-10}
& ACC & NMI & ARI & ACC & NMI & ARI & ACC & NMI & ARI \\ \hline
IMC-SwAV-18 (Avg) & 89.1 & 81.1 & 79.0 & 49.0 & 50.3 & 33.7 & 83.1 & 72.9 & 68.5 \\
IMC-SwAV-18 (Best) & 89.7 & 81.8 & 80.0 & 51.9 & 52.7 & 36.1 & 85.3 & 74.7 & 71.6 \\
IMC-SwAV-34 (Avg) & 89.5 & 81.7 & 79.9 & 50.2 & 51.2 & 34.6 & 84.5 & 74.8 & 70.9 \\
IMC-SwAV-34 (Best) & 90.2 & 82.4 & 80.9 & 52.1 & 53.2 & 36.3 & 86.0 & 76.7 & 73.3 \\
IMC-SwAV-50 (Avg) & 91.0 & 83.8 & 82.6 & 51.2 & 52.6 & 35.6 & 86.3 & 77.3 & 73.8 \\
IMC-SwAV-50 (Best) & 91.4 & 84.1 & 82.9 & 52.7 & 54.0 & 37.0 & 87.1 & 77.9 & 75.0 \\
\hline
\end{tabularx}
\label{tab: effects resnet}
\end{table}
\subsubsection{Prototypes}
We examine the performance of IMC-SwAV by varying the number of prototypes, $k'$, in the range of $[100,2000]$ while keeping all other parameters unchanged. Figure \ref{fig:prototypes illustration} shows the results from five independent runs of the unsupervised learning task. From these results, we conclude that IMC-SwAV shows robustness over different numbers of prototypes with the lowest mean value above $86.5\%$. We also note that for $k'$ larger than $300$, all three performance metrics become less sensitive to $k'$. Additionally, we note that our choice to set $k'$ to $1000$ prototypes in the main settings is not the optimum in terms of all metrics on CIFAR-10. However, in unsupervised learning, the ground truth is unknown and hence it is more realistic to compare the performance without using the optimal parameter setting.
\subsubsection{Experiment with different encoder architectures}
In this subsection, we examine the performance of the proposed IMC-SwAV by varying the encoder's architecture by using different types of the ResNet \cite{he2015deep} models. Specifically, we compare the performance of the proposed method when ResNet34 (21.3M parameters) and ResNet50 (23.5M parameters), respectively, is adopted to replace ResNet18 (11.1M parameters) used in the main experiments. All models are evaluated over five independent runs (the results of ResNet18 are averaged over 15 runs as in the main experiment) and the average and best recorded performances are reported for each model. The reported metrics are on the unseen test subset, and all frameworks are optimised with training subset only. The results are listed in Table \ref{tab: effects resnet}. As expected, larger models achieve slightly better performance in terms of all metrics. Note also that the discrepancies between different metrics of each model are minor, with the highest differences being observed difference on CIFAR-10.
\subsubsection{Multiple Crops}
We demonstrate the impact of the multiple crops strategy on the clustering performance on CIFAR-10. We change the number of low resolution cropped views with all other settings unchanged. First, a basic setting with two main views of the original image dimensions is evaluated. Then, we add two, four, and six additional low resolution views, respectively. The size of the two main views is decreased to $28\text{x}28$ to reduce the computation time. The resizing ratio remains unchanged: $[ 0.2,1.0]$ for the two main views, and $[0.08,0.4]$ for the low resolution views. We report the mean and STD of the metrics over five independent experiments. Additionally, the accuracy of a single layer net, which is trained independently on generated features through a supervised mode, is also included. \par
\begin{table}[t]
\caption{Effectiveness of Multiple Crops Strategies}
\resizebox{\columnwidth}{!}{
\begin{tabular}{l|ccccc}
\hline
\textbf{Crops} & \textbf{ACC} & \textbf{NMI} & \textbf{ARI} & \textbf{Sup.}& \begin{tabular}{@{}c@{}}\textbf{Train} \\ \textbf{Time}\end{tabular} \\ \hline
2x32 & 78.7$\pm$1.0 & 66.2$\pm$1.1 & 61.2$\pm$1.3 & 86.1 & $1\text{x}$ \\
2x28+2x18 & 87.2$\pm$1.1 & 78.6$\pm$1.4 & 75.8$\pm$2.0 & 92.1 & $1.1\text{x}$ \\
2x28+4x18 & 89.1$\pm$0.5 & 81.1$\pm$0.7 & 79.0$\pm$1.0 & 92.8 & $1.7\text{x}$ \\
2x28+6x18 & 89.4$\pm$0.6 & 81.2$\pm$0.9 & 79.0$\pm$1.8 & 93.2 & $2.4\text{x}$ \\
\hline
\end{tabular}
}
\label{tab: effects multiple}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=0.45 \textwidth]{figs/training_epochs.png}
\caption{Illustration presents the accuracy per epoch of each multi-crop implementation of Table \ref{tab: effects multiple}.
}
\label{fig:training epochs}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.45 \textwidth]{figs/view_sizes.png}
\caption{Presentation with various combinations of high and low crop sizes. $x\text{-axis}$ represent different low resolution crop sizes and high resolution are associated with different colours and symbols.
}
\label{fig:views resolution}
\end{figure}
Table \ref{tab: effects multiple} presents the results of the multiple crop experiments. From results, we notice that the additional low resolution views increase the performance in terms of all metrics, for both the unsupervised and supervised models. The last column presents the ratio of training time based on a basic training strategy of $2\text{x}32$. The training accuracy over the iterations are presented in Fig. \ref{fig:training epochs} for each experiment. We note that the mini-crop strategy of ``$2\text{x}28\text{+}4\text{x}18$" already reaches satisfactory results ($87.7$ ACC) within $250$ iterations, in comparison to the full experiment of a simple setup. Recall, the two-crop strategy holds an advantage by using the original resolution of $2\text{x}32$ instead of $2\text{x}28$.\par
\subsubsection{Combinations of crop sizes}
To further examine the impact of multi-crop's strategy, we vary the resolution of crop sizes of main views (high) and smaller views (low) within the intervals of $[20,32]$ and $[12,22]$, respectively. All experiments are performed on CIFAR-10 based on the main setting. We also preserve the same crop ratios and the number of views (two main views and four smaller views). Since, the resizing ratio of the smaller view is set to $0.4$, the difference between high and low resolutions does not exceed the absolute difference of the maximum $14$ pixels and minimum four pixels, to keep proportional to the crop size of main view. Figure \ref{fig:views resolution} presents our findings in terms of ACC. The performance of IMC-SwAV slightly degrades in a very low resolution in combination with ``$2\text{x}20\text{+}4\text{x}12$", probably because the model fails to capture sufficient semantic details. On the other hand, all combinations above the crop sizes of $\text{(high view)} \geq 26$ and $\text{(low view)}\geq18$ exhibit a satisfactory accuracy above $86\%$.
\subsubsection{Other representation learning strategies}
In this section, we aim to demonstrate that our clustering strategy is not restricted to the SwAV \cite{caron2020unsupervised} framework and is also applicable to other self-supervised learning strategies. Here, we evaluate our single-phase clustering method that adopts SeLa \cite{asano2020self}, another grouping based self-supervised learning method, to replace SwAV. Similar to IMC-SwAV, the components of the encoder and the classifier are jointly trained in a single-phase mode. We compare the performance of the variant of our clustering method using Sela (called IMC-SeLa) with the K-means algorithm, and a supervised classifier net trained on same extracted features. The parameters of the encoder model are optimized with the self-supervised loss function of the adopted strategy only without using any additional loss functions or pre-trained models. Similar to the main experiments, the framework is optimized on training subset and the performance evaluation is made on the validation subset. Each method is performed for five independent runs and the best result is reported. For a fair comparison, the architecture is similar to the main experiment with the encoder being based on a ResNet18 network. \par
The comparative results are listed in Table \ref{tab: sela}, from which a performance degradation of all methods under comparison can be observed, implying that the extracted features are less separable. This can be easily validated by examining the results of the supervised linear classifier net (Supervised in the Tables) trained on the extracted features (for SwAV \cite{caron2020unsupervised} in Table \ref{tab:Unsupervised Clustering} and for SeLa \cite{asano2020self} in Table \ref{tab: sela}). This can also be confirmed by the K-means clustering approach on the same features extracted by the two self-supervised techniques.
These results indicate that the final performance of the proposed framework also depends on the performance of the self-supervised learning algorithm it adopts. Nevertheless, IMC-SeLa still outperforms "SeLA + K-means'' at a large margin on CIFAR-10, CIFAR-100/20 and Tiny-ImageNet/200.
\begin{table}[t]
\caption{Clustering strategy adopted to SeLA training}
\setlength\tabcolsep{3.5pt}
\begin{tabularx}{250.pt}{l|YYY|YYY|YYY}
\hline
\multirow{2}{*}{\begin{tabular}{@{}c@{}}\textbf{Method/}\\ \textbf{Dataset.}\end{tabular}}& \multicolumn{3}{c}{\textbf{CIFAR-10}} & \multicolumn{3}{c}{\textbf{CIFAR-100/20}} & \multicolumn{3}{c}{\textbf{Tiny-ImgNet200}} \\
\cline{2-10}
& ACC & NMI & ARI & ACC & NMI & ARI & ACC & NMI & ARI \\ \hline
IMC-SeLa & 74.5 & 65.2 & 59.1 & 39.4 & 40.7 & 25.1 & 19.1 & 45.2 & 7.9 \\
SeLA \cite{asano2020self} + Kmeans & 64.0 & 56.4 & 47.3 & 34.7 & 34.3 & 18.0 & 17.1 & 43.2 & 6.7 \\
Supervised & 87.1 & - & - & 66.6 & - & - & 34.2 & - & - \\
\hline
\end{tabularx}
\label{tab: sela}
\end{table}
\section{Conclusion}
This study presents a single-phase framework for image clustering, called IMC-SwAV. We introduce a modified mutual information to jointly train the framework by maximizing the dependency of a self-labelling assignment strategy and an implemented classifier. IMC-SwAV achieves highly competitive performance on challenging datasets compared to the state-of-the-art. In addition to the encouraging clustering performance, we demonstrate the robustness of the proposed method to the parameter settings and the possibility to adopt a different self-supervised learning technique. \par
Although the proposed model achieves significant capabilities in clustering, the multi-crop training strategy adopted in the framework slightly increases the time complexity. Additionally, we observe a dependency of its performance on the grouping based self-supervised learning method. Our future work will aim to remove the requirement of using the multi-crop views without deteriorating the performance of the overall framework. Moreover, we will examine the effectiveness of other grouping based self-supervised learning algorithms in the proposed single-phase training framework to make it available for a wider range of applications.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Motivation and significance}
\label{}
Despite the wide use of open-source libraries for solving systems of linear algebraic equations in
computational software, in some cases there is still the need for developing novel libraries, which can
extend the functionality of well-known ones or add some features, allowing to improve
performance of the calculations. The \textit{hypre} library~\cite{hypre} is an example of an outstanding piece of
software containing powerful and highly scalable algorithms for solving large sparse systems of
linear algebraic equations~(SLAEs). The classical algebraic multigrid method~\cite{Trottenberg}
implemented in \textit{hypre} is a robust method widely applied for solving SLAEs derived from elliptic or
parabolic differential equations. However, the implementation of the methods in \textit{hypre}, as well as
many other libraries containing various modifications of multigrid methods,
is limited by algorithms performing calculations with a single right-hand side vector (RHS).
Meanwhile, the solution of systems with multiple RHSs,
$$A \, X = B,$$
where $A \in \mathbb{R}^{n \times n}$, and $X,B \in \mathbb{R}^{n \times m}$, $m \ll n$, in terms of
computational efficiency can be preferable compared to multiple solutions of systems with single RHS.
The reasons to this include: (i)~increasing the arithmetic intensity, (ii)~regularization of the memory access
pattern, (iii)~vectorization of calculations, and others.
The SLAEs with multiple RHSs occur in a variety of mathematical physics applications. For
example, the need for solving the corresponding systems of linear algebraic equations appears in
structural analysis applications~\cite{Feng1995}, in uncertainty quantification
problems~\cite{Kalantzis2018}, in Brownian dynamics simulations~\cite{Liu2012}, quantum
chromodynamics~\cite{Clark2018}, {computational fluid
dynamics~\cite{KrasnopolskyCPC2018},} and others.
Despite the variety of applications and several advantages mentioned above, implementations of
iterative methods for solving SLAEs with sparse matrices and multiple RHSs can rarely be
found in publicly available libraries of numerical methods. {Among the only few
exceptions is the Trilinos library~\cite{Trilinos} containing the implementation of several Krylov subspace and
aggregation-based algebraic multigrid methods. These, however, do not fully cover the entire
set of the methods used in numerical modeling, e.g. the classical algebraic
multigrid methods, provided by the \textit{hypre} library. This issue complicates the development of the} computational
algorithms performing calculations with multiple RHSs.
The XAMG library focuses on the solution of a series of systems of linear
algebraic equations with multiple RHSs (that, e.g. occur in incompressible turbulent flow
simulations~\cite{KrasnopolskyCPC2018}). The code extends the functionality of the classical
algebraic multigrid method, implemented in \textit{hypre}, and provides implementations of both classical and
merged formulations of the Krylov subspace iterative methods~\cite{Krasnopolsky_CAMWA2020} to solve
systems with multiple RHSs. The XAMG library aims to develop an optimized implementation for the
solve phase of the methods, while the \textit{hypre} library can be reused ``as is'' to construct
a multigrid matrix hierarchy.
The library provides additional optimization features, including the hierarchical multilevel
parallelization, low-level intra-node shared memory communications featuring synchronization
primitives based on POSIX shared memory and atomics, multi-block data segmentation in accordance
with hierarchical parallelization, mixed-precision floating-point calculations, compression of
indices, vector status flags, and others, which allow to outperform the \textit{hypre} library calculation
times.
\section{Software description}
\label{}
\subsection{Mathematical methods}
\label{lab:meth}
The XAMG library contains a set of numerical methods typically used for solving systems of linear
algebraic equations coming from the discretization of elliptic partial differential equations. These
include several Krylov subspace iterative methods (CG~\cite{Hestenes1952} and
BiCGStab~\cite{Vorst1992}), classical algebraic multigrid method~\cite{Trottenberg}, block Jacobi
and symmetric Gauss-Seidel methods, Chebyshev iterative method, and a direct solver. All these methods
are adopted to solve SLAEs with multiple RHSs. In addition to the classical BiCGStab method, the
library also contains several modified methods, including the Reordered BiCGStab~\cite{RBiCGStab},
Pipelined BiCGStab~\cite{Cools2017} and merged formulations of these
methods~\cite{Krasnopolsky_CAMWA2020}, which combine vector updates and dot products to minimize the
overall amount of vector read/write operations. The possible combinations of the methods, which can
be used as a standalone solver, preconditioner, or smoother with the multigrid method, are
summarized in Tab.~\ref{tab:methods_usage}. The full list of the methods and corresponding method
parameters is presented in~\cite{xamg_params_ref}.
\begin{table*}[!th]
\caption{Combinations of the methods and solver types implemented in the XAMG library.}
\label{tab:methods_usage}
\centering
\begin{tabular}{ | c | c | c | c | }
\hline
Method & Solver & Preconditioner & Pre-/post-smoother \\
\hline
CG & + & - & - \\
\hline
BiCGStab & + & + & + \\
\hline
MultiGrid & + & + & - \\
\hline
Jacobi & + & + & + \\
\hline
Gauss-Seidel & + & + & + \\
\hline
Chebyshev & - & + & + \\
\hline
Direct & + & - & - \\
\hline
\end{tabular}
\end{table*}
\subsection{Advantages of calculations with multiple RHSs}
\label{lab:advantages}
The solution of SLAEs with multiple RHSs opens up opportunities for several optimizations.
The main advantage of using iterative methods with performing simultaneous independent solutions for
multiple RHSs is an increasing arithmetic intensity of calculations. The flop per
byte ratio is a measure of floating-point operations relative to the amount of memory accesses.
Basic linear algebra operations like vector
updates, dot products, and sparse matrix-vector multiplications (SpMV) are characterized by flop per byte ratio
of only about~0.1. This means that the corresponding operations are memory bound~\cite{Williams2009,
Krasnopolsky_CAMWA2020}, and its performance mostly depends on the
amount of memory traffic and memory bandwidth. The solution of SLAEs with multiple RHSs
allows to load a matrix for SpMV operations with $m$~right-hand sides only once, compared
to $m$~loads when performing $m$ SLAE solutions with a single RHS. This typically
leads to {about a twofold} reduction in the total amount of memory traffic for SLAE
solver per each RHS, an increase of the flop per byte ratio, and, subsequently, to a significant calculation speedup.
The use of specialized data storage formats for sparse matrices is necessary for performing the
operations with large matrices arising in numerical simulations. The compressed sparse row~(CSR)
format~\cite{Saad2003} is the universal one, which is widely used for the matrices of general form.
However, the implementation of SpMV operation for CSR format requires indirect memory accesses,
which reduces cache efficiency and gives no chance for loop vectorization. The generalized SpMV
operation performed for multiple RHSs allows to partially regularize an access to the memory
and vectorize the computations over the right-hand sides.
\subsection{Software architecture}
\label{sec:arch}
The XAMG library is designed as a header-only template C++ library. The template polymorphism is
used to implement variability of all basic data types for matrices and vectors, both interface-level
and internal ones. The number of RHSs is also given as a template parameter.
The specialization of the number of RHSs at compile time is a key design point.
This approach allows to vectorize all the relevant subroutines with respect to a constant
range of loops iterating over the RHSs. Keeping in mind the appropriate vectors' data
placement and some pragma directives code instrumentation, such a design allows a regular
C++ compiler to vectorize subroutines to the greatest extent possible.
\begin{figure*}[t!]
\centering
\includegraphics[width=13.0cm]{mpi-shm-decomp_v7.eps}
\caption{The hierarchical matrix decomposition.}
\label{fig:hierarchy}
\end{figure*}
One can highlight the following XAMG program code elements:
\begin{itemize}[noitemsep]
\item matrix and vector data structures;
\item basic sparse linear algebra subroutines gathered in ``blas'' and ``blas2'' groups;
\item solver classes inherited from an abstract interface, which exposes the main library
functions: \texttt{setup()} and \texttt{solve()};
\item solver parameter classes.
\end{itemize}
The matrix and vector data structures are used to store the corresponding data objects.
The ``blas'' group of subroutines consists of a collection of primitives operating with vectors
(linear updates, dot products, etc.) The ``blas2'' group of subroutines includes several SpMV-like
matrix-vector operations. Solver classes implement various iterative methods used to solve systems
of linear algebraic equations. Finally, the solver parameter classes are used to store the lists of
numerical method parameters, specifying each numerical method's configuration.
The \verb|setup()| virtual function of the MultiGrid solver class implements the
classical algebraic multigrid ``setup'' phase. This is done by wrapping up the \textit{hypre}
functions to construct the multigrid matrix hierarchy. This is the only place where the
\textit{hypre} library codebase is used inside the XAMG library program code.
An important design feature of the XAMG library is the hierarchical hybrid parallel programming
model. This model combines the Message Passing Interface (MPI) standard~\cite{MPI_standard} for
cross-node communication and a special kind of shared memory parallel programming model on the
intra-node level, based on POSIX shared memory (ShM) functionality. {The MPI+ShM
model implies that MPI ranks, which reside on the same compute node, allocate and use the common
POSIX shared memory regions to store data objects. Unlike many other MPI+X programming models,
MPI+ShM does not use threading within a node. The MPI rank to CPU core mapping is one-to-one.}
To benefit from better data locality, most well-known distributed linear algebra algorithms
implement a data decomposition that splits the calculations into fully local and fully non-local
parts~\cite{Gorobets2018}.
The MPI+ShM programming model also implements a specific data decomposition for matrix and vector
data structures. The idea behind this decomposition is the hierarchical segmentation of the data
into: (i) those with only local algorithmic dependencies, and (ii) those requiring some parallel
communication to work with. In Figure~\ref{fig:hierarchy}, the blue-colored boxes represent the
``local dependency'' blocks, whereas the white-colored ones are for ``remote dependency''. The local
dependency blocks can also be called ``diagonal blocks'' since they contain matrix diagonal
elements for square matrices. The Figure~\ref{fig:hierarchy} also illustrates why the data
decomposition is called hierarchical. In fact, the recursive data decomposition introduces three
logical layers: the node layer, the numa layer, and the core layer. They are meant to reflect the
natural hierarchy of a modern CPU-based high performance computing system hardware, consisting of
the compute nodes, the NUMA blocks inside each node, and the CPU cores.
The hierarchical decomposition principle makes it possible to split the local and non-local data
dependencies. As a practical result of this, the number of non-local MPI communications
(on a cross-node layer) as well as non-local data accesses (for a cross-NUMA layer) is reduced
significantly.
To summarize, the hybrid MPI+ShM parallel programming model implies: (i)~the direct shared memory
data access instead of MPI communications on an intra-node scope, and (ii)~a specific hierarchical data
and algorithms decomposition. Both features combined significantly improve the single-node parallel
performance and the multi-node scalability compared to the pure MPI model.
{The data compression features of the XAMG library try to benefit from reducing
the number of data transfers in the main library algorithms and communication procedures. These features
include the turnover to the reduced precision for floating-point numbers and using lesser bits for integer
indices for the matrices stored in the CSR format. The reduced precision floating-point numbers
typically cannot be used for the whole solver without losing the solution accuracy. However, it is
known that switching to the reduced precision when performing preconditioner
calculations~\cite{Buttari2008, Sumiyoshi2014}
may be a compromise variant, preserving the basic precision of the resulting vector and having only a
minor or no reduction of the convergence rate. Depending on the fraction of the calculations with
the reduced precision in the overall SLAE solver calculations, the 5-30\% cut on the full amount of
data to transfer allows to obtain the proportional calculations speedup. Since the reduced
precision feature is only applied to a part of the multigrid matrices hierarchy, this data
compression model is called ``mixed-precision''.}
{The second type of optimizations related to compression of indices for the CSR
matrix storage format is applicable to all matrices. The hierarchical data decomposition leads to a
multiblock matrices representation, and for the small blocks CSR indices data types can be changed
to the smaller ones like uint16\_t or uint8\_t. This modification also results in reducing the
amount of memory accesses.}
\subsection{Software usage examples}
\label{}
The main API of the XAMG library has a header-only C++11 template-based interface, but the C
language binding also exists. The below overview shows an example of using the main C++ library
interface.
\begin{figure*}[t!]
\begin{lstlisting}[language=C++]
1 #include <xamg/xamg_headers.h>
2 #include <xamg/xamg_types.h>
3 #include <xamg/init.h>
4 #include <xamg/blas/blas.h>
5 #include <xamg/blas2/blas2.h>
6 #include <xamg/solvers/solver.h>
7
8 int main(int argc, char *argv[]) {
9 using FP = double;
10 using matrix_t = XAMG::matrix::csr_matrix<FP, uint32_t, uint32_t,
11 uint32_t, uint32_t>;
12 constexpr uint16_t NV = 1;
13
14 XAMG::init(argc, argv, "nnumas=1:ncores=1");
15
16 matrix_t csr_file_mtx;
17 XAMG::vector::vector csr_file_x, csr_file_b;
18 XAMG::io::read_system<matrix_t, NV>(csr_file_mtx, csr_file_x,
19 csr_file_b, "matrix.csr");
20
21 XAMG::matrix::matrix A(XAMG::mem::DISTRIBUTED);
22 XAMG::vector::vector x(XAMG::mem::DISTRIBUTED);
23 XAMG::vector::vector b(XAMG::mem::DISTRIBUTED);
24 auto part = XAMG::part::make_partitioner(csr_file_mtx.nrows);
25 XAMG::matrix::construct_distributed(part, csr_file_mtx, A);
26 XAMG::vector::construct_distributed<FP, NV>(part, csr_file_x, x);
27 XAMG::vector::construct_distributed<FP, NV>(part, csr_file_b, b);
28
28
29 XAMG::params::global_param_list params;
30 params.add("solver", {"method", "PBiCGStab"});
31 params.add("preconditioner", {"method", "Jacobi"});
32 params.set_defaults();
33
34 auto solver =
35 XAMG::solver::construct_solver_hierarchy<FP, NV>(params, A, x, b);
36
37 solver->solve();
38
39 XAMG::finalize();
40 return 0;
41 }
\end{lstlisting}
\vspace{-0.5cm}
\caption{The XAMG library general usage example, C++ API.} \label{fig:generalusage}
\end{figure*}
The solver code is configured at compile-time depending on the classes' and functions' concrete
template parameters. The C++ template instantiation is a mechanism behind this. A basic matrix
type is instantiated with integer types (uint8\_t, uint16\_t, and uint32\_t are allowed) and
floating-point type (float or double). The integer template types specify the storage types for
nonzero matrix elements indices. The floating-point template type defines the storage type for
the values of these elements. The same floating-point type is also used to instantiate the functions
operating with input and output vectors. The majority of the XAMG data structures and library
functions have an additional template parameter $NV$\footnote{The number of RHSs, $m$,
is denoted in the source code as a template parameter $NV$.}, the number of RHSs within
a single SLAE solution operation.
The code listing in Figure~\ref{fig:generalusage} presents a basic working code portion that uses
the XAMG library to solve a linear system with a sparse matrix stored in the CSR format.
Lines 9-12 set up the compile time definitions for the above mentioned basic data types and the
number of RHSs for a solution. The library initialization call in line~14 passes an optional
colon-separated configuration string. This string can be used to specify the
parameters of the hybrid three-level parallel programming model. The details on these parameters
are given in the program documentation.
Lines 16-19 represent a basic code to read a matrix from a binary file. The file is supposed to
hold a CSR format matrix, an RHS, and an initial guess vector (as an option).
Lines 21-27 fill in the distributed versions of the matrix and the vectors with a correct
decomposition, and map the data structures to
the MPI+ShM data hierarchy. The desired solver parameters are set in lines 29-32. A preconditioned
BiCGStab method is set for usage in this example, with the Jacobi method serving as a
preconditioner. All the default parameters for the solver/preconditioner pair are set up in line 32.
Lines 34-35 create a C++ object representing a solution plan based on the solver and preconditioner
parameters, and this plan becomes connected to the distributed matrix and the vectors $A$, $x$, and
$b$. The SLAE solution process for $NV$ RHSs is done in line 37. The solution result is
stored in $x$ vector, and some solution statistics are held in the $stats$ variable of the $solver$
object.
\begin{figure*}[t!]
\begin{lstlisting}[language=C++]
1 XAMG::params::global_param_list params;
2 params.add("solver", {"method", "PBiCGStab"});
3 params.add_map("solver", {{"max_iters", "20"}});
4 params.add_map("preconditioner", {{"method", "MultiGrid"},
5 {"max_iters", "1"},
6 {"mg_agg_num_levels", "2"},
7 {"mg_coarse_matrix_size", "500"},
8 {"mg_num_paths", "2"}});
9 params.add_map("pre_smoother", {{"method", "Chebyshev"},
10 {"polynomial_order", "2"}});
11 params.add_map("post_smoother", {{"method", "Chebyshev"},
12 {"polynomial_order", "2"}});
13 params.set_defaults();
\end{lstlisting}
\vspace{-0.5cm}
\caption{An example of the numerical method parameters configuration with the BiCGStab method,
supplemented with the algebraic multigrid preconditioner.} \label{fig:mgparams_src}
\end{figure*}
The example set of parameters for a simple algebraic multigrid configuration of a solver is shown in
Figure~\ref{fig:mgparams_src}. In lines 4-8, the multigrid method is set as a preconditioner for a
preconditioned BiCGStab solver. A few parameters which tune the multigrid hierarchy setup process
are set explicitly. In lines 9-12, we take care of an explicit smoothers set up for the
multigrid method. Specifically, the Chebyshev method with a polynomial order of~2 is set. The
\verb|set_defaults()| function call in line~13 ensures that all the solver parameters, as well as
preconditioner and smoother methods, which are not set explicitly in previous lines, are correctly
initialized with the default values.
\section{Example case: performance evaluation}
\label{sec:example_case}
The example case presents the study on the efficiency of the XAMG library and the impact of
implemented optimization features on the overall code performance. {It demonstrates
(i)~the XAMG library performance compared to \textit{hypre}, (ii)~the effect of data compression
optimizations on the calculation time reduction, (iii)~the performance gain due to calculations with
multiple RHSs, and (iv)~the parallel efficiency results for various numbers of RHSs}. The
Lomonosov-2 supercomputer is used for the performance evaluation to demonstrate the potential of the
developed library. It consists of compute nodes with a single Intel Xeon Gold~6126 processor and
InfiniBand FDR interconnect. In all the cases all available 12~physical CPU cores per node are
utilized during the calculations. {The XAMG library is compiled with the Intel
C/C++ compiler~2019\footnote{The compilation options are: \texttt{-ip -O3 -no-prec-div -static
-fp-model fast=2 -xCORE-AVX512}} and Intel MPI Library, Version~2019 Update~9.} The internal
integration test application \verb|xamg_test| is used for the tests; the YAML configuration files
are provided within the XAMG source code repository~\cite{xamg_repo} in \verb|examples/test/yaml/|
directory.
The test runs are performed for several test SLAEs. They correspond to the systems obtained as a
result of spatial discretization of the 3D Poisson's equation $$\Delta u = f$$ with the 7-point
stencil in various computational domains, including the regular grid in the cube (grids with $150^3$
and $250^3$~cells), and the computational grid, corresponding to the direct numerical simulation of
the incompressible turbulent flow in a channel with a matrix of wall-mounted cubes (matrix of 9.7~mln.
unknowns~\cite{KrasnopolskyCPC2018}). The corresponding matrix generators are embedded into the
\verb|xamg_test| integration test application.
\subsection{Calculations with single RHS}
\label{sec:1rhs}
The first set of experiments is performed with two test SLAEs corresponding to cubic computational
domain with the grids of $150^3$ and $250^3$ unknowns. The numerical method configuration, identical
for both XAMG and \textit{hypre}, comprises the BiCGStab method supplemented with an algebraic multigrid
preconditioner and hybrid symmetric Gauss-Seidel smoother. The runs performed include five
different test series:
the \textit{hypre} library tests\footnote{The most recent \textit{hypre} library version~2.20.0
available to date is used in the performance comparison tests.} in pure MPI and hybrid MPI+OpenMP execution
modes, the XAMG library tests in pure MPI and MPI+ShM execution modes, and the XAMG library
tests in hybrid MPI+ShM mode with data compression optimizations.
The obtained performance evaluation results are shown in Figure~\ref{fig:scal}. The data is
presented in terms of the relative speedup, which is defined as a ratio of the single node
calculation time for the \textit{hypre} library in a pure MPI
mode, to the calculation time with the specific number of compute nodes and execution mode:
$$
S^i_p = \frac{T_1^{hypre, MPI}}{T_p^i}.
$$
Results presented in Figure~\ref{fig:scal} show similar scalability
for the pure MPI implementations of XAMG and \textit{hypre}.
However, the MPI+ShM hybrid programming model of the XAMG library provides a big scalability
improvement: with MPI+ShM programming model enabled, XAMG outperforms
\textit{hypre}{, employing the MPI+OpenMP model,} more than twice.
Additionally, the data compression optimization of XAMG adds an extra 10\% speedup.
{The difference in the hybrid programming model benefits, observed
when comparing XAMG and \textit{hypre} libraries, is related to different
hybridization principles for these codes. The MPI+OpenMP hybrid programming model
is known to suffer from fundamental multithreading issues within the
MPI parallel paradigm~\cite{Amer2015}. The MPI+ShM programming model
avoids these pitfalls with the single-threaded design. Among the other reasons can be the difference
in data decomposition principles, solver implementation aspects, implementation of inter-node
communications, or some other algorithmic or programming aspects. However, this topic cannot be elaborated
in detail within the scope of this article.}
\begin{figure*}[t!]
\centering
\includegraphics[width=6.0cm]{scal_Lom2-cubic150.eps}
\includegraphics[width=6.0cm]{scal_Lom2-cubic250.eps}
\caption{Relative calculation speedup for the XAMG and \textit{hypre} libraries in various execution modes,
test matrices with $150^3$ (left) and $250^3$ (right) unknowns.}
\label{fig:scal}
\end{figure*}
\subsection{Calculations with multiple RHSs}
\subsubsection{Single-node performance}
\label{sec:mrhs_1node}
{
The performance gain parameter is a key characteristic indicating the benefit of solving the SLAE
with multiple RHSs over multiple solutions with a single RHS. This parameter is defined as:
$$
P_m = \frac{m T_1}{T_m},
$$
where $m$ is the number of RHSs, $T_1$ is the calculation time for a single RHS, and $T_m$ is the
calculation time for $m$ RHSs. The number of RHSs in the tests varied in the range 1--64 for the
smaller test matrix, and in the range 1--16 for the bigger ones (due to the single node memory
capacity limitations). The corresponding results are presented in Figure~\ref{fig:scal_mrhs}. The
plot shows at least twofold performance gain when performing calculations with
multiple RHSs, which is in agreement with the theoretical estimates~\cite{KrasnopolskyCPC2018}. The
peak values vary in the range 2-2.5, depending on the fraction of SpMV-like operations
in the overall SLAE solver calculation loop.}
\begin{figure}[t!]
\centering
\includegraphics[width=6.0cm]{1node_mrhs_ed.eps}
\caption{Performance gain for the calculations with multiple RHSs.}
\label{fig:scal_mrhs}
\end{figure}
\subsubsection{Parallel efficiency}
The XAMG library performance is also evaluated by the multiple RHS calculations parallel
efficiency. The corresponding tests are performed with the test matrix of 9.7~mln. unknowns with 1, 4, and 16~RHS
vectors. The parallel efficiency results,
$$
E_p = \frac{T_1}{p \, T_p},
$$
are obtained for the hybrid execution mode. The corresponding results are presented in
Figure~\ref{fig:scal_mrhs}. These results show that the parallel efficiency for multiple RHSs
is, at least, not worse compared to the one for the single RHS calculations.
\begin{figure}[t!]
\centering
\includegraphics[width=6.0cm]{scal_Lom2-mrhs.eps}
\caption{The parallel efficiency for the multiple RHSs calculations with the XAMG library, 9.7~mln.
unknowns test matrix.}
\label{fig:scal_mrhs}
\end{figure}
\section{Impact and conclusions}
\label{}
The new linear solver code XAMG is a modern C++ project which adds a rare feature for the libraries
of iterative methods, as it solves systems with multiple right-hand side vectors.
The key advantage of the solver
with multiple RHSs is a potential speedup over multiple solutions with single RHS, and this feature
can be effectively used in some well-known mathematical methods for structural analysis,
uncertainty quantification, Brownian dynamics simulations, quantum chromodynamics, modeling of
incompressible turbulent flows, and others. However, despite evident
advantage in time to solution, this functionality is not implemented in most
libraries containing robust and scalable iterative methods, like algebraic multigrid.
Among the real applications already taking up this advantage, we highlight an in-house code for
incompressible turbulent flow simulations~\cite{KrasnopolskyCPC2018}, since the XAMG library is
already successfully integrated with it. The integration can be done as well for any other project
written in C/C++ that is able to exploit the multiple RHS feature of a linear solver because the
initial release of the XAMG library has been recently published as an open repository
on gitlab.com~\cite{xamg_repo}.
The XAMG project also features a number of leading source code and computational method optimizations
which impact the generic problem of a linear solver speedup on modern multicore and manycore
HPC systems. This problem stays aside from the multiple RHS feature. These optimizations like a
novel hybrid parallel programming model (MPI+ShM), mixed-precision calculations, matrix indices
compression, and others make it possible to improve the productivity of a linear solver compared to
other well-known libraries containing robust and scalable iterative methods (e.g. \textit{hypre}). The
modern C++11 codebase allows for broad future research prospects in this direction.
{The experimental performance evaluation shows the XAMG library speedup against
\textit{hypre} by a factor of~2 for the multi-node runs performed. The comparison of calculations
with multiple RHSs against multiple runs with single RHS shows 2-2.5 times speedup for single-node
runs. Additionally, the multi-node parallel efficiency of the multiple RHSs calculations is, at
least, not worse compared to the one for the single RHS.}
\section*{Declaration of competing interest}
The authors declare that they have no known competing financial interests or personal relationships
that could have appeared to influence the work reported in this paper.
\section*{Acknowledgments}
\label{lab:ack}
The presented work is supported by the RSF grant No. 18-71-10075. The research is carried out using
the equipment of the shared research facilities of HPC computing resources at Lomonosov Moscow State
University.
\bibliographystyle{elsarticle-num}
|
\section*{Acknowledgment}
The authors would like to thank Jasan Zughaibi, Michael Egli, Matthias Müller, and Helen Hanimann for their contributions to this work.
\section*{Appendix}
\section{Ball Catching Application}\label{sec:application}
In this section, the ball catching application using the offset-free MPC is discussed. First, the intersection point between the thrown ball and the spherical range of the robotic arm is predicted with an estimator. Subsequently, the predicted intersection point is used in a motion planner to generate a trajectory of angle set points for the offset-free MPC. Finally, the set point trajectory guides the controller to catch the ball.
\subsection{Intersection Estimator}
An estimator is designed to predict where the thrown ball intersects with the sphere of the arm. The current position of the ball is measured using the motion capture system. An extended Kalman filter \cite{simon2006optimal} is applied to smooth the ball position measurements and estimate the velocity of the ball as well as the aerodynamic drag coefficient. The predicted intersection point between the ball and the spherical range of the arm is computed by forward-predicting the ball model. \par
The ball is modeled as a point mass under the
influence of gravity and aerodynamic drag where spin is neglected \cite{muller2011quadrocopter}. With the ball's position denoted as ${r_B}$, the equation of motion is given by,
\begin{equation}
\Ddot{{r}}_B = {g} \MINUS K_D ||\Dot{{r}}_B||\Dot{{r}}_B \text{\,,}
\end{equation}
where $K_D$ is the constant aerodynamic drag coefficient, and ${g} = [0,\,0,\,\MINUS9.81]^T$ denotes the gravity vector. As the aerodynamic drag coefficient $K_D$ varies for different balls \cite{muller2011quadrocopter}, it is included in the state to be estimated online. The ball state is defined as
\begin{equation}
x_B = (r_B,\, \Dot{r}_B,\, K_D) \text{\,.}
\end{equation}
The dynamic model of the ball is formulated as
\begin{equation}
\label{eq: ball's dynamic model}
\Dot{x}_B =
\begin{bmatrix}
\Dot{r}_B\\
{g} \MINUS K_D ||\Dot{r}_B||\Dot{r}_B\\
0
\end{bmatrix} \text{\,.}
\end{equation}
The measurement model of the ball is given by
\begin{equation}
\label{eq: ball's measurement model}
z_B = [{I}_{3\times3} \quad {0}_{3\times4}]\,x_B \text{\,,}
\end{equation}
where $z_B$ denotes the measurement of the ball position. The dynamic model (\ref{eq: ball's dynamic model}) is discretized using the Runge-Kutta 4th order method with a sampling time of 0.005 s. With the discretized ball model and the measurement model (\ref{eq: ball's measurement model}), the standard extended Kalman filter formulation is applied to design an estimator for the ball state $x_B$.\par
Once the ball is detected in the air, the extended Kalman filter starts to update. Given the estimate of the ball state from the filter, the state is forward predicted until the ball position intercepts with the sphere, as shown in Fig. \ref{fig:MotionPlanning}. The predicted intersection point is expressed in the angular space as $(\tilde{\alpha},\,\tilde{\beta})$.
\begin{figure}[t]
\centering
\resizebox{0.45\textwidth}{!}{\input{img/MotionPlanning.tikz}}
\caption{The intersection point $(\tilde{\alpha},\,\tilde{\beta})$ is predicted using the ball's state estimate and its forward prediction. In the motion planner, the shortest path between the current end-effector position $(\alpha,\,\beta)$ and the predicted intersection point $(\tilde{\alpha},\,\tilde{\beta})$ is computed, which lies on the great circle of the sphere. The shortest path is denoted as the dashed line between these two positions.}
\label{fig:MotionPlanning}
\end{figure}
\subsection{Motion Planner}
At each time step, a desired set point trajectory that connects the current end-effector position $(\alpha,\,\beta)$ and the estimated intersection point $(\tilde{\alpha},\,\tilde{\beta})$ is commanded to the MPC controller. Although it is possible to command the offset-free MPC to directly track the estimated intersection point $(\tilde{\alpha},\,\tilde{\beta})$ as a single step, planning a smoother trajectory for the controller can reduce overshoot and consequently increase the success rate of catching a ball. As the end-effector is moving on a sphere, the shortest path between $(\alpha,\,\beta)$ and $(\tilde{\alpha},\,\tilde{\beta})$ is the circular arc that passes these two positions and lies on a great circle \footnote{The great circle of a sphere is the intersection of the sphere and a plane that passes through the center point of the sphere.} of the sphere (see Fig. \ref{fig:MotionPlanning}). The angle of this circular arc is denoted as $\theta$. By setting a constant angular velocity $\omega_{SP}$, this circular arc is divided into $M$ pieces with equal length, where $M = \theta/(\omega_{SP}T_s)$. If $M$ is not an integer, ceiling is applied. The evenly divided positions on the arc are denoted as $(r_0, \dots, r_M)$,
where $r_0 = (\alpha,\,\beta)$ and $r_M = (\tilde{\alpha},\,\tilde{\beta})$. It is worth emphasizing that the trajectory $(r_0, \dots, r_M)$ is the shortest path between the current arm position and the predicted intersection point. Note that this path does not coincide with a straight line in the $\alpha$-$\beta$-plane.
The first $N+1$ elements of $(r_0, \dots, r_M)$ are set as the set point trajectory for the offset-free MPC. If M is smaller than $N$, the last set point $r_M$ is repeated. In practice, $\omega_{SP}$ is set to $240 \,\mathrm{deg/s}$ to generate a fast trajectory for ball catching. The angular velocity of $240 \,\mathrm{deg/s}$ has been assessed in the trajectory tracking experiment shown in Section \ref{sec:MPC}, and it is within the tracking capability of the offset-free MPC.
\begin{figure}[t]
\vspace{5 pt}
\centering
\begin{subfigure}{.245\textwidth}
\centering
\includegraphics{img/BallCatch_SuccessFail_plot.pdf}
\end{subfigure}%
\begin{subfigure}{.245\textwidth}
\centering
\includegraphics{img/BallCatch_Precision_Projection.pdf}
\end{subfigure}
\caption{Statistical analysis of the ball catching application. In the left hand plot, the intersection points are expressed in the angular space. The green dashed boundary denotes the reachable range of the robotic arm given the input constraints listed in Section \ref{sec:MPC}, which is identified from experiments. The blue points are the intersection points of the successful catches, while the red points are the intersection points of the unsuccessful ones. Because the balls are thrown from a fixed position, the intersection points concentrate on a corner of the angular space. In the right hand plot, the intersection points of all the successful throws are projected on to the plane of the net opening.}
\label{img:statistic analysis}
\end{figure}
\begin{figure}[t]
\vspace{7 pt}
\centering
\includegraphics{img/BallCatch_MotionPlanner_Windy.pdf}
\caption{A successful catch when a wind gust disturbance acts on the ball. The $x$ coordinate is expressed in the world frame, with the origin being re-centered to the true intersection point. The blue line denotes the position of the net opening, the orange line the position of the thrown ball, and the green line the predicted intersection point between the ball and the sphere. The red dashed lines denote $\pm31$ mm (the radius of the net opening) from the true intersection point. As the predicted intersection point is used to generate the set point trajectory for the offset-free MPC, the net is controlled toward the predicted intersection point and eventually catches the ball.}
\label{img:BallCatch_MotionPlanner_Windy}
\end{figure}
\subsection{Experimental Results}
The same controller settings are used for the ball catching experiments as for the evaluation experiment shown in Fig. \ref{img:soft step trj}.
For a statistical analysis of the success rate and the accuracy, the ball catching experiments where a person throws the ball at a fixed position two meters away from the robotic arm are repeated multiple times. The throws that do not intercept with the sphere are not considered in the analysis because they are physically impossible for the robotic arm to catch. The experimental results are shown in Fig. \ref{img:statistic analysis} (left). As the ball is thrown from the positive to the negative $\alpha$-direction, a throw may align with the net opening poorly with a trajectory nearly tangential to the sphere if it lands in the region with negative $\alpha$.
Additionally, a throw may collide with the net's frame when it lands closer to the boundary of the system's range of motion. As a result, these regions have a lower success rate. Among 68 throws, 62 throws were successfully caught by the robotic arm while six throws were not, which equals a success rate of 91\%.
In Fig. \ref{img:statistic analysis} (right), the intersection points of the successful throws are projected on to the plane of the net opening. The average distance from these projected points to the center of the net opening is 9.7 mm, around one-third of the radius of the net opening (31 mm).\par
In order to test the robustness against disturbances of the proposed control approach, ventilation fans are used to generate wind gusts acting as disturbances in the ball catching experiment. As the predicted intersection point and the set point trajectory from the motion planner are recomputed at every time step, it allows the controller to reject disturbances applied on the flying ball. Because the wind mainly acts in the positive $x$-direction, the predicted intersection point is constantly re-adjusted in this direction, as shown in Fig. \ref{img:BallCatch_MotionPlanner_Windy}. When the ball gets closer to the sphere, the predicted intersection point becomes more accurate. As a result, the net is controlled toward the true intersection position and eventually successfully catches the ball. A video showing the ball catching experiment under the wind gust disturbance is available in the video attachment (https://youtu.be/b8ov1Jzd89k).\par
We also tried to implement the ball catching application with the standard MPC. However, the success rate is below 30\% for the same experiment as shown in Fig. \ref{img:statistic analysis} due to the significant tracking offsets.
\section{Conclusion}\label{sec:conclusion}
An offset-free MPC controller is implemented on a soft robotic arm. This controller employs a simple linear model and compensates for the residual model errors with a disturbance estimation scheme. In the trajectory tracking evaluation, it outperforms the standard MPC with a 35\% smaller tracking error. Furthermore, the offset-free MPC is used to realize a ball catching application. It is shown that this controller can successfully catch a ball while rejecting external disturbances, which demonstrates the responsiveness, accuracy, and robustness of the control approach proposed. These results indicate that the offset-free MPC can cope with control challenges associated with the use of soft materials despite the simplicity of the employed model.\par
Currently, the stiffness-related variable $\bar{p}$ is set to a constant value in modeling and control, which limits the joint stiffness of the soft robotic arm. Future work will apply offset-free MPC to a model parametrized by $\bar{p}$ for stiffness control of the robotic manipulator.
\section{Introduction}\label{sec:Introduction}
Pneumatically actuated soft robots show potential in various applications \cite{polygerinos2017soft}, \cite{gaiser2012compliant}, including prosthetics \cite{pneumaticprosthetichand} and
general automation tasks like pick and place \cite{SoftPickAndPlace}. Their inherent compliance and low
inertia make them safe to work alongside humans \cite{sanan2011physical}, and the use of pneumatic actuation allows for fast maneuvers \cite{mosadegh2014pneumatic}. However, these advantages come at the cost of a more challenging control task. The use of soft materials introduces dynamical effects such as material relaxation that are hard to describe with compact models suitable for control and can lead to tracking offsets and an overall degradation of control performance.\par
Model predictive control (MPC) is an optimization-based control approach: At each time step, the future evolution of the system is predicted based on a model of the system \cite{morari1999model}. The optimal input sequence is computed by solving an optimization problem where a control objective is minimized along the prediction horizon. Thereby, state and input constraints can be explicitly incorporated, making MPC a systematic control approach for handling constraints, while providing an optimal control input for a given objective function.\par
The authors of \cite{gillespie2016simultaneous} apply MPC to a single degree of freedom antagonistic soft inflatable robot. By penalizing the tracking error and the pressure magnitude in the objective function, the resulting model predictive controller can achieve simultaneous position and stiffness control. As a continuation of \cite{gillespie2016simultaneous}, a neural network is used to describe the dynamics of the soft inflatable robot and MPC is applied to the learned model \cite{gillespie2018learning}. To control a six degrees of freedom soft robotic platform, a model predictive controller based on a neural network model is implemented in \cite{hyatt2019model}, where training data is generated from a first principle model.\par
\begin{figure}[!th]
\centering
\includegraphics[height=2in]{img/V3BallCatch.png}
\caption{The inflatable soft robotic arm used for the experimental evaluation of the proposed control approach. The system is pneumatically actuated and a 3D printed net is attached to the arm as an end-effector for catching a ball.}
\label{img:V3photo}
\end{figure}
MPC is capable of rejecting impulse disturbances as it introduces state feedback into the optimization problem in a receding horizon fashion. However, MPC can not track a set point in steady state with zero offset in the presence of persistent model errors or external disturbances. The reason for these tracking offsets is that MPC is not able to correctly predict the system's behaviors without the knowledge of the model errors or the external disturbances. Integral action can be applied to MPC to compensate for occurring offsets as demonstrated in \cite{gillespie2018learning} and \cite{hyatt2019model}. Adaptive control schemes, which adapt model parameters online, can be combined with MPC to eliminate steady state tracking errors, as shown in \cite{hyatt2020model}. The authors combine model reference adaptive control with MPC on a continuum joint robot to compensate for model mismatch. An alternative method to eliminate tracking offsets is the control approach referred to as offset-free MPC (\cite{borrelli2007offset}, \cite{pannocchia2015offset}). Thereby, the system model is augmented with a disturbance state that is estimated online and compensated for in the MPC optimization problem. This allows the elimination of tracking offsets in a systematic way.\par
Ball catching has been extensively studied as a benchmark task to evaluate robotic systems, because it requires fast and accurate control performance. Various examples of ball catching have been demonstrated for rigid body robots (\cite{ExperimentsinRoboticCatching}, \cite{bauml2010kinematically}) and for a robot arm composed of flexible links \cite{link_flexible}. A soft prosthetic hand capable of grasping a ball thrown towards the hand is discussed in \cite{2018elastomeric}. However, to the best of our knowledge, no ball catching application has been demonstrated with a fully soft robotic system that can autonomously catch a thrown ball by moving to the predicted intersection point.\par
In this work, ball catching is realized with a spherical soft robotic arm (see Fig. \ref{img:V3photo}). As the application requires the controller to track a trajectory that is unknown beforehand, previously explored control approaches such as iterative learning control (as used in \cite{zughaibi2020fast} for realizing a pick and place application) are not feasible for this application. Therefore, an offset-free MPC approach that can deal with an a priori unknown reference trajectory is employed for realizing the ball catching application. \par
The remainder of this paper is organized as follows: A brief overview of the soft robotic system is given in Section \ref{sec:platform}, followed by the modeling of the robotic manipulator in Section \ref{sec:Modeling}. The offset-free MPC approach is discussed in Section \ref{sec:MPC} and experimentally evaluated through a comparison to a standard MPC approach. The realization of the ball catching application is presented in Section \ref{sec:application} and a conclusion is drawn in Section \ref{sec:conclusion}.
\section{System Modeling}\label{sec:Modeling}
In this section, the modeling of the soft robotic arm is presented. First, a control allocation strategy that simplifies the modeling process is discussed. Subsequently, a linear state-space model is proposed where its parameters are determined from system identification experiments.
\subsection{Control Allocation}
The three actuator pressures $(p_A,\, p_B,\, p_C)$ form the inputs to the robotic arm. Beside the two orientation degrees of freedom, there is an additional degree of freedom related to the joint stiffness of the system (see \cite{zughaibi2020fast}, \cite{hofer2020design}). In order to simplify the modeling of the system, an alternative representation of the three inputs is employed using two pressure differences $(\Delta{p}_{\alpha},\, \Delta{p}_{\beta})$ and a lower pressure bound $\bar{p}$ (see \cite{zughaibi2020fast} for a detailed description). The two pressure differences are related to the orientational degrees of freedom and the lower pressure bound to the joint stiffness.\par
The bijective mapping relates the two representations, namely
\begin{equation}
(\Delta{p}_{\alpha},\, \Delta{p}_{\beta},\, \bar{p}) = \xi(p_A,\, p_B,\, p_C) \text{\,,}
\end{equation}
which is defined by concatenating
\begin{equation}
\label{}
\begin{bmatrix}
\Delta{p}_{\alpha}\\
\Delta{p}_{\beta}\\
\end{bmatrix} = T
\begin{bmatrix}
\Delta{p}_{AB}\\
\Delta{p}_{BC}\\
\end{bmatrix} \quad \text{with} \quad T =
\begin{bmatrix}
0 & \sqrt{3}/2\\
\MINUS1 & \MINUS1/2
\end{bmatrix}
\end{equation}
and
\begin{equation}
\label{transformation:absolute to delta}
\begin{aligned}
&\Delta{p_{AB}} = p_{A} \MINUS p_{B} \\
&\Delta{p_{BC}} = p_{B} \MINUS p_{C} \text{\,.}
\end{aligned}
\end{equation}
The lower pressure level is defined as
\begin{equation}
\Bar{p} = \text{min} \{p_{A}, p_{B}, p_{C}\} \text{\,.}
\end{equation}
The mapping $\xi$ roughly aligns the resulting $\Delta{p}_{\alpha}$ and $\Delta{p}_{\beta}$ with the $\alpha$-direction and the $\beta$-direction, respectively. More specifically, for small values of $\alpha$ and $\beta$, a change in $\Delta{p}_{\alpha}$ only causes a deflection in the $\alpha$-direction, and a change in $\Delta{p}_{\beta}$ only causes a deflection in the $\beta$-direction. Therefore, applying the $\xi$ mapping to the actuator pressures and their set points can decouple the system dynamics in the $\alpha$- and $\beta$-directions and simplify the modeling process.\par
The inverse mapping,
\begin{equation}
(p_A,\, p_B,\, p_C) = \xi^{\MINUS1}(\Delta{p}_{\alpha},\, \Delta{p}_{\beta},\, \bar{p}) \text{\,,}
\end{equation}
is given by
\begin{equation}
\begin{aligned}
&p_{A} = \text{max}\{\Bar{p},\, \Bar{p} + \Delta{p_{AB}},\, \Bar{p} + \Delta{p_{AB}} + \Delta{p_{BC}} \}\\
&p_{B} = \text{max}\{\Bar{p},\, \Bar{p} + \Delta{p_{BC}},\, \Bar{p} \MINUS \Delta{p_{AB}} \}\\
&p_{C} = \text{max}\{\Bar{p},\, \Bar{p} \MINUS \Delta{p_{BC}},\, \Bar{p} \MINUS \Delta{p_{AB}} \MINUS \Delta{p_{BC}} \}\\
&\begin{bmatrix}
\Delta{p}_{AB}\\
\Delta{p}_{BC}\\
\end{bmatrix}
= T^{\MINUS1}
\begin{bmatrix}
\Delta{p}_{\alpha}\\
\Delta{p}_{\beta}\\
\end{bmatrix} \text{\,.}
\end{aligned}
\end{equation}
\subsection{Inputs and States}
The input vector $u$ of the system is defined as the set points of $\Delta{p}_{\alpha}$ and $\Delta{p}_{\beta}$, namely
\begin{equation}
\label{eq:input vector}
u :=
(\Delta{p}_{\alpha SP},\,
\Delta{p}_{\beta SP}) \text{\,.}
\end{equation}
The pressure set points applied to the low-level PID controller $p_{ASP}$, $p_{BSP}$, and $p_{CSP}$ are retrieved by applying the inverse mapping $\xi^{\MINUS1}$ to $u$.
The stiffness-related variable $\bar{p}$ is set to a constant value of $\bar{p} = 1.05$ bar throughout this work, as the control of joint stiffness is not relevant in the ball catching application.
\par
The state vector $x$ of the system is defined as
\begin{equation}
\label{eq:state vector}
x :=
(\alpha,\,
\Dot{\alpha},\,
\Delta{p}_{\alpha},\,
\beta,\,
\Dot{\beta},\,
\Delta{p}_{\beta})
\text{\,.}
\end{equation}
All states in $x$ are computed from sensor data. The angles $\alpha$ and $\beta$ are retrieved from position data provided by the motion capture system, and $\Dot{\alpha}$ and $\Dot{\beta}$ are computed by applying the finite difference method to the angle measurements. The pressure states $\Delta{p}_{\alpha}$ and $\Delta{p}_{\beta}$ are retrieved by applying the $\xi$ mapping to the pressure measurements.
\subsection{Model Structure}
The proposed model consists of the decoupled arm dynamics describing the motion of the robotic arm and the decoupled pressure dynamics in the $( \Delta{p}_{\alpha},\,
\Delta{p}_{\beta})$ space. \par
As a result of the $\xi$ mapping, the dynamics of $\alpha$ and $\beta$ are assumed to be decoupled. The arm dynamics in $\alpha$- and $\beta$-directions can therefore be modeled as two spring-damper systems,
\begin{equation}
\begin{aligned}
\label{V3 linear arm model}
\Ddot{\alpha} &= \MINUS k_\alpha \alpha \MINUS d_\alpha \Dot{\alpha} + h_{\alpha}\Delta{p}_{\alpha}\\
\Ddot{\beta} &= \MINUS k_\beta \beta \MINUS d_\beta \Dot{\beta} + h_{\beta}\Delta{p}_{\beta} \text{\,,}
\end{aligned}
\end{equation}
where $k_\alpha$, $k_\beta$ are the stiffness coefficients, $d_\alpha$, $d_\beta$ the damping coefficients, and $h_{\alpha}$ and $h_{\beta}$ are coefficients that map pressure differences to angular excitations. All coefficients are mass-normalized. Note that the decoupling property of the $\xi$ mapping is only valid for small values of $\alpha$ and $\beta$. Consequently, the decoupled arm model (\ref{V3 linear arm model}) is an approximation of the true dynamics for large angles.\par
Given that the pressure dynamics are controlled in the inner control loops, they are modeled as two decoupled first-order systems. Thereby, interactions with the arm movement are taken into account. In particular, $\Dot{\alpha}$ is assumed to linearly affect $\Delta\Dot{p}_\alpha$, and analogously for $\Dot{\beta}$ and $\Delta\Dot{p}_\beta$. Therefore, the pressure dynamics can be written as,
\begin{equation}
\begin{aligned}
\label{V3 interacting pressure model}
\Delta\Dot{p}_\alpha &= 1/\tau_{\alpha}(\Delta{p}_{\alpha SP} \MINUS \Delta{p}_{\alpha}) + c_{\alpha} \Dot{\alpha}\\
\Delta\Dot{p}_\beta &= 1/\tau_{\beta}(\Delta{p}_{\beta SP} \MINUS \Delta{p}_{\beta}) + c_{\beta} \Dot{\beta} \text{\,,}
\end{aligned}
\end{equation}
where $\tau_{\alpha}$, $\tau_{\beta}$ are the time constants of the closed-loop pressure dynamics and $c_{\alpha}$, $c_{\beta}$ are related to the interaction with the arm movements.\par
\subsection{System Identification}
A series of sinusoidal signals with frequencies ranging from 0.5 Hz to 5 Hz are applied to both inputs $\Delta{p}_{\alpha SP}$ and $\Delta{p}_{\beta SP}$ to generate training data to identify the arm dynamics (\ref{V3 linear arm model}). The sinusoidal signals of different frequencies are applied with equal time duration to generate balanced data for each frequency. Step inputs, which excite all frequencies simultaneously, are used to generate training data to identify the pressure dynamics (\ref{V3 interacting pressure model}). The magnitudes of the applied steps cover a large range of the input space $(\Delta{p}_{\alpha SP},\,\Delta{p}_{\beta SP})$. \par
The recorded data sets contain time series of variables $\alpha$, $\beta$, $\Delta{p}_{\alpha}$, $\Delta{p}_{\beta}$, $\Delta{p}_{\alpha SP}$ and $\Delta{p}_{\beta SP}$. The differentiated values $\Dot\alpha$, $\Ddot{\alpha}$, $\Dot\beta$, $\Ddot{\beta}$, $\Delta\Dot{p}_\alpha$ and $\Delta\Dot{p}_\beta$ are obtained through differentiation of the spline interpolations of $\alpha$, $\beta$, $\Delta{p}_{\alpha}$ and $\Delta{p}_{\beta}$. All variables are normalized to $[\MINUS1, 1]$. Finally, the parameters in (\ref{V3 linear arm model}) and (\ref{V3 interacting pressure model}) are identified via linear regression.
\section{Offset-free Model Predictive Control}\label{sec:MPC}
In this section, the implementation of the offset-free MPC approach is discussed. First, the model identified in the previous section is augmented with a disturbance state, and then the disturbance is estimated with a steady state Kalman filter \cite{simon2006optimal}. Second, the tracking target that accounts for the estimated disturbance and ensures offset-free tracking is computed in the target calculation problem. Third, the constraint set for the MPC optimization problem is defined. Finally, the three components are combined to formulate the MPC optimization problem.
The effectiveness of the offset-free MPC approach is evaluated and compared to a standard MPC without a disturbance compensation scheme.
\subsection{Disturbance Augmentation and Estimation}
The system model combining the arm dynamics (\ref{V3 linear arm model}) and the pressure dynamics (\ref{V3 interacting pressure model}) can be reformulated as
\begin{equation}
\label{eq: V3 linear Model Reformiulated}
\begin{aligned}
\Dot{x} &=
\underbrace{\begin{bmatrix}
0 &\mkern-15mu 1 &\mkern-15mu 0 & & &\\
\MINUS k_\alpha &\mkern-15mu\MINUS d_\alpha &\mkern-15mu h_{\alpha} & & &\\
0 &\mkern-15mu c_{\alpha} &\mkern-15mu \MINUS 1/\tau_{\alpha} & & &\\
& & &\mkern-15mu 0 &\mkern-15mu 1 &\mkern-15mu 0\\
& & &\mkern-15mu\MINUS k_\beta &\mkern-15mu\MINUS d_\beta &\mkern-15muh_{\beta}\\
& & &\mkern-15mu 0 &\mkern-15mu c_{\beta} &\mkern-15mu \MINUS 1/\tau_{\beta}
\end{bmatrix}}_{=:A^c} x +
\underbrace{\begin{bmatrix}
0 &\\
0 &\\
1/\tau_{\alpha} &\\
&\mkern-15mu 0\\
&\mkern-15mu 0\\
&\mkern-15mu 1/\tau_{\beta}\\
\end{bmatrix}}_{=:B^c}u \text{\,,}
\end{aligned}
\end{equation}
where $u$ and $x$ are defined in (\ref{eq:input vector}) and (\ref{eq:state vector}).
Assuming that the disturbance acting on the system remains constant across the MPC prediction horizon, the model is augmented with the following constant disturbance $d$,
\begin{equation}
\label{eq:disturbance augmented model}
\begin{aligned}
\Dot{x} &= A^cx + B^cu + d\\
\Dot{d} &= 0 \text{\,.}
\end{aligned}
\end{equation}
The augmented model is discretized using the exact discretization method with a sampling time $T_s$, which results in the following discrete time linear model,
\begin{equation}
\label{discrete dis augmented model}
\begin{bmatrix}
x\idxk\\
d\idxk
\end{bmatrix} =
\begin{bmatrix}
A &E\\
0 &I
\end{bmatrix}
\begin{bmatrix}
x\idxkmo\\
d\idxkmo
\end{bmatrix} +
\begin{bmatrix}
B\\
0
\end{bmatrix}u\idxkmo \text{\,,}
\end{equation}
where $A$, $B$ and $E$ are the discrete time matrices, $I$ the identity matrix, and $k$ the time index. \par
As discussed in Section \ref{sec:Modeling}, all states in $x$ are retrieved from sensor data, therefore the measurement model of the augmented system is as follows,
\begin{equation}
\label{measurement model}
z\idxk =
\begin{bmatrix}
I &0
\end{bmatrix}
\begin{bmatrix}
x\idxk\\
d\idxk
\end{bmatrix}\text{\,,}
\end{equation}
where $z\idxk$ denotes the measurement at time step $k$.\par
We assume additive zero-mean Gaussian noise in the process model (\ref{discrete dis augmented model}) and the measurement model (\ref{measurement model}). The standard steady state Kalman filter formulation \cite{simon2006optimal} is applied to design an estimator for the augmented state, and the resulting estimate is denoted as $(\hat{x}\idxk,\,\hat{d}\idxk)$. The update equations of the steady state Kalman filter is given by,
\begin{equation}
\begin{bmatrix}
\hat{x}\idxk\\
\hat{d}\idxk
\end{bmatrix}
= \hat{A}
\begin{bmatrix}
\hat{x}\idxkmo\\
\hat{d}\idxkmo
\end{bmatrix} + \hat{B}u\idxkmo + K_{\infty}z(k) \text{\,,}
\end{equation}
where $K_{\infty}$ is the steady state Kalman filter gain \cite{simon2006optimal} and $\hat{A}$, $\hat{B}$ are defined as follows,
\begin{equation}
\begin{aligned}
\hat{A} &= (I \MINUS K_{\infty}\begin{bmatrix}I&0\end{bmatrix}) \begin{bmatrix}
A &E\\
0 &I
\end{bmatrix}\\
\hat{B} &= (I \MINUS K_{\infty}\begin{bmatrix}I&0\end{bmatrix})
\begin{bmatrix}
B\\
0
\end{bmatrix} \text{\,.}
\end{aligned}
\end{equation}
\subsection{Target Calculation}
Given a single desired angle set point $r = (\alpha_{SP}, \,\beta_{SP})$ and the current disturbance estimate $\hat{d}$, the target state and input $(\bar{x},\,\bar{u})$ that account for the estimated disturbance and ensure offset-free tracking are computed by solving the following linear equations (see \cite{borrelli2007offset} for more details),
\begin{equation}
\label{eq: V3 linear ss computation}
\begin{bmatrix}
A\MINUS{I} &B\\
H &0
\end{bmatrix}
\begin{bmatrix}
\bar{x}\\
\bar{u}
\end{bmatrix} =
\begin{bmatrix}
\MINUS E \hat{d} \,\\
r
\end{bmatrix} \text{\,,}
\end{equation}
where $H$ selects the first and the fourth entries of $\bar{x}$,
\begin{equation}
H = \begin{bmatrix}
1 &0 &0 &0 &0 &0\\
0 &0 &0 &1 &0 &0
\end{bmatrix} \text{\,.}
\end{equation}
If no unique solution exists, the pseudo-inverse is applied to compute $(\bar{x},\,\bar{u})$. The resulting target $\bar{x}$ and $\bar{u}$ take into account the disturbance acting on the state evolution and the tracking of the set point.\par
In most tracking tasks, however, the goal is not to track a single set point, but rather a trajectory of desired set points. In this case, the target calculation for a single set point is extended to address a trajectory of set points. At each time step, a set point trajectory denoted as $(r_0, \dots ,r_N)$ is required to be tracked by the MPC controller, where $N$ is the MPC prediction horizon length, and $r_i$ represents the desired set point $i$ time steps ahead of the current time step. A trajectory of target states and inputs is computed by repeatedly solving (\ref{eq: V3 linear ss computation}) for each set point,
\begin{equation}
\begin{aligned}
\label{eq: V3 linear ss computation trj}
\begin{bmatrix}
A\MINUS{I} &B\\
H &0
\end{bmatrix}
\begin{bmatrix}
\bar{x}_{i}\\
\bar{u}_{i}
\end{bmatrix} =
\begin{bmatrix}
\MINUS E \hat{d} \,\\
r_i
\end{bmatrix}\text{\,,} \quad
i = 0, \dots, N \text{\,,}
\end{aligned}
\end{equation}
where $\hat{d}$ is assumed to be constant over the prediction horizon, and $(\bar{x}_{i},\bar{u}_{i})$ denotes the target at prediction step $i$. The target trajectory $(\bar{x}_{0},\bar{u}_{0},\dots,\bar{x}_{N},\bar{u}_{N})$ allows the MPC to optimally plan for the future set points and improve trajectory tracking performance.
\subsection{Constraint Formulation}
The following constraints apply to the pressure set point of each individual actuator,
\begin{equation}
\label{eq:V3 Input Constraints}
\begin{aligned}
p_{\text{min}} \leq \, &p_{ASP} \leq p_{\text{max}}\\
p_{\text{min}} \leq \, &p_{BSP} \leq p_{\text{max}}\\
p_{\text{min}} \leq \, &p_{CSP} \leq p_{\text{max}} \text{\,,}
\end{aligned}
\end{equation}
where $p_{\text{min}}$ is set to ambient pressure (1.0 bar) and $p_{\text{max}}$ to the maximum allowed pressure (1.9 bar). These constraints limit the pressure set points of the actuators to a safe range. Using the mapping $\xi$, these constraints can be converted to constraints applying to $\Delta{p}_{\alpha SP}$ and $\Delta{p}_{\beta SP}$. A visualization of the resulting constraint set $\mathbb{U}$ is given in Fig. \ref{img:V3_Input_Constraint_Set}. Note that the state $x$ is unconstrained, but a limited set of angles $\alpha$ and $\beta$ is reachable as a consequence of the input constraints.
\begin{figure}[!t]
\vspace{5 pt}
\centering
\includegraphics{img/InputConstraintPlot.pdf}
\caption{A visualization of the input constraint set $\mathbb{U}$. The feasible inputs are indicated by the gray area.}
\label{img:V3_Input_Constraint_Set}
\end{figure}
\begin{figure*}[t]
\vspace{5 pt}
\centering
{
\resizebox{1\textwidth}{!}{
\input{img/ControlArchitecture.tikz}
}
\caption{The cascaded control architecture employed for the robotic arm. The first input from the MPC optimization problem is converted to the pressure set points via the inverse mapping $\xi^{\MINUS1}$. The set points for the actuator pressures are tracked in separate inner control loops at a higher rate.}
\label{fig:ControlArchitecture}
}
\end{figure*}
\subsection{MPC Optimization Problem}
Combining the previous building blocks, the MPC optimization problem is formulated as
\begin{equation}
\label{eq: V3 linear MPC optimization}
\begin{aligned}
\min_{x_i,u_i}||x_N\MINUS\bar{x}_{N}||_P&\\
+ \sum_{i=0}^{N \MINUS 1} ||x_i \MINUS \bar{x}_{i}||_Q
+ ||u_i \MINUS \bar{u}_{i}||_R &
+ ||u_i \MINUS u_{i\MINUS1}||_{R_d}\\
\text{s.t.} \quad x_i = {x}\idxk ,\, u_{i\MINUS1} = u\idxkmo,\,&\text{for $i$ = 0}\\
x_{i+1} = A x_{i} + B u_i + E \hat{d}\idxk,\,&\text{for $i = 0, \dots ,N\MINUS 1$}\\
u_i \in \mathbb{U},\, &\text{for $i = 0, \dots ,N\MINUS 1$\,,}
\end{aligned}
\end{equation}
where $||x||_M$ denotes the induced norm of a vector $x$ by the weighting matrix $M$.
At time step $k$, ${x}\idxk$ is the measurement of the current state, and
$\hat{d}\idxk$ denotes the disturbance estimate from the steady state Kalman filter. Because the state measurement ${x}\idxk$ has high accuracy and is updated with a higher frequency than the steady state Kalman filter, it has a lower latency and is directly used as the state feedback. The cost matrix $Q$ penalizes a deviation of the state from its target, the matrix $R$ similarly for the input and the cost matrix $P$ a deviation from the terminal target. The last term in the cost function including the cost matrix $R_d$ introduces a cost on the input rate. By penalizing the change between consecutive input values, the smoothness of the resulting input is improved, which limits control action at high frequencies that are beyond the bandwidth of the low-level pressure controllers.\par
The overall control architecture is shown in Fig. \ref{fig:ControlArchitecture}. The MPC optimization problem (\ref{eq: V3 linear MPC optimization}) is solved in a receding-horizon fashion at each time step. From the first control input $u_0$, the three pressure set points $p_{ASP}$, $p_{BSP}$, and $p_{CSP}$ are retrieved via the $\xi^{\MINUS1}$ mapping. The pressure set points $p_{ASP}$, $p_{BSP}$, and $p_{CSP}$ are then tracked in the inner control loops.
\subsection{Trajectory Tracking Experimental Results}
Besides the offset-free MPC discussed above, a standard MPC controller is implemented for comparison. The standard MPC is realized by disabling the disturbance compensation scheme. The prediction horizon length is set to $N = 50$ and the sampling time is set to $T_s\mkern-2mu = \mkern-2mu 0.02s$ for both controllers. The optimization problems are solved using the interior-point-based solver FORCES PRO \cite{FORCESPro}. The controllers are implemented in C++ and are executed on a laptop computer (Intel i7, quad-core, 1.8 GHz).\par
Both controllers are commanded to track the same angle trajectory consisting of
a ramp trajectory with magnitudes of both angles randomly generated in the interval $[\MINUS30\,\mathrm{deg},\,30\,\mathrm{deg}]$ and angular velocities randomly generated in the interval $[60 \,\mathrm{deg/s},\,300 \, \mathrm{deg/s}]$,
a soft step trajectory with step magnitudes of both angles randomly generated in the interval $[\MINUS30\,\mathrm{deg},\,30\,\mathrm{deg}]$ (a section of this trajectory is shown in Fig. \ref{img:soft step trj}) and
a sinusoidal trajectory with frequencies in the range between $0.5 \,\mathrm{Hz}$ and $3 \,\mathrm{Hz}$. The two controllers are tuned for the best individual tracking performance over the entire trajectory.\par
The Root Mean Square Errors (RMSE) of both $\alpha$- and $\beta$-directions are computed and then averaged. The offset-free MPC controller has an average RMSE of $1.8^{\circ}$, which is 35\% smaller than the average RMSE of $2.8^{\circ}$ from the standard MPC. Moreover, the offset-free MPC can track step set points without offset, while the standard MPC shows static offsets, as shown in Fig. \ref{img:soft step trj}.\par
We also implemented an MPC with integral action by including the integral of both angles in the state vector. Although this controller shows zero steady state offset, there is a trade off between a fast rise time (high cost on integral states) and little overshoot (small cost on integral states). The overall tracking performance was better than the standard MPC but worse than the offset-free MPC.
\begin{figure}[t]
\centering
\includegraphics{img/CE_SoftStep_trj_linear.pdf}
\caption{A section of the evaluation trajectory. The black dashed line denotes the set point trajectory (SP), the orange line the trajectory from the offset-free MPC (OSFMPC) and the blue line the trajectory from the standard MPC (MPC). The green dotted line in the upper plot denotes the estimated disturbance in the $\alpha$-direction of (\ref{V3 linear arm model}), and the one in the lower plot denotes the estimated disturbance in the $\beta$-direction. As these two disturbances act in the acceleration directions, they can be interpreted as mass-normalized torques acting on the robotic arm. Without the disturbance compensation scheme, the standard MPC has tracking offsets in the direction of the estimated disturbances. Note that the offsets are not linearly proportional to the disturbances due to the nonlinear relationship between the actuator pressures and the resulting actuation torques. For the tracking of step set points in both the $\alpha$- and $\beta$-directions, the standard MPC shows a maximum steady state offset of $5^{\circ}$ and the offset-free MPC a maximum offset below $0.5^{\circ}$.}
\label{img:soft step trj}
\end{figure}
\section{Soft Robotic Platform Overview}\label{sec:platform}
\begin{figure}[!t]
\vspace{5 pt}
\centering
\resizebox{0.5\textwidth}{!}{\input{img/angleConvention.tikz}}
\caption{The left hand plot shows the soft robotic arm and the different components it is composed of. A static link (1) is connected to a movable inflatable link (2) over a soft joint (3). The links are made of lightweight fabric material, and the soft joint is made from silicone rubber. The system is pneumatically actuated by soft bellow actuators (4) that are fabricated from soft material with high tensile strength. A 3D printed net (5) is attached to the movable link for the ball catching application. The middle plot shows the orientation parametrization using the extrinsic Euler angles $\alpha$ and $\beta$. The right hand plot shows the antagonistic configuration of the three symmetrically arranged bellow-type actuators.}
\label{fig:angleConvention}
\end{figure}
The spherical soft robotic arm originates from \cite{hofer2020design} and an improved version is used in \cite{zughaibi2020fast}. The individual components of the soft manipulator are explained in Fig. \ref{fig:angleConvention} (left). Connected to a flexible soft joint, the movable link can move on a section of a sphere parametrized by the extrinsic Euler angles $\alpha$ and $\beta$ as shown in Fig. \ref{fig:angleConvention} (middle). The three actuators have an antagonistic configuration as shown in Fig. \ref{fig:angleConvention} (right). Adjusting the air pressures of the actuators allows for control of their expansion and consequently the orientation of the movable link.\par
The angles $\alpha$, $\beta$ are retrieved from an infrared motion capture system running at 200 Hz. The pressure in each actuator is measured with a pressure transducer sampled at 200 Hz and is controlled by a separate proportional–integral–derivative controller (PID controller).
|
\section{Introduction}
Our problem in this paper is the \emph{online car-sharing problem}.
In car-sharing (not only for cars, but also for other resources like bikes and shuttle-buses), there are several service stations in the city, for instance in residential areas and downtown, at popular sightseeing spots, and so on.
Customers can make a request with a pick-up time and place and a drop-off time and place. The decision for accepting or rejecting a request should be made in an online fashion and we want to maximize the profit by accepting as many requests as
possible.
Relocation of
(unused) resources
is usually possible with a much smaller or even negligible costs. (It is seen occasionally that a truck is carrying bikes for this purpose.)
Theoretical studies of this problem have started rather recently and turned out
to be nontrivial even for two locations.
\medskip
\noindent
{\bf Model.}
We basically follow the problem
setting of previous studies by Luo et
al.~\cite{luo2018,luo2018car,luo2019car,LEX_kS2L} with main focus to
that of two locations and $k$ servers (i.e. cars).
The two location are denoted by 0 and 1
and $k (\geq 2)$ servers are initially located at location 0.
The travel time from 0 to 1 and 1 to 0 is the same, denoted by $t$.
The problem for $k$ servers and two locations is called the $k$S2L problem for short.
\begin{figure}[h]
\centering
\includegraphics[scale=0.4]{2L.PNG}
\caption{The car-sharing problem with two locations}
\label{2location}
\end{figure}
We denote the $i$-th request by $r_i = (\tilde{t_i}, t_i, p_i)$ which
is specified by the \emph{release time} or the \emph{booking time}
$\tilde{t_i}$, the \emph{start time} $t_i$, and the \emph{pick-up
location} $p_i \in \left\{0, 1\right\}$ (the \emph{drop-off location}
is $1-p_i$). If $r_i$ is accepted, the server
must pick up the customer at $p_i$ at time $t_i$ and drop off the
customer at $1-p_i$ at time $t_i + t$. Suppose for each $r_i$,
$t_i$ is an integer multiple of the travel time between location 0 and
1, i.e., $t_i = vt$ for some $v \in \mathbb{N}$.
We assume that $t_i
- \tilde{t_i}$ is equal to a fixed value $a$, where $a \ge t$ for all
requests. Without loss of generality, assume $a=t$.
Then we are only interested in
a discrete-time stage,
denoted by 0, 1, 2, $\ldots$.
Each server can only serve one request at a time. Serving a request yields
a fixed positive profit $y$.
A server used for a request with $p_i=0$ ($p_i=1$, resp.) cannot be used for a
request with $p_i=0$ ($p_i=1$, resp.) in the next stage.
We allow \emph{empty movement}s, i.e., a server can be moved from one location to
the other
without serving any request. An empty movement spends time $t$,
but takes no cost. The goal of the $k$S2L problem is to maximize the total profit
by serving a set of online requests.
Note that the performance of an online scheduling algorithm is typically evaluated by
\emph{competitive analysis}.
More precisely, the quality of an online algorithm is measured by the worst case ratio, called
\emph{competitive ratio} (CR), which is defined to be the fraction
of the profit of the offline optimal algorithm over that of the
online algorithm. The offline algorithm is aware of all requests in
advance.
If an
online algorithm is randomized, we use expected values for the output
of the online algorithm.
The online algorithm is called
$1/\delta$-competitive, if for any instances, the profit of the
algorithm is at least $\delta$ times the offline optimal profit.
So far the current model is exactly the same as $k$S2L-F
in~\cite{luo2018,luo2018car,luo2019car,LEX_kS2L}, although no
randomized cases were discussed in these papers.
\medskip
\noindent
{\bf Simultaneous Decision Model.}
Recall that in the $k$S2L model, two (or more) inputs with \emph{the
same booking time} still have an order.
Thus we can equivalently think that if $r_1, \ldots, r_d$ are requests
with booking time $t$, they are coming later than $t-1$ and before or
at $t$, one by one. Each of them should get a
decision (accept or reject)
immediately before the next request. The adversary can change
$r_i$ after looking at the response of the online algorithm against $r_1,
\ldots, r_{i-1}$.
This setting sounds reasonable as an online model, but the following
question seems also natural; what if requests with the same booking
time come exactly at the same time, the online player can see
all of them and can make decisions all together simultaneously at the booking moment
(equivalently the requests arrive in the same fashion as above but the
player can delay his/her online decisions
until the booking moment).
In this study
we also consider this
new model, denoted by $k$S2L-S. We further extend the
model, assuming that the number of requests with the same booking time
is at most $Rk$ for some constant $R$.
We call the generalized model $Rk$S2L-S.
Notice that having more than $k$ requests at the same location with the same booking time
never helps. Therefore, we only need to study the range $0\leq R\leq 2$
and $k$S2L-S means the special case that $R=2$. Our
algorithm for $Rk$S2L-S is \emph{adaptive} in the sense that it
automatically accommodates the value of $R$, which does not have to be
known in advance.
\medskip
\noindent
{\bf Prior work.}
The car-sharing problem has received a considerable amount of attention in recent years.
Luo et al.~\cite{luo2018} studied the problem with a single server and
two locations with both fixed booking time and variable
booking time. Here ``variable booking time'' essentially means that requests
with start time $t$ may come after requests with start time $t-1$.
They gave lower bounds on the CR for both fixed and variable
booking time under the positive
empty movement assumption.
Later,
Luo et al.~\cite{luo2018car} studied the car-sharing problem with two servers and two locations, i.e. 2S2L.
They considered
only the problem with fixed booking time and proposed an online algorithm which
can achieve a tight bound of two.
Luo et al.~\cite{LEX_kS2L} studied the car-sharing problem with $k$
servers and two locations, for both fixed booking time ($k$S2L-F) and variable
booking time ($k$S2L-V). Namely they showed the CR is
at least 1.5 for all $k$ and at most 1.5 for $k=3i$ for $k$S2L-F and
at least $5/3$ for all $k$ and at most $5/3$ for $k=5i$ for
$k$S2L-V.
Very recently, Luo et al.~\cite{luo2019car} studied the car-sharing
problem on a star network with $k$ servers as well as
two types of travel time: a unit travel time and an arbitrary travel time.
In comparison with the online setting,
B{\"o}hmov{\'a} et al.~\cite{bohmova2016} considered the offline car-sharing problem
in which all input requests are known in advance.
The objective is to minimize the number of vehicles
while satisfying all the requests.
The offline (i.e. static) problem can be solved in polynomial time.
Another closely related problem is the on-line dial-a-ride problem
(OLDARP),
where objects are transported between given points in a metric space.
The problem has been studied widely.
The goal is to minimize the total makespan~\cite{Ascheuer2000,Antje2017}
or the maximum flow time~\cite{Krumke2005}.
Christman et al.~\cite{christman2018} studied a variation of OLDARP where each request yields a revenue.
Yi et al.~\cite{Yi2005} studied another variation of OLDARP where each
request has a deadline, having a similar flavor as car sharing.
\medskip
\noindent
{\bf Our contribution.}
Recall that the tight CR of $k$S2L-F is 1.5
for $k=3i$~\cite{LEX_kS2L} and 2 for
$k=2$~\cite{luo2018car}, but open for other $k$'s.
In this paper, we show that it is $\frac{2k}{k+\lfloor k/3\rfloor}$
for all $k\geq 2$ and 1.5 for all $k\geq 2$ if randomization is
allowed. For $k$S2L-S that allows the online player to delay its
decision, it is shown that we can indeed take this advantage. Namely
the tight CR for $k$S2L-S is $\frac{2k}{k+\lfloor
k/2\rfloor}$ for all $k\geq 2$ and 4/3 for all $k\geq 2$ if
randomization is allowed.
For $Rk$S2L-S (we can assume $1\leq R\leq 2$
without loss of generality), it is shown that the CR is
strictly improved if $R<2$, namely the tight CR
(for randomized algorithms) is improved to $(2+R)/3$.
Note that if $R= 1.1$
(the number of requests at each stage exceeds $k$ by at most $10\%$),
the CR becomes at most $1.034$.
The basic idea of our algorithms is ``greedy'' and ``balanced''. Both notions
have already appeared in~\cite{LEX_kS2L}, but our implementation of
them is significantly different from theirs. More importantly, our
analysis is completely new; namely we use a simple mathematical
induction (augmented by two interesting parameters other than the
profit itself) while a classification of request types
was used in~\cite{LEX_kS2L}.
\begin{table}[h]
\normalsize
\begin{center}
\renewcommand{\arraystretch}{1.15}
\resizebox{\textwidth}{!}{
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
{\bf Problem} & \begin{tabular}[c]{@{}c@{}}{\bf Booking} \\ {\bf Time}\end{tabular} & {\bf Start Time} & \begin{tabular}[c]{@{}c@{}} {\bf The Cost Of} \\ {\bf Empty Move}\end{tabular} & \begin{tabular}[c]{@{}c@{}}{\bf Types of} \\ {\bf algorithms}\end{tabular} & \begin{tabular}[c]{@{}c@{}} {\bf Lower} \\ {\bf Bound}\end{tabular} & \begin{tabular}[c]{@{}c@{}} {\bf Upper} \\ {\bf Bound}\end{tabular} & {\bf Reference} \\ \hline
2S2L & Fixed & $t_i$ & $c=y$ & Deterministic & --- & 1 & MFCS'18 ~\cite{luo2018car} \\ \hline
2S2L & Fixed & $t_i$ & 0 & Deterministic & 2 & 2 & MFCS'18 ~\cite{luo2018car} \\ \hline
2S2L & Fixed & $t_i$ & $0<c<y$ & Deterministic & 2 & 2 & MFCS'18 ~\cite{luo2018car} \\ \hline
$k$S2L-F & Fixed & $t_i=vt $ for $v \in \mathbb{N}$ & 0 & Deterministic & 1.5 & $1.5 (k=3i, i\in \mathbb{N})$ & ISAAC'18 ~\cite{LEX_kS2L} \\ \hline
$k$S2L-F & Fixed &
$t_i=vt $ for $v \in \mathbb{N}$ & 0 &
Deterministic & $\frac{2k}{k
+ \lfloor k/3 \rfloor}$ & $\frac{2k}{k + \lfloor
k/3 \rfloor}$ & this paper \\ \hline
$k$S2L-F & Fixed &
$t_i=vt $ for $v \in \mathbb{N}$ & 0 &
Randomized & 1.5 & 1.5 & this paper \\ \hline
$k$S2L-V & Variant & $t_i=vt $ for $v \in \mathbb{N}$ & 0 & Deterministic & 1.5 & $1.5 (k=3i, i\in \mathbb{N})$ & ISAAC'18 ~\cite{LEX_kS2L} \\ \hline
$k$S2L-V & Variant & $t_i=vt $ for $v \in \mathbb{N}$ & 0 & Deterministic & 5/3 & $5/3 (k=5i, i\in \mathbb{N})$ & ISAAC'18 ~\cite{LEX_kS2L} \\ \hline
\hline
$k$S2L-S & Fixed
& $t_i=vt $ for $v \in \mathbb{N}$ & 0 &
Deterministic &
$\frac{2k}{k + \lfloor k/2\rfloor}$ & $\frac{2k}{k + \lfloor k/2\rfloor}$ & this paper \\ \hline
$k$S2L-S & Fixed & $t_i=vt $ for $v \in \mathbb{N}$ & 0 & Randomized & 4/3 & 4/3 & this paper
\\ \hline
$Rk$S2L-S ($1\le R \le 2$) & Fixed & $t_i=vt$ for $v \in \mathbb{N}$ & 0 & Randomized & (2+$R$)/3 & (2+$R$)/3 & this paper \\ \hline
\end{tabular}}
\end{center}
\caption{Overview of known and new results}
\label{table_summary}
\end{table}
The merit of our new analysis is demonstrated more clearly in $k$S2L-S
than in the original $k$S2L-F. Therefore we present the results for
$k$S2L-S first and then those for $k$S2L-F; the deterministic case in
Section 2 and the randomized case in Section 3. $Rk$S2L-S is discussed
in Section 4, where we introduce two magic numbers calculated from the
number of requests in each stage. Finally all matching lower bounds are
given in Section 5.
\section{Deterministic algorithms}
As mentioned in the previous section, we first discuss the basic GBA that
works for the $k$S2L-S model and then its accept/reject version
that works for the original $k$S2L-F model. The analysis for the
former will carry over to that of the latter pretty well. The
following table summarizes our notations which are used in the rest of
the paper.
\begin{table}[H]
\begin{tabular}{ll}
\textbf{Notation} & \\
$k$ &The number of total servers \\
$(0, 1)$: & Requests from location 0 to 1 \\
$(1, 0)$: & Requests from location 1 to 0
\\
$I\kern-.07em\ell_i$: & The number of (0,1)'s requested in stage $i$ with start time $i$ \\
$Ir_i$: & The number of (1,0)'s requested in stage $i$ with start time $i$ \\
$G\ell_i$: & The number of (0,1)'s accepted by the algorithm in stage $i$ \\
$Gr_i$: & The number of (1,0)'s accepted by the algorithm in stage $i$ \\
$G\kern-.1emf_i$: & The number of servers not used, i.e., $k-Gr_i-G\ell_i$ \\
$O\ell_i$: & The number of (0,1)'s accepted by OPT in stage $i$ \\
$Or_i$: & The number of (1,0)'s accepted by OPT in stage $i$ \\
$O\kern-.1emf_i$: & The number of servers not used, i.e., $k-Or_i-O\ell_i$ \\
\end{tabular}
\end{table}
\longdelete{
The idea of this greedy balanced algorithm can be illustrated as follows.
All servers are at location 0 in stage 0.
An arbitrary number of servers can be moved to location 1 if needed in stage 1.
We can think of all servers as ``floating'', available at either
location 0 or 1 in stage 1. Let $k=100$ and suppose requests in stage 1
are $(I\kern-.07em\ell_1, Ir_1)=(100,100)$. Thanks to the floating
servers, we can assign 100
servers freely to locations 0 and 1,
and hence accepted requests, $(G\ell_1,Gr_1)$, can be anything like
(100,0), (75,25) or (0,100).
However, if (100,0) is selected, then the server allocation in stage 2
is $\bigl[0,0,100\bigr]$.
Here we use notation $\bigl[x,y,z\bigr]$ to indicate $x$ and $z$ servers at locations
0 and 1, respectively and $y$ servers floating. So
the adversary would send
$(I\kern-.07em\ell_2, Ir_2)=(100,0)$ for stage 2, and no servers are
available for the online player. Since the almighty
adversary can select (0,100) in
stage 1, the CR would be 2.
Thus one can easily see that the best thing an algorithm can do is to accept
$(G\ell_1, Gr_1)=(50,50)$ in stage 1 to secure a CR of 1.5.
This is the notion of ``\emph{Balanced}''.
Note that requests denoted by
$(I\kern-.07em\ell_2, Ir_2)=(100,0)$ have booking time 1, but a
booking time of requests (and when they actually come)
is no longer important once we know $(I\kern-.07em\ell_i, Ir_i)$ and the server
allocation $\bigl[Gr_{i-1}, G\kern-.1emf_{i-1}=k-G\ell_{i-1}-Gr_{i-1}, G\ell_{i-1}\bigr]$
in stage $i$. Thus we will simply use ``requests for stage
$i$'' without mentioning their booking time.
}
Suppose there are $x$ and $x'$ servers at location 0 and $y$ and $y'$ servers at location 1 at time $i$,
where $x$ servers will serve (0,1)'s with start time $i$ but $x'$ servers are not used.
Similarly, $y$ servers will serve (1,0)'s but $y'$ servers not.
Then at time $i+1$ we can use $x$ servers for (1,0)'s and $y$ servers for (0,1)'s.
Furthermore, $x'+y'$ servers are available for requests of both directions since the requests with start time $i+1$ come at time $i$ and we can move $x'+y'$ servers to whichever locations as we like.
Now we introduce {\em stages} and say we have $x$~$(=G\ell_i)$ servers at location 1, $y$~$(=Gr_i)$ servers at location 0,
and $x'+y'$ $(=G\kern-.1emf_i)$ ``floating'' servers in stage $i+1$.
We also denote the {\em server allocation} at time $i+1$ as $\bigl[y,x'+y',x \bigr]$ using the new notation $\bigl[-,-,- \bigr]$.
Note that we need to have artificial $Gr_0, G\ell_0$ and $G\kern-.1emf_0$ for stage 1 whose values are 0, 0, and $k$, respectively.
Now the idea of this new greedy balanced algorithm can be illustrated as follows.
Let $k=100$ and suppose requests in stage 1 are $(I\kern-.07em\ell_1, Ir_1)=(100,100)$.
As described above, the server allocation in stage 1 is $\bigl[0,100,0\bigr]$, and hence accepted requests, $(G\ell_1,Gr_1)$, can be anything like (100,0), (75,25) or (0,100).
However, if (100,0) is selected, then the server allocation in stage 2 is $\bigl[0,0,100\bigr]$ and the adversary would send $(I\kern-.07em\ell_2, Ir_2)=(100,0)$ for stage 2,
by which no servers are available for the online player. Since the almighty adversary can select (0,100) in
stage 1
and he/she can accept all the requests in stage 2,
the CR would be 2.
Thus one can easily see that the best thing an algorithm can do is to accept $(G\ell_1, Gr_1)=(50,50)$ in stage 1 to secure a CR of 1.5.
This is the notion of ``\emph{Balanced}''.
Note that requests denoted by $(I\kern-.07em\ell_2, Ir_2)=(100,0)$ have booking time 1,
but a booking time of requests (and when they actually come)
is no longer important once we know $(I\kern-.07em\ell_i, Ir_i)$ and the server allocation $\bigl[Gr_{i-1}, G\kern-.1emf_{i-1}=k-G\ell_{i-1}-Gr_{i-1}, G\ell_{i-1}\bigr]$ in stage $i$.
Thus we will simply use ``requests for stage $i$'' without mentioning their booking time.
What if $(I\kern-.07em\ell_1, Ir_1)=(60,20)$? In this
case,
$(G\ell_1, Gr_1)=(60,20)$
is the best, i.e., the strategy is
a simple ``\emph{Greedy}'' one. If $(I\kern-.07em\ell_1, Ir_1)=(100,30)$, our
selection is
$(G\ell_1, Gr_1)=(70,30)$,
namely ``Greedy'' but as
``Balanced'' as possible. Algorithm~\ref{gba} realizes this idea
almost as it is and it will also be a core of all the subsequent
algorithms in this paper.
\begin{algorithm}
\caption{GBA($k$): Greedy Balanced Algorithm}
\label{gba}
\begin{algorithmic}[1]
\Require $I\kern-.07em\ell_i$ and $Ir_i$ are the numbers of (0,1)'s and (1,0)'s in stage
$i$, respectively.
An integer $k$ is the number of total servers.
The server allocation at the beginning of stage $i$ is
$\bigl[Gr_{i-1},G\kern-.1emf_{i-1}, G\ell_{i-1}\bigr]$, namely
$Gr_{i-1}$ and $G\ell_{i-1}$ servers at locations 0 and
1, respectively and $G\kern-.1emf_{i-1}=k-G\ell_{i-1}-Gr_{i-1}$ floating servers.
\Ensure
$G\ell_i$ and $Gr_i$ are the numbers of accepted (0,1)'s and
(1,0)'s, respectively.
\If{$Gr_{i-1}+G\kern-.1emf_{i-1} \leq \lfloor k/2 \rfloor$ or $I\kern-.07em\ell_i \leq \lfloor k/2 \rfloor$}
\State $G\ell_i \leftarrow {\rm min}\{I\kern-.07em\ell_i, Gr_{i-1}+G\kern-.1emf_{i-1}\}$; $Gr_i
\leftarrow{\rm min}
\{Ir_i,G\ell_{i-1}+G\kern-.1emf_{i-1}, k-G\ell_i \}$;
\Else
\If{$G\ell_{i-1}+G\kern-.1emf_{i-1} \leq \lfloor k/2 \rfloor$ or $Ir_i \leq \lfloor k/2 \rfloor$ }
\State $Gr_i\leftarrow{\rm min}\{Ir_i, G\ell_{i-1}+G\kern-.1emf_{i-1}\}$;\; $G\ell_i\leftarrow{\rm min} \{I\kern-.07em\ell_i, Gr_{i-1}+G\kern-.1emf_{i-1}, k-Gr_i \}$;
\Else
\State $Gr_i\leftarrow \lfloor k/2 \rfloor$;\; $G\ell_i\leftarrow\lceil k/2 \rceil$;
\EndIf
\EndIf
\State \Return $G\ell_i$ and $Gr_i$
\end{algorithmic}
\end{algorithm}
Recall that floating servers in stage $i$ are actually sit at location 0 or 1
at time $i-1$, say 30 ones at location 0 and 10 at location 1 among 40
floating ones. So if we need 20 floating servers in stage $i$ at
location 1, we need to move 10 servers from location 0 to 1 using the
duration from time $i-1$ to $i$. However, we do not describe this
empty movement in our algorithms since it is easily seen and its cost is free in our model.
Observe that the greedy part of Algorithm~\ref{gba} appears in lines 1
and 2 for (0,1)'s and in lines 4 and 5 for (1,0)'s. If
the condition in line 1 is met, then we accept (0,1)'s until we have
exhausted the servers or the (0,1)'s available in this stage. After
that we accept as many (1,0)'s as possible.
It works similarly for lines 4 and
5. If neither the condition in line 1 nor the one in line 4 is met, we
just split the requests almost evenly in line 7. The following theorem shows
that GBA achieves the optimal $\frac{4}{3}$-competitiveness for all even $k$ and approaches this value when $k$ is a large odd number.
In the rest of this paper, we use ALG to denote an online algorithm
and OPT an offline optimal scheduler in general.
\begin{figure}[t]
\centering
\includegraphics[scale=0.6]{GBA.png}
\caption{Server allocation in GBA}
\label{fig:gba}
\end{figure}
\begin{theorem}
\label{thm:gba}
GBA is a $1/\delta$-competitive algorithm for $k$S2L-S for any $k\geq 2$, where $\delta = \frac{k+ \lfloor k/2 \rfloor}{2k}$.
\end{theorem}
\begin{proof}
In order to prove the theorem, we consider the following six key values:
$$A_i=\sum_{j=1}^{i} (Gr_j+G\ell_j),
\quad B_i=\sum_{j=1}^{i} (Or_j+O\ell_j),$$
$$X_i=A_i + Gr_i + Gf_i, \quad Y_i=B_i+Or_i + Of_i,$$
$$U_i=A_i + G\ell_i + Gf_i, \quad V_i=B_i+O\ell_i + Of_i.$$
Our goal is to bound $A_i$ by $B_i$. To do so, it
is popular to use a potential function for competitive analysis, which is typically the difference between configurations of ALG and OPT.
In our present case, it may be the difference between server allocations of GBA and OPT.
It turns out, however, that this configuration difference or a similar one is unlikely to work since we still have a freedom for server selection which is not controlled by this difference strongly.
Instead we introduce four parameters, $X_i$, $Y_i$, $U_i$ and $V_i$, which play a key role in our proof.
Note that $X_i$ and $Y_i$ denote the total revenue of GBA and OPT respectively
for the first $i+1$ stages
assuming that the adversary tries to penalize the algorithm choice by
introducing
$k$ (0,1)'s in stage $i+1$;
the last two values, $U_i$ and $V_i$, denote the total revenue of GBA and OPT respectively
for the first $i+1$ stages
assuming that the adversary tries to penalize the algorithm choice by
introducing
$k$ (1,0)'s in stage $i+1$.
Intuitively, GBA
balances the accepted requests in both directions and guarantees that
the CR's in these two instances ($\frac{Y_i}{X_i}$ and
$\frac{V_i}{U_i}$, respectively) are not too large. It turns out that
taking care of these two extreme instances is sufficient to keep the
CR low for all instances.
In order to prove that the algorithm is
$1/\delta$-competitive,
we show that the set of the following inequalities (i) to (iii), denoted by $S(n)$,
$$\text{ (i) } A_n \geq \delta B_n, \quad \text{ (ii) } X_n \geq \delta Y_n, \quad \text{ (iii) } U_n \geq \delta V_n$$
hold for every $n$ by induction.
For the base case, $n=0$, we have $A_0=B_0=Gr_0=G\ell_0=Or_0=O\ell_0=0$ and
$Gf_0=Of_0=k$. Thus the three inequalities hold since $\delta \leq 1$.
Now the main part of the proof is proving $S(n)$ assuming, as stated in the
induction hypothesis, that $S(j)$
holds for all $0\leq j\leq n-1$. Note that we can rewrite $A_i$,
$B_i$, $X_i$ and so on as follows:
$$A_i=A_{i-1}+G\ell_i+Gr_i, \quad B_i=B_{i-1}+O\ell_i+Or_i,$$
$$X_i=A_{i-1}+ k + Gr_i, \quad Y_i=B_{i-1}+ k +Or_i,$$
$$U_i=A_{i-1}+k + G\ell_i, \quad V_i=B_{i-1}+k+O\ell_i.$$
Since $O\ell_i\leq {\rm min} \{ I\kern-.07em\ell_i, Or_{i-1}+Of_{i-1} \}$ and
$Or_i\leq {\rm min} \{ Ir_i, O\ell_{i-1}+Of_{i-1} \}$, the
following lemma is obvious, but will be used frequently.
\begin{lemma}
\label{optvalue}
$O\ell_i\leq I\kern-.07em\ell_i$, $O\ell_i \leq Or_{i-1}+Of_{i-1}$, $Or_i \leq Ir_i$, and $Or_i \leq O\ell_{i-1}+Of_{i-1}$.
\end{lemma}
Now we are ready to prove the theorem.
Suppose line 2 is executed. Then the following (L1) or (L2) holds for
the value of $G\ell_i$ and (R1), (R2) or (R3) for the value of
$Gr_i$. Similarly if line 5 is executed, (R1) or (R2) holds for
$Gr_i$ and (L1), (L2) or (L3) for $G\ell_i$.
\begin{eqnarray*}
&&\text{(L1)}\: G\ell_i=I\kern-.07em\ell_i,\:\: \text{(L2)}\: G\ell_i=Gr_{i-1}+G\kern-.1emf_{i-1},\: \text{(L3)}\:
G\ell_i =k-Gr_i (\geq \lfloor k/2 \rfloor),\\
&&\text{(R1)}\: Gr_i=Ir_i,\: \text{(R2)}\: Gr_i=G\ell_{i-1}+G\kern-.1emf_{i-1},\: \text{(R3)}\:
Gr_i =k-G\ell_i (\geq \lfloor k/2 \rfloor).
\end{eqnarray*}
Note that the condition ``$\geq \lfloor k/2 \rfloor$'' in (L3) and (R3)
comes from the conditions in lines 1 and 4, respectively. Now we consider stage $n$ and show that if
(L1), (L2) or (L3) holds, the induction (iii) holds, if (R1), (R2)
or (R3) hold, the induction (ii) holds and if any one of the nine
combinations \{(L1), (L2), (L3)\}$\times$\{(R1), (R2), (R3)\} holds,
(i) holds. First suppose (L1) holds. Then since $O\ell_n\leq
I\kern-.07em\ell_n$ by Lemma~\ref{optvalue}
\begin{eqnarray*}
&&U_n = A_{n-1}+ k + G\ell_n= A_{n-1}+ k + I\kern-.07em\ell_n,\\
&&V_n = B_{n-1}+k+O\ell_n\leq B_{n-1}+k+I\kern-.07em\ell_n.
\end{eqnarray*}
Thus (iii) is true by the induction hypothesis on (i). Similarly for
(L2), i.e., by Lemma~\ref{optvalue}
\begin{eqnarray*}
&&U_n = A_{n-1}+ k + G\ell_n= A_{n-1}+ k + Gr_{n-1}+G\kern-.1emf_{n-1}, \\
&&V_n = B_{n-1}+k+O\ell_n\leq B_{n-1}+k+ Or_{n-1}+Of_{n-1}.
\end{eqnarray*}
Thus (iii) is proved by the hypothesis on (ii). Finally for (L3),
\begin{eqnarray*}
&&U_n = A_{n-1}+ k + G\ell_n\geq A_{n-1}+ k + \lfloor k/2 \rfloor, \\
&&V_n = B_{n-1}+k+O\ell_n\leq B_{n-1}+k+ k,
\end{eqnarray*}
then use the hypothesis on (i) to claim (iii).
The proof that (R1) or (R2) or (R3) implies (ii) is similar and
omitted.
Finally we show that (i) follows from any combination. Observe
that (L1) and (R1) obviously implies (i)
since no algorithms accept more requests than requested.
All combinations including (L3)
or (R3) are also obvious since GBA accepts $k$
requests. Similarly for (L2) and (R2) when
$G\kern-.1emf_{i-1}=0$ (otherwise impossible). The remaining cases are (L1) and
(R2), and (L2) and (R1). For the former, by Lemma~\ref{optvalue}
\begin{eqnarray*}
&&A_n = A_{n-1}+ G\ell_n+Gr_n = A_{n-1}+I\kern-.07em\ell_n+G\ell_{n-1}+G\kern-.1emf_{n-1},\\
&&B_n = B_{n-1}+ O\ell_n+Or_n \leq B_{n-1}+I\kern-.07em\ell_n+O\ell_{n-1}+Of_{n-1},
\end{eqnarray*}
and we can use the hypothesis on (iii). (L2) and (R1) is similar and
omitted.
What remains is the case that line 7 is executed. Observe that line 7
gives us $G\ell_n\geq \lfloor k/2 \rfloor$, $Gr_n\geq \lfloor k/2
\rfloor$ and $G\ell_n+Gr_n=k$. We have already shown that the first one
implies (iii), the second one (ii) and
the third one means all the servers accept requests and is obviously
enough for (i).
Thus the theorem is proved.
\end{proof}
As seen in GBA and its analysis, a dangerous situation for the
online player is that ALG accepts too many requests of one direction
when it is possible. If ALG knows the total number of requests in each
direction in advance, we can avoid this situation rather easily. Now we discuss
$k$S2L-F, in which ALG does not know the total number of requests in
advance. A simple and apparent solution is to stop accepting requests
of one direction when its number gets to some value, even if more
requests of that direction are coming and could be accepted. In the next algorithm, ARGBA,
we set this value as $2k/3$. It then turns out, a little surprisingly,
that the analysis for Theorem~\ref{thm:gba} is also available for the
new algorithm almost as it is.
\longdelete{
\begin{algorithm}[H]
\caption{ARGBA($k$): Accept or reject GBA}
\label{argba}
\begin{algorithmic}[1]
\Require The server location is $\bigl[Gr_{i-1},
G\kern-.1emf_{i-1}=k-G\ell_{i-1}-Gr_{i-1},G\ell_{i-1}\bigr]$
at the
beginning of this stage $i$. $r_i^1, r_i^2,$ $\ldots, r_i^j,
\ldots, r_i^m$ are a sequence of requests, each of which is (0,1) or (1,0) in this stage. (The algorithm does not know the value of $m$ in advance.) $k$ is the total number of servers.
\Ensure immediate accept or reject for $r_i^j$. $G\ell_i$ and $Gr_i$
for the next server allocation.
\State $G\ell_i\leftarrow 0$; $Gr_i\leftarrow 0$, $A\ell^{0}\leftarrow 0$; $Ar^{0}\leftarrow 0$
\While{there are still requests in stage $i$, i.e. $j \leq m$}
\State Let $A\ell^{j-1}$ ($Ar^{j-1}$, resp.) be the number of
(0,1)'s ((1,0)'s, resp.) in $r_i^1,\ldots, r_i^{j-1}$;
\If{$r_i^j$ is (0,1)}
\If{$A\ell^{j-1}<Gr_{i-1}+G\kern-.1emf_{i-1}$ and $A\ell^{j-1}<2k/3$ and
$G\ell_i+Gr_i<k$}
\State accept $r_i^j$; $\:G\ell_i \leftarrow G\ell_i+1$; $j \leftarrow j+1$;
{\bf continue}
\EndIf
\Else $\:\:$ (namely, $r_i^j$ is (1,0))
\If{$Ar^{j-1}<G\ell_{i-1}+G\kern-.1emf_{i-1}$ and $Ar^{j-1}<2k/3$ and
$G\ell_i+Gr_i<k$}
\State accept $r_i^j$; $\:Gr_i \leftarrow Gr_i+1$; $j \leftarrow j+1$;
{\bf continue}
\EndIf
\EndIf
\State reject $r_i^j$; $j \leftarrow j+1$;
\EndWhile
\State \Return $G\ell_i$ and $Gr_i$
\end{algorithmic}
\end{algorithm}
}
\begin{algorithm}[h]
\caption{ARGBA($k$): Accept or reject GBA}
\label{argba}
\begin{algorithmic}[1]
\Require The server location is $\bigl[Gr_{i-1},
G\kern-.1emf_{i-1}=k-G\ell_{i-1}-Gr_{i-1},G\ell_{i-1}\bigr]$
at the
beginning of this stage $i$.
Requests are coming sequentially, each of which, $r$, is (0,1) or
(1,0). $k$ is the total number of servers.
\Ensure Immediate accept or reject for $r$. $G\ell_i$ and $Gr_i$
for the next server allocation.
\State $G\ell_i\leftarrow 0$; $Gr_i\leftarrow 0$,
$A\ell_i\leftarrow 0$; $Ar_i\leftarrow 0$ ($A\ell_i$
($Ar_i$, resp.) is the number of
(0,1)'s ((1,0)'s, resp.) received in this stage so far.)
\While{a new request $r$ comes}
\If{$r$ is (0,1)}
\State $A\ell_i \leftarrow A\ell_i+1$;
\If{$A\ell_i<Gr_{i-1}+G\kern-.1emf_{i-1}$ and $A\ell_i<2k/3$ and
$G\ell_i+Gr_i<k$}
\State accept $r$; $\:G\ell_i \leftarrow G\ell_i+1$;
\Else
\State reject $r$;
\EndIf
\Else $\:\:$ (namely, $r$ is (1,0))
\State $Ar_i \leftarrow Ar_i+1$;
\If{$Ar_i<G\ell_{i-1}+G\kern-.1emf_{i-1}$ and $Ar_i<2k/3$ and
$G\ell_i+Gr_i<k$}
\State accept $r$; $\:Gr_i \leftarrow Gr_i+1$;
\Else
\State reject $r$;
\EndIf
\EndIf
\EndWhile
\State \Return $G\ell_i$ and $Gr_i$
\end{algorithmic}
\end{algorithm}
\begin{theorem}
\label{thm:argba}
ARGBA is a $1/\delta$-competitive algorithm for $k$S2L-F for any
$k\geq 2$, where $\delta=\frac{k + \lfloor k/3 \rfloor}{2k}$.
\end{theorem}
\begin{proof}
Observe that once a (0,1) (similarly for (1,0)) is rejected,
then subsequent ones are all rejected. Let $X_i$ and $Y_i$ be the last
(0,1) and (1,0), respectively, that are accepted and $G\ell_i$
and $Gr_i$ be their numbers after lines 6 and 10,
respectively. Also, let $I\kern-.07em\ell_i$ and $Ir_i$ be the total numbers of
(0,1)'s and (1,0)'s in stage $i$, respectively (only used for analysis).
Suppose the last accepted (0,1) has gone through the
conditions in line 5.
Then, one can see that the next (0,1), if any, is blocked by one of these
conditions and thus
one of the following four conditions, (L1) through
(L4), is met. Similarly for (1,0)'s, one of (R1)
through (R4) is met.
\begin{eqnarray*}
&&\text{(L1)}\: G\ell_i=I\kern-.07em\ell_i,\:\: \text{(L2)}\: G\ell_i=Gr_{i-1}+G\kern-.1emf_{i-1},\: \text{(L3)}\:
G\ell_i =\lceil 2k/3 \rceil,\\
&&\qquad\qquad\qquad\qquad \text{(L4)}\: G\ell_i+Gr_i=k \text{ and }G\ell_i\geq
\lfloor k/3\rfloor \text{ and } Gr_i\geq \lfloor k/3\rfloor, \\
&&\text{(R1)}\: Gr_i=Ir_i,\: \text{(R2)}\: Gr_i=G\ell_{i-1}+G\kern-.1emf_{i-1},\: \text{(R3)}\:
Gr_i =\lceil 2k/3 \rceil,\\
&&\qquad\qquad\qquad\qquad \text{(R4)}\: G\ell_i+Gr_i=k \text{ and }G\ell_i\geq
\lfloor k/3\rfloor \text{ and } Gr_i\geq \lfloor k/3\rfloor.
\end{eqnarray*}
Note that the lower bound condition in (L4) and (R4) is correct since
otherwise the second condition in line 5 or 11 should have been met before.
Now consider stage $n$. In a way similar to the proof of
Theorem~\ref{thm:gba}, we can show that one of (L1) to (L4)
(with subscript $n$ replacing $i$) implies the induction (iii).
In fact, the reason is exactly the same for (L1) and (L2) as before.
Using
$G\ell_i\geq\lfloor k/3\rfloor$ in (L4) we have
\begin{eqnarray*}
&&U_n = A_{n-1}+ k + G\ell_n\geq A_{n-1}+ k + \lfloor k/3 \rfloor, \\
&&V_n = B_{n-1}+k+O\ell_n\leq B_{n-1}+k+ k,
\end{eqnarray*}
and then use the hypothesis on (i) to claim (iii) (recall that
our target CR is relaxed to $\frac{2k}{k + \lfloor k/3
\rfloor}$). (L3) obviously implies $G\ell_i\geq\lfloor k/3\rfloor$.
Also we can show, though omitted, that one of (R1) to (R4)
(with subscript $n$ replacing $i$) implies the induction (ii).
We can furthermore show that any one of the 16 combinations of (L1) to
(L4) and (R1) to (R4) implies (i): If a combination includes one of (L3), (L4), (R3) and (R4), then (i) is obvious since ARGBA accepts at least
$\lceil 2k/3 \rceil$ requests in this stage. So we only have to
consider the four combinations ((L1) or (L2), and (R1) or (R2)), and
these cases already appeared in the proof of Theorem~\ref{thm:gba},
which concludes the proof.
\end{proof}
\section{Randomized Algorithms}
Notice that the CR of GBA is 2 when $k=2$. The reason is simple, i.e., the
existence of the ceiling function,
namely if we can accept a fractional request, our CR would be 4/3.
Of course it is impossible to accept a request by one third, but
it is possible to accept that request with probability $1/3$, which
has the same effect as accepting it by one third in terms of an
expected number.
We define the following function for probabilistic rounding. Let $x$
be a (possibly fractional) non-negative number. Then define
$$\prrd(x)=\begin{cases}
\lceil x \rceil\quad \text{with probability}\: x-\lfloor x \rfloor\\
\lfloor x \rfloor\quad \text{with probability}\: 1-(x-\lfloor x \rfloor)
\end{cases}
$$
For instance, $\prrd(3.3)$ is 4 with probability 0.3 and 3 with
probability 0.7. $\prrd(3)$ is always 3. Note $E[\prrd(x)] =x$.
Now we are ready to introduce the Probabilistic GBA.
\begin{algorithm}
\caption{PrGBA($k$): Probabilistic GBA}
\label{prgba}
\begin{algorithmic}[1]
\State The same as Algorithm~\ref{gba} except that line 7 is
replaced as follows: $Gr_i \leftarrow \prrd(k/2)$ and $G\ell_i
\leftarrow k-Gr_i$
\end{algorithmic}
\end{algorithm}
\begin{theorem}
\label{thm:prgba}
PrGBA is a $4/3$-competitive algorithm for $k$S2L-S for any $k\geq 2$.
\end{theorem}
\begin{proof}
Observe the induction in the proof of the deterministic case. The base case is fine
with $\delta = 3/4$, and we can keep using this 3/4 unless line 7 is executed. Since the expected
value of $Gr_i$ and $G\ell_i$ are both $k/2$ when the modified line 7 is executed, we can remove the ceiling sign
from the description of the algorithm. Thus our new CR is $2k/(k + k/2) = 4/3$.
\end{proof}
\begin{algorithm}
\caption{PrARGBA($k$): Probabilistic ARGBA}
\label{prargba}
\begin{algorithmic}[1]
\State ARGBA is modified as follows: Suppose the three conditions in
line 5 are all met and $0<2k/3-A\ell^{j-1}<1$. Then accept
$r_i^j$ and increase $G\ell_i$ if $\prrd(2k/3-A\ell^{j-1})=1$,
otherwise reject it.
Similarly for line 11.
\end{algorithmic}
\end{algorithm}
\begin{theorem}
\label{thm:prargba}
PrARGBA is a 1.5-competitive algorithm for $k$S2L-F for any $k\geq 2$.
\end{theorem}
\begin{proof}
The same idea as the proof of Theorem~\ref{thm:prgba}. If this modified
part is executed, the expected value of $G\ell_i$ is $2k/3$ and hence the
expected value of $Gr_i$ is $k/3$ if $G\ell_i+Gr_i=k$. Thus the worst
case of deterministic ARGBA, $G\ell_i=\lceil 2k/3 \rceil$ and $Gr_i=\lfloor k/3
\rfloor$, can be avoided.
\end{proof}
\section{Adaptive GBA}
$k$S2L-S gives the online player the advantage of an advance
knowledge of the number of requests in the current stage.
GBA does exploit this advantage, but not fully.
Suppose $(I\kern-.07em\ell_1, Ir_1)=(50,100)$. GBA accepts the
same number, 50, of (0,1)'s and (1,0)'s in stage~1. Then the adversary sends
$(I\kern-.07em\ell_2, Ir_2)=(100,0)$, resulting in that only 50 (0,1)'s can
be accepted by GBA in stage 2, but 100 (0,1)'s by OPT which could accept 100 (1,0)'s
in stage~1. Thus the CR in these two steps is 4/3.
Now what about accepting roughly 28.57 (0,1)'s and 71.43
(1,0)'s in stage 1 (recall we can handle fractional numbers
due to randomized rounding)?
Then the best the adversary can do is to provide
$(I\kern-.07em\ell_2, Ir_2)=(100,0)$ or $(0,50)$, in both of which the CR is
$200/171.43\approx 150/128.57 \approx 1.17$, significantly better than 1.5
of GBA. This is the basic idea of our new algorithm, AGBA. These
key values 28.57 and 71.43 are denoted by
$\alpha_i$ and $\beta_i$, respectively.
The ultimate goal of AGBA is
to accept exactly $\alpha_i$ (0,1)'s and $\beta_i$ (1,0)'s while the
ultimate goal of GBA was to accept $k/2$ (0,1)'s and $k/2$ (1,0)'s. If
this goal is unachievable, to be figured out from the
values of $I\kern-.07em\ell_i$, $Ir_i$, $G\ell_{i-1}$, $Gr_{i-1}$ and $G\kern-.1emf_{i-1}$,
both algorithms simply turn greedy.
\begin{algorithm}
\caption{AGBA($k$): Adaptive GBA}
\label{agba}
\begin{algorithmic}[1]
\Require The same as GBA
\Ensure The same as GBA
\If{$R_i=(I\kern-.07em\ell_i+Ir_i)/k \geq 1$}
\State $\alpha_i=\frac{(1-R_i)k+3I\kern-.07em\ell_i}{2+R_i}$;\:\:
$\beta_i=\frac{(1-R_i)k+3Ir_i}{2+R_i}$;
\Else
\State $\alpha_i=I\kern-.07em\ell_i$; $\beta_i=Ir_i$;
\EndIf
\If{$Gr_{i-1}+G\kern-.1emf_{i-1} < \alpha_i$}
\State $G\ell_i\leftarrowGr_{i-1}+G\kern-.1emf_{i-1}$;\:\:$Gr_i\leftarrow{\rm min} \{ Ir_i,G\ell_{i-1} \}$;
\Else
\If{$G\ell_{i-1}+G\kern-.1emf_{i-1} < \beta_i$}
\State $G\ell_i\leftarrow{\rm min} \{ I\kern-.07em\ell_i,Gr_{i-1} \}$;\:\:$Gr_i\leftarrowG\ell_{i-1}+G\kern-.1emf_{i-1}$;
\Else
\State $G\ell_i\leftarrow\prrd(\alpha_i)$;
\State $Gr_i\leftarrow k-G\ell_i$;
\EndIf
\EndIf
\State \Return $G\ell_i$ and $Gr_i$
\end{algorithmic}
\end{algorithm}
Suppose $I\kern-.07em\ell_i+Ir_i$ is at most $kR_i$ for stage $i$. As
mentioned in a moment, we do not lose generality if $R_i$ is restricted to
$1\leq R_i\leq 2$, under which the CR of AGBA is bounded by $(2+R)/3$, where
$R$ is the maximum value of $R_i$ in the whole stages. We do
not know $R$ in advance and AGBA does not have to, either.
The competitive analysis is given by the following theorem. Note that
if the input in stage $i$ includes more than $k$ (0,1)'s, we can select
an arbitrary subset of size $k$ and similarly for (1,0)'s. This
guarantees that $R\leq 2$ and the case that $I\kern-.07em\ell_i+Ir_i <k$ is
covered by $R=1$. Thus the restriction of $R$, $1\leq R\leq 2$, makes
sense. Also, note that $\alpha_i + \beta_i = k$ whenever $R_i \geq 1$.
AGBA uses $R_i$ but not $R$. $R={\rm max}\{R_i\}$ by definition and so $\delta\leq \delta_i$ (where $\delta_i$ is the local value of $\delta$ in this stage, see its definition at Lemma 8) for all $i$.
\begin{theorem}
\label{thm:agba}
Suppose the number of requests is limited to at most $Rk$ in all
stages for some $R$ such that $1\leq R\leq 2$ and $Rk$ is an
integer.
Then AGBA solves $Rk$S2L-S and is
$1/\delta$-competitive, where $\delta=3/(2+R)$.
\end{theorem}
\begin{proof}
Because of the probabilistic rounding used in lines 12,
the expected
values of $G\ell_i$ and $Gr_i$ are $\alpha_i$ and $\beta_i$,
respectively, if this part is executed. We need two lemmas; the second
one illustrate a tricky nature of $\alpha_i$ and $\beta_i$.
\begin{lemma}
\label{alphabeta}
For any $R_i\geq 0$, $\alpha_i\leq I\kern-.07em\ell_i$ and $\beta_i\leq
Ir_i$. Also, $\alpha_i+\beta_i\leq k$.
\end{lemma}
\begin{proof}
A simple calculation of the formulas in line 2 is enough if the
condition $I\kern-.07em\ell_i+Ir_r\geq k$ is met. Otherwise it is also obvious by
line 4.
\end{proof}
\begin{lemma}
\label{trick}
Let $\delta_i={\rm min}\{1,3/(2+R_i)\}$. Then $k+\beta_i= \delta_i(k+Ir_i)$
and $k+\alpha_i= \delta_i(k+I\kern-.07em\ell_i)$.
\end{lemma}
\begin{proof}
If $R_i<1$ then $\delta_i=1, \alpha_i=I\kern-.07em\ell_i$ and $\beta_i=Ir_i$, so
the lemma is obviously true. Otherwise, a simple calculation:
$$k+\alpha_i=k+\frac{(1-R_i)k+3I\kern-.07em\ell_i}{2+R_i}=\frac{3k+3I\kern-.07em\ell_i}{2+R_i}=\delta_i(k+I\kern-.07em\ell_i).$$
Similarly for the other.
\end{proof}
The basic strategy of the proof is the same as
Theorem~\ref{thm:gba}. We consider the following four conditions for
the values of $G\ell_i$ and $Gr_i$. Suppose line 7 is executed. Then
$G\ell_i$ satisfies (L2) and $Gr_i$ satisfies (R1). Similarly if line 10 is executed, (L1) and (R2) hold.
\begin{eqnarray*}
&&\text{(L1)}\: G\ell_i={\rm min}\{I\kern-.07em\ell_i,Gr_{i-1}\},\: \text{(L2)}\: G\ell_i=Gr_{i-1}+G\kern-.1emf_{i-1}, \\
&&\text{(R1)}\: Gr_i={\rm min}\{Ir_i,G\ell_{i-1}\},\: \text{(R2)}\: Gr_i=G\ell_{i-1}+G\kern-.1emf_{i-1}.
\end{eqnarray*}
Now consider stage $n$. It is shown that (L1) or (L2) implies (iii) of the induction. For
(L2), the analysis is the same as before and omitted.
For (L1), using Lemma~\ref{optvalue}
for $V_n$, we have
\begin{eqnarray*}
&&U_n = A_{n-1}+k+G\ell_n = A_{n-1}+ k+{\rm min}\{I\kern-.07em\ell_n,Gr_{n-1}\},\\
&&V_n =B_{n-1}+k+O\ell_n\leq B_{n-1}+k+I\kern-.07em\ell_n.
\end{eqnarray*}
If ${\rm min}\{I\kern-.07em\ell_n,Gr_{n-1}\}=I\kern-.07em\ell_n$, then we are done using the
hypothesis (i). Otherwise recall that the condition of line 9,
$G\ell_{n-1}+Gr_{n-1} < \beta_n$, is met. So we have
$Gr_{n-1}=k-(G\ell_{n-1}+G\kern-.1emf_{n-1})\geq k-\beta_n\geq \alpha_n$ (by
Lemma\ref{alphabeta} for the last inequality)
and thus we can use Lemma~\ref{trick} and the hypothesis on (i)
to claim (iii).
The proof that (R1) or (R2) implies
(ii) is very similar and omitted.
Next we prove that each of the four
combinations implies (i). (L1) and (R1), and (L2) and (R2) are obvious
since AGBA is as efficient as OPT or accepts $k$ requests (recall
${\rm min}\{I\kern-.07em\ell_n,Gr_{n-1}\}\geq \alpha_n$ and
${\rm min}\{Ir_n,G\ell_{n-1}\}\geq \beta_n$ mentioned above and by Lemma~\ref{alphabeta}). For (L1) and (R2), using
Lemma~\ref{optvalue} and $\alpha_n\leq I\kern-.07em\ell_n$, we have
\begin{eqnarray*}
&&A_n = A_{n-1}+ G\ell_n+Gr_n = A_{n-1}+{\rm min}\{I\kern-.07em\ell_n, Gr_{n-1}\}+G\ell_{n-1}+G\kern-.1emf_{n-1},\\
&&B_n =B_{n-1}+O\ell_n+Or_n \leq B_{n-1}+I\kern-.07em\ell_n + O\ell_{n-1}+Of_{n-1}.
\end{eqnarray*}
Thus the hypothesis on (iii) or $Gr_{n-1}+G\ell_{n-1}+G\kern-.1emf_{n-1}=k$
implies that $A_n\geq \delta B_n$. (L2)
and (R1) are similar.
The remaining case is the one that lines 12 and 13 are executed.
If $G\ell_n=\alpha_n$, we have $U_n=A_{n-1}+k+\alpha_n$ and thus we can
use Lemma~\ref{trick} to claim (iii) as shown above. Similarly for
$Gr_n$ and (ii).
(i) is obvious since
$G\ell_i+Gr_i=k$, completing the proof.
\end{proof}
\section{CR Lower Bounds}
The CR's given so far are all tight. In this section we prove matching
lower bounds for $k$S2L-S, for $k$S2L-F, for $k$S2L-F with randomization,
and for $Rk$S2L-S with randomization (including for
$k$S2L-S with randomization as a special case).
\begin{theorem}
\label{lb:gba}
No deterministic online algorithms for the $k$S2L-S problem can achieve a CR of less than $\frac{2k}{k + \lfloor k/2 \rfloor}$.
\end{theorem}
\begin{proof}
Let $\mathcal{A}$ be any deterministic
algorithm. The adversary requests $k$ (0,1)'s and
$k$ (1,0)'s in stage 1.
$\mathcal{A}$ accepts $k_\ell$ (0,1)'s and $k_r$ (1,0)'s.
If $k_\ell \leq \lfloor k/2 \rfloor$, then the adversary requests
$k$ (1,0)'s (and zero (0,1)'s) in stage 2. The profit of $\mathcal{A}$
is $k_\ell + k_r$ in stage 1, and at most
$(k-k_\ell-k_r) + k_\ell$
in stage 2.
Therefore, the total profit of $\mathcal{A}$
is at most
$k+k_\ell \leq k+\lfloor k/2 \rfloor.$
The profit of OPT is $2k$, and the theorem is proved.
If $k_\ell > \lfloor k/2 \rfloor$, then
$k_r \leq \lfloor k/2 \rfloor$.
Now the adversary requests $k$
(0,1)'s in stage 2. The profit of $\mathcal{A}$ and OPT are exactly the same as above and we may omit
the rest of calculation. Thus the bound is tight.
\end{proof}
\begin{theorem}
\label{lb:argba}
No deterministic online algorithms for the $k$S2L-F problem can achieve a CR of less than $\frac{2k}{k + \lfloor k/3 \rfloor}$.
\end{theorem}
\begin{proof}
Let $\mathcal{A}$ be any deterministic algorithm. The basic idea is
similar to \cite{LEX_kS2L}. The adversary gives $k$ (0,1)'s (sequentially) for stage 1.
If $\mathcal{A}$ accepts at most
$\lfloor 2k/3\rfloor$ ones, then the adversary stops his/her requests
and the game ends. Thus the CR is at least $\frac{k}{\lfloor
2k/3\rfloor}$. Otherwise, if $\mathcal{A}$ accepts $\lceil 2k/3\rceil$
or more, then the adversary gives another $k$ (1,0)'s for stage
1 and $k$ (0,1)'s for stage 2. Since $\mathcal{A}$ has accepted at
least $\lceil 2k/3\rceil$ (0,1)'s in stage 1, $\mathcal{A}$ cannot use
those servers for the (0,1)'s for stage 2. Hence $\mathcal{A}$
can accept at most $k$ requests in stage 1 and at most $k-\lceil 2k/3\rceil$
requests in stage 2, meaning at most $2k-\lceil 2k/3\rceil=k+\lfloor
k/3 \rfloor$ requests in total. OPT can accept $k$ (1,0)'s in stage 1
and $k$ (0,1)'s in stage 2, i.e., $2k$ in total. Thus the CR is at
least $\frac{2k}{k+\lfloor k/3 \rfloor}$. Since $\frac{2k}{k+\lfloor
k/3 \rfloor} \leq\frac{k}{\lfloor 2k/3\rfloor}$ for all $k$ (this can be
verified by checking for $k=3j$, $k=3j+1$ and $k=3j+2$), the theorem is
proved.
\end{proof}
\begin{theorem}
\label{lb:prargba}
No randomized online algorithms for the $k$S2L-F problem can achieve a
CR of less than 1.5.
\end{theorem}
\begin{proof}
The proof is almost the same as that of Theorem~\ref{lb:argba}.
Since the adversary has the full information (other than random values) of ALG,
he/she can compute the expected value of ALG's output. So what we have
to do is just removing the floor and ceiling signs from the previous
proof and considering the resulting numbers as expected values.
The proof is complete since the previous deterministic OPT has a profit of at least $2k$.
\end{proof}
\begin{theorem}
\label{lb:agba}
No randomized online algorithms for the $Rk$S2L-S problem can
achieve a CR of less than $\frac{2+R}{3}$.
\end{theorem}
\begin{proof}
Let $\mathcal{A}$ be any randomized
algorithm. The adversary requests $\lfloor Rk/2\rfloor$ (0,1)'s and
$\lceil Rk/2\rceil$ (1,0)'s in
stage 1 ($\lfloor Rk/2\rfloor + \lceil Rk/2\rceil=Rk$ by the integrality
condition).
$\mathcal{A}$ accepts $k\ell$ (0,1)'s and $kr$ (1,0)'s.
Let $\alpha= \frac{(1-R)k+3\lfloor
\frac{Rk}{2}\rfloor}{2+R}$ and if $E[k\ell]\leq \alpha$ (note that the adversary has the full information of
$\mathcal{A}$, so it can compute $E[k\ell]$), then the adversary requests
$k$ (1,0)'s (and zero (0,1)'s) in stage 2. The profit of $\mathcal{A}$
is at most
$$k+E[k\ell]\leq k+\alpha=k+ \frac{(1-R)k+3\lfloor
\frac{Rk}{2}\rfloor}{2+R}=\frac{3(k+\lfloor
\frac{Rk}{2}\rfloor)}{2+R}.$$
The profit of OPT is $k+\lfloor Rk/2\rfloor$, and the theorem is proved.
If $E[k\ell]> \alpha$ then let $\beta=\frac{(1-R)k+3\lceil
\frac{Rk}{2}\rceil}{2+R}$, and it is easy to see that $\alpha+\beta=k$.
Hence we have $E[kr]\leq \beta$ because
$E[k\ell]+E[kr]=E[k\ell+kr]\leq k$. Now the adversary requests $k$
(0,1)'s. The profit of
$\mathcal{A}$ and OPT are exactly the same as above by replacing $\lfloor
\frac{Rk}{2}\rfloor$ with $ \lceil\frac{Rk}{2}\rceil$.
We may omit
the rest of calculation.
Note that although the two input instances provide a tight lower bound for the competitive ratio, applying Yao's Minimax theorem on any probability distribution over these two input instances does not provide the same tight bound.
\end{proof}
\section{Concluding Remarks}
\longdelete{
Our analysis fully exploits the notion of ``floating servers,'' and
the arguments of the mathematical induction using supplementary parameters $X_i$
and $U_i$ (and their OPT counterparts), which seems quite powerful and available for future studies of the problem. The current car sharing
model is still elementary, and
it would be worthwhile to investigate its many extensions.
In particular, relaxation
of the rental period condition should be challenging and important for
practical applications.
}
We have presented a different greedy and balanced algorithm with a new analysis which fully exploits the notion of “floating servers,” and the arguments of the mathematical
induction using supplementary parameters $X_i$ and $U_i$ (and their OPT counterparts). We believe that the analysis technique is powerful for future studies of many extensions of the car sharing problem, which would be worthwhile to investigate for practical purposes.
In particular, relaxation of the rental period condition should be challenging and important.
\newpage
\nocite{*
|
\subsection{Longitudinal Registration}
Due to the nature of chest CT scans, the initial volumes of data between different time points are highly misaligned. Aspects like patient positioning, variations of the imaging parameters or devices, different phases in the breathing cycle, and the disease progression are the main reasons. This misalignment cannot be described as a linear transformation composed of translation and rotation between the time points of data and can only be expressed through non-linear transformations. The misalignment can make the network incapable of using the longitudinal information present between different time points of data.
As a solution for this problem, we utilize a deformable registration algorithm~\cite{lowekamp2013design} where a BSpline Transform is defined using a sparse set of grid points overlaid onto the fixed domain of the image domain to deform it. Using this algorithm we register the reference scan lung mask $M_0$ to the follow up scan lung mask $M_1$ and this transform function is defined by $R_{M_0 \rightarrow M_1}(\cdot)$. Based on this function, the transformations are applied to the respective CT-scans as ${X}^{reg}_{0}=R_{M_0 \rightarrow M_1}(X_0)$. Using the lung masks, we avoid registration errors due to the pathological changes in the lung parenchyma while compensating for positioning, breathing phase, and acquisition-related differences.
\subsection{Longitudinal Segmentation}
Due to the challenge of training a 3D model with a limited number of training data, 2.5D approaches~\cite{aslani2019multi,zhang2019multiple,roy2019quicknat,denner2020spatio} have shown state-of-the-art results on various medical segmentation problems. For the 3D approach, it is challenging to directly process a full 3D volume by using current GPU memory limitations~\cite{roy2019quicknat}, which forces people to operate on 3D patches~\cite{wachinger2018deepnat,hashemi2018asymmetric}. However, patch-wise training limits the overall spatial context for accurate semantic segmentation.
In this study, we adopt the 2.5D approach of~\cite{denner2020spatio} with a fully convolutional (FC) DenseNet~\cite{jegou2017one} as a baseline 2D segmentation model. The FC DenseNet consists of a downsampling path (Encoder) with 5 Transitions Down blocks, each with 4 layers and an upsampling path (Decoder) with 5 Transitions Up blocks, each with 4 layers. The model is trained for all three views (coronal, sagittal, and axial view). At test time, for each given voxel, the segmentation is conducted on all three orthogonal views. Afterward, the predicted probability of a given voxel is averaged among views to assign a final predicted probability.
We extend the aforementioned 2.5D segmentation to deal with longitudinal information by modifying its architecture. To capture subtle spatio-temporal cues, we concatenate two registered scans from two different time-points as input for the longitudinal segmentation network: $[X^{reg}_0,X_1]$ for segmenting pathologies on $X_1$ or $[X_1, X^{reg}_0]$ for $X_0$. This enables the segmentation network to capture temporal changes as shown in Figure~\ref{fig:methods}. For subjects who have more than two scans, we select consecutive scans for the segmentation in each step.
\subsection{Progression Analysis}
To monitor the progression of the COVID-19 infection and the response to therapy, we extract the segmented pathologies from consecutive longitudinal CT scans and quantify the volume differences between them.
Our approach is capable of quantifying all combinations of changes between different pathologies and healthy lung parenchyma.
In this work’s scope, we define two classes as consolidation and non-consolidation since consolidation has shown to be a robust biomarker for COVID-19~\cite{li2020clinical}.
The progression of consolidation is computed by subtracting two registered longitudinal CT-scans. The resulting residual voxel values could be either positive or negative. Positive voxels indicate that a healthy, GGO or pleural effusion region progresses to consolidation (Progression), and negative voxels suggest that an area recovers from the severe infection of consolidation (Recovery).
\subsection{Training with Progression Information}
For the optimization of our model, we define an overall loss function combining a segmentation loss $\mathcal{L}_{seg}$ and a progression loss $\mathcal{L}_{prog}$. Note that as shown in Fig.~\ref{fig:methods}, the outputs of our model consist of the segmentation masks for reference (t=0) and follow-up (t=1) scans and the subtracted volume between reference and follow-up scans for presenting progression of the COVID-19 infection. The overall loss is defined as:
$\mathcal{L} = \mathcal{L}_{seg} + \mathcal{L}_{prog}.$
The segmentation loss is defined as $\mathcal{L}_{seg}=L_{MSE}(Y_0,\hat{Y_0})+L_{MSE}(Y_1,\hat{Y_1})$ where $L_{MSE}$ denotes a mean squared error loss \cite{zhang2019multiple,denner2020spatio}. In our evaluation, this metric yielded better results in comparison to the dice score. $Y_0$ and $\hat{Y_0}$ denote a ground truth pathology segmentation map and a predicted segmentation mask for t=0 scan, respectively. $Y_1$ and $\hat{Y_1}$ denote a ground truth pathology segmentation map and a predicted segmentation mask for t=1 scan.
The progression loss is defined as $\mathcal{L}_{prog}=\mathcal{L}_{MSE}(Y^{con}_1-Y^{con}_0,\hat{Y}^{con}_1-\hat{Y}^{con}_0)$ where $Y^{con}$ denotes a ground truth consolidation map and $\hat{Y}^{con}$ denotes a predicted consolidation map. In those binary maps, consolidation is mapped to 1 and non-consolidation to 0. As explained above, the progression map is calculated by $Y^{con}_1-Y^{con}_0$. Note that, the progression loss does not minimize the distance between the two segmentations. Instead it explicitly uses the structural changes of pathologies through times as cues for modeling the optimization. In other words, if there are large changes over time, the progression loss encourages the model to predict those large changes also in the segmentation.
\subsection{Dataset}
To our knowledge there is no publicly available longitudinal CT dataset for COVID-19. Accordingly, to evaluate the proposed method, we used an in-house clinical dataset which consists of longitudinal low-dose CT-scans from 38 patients (64$\pm$18 years old, 16 females, 22 males) with positive PCR from the first COVID-19 wave (March-June 2020). 28 patients had two scans and 10 had three. The CTs were separated 17$\pm$10 days (1-43 days) and were taken at admission and during the hospital stay (33$\pm$21 days, 0-71 days). 8 patients of the 38 died; 30 recovered from COVID-19, 20 of them needing intensive care. All scans were performed in-house using two different CT devices (IQon Spectral CT and iCT 256, Philips, Hamburg, Germany) with the same parameters (X-ray current 140-210 mA, voltage 120 kV peak, slice thickness 0.9mm, no contrast media) and covered the complete lung. The data was collected retrospectively with the approval of the institutional review board of our institution (ethics approval 111/20 S-KH).
The dataset was annotated at a voxel-level by a single expert radiologist (5 years experience), generating lung masks (lung parenchyma vs. other tissues) and pathology masks including four classes: healthy lung (HL), GGO, consolidation (CONS), and pleural effusion (PLEFF). For segmentation, the radiologist used the software ImFusion Labels (ImFusion, Munich, Germany).
The dataset was split into a training set of 16 patients (37 volumes) and an independent test set of 22 patients (49 volumes). From the training set, 12 patient scans are used for model training and 4 patient scans are used for validation. The model is finally evaluated on the unseen test set.
\subsection{Implementation Details}
The raw CT volumes highly vary in intensity range, size, and alignment. Therefore, we perform the following pre-processing steps on the raw volumes to enable effective use of the longitudinal data:
\noindent \textbf{Cropping.}
Since different body regions can be presented between time points and patients, we crop the volumes to the lung regions, using the manually-annotated lung masks.
\noindent \textbf{Clipping and Normalization.} To alleviate different intensity ranges among CT-scans, intensity values outside the range $(-1024, 600)$ are clipped and then min-max normalization is performed on each volume.
\noindent \textbf{Resizing.} After Cropping, resulting volumes vary in size in all three dimensions, ranging from 100 pixels to 580 pixels. Therefore all volumes are resized to a fixed size of 300$\times$300$\times$300 with 300 being the median among the cropped-volume sizes.
\noindent \textbf{Slicing and Removing Empty Slices.} Finally, the volumes are sliced in each of the three dimensions to 300 slices, generating sagittal, coronal, and axial views of the lung. Slices that have a voxel-value variation smaller than $0.001\%$ between their maximum and the minimum value are considered empty and are removed.
\noindent \textbf{Model Training.} For training, Adam optimizer~\cite{kingma2014adam} with a learning rate of 0.0001 and a decay rate of 0.1 for every 50 steps was used. The model was trained over 100 epochs with early stopping if no decrease in the validation loss was computed for 5 epochs.
Our method was implemented in PyTorch 1.4 and our models were trained on an NVIDIA Titan V 12GB GPU using Polyaxon\footnote[2]{https://polyaxon.com/}. The source code is publicly available\footnote[3]{https://github.com/lilygoli/longitudinalCOVID}. Our longitudinal model had 1.3752M parameters in comparison to its static counterpart 1.3748M.
\subsubsection{Effectiveness of Longitudinal Segmentation}
First, we conduct comparative experiments to verify the effectiveness of our longitudinal segmentation method. In Table~\ref{Table_Comparison_Main}, we compare our method with a static network~\cite{zhang2019multiple} based on FC-DenseNet~\cite{jegou2017one} and with a longitudinal network without progression loss. This simpler longitudinal network has the same architecture as our proposed one, i.e., it concatenates longitudinal CT scans as an input for the segmentation model, but it is trained using only the segmentation loss.
As shown in Table~\ref{Table_Comparison_Main}, both longitudinal networks achieved a higher Dice Similarity Coefficient (DSC) than the static network. The difference was statistically significant for all classes (\textit{p}$<$0.05 by paired t-test~\cite{altman1990practical}). This implies that using longitudinal information from the reference CT scan is informative to segment pathology on the target CT scan. In our longitudinal network with progression loss, the DSC was further improved for HL, CONS, and GGO. But for PLEFF, the performance slightly decreased. This can be attributed to the fact that the progression loss encourages the model to focus on CONS rather than on PLEFF. Additionally, PLEFF is a challenging, under-represented class in our dataset (only 2.17\% voxels).
\subsubsection{Effect of Longitudinal Registration.}
\input{Figures/Table2}
To showcase the importance of registration among the longitudinal scans, we report results with and without deformable registration. As seen in Table~\ref{Table_Analysis_Registration} the performance after registration substantially improves across the board with the increase ranging from 0.07 to 0.10.
\subsubsection{Importance of Temporal Information in Longitudinal Network.}
In this experiment, we highlight the importance of the longitudinal scans for the performance of the model. Specifically, we concatenate two duplicates of the reference scan instead of the reference and follow-up scan as input to our model ('static input'). As shown in Table~\ref{Table_Analysis_Registration}, the longitudinal input (proposed method) outperforms the static input for all classes.
\subsubsection{Progression Analysis.}
Finally, we evaluate our method for progression analysis by comparing with a static network~\cite{pu2021automated}, a longitudinal network with multi-view approach~\cite{birenbaum2017multi} and our model trained without the progression loss.
As shown in Table~\ref{Table_Progression}, our longitudinal architecture has a 3.4\% increase compared to the static network. The proposed model with the progression loss has a 4.8\% increase with respect to static network. Note that the model using the progression loss significantly outperformed the static network~\cite{pu2021automated} and the longitudinal network with multi-view approach~\cite{birenbaum2017multi} (\textit{p}$<$0.05). Moreover, the progression loss had statistically significant improvement for the recovery and average progression prediction compared to the longitudinal model without the progression loss.
\input{Figures/Table3}
Figure~\ref{fig:qualitative} showcases qualitative results of the segmentation and progression analysis of our method for all 3 different views. As shown in Figure~\ref{fig:qualitative}, our method successfully provides segmentation and progression maps for both reference and follow-up scans across views and patients. Even the under-represented class of PLEFF is successfully segmented. Regarding the progression, the fine-grained regions of recovered and progressed consolidation are also correctly identified.
\begin{figure}[t]
\centering
\includegraphics[width=0.98\textwidth]{Figures/Figure2.pdf}
\caption{Qualitative results of our method for 3 patients from different views. For the segmentation maps, blue, green, red and yellow denote healthy lung, consolidation, ground-glass opacity, and pleural effusion, respectively. For the progression map, red denotes Progression and the green Recovery.}
\label{fig:qualitative}
\end{figure}
\section{Introduction}
\input{1.Introduction}
\section{Methodology}
\input{2.Methodology}
\section{Experiment Setup}
\input{3.Experiments}
\section{Results and Discussion}
\input{4.Results}
\section{Conclusion}
\input{5.Conclusion}
\section*{Acknowledgements}
This paper was funded by the Bavarian Research Foundation (BFS) under grant agreement AZ-1429-20C. We would like to thank NVIDIA for the GPU donation.
\label{sec:acknowledgements}
\bibliographystyle{splncs04}
|
\section{Introduction}
The quest for ultimate limits serves as unfathomable source of inspiration for theoretical as well as experimental research.
The "ultimate limits" usually do not represent a final goal but rather show the direction worth pursuing. It is intriguing to note that coherence (and coherent states) provides one of the keystones for quantum optics but the role of (partial) coherence in metrology has not been fully exploited yet.
Recent developments have been inspired by reconsideration of well-known concepts of classical optics such as Rayleigh \cite{Rayleigh} or Sparrow \cite{Sparrow} resolution criterion from the point of view of quantum estimation theory \cite{Tsang-review} and particularly the concept of Fisher information \cite{Helstrom}. The problem of classical resolution can be paraphrased as the question of how well can we distinguish two
bright spots, or more generally two elementary signals, in which is coded the information about the parameter of interest.
The celebrated classical resolution criteria suggest that this can be done up to the
distances when two blurred spots start to overlap.
{ This rule of thumb can be justified by the analysis of the intensity pattern, since the Fisher information vanishes quadratically for separations approaching zero. }
As shown by Tsang and coworkers \cite{Tsang} and demonstrated experimentally on different platforms
\cite{Steinberg,Lvovsky,Tang,Opt, TF}, this behaviour can be avoided if quantum estimation theory is
adopted for the estimation of geometrical parameters, namely the transversal
separation and the centroid positions of two equally bright spots with known
intensities.
In this context, the Fisher information refers to quantum measurements and
converts to the Quantum Fisher Information (QFI) upon optimising over all
possible measurement schemes \cite{Helstrom, Liu20}.
The role of coherence in quantum estimation problems has been discussed in recent publications \cite{Saleh,Comment_Tsang,Reply_Larson} without leading to a consensus.
More light was shed on this problem in the paper \cite{coherence} showing that if the separation between two sources is coded into the coherent superposition state, the QFI may become unlimited. Such an effect, however, is always associated with the measurement on the dark channel with vanishing signal. In this Letter we extend the theory to partially coherent states addressing several issues.
At first we show that there are non-equivalent models for the description of partially coherent states built on the mixing of normalised states or alternatively on the mixing of coherent amplitudes, when any fusion of both those generic models is allowed as well. When the
strength of the signal is taken into account, none of the models presents a metrological advantage since precision of any of those partial superpositions is limited by the precision of incoherent mixtures though the Quantum Fisher Information itself may become infinite.
{ In this sense our conclusion can be seen as an optical illustration of the profound and more general observation that probabilistic metrology or any post-selection scheme can never improve quantum limits on estimation of a single parameter, though some estimates may appear occasionally excellent \cite{Caves2014}. In addition to this }
we show, how the limiting precision of incoherent mixtures can {be overcome} by special state preparation.
This will be achieved by careful tailoring of the degree of partial coherence depending on the separation. This scenario surely lacks practical relevance for estimating an unknown separation, { but this is a common issue for all model-dependent schemes, when the structure cannot be assumed ad hoc as unknown without additional verification steps. The example with enhanced resolution is just stressing the importance of the "classical" part of the signal- namely its strength- which may depend on the estimated parameter. This can be the result of some (presumed) internal process responsible for generation of such a signal, or a part of the controlled experiment: In the first state-preparation step, parameters are coded into the signal and prior information about parameters is used. In the second detection step, where we use only knowledge of the model without using knowledge of the parameters, these parameters are retrieved from data with an increased precision.
Such a protocol can be viewed as testbed for experimentalists, and requires the mastering of several advanced steps including state engineering, detection on demand, noise control and data processing.
In this perspective, the coherence spreads information into different interfering channels, into the (modulated) strength or probability distribution. As a part of this scenario we will finally show that coherence may allow efficient sorting of information about parameter of separation into the anti-phase superposition, and centroid position into the in-phase superposition.
The analysis will clarify some conceptual misunderstandings related to use of unnormalised states in quantum information processing \cite{Saleh,Comment_Tsang,Reply_Larson}. }
In the following, the quantum formalism of Dirac notation will be adopted, though the motivation for this research stems from classical optics. The method used here will rely on quantum optimisation and was developed in the context of quantum theory. The conclusions derived here will be still valid in the broader context of quantum metrology. To highlight the optical interpretation, the relevant conclusions will be framed in terms of classical optics.
\section{Method and Results}
Let us review some basic facts from quantum estimation theory for unbiased estimators of a variable $\theta.$ The Fisher Information $F$ quantifies the content of
information per detected particle; the Cr\'amer--Rao inequality (CR) \cite{Helstrom} (here for a
single parameter $\theta$ and sequential detection),
\begin{equation}
(\Delta \theta )^2\geq 1/ {\cal H } \equiv \frac{1}{{n {\cal F}}}\,,
\label{precision}
\end{equation}
sets a bound given by the precision ${\cal H} $ with which $\theta$ can be estimated from
the data. Here, $(\Delta \theta )^2$ is the expected value of the variance of the unbiased
estimator, and $n$ is the number of repetitions of the detection. { The Fisher information $ \cal F$ plays the pivotal role in quantum metrology and can be calculated for the given model by means symmetric logarithmic derivative (for more information see the topical review \cite{Liu20}, where the mathematically rigorous formulation is illustrated with numerous examples for applications in quantum science). } The Fisher information quantifies well the resources needed for reaching some target precision: small values of $ \cal F$ must be compensated by an increase of the data set.
As argued in \cite{coherence} special attention must be paid to coherent superpositions, where some extra cost are associated with preparation of a particular superposition of displaced states:
\begin{eqnarray}
\label{Phi12}
& |\Phi_{1,2}\rangle&=\frac{1}{2}\bigl(|\Psi_{+}\rangle
\pm e^{i\varphi}|\Psi_{-}\rangle\bigr) ,
|\Psi_{\pm}\rangle =\exp(\pm iPs/2)|\Psi\rangle,
\end{eqnarray}
where $P$ is the operator of transverse momentum, $s$ parametrises the separation, { $\varphi = 0$ and 1D geometry is assumed for the simplicity. The superposition state with $\pm$ sign will be called in-phase and anti-phase $ |\Phi_{1,2}\rangle, $respectively.}
As reviewed in the Supplementary Materials, costs for preparing such states are simply given by the corresponding norms $ | \Phi_{1,2} |^2. $
The normalisation used here $ | \Phi_{1} |^2 + | \Phi_{2} |^2 = 1 $ guarantees that costs $C $ for preparing the incoherent mixture equal to one. In general, the costs for preparing the partial coherent mixture enter the CR inequality as a multiplicative rescaling $n$ into $nC$ in the inequality Eq. \ref{precision} with the obvious meaning: if $n$ pairs of the separated states $|\Psi_{\pm}\rangle $ are at our disposal, the desired superposition is generated $nC $ times. Mathematical models leading to specific choice of $C$ will be specified later.
Partially coherent states and the corresponding rate factors can be generated by many ways, all of them legitimate in the sense of estimation theory but just a few of them useful. The following two basic models (A, B) are useful for understanding of the role of coherence in quantum metrology. In Model A, the states are normalised before they are added, whereas in Model B the unnormalised states ( "amplitudes") are added. The way how to sort out those states from a generic scheme is clarified in Supplementary Materials \cite{Supp}.
\begin{eqnarray}
\label{state_A}
\rho_A = \frac{p_1}{|| \Phi_1 ||^2} |\Phi_1\rangle \langle \Phi_1 | + \frac{ p_2}{|| \Phi_2 ||^2} |\Phi_2\rangle \langle \Phi_2 | , \\
\label{state_B}
\rho_B = \frac{1}{C} \biggl[ {p_1} |\Phi_1\rangle \langle \Phi_1 | + p_2 |\Phi_2\rangle \langle \Phi_2 | \biggr] , \\
\label{rate}
C = p_1 || \Phi_1||^2 + p_2 || \Phi_2||^2 , \quad p_1 + p_2 = 1 .
\end{eqnarray}
The rate $C$ is equal in both the cases.
As shown in Supplementary Materials, such states can be generated "blindly," meaning without the knowledge about the true value of the estimated parameter of separation $s.$ { The difference between these models was a subject of some misunderstanding. }
The theory elaborated by Larson and Saleh in Refs. \cite{Saleh,Reply_Larson} represents a mixture of models A and B, whereas Tsang and Nair \cite{Tsang-review,Comment_Tsang} have considered just Model B as the only option relevant for optics. We also point the reader to the arguments from the Appendix 3 of the Ref. \cite{Tsang-review} based on Helstrom book \cite{Helstrom} advocating the use of unnormalised function of mutual coherence instead of its normalised form. This may be but need not be always so, depending whether the normalisation factor $C$ is or is not available from the measurement. { These issues are detailed in Supplementary Materials \cite{Supp}. As a brief message for experimentalists, it matters how the partially coherent state is created - either by mixture of coherent superpositions or by an incoherent mixture superimposed with coherent signal, whether or not the intensities are fixed or depend on parameters, and particularly the strength of the detected signal is important. Coherent effects allow to distribute information in different channels (i.e. partially coherent states) and/or modulation of the strength of the signal. "Rayleigh's curse" reappears only if some piece of this information is omitted but coherence cannot be blamed for that. }
{ The central role in quantum metrology is played by QFI matrix associated with the given model. Though this is in general a rather involved problem, the off-the-shelf formulae are at disposal \cite{Liu20}. For the understanding of the role of coherence, the rank-2 density matrix is enough and QFI can be cast in the form }
\begin{eqnarray}
{\cal F}_{2-rank} = \sum_{i=1,2} \frac{ (\partial _s \lambda_i)^{2}}{\lambda_i} + 4 \sum_{i=1,2} \lambda_i [ \frac{ ||\partial \Phi_i ||^2 }{ ||\Phi_i ||^2 } - \frac{ |\langle \Phi_i | \partial \Phi_i \rangle |^2 }{ ||\Phi_i ||^4} ] .
\label{QFI-2}
\end{eqnarray}
where $\lambda_{1,2}$ are the eigenvalues of the density matrix and $|\partial \Phi\rangle $ is the derivative of the state with respect to the parameter. As an important distinction the
QFI in Model A is given just by adding Fisher informations of individual coherent superpositions,
whereas in Model B an extra term appears as a consequence of modulating eigenvalues of those coherent superpositions. { This contribution is large for states close to anti-phase superposition, where the signal is vanishing for small separation, see Supplementary Materials \cite{Supp}. } However, if the QFI is accompanied by the rate factor C (\ref{rate}) , the overall precision is limited by the resolution of the incoherent mixtures.
This limit can be overcome in some specific scenarios when the parameter is coded into the signal in order to distinguish it more efficiently afterwards. { Assume the scheme when the partially coherent state is created in such a way that normalised superposition states $|\Phi_1\rangle , |\Phi_2 \rangle $ are normalised and mixed with the parameter -dependent weights $\bar{ p}_1(s), \bar{ p}_2(s).$ If we choose
\begin{eqnarray}
\bar{ p}_1(s) = \frac{ p_1 ||\Phi_1||^2 }{ C }, \quad \bar{ p}_2(s) = \frac{ p_2 ||\Phi_2 ||^2 }{ C } , \\
\rho_E = \frac{ \bar{p}_1}{|| \Phi_1 ||^2} |\Phi_1\rangle \langle \Phi_1 | + \frac{\bar {p}_2}{|| \Phi_2 ||^2} |\Phi_2\rangle \langle \Phi_2 | = \rho_B, \\
C = p_1 || \Phi_1||^2 + p_2 || \Phi_2||^2 , \quad \bar{p}_1 + \bar{p}_2 = p_1 + p_2 = 1 ,
\end{eqnarray}
we get the same state as in Model B! What has changed is just the rate factor which will be given as
\begin{eqnarray}
\label{E-rate}
C_E = \frac{1}{C} [ p_1 || \Phi_1||^4 + p_2 || \Phi_2||^4 ] ,
\end{eqnarray}
with all the coefficients having the same meaning in all models. QFI remains large even when multiplied by the rate factor $C_E.$ Here the product $ C_E {\cal F}$ scales with the small separation like $1/s^2 $ and may exceed considerably the precision of an incoherent mixture. This is the essence of models with enhanced resolution (Model E). This behaviour is exemplified in Fig. 1 in comparison to Model B, which has just indicative meaning and demonstrates that such resolution is limited by incoherent superposition. }
Though this result might appear paradoxical, there is even some classical rationale behind it as explained by behaviour of ``anti-phase''
superpositions, which may be considered as always "distinguishable" due to the gap between the peaks \cite{Cesini},
but such a signal is vanishing when approaching anti-phase.
{ If we analyse the resolution in Model E in dependance on the weight parameter $ p , p_1 = p, p_2= 1-p $ (see Supplementary Materials for details) \cite{Supp}, we conclude that maximum is achieved for states close to the mixtures when in-phase component $ \Phi_1$ is taken with the weight of anti-phase phase component, i. e $ | \Phi_2|^2 $ and the other way around. This defines "nearly optimal" point on the plot in Fig. 1 close to maximum of resolution. For small separations, the anti-phase component is vanishing and in-phase component is dominating, hence the optimal state requires to combine strong signal with vanishing weight and vanishing signal with strong weight factors. This explains the "equal roles" of both the "anti-phase" and "in-phase" components in models yielding enhanced resolution.}
For such a state $C_E = 1/2,$ what means that just half of the signal in comparison to incoherent mixture is used. However, the increase of QFI is enormous and significantly outperforms the precision of QFI for incoherent mixtures.
\begin{figure}
\includegraphics[width=1\columnwidth]{fig1}
\caption{
Dependence of the precision $C {\cal F} $ on the weights $ p_1= p, p_2=1- p$ of partially coherent mixtures for values of the separations $ s = 0.5, 0.7, 1 $ { (dashed-doted, dashed, full ) listed from the bottom up.} Red curves indicate enhanced resolution (E). { The purple curves corresponding to Model B are below the threshold $1/4$- the limit of incoherent mixtures, and have just indicative meaning. } Black points correspond to the "near optimal" $ p$ -values, where $C = 1/2, $ see the explanation in the text. The displacement is in units of $\sigma, {\cal F}$ is in units of
$\sigma^{-2}$ and numerical calculations are done for Gaussian distribution. The inset shows the ratio of corresponding rate factors C { for separations $ s = 0.5, 0.7, 1 $ listed from the bottom up. } }
\end{figure}
To exemplify the enhancement in numbers let us consider the parameters at the separation $ s= 0.1 \sigma $ where super-resolution has been already demonstrated on different experimental platforms \cite{Opt,TF} for incoherent mixtures.
For optimal enhanced super-resolution one can get the numerical factors $C= 0.02 , C_E =0.67, $ whereas $ {\cal F} = 89 $ and the precision scales with the factor $ C_E {\cal F} \approx 59.$ In comparison the precision for the incoherent mixtures of separated sources is constant $ (\Delta P)^2 = 0.25 .$
More detailed calculations are provided in the Supplementary Materials.
The enhanced resolution need not be the only potential benefit of coherence manifested by constructive and destructive interference. Coherent effects may also facilitate new ways for sorting of information. We demonstrate how to exploit the full information content of the signal including the information contained in the modulated rate factor $C$- the norm of the state, and in the normalised state.
As calculated in Supplementary Material \cite{Supp} the full information carried by unnormalised state $ | \Phi \rangle $ { consists of QFI for a pure normalised state (e.g. given by Eq. ( \ref{QFI-2} ) for single eigenvalue $ \lambda_1 = 1 $ ), rescaled by the normalisation $ C = || \Phi ||^2 $ with an extra added Fisher information term corresponding to the modulation $ {(\partial C)^2}/{C},$ all together }
\begin{eqnarray}
{\cal F}_{total} = 4 ||\partial \Phi ||^2 + \frac{ [ \langle \Phi | \partial \Phi \rangle - \langle \partial \Phi| \Phi \rangle ]^2 }{ ||\Phi ||^2} .
\end{eqnarray}
Applying this result to specific in-and anti-phase superpositions $|\Phi_{1,2}\rangle $ we conclude that for small separations the Fisher information for the separation is carried dominantly by the anti-phase component, and particularly is coded into the norm, whereas the information about centroid position is carried by the in-phase superposition. This simple observation may serve as inspiration for future complex protocols for distributing and sorting of information by means of multi port devices. { This effect is graphically illustrated in Fig.2 plotting the Fisher information for separation of Gaussian profiles for in-phase superposition $ {\cal F}_1$ and anti-phase superposition ${\cal F}_2, $ and particularly, its dominant part corresponding to the Fisher information attributed to the norm of anti-phase component $ {\cal F}_{C}.$}
{
\begin{figure}
\includegraphics[width=1\columnwidth]{fig2}
\caption{
Fisher information for separation as a function of separation for in-phase superposition $ {\cal F}_1$ (full blue) and anti-phase superposition $ {\cal F}_2$ (dashed red). Dominant contribution of $ {\cal F}_2 $ for small separations stems from the modulation of the norm of anti-phase superposition plotted as $ {\cal F}_C$ (dashed-dotted purple). The sum of individual terms $ {\cal F}_1$ and $ {\cal F}_2$ is conserved and saturates the limiting resolution for incoherent mixture given as $1/4.$ The separation is in units of $\sigma, $ ${\cal F}$ is in units of $\sigma^{-2}$ and numerical calculations are done for Gaussian distribution. }
\end{figure}
}
\section{Discussion}
The essence of the proposed scheme for enhanced resolution hinges upon the manifestation of genuine partial coherence, { when both the in- and anti-phase coherent channels and their intensities depend on the estimated parameter. The enhancement of the precision stems from the strong modulation of the intensity terms near the point of the almost vanishing signal. The scenario with particular choice of weights considered here serves as an example of how the modulation of coherent signals can be utilised for enhancing the precision above the limit of incoherent mixtures.
There is even a certain classical rationale behind this effect. As shown in Refs. \cite{tempering, reading} points where the intensity of the PSF is vanishing contribute to increased precision even in the case of the intensity detection. The improvement is manifested by the linear scaling of Fisher information. The theoretical analysis presented here is actually the quantum analogue of the same effect: since the signal must vanish for any observation, the signal itself must vanish. Such arguments cannot be used for estimating the separation between unknown sources; however
the analysis is relevant for coding the information into the prepared state which will serve for reading out the information afterwards. }
Though the effect of enhanced super-resolution is theoretically possible, it will be challenging to observe it experimentally. Several steps must be mastered simultaneously: precise state engineering of phase sensitive states, phase sensitive measurements as projection on the eigenstates of the symmetric logarithmic derivative, and all this must be done with extremely weak fields. These steps have been implemented to a certain extent separately on various experimental set-ups. In particular, the experiments in Refs. \cite{tempering,reading} demonstrated that it is possible to detect faithfully the zero intensity signal and based on this to see the region of linear scaling of classical Fisher information. New progress of sophisticated optical methods using quantum pulse gates \cite{QPG} or mode multiplexing \cite{multiplexer} are promising tools for implementation of the phase sensitive detection on demand.
\section*{Conclusions}
The analysis of partial coherence identifies metrological advantages of future protocols inspired by quantum information processing. The ultimate precision of incoherent mixtures provides a natural bound which can be overcome in controlled experiments with information additionally coded into the partially coherent signal. Coherence may also provides additional benefits in the form of sorting of information for different parameters into different channels as demonstrated on the separation attributed to the anti-phase component and the centroid position linked with the in-phase component.
\noindent
{\bf Acknowledgments}
The authors are indebted to Hubert de Guise for valuable comments.
We acknowledge
financial support from the Grant Agency of the Czech Republic (Grant
No.~18-04291S), the Palack\'y University (Grant
No. IGA\_PrF\_2020\_004) the project ApresSF supported be the MEYS, Czech Republic, under the QuantERA programme, which has received funding from the European Union's Horizon 2020 research and innovation programme " and H2020-FETOPEN-2018-2019-2020-01 StormyTune.
\noindent
{\bf Disclosures} The authors declare no conflicts of interest.
\noindent
See Supplement for supporting content.
\newcommand{\DOI}[2]{\href{#1}{#2}}
|
\section*{Introduction}
The processes of electron scattering with spin flip play a dominant role in the Kondo problem, they are determined by the exchange s-d interaction \cite {W1, W2}. They should also be explicitly taken into account when considering the behavior of the Kondo lattice, since at half-filling the insulator state is also determined by this interaction. The one-particle modifications of the Kondo lattice (see for example \cite{1}), exact solvable many-particle models \cite{K0,KS}, which do not taken into account the electron scattering with spin flip unable to describe the Kondo insulator state. Numerical calculations of the Anderson and Kondo lattices take into account clusters with a small number of particles, which significantly affects the adequacy of the results obtained \cite{A1,A2,A3}. This leads to the fact that the mechanism of the formation of the large Fermi surface and the gap in the electronic spectrum in the Kondo lattice at half filling is still unclear. In contrast to the single-impurity Anderson and Kondo models in their lattice version, the processes of electron scattering with spin flip cannot be calculated exactly.
In the weak s-d hybridization limit the Hamiltonian of the Anderson lattice reduces to that of the Kondo lattice. It should also be taken into account, that the local density of localized electrons must be equal to unity. In the weak s-d hybridization limit, the local density of d-electrons in the symmetric Anderson lattice is equal to unity. d-electrons determine the local spin-$\frac{1}{2}$ at the sites of the lattice, only in this case we can talk about the spin-$\frac{1}{2}$ Kondo lattice. Thus the symmetric Anderson model is similar to the Kondo lattice.
In the strong coupling limit of the Hubbard model, when $U \gg t$ ( $U$ and $t$ are on-site repulsion and hopping integral),
it is shown that the Hubbard model and the Kondo lattice model become identical \cite {3}. It gives possibility to use the formalism, proposed for calculation of the Mott transition in the Hubbard model \cite{K1}, for solution of the Kondo lattice problem.
In the paper, we consider the solution of the symmetric Anderson model, using a mean field approach. An effective $\lambda$-field connects the states of d-electrons with the different spins and momenta $\textbf{k}$, $\textbf{k}+\overrightarrow{\pi}$. Due to the s-d hybridization, the states of s-electrons
with different spins and momenta $\textbf{k}$, $\textbf{k}+\overrightarrow{\pi}$ also hybridize, the gap opens in the electron spectrum at half filling. The value of the gap is determined by the magnitude of the $\lambda$-field (which in turn depends on the on-site repulsion) and the value of the s-d hybridization.
We shall show that breaking spontaneous symmetry in the Anderson lattice Hamiltonian makes it possible to take into account the scattering of conduction electrons on d-electrons with a spin flip. They determine the ground state of symmetric Anderson and Kondo lattices, the value of the gap in the electron spectrum, cause formation of an insulator state at half-filling. The electron spectrum is symmetric about the zero energy, it corresponds to the symmetric Anderson model for a nontrivial solution of the $\lambda$-field. The gap in the electron spectrum has a many-particle nature.
\section*{Model}
The Hamiltonian of the Anderson lattice is the sum of two terms, the first of which is determined by energy of the bands of s- and d-electrons and hybridization between them, the second takes into account the on-site repulsion of d-electrons ${\cal H}={\cal H}_{0}+{\cal H}_{int}$
\begin{eqnarray}
&&{\cal H}_0= -
\sum_{j=1}^{N-1}\sum_{\sigma=\uparrow,\downarrow}
(c^\dagger_{j,\sigma} c_{j+1,\sigma}+ c^\dagger_{j+1,\sigma} c_{j,\sigma})
\epsilon_g \sum_{j=1}^{N}\sum_{\sigma=\uparrow,\downarrow}n_{j,\sigma}+v\sum_{j=1}^{N}\sum_{\sigma=\uparrow,\downarrow}
(c^\dagger_{j,\sigma} d_{j,\sigma}+ d^\dagger_{j,\sigma} c_{j,\sigma}),\nonumber\\&&
{\cal H}_{int}= U\sum_{j=1}^{N}n_{j,\uparrow}n_{j,\downarrow},
\label{eq:H}
\end{eqnarray}
where $c^\dagger_{j,\sigma},c_{j,\sigma}$ and $d^\dagger_{j,\sigma},d_{j,\sigma}$ ($\sigma=\uparrow,\downarrow)$ are the fermion operators determined on a lattice site $j$, $U$ is the value of the on-site Hubbard interaction determined by the density operator $n_{j,\sigma}=d^\dagger_{j,\sigma}d_{j,\sigma}$, the band width of c-fermions is determined by the hopping integral equal to one, the energy of flat band of d-fermions equal to $\epsilon_g$, $v$ defines the hybridizations of s- and d-electrons, N is the total number of atoms.
First of all, we will focus on the consideration of the model for single Anderson impurity, exact solution of which has been obtained by Wiegmann \cite{W1}. At $\epsilon_g=-\frac{U}{2}$ and $U \gg \Gamma $ ($\Gamma= v^2$) trickly one electron is localized on the impurity, so we can talk about an impurity with spin-$\frac{1}{2}$. This case corresponds to the symmetric Anderson model. According to \cite{W2} the behavior of impurity in the symmetric Anderson model is equivalent to the spin-$\frac{1}{2}$ Kondo impurity. In this context, in a non-magnetic (paramagnetic) state the behavior of an electron liquid in the symmetric Anderson lattice is similar to that in the spin-$\frac{1}{2}$ Kondo lattice.
Should be notes also, when the hybridization is small the Hamiltonian (1) can be mapped into the Kondo lattice model with the exchange imtegral
$J\simeq -\frac{2\Gamma U}{\epsilon_g(\epsilon_g+U)}$.
\begin{figure}[tp]
\centering{\leavevmode}
\includegraphics[width=.7\linewidth]{1.pdf}
\caption{(Color online)
Electron spectrum of the chain a) and low energy excitations b) (where $\Delta =0.004$) as function of wave vector calculated at $\epsilon_g=-1$ (or $U=2$), $v=0.1$, $\lambda=0.2$.
}
\label{fig:1}
\end{figure}
\begin{figure}[tp]
\centering{\leavevmode}
\includegraphics[width=.7\linewidth]{2.pdf
\caption{(Color online)
Electron spectrum of the square lattice a) and low energy excitations b) (where $\Delta =0.004$) as function of the wave vector calculated at $\epsilon_g=-1$ (or $U=2$), $v=0.1$, $\lambda=0.2$.
}
\label{fig:2}
\end{figure}
\begin{figure}[tp]
\centering{\leavevmode}
\includegraphics[width=.5\linewidth]{3.pdf}
\caption{(Color online)
The gap $\Delta$ in the electron spectrum as function of $v$ and $\lambda$ calculated at $\epsilon_g=-1$ (or $U=2$), in the coordinates $v$, $\lambda$
(the value of $\Delta$ does not depend on the dimension of the lattice).
}
\label{fig:3}
\end{figure}
\section*{The ground-state}
$\lambda$-field breaks the spontaneous symmetry \cite {K3,K4}, the magnetic moments of conduction and d-electrons are not conserved, the total number of electrons is conserved \cite{K1}. Only for $\textbf{q}=\overrightarrow{\pi}$, the electron spectrum corresponds to the symmetric Anderson lattice. Thus, we are talking about the lattice with the doubled period, as in the case of the Peierls phase-transition.
In this case the electron spectrum is symmetric with respect to zero and has the following form for the chain (see in Fig 1a) and square lattice (see in Fig 2a). Four branches of the spectrum $\pm E_\gamma (\textbf{k)} $ ($\gamma =1,2$) are determined by the following expression (see section Methods))
\begin{equation}
E_\gamma(\textbf{k}) = \frac{1}{\sqrt2} \sqrt{\alpha(\textbf{k})+(-1)^\gamma \sqrt{\alpha^2(\textbf{k})-4\beta^2(\textbf{k})}},
\label{eq:H5}
\end{equation}
where $\alpha (\textbf{k})= \varepsilon^2(\textbf{k})+ \epsilon_g^2 +\lambda^2 +2\Gamma$, $\beta^2 (\textbf{k})=(\lambda^2 +\epsilon_g^2)\varepsilon^2(\textbf{k})-2\epsilon_g \Gamma \varepsilon (\textbf{k})+\Gamma^2$.
For $v \neq 0$, $\lambda \neq 0$ and an arbitrary dimension of the lattice the spectrum is gapped at half-filling. The gap is equal to zero at $v=0$ or $\lambda=0$ (see in Fig 3), $v =\lambda =0$ corresponds to the atomic limit of the symmetric Anderson lattice with a bare on-site repulsion.
The dependence of the gap on $v$ and $\lambda$ has quite universal in nature, its value does not depend on the dimension of the model (see in Fig 3). As shown below, the value of $\lambda$ is the solution of Eq (3), at given $U$ its value is determined by $v$, depends on dimension of the lattice.
In the chain the low energy spectrum is determined by two special points (see in Fig 1b). In the square lattice, they form a square in the $\textbf{k}$-plane (see in Fig 2b). The k-surface, formed by the points on the $ \textbf {k} $-plane of the low-energy spectrum, is constructed as follows: each projection of the wave vector corresponds to two values of its other projections. The $\gamma$-branches of the spectrum are hybridized, the electron density is defined as $\sum_{\gamma=1,2}\int d\textbf{k} =n_s +1$, here $n_s$ is the density of s-electrons ($n_s=1$ at half-filling). Thus, we can talk about a large Fermi surface in the Kondo lattice also, the behavior of which is similar to the symmetric Anderson lattice.
In the ground state the equation for $\lambda$ follows from the form of the action (6) and the spectrum (2)
\begin{equation}
\frac{\lambda}{U}= \frac{1}{4}\sum_{\gamma=1,2}\int d \textbf{k} \frac{1}{E_\gamma(\textbf{k})} \frac {\partial E_\gamma^2(\textbf{k})}{\partial \lambda},
\label{eq:H6}
\end{equation}
where the value of the on-site repulsion corresponds to a condition $U=-2\epsilon_g$. The energy of the quasi-particle excitations depend on $\epsilon_g$ (2), thus (3) is self-consistent equation. As a result, the state of the symmetric Anderson lattice can be realized for $\lambda \neq 0$.
As we noted above the gap in the spectrum depends both $v$ and $\lambda$ (in the Anderson model the scattering matrix depends on $v$ and $U$ \cite{W1}). Numerical solutions for $\lambda$ as function of $v$ are presented in Fig 4 for different dimension of the lattice. Results of calculations are obtained at $\epsilon_g=-1$, $U=2$ (see in Fig 4a)) and $\epsilon_g=-1/2$, $U=1$ (see in Fig 4b)). Trivial solution $\lambda =0$ and $v=0$ corresponds to the symmetric Anderson model with $\epsilon_g=-\frac{U}{2}$ in which s- and d-electrons are not coupled. Nontrivial solution for $\lambda$ and $v$ determines state of electron liquid in the symmetric Anderson model. Solution for $\lambda$ takes place at finite values of $v$, the $\lambda$-value increases with increasing $v$ (see in Fig 4). Such at $v=0.45$ the gaps in the spectrum have the following values for U=2 (U=1): $\Delta =0.015$ ($\Delta =0.093$) in the chain, $\Delta= 0.018$ ($\Delta =0.074$) in square and $\Delta =0.005$ ($\Delta =0.088$) in cubic lattices. The value of the gap decreases with increasing $U$.
The stability of gapped ground state also depends on the magnitude of the on-site repulsion. Comparing actions (6) at T = 0K for the gapped (paramagnetic) state and the phase state with the maximum magnetic moment $M=M_s+M_d=\frac{N}{2}+\frac{N}{2}$, we determine the region of stable insulator state. The minimum value of the on-site Hubbard repulsion $ U_c $, at which the gapped state is realized, is $ U_c = 0.999205$ (1D), $U_c = 0.99933$ (2D), $ U_c = 0.99914$ (3D) at $v=\frac{1}{2}$ and $ U_c = 0.96987$ (1D), $U_c = 0.98969$ (2D), $U_c = 0.97796 $ (3D) at $v=1$. The value of $ U_c $ weakly depends on the dimension of the lattice and the magnitude of the s-d hydridization $ v $. The non-trivial solution for $ \lambda $, shown in Fig. 4, corresponds to stable insulator state at $ U = 2 $, the gapped state at $ U = 1 $, is unstable.
Calculating the actions $ \delta S = \frac {S (\lambda)} {\beta} - \frac {S (0)} {\beta} $ at T = 0K, which correspond to the states with the gap (for $ \lambda \neq 0 $) and gapless with a maximum magnetic moment (for $ \lambda = 0 $), we determine the stability of the insulator state at $ U = 2 $. The numerical calculations are shown in Fig 5 for different dimensions of the lattice. The region of stability of the insulator state is highlighted in color, the solution for $ \lambda $ obtained for $ U = 2 $ (see Fig. 4a) is in the colored region. Note, that $\delta S=0$ at $\lambda =0$ and $\delta S \sim \lambda^2$ at $\lambda\to \infty$, thus intermediate values of $0<\lambda < 0.1$ correspond to the stable insulator state (see in Fig 5).
Taking into account the spectrum of quasi particle excitations (2) in the action (6), we determine the low temperature behavior of the specific heat (its electronic part) for 3D system (cubic lattice). In the insulator state, the value of $ C $ exponentially depends on the value of the gap. Numerical calculations of the temperature dependence $\frac{C}{T}$ are shown in Fig. 6. The calculations are made for $ U = 2 $ and $ v = 0.43$, $ \lambda = 0.0075$, $ \Delta = 0.0023$ (see Fig. 6a) and $ v = 0.5$, $ \lambda = 0.07 $, $ \Delta = 0.0279 $ (see Fig.6b). Note that for these parameters, the gaps in the electron spectrum differ by an order of magnitude.
The experimentally measured heat capacity $ \frac {C} {T} $ $ SmB_6$ \cite {C} shows the anomalous behavior in the low temperature range $ 2K <T <10K $, namely at T = 4K there is a minimum. According to numerical calculations (see insets in Fig. 6), a local minimum of $ \frac {C} {T} $ is realized at low temperatures. The temperature range and the minimum value depend on the value of the gap.
\begin{figure}[tp]
\centering{\leavevmode}
\begin{minipage}[h]{.75\linewidth}
\center{
\includegraphics[width=\linewidth]{4.pdf
}
\end{minipage}
\caption{(Color online)
$\lambda$ as a function of $v$ calculated at $\epsilon_g=-1$, $U=2$ a) and $\epsilon_g=-0.5$, $U=1$ b) for the chain, square and cubic lattices.
}
\label{fig:4}
\end{figure}
\begin{figure}[tp]
\includegraphics[width=\linewidth]{5.pdf
\caption{(Color online)
$\delta S \times 10^5 $ as a function of $v$ and $\lambda$ calculated at $\epsilon_g=-1$ ($U=2$) for the chain a), square b) and cubic c) lattices,
the region of stability of the gapped state is highlighted in color.
}
\label{fig:5}
\end{figure}
\begin{figure}[tp]
\centering{\leavevmode}
\begin{minipage}[h]{.9\linewidth}
\center{
\includegraphics[width=\linewidth]{6.pdf}
}
\end{minipage}
\caption{(Color online)
The temperature dependence of the specific heat $\frac{C}{T}$ calculated for cubic lattice at $U=2$ : $ v = 0.43$, $ \lambda = 0.0075$, $ \Delta = 0.0023$ a) and $ v = 0.5$, $ \lambda = 0.07 $, $ \Delta = 0.0279 $. Inserts illustrate the regions of abnormal behavior of the specific heat at low temperatures.
}
\label{fig:5}
\end{figure}
\section*{Methods}
\subsection*{Canonical functional of the symmetric Anderson model}
Let us introduce the operator $\chi_{j}^\dagger= d^\dagger_{j, \uparrow}d_{j,\downarrow}$ and redefine the term ${\cal H}_{int}$ (1) is the following form
${\cal H}_{int}= -U\sum_{j}\chi_{j}^\dagger\chi_{j}$ \cite{K1}.
The Hubbard-Stratonovich transformation converts the interacting problem into a non-interacting one in a stochastic external field (hereinafter we will define it as the $ \lambda $-field). We define the interaction term, taking into account the action ${S}_{0}$
\begin{eqnarray}
&& S=S_{0}+\sum_{j} \frac{\lambda^\ast_{j}\lambda_{j}}{U}+\sum_{j}(\lambda_{j} \chi_{j}+ H.c.)
\label{A1}
\end{eqnarray}
The canonical functional is determined as $${\cal Z}=\int {\cal D}[\lambda] \int {\cal D}[\chi\dagger,\chi] e^{-S},$$ where the action $S=\frac{1}{U}\sum_{j}\lambda^\ast_{j}\lambda_{j}+
\int_0^\beta d\tau \sum_\textbf{k}\Psi_\textbf{k}^\dagger (\tau)[\partial_\tau + {\cal H}_{eff}(\textbf{k})]\Psi_\textbf{k} (\tau)$, $\Psi_\textbf{k} (\tau)$ is the wave function, $\textbf{k}$ is the wave vector of an electron. We expect that $ \lambda_{j} $ does not depend on $ \tau $, since the translation invariance is conserved in an electron liquid state.
At the on-site hybridization between d-states of electrons with different spins and due to translation invariance of the model Hamiltonian, only the phases of $\lambda_j$ depends on $j$, a namely $\lambda_{j}=\exp(i \textbf{q }\textbf{j}) \lambda$ \cite{K1}, where $\textbf{q}$ is an unknown wave vector.
The phase can also fluctuate at the lattice site; subsequent averaging over local fluctuations restores translational invariance \cite{D}.
The task is reduced to moving fermions in a static inhomogeneous $\lambda$-field that sets the form of ${\cal H}_{eff}(\textbf{k})$
\begin{equation}
{\cal H}_{eff}(\textbf{k}) = \left(
\begin{array}{cccc}
-\varepsilon (\textbf{k}) & v & 0 & 0\\
v & \epsilon_g & \lambda & 0\\
0 & \lambda^*& - \epsilon_g& v \\
0 & 0& v & -\varepsilon (\textbf{k}+\textbf{q})
\end{array}
\right)
\end{equation}
The spectrum of non-interacting s-electrons is given by $\varepsilon (\textbf{k})=-2\sum_{i=1}^D \cos k_i$, here D is dimension of the model.
We can integrate out fermions to obtain the following action $S$ per an atom
\begin{eqnarray}
\frac{S(\lambda)}{\beta}=-\frac{{T}}{{N}}\sum_{\textbf{k}}\sum_n \sum_{\gamma=1}^{4} \ln [-i \omega_n+E_\gamma(\textbf{k},\textbf{q})]
+\frac{|\lambda|^2}{{U}},
\label{A2}
\end{eqnarray}
where $\omega_n =T(2n+1)\pi$ are the Matsubara frequencies, $\textbf{k}$, $\textbf{q}$ are the momenta of electrons, four quasiparticle excitations $E_\gamma(\textbf{k},\textbf{q})$ ($\gamma =1,...,4$) determine the electron states in the $\lambda$-field. In the saddle point approximation the canonical functional ${\cal Z}$ will be dominated by the minimal action $S$ (6), that satisfies the following equation $\partial S/\partial \lambda =0$.
In the Kondo problem, the processes of electron scattering from spin flip dominate, in our case, the $\lambda$-field connects the states of s-electrons with opposite spins and different momenta $\textbf{k}$ and $\textbf{k}+\textbf{q}$. Due to the on-site repulsion between electrons in the Hubbard model\cite{K1}, an effective field connects states of electrons with different spins and different momenta $\textbf{k}$, $\textbf{k}+\overrightarrow{\pi}$. This state is stable at half-filling, a gap opens in the electron spectrum \cite{K1}. In this case, the same mechanism of coupling between s-electrons is realized indirectly through an intermediate subsystem of d-electrons. The gap in the electron spectrum will naturally be smaller.
\section*{Conclusion}
Using mean field approximation we have considered the solution of the symmetric Anderson lattice at half-filling for different dimensions of the lattice.
It is shown that an effective field, which binds the states of s-electrons with different spins and momenta, leads to the gap in the electron spectrum. The electron spectrum is mirror symmetric (with respect to zero energy), has the type of the Majorana spectrum. The states of s- and d-electrons are hybridized, therefore the Fermi surface is determined by the total density of electrons. In the symmetric Anderson lattice model the gaped state of electron liquid corresponds to breaking spontaneous symmetry.
This made it possible to take into account the processes of spin flip scattering of conduction electrons on localized electrons. This is important because it is these processes that lead to the appearance of the Abrikosov - Suhl resonance in the Kondo problem and an insulator state in the Kondo lattice.
The gapped state is formed at finite values of the s-d hybridization and on-site repulsion.
Numerical calculations of the low-temperature behavior of the heat capacity are carried out in the gapped state.
The results explained the presence of a local minimum of the heat capacity $ \frac {C} {T} $ at low temperatures observed in $SMB_6$.
In the symmetric Anderson lattice, a local spin-$\frac{1}{2}$ is realized at lattice site, therefore its behavior is similar to the spin-$\frac{1}{2}$ Kondo lattice. The proposed approach allows us to describe the symmetric Anderson and Kondo lattices in one formalism. One can speak of a large Fermi surface and the gapped state of an electron liquid in the Kondo lattice at half filling.
|
\section{Background}
In this section, we introduce the context and core concepts that concern this review, namely test case generation and agent-based models. An informal definition of test case generation will be provided and then revisited in the context of an agent-based model. In addition, the challenges associated with testing an agent-based model and their impact on the test case generation process are highlighted.
\subsection{Agent-based models}
An agent-based model is a system of autonomous processes characterised by sets of individual and collective goals which, upon execution, typically result in complex emergent behaviour. Whilst there is no universal definition of an agent, there are several criteria introduced by Crooks et al. \cite{crooks2012introduction} that are generally agreed upon as defining features of an agent:
\begin{enumerate}
\item Autonomy: agents are self-governing individuals that make decisions and act without centralised influence.
\item Heterogeneity: agents are diverse individuals that are distinguished by differences in their characteristics (e.g. age, gender, height).
\item Active: agents percept and affect their environment, including other agents, to achieve their goals.
\end{enumerate}
Many different forms of agent-based model and similar paradigms exist, from the beliefs, desires and intentions framework \cite{rao1995bdi} (BDI agents) to cellular automata \cite{wolfram1983statistical}, however, in practice the concepts tend to overlap. In this review we consider an agent-based model to be an umbrella term that captures all of the above criteria.\footnote{For reasons of practicality a line is drawn at physical and cyber-physical systems as these topics fall into the field of robotics, though one could argue that they also fall into the agent-based category.}
Agent-based models lend themselves to modelling theory-driven phenomena \cite{arnold2019dag} such as the impact of social policies \cite{chao2015dynamic} or the spread of infectious disease \cite{frias2011agent}, where the underpinning mechanisms are well-understood, but their larger-scale implications are not. A recent example of a significant agent-based model is the COVID-19 pandemic model developed at Imperial College London; a complex simulation that has been used to inform the United Kingdom's response to the COVID-19 pandemic \cite{flaxman2020report}. However, COVID-19 models have received criticism \cite{chatterjee2020transparency, wynants2020prediction} highlighting a lack of transparency, reproducibility, and an urgent need for more rigorous validation. To this end, software testing techniques may be applied to generate test cases that exercise the system under different conditions to reveal software faults and validate the model.
\subsection{Test case generation}
Informally, test case generation is the activity in which test cases are created either manually, semi-automatically or automatically to expose faulty behaviour in a software system. In general, an approach to testing relies on four key components: an information artifact, a generation mechanism, test data adequacy criteria and a test oracle. Here we define these components alongside fundamental testing concepts that will be mentioned throughout the paper and present an overview of the test case generation process used to frame and motivate our research questions.\newline
\begin{itemize}
\item \textbf{Test} - The act of exercising software with test cases with the goal of discovering faults or demonstrating correctness \cite{jorgensen2018software}.
\item \textbf{Test case} - A set of inputs, preconditions and expected outcomes associated with a particular program behaviour \cite{jorgensen2018software}.
\item \textbf{Test suite} - A collection of test cases \cite{harrold2001regression}.
\item \textbf{Test execution} - An execution of the system-under-test constrained to the conditions specified by a given test case.
\item \textbf{Information artifact} - A source of information from which test cases may be derived. An information artifact can take the form of a formal specification, a design document or source code, and provides information about the functionality or goals of the system that are necessary to produce meaningful test cases.
\item \textbf{Generation mechanism} - An algorithm or strategy used to produce a test case from an information artifact. A generation mechanism is central to the creation of test cases, such as the application of mutation operators to seed test cases or the use of user-defined input generation rules.
\item \textbf{Test data adequacy criterion} - A measure of test case quality with respect to a testing goal such as coverage \cite{zhu1997software}. A test data adequacy criterion is often used to guide test case generation towards achieving the defined goal.
\item \textbf{Test oracle} - A procedure that determines whether a given test execution of the system-under-test is correct or not \cite{mcminn2015oracle}. A test oracle may be specified, derived, implicit, or human.
\end{itemize}
In light of these definitions, test case generation may be described as the process by which a test suite is generated from an information artifact using a generation mechanism, often guided by a test data adequacy criterion. Following the generation of a test case, its execution on the system-under-test should be judged by a test oracle to determine whether it passes or fails. However, it is important to note that these features are not essential for test case generation, but are necessary to consider as part of the larger testing process. A generalised overview of the testing process is presented in Figure \ref{fig:tcgProcess} below.
\begin{figure}[!htb]
\centering
\includegraphics[width=\linewidth]{images/testCaseGeneration/testCaseGenerationProcess.png}
\caption{An overview of the testing process.}
\label{fig:tcgProcess}
\end{figure}
\subsection{Testing agent-based models}
Software testing is a significant challenge in the context of agent-based models. This is partly due to the challenging features of an agent-based model \cite{nguyen2008experimental}, but may also be attributed to the domain in which agent-based models are typically developed, such as epidemiology and social sciences, where there are few formally trained software engineers with testing experience \cite{kanewala2014testing}. Additionally, with multiple levels of abstraction, an agent-based model requires the consideration of several different testing objectives. Nguyen et al. propose five levels of abstraction to consider when testing agent-based models \cite{nguyen2009testing}, these are as follows:
\begin{enumerate}
\item Unit: Testing of individual units that make up an agent, including code, goals and plans.
\item Agent: Testing the integration of the different units of an agent, amounting to their ability to act as an individual.
\item Integration: Testing the interaction of agents, the interaction of agents with their environment, and communication protocol and semantics.
\item Society: Testing the multi-agent system as a collective, including the macroscopic and emergent behaviour of the overall system.
\item Acceptance: Testing the multi-agent system in the real execution environment to verify whether the implementation meets stakeholder goals.
\end{enumerate}
Agent-based models have unique properties that pose significant challenges for software testing. These challenges affect all levels of abstraction, including individual agents, interacting agents, and the resulting society of agents.
Agents are autonomous; they can proactively plan and adapt their behaviour to achieve specific goals \cite{padgham2005developing}. With control over their state \cite{adra2010mutation}, agents may derive new plans of their own accord to tackle problems in previously unknown ways \cite{tiryaki2006sunit}. However, autonomous behaviour also exacerbates the test oracle problem \cite{mcminn2015oracle}: how can we determine whether newly introduced behaviour is correct or incorrect, given that it is absent in the specification?
Agents can interact with many other agents and their environment concurrently, leading to a potentially intractable number of different interactions and scenarios \cite{zheng2005conformance}. As a consequence, integration-level testing suffers from the problem of controllability \cite{freedman1991testability}. Due to the large number of autonomous interactions, it is difficult to instantiate test cases that will consistently lead to the desired output. Therefore, it is particularly difficult to control and test social behaviour in an agent-based system.
On a macroscopic scale, autonomy and agent interaction lead to complex patterns of social behaviour, known as emergent behaviour. Emergent behaviour is difficult to test as it may require significant time to develop, is not always guaranteed to occur, and cannot be predicted from the characteristics of the involved agents alone \cite{scholl2001agent}. Additionally, a tester must first address the challenges that hinder the lower levels of testing to achieve society level testing.
Testing an agent-based model generally involves executing the model under a particular set of conditions for the agents and their environment. To ascertain whether the model is correct, the resulting outcome is then compared to the expected outcome, which may be captured in either a formal or informal specification. However, due to the wide variety of different agent-based modelling paradigms, frameworks, and tools, the testing process will vary. For example, a technique developed for testing the conformance of messages sent between message-passing agents may not be compatible with a model involving proximity-based interactions.
In the context of an agent-based model, test case generation involves the creation of valid and invalid input data that trigger a particular sequence of events \cite{nguyen2008ontology}. Test case generation may be manual, or may be automated in various ways to support the testing process outlined in Section 2.2. A comprehensive technique may also cover other aspects of the testing process, such as an agent-specific test adequacy criterion to guide test case generation towards achieving a testing goal. A test oracle may also be provided to assign verdicts to test cases.
\subsection{Related work}
To the best of the authors' knowledge, there are no existing systematic literature reviews on the topic of test case generation for agent-based models. However, there are several systematic reviews of related topics.
Bakar and Selamat \cite{bakar2018agent} conducted a systematic literature review and mapping on the topic of agent systems verification, focusing on techniques for checking agent properties and fault detection during the stages of the agent-based development process. In comparison to our own literature review, Bakar and Selamat's review has a wider scope as it investigates testing techniques for agent systems in general, mapping the different properties and faults that existing techniques may detect, whereas our review focuses on test case generation specifically and explores this process in greater detail. Despite having a different scope, many of the results from Bakar and Selmat's review support our own observations. For example, Bakar and Selmat's mapping of the types of properties tested by existing techniques reveal that functional requirements can be tested using many techniques, whilst non-functional requirements cannot. This aligns with our own observation that techniques for generating functional test cases are well-established, whilst there are fewer techniques that target non-functional requirements.
Blanes et al. \cite{blanes2009requirements} present a systematic review of requirements engineering in the development of multi-agent systems. Whilst this work does not focus on testing or test case generation for agent-based models, it covers different aspects of requirements engineering including specification and validation that have some relevance to our own review. In this work, ontologies are recognised as a technique to deal with requirements, supporting the use of ontologies as a formal information artifact in our taxonomy. Blanes et al. also indentified that roughly a quarter of the literature they reviewed provided some form of automated validation, highlighting the need for new automatic approaches to test agent-based models.
Arora and Bhatia \cite{arora2018systematic} conduct a systematic review of agent-based test case generation for regression testing. Whilst this approach is relevant to test case generation, it reviews agent-based approaches to test case generation, whereas our literature review is concerned with test case generation approaches for agent-based models. That is, this review investigates the use of agents for testing rather than how agents can be tested. Although this review does not have any significant relevance to our own, it is necessary to highlight the difference as, based on the title alone, it would appear to be similar.
Existing reviews have covered verification in agent systems, requirements engineering for multi-agent systems, and the use of agents to generate test cases for regression testing. However, these reviews do not consider the components supporting test case generation, such as specification languages or test oracles, in-depth. Our review addresses this gap by reviewing the artifacts used to form tests cases, generation mechanisms, test oracles, test adequacy metrics, and levels of testing abstraction achieved. To that end, our review covers not only the state-of-the-art test case generation techniques for agent-based models but also the individual components on which these techniques depend.
\section{Discussion}
This section provides a collective discussion of the reviewed literature, highlighting the advantages and disadvantages of the reviewed techniques, interesting observations from each research questions, and future research directions.
\subsection{Advantages and disadvantages}
In response to RQ1 and RQ2, we reviewed a range of information artifacts and generation mechanisms, each having different advantages and disadvantages that make it more or less suitable for particular applications. In Tables \ref{tab:advantagesAndDisadvantagesIA} and \ref{tab:advantagesAndDisadvantagesGM}, we summarise the relevant advantages and disadvantages of the reviewed information artifacts and generation mechanisms respectively.
{
\renewcommand{\arraystretch}{1.2}
\begin{table}[htb]
\footnotesize
\begin{tabularx}{\linewidth}{XXX}
\toprule
Information Artifact & Advantage & Disadvantage \\ \midrule
Design documents &
Widely used in agent-oriented development and therefore applicable to many existing systems \cite{padgham2013model}. &
Requires the user to follow particular design methodologies which are time consuming to learn \cite{hadar2010empirical}. \\
State machines &
Supports the use of a range of established test generation techniques \cite{kalaji2009generating}. &
Requires the user to learn how to formally specify state machines \cite{friedman2002projected}. \\
Contracts and assertions &
Users can specify behaviour independent of the implementation i.e. what the model should do, not how the model should do it \cite{thaler2019show}. &
No guarantee that a test case will be generated that, upon execution, covers
a specified contract or assertion \cite{leitner2007contract}. \\
Ontologies &
Supported by many popular agent development platforms including JADE \cite{nguyen2008experimental}. &
Converting domain expertise into an ontology is difficult and time consuming \cite{tan2019lessons}. \\
\bottomrule
\end{tabularx}
\caption{Advantages and disadvantages of the different information artifacts}
\label{tab:advantagesAndDisadvantagesIA}
\end{table}
\clearpage
\begin{table}[h!]
\footnotesize
\begin{tabularx}{\linewidth}{XXX}
\toprule
Generation Mechanism & Advantage & Disadvantage \\ \midrule
Path-traversal (State machine) &
Under specific design-for-test conditions, certain techniques can guarantee correctness of the implementation \cite{eleftherakis2011methodology}. &
Suffers from state-explosion problem, particularly in concurrent systems \cite{friedman2002projected}. \\
Path-traversal (Petri net) &
Provides a natural representation for concurrent behaviour \cite{kissoum2008recursive}, a defining feature of agent-based models. &
Suffers from the complexity problem. That is, they quickly become too large to analyse for even modest-size systems \cite{murata1989petri}.\\
Path-traversal (AUML) &
Effective for generating test cases that exercise agent interactions and scenarios \cite{kissoum2008recursive, dehimi2019novel}. &
Agents are represented as objects in AUML, limiting the ability to express and test proactive behaviour \cite{gonccalves2015mas}. \\
Information extraction (Design documents) &
Supports class partitioning and boundary value analysis to efficiently achieve a particular level of coverage \cite{zhang2009automated}. &
Additional testing information must be added manually to design documents to enable test case generation \cite{zhang2009automated, thangarajah2011scenarios}. \\
Information extraction (Ontology) &
Supports ontology alignment to expand and diversify the available testing information \cite{nguyen2008experimental}. &
Ontology-based approaches require greater effort than manual testing to set up initially, but are more efficient in the long term \cite{moser2010ontology}. \\
Random testing &
Capable of generating large amounts of test cases for a relatively small cost \cite{arcuri2011random}. &
Requires manual intervention or custom data generators to handle complex, non-numerical inputs \cite{thaler2019show}. \\
Rule-based (Transformation rules) &
Enables the automatic generation of domain-relevant test cases using reusable transformation rules \cite{shahir2012generating}. &
Transformation rules must be manually specified based on domain expertise \cite{shahir2012generating}, often using a formal language such as Maude \cite{mokhati2013novel}. \\
Rule-based (Genetic operators) &
Enables the generation of test cases that are challenging with respect to a specific user-defined testing objective \cite{nguyen2012evolutionary}. &
Performance is dependent on how accurately the fitness function captures the testing objective \cite{nguyen2012evolutionary}. \\
\bottomrule
\end{tabularx}
\caption{Advantages and disadvantages of the different generation mechanisms}
\label{tab:advantagesAndDisadvantagesGM}
\end{table}
}
\subsection{Discussion of results}
In response to RQ1, two different groups of information artifacts were identified: formal and informal. The results to RQ1 showed that significantly more approaches use formal information artifacts than informal information artifacts as the basis for test generation. As a consequence, the majority of the reviewed approaches focus on well-defined, functional properties of an agent-based system that can be verified using existing, well-established testing techniques. For example, where state machines have been used as a formal specification of the system-under-test, the W-method may be used to generate test cases. However, a state machine does not support the specification and testing of non-functional, qualitative properties intrinsic to agent-based models such as efficiency and reliability. Conversely, only a small number of approaches have used informal information artifacts to drive test case generation, that instead focus on approximating soft goals and testing whether they are fulfilled. \\
\noindent\fbox{%
\parbox{\linewidth}{%
\textbf{Observation 1:} The majority of reviewed information artifacts do not facilitate specification and testing of non-functional requirements such as soft goals.
}%
} \\[0.2cm]
In response to RQ2, four groups of generation mechanism were found: information extraction, path traversal, rule-based and random testing. Following the outcome of RQ1, most of these generation mechanisms harness information expressed by a formal information artifact, such as a Petri net, state machine or ontology, to form new test cases that test different behaviour. Whilst path-traversal, information extraction and rule-based mechanisms were extensively used, only a small number of approaches used random testing as a mechanism for test case generation. \\
\noindent\fbox{%
\parbox{\linewidth}{%
\textbf{Observation 2:} Information extraction, path traversal and rule-based mechanisms have been used as techniques for generating test cases that exercise functional requirements.
}%
} \\[0.2cm]
We found that in the few approaches where informal artifacts had been used, genetic algorithms had been adopted as a solution to generate test cases that satisfy a non-functional requirement. In these approaches, the genetic operators known as mutation and crossover are repeatedly applied to seed test cases to form new, potentially more challenging tests with respect to a non-functional requirement expressed as a fitness function. \\
\noindent\fbox{%
\parbox{\linewidth}{%
\textbf{Observation 3:} Genetic algorithms have been used as a technique for evolving test cases that exercise non-functional requirements.
}%
} \\[0.2cm]
In response to RQ3, three forms of test oracle were discussed: specified, derived, and human. Due to most approaches involving some form of specification as an information artifact, the vast majority of approaches involved a specified oracle. Where informal specifications of non-functional requirements were used, oracles were generally derived from the system-under-test or domain knowledge in the form of fitness functions that approximate soft goals. One approach also required a human oracle to judge whether the visual output of an agent-based model matched the specification. This research question proved particularly difficult to answer as most of the reviewed literature did not discuss the test oracle in great detail, neglecting information about how the outcome of a test execution is compared to the specification in order to assign a verdict. Whilst in some cases this is a trivial comparison, in others, it is not so straightforward and it appears that domain expertise are needed to compare the actual execution to the expected execution, raising questions over whether the oracle should be considered specified or human. Therefore, it is necessary to communicate all details of the oracle. \\
\noindent\fbox{%
\parbox{\linewidth}{%
\textbf{Observation 4:} Most of the reviewed approaches do not provide sufficient details regarding the test oracle.
}%
} \\[0.2cm]
In response to RQ4, two forms of test data adequacy criterion were identified: coverage criteria and fitness functions. Where a formal information artifact has been used to specify the behaviour of the system-under-test, we found that a variety of different agent-based coverage metrics were used as a measure of test data adequacy. For example, in some BDI agent systems, test suites have been created with the aim of covering various different artifacts specific to the BDI agent architecture such as agent plans. \\
\noindent\fbox{%
\parbox{\linewidth}{%
\textbf{Observation 5:} The adequacy of a functional test suite is generally measured by some form of agent-based coverage.
}%
} \\[0.2cm]
Where an informal information artifact has been used, fitness functions capturing application-specific soft goals have generally been used as a measure of test adequacy, such that an adequate test suite is one that poses a significant challenge with respect to the non-functional requirement in question. Therefore, the goal of techniques driven by informal artifacts is not to cover a particular component of the system-under-test, but to develop test cases that force the system-under-test into extreme conditions under which soft goals such as efficiency become increasingly difficult to satisfy. \\
\noindent\fbox{%
\parbox{\linewidth}{%
\textbf{Observation 6:} The adequacy of a non-functional test case is generally measured using a fitness function that represents fulfilment of the non-functional requirement.
}%
} \\[0.2cm]
In response to RQ5, we classified each of the reviewed approaches according to the level of abstraction that the generated test suites target: unit, agent, integration, society, and acceptance. We found that the significant majority of approaches target the agent level of testing, generating test cases that target the functionality of individual agents. Several approaches also targeted the unit, integration and acceptance levels of testing. However, the society level of testing has received far less attention than the other levels of abstraction. \\
\noindent\fbox{%
\parbox{\linewidth}{%
\textbf{Observation 7:} Most of the reviewed approaches target the agent-level of abstraction, whilst the society-level has received the least attention.
}%
} \\[0.2cm]
After conducting the quality assessment, we found that the quality of the reviewed approaches was generally high, with most of the reviewed literature scoring highly in QA1 and QA2 reflecting clear definition and fulfilment of aims. However, we found that in most cases there were insufficient details to replicate the approach, and in particular, only a few approaches were evaluated on ``real-world'' agent-based systems.
Whilst the majority of techniques were demonstrated on toy examples, one technique was evaluated thoroughly on a complex production automation system \cite{moser2010ontology}. The evaluation focused on the feasibility, cost-reducing capability, and coverage of possible scenarios achieved using an ontology-based test case generation technique. The authors demonstrated that in the long-term, ontology-based test case generation outperforms manual test case generation in all dimensions. Given the success of this technique in an industrial context, future work should also consider real-world case studies for evaluation in order to demonstrate their potential in industry. \\
\noindent\fbox{%
\parbox{\linewidth}{%
\textbf{Observation 8:} Evaluation is rarely performed on real-world systems that include typical agent-based properties such as concurrency and non-determinism.
}%
} \\[0.2cm]
Most of the reviewed literature did not include an experimental evaluation. Instead, most techniques were demonstrated on example systems ranging from bibliographic management systems to implementations of agent interaction protocols. However, these did not incorporate a controlled comparison with other techniques. In the few papers that did conduct an experimental evaluation, mutation testing was generally applied to an example system. These works report the mutation score or fault detection rate of their test generation technique with respect to a number of known or seeded faults.
Whilst mutation testing provides a measure of fault detection capability, multiple confounding factors must be considered in order to provide a robust comparison between techniques \cite{papadakis2019mutation}. In particular, the results of mutation testing depend on the introduced faults and the system-under-test. Therefore, to achieve a fair comparison, techniques should be compared in terms of their ability to reveal faults in identical systems. To this end, future work would benefit from a database of benchmark agent-based systems seeded with faults, similar to the Defects4J framework in Java \cite{just2014defects4j}. \\
\noindent\fbox{%
\parbox{\linewidth}{%
\textbf{Observation 9:} Existing techniques are difficult to compare due to a lack of consistent experimental evaluation methodology.
}%
} \\[0.2cm]
\subsection{Impact on research and practice}
The findings of this literature review have several implications for both research and practice concerning agent-based models:
\begin{enumerate}
\item Testing of non-functional requirements needs greater attention from researchers. There are comparatively few test case generation techniques that support the specification and testing of non-functional requirements in agent-based models. This is a significant limitation as agent-based models often have soft goals as requirements, such as reliability and safety, that the majority of approaches cannot test.
\item To advance the reproducibility and comparability of techniques, researchers should provide greater detail about the test oracle. Without full details of the test oracle, the reader is left to make implicit assumptions about the testing procedure, making the comparison of research difficult \cite{staats2011programs}.
\item Additionally, future work would be more comparable if experimental evaluations used the same subject systems. To that end, we recommend that a collection of faulty agent-based models, similar to the Defects4J framework in Java \cite{just2014defects4j}, should be developed to be used as a baseline for comparing agent-based testing techniques.
\item Due to a lack of rigorous experimental evaluation, it is currently unclear how effective the reviewed approaches would be in practice. In order to achieve more rigorous and realistic evaluations, future work should consider using real-world case studies as examples, such as the complex automation system studied by Moser et al. \cite{moser2010ontology}.
\item In RQ1, we identified a number of forms of information artifact. A potential barrier to their adoption in industry is the time required to learn and construct them. Future work should investigate methodologies for developing user-friendly artifacts in a timely manner so as to encourage wider adoption.
\item Only one of the reviewed test case generation techniques covered both functional and non-functional test input generation. Therefore, practitioners will usually have to adopt multiple techniques in order to generate test cases that exercise both functional and non-functional requirements. To address this, future tools would benefit from providing both functional and non-functional test case generation.
\item Most of the reviewed test case generation techniques are only applicable to a particular form of agent-based model or development methodology. Future work should consider techniques that could be applied more generally to any form of agent-based model, without requiring significant additional effort. We believe that this is one of the necessary steps towards adoption in industry.
\end{enumerate}
\subsection{Threats to validity}
Following Kitchenham and Charter's guidelines for conducting systematic literature reviews, we have considered potential threats to validity that may affect the results presented in this systematic literature review. We discuss the following categories of threats to validity defined by Zhu et al. \cite{zhou2016map}: construct validity, internal validity, external validity and conclusion validity.
\subsubsection{Construct validity}
Construct validity primarily concerns the review protocol and how inclusive the selected search string, data sources and screening criteria were with respect to the research questions. To minimise threats to construct validity, we included a validation procedure for each of the components of the review protocol. For example, when designing the search string the authors each selected a sample of three papers that they believed the search should retrieve. The search string was then iteratively improved until it could retrieve all of the selected literature. However, it is still possible that some relevant literature may have been missed by the search due to the choice of search terms, data sources and screening criteria.
\subsubsection{Internal validity}
Internal validity concerns the trustworthiness of the cause-effect relationship between the results and the review methodology. In this review, the search procedure was primarily conducted by a single researcher, and therefore the results may be subject to selection bias. To reduce this, a validation procedure was conducted where a randomly selected collection of papers were reviewed by the other authors, such that each paper was reviewed by both the first author and at least one other author. Out of a set of 32 papers, 84\% yielded the same verdict by both authors, and after group discussion over the studies yielding a disagreeing verdict, 90\% of the other authors' verdicts were in agreement with the primary authors. Furthermore, the search procedure has been conducted to favour false positives over false negatives, such that a study was only rejected if the author was certain that it was irrelevant according to the screening criteria. In the case of any doubt, a paper would be moved forwards to the next stage, and if after a full-text review the verdict of a paper was still unclear, it would be reviewed by all authors independently and discussed until a group consensus was reached.
\subsubsection{External validity}
External validity concerns the applicability of the results to a given domain, which in our case is testing for agent-based models. Due to the lack of an agreed-upon definition of an agent-based model, it is possible that some of the reviewed approaches do not conform to a particular definition of an agent-based model. To minimise this threat, we provided an inclusive definition of an agent-based model in Section 2.1 and designed our review protocol accordingly. Furthermore, we carefully designed a taxonomy to group similar responses to each research question under a generalised category, providing results that generalise to various forms of agent-based model regardless of their exact details.
\subsubsection{Conclusion validity}
Conclusion validity concerns the reproducibility of the results. In order to maximise the reproducibility of our results, we followed Kitchenham and Charter's guidelines for systematic literature reviews to provide details including the choice of search string, data sources and screening criteria. We also introduced a taxonomy based on the reviewed literature to provide a more structured approach for answering the qualitative research questions. However, there remains subjectivity over the classification of an approach according to the taxonomy and therefore we cannot guarantee that upon replication the exact same conclusions would be met.
\section{Conclusion}
With the need for drastic improvements in model validation and testing being highlighted recently \cite{wynants2020prediction, squazzoni2020computational}, it is clear that more effective and accessible techniques must be developed for testing agent-based models. In this review, we have collected, organised, and presented the literature on test case generation for agent-based models, providing insights into the current state-of-the-art for test case generation in agent-based models. The systematic literature review was conducted using five digital libraries, yielding a total of 464 papers from which 24 relevant studies were identified. We used the relevant studies to answer five research questions related to the key aspects of test case generation: the information artifact, the generation mechanism, the test oracle, the test data adequacy criterion, and the level of abstraction tested.
The results showed that 88\% of the reviewed approaches focus on generating test cases that test functional requirements, and conversely, that only 12\% of approaches focus on testing non-functional requirements. This statistic suggests that testing of soft goals is not well supported by the majority of the existing approaches, limiting the potential of these techniques in practice. We also found that the majority of the techniques could be classified as information extraction, path-traversal or rule-based in terms of the mechanism used to create test cases. Furthermore, we identified a number of agent-based coverage criteria used to measure the adequacy of functional test suites and described the use of fitness functions to measure the adequacy of non-functional test suites.
This review has also revealed a lack of information regarding test oracles in the agent-based literature. To increase reproducibility and comparability of results and techniques, the procedure for assigning a verdict to a test case based on the provided information must be clearly communicated. It is not enough to state that ``the expected outcome can then be compared to the actual outcome''. Most of the reviewed approaches were also found to generate test cases that exercise agent-level behaviour, whereas the emergent societal behaviours inherent to agent-based systems have been largely ignored. Similarly, we found that evaluation of the reviewed approaches were often conducted using simplistic examples that do not exhibit properties typical of an agent-based system that would make test case generation considerably more difficult. Therefore, future techniques should consider society-level testing and use more realistic examples for evaluation that exhibit challenging agent-based properties such as non-determinism and concurrency.
\section{Research methodology}
In recent years agent-based technology has increased in popularity across a variety of domains ranging from epidemiology \cite{keyes2017invited} to economics \cite{platt2020comparison}. Meanwhile, significant advances in software testing have revolutionised the way we test complex systems, such as the emergence of fuzzing techniques to enhance the automatic discovery of security vulnerabilities \cite{godefroid2012sage}. Despite significant progress in both agent-based modelling and software testing, the need for better validation of agent-based models has been highlighted recently \cite{fagiolo2019validation, squazzoni2020computational, wynants2020prediction}.
This paper reviews existing state-of-the-art test case generation techniques for agent-based models. In this section, we outline the research methodology used to conduct the review, including the review protocol and our research questions.
\subsection{Research questions}
We pose five research questions that focus on the individual aspects of the testing process outlined in Section 2.2. To help answer each question, a taxonomy grouping similar techniques will be introduced in Section 4, facilitating a collective discussion of the key similarities and differences amongst the state-of-the-art.
\subsubsection{RQ1: What artifacts are used to drive test case generation for agent-based models?}
An approach to test case generation requires knowledge of the system-under-test if it is to produce meaningful inputs, identify errors or measure the adequacy of the generated test cases. This knowledge is provided by some information artifact (defined in Section 2.2). In this research question, we identify the information artifacts specific to agent-based models that have been used to drive test case generation.
\subsubsection{RQ2: What mechanisms are used to generate tests?}
An approach to test case generation requires some form of generation mechanism (defined in Section 2.2) that can utilise the knowledge contained in an information artifact to create test cases. Given that there are different types of agent-based models, each with potentially different purposes, it follows that there are also many different forms of generation mechanisms. Additionally, due to the lack of software testing experience amongst scientific developers \cite{kanewala2014testing}, commonplace approaches to test case generation that rely on knowledge of specific technology or languages, such as unit testing in JUnit, may be unsuitable for testing agent-based models. In this research question, we look at the existing approaches used to generate test cases for agent-based models.
\subsubsection{RQ3: How are verdicts assigned to generated test cases?}
Given the often exploratory nature of agent-based models \cite{thaler2019show}, the expected outcome of an execution may be unknown and it may be difficult to assign a verdict to a test case. This is known as the test oracle problem \cite{mcminn2015oracle}. In evaluating an approach to test case generation it is essential to consider the test oracle, as a test case without a verdict provides no information about the validity or quality of the system-under-test. This issue has been discussed in detail by Staats et al. \cite{staats2011programs}. In the context of agent-based models, it is unclear how oracles are used to assign verdicts to test cases. In this research question, we investigate this issue.
\subsubsection{RQ4: How is the adequacy of a generated test suite measured?}
Test data adequacy is a measure of how good a test suite is with respect to some testing goal, such as coverage (see Section 2.2). The definition of a test data adequacy criterion is essential to the fault detection capabilities of a generated test suite \cite{zhu1997software}, as it can be used to drive test case generation towards areas of the system-under-test that are yet to be tested. It follows from our previous research questions that different test case generation techniques will involve different information artifacts and mechanisms, and therefore are likely to have different testing goals too. In this research question, we aim to identify the test data adequacy criterion used to judge the extent to which such goals have been met.
\subsubsection{RQ5: What level of abstraction do the generated test cases target?}
Due to the complex, multi-layered nature of agent-based models, Nguyen et al. propose five levels of abstraction in terms of which agent-based models should be tested: unit, agent, integration, society and acceptance as described in Section 2.3. Previous work has identified a lack of testing approaches for agent-based models that target the society and acceptance levels of testing \cite{nguyen2009testing}. However, this is yet to be investigated for the activity of test case generation. In this research question, we aim to investigate the different levels of abstraction that generated test suites are capable of testing.
\subsection{Review protocol}
To conduct this systematic literature review, we followed the guidelines presented by Kitchenham et al. \cite{kitchenham2007guidelines} and Petersen et al. \cite{Petersen2015}. This involved the following steps:
\begin{itemize}
\item Selecting data sources from which the primary publications were collected.
\item Constructing a search string to capture all relevant publications.
\item Defining inclusion and exclusion criteria to systematically determine the relevance of publications.
\item Designing a quality assessment questionnaire to assess the quality of the relevant publications with respect to the goals of the review.
\end{itemize}
To conduct the search procedure, we initially considered the title and abstract of all papers retrieved using the search string, removing only those that were irrelevant according to the screening criteria. In the next phase we considered the full-text of the remaining papers, and where there was any doubt over the relevance of a study, each researcher would independently review and discuss the paper until a consensus could be met. After identifying the primary publications, forward snowballing was conducted on the primary publications by searching their references to identify any literature missed by the search that may satisfy the screening criteria. The search procedure was then repeated for the snowballed papers. To reduce any potential bias, the components of the protocol have been independently reviewed by each author and a validation procedure has been conducted where necessary. An overview of the search procedure is given in Figure \ref{fig:searchProcess}.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{images/searchProcess/searchProcessHorizontal.png}
\caption{Overview of the search procedure.}
\label{fig:searchProcess}
\end{figure}
In the following sections, we will explain each stage of the review protocol in greater detail.
\subsection{Data sources}
Following the guidelines presented by Petersen et al. \cite{Petersen2015}, we selected five reputable digital libraries as our data sources, including both scientific and indexed databases. The data sources were selected based on their coverage of high-impact agent-based test case generation papers, in addition to the search functionality offered. To validate the choice of data sources, each author identified a number of papers that were relevant to the goals of the review, and we ensured that each paper was available from at least one of the data sources. The selected data sources alongside the number of studies retrieved from each are shown in Table \ref{tab:retrievedStudies}.
\begin{table}[htbp]
\centering
\begin{tabular}{lll}
\toprule
Digital Library & URL & Studies Retrieved \\ \midrule
ScienceDirect & \url{https://www.sciencedirect.com/} & 77 \\
IEEE & \url{https://ieeexplore.ieee.org/} & 58 \\
Scopus & \url{https://www.scopus.com/} & 42 \\
ACM & \url{https://dl.acm.org/} & 91 \\
SpringerLink & \url{https://link.springer.com/} & 196 \\ \midrule
Total & & 464 \\
\bottomrule
\end{tabular}
\caption{Literature retrieved from each data source.}
\label{tab:retrievedStudies}
\end{table}
\subsection{Search string}
We defined a targeted search string to retrieve the relevant studies from each data source. To form the search string, we adapted the \emph{PICO} criteria suggested by Kitchenham et al. \cite{kitchenham2007guidelines}. \emph{PICO} is an acronym designed to help produce a more effective search string by breaking up the focus of the search into the categories: \emph{Population}, \emph{Intervention}, \emph{Comparison} and \emph{Outcome}.
\emph{Population} concerns the target group of the publication, such as a certain demographic, industry or users of a particular tool. For this survey, we identified the population as publications that involve a form of agent-based model.
\emph{Intervention} covers the technique, technology or software used to solve a particular problem in the context of the selected population. We identified the intervention as a test case generation technique, given that we are investigating test case generation for agent-based models.
\emph{Comparison} concerns an alternative intervention that is to be compared with the identified intervention. Given that this review is exploratory in nature, we decided that it would not be appropriate to include the comparison category in the search. For instance, if we included \emph{``search-based''} and \emph{``random testing''} as terms for comparison, any literature that does not mention one of these techniques would be excluded.
\emph{Outcome} relates to the qualities or results of the study focus that a researcher or practitioner would identify as important. In this study, we identified the outcome as a test, given that any test case generation technique should produce at least one test case if successful.
Using the above criteria, we formulated the following search string by identifying synonyms for each group of search terms, joining these by the boolean $OR$ operator, and then joining the resulting clauses with the boolean $AND$ operator, as shown below in Listing \ref{lst:searchString}. The one exception is the population search terms, which are split into two $OR$ clauses joined by an $AND$, such that a study may refer to the population as an agent-based model, agent-based system, agent-based simulation, multi-agent system, multi-agent model, or a multi-agent simulation.
\begin{figure}
\begin{codebox}{Java}{Search string for retrieving relevant studies}{lst:searchString}
("agent-based" OR "multi-agent") AND
("system" OR "model" OR "simulation") AND
("test generation" OR "test case generation" OR "test input generation" OR "parameter selection") AND "test"
\end{codebox}
\end{figure}
Similar to the validation procedure employed for the selection of data sources, each researcher independently identified a collection of papers relevant to the literature review that was used as a benchmark to judge the validity of the search string. The search string was modified until it was capable of obtaining all of the benchmark studies from the selected data sources, resulting in the search string displayed in Listing \ref{lst:searchString}. This included the separation of the population terms by the $OR$ operator as described above, inclusion of the \emph{``simulation''} term, and inclusion of the term \emph{``parameter selection''} which was often used to describe a process that resembles test case generation. During this process the format of the search string was also modified to fit the syntax of each individual search engine, preserving the same meaning as much as possible, within the constraints of the differing functionality of the data sources.
Whilst the validation procedure improves our confidence in the search string, it does not guarantee that it will capture all relevant papers. In fact, we identified a further 4 relevant papers through forward snowballing. These papers were not captured by the initial search as they included population or intervention terms that were too general or niche to include in the search string. Inclusion of the former would increase the rate of false positives, making it infeasible to review each paper. Inclusion of the latter would not have improved the precision beyond the snowballed paper, making the search string overly-specific (an issue for digital libraries constrained to a maximum number of search terms such as ScienceDirect).
\subsection{Screening criteria}
In order to systematically decide which of the retrieved studies were relevant, we applied the inclusion and exclusion criteria shown in Table \ref{tab:inclusionCriteria}. In order to satisfy the inclusion criteria, a paper must satisfy both \textbf{IC1} and \textbf{IC2} in addition to at least one of \textbf{IC3}, \textbf{IC4} or \textbf{IC5}. This ensures that a relevant paper focuses on the topic of test case generation for agent-based models and has been peer-reviewed. On the other hand, if any of the exclusion criteria are satisfied then the paper is deemed irrelevant. The criteria were applied during each phase of the search process, namely the title and abstract review, the full-text review and snowballing as shown in Figure \ref{fig:searchProcess}.
After applying the inclusion and exclusion criteria to an initial selection of 464 papers, we identified 20 relevant papers. A large number of false positives were obtained as there exists a significant body of research on the application of agents to testing activities, such as agent-based test case generation \cite{arora2018systematic}. This research concerns a similar but irrelevant topic and thus uses terminology that is difficult to distinguish from relevant research. To address this issue, we introduced a specific exclusion criterion: \textbf{EC2}.
\begin{table}[htbp]
\centering
\begin{tabular}{ll}
\toprule
ID & Criteria \\ \midrule
\textbf{IC1} & Evaluates or proposes at least one test case generation \\ & method for an agent-based model \\
\textbf{IC2} & Study is published in either a journal or conference \\ & proceedings \\
IC3 & Mentions how the expected outcomes for generated test \\ & cases are known i.e. whether a test case should pass \\ & or fail for a given input \\
IC4 & Mentions the information artifact from which test cases \\ & are derived e.g. a specification \\
IC5 & Mentions a test adequacy measure \\
\midrule
EC1 & Focus of the study does not relate to test case generation \\ & for agent-based models \\
EC2 & Study involves using agent-based models to generate test \\ & cases for a system that is not agent-based \\
EC3 & Non-English study \\
EC4 & Duplicate study \\
EC5 & Study is unavailable in hard or electronic format \\
\bottomrule
\end{tabular}
\caption{Inclusion and exclusion criteria for determining relevance of retrieved papers.}
\label{tab:inclusionCriteria}
\end{table}
\subsection{Quality assessment}
To appraise the quality of the selected studies, we produced a quality assessment questionnaire adapted from the example questionnaires presented by Kitchenham and Charters \cite{kitchenham2007guidelines}, as shown in Table \ref{tab:qualityAssessmentCriteria}. We primarily judge the credibility of a study based on the clarity of its objectives and evaluation, its reproducibility and its real-world impact. Concerning each criterion, a study may be scored as either 0, 0.5 or 1, corresponding to not fulfilled, partially fulfilled and fulfilled respectively. To ensure the credibility of selected studies, we set a minimum quality threshold of 1.5 such that any paper with a quality score less than or equal to 1.5 is ignored due to a perceived lack of credibility.
\begin{table}[!htb]
\centering
\begin{tabular}{ll}
\toprule
ID & Criteria \\ \midrule
QA1 & Are the aims of the study clear? \\
QA2 & Have all aims of the study been addressed in the \\ & evaluation? \\
QA3 & Is sufficient information available to reproduce the \\ & study? \\
QA4 & Does the study consider real-life applications? \\
\bottomrule
\end{tabular}
\caption{Quality assessment criteria for assessing the quality of relevant literature.}
\label{tab:qualityAssessmentCriteria}
\end{table}
\section{Results}
In this section, we outline the results for each research question according to a derived taxonomy. It is important to note that the purpose of the proposed taxonomy is to group and discuss similar approaches to test case generation for agent-based models, rather than to act as a concrete definition. As a result, there may be some overlap between categories within a taxonomy. A separate bibliography for the reviewed literature is included as an Appendix, where citations to this bibliography are prefixed with ``P'' e.g. [P1].
\input{sections/subsections/informationArtifact.tex}
\input{sections/subsections/generationMechanism.tex}
\input{sections/subsections/testOracle.tex}
\input{sections/subsections/testAdequacy.tex}
\input{sections/subsections/levelOfAbstraction.tex}
\input{sections/subsections/publicationTrends.tex}
\section*{Acknowledgements}
Walkinshaw and Hierons are funded by the EPSRC CITCOM grant EP/T030526/1.
\section{Introduction}
In recent years, computational models have become increasingly popular in a variety of domains, with applications ranging from economic modelling \cite{liu2020interbank} to public health \cite{tracy2018agent}. The ability to model and understand vast quantities of data, coupled with the increased availability of data, have helped to establish the computational model as an essential research tool. Agent-based models are one variety of computational model that are used to simulate complex scenarios involving many individuals, often to inform social and economic policies, including national responses to the recent COVID-19 pandemic \cite{flaxman2020report, panovska2020determining, kerr2020covasim}. With such critical applications, poor modelling assumptions or software faults can lead to serious repercussions.
Software testing is an essential software engineering practice that aims to ensure the quality of software products, including the identification and removal of faults. The time and effort necessary to test software and achieve such quality are significant, often accounting for over 50\% of total project costs \cite{ramler2006economic}. In recent years, a variety of techniques have been proposed to alleviate the cost of testing, from search-based testing \cite{mcminn2004search} to fuzzing \cite{godefroid2008automated}, that automate or partially-automate various testing activities such as the creation of test cases.
Of all software testing activities, perhaps the most labour-intensive and time-consuming activity is the creation of test cases \cite{prasanna2005survey}. As a consequence, the topic of test case generation has received significant attention in recent years, leading to a variety of surveys, techniques and tools supporting the automatic or semi-automatic derivation of test suites \cite{anand2013orchestrated}. However, for many real-world applications, it remains unclear how effective these tools and techniques are, with areas such as scientific software requiring further empirical study to establish their applicability \cite{kanewala2014testing}.
Agent-based modelling is one area where the applicability of existing test case generation techniques is unclear. Due to their often non-deterministic nature coupled with their complex emergent behaviour, agent-based models present significant challenges for test case generation \cite{russell2002artificial, luck2005agent}. For example, in response to the ongoing COVID-19 pandemic, the UK government has used an agent-based model developed at Imperial College London to inform an economic policy estimated to cost \pounds192.3bn over the 2020-2021 financial year \cite{flaxman2020report, obr2020coronavirus:online}. Despite the importance of testing such critical software, the model contains only basic integration tests configured to a specific set of parameters \cite{mrcideco79:online}. To this end, there have been several recent calls for improved techniques for validation and testing of predictive models \cite{squazzoni2020computational, wynants2020prediction}.
Clearly, there is a need for tools and techniques that make software testing an accessible practice for both researchers and practitioners regardless of their discipline. In this review we aim to identify the current state-of-the-art techniques and future research directions for test case generation in agent-based modelling, paying particular attention to their potential real-world applications. In order to achieve a fair and comprehensive survey of the existing test case generation literature in the agent-based modelling field, this paper reports the results of a systematic literature review in accordance to the guidelines proposed by Kitchenham and Charters \cite{kitchenham2007guidelines}.
The remainder of this work is organised as follows: in Section 2, we provide an overview of agent-based models and test case generation. In Section 3, we present the research methodology including the research questions and review procedure. In Section 4, the results of the systematic literature review are presented, followed by a discussion of some particularly interesting observations in Section 5. In Section 6 we conclude the review and present potential future directions for research.
\subsection{RQ3: How are verdicts assigned to generated test cases?}
Due to the exploratory nature of an agent-based model, the ``correct'' outcome of an execution is often unknown and it may be particularly difficult to assign a verdict to a test case. This is known formally as the test oracle problem, where the test oracle is the mechanism behind the decision. In the reviewed literature, three of the four different categories of oracle introduced by McMinn et al. \cite{mcminn2015oracle} have been identified: specified, human and derived (see Figure \ref{fig:testOracleTaxonomy}).
\begin{figure}[h!]
\centering
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=0.35\linewidth]{images/taxonomies/testOracle.png}
\caption{}
\label{fig:testOracleTaxonomy}
\end{subfigure}
\begin{subfigure}{\textwidth}
\centering
\includegraphics[width=0.6\linewidth]{images/results/testOracle/oracle_pie.png}
\caption{}
\label{fig:testOracleDistribution}
\end{subfigure}
\caption{A taxonomy (a) and the distribution (b) of test oracles that appear in the literature.}
\end{figure}
\subsubsection{Specified}
Specified test oracles are mechanisms that assign test verdicts based on a formal specification of the system-under-test \cite{mcminn2015oracle}. Formal specifications, including design documents, state machines, contracts and assertions, and ontologies, express the expected behaviour of the system-under-test under various conditions, defining what should happen and implicitly what should not. Given that the majority of the reviewed approaches involve a formal specification (discussed in Section 4.1.1), it follows that most of the approaches also include a specified oracle as shown in Figure \ref{fig:testOracleDistribution}. Whilst the oracle information is present in these approaches, several have not explained how it can be compared to the test execution in order to assign a verdict [P3, P12, P14, P19, P20, P22, P23]. On the other hand, one of the reviewed approaches focuses on the automatic generation of partial, passive test oracles from Prometheus agent design documents for BDI agent-based systems alongside a technique for test case generation [P7]. Using this approach, a selection of faults in BDI agent plans, events and beliefs can be automatically detected in an implementation of a BDI agent system, based on the expected behaviour specified in Prometheus design documents.
\subsubsection{Derived}
Derived test oracles are mechanisms that assign test verdicts based on information derived from artifacts of the system-under-test, rather than information that is explicitly defined in a formal specification \cite{mcminn2015oracle}. Whilst most of the reviewed approaches use a formal specification as an information artifact, three approaches focus on generating test cases that satisfy soft goals, or non-functional requirements, of the system-under-test that cannot be formally expressed [P15, P16, P24]. To address this issue, fitness functions have been used to approximate the fulfilment of a particular soft goal, acting as an oracle that measures whether a certain quality or non-functional requirement is satisfied by agents. Fitness functions are typically derived from qualities of the system-under-test, rather than a feature of a formal specification. For example, the efficiency of an agent cleaning a grid can be measured by a pair of fitness functions [P16]:
$$f_{power}=\frac{1}{Total \: power \: consumption}$$ $$f_{obs}=\frac{1}{Number \: of \: obstacles \: encountered}$$ In this scenario, a threshold value can be set such that any agent that exceeds the threshold is to be considered efficient, providing an approximation of correctness and a goal for test case generation.
\subsubsection{Human}
A human oracle is used where no information artifact can be used as the basis of a test oracle, and instead, a human must make the decision as to whether a test case should pass or fail using domain expertise \cite{mcminn2015oracle}. In the reviewed literature, all of the approaches use a formal or informal specification as an information artifact and thus provide at least a derived oracle. However, many of the approaches focus entirely on how test cases are generated and not how verdicts are assigned. Even in the case that the information artifact contains the information necessary to determine whether an execution is as expected, human judgement may still be required to compare the execution to the specification. Whilst this comparison is straightforward in some instances, for models with visual or otherwise complex outputs, the comparison may be subjective. In such cases, the distinction between different forms of an oracle is less clear. For example, in one state machine-based approach, a human must observe the visual output produced by the system-under-test and compare it with the expected output expressed by a stream X-machine specification [P12]. In this situation, the type of oracle used is unclear: the stream X-machine specification is a specified oracle, but it requires a subjective human decision which could itself constitute a human oracle. Whilst only one approach explicitly mentions the need for human comparison [P12], several others may also require the same treatment [P3, P14, P19, P20, P22, P23].
\begin{table}[htbp]
\centering
\begin{tabular}{ll}
\toprule
Test Oracle Type & Papers \\ \midrule
Specified & {[}P1-P14, P17-P23{]} \\
Derived & {[}P15, P16, P24{]} \\
Human & {[}P12{]} \\
\bottomrule
\end{tabular}
\caption{Summary of test oracles used in the evaluated literature.}
\end{table}
\subsection{RQ4: How is the adequacy of a generated test suite measured?}
An effective approach to test case generation should be guided by some measure of how \emph{good} a test suite is \cite{mathur1994empirical}, known more formally as a measure of test adequacy. In the literature, we have identified two core measures of test adequacy that have been used as the goal for test case generation: coverage criteria and fitness functions (see Figure \ref{fig:testAdequacyTaxonomy}).
\begin{figure}[h!]
\centering
\begin{subfigure}{0.9\textwidth}
\centering
\includegraphics[width=0.6\linewidth]{images/taxonomies/testAdequacy.png}
\caption{}
\label{fig:testAdequacyTaxonomy}
\end{subfigure}
\begin{subfigure}{0.9\textwidth}
\centering
\includegraphics[width=0.5\linewidth]{images/results/testDataAdequacy/adequacy_criteria_pie.png}
\caption{}
\label{fig:testAdequacyDistribution}
\end{subfigure}
\caption{A taxonomy (a) and the distribution (b) of test adequacy criteria that appear in the literature.}
\end{figure}
\subsubsection{Coverage criteria}
Coverage criteria are a measure of test adequacy that focus on how thoroughly a test suite covers a specified portion of the information artifact that represents some functionality in the system-under-test. Where a state machine specification is available, test adequacy is often measured using state and transition coverage [10, 12], or alternatively, test adequacy can be measured using statement coverage, branch coverage and path coverage \cite{zhu1997software}. However, the latter criteria are not directly applicable to agent-based models as the notions of branches and paths from other programming paradigms do not translate to agent-based models clearly \cite{low1999automated}. To address this, several works have proposed coverage criteria specific to agent-based frameworks and testing approaches, including the following particularly interesting criteria: BDI artifact coverage, data-flow coverage and ontology coverage.
\emph{BDI artifact coverage} - A series of belief-desire-intention coverage criteria have been defined which draw parallels to existing coverage criteria such as branch and statement coverage, but instead apply to the equivalent BDI artifacts: plan-based and node-based criteria [P19]. The group of node-based criteria are designed to generate test cases that cover all possible states of the system, whereas the group of path-based criteria aim to test how agent plans are related to one another. Additionally, a subsumption hierarchy is defined, indicating the minimal necessary coverage criteria for generating a satisfactory test suite. Alternative definitions for coverage of BDI events, beliefs, context conditions and plans have also been given in later works [P7], including a series of state machine-based coverage criteria that are applicable when BDI agents can be modelled as an extended state machine [P10].
\emph{Data-flow coverage} - An alternative test case generation technique has been developed that focuses on data flow \cite{cavarra2011data}. In this approach, a series of data flow coverage criteria are introduced to enable data flow testing of multi-agent abstract state machines, focusing on the interaction of agent components. Data flow testing \cite{rapps1985selecting} concerns a family of testing techniques that explore how variables are defined and subsequently used in a system to reveal unreasonable data usage, such as declaring but never using a variable. Typically, data flow testing is based on the control flow graph of the system-under-test, however, an agent-based model does not have a one-to-one mapping from code to control flow graph and therefore traditional approaches to data flow testing are not directly applicable. Instead, this approach adapts traditional data flow testing techniques to handle abstract state machine representations of agent-based models, including a series of abstract state machine-specific data flow coverage criteria [P18].
\emph{Ontology coverage} - Test case generation has also been guided by the coverage of an ontology in ontology-based approaches [P13]. In this approach, concepts and instances from an ontology that have not been previously selected for test cases are prioritised to provide greater coverage of the input space, instead of focusing on a narrow subset of information. In the case that an instance must be reused in a test case, the most infrequently used instance is selected in order to maximise the diversity of the test suite. As a consequence of the thorough exploration of the input space, the fault detection capability of the generated test suite is shown to be superior to manual testing. In other approaches, coverage of concepts derived from context ontologies [P21] and coverage of the possible test cases that can be formed from a given testing scenario derived from a domain ontology [P20] have been used as measures of test data adequacy.
\subsubsection{Fitness functions}
A fitness function evaluates how well a particular solution satisfies a problem-specific objective, providing a ranking over all potential solutions \cite{whitley1994genetic}.
For the activity of test case generation, a fitness function can be used to measure the adequacy of a test case with respect to a problem-specific objective, including satisfaction of stakeholder goals such as collision avoidance [P16] or the ability to reveal known faults [P21]. In the literature, fitness functions have also been defined to capture soft goals of agent-based systems, where the adequacy of a generated test suite is measured by the fulfilment of non-functional requirements such as efficiency and reliability [P15, P16, P24]. Consequently, fitness functions have been used to guide evolutionary approaches to test case generation in which genetic operators are applied to evolve test suites that are challenging with respect to a defined objective, as described in Section 4.2.4. As part of the eCAT test case generation tool for multi-agent systems, mutation adequacy score has been used as a measure of adequacy for test suites generated using the \lstinline{EVOL-MUTATION} approach [P1]. Mutation adequacy is the fraction of injected faults that are detected by the test suite, summarising the ability of a test suite to detect known faults in the system-under-test \cite{jia2010analysis}.
\begin{table}[htbp]
\centering
\begin{tabular}{ll}
\toprule
Test Data Adequacy Criterion & Papers \\ \midrule
Coverage criteria & {[}P1, P3-P10, P12, P13, P18-P22{]} \\
Fitness function & {[}P1, P15, P16, P21, P24{]} \\
\bottomrule
\end{tabular}
\caption{Summary of test data adequacy critria used in the evaluated literature.}
\end{table}
\subsection{RQ5: What level of abstraction do the generated test cases target?}
Agent-based models comprise multiple levels of abstraction, from the individual units that form a single agent up to a society composed of multiple interacting agents, each containing different functionality and testing objectives. Nguyen et al. have distinguished five different testing objectives that should be considered when testing a typical agent-based model: unit, agent, integration, society and testing \cite{nguyen2009testing} as shown in Figure \ref{fig:levelsOfAbstractionTaxonomy}.
\begin{figure}[h!]
\centering
\begin{subfigure}{0.9\textwidth}
\centering
\includegraphics[width=0.6\linewidth]{images/taxonomies/levelOfAbstraction.png}
\caption{}
\label{fig:levelsOfAbstractionTaxonomy}
\end{subfigure}
\begin{subfigure}{0.9\textwidth}
\centering
\includegraphics[width=0.9\linewidth]{images/results/levelsOfAbstraction/levels_of_abstraction.png}
\caption{}
\label{fig:levelsOfAbstractionDistribution}
\end{subfigure}
\caption{A taxonomy (a) and the distribution (b) of levels of abstraction that have been tested in the literature.}
\end{figure}
\subsubsection{Unit}
The lowest level of abstraction to consider when testing an agent-based model is that of the constituent units that make up an agent, such as goals, plans and rules. Several approaches from the reviewed literature have focused on the generation of unit tests for agent-based systems alone [P7, P19, P22], whilst a further three approaches have also considered additional levels of abstraction, such as the agent-level [P2, P6, P15]. For example, the beliefs, plans and events of a BDI agent-based model have been subjected to boundary value analysis and equivalence class partitioning in order to generate variable-value pairs as test cases that exercise these units of the system-under-test [P22].
\subsubsection{Agent}
Agent-level testing concerns the integration of the units that compose an agent (covered individually in unit testing), such as goals and plans, that collectively make up the behaviour of an individual. A significant majority of the approaches from the reviewed literature focus on generating test cases that exercise agent-level behaviour alone or in addition to other levels of abstraction [P2, P5, P6, P8, P10-P13, P15, P16, P18, P21, P23, P24]. For example, using the random testing tool QuickCheck, test cases focusing on unit and agent-level behaviour have been generated for property-based testing \cite{fink1997property} of agent-based models [P2].
\subsubsection{Integration}
Integration-level testing concerns the interaction between individual agents and their environments towards cooperative behaviour. Several works have focused on both the agent-level and integration-level of interaction [P6, P10, P18, P23], where test cases target both individual agent behaviour and behaviour that is the result of agent-to-agent or agent-to-environment interactions. In addition, a pair of techniques focusing solely on integration-level testing have been proposed [P9, P14]. For example, BDI agents have been modelled as extended state machines where pair-wise interaction can be captured by the product machine composed of two individual extended state machines. As a consequence, existing testing techniques can be used to generate both agent and integration-level test cases corresponding to testing individual and cooperative behaviour [P10].
\subsubsection{Society}
Perhaps the most complicated level of abstraction to test is the society level, concerning the complex emergent behaviours that are the product of the lower levels of abstraction. Emergence in agent-based systems can be described as ``the arising of novel and coherent structures, patterns and properties during the process of self-organisation in complex systems'', as defined by Goldstein \cite{goldstein1999emergence}. Society level test case generation has received little attention with only three approaches considering this level of abstraction [P4, P6, P23]. One of the approaches targets test case generation for holonic multi-agent systems \cite{gerber1999holonic}, focusing on testing emergent behaviours that have been acquired over time in order to deal with challenging and previously unseen situations [P4]. Another approach introduces a testing framework for agent-based models that supports the testing of agent, integration and society-level testing through the generation of mock agents from XML protocol specifications that trigger a testing scenario [P23]. The third approach involves conformance testing of agent-interaction protocols specified in Maude, where the specified transition sequences are extracted and used to conduct testing at the unit, agent, integration and society levels of testing [P6].
\subsubsection{Acceptance}
The final level of abstraction to consider when testing an agent-based model is acceptance. Acceptance testing is a well-known software testing practice that involves testing whether high-level acceptance criteria are met, such as the fulfilment of user requirements \cite{miller2001acceptance}. In the reviewed literature, a couple of approaches for acceptance testing have been proposed that focus on generating alternative versions of the same user-specified scenario. One approach focuses on generating different variations of a specified marine safety and security scenario as test cases [P17], whilst the other approach is a more general testing framework that enables acceptance testing of scenarios specified as part of the Prometheus agent development methodology [P3]. In addition, evolutionary approaches to test case generation have been proposed that are suitable for requirements testing [P16, P24]. In these approaches, stakeholder goals are formulated as fitness functions which act as both the generation mechanism, as discussed in Section 4.2.4, and an oracle for evaluating whether a particular execution meets the stakeholder goal, as discussed in Section 4.3.2. This approach has also been used as a complementary testing methodology for the AgentTest specification language that covers unit, agent and acceptance level testing [P15].
\begin{table}[htbp]
\centering
\begin{tabular}{ll}
\toprule
Level of abstraction & Papers \\ \midrule
Unit & {[}P2, P6, P7, P15, P19, P22{]} \\
Agent & {[}P2, P5, P6, P8, P10-P13, P15, P16, P18, P21, P23, P24{]} \\
Integration & {[}P6, P9, P10, P14, P18, P23{]} \\
Society & {[}P4, P6, P23{]} \\
Acceptance & {[}P3, P15-P17, P24{]} \\
\bottomrule
\end{tabular}
\caption{Summary of levels of abstraction tested in the evaluated literature.}
\end{table}
\subsection{Publication trends}
In this section, the publication trends collected during the data extraction are presented, including publication venues and the results of the quality assessment. Notice that the quality assessment presented in Table \ref{tab:qualityAssessmentScores} does not include any literature with $quality \leq 1.5$, as any study below this threshold would be excluded due to a perceived lack of quality in accordance to the quality assessment procedure discussed in Section 3.6.
\begin{table}[!htbp]
\footnotesize
\centering
\begin{tabular}{lc}
\toprule
Venue & Count \\ \midrule
International Conference on Autonomous Agents and Multi-Agent Systems & 4 \\
International Workshop on Agent-Oriented Software Engineering & 2 \\
IEEE/ACS International Conference on Computer Systems and Applications & 2 \\
International Conference on Software Engineering and Knowledge Engineering & 1 \\
International Conference on Informatics in Control, Automation and Robotics & 1 \\
International Conference on Agents and Artificial Intelligence & 1 \\
IEEE Transactions on Software Engineering & 1 \\
IEEE International Conference on Software Maintenance & 1 \\
Summer Simulation Conference & 1 \\
Engineering Applications of Artificial Intelligence & 1 \\
Science and Information Conference & 1 \\
Federated Conference on Computer Science and Information Systems & 1 \\
International Conference on Networking and Advanced Systems & 1 \\
International Symposium on Search Based Software Engineering & 1 \\
Asia-Pacific Software Engineering Conference & 1 \\
Journal of Computer Science & 1 \\
Neurocomputing & 1 \\
Security Informatics & 1 \\
Formal Aspects of Computing & 1 \\
\bottomrule
\end{tabular}
\caption{Publication venues of reviewed work.}
\label{tab:venues}
\end{table}
As shown in Table \ref{tab:venues}, the reviewed literature has been published in a variety of conferences and journals spanning multiple subject areas. Whilst this demonstrates that agent-based models and test case generation are multidisciplinary research topics, it also highlights the lack of an established publication channel.
\begin{table}[!htbp]
\footnotesize
\centering
\begin{tabular}{llllll}
\toprule
ID & QA1 & QA2 & QA3 & QA4 & Total \\ \midrule
{[}P1{]} & 1 & 1 & 0 & 1 & 3 \\
{[}P2{]} & 1 & 1 & 0.5 & 1 & 3.5 \\
{[}P3{]} & 1 & 1 & 0.5 & 0.5 & 3 \\
{[}P4{]} & 1 & 1 & 1 & 1 & 4 \\
{[}P5{]} & 1 & 1 & 0.5 & 0.5 & 3 \\
{[}P6{]} & 1 & 1 & 0.5 & 0.5 & 3 \\
{[}P7{]} & 1 & 1 & 0.5 & 1 & 3.5 \\
{[}P8{]} & 1 & 1 & 0.5 & 0.5 & 3 \\
{[}P9{]} & 1 & 0.5 & 1 & 0 & 2.5 \\
{[}P10{]} & 1 & 1 & 1 & 0 & 3 \\
{[}P11{]} & 1 & 1 & 0.5 & 0.5 & 3 \\
{[}P12{]} & 1 & 1 & 0.5 & 0.5 & 3 \\
{[}P13{]} & 1 & 1 & 0.5 & 0.5 & 3 \\
{[}P14{]} & 1 & 0.5 & 0.5 & 0.5 & 2.5 \\
{[}P15{]} & 1 & 1 & 0.5 & 0.5 & 3 \\
{[}P16{]} & 1 & 1 & 1 & 0.5 & 3.5 \\
{[}P17{]} & 1 & 1 & 0.5 & 1 & 3.5 \\
{[}P18{]} & 1 & 1 & 0.5 & 0.5 & 3 \\
{[}P19{]} & 1 & 1 & 0.5 & 0.5 & 3 \\
{[}P20{]} & 0.5 & 0.5 & 0.5 & 0.5 & 2 \\
{[}P21{]} & 0.5 & 0.5 & 0.5 & 0.5 & 2 \\
{[}P22{]} & 1 & 1 & 0.5 & 0.5 & 3 \\
{[}P23{]} & 1 & 1 & 1 & 0.5 & 3.5 \\
{[}P24{]} & 1 & 1 & 0.5 & 0.5 & 3 \\
\bottomrule
\end{tabular}
\caption{Quality assessment scores of the reviewed literature.}
\label{tab:qualityAssessmentScores}
\end{table}
As shown in Table \ref{tab:qualityAssessmentScores}, the majority of the reviewed publications scored highly in QA1 and QA2 as a result of having well-defined aims that were clearly met. However, many of the publications did not include a thorough evaluation of the proposed technique and did not include sufficient details for replication. Consequently, studies did not score as highly in QA3 and QA4 in general.
\subsection{RQ1: What artifacts are used to drive test case generation?}
All of the reviewed approaches use some form of specification as an information artifact to drive test case generation. The types of information artifact can be divided into two different categories: formal and informal artifacts.
\begin{figure}[h!]
\centering
\begin{subfigure}{0.9\textwidth}
\includegraphics[width=1\linewidth]{images/taxonomies/informationArtifactHierarchy.png}
\caption{}
\label{fig:informationArtifactTaxonomy}
\end{subfigure}
\begin{subfigure}{0.9\textwidth}
\includegraphics[width=1\linewidth]{images/results/informationArtifact/formal_vs_informal_bar.png}
\caption{}
\label{fig:informationArtifactDistribution}
\end{subfigure}
\caption{A taxonomy (a) and the distribution (b) of information artifacts that appear in the literature.}
\end{figure}
\subsubsection{Formal artifacts}
Formal artifacts express unambiguous behaviour or functionality of the system-under-test in a well-defined format. Four different forms of formal artifact have been identified in the literature: design documents, state machines, contracts and assertions, and ontologies as shown in Figure \ref{fig:informationArtifactTaxonomy}.
\emph{Design documents} - Design documents comprise a series of structured artifacts that are used in development methodologies such as Prometheus \cite{padgham2002prometheus} and Tropos \cite{bresciani2004tropos} to specify requirements, goals and scenarios for the design and development of agent-based systems. These documents contain details about the system-under-test that span the system architecture, agent communication protocol and testing goals, often taking the form of a directed graph or an AUML document. AUML is an agent-based variant of the Unified Modelling Language \cite{bauer2001agent} that is predominantly used to specify agent interaction protocols describing the semantics and structure of agent communication [P4, P9, P11]. Design documents from the Prometheus design methodology can be enriched with ``testing descriptors'' that map design variables to implementation variables, highlighting the necessary information for test input generation [P22]. Prometheus design documents have also been used as the basis for test generation in [P3, P7], and similarly, Tropos design documents have appeared in [P1, P14]. Prior to the development of the Tropos design methodology, BDI agent plans had also been used to drive test case generation [P19].
\emph{State machines} - An alternative approach is to develop state machine-based agents. In these approaches the expected behaviour of an agent-based system can be expressed using a state machine, such as an X-machine \cite{eilenberg1974automata} or stream X-machine \cite{laycock1993theory}, which enables existing automata testing theory to be applied. Regardless of their exact form, state machine specifications all contain the same basic components; states joined by transitions, where a transition corresponds to an action that moves the system from one state to another. A path through the specification corresponds to an execution of the system-under-test, and therefore a state machine specification expresses the paths through the system that may be tested \cite{chow1978testing}. A variety of different state machine information artifacts have been seen in the literature, including extended and abstract state machines [P10, P18], as well as X-machines and stream X-machines [P8, P12].
\emph{Contracts and assertions} - A different approach to formally specifying the expected behaviour of an agent-based system is the use of contracts and assertions. In contract-based testing, a contract specifies a number of pre-conditions, post-conditions, and invariants that assert the expected behaviour of a system before, after, and during execution \cite{heckel2005towards}. These conditions are a form of assertion - a logical statement that places a constraint on some state of computation \cite{korel1996assertion} - that can be used as the basis for both test case generation and an oracle. Using the formal programming language Maude \cite{clavel2002maude}, [P5] and [P6] have used rewrite rules to specify the functional behaviour of an agent-based system in terms of pre-conditions and post-conditions. The functional programming language Haskell has also been used as an executable specification to assert properties that must be satisfied during an execution and drive test case generation [P2]. Additionally, structured test scenarios containing assertion-like constructs that describe the expected outcome for a given scenario have been used as the basis for test case generation in [P17] and [P23].
\emph{Ontologies} - Ontologies are another form of formal artifact that have appeared in several of the reviewed techniques. An ontology is a collection of concepts from a given domain, including properties of the concepts and relationships between them \cite{calero2006ontologies}. In the context of agent-based models, two forms of ontology have been used as an information artifact: ``interaction ontologies'' and ``domain ontologies'' [P13]. An interaction ontology defines the concepts, relationships and interaction semantics of a model as entities and relations, providing a specification of the system that dictates how agents can interact. Whereas a domain ontology contains information specific to a domain or subject, for example, ontologies describing a production system [P20] or an environment [P21]. Using a technique known as ontology alignment, publicly available domain ontologies can be harnessed to supplement existing ontologies, enhancing the information artifact used for test case generation [P13].
\begin{table}[htbp]
\centering
\begin{tabular}{ll}
\toprule
Formal Information Artifact Type & Papers \\ \midrule
Design documents & {[}P1, P3, P4, P7, P9, P11, P14, P19, P22{]} \\
State machine & {[}P8, P10, P12, P18{]} \\
Contracts and assertions & {[}P2, P5, P6, P17, P23{]} \\
Ontologies & {[}P1, P13, P20, P21{]} \\
\bottomrule
\end{tabular}
\caption{Summary of formal information artifacts used in the evaluated literature.}
\end{table}
\subsubsection{Informal artifacts}
Informal artifacts concern qualitative features of the system-under-test that are difficult to express formally. In comparison to formal artifacts, informal artifacts appeared significantly less in the literature, as shown in Figure \ref{fig:informationArtifactDistribution}, and therefore we do not divide them further. However, it is important to distinguish between formal and informal artifacts based on the type of information they contain and the techniques that use them.
Informal artifacts typically capture qualitative features of an agent-based model such as stakeholder requirements expressed as soft goals. A soft goal is a goal that does not have a clear satisfiability definition and is often used for modelling non-functional requirements such as security, usability and flexibility \cite{padgham2008unified}. Therefore, it is not possible to definitively prove whether a soft goal has been satisfied as the threshold for satisfaction is subjective. Instead, judgements are informed by an expert opinion or a fitness function that approximates the soft goal, transforming the task of test case generation into a search problem where the objective is to find test cases that satisfy the fitness function.
Informal artifacts do not provide the same level of detail about the system-under-test as their formal counterparts, but instead, approximate desired qualities of the system-under-test that cannot be expressed formally. As a consequence, informal artifacts are generally expressed using natural language rather than a formal specification language, however, soft goals may also appear in design documents as additional information \cite{padgham2008unified}. In the reviewed literature, informal artifacts have been used to drive evolutionary approaches to test case generation, where fitness functions representing soft goals are used to continuously evolve more challenging test cases [P16, P24]. Examples of soft goals seen in the agent-based test case generation literature include reliability, efficiency and robustness [P15, P16].
\subsection{RQ2: What approaches are used to generate tests from the artifacts?}
Four different groups of generation mechanism have been identified in the agent-based literature: path traversal, information extraction, random testing and rule-based, as shown in Figure \ref{fig:generationMechanismTaxonomy}.
\begin{figure}[h!]
\centering
\begin{subfigure}{\textwidth}
\includegraphics[width=1\linewidth]{images/taxonomies/generationMechanism.png}
\caption{}
\label{fig:generationMechanismTaxonomy}
\end{subfigure}
\begin{subfigure}{\textwidth}
\includegraphics[width=1\linewidth]{images/results/generationMechanism/mechanisms_bar.png}
\caption{}
\label{fig:generationMechanismDistribution}
\end{subfigure}
\caption{A taxonomy (a) and the distribution (b) of generation mechanisms that appear in the literature.}
\end{figure}
\subsubsection{Path-traversal}
Path-traversal generation mechanisms are approaches that explore the different possible paths through a traversable representation of the system-under-test, such as a Petri net \cite{petri1962kommunikation} or a state machine \cite{gill1962introduction}, where different paths through the system correspond to different possible executions and therefore test cases. The identified path-traversal generation mechanisms can be further divided into three categories based on their subject: state machines, Petri nets and AUML.
\emph{State machine} - Several of the reviewed approaches adopt state machine-based specifications, where different paths correspond to different sequential behaviours of the system-under-test. State machines are supported by a range of automated testing techniques that make them particularly effective for specifying and testing complex systems such as agent-based models \cite{lee1996principles}. Informally, these techniques derive sequences of inputs from the specification representing paths through the system, that are used to determine whether the implementation conforms to the specification. One such technique known as the W-method \cite{chow1978testing} has been used as a generation mechanism for agent-based systems that are equipped with X-machine specifications [P8, P12]. A similar approach has also been used to generate test cases from an extended state machine specification \cite{cheng1993automatic} for the unit and pair testing of BDI agents [P10]. Test cases have also been generated using a model-checker to identify paths through a multi-agent abstract state machine that satisfy a series of data-flow coverage criteria [P18].
\emph{Petri net} - Path-traversal mechanisms also appear in the literature where Petri nets have been used as a graphical representation of the system-under-test. A Petri Net is an abstract, formal model used to describe the flow of information in systems, where dynamic components are captured by tokens that can move between states, enabling asynchronous and concurrent behaviour to be captured \cite{peterson1977petri}. Given that agents are often required to perform multiple tasks concurrently \cite{shehory2001evaluation}, Petri nets are a particularly effective tool for specifying and testing the behaviour of agent-based systems. In the reviewed literature, Petri nets have been used to represent and test the structure of BDI agent plans [P5]. Recursive coloured Petri nets \cite{haddad1999theoretical} have also been used, enabling multiple levels of abstraction of an agent-based model to be captured and accessed via recursive unfolding of nested Petri nets [P11]. To automate the generation of test cases from Petri nets, algorithms have been used that first transform the nets into a machine-readable format such as a state table, before traversing the individual paths and extracting their corresponding input sequences [P11].
\emph{AUML} - Another form of graphical representation appearing in the literature that facilitates test generation through path-traversal is AUML. Similar to Petri Nets, AUML also provides support for concurrent behaviour amongst other agent-related characteristics \cite{juneidi2010survey}, however, AUML is arguably more accessible due to its similarity to UML \cite{peres2005experiencing}. In the reviewed literature, algorithms have been presented that convert AUML sequence diagrams into a traversable format in order to identify paths through the system corresponding to test cases. For example, in an approach for testing holonic multi-agent systems \cite{gerber1999holonic}, a hierarchical AUML sequence diagram is converted into a traversable format known as a graph sequence diagram. A series of algorithms are presented to identify the different paths through the graph, before extracting the pre-conditions and post-conditions necessary to form a covering test suite [P4]. Later this approach was extended to include OCL annotations, allowing tests to identify which agent caused an error in a scenario or interaction that contains multiple agents [P9]. AUML sequence diagrams have also been translated into recursive coloured Petri nets, using its corresponding state table as the basis for test case generation [P11].
\subsubsection{Information extraction}
Information extraction generation mechanisms comprise approaches to test case generation that are driven by the extraction of test information from an information artifact, such as the extraction of variable-value pairs from design documents. In the literature, information extraction techniques have generally been applied to two different forms of information artifact: design documents and ontologies (see Figure \ref{fig:generationMechanismDistribution}).
\emph{Design documents} - As discussed in Section 4.1.1, design documents comprise a series of artifacts that are used for the design of agent-based models in agent development methodologies. For testing, useful information can be extracted from design documents and used to form test cases, such as agent goals and plans. For example, by following a structured approach, integration test suites can be derived from Tropos design documents concerning the system goals of an agent-based system [P14]. Design documents can also be enriched with test descriptors to provide additional information for test case generation, such as highlighting variables that influence a testing scenario. With this additional information, a test case specification can be extracted from design documents that highlights any initialisation procedures and relevant variables involved in a scenario that can be modified to form new test cases [P3]. Whilst this approach is only semi-automatic, a similar approach automates the extraction of variable-value pairs from Prometheus design documents. This is achieved using equivalence class partitioning and boundary value analysis to generate a minimal set of values for each influencing variable [P22]. Later, this approach was used alongside a model-based test oracle generation technique to provide automatic unit testing for agent systems [P7].
\emph{Ontology} - Another form of artifact containing useful information that may be extracted for testing is an ontology. An ontology \cite{smith2012ontology} is a formal model of knowledge representation for a given domain, detailing a set of concepts and how they are related to facilitate the exchange and organisation of knowledge \cite{moser2010ontology}. In the context of an agent-based model, ontologies are used to model a vast variety of information ranging from properties of agents and their context to interaction protocols. It follows that information extraction approaches have been proposed in the literature to utilise the information hosted by ontologies for the purpose of test case generation. One approach to ontology-based test case generation focusing on message-passing agents combines both agent interaction ontologies and domain ontologies, generating test cases in the form of messages that conform to an interaction protocol. The task of test generation then consists of generating meaningful message content that is to be sent by a \lstinline{TesterAgent} to an agent-under-test in order to provoke a particular behaviour. To generate the message content, valid and invalid inputs are generated by selecting an existing instance of a concept from the ontology that conforms to OCL constraints. Alternatively, if there is no existing concept then a new one is generated according to a series of input generation rules which will be discussed in Section 4.2.4. A technique known as ontology alignment (or ontology matching) \cite{euzenat2007ontology} is also used to expand existing ontologies by supplementing them with additional information from publicly available ontologies [P13]. Ontology-based techniques have also been used to generate test cases that focus on industry specific applications [P20] and the context of an agent-based system [P21].
\subsubsection{Random testing}
Random testing \cite{hamlet2002random} is a well-known testing strategy where inputs are generated at random to exercise a system-under-test. As the name suggests, the generation mechanism behind random testing is the process by which random inputs are generated or selected. In comparison to other test case generation mechanisms, random testing has not received as much attention in the agent-based testing literature.
Random generation of numerical inputs is the most basic example of random testing, where in general, generation of a test case corresponds to randomly selecting a series of numerical values as input to the system-under-test. Depending on the system-under-test, the domain could be restricted to a user-defined range of numerical values or a particular set such as the natural or real numbers. Perhaps due to the challenging characteristics of an agent-based model such as communication through message-passing \cite{padgham2005developing}, richer inputs are required and therefore random generation of numerical inputs has not been used as a generation mechanism alone.
Instead, this approach has been used to support other forms of generation mechanism, such as an ontology-based technique, by selecting numerical values within a range specified as an OCL constraint to form valid and invalid inputs [P13]. Random testing can also involve sampling non-numerical, user-specified domains to generate more sophisticated test cases. For example, one approach uses random testing to sample a domain data model which defines the range and structure of messages permitted by a specific agent-interaction protocol [P1]. In addition, QuickCheck \cite{claessen2011quickcheck} - a library for random testing of program properties - has also been used to check whether a series of formally specified properties hold in an agent-based system. To achieve this, QuickCheck can automatically generate simple test inputs and provides the functionality to define custom test data generators for more complex inputs. In this work, QuickCheck is demonstrated on a simple SIR (susceptible-infectious-recovered) model \cite{bailey1975mathematical} and on a more complex model of an artificial society, SugarScape \cite{epstein1996growing}. In the former, QuickCheck generates a series of agent populations as test cases, whilst in the latter, a custom data generator is used to generate complex SugarScape environments [P2].
\subsubsection{Rule-based}
The final group of test generation mechanisms used in the agent-based literature concern rule-based approaches. Rule-based generation mechanisms involve a series of user-defined or automatically derived rules that describe how a test case can be manipulated to create new ones. Two types of rule-based generation mechanism have been identified in the literature: transformation rules and genetic operators.
\emph{Transformation rules} - Transformation rule-based generation mechanisms concern test generation techniques that are driven by a series of rules describing how one test case can be transformed to form a new one, by a logical transformation. User-defined transformation rules have been applied to testing agent-based models for infrastructure protection and emergency response simulation, targeting models of marine safety and security operations in particular [P17]. To this end, structured textual descriptions of example scenarios, known as vignette specifications, are modified by a library of user-defined transformation rules in order to produce different variations of the same scenarios as test cases. These transformation rules can be applied to selected variables in the vignette specification such as the number of boats involved in a scenario, their positions and their physical dimensions. Similar approaches have also used transformation rules as a generation mechanism to manipulate ``test data models'' to create new test data [P21], as well as applying them to XML definitions of an interaction protocol to form ``mock agents'' [P23].
A number of the reviewed techniques used Maude rewrite rules as a generation mechanism. In rewriting logic \cite{meseguer1992conditional}, a rewrite rule is a procedure which describes how an object can be transformed into another object. Maude \cite{clavel2002maude} is a formal language based on rewriting logic which allows users to specify components of the system-under-test and their expected behaviour as an executable specification using rewrite rules. In Maude, a rewrite rule is a local transition rule \lstinline{t => t'} that describes a valid transformation of term \lstinline{t} into another term \lstinline{t'}; for example, the rule \lstinline{rl [buy-drink] : coin => drink} describes the action of buying a drink for a coin. In the agent-based literature, Maude rewrite rules have been used as a rule-based mechanism for generating test cases alongside their expected results from a specification expressed in Maude [P5, P6].
\emph{Genetic operators} - In the context of software testing, genetic algorithms \cite{davis1991handbook} can be harnessed to evolve new test cases. Informally, a genetic algorithm is an algorithm inspired by evolution in which a solution to a specific problem, such as a test case for test case generation, is encoded as a data structure that resembles a chromosome. A series of genetic operators, namely mutation and crossover, are then applied to chromosomes repeatedly to modify them in such a way that preserves favourable features \cite{whitley1994genetic}. Mutation is the process by which the chromosomes are randomly changed, and crossover is the process in which the genetic information of two parent chromosomes are combined. To guide the algorithm towards an optimal solution, it is also necessary to define a fitness function as a measure of how well the solution attains its objective.
For testing agent-based models, genetic operators can be considered as rule-based generation mechanisms as they are applied to existing test cases to form new, potentially more challenging ones. For example, an evolutionary approach to test case generation has been proposed in the agent-based literature where the aim is to evolve test cases that are challenging with respect to soft goals such as efficiency, as discussed in Section 4.1.2. This is achieved using a multi-objective genetic algorithm, NSGA-II \cite{deb2002fast}, with multiple fitness functions that represent the soft goals. The genetic algorithm applies mutation operators and crossover with a given probability to a chromosome representing the agent-under-test's environment, producing new environments as test cases that are potentially more challenging. The approach is evaluated in an example scenario where a \lstinline{CleanerAgent} is tasked with cleaning a grid whilst maintaining a certain power level and avoiding obstacles efficiently. Test cases are generated with multiple obstructions and difficult to reach charging stations, presenting a significant challenge for the agents [P16]. This approach was later adopted to support AgentTest, a specification language for agent-based system testing [P15], before being improved with a preference-based multi-objective algorithm, r-NSGA-II, that is capable of prioritising a subset of the multiple objectives, such that generated test cases are most challenging with respect to those objectives [P24].
\begin{table}[htbp]
\centering
\begin{tabular}{ll}
\toprule
Generation Mechanism Type & Papers \\ \midrule
Path-traversal & {[}P4, P5, P8-P12, P18{]} \\
Information extraction & {[}P1, P3, P7, P13, P14, P20-P22{]} \\
Random testing & {[}P1, P2, P13{]} \\
Rule-based & {[}P1, P5, P6, P15-P17, P21, P23, P24{]} \\
\bottomrule
\end{tabular}
\caption{Summary of different generation mechanisms used in the literature.}
\end{table}
\section*{Declaration of competing interest}
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
\section{Introduction}
In recent years, computational models have become increasingly popular in a variety of domains, with applications ranging from economic modelling \cite{liu2020interbank} to public health \cite{tracy2018agent}. The ability to model and understand vast quantities of data, coupled with the increased availability of data, have helped to establish the computational model as an essential research tool. Agent-based models are one variety of computational model that are used to simulate complex scenarios involving many individuals, often to inform social and economic policies, including national responses to the recent COVID-19 pandemic \cite{flaxman2020report, panovska2020determining, kerr2020covasim}. With such critical applications, poor modelling assumptions or software faults can lead to serious repercussions.
Software testing is an essential software engineering practice that aims to ensure the quality of software products, including the identification and removal of faults. The time and effort necessary to test software and achieve such quality are significant, often accounting for over 50\% of total project costs \cite{ramler2006economic}. In recent years, a variety of techniques have been proposed to alleviate the cost of testing, from search-based testing \cite{mcminn2004search} to fuzzing \cite{godefroid2008automated}, that automate or partially-automate various testing activities such as the creation of test cases.
Of all software testing activities, perhaps the most labour-intensive and time-consuming activity is the creation of test cases \cite{prasanna2005survey}. As a consequence, the topic of test case generation has received significant attention in recent years, leading to a variety of surveys, techniques and tools supporting the automatic or semi-automatic derivation of test suites \cite{anand2013orchestrated}. However, for many real-world applications, it remains unclear how effective these tools and techniques are, with areas such as scientific software requiring further empirical study to establish their applicability \cite{kanewala2014testing}.
Agent-based modelling is one area where the applicability of existing test case generation techniques is unclear. Due to their often non-deterministic nature coupled with their complex emergent behaviour, agent-based models present significant challenges for test case generation \cite{russell2002artificial, luck2005agent}. For example, in response to the ongoing COVID-19 pandemic, the UK government has used an agent-based model developed at Imperial College London to inform an economic policy estimated to cost \pounds192.3bn over the 2020-2021 financial year \cite{flaxman2020report, obr2020coronavirus:online}. Despite the importance of testing such critical software, the model contains only basic integration tests configured to a specific set of parameters \cite{mrcideco79:online}. To this end, there have been several recent calls for improved techniques for validation and testing of predictive models \cite{squazzoni2020computational, wynants2020prediction}.
Clearly, there is a need for tools and techniques that make software testing an accessible practice for both researchers and practitioners regardless of their discipline. In this review we aim to identify the current state-of-the-art techniques and future research directions for test case generation in agent-based modelling, paying particular attention to their potential real-world applications. In order to achieve a fair and comprehensive survey of the existing test case generation literature in the agent-based modelling field, this paper reports the results of a systematic literature review in accordance to the guidelines proposed by Kitchenham and Charters \cite{kitchenham2007guidelines}.
The remainder of this work is organised as follows: in Section 2, we provide an overview of agent-based models and test case generation. In Section 3, we present the research methodology including the research questions and review procedure. In Section 4, the results of the systematic literature review are presented, followed by a discussion of some particularly interesting observations in Section 5. In Section 6 we conclude the review and present potential future directions for research.
\section*{Appendix. Research papers evalutated in the review}
\begin{xltabular}{\linewidth}{lX}
{[}P1{]} & Nguyen, C.D., Perini, A. and Tonella, P., 2008, May. eCAT: a tool for automating test cases generation and execution in testing multi-agent systems. In Proceedings of the 7th international joint conference on Autonomous agents and multiagent systems: demo papers (pp. 1669-1670). \\
{[}P2{]} & Thaler, J. and Siebers, P.O., 2019, July. Show me your properties: the potential of property-based testing in agent-based simulation. In Proceedings of the 2019 Summer Simulation Conference (p. 1). Society for Computer Simulation International. \\
{[}P3{]} & Thangarajah, J., Jayatilleke, G. and Padgham, L., 2011. Scenarios for system requirements traceability and testing. In Autonomous Agents and MultiAgent Systems (pp. 285-292). IFAAMAS. \\
{[}P4{]} & Dehimi, N.E.H., Mokhati, F. and Badri, M., 2015. Testing HMAS-based applications: An ASPECS-based approach. Engineering Applications of Artificial Intelligence, 46, pp.232-257. \\
{[}P5{]} & Kissoum, Y. and Sahnoun, Z., 2007, May. A formal approach for functional and structural test case generation in multi-agent systems. In 2007 IEEE/ACS International Conference on Computer Systems and Applications (pp. 76-83). IEEE. \\
{[}P6{]} & Mokhati, F., Badri, M. and Zerrougui, S., 2013, October. A novel conformance testing technique for Agent Interaction Protocols. In 2013 Science and Information Conference (pp. 485-495). IEEE. \\
{[}P7{]} & Padgham, L., Zhang, Z., Thangarajah, J. and Miller, T., 2013. Model-based test oracle generation for automated unit testing of agent systems. IEEE Transactions on Software Engineering, 39(9), pp.1230-1244. \\
{[}P8{]} & Eleftherakis, G., Kefalas, P. and Kehris, E., 2011, September. A methodology for developing component-based agent systems focusing on component quality. In 2011 Federated Conference on Computer Science and Information Systems (FedCSIS) (pp. 561-568). IEEE. \\
{[}P9{]} & Dehimi, N.E.H. and Mokhati, F., 2019, June. A Novel Test Case Generation Approach based on AUML sequence diagram. In 2019 International Conference on Networking and Advanced Systems (ICNAS) (pp. 1-4). IEEE. \\
{[}P10{]} & Zheng, M. and Alagar, V.S., 2005, December. Conformance testing of BDI properties in agent-based software. In 12th Asia-Pacific Software Engineering Conference (APSEC'05) (pp. 8-pp). IEEE. \\
{[}P11{]} & Kissoum, Y. and Sahnoun, Z., 2008, March. A Recursive Colored Petri Nets semantics for AUML as base of test case generation. In 2008 IEEE/ACS International Conference on Computer Systems and Applications (pp. 785-792). IEEE. \\
{[}P12{]} & Sakellariou, I., Dranidis, D., Ntika, M. and Kefalas, P., 2015, January. Stream X-Machines for Agent Simulation Test Case Generation. In International Conference on Agents and Artificial Intelligence (pp. 37-57). Springer, Cham. \\
{[}P13{]} & Nguyen, C.D., Perini, A. and Tonella, P., 2008, May. Experimental evaluation of ontology-based test generation for multi-agent systems. In International Workshop on Agent-Oriented Software Engineering (pp. 187-198). Springer, Berlin, Heidelberg. \\
{[}P14{]} & Houhamdi, Z. and Athamena, B., 2011. Structured integration test suite generation process for multi-agent system. Journal of Computer Science, 7(5), p.690. \\
{[}P15{]} & Babac, M.B. and Jevtić, D., 2014. AgentTest: A specification language for agent-based system testing. Neurocomputing, 146, pp.230-248. \\
{[}P16{]} & Nguyen, C.D., Miles, S., Perini, A., Tonella, P., Harman, M. and Luck, M., 2012. Evolutionary testing of autonomous software agents. Autonomous Agents and Multi-Agent Systems, 25(2), pp.260-283. \\
{[}P17{]} & Shahir, H.Y., Glässer, U., Farahbod, R., Jackson, P. and Wehn, H., 2012. Generating test cases for marine safety and security scenarios: a composition framework. Security Informatics, 1(1), p.4. \\
{[}P18{]} & Cavarra, A., 2011. A data-flow approach to test multi-agent ASMs. Formal aspects of computing, 23(1), pp.21-41. \\
{[}P19{]} & Low, C.K., Chen, T.Y. and Rónnquist, R., 1999. Automated test case generation for BDI agents. Autonomous Agents and Multi-Agent Systems, 2(4), pp.311-332. \\
{[}P20{]} & Moser, T., Dürr, G. and Biffl, S., 2010. Ontology-Based Test Case Generation For Simulating Complex Production Automation Systems. In SEKE (pp. 478-482). \\
{[}P21{]} & Szatmári, Z., Oláh, J. and Majzik, I., 2011, July. Ontology-based Test Data Generation using Metaheuristics. In ICINCO (2) (pp. 217-222). \\
{[}P22{]} & Zhang, Z., Thangarajah, J. and Padgham, L., 2009, May. Automated testing for intelligent agent systems. In International Workshop on Agent-Oriented Software Engineering (pp. 66-79). Springer, Berlin, Heidelberg. \\
{[}P23{]} & Coelho, R., Cirilo, E., Kulesza, U., von Staa, A., Rashid, A. and Lucena, C., 2007, October. Jat: A test automation framework for multi-agent systems. In 2007 IEEE International Conference on Software Maintenance (pp. 425-434). IEEE. \\
{[}P24{]} & Kalboussi, S., Bechikh, S., Kessentini, M. and Said, L.B., 2013, August. Preference-based many-objective evolutionary testing generates harder test cases for autonomous agents. In International Symposium on Search Based Software Engineering (pp. 245-250). Springer, Berlin, Heidelberg. \\
\end{xltabular}
|
\section{Introduction}
\label{sec:intro}
Voice and face share various potential characteristics, \eg, gender, ethnicity, age, which are helpful for identification and matching. Literatures \cite{smith2016matching, kamachi2003putting, mavica2013matching} show that humans can hear the voice of an unknown person and match the corresponding face with higher accuracy than chance, and vice versa. From the perspective of brain science, multimodal brain regions exist in the human brain, which process both voices and faces to form person identity representations \cite{tsantani2019brain}. Can machines learn such ability to recognize the face with the same identity only by hearing the voice, or recognize the voice from the face? In recent years, researchers have started to seek an answer to this interesting question \cite{nagrani2018seeing, wen2018disjoint}. The research of this technology is beneficial to many application scenarios, including criminal investigation, synthesis or retrieval of human faces from voices \cite{oh2019speech2face, wen2019face, bai2020speech, choi2019inference}, \etc. This task can be specialized as cross-modal matching, verification and retrieval problems. Different from the audio-visual speech recognition task \cite{zhang2020can}, the voice-face association problem is aim to find the identity relationship between face and voice, rather than the relationship between voice and facial action.
In recent years, we have witnessed some progress of early studies along this line. As a representative example, SVHF \cite{nagrani2018seeing} regards the matching problem as a binary classification problem, and has achieved comparable performance with human baseline in both voice-to-face matching and face-to-voice matching. Benefit from the development of deep learning and the cross-modal retrieval technology, some recent work \cite{kim2018learning, wang2020learning, horiguchi2018face, xiong2019voice, nagrani2018learnable} has further verified the feasibility of this problem through deep metric learning. Wen \etal \cite{wen2018disjoint} boost the performance with multiple supervision.
Despite previous methods have been able to reach the same level as untrained humans, there are still two problems in learning voice-face association. \textbf{(a)} The first problem is that contrastive loss functions used in previous work only use local information in a mini-batch, which leads to slow convergence. \textbf{(b)} The second one is that the diversity of difficulty across identities is ignored. Here the diversity of difficulty means that there are obvious differences in the difficulty of learning voice-face association among different subjects. To illustrate this problem, we train a model and test its accuracy on 1:2 voice-to-face matching for different subjects. The result is shown in \figref{fig:acc_id}, from which it can be noticed that the accuracy of identities is significantly different. This phenomenon coincides with what we find in reality. For example, not all male voices are low and rough, and there is an unignorable fraction of male voices that have their own characteristics. In \figref{fig:acc_id}, we show some easy and hard identities in the obtained results. We observe that the accuracy distribution is relatively uniform, which validates our assumption that the learning difficulty is diverse. Moreover, there exists an unignorable fraction of identities suffering from a accuracy worse than random guess. This validates the existence of extremely hard identities.
Identities of this kind are hardly learnable. What is worse, they might even confuse the model and shift the correct decision boundary. In this paper, we name the hard but learnable identity as
\emph{hard identity}, and the extremely hard identity as \emph{personalized identity}. In this sense, a reasonable learning method should explore deeper into the hard identities while filtering out the personalized ones.
\begin{figure}[t]
\includegraphics[scale=0.43]{fig_acc_id.pdf}
\centering
\caption{Accuracy of different identities in the validation set under the 1:2 voice-to-face matching setting. There is a significant gap between identities, performances of some identities are even lower than chance (50\%).}
\label{fig:acc_id}
\end{figure}
Based on the above consideration, in this paper, we propose an adaptive framework for the voice-face association learning. To overcome \textbf{(a)}, we introduce a two-level modality alignment, which consists of implicit and explicit modality alignment. The implicit alignment is implemented with an identity-classification-driven loss. With the theoretical analysis, we show that minimizing the implicit alignment loss could maximize the distance of embedding across modalities and identities and minimize the distance of embeddings across modalities but belong to the same identity. Moreover, the distance is measured from a global perspective instead of a local mini-batch. In this way, the implicit alignment introduces global information and identity semantics in the embeddings. Moreover, the explicit alignment, as a complementary component, aligns the two modalities in a mini-batch directly. For \textbf{(b)}, we propose an adaptive framework to handle the hard identities and personalized identities with dynamic identity weights. The hard identities obviously contribute to the bottleneck of the performance of the learning methods. We propose an adaptive weighting strategy to gradually increase the weights of the hard identities. This encourages the network to dive deeper into the associations between voice and face. Since personalized identities are extremely hard to learn, and their gradients are larger throughout the training phase, forcing the model to learn these samples will reduce the generalization of the model. Therefore, our proposed strategy adaptively assigns zero weights to personalized identities.
In a nutshell, the main contributions of this work can be summarized as follows:
\begin{itemize}
\item We propose explicit modality alignment and implicit modality alignment to effectively learn the voice-face association in a comprehensive manner.
\item We propose an adaptive identity re-weighting framework to better explore cross-modal associations from hard identities, and excluding personalized identities for generalization.
\item Experiments under various settings are conducted to illustrate the effectiveness of the proposed framework.
\end{itemize}
\section{Related Work}
\label{related_work}
\begin{figure*}[h]
\centering
\vspace{-1mm}
\includegraphics[scale=0.82]{fig_overview.pdf}
\caption{An overview of the proposed method. Face images and voice audio clips are fed into the face encoder network and the voice encoder network, respectively. The extracted embeddings are then assigned different weights according to the average loss of identities, and personalized samples are filtered out. Finally, the parameters of the network are updated with two-level modality alignment.}
\label{fig:overview}
\end{figure*}
\subsection{Learning voice-face association}
In recent years, learning the voice-face association has aroused the interest of researchers. To the best of our knowledge, SVHF \cite{nagrani2018seeing} is the first work to propose a machine learning algorithm for the voice-face association learning, which focuses on the task of matching voices and faces. It poses the matching as a binary classification problem and simply uses the concatenation of the features as input of the classifier. Then the researchers turn to solve the problem from the cross-modal learning perspective, and include additional tasks such as cross-modal verification and retrieval. Inspired by existing cross-modal retrieval methods, some studies \cite{nagrani2018learnable,kim2018learning,xiong2019voice} adopt the metric learning technique, \eg, contrastive loss \cite{hadsell2006dimensionality} or triplet loss \cite{weinberger2006distance}, to help bridge the semantic gap between the two modalities. To further exploring the relationship between samples in a mini-batch, Horiguchi \etal utilize N-pair loss \cite{horiguchi2018face, sohn2016improved} and Wang \etal propose ranking loss \cite{wang2020learning} in the learning. Different from these approaches that are only supervised by the identity information, Wen \etal explicitly introduce more information like gender and nationality to supervise the learning in a multi-task manner \cite{wen2018disjoint}.
A shortcoming of most existing metric learning methods is that the losses they adopted only utilize the local information in a mini-batch. Yet in this work, we consider a two-level modality alignment where the global information in the whole dataset and the local information in the mini-batch are simultaneously employed.
\subsection{Sample re-weighting in deep learning}
In deep learning, different samples have different effects on model performance.
To evaluate how samples affect models, Koh \& Liang \cite{koh2017understanding} use influence function to test the influence of sample weighting on loss without retraining. Based on the influence function, Wang \etal \cite{wang2018data} locate samples which are not helpful after the first round of training, and retrain the model from scratch without these samples. For the same purpose, Ren \etal \cite{ren2018learning} utilize meta learning method to assign weights based on the gradient direction of samples, and Fan \etal \cite{fan2017learning} explores deep reinforcement learning to automatically select samples in the training process. Sample re-weighting strategies are also applied in the object detection task \cite{shrivastava2016training,lin2017focal,li2019gradient}. One of the most well-known works is focal loss \cite{lin2017focal}, which alleviates class imbalance with a weighted loss function.
These re-weighting strategies only consider the instance-level difficulty, thus weight each sample separately. However, in the voice-face association learning task, the difficulty is identity-level. To address this issue, we propose a novel strategy that assigns identity-specific weights, where personalized identities are further excluded with zero weights adaptively. The advantages are two-fold: on one hand, it can reduce the number of weights and thus reduce the algorithm's complexity; on the other hand, the exclusion of personalized identities ensures the generalization of the network.
\section{Methodology}
Our goal is to learn generic vector representations to bridge the semantic gap across voices and faces such that a series of tasks ranging from face-voice/voice-face matching, verification to retrieval could be made available. To achieve this goal, we consider two factors. First of all, to narrow the gap across modalities, we expect to perform effective modality alignment. To this end, we introduce a loss-driven alignment mechanism in \secref{sec:softmax} where the semantic gap is indirectly reduced by minimizing the identity classification loss function, which is called implicit modality alignment loss.
Besides indirect information, we also consider an explicit modality alignment loss in \secref{sec:cross} based on the contrastive learning framework. As for the second factor, we consider the diversity of learning difficulty mentioned in \secref{sec:intro}. It leads us to an adaptive learning framework with dynamic identity weights, which will be described in \secref{sec:noise}.
An overview of our framework is provided in \figref{fig:overview}, where
we adopt the convolutional networks as our backbone.
After extracting embeddings, we perform an adaptive learning procedure via identity re-weighting. The loss function is the sum of the explicit and implicit modality alignment loss. When the identity weights are obtained, the final model is obtained by retraining the network based on the weights.
\subsection{Identity Recognition and Implicit Modality Alignment}
\label{sec:softmax}
First of all, we start with an implicit formulation of the modality alignment loss.
To make sure that the face/voice embeddings are consistent with the semantics of identity, we expect that the learned embeddings should lead to accurate identity recognition. This motivates us to minimize the softmax loss of the identity classification for both face and voice embeddings. Recalling Fig.\ref{fig:overview}, our backbone includes the deep encoders to leverage the face/voice embeddings and a linear classifier, i.e., the last FC layer of the architecture. We assume both voice and face embeddings share a common identity linear classifier. Its weight matrix is denoted as $\bm{W} = [\bm{\omega}_1, \bm{\omega}_2, \cdots, \bm{\omega}_M] \in \mathbb{R}^{D\times M}$, where $M$ and $D$ represent the number of identities and feature dimensions respectively.
Given the training data $\mathcal{D} = \{(\bm{v}_i, \bm{x}_i, {y}_i)\}_{i=1}^{N}$, the identity classification loss is presented as follows:
\begin{eqnarray}
\label{eq:l_face}
&&\mathcal{L}_{face} = - \frac{1}{N} \sum\limits_{i=1}^N \log \frac{\exp(\bm{\omega}_{\yi}^T \bm{x}_i)}{\sum_{j=1}^M \exp(\bm{\omega}_{j}^T \bm{x}_i)} \\
\label{eq:l_voice}
&&\mathcal{L}_{voice} = - \frac{1}{N} \sum\limits_{i=1}^N \log \frac{\exp(\bm{\omega}_{\yi}^T \bm{v}_i)}{\sum_{j=1}^M \exp(\bm{\omega}_{j}^T \bm{v}_i)} \\
\label{eq:l_exp}
&&\mathcal{L}_{implicit} = \mathcal{L}_{face} + \mathcal{L}_{voice}
\end{eqnarray}
As an interesting fact, we can prove that adopting a common classifier could lead to an implicit modality alignment mechanism. This is shown in the following proposition.
\begin{prop}\label{prop:implicit}
Supposing that, for any $k \in \left\{1,2,\cdots,M \right\}$, the weight decay strategy ensures $\|\bm{\omega}_k\| \leq C$, we have a lower bound of $\mathcal{L}_{implicit}$ written as follows:
\begin{align*}
\mathcal{L}_{implicit}
&\geq 2\log M - \frac{C}{MN}\sum_{j=1}^M D_j
\end{align*}
where
\begin{align*}
D_j &= \left\|(M - 1)\sum_{{y}_i = j}(\bm{x}_i + \bm{v}_i) - \sum_{{y}_i \neq j}(\bm{x}_i + \bm{v}_i)\right\|.
\end{align*}
\end{prop}
\noindent Prop. \ref{prop:implicit} shows the following properties of $\mathcal{L}_{implicit}$:
\begin{itemize}
\item According to the inequality in this proposition, minimizing $\mathcal{L}_{implicit}$ leads a smaller value of its lower bound $\log M - \frac{C}{MN}\sum_{j=1}^M D_j$, which equivalently leads to a larger value of $\frac{C}{MN}\sum_{j=1}^M D_j$.
\item For a fixed $j$, $D_j$ is the overall distance between the face and voice embeddings belonging to $j$, and the embeddings that do not belong to $j$. Maximizing $\frac{C}{MN}\sum_{j=1}^M D_j$ eventually enforces all $\bm{x}_i, \bm{v}_i$ from the same class to be close to each other and enforces all $\bm{x}_i, \bm{v}_i$ from different classes to be far away from each other. In this sense, minimizing $\mathcal{L}_{implicit}$ provides an \emph{implicit modality alignment} mechanism for our method.
\item Compared with directly maximizing $\frac{C}{MN}\sum_{j=1}^M D_j$, minimizing $\mathcal{L}_{implicit}$ introduces $\bm{W}$ to the model which leverages global information shared across all sample points. From the efficiency perspective, optimizing $D_k$ directly requires traversing the entire training set, while the implicit optimization can be accelerated by mini-batch training.
\end{itemize}
\subsection{Explicit modality alignment}
\label{sec:cross}
In the previous subsection, we show that the identity recognition loss behaves like a global and implicit modality alignment loss. To obtain a comprehensive loss, we introduce a local and explicit modality alignment loss to include the complementary information of the implicit loss.
Instead of using paired data directly, we implement the explicit modality alignment with N-pair loss \cite{sohn2016improved, sun2020circle}, which explores the local relationship among a mini-batch of instances. It is formulated as follows:
\begin{equation}
\label{eq:l_cross}
\begin{aligned}
\mathcal{L}_{explicit} &= \frac{1}{N}\sum\limits_{i=1}^{N} \log(m + \frac{\sum_{{y}_j\neq {y}_i}\exp(\bm{v}_i\hat{\bm{x}}_j)} {\exp(\bm{v}_i\hat{\bm{x}}_i)}) \\
&+ \frac{1}{N}\sum\limits_{i=1}^{N} \log(m + \frac{\sum_{{y}_j\neq {y}_i}\exp(\bm{x}_i\hat{\bm{v}}_j)} {\exp(\bm{x}_i\hat{\bm{v}}_i)})
\end{aligned}
\end{equation}
where $m$ is a hyper parameter to control inter-class margins, and $\hat{\bm{x}} = \frac{\bm{x}}{\|\bm{x}\|}$, $\hat{\bm{v}} = \frac{\bm{v}}{\|\bm{v}\|}$.
This loss function adds explicit constraints on the embeddings of the two modalities, as a powerful complementary to the implicit alignment. At first glance, the ratio of positive samples and negative samples is $1:($N$-1)$, which might lead to the imbalance issue. Nonetheless, according to the following analysis, we can see that the loss terms will not be affected by the overwhelming ratio of the negative instances. In fact, $\mathcal{L}_{explicit}$ can be approximately written as
\begin{equation}
\begin{split}
\mathcal{L}_{explicit}
&\approx \frac{1}{N}\sum\limits_{i=1}^{N}[\max\limits_{{y}_j\neq {y}_i}\{\bm{v}_i\hat{\bm{x}}_j\} - \bm{v}_i\hat{\bm{x}}_i + m-1]_+ \\
& + \frac{1}{N}\sum\limits_{i=1}^{N}[\max\limits_{{y}_j\neq {y}_i}\{\bm{x}_i\hat{\bm{v}}_j\} - \bm{x}_i\hat{\bm{v}}_i + m-1]_+
\end{split}
\end{equation}
where $[x]_+$ indicates $\max(x,0)$. This shows that $\mathcal{L}_{explicit}$ behaves like a hinge loss to punish $\max\limits_{{y}_j\neq {y}_i}\{\bm{v}_i\hat{\bm{x}}_j\} - \bm{v}_i\hat{\bm{x}}_i$. In this sense, only the one negative instance realizing the maximum is activated for each loss term. Hence, the imbalance issue could be naturally avoided.
\subsection{Adaptive identity re-weighting}
\label{sec:noise}
In this subsection, we propose an adaptive learning algorithm to deal with the diversity of the learning difficulties. The proposed method has three stages. During the first stage, we perform a warm-up training, where a pretrained model is learned without identity weights. In the second stage, the pretrained model is updated together with the identity weights. In the third stage, we train the final model based on the learned identity weights. A summary of all the details is shown in \AlgRef{alg:noise}.
\begin{algorithm}
\caption{Training with identity weights}
\begin{algorithmic}[1]
\REQUIRE {Training data $\mathcal{D}$, warm up iteration $T_{warm}$, update iteration $T_{update}$, max iteration $T_{max}$, batch size $N$, number of identities $M$, ratio of data retained $R_{keep}$.}
\ENSURE {model parameters $\theta_f, \theta_g$.}
\STATE{\color{org}\COMMENT{\texttt{First Stage}}}
\STATE {\texttt{Train} $F_{\theta_f}, G_{\theta_v}$ \texttt{with} $\mathcal{D}$ \texttt{for} $T_{warm}$ \texttt{iterations.}}
\STATE{}{\color{org}\COMMENT{\texttt{Second Stage}}}
\WHILE {$\sum\limits_{i=0}^M I[s_i^{t-1} > 0] < R_{keep}\times M$}
\STATE {$\{(A_i, I_i, y_i)\}_{i=1}^N \leftarrow $\texttt{SampleMiniBatch}($\mathcal{D}, N$)}
\STATE {$\bm{x}_i, \bm{v}_i \leftarrow F_{\theta_f}(I_i), G_{\theta_g}(A_i)$.}
\STATE {\texttt{Calculate} $\mathcal{L}_{implicit}$ \texttt{with} \equref{eq:l_face} \texttt{and} (\ref{eq:l_voice}).}
\STATE {\texttt{Update} $H^{t}$ \texttt{with} \equref{eq:mean_loss}.}
\IF {$t \% T_{update} = 0$}
\STATE {\texttt{Update} $s^{t}$ \texttt{with} \equref{eq:s_update}.}
\ELSE
\STATE {$s^{t} \leftarrow s^{t-1}$.}
\ENDIF
\STATE {\texttt{Update} $\theta_f,\theta_g$ \texttt{with} \equref{eq:l_theta}.}
\ENDWHILE
\STATE{}{\color{org}\COMMENT{\texttt{Third Stage}}}
\STATE {\texttt{Reinitialize} $\theta_f, \theta_g$.}
\FOR{$t=1$ to $T_{max}$}
\STATE {\texttt{Update} $\theta_f,\theta_g$ \texttt{with} \equref{eq:l_theta}.}
\ENDFOR
\end{algorithmic}
\label{alg:noise}
\end{algorithm}
The details for the first and third stage are obvious. We only elaborate on the second stage next.
During this training phase, we update the weights in an iterative manner. In the $t$-th iteration, we first sample a mini-batch (\texttt{Line 5} in \AlgRef{alg:noise}) and perform a standard inference to obtain the embeddings $\bm{x}_i, \bm{v}_i$ for all the instances in the mini-batch (\texttt{Line 6} in \AlgRef{alg:noise}). Next, we evaluate the hardness of each involved identity in the current mini-batch. To ensure the robustness of the learned weights, we only employ the more stable implicit loss (\texttt{Line 7} in \AlgRef{alg:noise}) $\mathcal{L}_{implicit}$ to measure the hardness. To integrate the hardness information from previous iterations and the current iteration, we perform a moving average strategy to calculate the hardness (\texttt{Line 8} in \AlgRef{alg:noise}). Above all, the hardness for the $i$-th identity, \ie, $H^{t}_{i}$, could be written as:
\begin{equation}
\label{eq:mean_loss}
H^{t}_{i} = \beta \cdot H^{t-1}_{i} + (1 - \beta) \cdot \mathcal{L}_{implicit}
\end{equation}
where $\beta \in (0,1)$ is a hyper parameter to control the importance of the current iteration.
Now, we assign a weight for each identity to gradually add the hard identities to the training set (\texttt{Line 11} in \AlgRef{alg:noise}). Specifically, in the $t$ iteration of the training process, the weight of the samples belonging to the $i$-th identity in the loss is denoted as $s^{t}_i$. As for the initialization, $s^{0}_i$ is set to $1$ for the bottom 30\% $H^{0}_{i}$, and is set to $0$ otherwise. To maintain
\noindent stability, the weight update is triggered after every $T_{update}$ iterations (\texttt{Line 9} in \AlgRef{alg:noise}). If the update is triggered,
$s^{t}_i$ will be updated as follows:
\begin{equation}
\label{eq:s_update}
s^{t}_i =
\begin{cases}
1, \emph{for all bottom} ~k~ H^{t}_{i}~\emph{ with} ~s^{t-1}_i = 0.
\\
\alpha \cdot s_i^{t-1}, \emph{otherwise.}
\end{cases}
\end{equation}
where $\alpha \in (0,1)$ and $k$ (a positive integer) are hyper parameters. Next, we explain how this equation works. First, we will pick out all the identities with the smallest $k$ $H^{t}_i$ and set their $s^t_i$ to $1$. For the rest of the identities, we simply employ a weight decay strategy by setting $s^t_i =\alpha\cdot s^{t-1}_i$.
Meanwhile, the identities with the largest $H^t_i$ are considered as personalized identities, which should be dropped out via a zero weight. To this end, we update the identity weights sequentially. We will terminate the second stage of training immediately after $R_{keep}$ percent of identities are assigned with non-zero weights (\texttt{Line 4} in \AlgRef{alg:noise}).
At the end of each iteration (if not terminated), the network parameters $\theta_f, \theta_g$ are updated with weighted loss functions (\texttt{Line 14} in \AlgRef{alg:noise}):
\begin{eqnarray}
\label{eq:l_face_denoise}
&&\mathcal{L}_{face} = - \sum\limits_{i=1}^N \hat{s}^{t}_{{y}_i} \log \frac{\exp(\bm{\omega}_{\yi} \bm{x}_i)}{\sum_{j=1}^M \exp(\bm{\omega}_{j} \bm{x}_i)} \\
\label{eq:l_voice_denoise}
&&\mathcal{L}_{voice} = - \sum\limits_{i=1}^N \hat{s}^{t}_{{y}_i} \log \frac{\exp(\bm{\omega}_{\yi}\bm{v}_i)}{\sum_{j=1}^M \exp(\bm{\omega}_{j} \bm{v}_i)}
\label{eq:l_cross_denoise}
\end{eqnarray}
\begin{eqnarray}
&&\begin{aligned}
\mathcal{L}_{explicit} = \sum\limits_{i=1}^{N} \hat{s}^{t}_{{y}_i} \log(m + \frac{\sum\limits_{{y}_j\neq {y}_i}\exp(\bm{v}_i\hat{\bm{x}}_j)} {\exp(\bm{v}_i\hat{\bm{x}}_i)}) \\
+ \sum\limits_{i=1}^{N} \hat{s}^{t}_{{y}_i} \log(m + \frac{\sum\limits_{{y}_j\neq {y}_i}\exp(\bm{x}_i\hat{\bm{v}}_j)} {\exp(\bm{x}_i\hat{\bm{v}}_i)})
\end{aligned} \\
\label{eq:l_theta}
&&\mathcal{L_{\theta}} = \mathcal{L}_{face} + \mathcal{L}_{voice} + \mathcal{L}_{explicit}
\end{eqnarray}
where $\hat{s}^t_i = s^t_i/\sum\limits_{j=1}^Ns^{t}_j$.
\section{Experiments}
\subsection{Datasets}
Following previous work \cite{wen2018disjoint, wang2020learning}, we evaluate the proposed method on a constructed dataset based on VoxCeleb \cite{nagrani2017voxceleb} and VGGFace \cite{parkhi2015deep} datasets. VoxCeleb is an audio-visual dataset of short human speech videos, which provides audios for our experiments. Meanwhile, VGGFace is a human face dataset of 2,622 identities. Since in our settings, it is not necessary to capture voices and faces from the same video, we use still faces from VGGFace instead of extracted faces from VoxCeleb. The intersection of VoxCeleb and VGGFace contains 1,225 identities after filtering low quality data. Then we split the data into train/validation/test sets without identity overlapping, and generate the queries for validation and test according to our evaluation protocols (See \secref{subsec:implement}). The statistics of the resulted dataset are reported in \tabref{tab:dataset}.
\begin{table}[h]
\caption{Details on the datasets.}
\centering
\setlength{\tabcolsep}{5pt}
\begin{tabular}{c|cccc}
\toprule
& train & validation & test & total \\
\midrule
audio clips & 113,322 & 14,182 & 21,850 & 149,354 \\
face images & 104,724 & 12,260 & 20,076 & 137,060 \\
identities & 924 & 112 & 189 & 1,225 \\
queries (V-F) & --— & 42,546 & 65,550 & 108,096 \\
queries (F-V) & --— & 36,780 & 60,228 & 97,008 \\
\bottomrule
\end{tabular}
\label{tab:dataset}
\end{table}
\begin{table*}[t]
\caption{Results (\%) on 1:2 matching, verification and retrieval. V-F: from voice to face; F-V: from face to voice; U: unrestricted; G: query restricted by gender. The best results of our models and competitors are highlighted in \first{soft red} and \second{soft blue}, respectively.}
\centering
\setlength{\tabcolsep}{5pt}
\begin{tabular}{c|cccc|cccc|cc}
\toprule
\multirow{2}{*}{Methods}
& \multicolumn{4}{c}{1:2 Matching (ACC)}
& \multicolumn{4}{|c}{Verification (AUC)}
& \multicolumn{2}{|c}{Retrieval (mAP)} \\
\cline{2-11} & V-F (U) & F-V (U) & V-F (G) & F-V (G)
& V-F (U) & F-V (U) & V-F (G) & F-V (G)
& V-F & F-V \\
\midrule
SVHF \cite{nagrani2018seeing} & 81.0 & 79.5 & 63.9 & 63.4 & --— &--—&--—&--—&--—&--—\\
DIMNet \cite{wen2018disjoint} & 81.3 & 81.9 & 70.6 & 69.9 & 81.0 & 81.2 & \second{70.4} & 69.3 & 4.3 & \second{3.8} \\
Wang's \cite{wang2020learning} & \second{83.4} & \second{84.2} & \second{71.7} & \second{71.1} & \second{82.6} & \second{82.9} & 70.3 & \second{70.1} & \second{4.4} & 3.4\\
\midrule
Ours (focal \cite{lin2017focal}) & 85.3 & 85.0 & 75.6 & 74.5 & 85.6 & 85.4 & 76.0 & 75.1 & 6.5 & 6.2 \\
Ours (-E -W) & 84.6 & 84.7 & 72.8 & 71.3 & 84.8 & 85.0 & 72.4 & 71.4 & 4.6 & 5.1 \\
Ours (-E) & 84.3 & 84.3 & 75.0 & 74.4 & 84.6 & 84.6 & 75.3 & 74.8 & 5.1 & 4.9 \\
Ours (-I -W) & 85.1 & 85.4 & 75.2 & 74.3 & 85.7 & 85.7 & 75.8 & 75.2 & 4.8 & 5.1 \\
Ours (-W) & 85.5 & 85.2 & 76.3 & 74.7 & 85.8 & 85.3 & 76.7 & 75.1 & \first{6.5} & \first{6.2} \\
Ours & \first{87.2} & \first{86.5} & \first{77.7} & \first{75.3} & \first{87.2} & \first{87.0} & \first{77.5} & \first{76.1} & 5.5 & 5.8 \\
\bottomrule
\end{tabular}
\label{tab:res_m_v_r}
\end{table*}
\subsection{Implementation details} \label{subsec:implement}
\noindent \textbf{Network architecture.}
The face feature extractor is implemented with SE-ResNet-50 \cite{hu2018squeeze}, which is a powerful backbone on multiple tasks of computer vision. The input is a face image of size $112\times112\times3$, which is normalized to $[-1, 1]$ by subtracting $127.5$ and dividing $127.5$, and the output is a 128-dimensional face embedding. The voice feature extractor is implemented with Thin-ResNet-34 \cite{arandjelovic2016netvlad, zhong2018ghostvlad}, which inputs a spectrogram of voice and outputs a 128-dimensional voice embedding. Spectrograms have 257 channels (256 frequency components and 1 DC component), which are generated with a hamming sliding window of width 25ms and of hop 10ms. The two networks are pre-trained with the face recognition task on MS-1M \cite{guo2016ms}, and the audio speaker recognition on VoxCeleb2 \cite{chung2018voxceleb2}, respectively.
\noindent \textbf{Training strategy.}
The training process is split into three stages: warm-up training, identity weight learning and retraining with fixed identity weights, as mentioned in \secref{sec:noise}. In order to balance the number of samples with different identities, we sample several identities in an iteration, and then randomly sample one face image and one voice audio clip for each identity. For data pre-processing, we apply data augmentation including random rotation (from $-15^{\circ}$ to $15^{\circ}$), random cropping ($112\times 112$) for images, and random cropping in time axis (from 2.5 seconds to 5.0 seconds) for audios. In our model, the hyperparameters are set as follows: $m = 3.4$ (\equref{eq:l_cross_denoise}), $\beta = 0.9$ (\equref{eq:mean_loss}), $\alpha = 0.99, k = 22$ (\equref{eq:s_update}), $T_{warm} = 500, T_{update} = 100, R_{keep}=0.9$ (\AlgRef{alg:noise}). We adopt the stochastic gradient descent (SGD) optimizer, where batch size and momentum are set to $64$ and $0.9$, respectively. The learning rate is initialized as $10^{-2}$, and decays by $0.1$ in the 2k and 3k iterations. The max iteration $T_{max}$ is 10k, and the best model on the validation set is preserved for evaluation.
\begin{figure*}[t]
\vspace{2mm}
\subfigure{
\begin{minipage}{1.0\textwidth}
\centering
\centerline{\includegraphics[scale=0.20]{legend.jpg}}
\end{minipage}
}
\vspace{-3.5mm}
\subfigure{
\begin{minipage}[b]{0.32\textwidth}
\includegraphics[scale=0.37]{match_v2f.pdf}
\centerline{(a) V-F (U)}
\end{minipage}
\label{fig:match_v2f}
}
\subfigure{
\begin{minipage}[b]{0.32\textwidth}
\includegraphics[scale=0.37]{match_f2v.pdf}
\centerline{(b) F-V (U)}
\end{minipage}
\label{fig:match_f2v}
}
\subfigure{
\begin{minipage}[b]{0.32\textwidth}
\includegraphics[scale=0.37]{match_g_v2f.pdf}
\centerline{(c) V-F (G)}
\end{minipage}
\label{fig:match_g_v2f}
}
\caption{Quantitative results on 1:$N$ matching task. Best viewed in color.}
\label{fig:match_n}
\end{figure*}
\noindent \textbf{Evaluation protocol.}
To show the overall capacity of the proposed model in voice-face association learning task, we conduct the evaluation under the following four settings:
\begin{enumerate}[(a)]
\item \textbf{1:2 matching}. Given an instance from one modality as the probe, and two candidates from the other modality (including one from the same identity as the probe) as the gallery, the task is to find out which candidate matches the probe. The performance is measured with accuracy (ACC).
\item \textbf{1:$\mathbf{N}$ matching}. This task is basically the same as 1:2 matching, except that the length of gallery $N$ ranges from $2$ to $10$ in our experiments. We report ACC on each $N$.
\item \textbf{Verification}. Given two instances from different modalities, the task is to judge whether they belong to the same person. The performance is measured with Area Under the ROC curve (AUC).
\item \textbf{Retrieval}. This task is extended from the 1:$N$ matching task, where the gallery contains one or more candidates that match the probe. The model is asked to rank the gallery samples so that the candidates matching the probe are ranked at the top. We report the performance with mean average precision (mAP).
\end{enumerate}
For each task, we report the metrics for two types of queries, from voice to face (\textbf{V-F}) and from face to voice (\textbf{F-V}). On matching and verification tasks, the queries are further divided into two subtypes: gallery samples have the same gender as the probe sample (\textbf{G}), or have unrestricted genders (\textbf{U}).
\noindent \textbf{Competitors.} We compare our proposed method with three models: SVHF \cite{nagrani2018seeing}, DIMNet \cite{wen2018disjoint} and Wang's model \cite{wang2020learning}. Note that SVHF takes dynamic faces as input, so its dataset is not exactly the same as ours.
\noindent \textbf{Ablated variants.} In order to quantify two main contributions of this paper: two-level modality alignment and identity re-weighting, we also implement four ablated variants: (1) the model without explicit alignment and re-weighting (denoted by \textbf{-E -W}), (2)the model without implicit alignment and re-weighting (denoted by \textbf{-I -W}) (3) the model without re-weighting (denoted by \textbf{-W}), and (4) a model trained with focal loss \cite{lin2017focal} instead of our re-weighting strategy (denoted by \textbf{focal}).
\subsection{Results and comparison}
\noindent \textbf{Quantitative results.} The results on the 1:2 matching, verification and retrieval are recorded in \tabref{tab:res_m_v_r}. We see that our model significantly outperforms the competitors over all three tasks, with an average improvement of about 2\%-7\%. Furthermore, comparing the performance on gender restricted and unrestricted queries, our method has a larger improvement in the case of gender restriction. These validate that our model could discover deeper associations between face and voice. On the other hand, the results of 1:$N$ matching are shown in \figref{fig:match_n}, which further verifies the advantage of our method. In this task, the accuracy decreases with the increase of N, but our method consistently has a higher performance and has less decrease than the competitors, which shows that the proposed framework is relatively more robust.
Moreover, we could make the following observations: (1) The explicit modality alignment can significantly improve performance, especially in the gender-restricted groups, which are more challenging. In the 1:2 matching task and the verification task, identity re-weighting brings an increase of about 1\%, and achieves the state-of-the-art performance. However, training with focal loss doesn't bring significant improvement. (2) For all the tasks, results on gender-unrestricted queries are obviously better than gender-restricted ones. This shows that facial and voice features are strongly related to gender, which is consistent with people's experience. Besides, even if the gender is restricted, machines can still learn the association between face and voice. (3) Despite that our identity re-weighting strategy brings improvements on the matching and verification tasks, the full model does not perform as well as its ablated counterpart on the retrieval task.
One possible reason is that the re-weighting strategy improves generalization at the expense of precision on the tail of gallery.
How to improve the performance of this situation more effectively is worthy of further study.
\begin{figure}[t]
\centering
\vspace{-1mm}
\subfigure[Embeddings on training set.]{
\includegraphics[scale=0.24]{emb_train.eps}
\label{fig:emb_train}
}
\subfigure[Embeddings on test set.]{
\includegraphics[scale=0.24]{emb_test.eps}
\label{fig:emb_test}
}
\caption{Visualization of learned embeddings. The same color points belong to the same person, and different shapes represent different modalities. Dimension reduction is implemented by multi-dimensional scaling (MDS) \cite{wickelmaier2003introduction}. Best viewed in color.}
\label{fig:emb}
\end{figure}
\noindent\textbf{Qualitative results.} A visualization of the face embeddings and the voice embeddings extracted with our model is provided in \figref{fig:emb}. It could be observed that embeddings from the same identities are close and discriminative in most cases, which shows the effectiveness of learned features intuitively.
\begin{figure}[t]
\centering
\includegraphics[scale=0.9]{fig_noise_sample.pdf}
\caption{Some examples of personalized samples. We show four cases of voice and face inconsistency as examples.}
\label{fig:noise_sample}
\end{figure}
\subsection{Ablation study for training strategy}
\noindent \textbf{Effect of identity weights.}
In our framework, the hard identities are filtered out by manually setting the ratio of data retained $R_{keep}$. As is shown in \tabref{tab:keep_ratio}, we evaluate the 1:2 matching accuracy of models learned with different $R_{keep}$, ranging from 0.6 to 1.0. Three conclusions can be drawn: (1) Weighting the difficult samples can bring about 1.2\% to 1.6\% performance improvement, even if extremely hard identities are not excluded. (2) After dropping extremely hard identities, the performance is further boosted, verifying the effectiveness of eliminating personalized identities. (3) After a certain point, the performance of the model decreases with the increase of the retaining ratio. This is due to the abandonment of some data might lower the data diversity. Some excluded samples are shown in \figref{fig:noise_sample}. Intuitively speaking, these samples are usually special in voice. See {\it Supplementary Materials} for more analysis.
\noindent \textbf{Effect of pre-training.} We also explored the effect of pre-training on the model in {\it Supplementary Materials}.
\begin{table}[t]
\caption{The effect of weighting and personalized identities filtering on 1:2 matching accuracy. The best results are shown in \textbf{bold}.}
\centering
\begin{tabular}{c|ccc}
\toprule
$R_{keep}$ &
re-weighting & V-F (U) & F-V (U)\\
\midrule
1.0 & $\times$ & 85.5 & 85.2 \\
1.0 & \checkmark & 86.7 & 86.8 \\
0.9 & \checkmark & \textbf{87.2} & \textbf{87.2} \\
0.8 & \checkmark & 87.0 & 87.1 \\
0.7 & \checkmark & 86.3 & 86.3 \\
0.6 & \checkmark & 85.9 & 85.3 \\
\bottomrule
\end{tabular}
\label{tab:keep_ratio}
\end{table}
\section{Conclusion}
In this paper, we propose a novel embedding framework for voice-face assoication learning, with \textbf{(a)} a comprehensive modality alignment loss embracing global and local information; \textbf{(b)} a dynamic re-weighting strategy to deal with the diversity of learning difficulty across identities. For \textbf{(a)}, we propose a two-level loss including implicit modality alignment loss and explicit modality alignment loss. In Prop. \ref{prop:implicit}, we prove that the implicit alignment loss could globally reduce the inconsistency of the embeddings across modality.
For \textbf{(b)}, we propose a re-weighting framework which can focus
on hard identities while filter out personalized identities via the evolution of the identity weights. Experiments on cross-modal matching, verification and retrieval show that compared with previous methods, our method can better learn the association between face and voice.
\section*{Acknowledgments}
This work was supported in part by the National Key R\&D Program of China under Grant 2018AAA0102003, in part by National Natural Science Foundation of China: 61620106009, 61931008, 61836002, and 61976202, in part by Youth Innovation Promotion Association CAS, and in part by the Strategic Priority Research Program of Chinese Academy of Sciences, Grant No. XDB28000000.
{\small
\balance
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
\label{section:introduction}
Solar spectral lines in the infrared, visible, and ultraviolet ranges are powerful diagnostics for the solar atmosphere. Many of the spectral lines originate in the solar atmosphere (the photosphere, chromosphere, and the chromosphere-corona transition region) and reveal the state and processes in this complex and highly dynamic region. The line profiles reflect the distribution and dynamics of physical parameters (density, temperature, macroscopic and turbulent velocities, etc.) as functions of height in the solar atmosphere. Correspondingly, analysis and interpretation of solar spectral lines becomes an extremely important task for understanding physical processes in the solar atmosphere.
With the launch of NASA's Interface Region Imaging Spectrograph (IRIS, \cite{DePontieu2014}), the diagnostics of the solar chromosphere and transition region was significantly advanced. IRIS observes a variety of lines formed in these regions (Mg\,II\,k\&h 2796\,\AA\, and 2803\,\AA\,, C\,II 1334\,\AA\, and 1335\,\AA\,, Si\,IV 1394\,\AA\, and 1403\,\AA), each sensitive to certain ranges of temperatures and, correspondingly, to certain atmospheric heights. In this work, our analysis will be focused on the Mg\,II\,k line profiles. It was previously demonstrated that Mg\,II\,k\&\,h lines may serve as diagnostic tools for probing parameters of the solar atmosphere in the upper chromosphere \cite{Leenaarts2013,Sadykov2021}. In addition to a variety of physics-based modeling, analysis, and inversion techniques, there have been several attempts to apply machine learning for enhancing the physics knowledge obtained from Mg\,II\,k line profiles. For example, more than 50000 {\it representative profiles} (RP) of the Mg\,II\,k\&h lines were computed by applying clustering algorithms in an assortment of IRIS observations \cite{SainzDalda2019}. The spectra were inverted using a physics-based approach, and the inversion results were used to train a deep learning model emulating the physics-based inversions. Thus, the physical mode can be obtained in a few minutes using a laptop or a desktop machine, which is $\sim$10$^5$-10$^6$ times faster than the classic physics-based inversions. The typical Mg\,II\,k\&h line profiles observed during solar flares were identified using unsupervised clustering algorithms \cite{Panos2018} and linked the appearance of certain types of clusters to specific physical processes occurring during solar flares (electron beam injection properties and flare ribbon propagation).
While the application of machine learning in solar physics is an expanding field \cite{Nita2020}, many questions are still not addressed. One of them is the question of compact embedding (extraction of the deep features) for spectral line profiles, which became of interest to the community just recently \cite{Panos2021}. Obtaining low-dimensional representations of spectra may potentially help address noise reduction, feature selection, and data compression and transmission problems. This manuscript describes an attempt to apply compact embedding analysis to spectral lines. The experiment shows that one needs only 4 dimensions to represent the studied data set. Section~\ref{section:data} describes the data set of Mg\,II\,k line profile observations, data cleaning, and normalization procedures. Section~\ref{section:autoencoder} illustrates the architecture of a neural network used to obtain a compact embedding of the line profiles. Analysis of reconstruction errors for different embedding layer dimensionalities is discussed in Section~\ref{section:results}. An interpretation of the results is summarized in Section~\ref{section:discussion} followed by conclusions in Section~\ref{section:conclusion}.
\section{Data}
\label{section:data}
We utilize IRIS observations of the quiet Sun taken on April 20, 2020 from 08:32:00 UT - 09:56:00 UT at the center of the solar disk. The observations were made in the sit-and-stare mode (the mode when the slit of the spectrograph is always in one position and ``tracks'' the solar features). During these observations, IRIS obtained more than 300,000 individual profiles of Mg\,II\,k lines sampled from the quiet Sun at different time moments and positions along the slit. Some examples of the obtained line profiles are presented in Figure~\ref{figure:examples}. One can notice that the Mg\,II\,k spectral line has a complex shape~-- it typically has a double-peak structure with a central-reversal signature if observed in quiet Sun regions. Both the locations of the peaks and the central reversal represent valuable features for diagnostics of the solar atmosphere \cite{Leenaarts2013}. It is important to note that we consider only the Mg\,II\,k line-related part of the observed spectra; it consists of intensity measurements at 110 wavelengths covering the 2794.9\,\AA\, -- 2797.7\,\AA\, spectral range; each wavelength represents a separate dimension.
\begin{figure*}[t]
\centerline{\includegraphics[width=1.0\linewidth]{Paperfigure1_examples.eps}}
\caption{Examples of the original (blue) and reconstructed (orange) Mg\,II\,k line profiles from the test subset for the case of an embedding layer size of 4 neurons. Lower right panel illustrates the corresponding autoencoder training process (mean squared error, MSE, as a function of the epoch number).}
\label{figure:examples}
\end{figure*}
The pool of obtained line profiles was cleaned. Any intensity points beyond the 2794.9\,\AA\, -- 2797.7\,\AA\, spectral range were removed from the analysis. The line profiles with spikes in the data (defined here as sudden increases of the intensity of 1500\,DN or more at a particular wavelength) and measurements with negative intensity values were removed from consideration. The 1500\,DN threshold was defined as a minimum possible threshold that keeps all the line profiles not affected by spikes in the data set. The resulting 290940 Mg\,II\,k line profiles were all normalized to the same number, which is the maximum peak intensity observed across the line profiles. The first 200000 normalized line profiles were assigned to the train data subset and the remaining 90940 profiles to the test data subset, giving an approximate partitioning of 70\% to 30\%.
\section{Autoencoder Configuration and Training}
\label{section:autoencoder}
For compression of spectroscopic line profiles we have utilized an autoencoder-type neural network architecture. The autoencoder is represented by a fully connected neural network with the numbers of neurons in the subsequent layers set at 110 - 64 - 32 - 16 - $n_{emb}$ - 16 - 32 - 64 - 110. The network was implemented in Python PyTorch \cite{Paszke2017}. Here $n_{emb}$ represents the dimensionality of the embedding layer and is varied from 1 to 15. The Rectified Linear Unit (ReLU) activation function was used for all neurons, and the Mean Squared Error (MSE) was utilized as a loss function. A batch size of 20,000 samples was used. Two network optimizers, ``adam'' \cite{Kingma2015} and stochastic gradient descent, and three learning rates (0.001, 0.00033, 0.0001) were utilized to progress the training. We found that introducing additional fully-connected layers into the network or increasing the number of neurons in the layers does not change the conclusions of this work.
To ensure the completeness of the training process and prevent the network from overfitting, we have implemented the early stopping criterion. Specifically, we assumed that the training process of the network for the current setting is complete if the MSE on the test subset is not improving for more than 0.5\% for two consequent training epochs. The strategy was sequentially applied for each optimizer and learning rate combination (with the learning rates arranged in decreasing order). The entire training process was repeated five times for each $n_{emb}$ to check how variations in the initial neuron weight distribution affect the training.
Figure~\ref{figure:examples} illustrates the reconstructed Mg\,II\,k line profiles and the corresponding training progress for $n_{emb}$=4. The network captures the main line profile features (the central-reversal, the peak asymmetry, and the line continuum behavior and features therein) relatively well, at least from a qualitative perspective.
\section{Results of Reconstruction}
\label{section:results}
\begin{figure*}[t]
\centerline{\includegraphics[width=1.0\linewidth]{Paperfigure2_dimvar.eps}}
\caption{Average error of the line profile reconstruction as a function of the embedding layer dimensionality. The upper-left panel shows the mean squared error (MSE) of the line profile intensity reconstruction, and the corresponding histogram of the distribution of such errors is presented in the upper right panel. The lower-left panel presents the relative mean error of the reconstruction of the odd statistical moments and the lower-right panel that of the even moments.}
\label{figure:dimvars}
\end{figure*}
The average MSE as a function of $n_{emb}$ (in the original intensity units of data numbers, DN) is presented in Figure~\ref{figure:dimvars}\textit{a}. One can notice that the MSE decreases sharply for $n_{emb} \leq{} 4$ and almost stops decreasing when $n_{emb} > 4$. One can state that there is no further progress in line compression for $n_{emb} > 4$ for the current training setting and autoencoder architecture. The distribution of MSE for the case of $n_{emb} = 4$ is presented in the upper-right panel of the same figure. For the majority of the line profiles, the average deviation of the reconstructed intensity from the original values was around 3\,DN, which is comparable with the intensity variations of the line continuum signal.
In addition to the MSE, we also consider the reconstruction of the statistical moments of the spectral lines. In this study, the n-th statistical moment of the line profile is defined as:
\begin{equation}
\mu{}_{n} = \left( \dfrac{ \int_{-\infty}^{\infty}(\lambda{}-\lambda{}_{0})^{n}I(\lambda{})d\lambda{} }
{ \int_{-\infty}^{\infty}I(\lambda{})d\lambda{} } \right)^{1/n},
\end{equation}
\begin{equation}
\lambda{}_{0} = \dfrac{ \int_{-\infty}^{\infty}\lambda{}I(\lambda{})d\lambda{} }
{ \int_{-\infty}^{\infty}I(\lambda{})d\lambda{} },
\end{equation}
Where $\lambda{}$ is a wavelength point and $I(\lambda)$ is the intensity at that wavelength. Typically, the first statistical moment serves as a proxy of for line Doppler shift, the second one as a proxy for the line width, etc. These characteristics are used in spectroscopic studies \cite{Sadykov2015}. Therefore, it is important to analyze the accuracy of the reconstruction for these moments. In this work, the statistical moments are computed only for the 2795.68\,$\AA$-2796.96\,$\AA$ spectral range, which covers the Mg\,II\,k line but removes most of the contribution from the continuum.
\begin{figure*}[t]
\centerline{\includegraphics[width=1.0\linewidth]{Paperfigure3_diminterp.eps}}
\caption{Variations of the decoded line profiles as results of perturbations of the embedding space. Black curves correspond to the line profiles reconstructed from the average embedding parameters. Red and orange curves correspond to the situation when one of the embedding parameters is one standard deviation larger or smaller than the average. The perturbed parameter is indicated in the plot title.}
\label{figure:diminterp}
\end{figure*}
The errors of reconstruction of the statistical moments are presented in Figure~\ref{figure:dimvars}\textit{c,d}. As one can see in the lower-right panel, the even statistical moments experience the same behavior as MSE does: the error of reconstruction of these line moments improves for $n_{emb} \leq{} 4$ and stops decreasing for $n_{emb} > 4$. The considered moments have an average relative error of the reconstruction of $\leq{}$\,2\%, and the error decreases for higher moments.
Surprisingly, the situation is different for the odd line profile moments. First, the odd (asymmetric) line moments experience significantly higher average errors, although, starting from $\mu{}_{5}$, the reconstruction error is below 40\% on average (Fig.~\ref{figure:dimvars}\textit{c}). Second, the reconstruction errors for the odd statistical moments does not decrease with increasing $n_{emb}$ as strongly as was observed for the even components (except the $\mu{}_{1}$ case, which also had a $\sim{}$\,50\% reconstruction error even for high $n_{emb}$).
\section{Discussion. Special case of $n_{emb}$=4.}
\label{section:discussion}
As mentioned in Section~\ref{section:introduction}, finding a compact embedding of spectroscopic line profiles has several important applications. In this work we consider, in particular, compression of the 110-dimensional Mg\,II\,k line profile into the $n_{emb} = [1; 15]$ space and analyze the resulting reconstruction error. One of the key results of such analysis is that the reconstruction error (in terms of MSE or the even statistical moments) does not decrease for $n_{emb} > 4$ (Fig.~\ref{figure:dimvars}). Such behavior may have two explanations. The first is that the entire physics of the line formation may be expressed in four free parameters (e.g., the temperature, density, the macroscopic and turbulent velocities in case of the homogeneous thermal media). Another possible explanation is that different neural network architectures and training strategies might achieve better performance of the autoencoder for $n_{emb} > 4$. Because there was a specific architecture of the network and a specific training strategy considered, this is a very probable scenario. One also can see that, for $n_{emb} = 8$ and $n_{emb} = 14$, there were breakthroughs for some particular trials. Specifically, one of the trials demonstrated a $\sim$20\% improvement of MSE for $n_{emb} = 8$, and one of the trials demonstrated a 55\% improvement of MSE for $n_{emb} = 14$. As one can see, it may possible to have better compressions of line profiles, but that will require more computational resources.
To understand better the meaning of the parameters that the autoencoder learned for the $n_{emb} = 4$ case, we conduct the following experiment on the embedding space. We calculate the means and the standard deviations for all parameters (dimensions) of the embedding space for the line profiles in the test data set. Next, we vary one of the four dimensions in the range of the mean $\pm$ two standard deviations, while keeping the other three parameters fixed and equal to their mean values. The result of the experiment is presented in Figure~\ref{figure:diminterp}.
As one can see, sometimes the meaning of the dimensions learned by the autoencoder could be intuitively interpreted. For example, dimension\,\#3 is mostly responsible for the line profile asymmetry, dimension\,\#4 is mostly responsible for the central-reversal feature depth, and dimension\,\#2 significantly affects the line profile width. The meaning of dimension\,\#1 is less clear, but it seems to control the line profile continuum with some effect on the line profile itself. Overall, it is very intriguing that the features learned by the network for the $n_{emb} = 4$ case can be interpreted and explained, to a certain degree, relatively simply. We would like to note that the optimal embedding layer dimensionality of $n_{emb} = 4$ and the interpretation of the derived features are done for the Mg\,II\,k line formed at the quiet Sun and should not be assumed by default for any other line profile or line formation conditions without testing.
Another application where autoencoders are often used is noise reduction. Figure~\ref{figure:examples} demonstrates that: the reconstructed line profiles (orange curves) look much smoother in comparison to the original line profiles (blue curves). Moreover, one can notice the small absorption feature on the left wing of the Mg\,II\,k line. This absorption is especially visible in Fig.~\ref{figure:examples}\textit{b,c}. This feature corresponds to the Mn\,I\,2795.64\,$\AA$ spectral line profile (see IRIS technical notes, \url{https://iris.lmsal.com/itn26/introduction.html}). It is again remarkable how the network captures it from the noisy spectroscopic data. Possibilities to reveal such weak spectral signatures from noisy data may also be utilized in the future.
Finally, a relatively good reconstruction of the Mg\,II\,k line profile with the $n_{emb} = 4$ setting may have several practical applications, especially in situations where the data flow capabilities are limited. Compressing the data on deep space missions may result in more data samples transmitted to the ground via limited telemetry bandwidth. Low dimensionality of the embedding space allows one to perform a similarity search-based data retrieval~--- one can search data sets based on compact embedding instead of the entire data volume. Finally, line profile denoising, as we saw, may be useful for extracting weak signals in the continuum variations or weak spectroscopic signatures from noisy data.
\section{Conclusions}
\label{section:conclusion}
In this paper we have studied the compact embedding of spectroscopic observations of the quiet Sun in the Mg\,II\,k line by NASA's IRIS satellite using the fully-connected autoencoders. The key outcomes of this study are as follows:
\begin{itemize}
\item It is possible to compress the data more than 27 times (i.e., to reduce the data dimensionality from 110 to 4) while having only a 4\,DN reconstruction error on average. The reconstruction error is somewhat comparable to the variations of the measurements in the line continuum observed for the line profiles;
\item The average error of reconstruction of the MSE and even statistical moments of the line profiles decreases sharply for the $n_{emb} \leq{} 4$ and barely drops for $n_{emb} > 4$. Reconstruction of the odd statistical line moments does not demonstrate such dependence on the dimensionality of the embedding layer (except for the $\mu{}_{1}$ statistical moment);
\item Some occasional improvements of the autoencoder training were observed for $n_{emb} > 4$, e.g., reducing MSE by $\sim$20\% for $n_{emb} = 8$ and by 55\% for $n_{emb} = 14$. However, these results were observed only twice among more than 100 runs, indicating that obtaining a better embedding requires, probably, longer training times and another training/architecture strategy;
\item The features learned for the $n_{emb} = 4$ case can be supported with an intuitively-understandable interpretation if variations in the embedding space are considered. Three of the identified features were found to be mostly responsible for defining the line width, asymmetry, and line dip formation respectively.
\end{itemize}
In the presented study we considered only the Mg\,II\,k line profiles observed at a quiet Sun disc center region with a certain configuration of the observing instrument (exposure and resolution). While we expect that the techniques of compression of line profiles may be extended to other line profiles, this will require a detailed assessment. The presented work considered a particular architecture of the neural network, a particular way of varying the embedding layer dimensionality, and a certain methodology to train the network. Using other architectures such as the convolutional neural networks and variational autoencoders \cite{Kingma14}, as well as other training strategies, may help to better compress the line profiles and will be investigated later. Nevertheless, the work illustrates the potential of line-profile compact embedding for both operational and scientific purposes.
\section*{Acknowledgment}
VMS is supported by NSF FDSS grant 1936361 and NSF grant 1835958. EI is supported by RSF grant 20-72-00106. The research is partially supported by NASA grants NNX12AD05A, 80NSSC20K0302, and NSF EarthCube grant 1639683. IRIS is a NASA small explorer mission developed and operated by LMSAL with mission operations executed at NASA Ames Research center and major contributions to downlink communications funded by ESA and the Norwegian Space Centre.
|
\section{Introduction}
Lexical Semantic Change (LSC) Detection has drawn increasing attention in the past years \citep{kutuzov-etal-2018-diachronic,2018arXiv181106278T,hengchen2021challenges}. Recently, SemEval-2020 Task 1 and the Italian follow-up task DIACR-Ita provided a multi-lingual evaluation framework to compare the variety of proposed model architectures \citep{schlechtweg-etal-2020-semeval,diacrita_evalita2020}. Both tasks demonstrated that type-based embeddings outperform token-based embeddings. This is surprising given that contextualised token-based approaches have achieved significant improvements over the static type-based approaches in several NLP tasks over the past years \citep{peters-etal-2018-deep,devlin-etal-2019-bert}.
In this study, we relate model results on LSC detection to results on the word sense disambiguation data set underlying SemEval-2020 Task 1. This allows us to test the performance of different methods more rigorously, and to thoroughly analyze results of clustering-based methods.
We investigate the influence of a range of variables on clusterings of BERT vectors and show that its low performance is largely due to orthographic information on the target word which is encoded even in the higher layers of BERT representations. By reducing the influence of orthography on the target word while keeping the rest of the input in its natural form we considerably improve BERT's performance.
\section{Related work}
Traditional approaches for LSC detection are type-based \citep{Dubossarskyetal19,Schlechtwegetal19}. This means that not every word occurrence is considered individually (token-based); instead, a general vector representation that summarizes every occurrence of a word (including polysemous words) is created. The results of SemEval-2020 Task 1 and DIACR-Ita \citep{diacrita_evalita2020,schlechtweg-etal-2020-semeval} demonstrated that overall type-based approaches \citep{asgari-etal-2020-emblexchange,kaiser-etal-2020-roots,prazak-etal-2020-uwb} achieved better results than token-based approaches \citep{beck-2020-diasense,kutuzov-giulianelli-2020-uiouva,laicher-etal-2020-volente}. This is surprising, however, for two main reasons: (i) contextualized token-based approaches have significantly outperformed static type-based approaches in several NLP tasks over the past years \citep{ethayarajh2019contextual}. (ii) SemEval-2020 Task 1 and DIACR-Ita both include a subtask on binary change detection that requires to discover small sets of contextualized usages with the same sense. Type-based embeddings do not infer usage-based (or token-based) representations and are therefore not expected to be able to find such sets \citep{schlechtweg-etal-2020-semeval}. Yet, they show better performance on binary change detection than clusterings of token-based embeddings \citep{kutuzov-giulianelli-2020-uiouva}.
\section{Data and evaluation}
We utilize the annotated English, German and Swedish datasets (ENG, GER, SWE) underlying SemEval-2020 Task 1 \citep{schlechtweg-etal-2020-semeval}. Each dataset contains a list of target words and a set of usages per target word from two time periods, $t_1$ and $t_2$ \citep{Schlechtweg2021dwug}. For each target word, a Word Usage Graph (WUG) was annotated, where nodes represent word usages, and weights on edges represent the (median) semantic relatedness judgment of a pair of usages, as exemplified in (\ref{ex:1}) and (\ref{ex:2}) for the target word \textit{plane}.
\begin{example}\label{ex:1}
Von Hassel replied that he had such faith in the \textbf{plane} that he had no hesitation about allowing his only son to become a Starfighter pilot.
\end{example}%
\begin{example}\label{ex:2}
This point, where the rays pass through the perspective \textbf{plane}, is called the seat of their representation.
\end{example}
The final WUGs were clustered with a variation of correlation clustering \citep{Bansal04}
(see Figure \ref{fig:graph1} in Appendix \ref{sec:wugs}, left) and split into two subgraphs representing nodes from $t_1$ and $t_2$ respectively (middle and right). Clusters are interpreted as senses, and changes in clusters over time are interpreted as lexical semantic change. \citeauthor{schlechtweg-etal-2020-semeval} then infer a binary change value $B(w)$ for Subtask 1 and a graded change value $G(w)$ for Subtask 2 from the two resulting time-specific clusterings for each target word $w$.
The evaluation of the shared task participants only relied on the change values derived from the annotation, while the annotated usages were not released. We gained access to the data set, which enables us to relate performances in change detection to the underlying data.\footnote{We had no access to the Latin annotated data. For the ENG clustering experiments we use the full annotated resource containing three additional graphs \citep{Schlechtweg2021dwug}.} We can also analyze the inferred clusterings with respect to bias factors, and compare their influence on inferred vs. gold clusterings. A further advantage of having access to the underlying data is that it reflects more accurately the annotated change scores. In SemEval-2020 Task 1 the annotated usages were mixed with additional usages to create the training corpora for the shared task, possibly introducing noise on the derived change scores.
\section{Models and Measures}
\label{sec:models}
\paragraph{BERT}
Bidirectional Encoder Representations from Transformers \citep[BERT,][]{devlin-etal-2019-bert} is a transformer-based neural language model designed to find contextualised representations for text by analysing left and right contexts. The base version processes text in 12 different layers. In each layer, a contextualized token vector representation is created for every word. A layer, or a combination of multiple layers (we use the average), serves as a representation for a token. For every target word, we feed the usages from the SemEval data set into BERT and use the respective pre-trained cased base model to create token embeddings.\footnote{We first clean the GER usages by replacing historical with modern characters.}
\paragraph{Clustering}
LSC can be detected by clustering the token vectors from $t_1$ and $t_2$ into sets of usages with similar meanings, and then comparing these clusters over time \citep[cf.][]{Schutze1998,Navigli09}. This section introduces the clustering algorithms and clustering performance measures that we used. \textbf{Agglomerative Clustering} (AGL) is a hierarchical clustering algorithm starting with each element in an individual cluster. It then repeatedly merges those two clusters whose merging maximizes a predefined criterion. We use Ward's method, where clusters with the lowest loss of information are merged \citep{ward1963hierarchical}. Following \citet{giulianelli-etal-2020-analysing} and \citet{Martinc2020evolution}, we estimate the number of clusters $k$ with the \textbf{Silhouette Method} \citep{rousseeuw1987silhouettes}: we perform a cluster analysis for each $2 \leq k \leq 10$ and calculate the silhouette index for each $k$. The number of clusters with the largest index is used for the final clustering.
The \textbf{Jensen-Shannon Distance} (JSD) measures the difference between two probability distributions \citep{Lin1991,DonosoS17}. We convert two time specific clusterings into probability distribution $P$ and $Q$ and measure their distance $JSD(P,Q)$ to obtain graded change values \citep{giulianelli-etal-2020-analysing,kutuzov-giulianelli-2020-uiouva}. If $P$ and $Q$ are very similar, the JSD returns a value close to 0. If the distributions are very different, the JSD returns a value close to 1.
\textbf{Spearman's Rank-Order Correlation Coefficient} $\rho$ measures the strength and the direction of the relationship between two variables \citep{bolboaca2006pearson} by correlating the rank order of two variables. Its values range from -1 to 1, where 1 denotes a perfect positive relationship between the two variables, and -1 a perfect negative relationship. 0 means that the two variables are not related.
\paragraph{Cluster bias}
We perform a detailed analysis on what the inferred clusters actually reflect. We test hypotheses on \textit{word form}, \textit{sentence position}, \textit{number of proper names} and \textit{corpus}. The influence strength of each of these variables on the clusters is measured by the \textbf{Adjusted Rand Index} (ARI) \citep{Hubert/Arabie:85}
between the inferred cluster labels for each test sentence and a labeling for each test sentence derived from the respective variable. For the variable \textit{word form}, we assign the same label to each use where the target word has the same orthographic form (same string). If ARI = 1, then the inferred clusters contain only sentences where the target word has the same form.
For \textit{sentence position} each sentence receives label 0, if the target word is one of the first three words of the sentence, 2, if the target word is one of the last three words, else 1.\footnote{We assume that especially the beginning and ending of a sentence have a strong influence.}
For \textit{proper names} a sentence receives label 0, if no proper names are in the sentence, 1, if one proper name occurs, else 2.\footnote{The influence of proper names is only measured for ENG, since no POS-tagged data was readily available for GER.} The hypothesis that proper names may influence the clustering was suggested in \citet{martinc-etal-2020-context}. For \textit{corpora}, a sentence is labeled 0, if it occurs in the first target corpus, else 1.
\paragraph{Average measures}
Given two sets of token vectors $V_1$ and $V_2$ from $t_1$ and $t_2$, \textbf{Average Pairwise Distance} (APD) is calculated by randomly picking $n$ vectors from both sets, calculating their pairwise cosine distances $d(x,y)$ where $x \in V_1$ and $y \in V_2$ and averaging over these.
\citep{Schlechtwegetal18,giulianelli-etal-2020-analysing}. We determine $n$ as the minimum size of $V_1$ and $V_2$.
\textbf{APD-OLD/NEW} measure the average of pairwise distances within $V_1$ and $V_2$, respectively. They are calculated as the average distance of max. $10,000$ randomly sampled unique combinations of vectors from either $V_1$ or $V_2$. \textbf{COS} is calculated as the cosine distance of the respective mean vectors for $V_1$ and $V_2$ \citep{kutuzov-giulianelli-2020-uiouva}.
\begin{table*}[ht]
\center
\begin{adjustbox}{width=0.43\textwidth}
\begin{tabular}{ l | l | c c c }
\toprule
& \textbf{Layer} & \textbf{Token} & \textbf{Lemma} & \textbf{TokLem} \\
\midrule
\multirow{3}{*}{\rotatebox[origin=c]{0}{\textbf{$\rho$}}}
& 1 & -.141 & -.033 & .100\\
& 12 & .205 & .154 & .168 \\
& 9-12 & .325 & \textbf{.345} & .293 \\
\midrule
\multirow{3}{*}{\rotatebox[origin=c]{0}{\textbf{ARI}}}
& 1 & .022 & .041 & .045\\
& 12 & .116 & .111 & .158 \\
& 9-12 & .150 & .159 & \textbf{.163} \\
\midrule
\multicolumn{4}{c}{} \\
\midrule
\multirow{3}{*}{\rotatebox[origin=c]{0}{\textbf{Form}}}
& 1 & \textbf{.907} & .014 & .014\\
& 12 & \textbf{.389} & \textbf{.018} & \textbf{.077} \\
& 9-12 & \textbf{.334} & \textbf{.018} & \textbf{.051} \\
\midrule
\multirow{3}{*}{\rotatebox[origin=c]{0}{\textbf{Position}}}
& 1 & .001 & \textbf{.026} & \textbf{.024}\\
& 12 & \textbf{.012} & \textbf{.012} & \textbf{.015} \\
& 9-12 & .002 & \textbf{.007} & \textbf{.003} \\
\midrule
\multirow{3}{*}{\rotatebox[origin=c]{0}{\textbf{Corpora}}}
& 1 & \textbf{.019} & \textbf{.021} & \textbf{.033}\\
& 12 & \textbf{.078} & \textbf{.056} & \textbf{.082} \\
& 9-12 & \textbf{.056} & \textbf{.044} & \textbf{.063} \\
\midrule
\multirow{3}{*}{\rotatebox[origin=c]{0}{\textbf{Names}}}
& 1 & -.007 & .010 & .010\\
& 12 & \textbf{.025} & \textbf{.027} & \textbf{.033} \\
& 9-12 & .019 & \textbf{.022} & \textbf{.026} \\
\bottomrule
\end{tabular}
\end{adjustbox}
\quad
\begin{adjustbox}{width=0.43\textwidth}
\begin{tabular}{ l | l | c c c }
\toprule
& \textbf{Layer} & \textbf{Token} & \textbf{Lemma} & \textbf{TokLem} \\
\midrule
\multirow{3}{*}{\rotatebox[origin=c]{0}{\textbf{$\rho$}}}
& 1 & -.265 & -.062 & -.170\\
& 12 & .123 & .427 & \textbf{.624}\\
& 9-12 & .122 & .420 & .533\\
\midrule
\multirow{3}{*}{\rotatebox[origin=c]{0}{\textbf{ARI}}}
& 1 & .033 & .002 & .003\\
& 12 & .119 & .159 & \textbf{.161}\\
& 9-12 & .155 & .142 & .154\\
\midrule
\multicolumn{4}{c}{} \\
\midrule
\multirow{3}{*}{\rotatebox[origin=c]{0}{\textbf{Form}}}
& 1 & \textbf{.706} & .024 & .004\\
& 12 & \textbf{.439} & \textbf{.056} & \textbf{.150}\\
& 9-12 & \textbf{.420} & \textbf{.047} & \textbf{.094}\\
\midrule
\multirow{3}{*}{\rotatebox[origin=c]{0}{\textbf{Position}}}
& 1 & .005 & \textbf{.023} & \textbf{.027}\\
& 12 & -.002 & .005 & -.002\\
& 9-12 & \textbf{.009} & \textbf{.018} & \textbf{.012}\\
\midrule
\multirow{3}{*}{\rotatebox[origin=c]{0}{\textbf{Corpora}}}
& 1 & .074 & .003 & .005\\
& 12 & \textbf{.110} & \textbf{.095} & \textbf{.096}\\
& 9-12 & \textbf{.107} & .068 & \textbf{.089}\\
\midrule
\multirow{3}{*}{\rotatebox[origin=c]{0}{\textbf{Names}}}
& 1 & - & - & -\\
& 12 & - & - & -\\
& 9-12 & - & - & -\\
\bottomrule
\end{tabular}
\end{adjustbox}
\caption{Overview of English clustering scores (left) and German clustering scores (right). Bold font indicates best scores for $\rho$ and ARI (top) or scores above all corresponding baselines for influence variables (bottom).}
\label{tbl: ClusterInfluence short}
\end{table*}
\section{Results}
\subsection{Clustering}
\label{clustering}
Because of the high computational load, we apply the clustering only to the ENG and the GER part of the SemEval data set. For this, we use BERT to create token vectors and cluster them with AGL, as described above. We then perform a detailed analysis of what the clusters reflect.\footnote{We also run most of our experiments with k-means \citep{Forgy:65}. Both algorithms performed similarly with a slight advantage for AGL. We therefore only report the results achieved using AGL.}
We report a subset of the clustering experiment results in Table \ref{tbl: ClusterInfluence short}, the complete results are provided in Appendix \ref{sec:performance}.
Table \ref{tbl: ClusterInfluence short} shows JSD performance on graded change ($\rho$), clustering performance (ARI) as well as the ARI scores for the influence factors introduced above, across BERT layers. For each influence factor we add two baselines: (i) The random baseline measures the ARI score of the influence factor using random cluster labels, and (ii) the actual baseline measures the ARI score between the true cluster labels and the influence factor. In other words, (i) and (ii) respectively answer the question of how strong the influence factor is by chance, and how strong it is according to the human annotation. The values of the two baselines are crucial: If an influence factor has an ARI score greater than both baselines, the clustering reflects the influence factor more than expected. If additionally the influence factor has an ARI score greater than the actual performance ARI score, the clustering reflects the partitioning according to the influence factor more than the clustering derived from human annotations.
\paragraph{Word form bias} As explained above, the word form influence measures how strongly the inferred clusterings represent the orthographic forms of the target word. Table \ref{tbl: ClusterInfluence short} shows that for both GER and ENG the form bias of the raw token vectors (column `Token') is extremely high and always yields the highest influence score for each layer combination of BERT. Additionally, the influence of the word form is significantly higher when using lower layers of BERT. This fits well with the observations of \citet{jawahar2019does} that the lower layers of BERT capture surface features, the middle layers capture syntactic features and the higher layers capture semantic features of the text. With the first layer of BERT the sentences are almost exclusively ($.9$) clustered according to the form of the target word (e.g. plural/singular division). Even in the higher layers word form influence is considerable in both languages (layer 12: $\approx .4$). This strongly overlays the semantic information encoded in the vectors, as we can see in the low $\rho$ and ARI scores, which are negatively correlated with word form influence.\footnote{Note that it is very difficult to reach high ARI scores because ARI incorporates chance.}
The word form bias seems to be lower in GER than in ENG (layer 1: $.7$ vs. $.9$). However, this is misleading, as our approach to measure word form influence does not capture cases where vectors cluster according to subword forms as in the case of \textit{Ackergerät}. Its word forms differ as to whether they are written with an `h' or not, as in \textit{Ackergerät} vs. \textit{Ackergeräth}. As a manual inspection shows this is strongly reflected in the inferred clustering. However, these forms then further subdivide into inflected forms such as \textit{Ackergeräthe} and \textit{Ackergeräthes}, which is reflected in our influence variable. For these cases, our approach tends to underestimate the influence of the variable.
In order to reduce the influence of word form we experiment with two pre-processing approaches: (i) We feed BERT with lemmatised sentences (Lemma) instead of raw ones. (ii) We only replace the target word in every sentence with its lemma (TokLem). TokLem is motivated by the fact that BERT is trained on raw text. Thus, we assume that BERT is more familiar with non-lemmatised sentences and therefore expect it to work better on raw text. In order to continue working with non-lemmatised sentences we only remove the target word form bias by exchanging the target word with its lemma.
As we can see in Table \ref{tbl: ClusterInfluence short}, lemmatisation strongly reduces the influence of word form, as expected.\footnote{In some cases it is however above the baselines, indicating that word form is correlated with other sentence features.} Accordingly, $\rho$ and ARI improve. However, it also leads to deterioration in some cases. Also, TokLem reduces the influence of word form and in most cases yields the overall maximum performance. The ARI scores for both languages are similar ($\approx .160$) while the $\rho$ performance varies very strongly between languages, achieving a very high score for GER ($.624$).
Replacing the target word by its lemma form seems to shift the word form influence in the different layers: Especially for GER, layers 1 and 1+12 show the highest influences ($.706$ and $.687$) with Token (see also Appendix \ref{sec:performance}). In combination with TokLem, both layers are influenced the least ($.004$ and $.046$). For ENG we see the same effect for layer 1.
\paragraph{Other bias factors}
We can see in Table \ref{tbl: ClusterInfluence short} that most influences are above-baseline. As explained above, the word form bias heavily decreases using higher layers of BERT. For all other influences the bias increases when using high layers of BERT. This may be because decreasing the word form influence reveals the existence of further --less strong but still relevant-- influences. The same is observable with the Lemma and TokLem results, since there the form influence is decreased or even eliminated. While for ENG the influence scores mostly increase using Lemma and TokLem, for GER only the position influence increases, while corpora influence decreases. This is probably because the corpora influence is to some extent related to word form, which often reflects time-specific orthography as in \textit{Ackergeräth} vs. \textit{Ackergerät}, where the spelling with the "h" mostly occurs in the old corpus.
Influence of position and proper names seems to be less important but the respective scores are still most of the times higher than the baselines. So overall the reflection of the two corpora seems to be the most influential factor apart from word form. Often the corpus bias is almost as high as the actual ARI score.
\subsection{Average Measures}
For the average measures we perform experiments for all three languages (ENG, GER, SWE).
\paragraph{Layers}
Because we observe a strong variation of influence scores with layers, as seen in Section (\ref{clustering}), we test different layer combinations for the average measures. The following are considered: 1, 12, 1+12, 1+2+3+4 (1-4), 9+10+11+12 (9-12). As shown in Table \ref{tab:performance1}, the choice of the layers strongly affects the performance. We see that for APD the higher layer combinations 12 and 9-12 perform best across all three languages, while the latter is slightly better ($.571$, $.407$ and $.554$). Interestingly, these two are the only layer combinations that do not include layer 1. All three layer combinations that include layer 1 are significantly worse in comparison.
While COS performs best with layer combination 1-4 for ENG ($.390$), for GER and SWE we see a similar trend as with APD. Again, the higher layer combinations perform better than the other three, which all include layer 1. For GER layer combination 12 ($.472$) performs best, while 9-12 yields the highest result for SWE ($.183$). Our results are mostly in line with the findings of \citet{kutuzov-giulianelli-2020-uiouva} that APD works best on ENG and SWE, while COS yields the best scores for GER.
\begin{table}[t]
\center
\begin{adjustbox}{width=0.48\textwidth}
\begin{tabular}{ l | c c c | c c c }
\toprule
\multirow{2}{*}{\textbf{Layer}}& \multicolumn{3}{c |}{\textbf{APD}} & \multicolumn{3}{c}{\textbf{COS}} \\
& ENG & GER & SWE & ENG & GER & SWE \\
\hline
1 & .297 & .205 & .228 & .246 & .246 & .089 \\
12 & .566 & .359 & .529 & .339 & \textbf{.472} & .134 \\
1+12 & .455 & .316 & .280 & .365 & .373 & .077 \\
1-4 & .431 & .227 & .355 & \textbf{.390} & .297 & .079 \\
9-12 & \textbf{.571} & \textbf{.407} & \textbf{.554} & .365 & .446 & \textbf{.183} \\
\bottomrule
\end{tabular}
\end{adjustbox}
\caption{Token performance for different layer combinations across languages.}\label{tab:performance1}
\vspace{-7ex}
\end{table}
\paragraph{Pre-processing}
As with the clustering, we try to improve the performance of the average measures by using the two above-described pre-processing approaches. We perform experiments only for three layer combinations in order to reduce the complexity: (i) 12 and (ii) 9-12 perform best and are therefore obvious choices. (iii) From the remaining combinations 1+12 shows the most stable performance across measures and languages. Table \ref{tab:pre-process} shows the performance of the pre-processings (Lemma, TokLem) over these three combinations. We can see that both APD and COS perform slightly worse for ENG when paired with a pre-processing (exception to this is 1+12 Lemma). In contrast, GER profits heavily: While APD with layer combinations 12 and 9-12 performs slightly worse with Lemma, and slightly better with TokLem, we observe an enormous performance boost for layer combination 1+12 ($.643$ Lemma and $.731$ TokLem). We achieve a similar boost for all three layer combinations with COS as a measure. We reach a top performance of $.755$ for layer 12 with TokLem. SWE does not benefit from Lemma. We observe large performance decreases, with the exception of combination 1+12 (APD). The APD performance of layers 12 and 9-12 is slightly worse with TokLem. However, layers 1+12, which performed poorly without pre-processing, reaches peak performance of $.602$ with TokLem. All COS performances increase with TokLem, but are still well below the APD counterparts. The general picture is that GER and SWE profit strongly from TokLem.
\begin{table}[t]
\centering
\begin{adjustbox}{width=0.43\textwidth}
\begin{tabular}{ c | c | l | c c c }
\toprule
& & \textbf{Layer} & \textbf{Token} & \textbf{Lemma} & \textbf{TokLem} \\
\hline
\multirow{6}{*}{\rotatebox[origin=c]{90}{\textbf{ENG}}} & \multirow{3}{*}{\rotatebox[origin=c]{90}{\textbf{APD}}} & 12 & \textbf{.566} & .483 & .494\\
& & 1+12 & .455 & \textbf{.483} & .455 \\
& & 9-12 & \textbf{.571} & .493 & .547 \\
\cline{2-6}
&\multirow{3}{*}{\rotatebox[origin=c]{90}{\textbf{COS}}} & 12 & \textbf{.339} & .251 & .331 \\
& & 1+12 & \textbf{.365} & .239 & .193 \\
& & 9-12 & \textbf{.365} & .286 & .353 \\
\midrule
\multirow{6}{*}{\rotatebox[origin=c]{90}{\textbf{GER}}} & \multirow{3}{*}{\rotatebox[origin=c]{90}{\textbf{APD}}} & 12 & .359 & .303 & \textbf{.456}\\
& & 1+12 & .316 & .643 & \textbf{.731} \\
& & 9-12 & .407 & .305 & \textbf{.516} \\
\cline{2-6}
&\multirow{3}{*}{\rotatebox[origin=c]{90}{\textbf{COS}}} & 12 & .472 & .693 & \textbf{.755} \\
& & 1+12 & .373 & .698 & \textbf{.729} \\
& & 9-12 & .446 & .689 & \textbf{.726} \\
\midrule
\multirow{6}{*}{\rotatebox[origin=c]{90}{\textbf{SWE}}} & \multirow{3}{*}{\rotatebox[origin=c]{90}{\textbf{APD}}} & 12 & \textbf{.529} & .214 & .505\\
& & 1+12 & .280 & .368 & \textbf{.602} \\
& & 9-12 & \textbf{.554} & .218 & .531 \\
\cline{2-6}
&\multirow{3}{*}{\rotatebox[origin=c]{90}{\textbf{COS}}} & 12 & .134 & -.019 & \textbf{.285} \\
& & 1+12 & .077 & .012 & \textbf{.082} \\
& & 9-12 & .183 & -.002 & \textbf{.284} \\
\bottomrule
\end{tabular}
\end{adjustbox}
\caption{Performance of pre-processing variants for three layer combinations.}
\vspace{-12mm}
\label{tab:pre-process}
\end{table}
\paragraph{Word form bias}
In order to better understand the effects of layer combinations and pre-processing, we compute correlations between word form and model predictions. To lessen the complexity, only layer combination 1+12 (which performed worst overall and includes layer 1), layer combination 9-12 (which performed best overall) in combination with Token and the superior TokLem are considered. The results are presented in Table \ref{tab:avm_wordform}. We observe similar findings for all three languages. The correlation between word form and APD predictions is strong ($.613$, $.554$ and $.730$) for layers 1+12 without pre-processing. The correlation is much weaker with layers 9-12 ($.068$, $.292$ and $.237$) or TokLem ($-.026$, $.105$ and $.176$). This is in line with the
performance development that also increases using layers 9-12 or TokLem. Both approaches (different layers, pre-processing) result in a considerable performance increase as described previously. Using layer combination 9-12 with TokLem further decreases the correlation (with the exception of ENG). However, the performance is better when only one of these approaches is used.
The correlation between word form and COS model predictions is weaker overall ($.246$, $.387$ and $.429$). We see a similar correlation development as for APD, however this time the performance of ENG does not profit from the lowered bias (see Table \ref{tab:pre-process}). Both GER and SWE see a performance increase when the word form bias is lowered by either using layers 9-12 or TokLem.
\begin{table}[]
\centering
\begin{adjustbox}{width=0.34\textwidth}
\begin{tabular}{c | c | l | c c }
\toprule
& & \textbf{Layer} & \textbf{Token} & \textbf{TokLem} \\
\hline
\multirow{4}{*}{\rotatebox[origin=c]{90}{\textbf{ENG}}} & \multirow{2}{*}{\rotatebox[origin=c]{90}{\textbf{APD}}} & 1+12 & .613 & -.026 \\
& & 9-12 & .068 & .090 \\
\cline{2-5}
& \multirow{2}{*}{\rotatebox[origin=c]{90}{\textbf{COS}}} & 1+12 & .246 & -.062 \\
& & 9-12 & .020 & .004 \\
\midrule
\multirow{4}{*}{\rotatebox[origin=c]{90}{\textbf{GER}}} & \multirow{2}{*}{\rotatebox[origin=c]{90}{\textbf{APD}}} & 1+12 & .554 & .271 \\
& & 9-12 & .292 & .105 \\
\cline{2-5}
& \multirow{2}{*}{\rotatebox[origin=c]{90}{\textbf{COS}}} & 1+12 & .387 & -.017 \\
& & 9-12 & .205 & -.008 \\
\midrule
\multirow{4}{*}{\rotatebox[origin=c]{90}{\textbf{SWE}}} & \multirow{2}{*}{\rotatebox[origin=c]{90}{\textbf{APD}}} & 1+12 & .730 & .176 \\
& & 9-12 & .237 & .048 \\
\cline{2-5}
& \multirow{2}{*}{\rotatebox[origin=c]{90}{\textbf{COS}}} & 1+12 & .429 & -.031 \\
& & 9-12 & .277 & -.035 \\
\bottomrule
\end{tabular}
\end{adjustbox}
\caption{Correlations of word form and predicted change scores.}
\vspace{-12mm}
\label{tab:avm_wordform}
\end{table}
\paragraph{Polysemy bias}
The SemEval data sets are strongly biased by polysemy, i.e., a perfect model measuring the true synchronic target word polysemy in either $t_1$ or $t_2$ could reach above $.7$ performance \citep{schlechtweg-etal-2020-semeval}. We use APD-OLD and APD-NEW (see Section \ref{sec:models}) to see whether we can exploit this fact to create a purely synchronic polysemy model with high performance. We achieve moderate performances for ENG and GER ($.274$/$.332$ and $.321$/$.450$ respectively) and a good performance for SWE ($.550$/$.562$). While the performance for ENG and GER is clearly below the high-scores, the performance is high for a measure that lacks any kind of diachronic information. And in the case of SWE, the performance of both APD-OLD and APD-NEW is just barely below the high-scores (cf. Table \ref{tab:pre-process}). Note that regular APD (in contrast to COS) is, in theory, affected by polysemy \citep{Schlechtwegetal18}. It is thus possible that APD's high performance stems at least partly from this polysemy bias. This is supported by comparing the SWE results of APD and COS in Table \ref{tab:pre-process}: COS is weakly influenced by polysemy and performs poorly, while APD has higher performance, but only slightly above the purely synchronic measures APD-OLD/NEW.
\section{Conclusion}
BERT token representations are influenced by various factors, but most strongly by target word form. Even in higher layers this influence persists. By removing the form bias we were able to considerably improve the performance across languages. Although we reach comparably high performance with clustering for graded change detection in German, average measures still perform better than cluster-based approaches. The reasons for this are still unclear and should be addressed in future research. Furthermore, we used BERT without fine-tuning. It would be interesting to see how fine-tuning interacts with influence variables and whether it further improves performance.
|
\section{Introduction}
\subsection{Basic Definitions And Notations}
Let $A \in \mathbb{Z}^{m \times n}$ be an integer matrix. We denote by $A_{ij}$ the $ij$-th element of the matrix, by $A_{i*}$ its $i$-th row, and by $A_{*j}$ its $j$-th column. The set of integer values from $i$ to $j$, is denoted by $\intint[i]j=\left\{i, i+1, \ldots, j\right\}$. Additionally, for subsets $I \subseteq \{1,\dots,m\}$ and $J \subseteq \{1,\dots,n\}$, the symbols $A_{IJ}$ and $A[I,J]$ denote the sub-matrix of $A$, which is generated by all the rows with indices in $I$ and all the columns with indices in $J$. If $I$ or $J$ are replaced by $*$, then all the rows or columns are selected, respectively. Sometimes, we simply write $A_{I}$ instead of $A_{I*}$ and $A_{J}$ instead of $A_{*J}$, if this does not lead to confusion.
The maximum absolute value of entries in a matrix $A$ is denoted by $\|A\|_{\max} = \max_{i,j} |A_{i\,j}|$. The $l_p$-norm of a vector $x$ is denoted by $\|x\|_p$.
\begin{definition}
For a matrix $A \in \ZZ^{m \times n}$, by $$
\Delta_k(A) = \max\{|\det A_{IJ}| \colon I \subseteq \intint m,\, J \subseteq \intint n,\, |I| = |J| = k\},
$$ we denote the maximum absolute value of determinants of all the $k \times k$ sub-matrices of $A$.
Clearly, $\Delta_1(A) = \|A\|_{\max}$. Additionally, let $\Delta(A) = \Delta_{\rank(A)}(A)$.
\end{definition}
\subsection{Description of Results and Related Work}
Let $A \in \ZZ_+^{m \times n}$, $b \in \ZZ_+^m$, $c \in \ZZ_+^n$ and $u \in \ZZ^{n}_+$. \emph{The bounded $m$-dimensional knapsack problem (shortly $m$-BKP)} can be formulated as follows:
\begin{gather}
c^\top x \to \max \notag\\
\begin{cases}
A x \leq b \\
0 \leq x \leq u \\
x \in \ZZ^n.
\end{cases}\tag{$m$-BKP}\label{main_prob}
\end{gather}
It is well known that the $m$-BKP is $NP$-hard already for $m = 1$. However, it is also well known that the $1$-BKP admits an FPTAS. The historically first FPTAS for the $1$-BKP was given in the seminal work of O.~Ibarra and C.~Kim \cite{IK}. The results of \cite{IK} were improved in many ways, for example in the works \cite{Chan,ParamKP,ParamWKP,FastUKP,CeJin,KP1,KP2,LAW,CarMKP,MO,Rhee}. But, it was shown in \cite{NoFPTAS} (see \cite[p.~252]{TheMKP} for a simplified proof) that the $2$-BKP does not admit an FPTAS unless $P = NP$. Due to \cite{NoEPTAS}, the $2$-BKP does not admit an EPTAS unless $W[1] = FPT$. However, the $m$-BKP still admits a PTAS. To the best of our knowledge, the state of the art PTAS is given in \cite{BestPTAS}. The complexity bound proposed in \cite{BestPTAS} is $O(n^{\lceil \frac{m}{\varepsilon} \rceil - m})$. The perfect survey is given in the book \cite{TheMKP}.
Within the scope of the article, we are interested in studying $m$-BKP problems with a special restriction on sub-determinants of the constraints matrix $A$. More precisely, we assume that all rank-order minors of $A$ are bounded in an absolute value by $\Delta$. We will call this class of $m$-BKPs as \emph{$\Delta$-modular $m$-BKPs}. The main result of the paper states that the $\Delta$-modular $m$-BKP admits an FPTAS, whose complexity bound depends on $\Delta$ linearly, for any fixed $m$.
\begin{theorem}\label{main_th}
The $\Delta$-modular \ref{main_prob} admits an FPTAS with the arithmetical complexity bound
$$
O(T_{LP} \cdot (1/\varepsilon)^{m+3} \cdot (2m)^{2m + 6} \cdot \Delta),
$$ where $T_{LP}$ is the linear programming complexity bound.
\end{theorem} Proof of the theorem is given in Section \ref{proof_sec}.
Due to the seminal work of N.~Megiddo \cite{MEG}, the linear program can be solved by a linear-time algorithm if $m$ is fixed.
\begin{corollary}\label{main_cor}
For fixed $m$ the complexity bound of Theorem \ref{main_th} can be restated as
$$
O(n \cdot (1/\varepsilon)^{m+3} \cdot \Delta).
$$
\end{corollary}
We need to note that results of the papers \cite{STEINITZILP,FPT18,ProximityUseSparsity} can be combined to develop an exact polynomial-time algorithm for the considered $\Delta$-modular $m$-BKP problem, and even more, for any $\Delta$-modular ILP problem in standard form with a fixed number of constraints $m$. But, the resulting algorithm complexity contains a non-linear dependence on $\Delta$ in contrast with the developed FPTAS. The precise formulation will be given in the following Theorem \ref{Delta_ILP_th} and Corollary \ref{Delta_ILP_cor}. First, we need to make some definitions:
\begin{definition}
Let $A \in \ZZ^{m \times n}$, $b \in \ZZ^m$, $c \in \ZZ^n$, $u \in \ZZ^n_{+}$, $\rank(A) = m$ and $\Delta = \Delta(A)$. \emph{The bounded $\Delta$-modular ILP in standard form} (shortly $m$-BILP) can be formulated as follows:
\begin{gather}
c^\top x \to \max \notag \\
\begin{cases}
A x = b \\
0 \leq x \leq u \\
x \in \ZZ^{n}.
\end{cases} \tag{$m$-BILP}\label{ILP_standard}
\end{gather}
The main difference between the problems \ref{ILP_standard} and \ref{main_prob} is that the input of the problem \ref{ILP_standard} can contain negative numbers. The inequalities of the problem \ref{main_prob} can be turned to equalities using slack variables.
\end{definition}
\begin{definition}\label{H_def}
Consider the problem \ref{ILP_standard}. Let $z^*$ be an optimal solution of \ref{ILP_standard} and $x^*$ be an optimal vertex-solution of the LP relaxation of \ref{ILP_standard}. \emph{The $l_1$-proximity bound $H$} of the problem \ref{ILP_standard} is defined by the formula
$$
H = \max_{x^*}\min_{z^*} \|x^* - z^*\|_1.
$$
It is proven in \cite{STEINITZILP} that
\begin{equation}\label{Delta1_H_bound}
H \leq m \cdot (2 m \cdot \Delta_1 + 1)^m,\quad\text{where $\Delta_1 = \Delta_1(A) = \|A\|_{\max}$.}
\end{equation}
It was noted in \cite[formula (4)]{ProximityUseSparsity} that this proximity bound \eqref{Delta1_H_bound} of the paper \cite{STEINITZILP} can be restated to work with the parameter $\Delta(A)$ instead of $\Delta_1(A)$. More precisely, there exists an optimal solution $z^*$ of the \ref{ILP_standard} problem such that \begin{equation*}\label{Delta_H_bound}
H \leq m\cdot(2m+1)^m \cdot \Delta,\quad\text{where $\Delta = \Delta(A)$.}
\end{equation*}
\end{definition}
\begin{theorem}\label{Delta_ILP_th}
The \ref{ILP_standard} problem can be solved by an algorithm with the following arithmetical complexity:
\begin{equation*}
n \cdot {O(H + m)}^{m+1} \cdot \log^2(H) \cdot \Delta + T_{LP}.
\end{equation*}
The previous complexity bound can be slightly improved in terms of $H$:
\begin{gather*}
n \cdot O(\log m)^{m^2} \cdot {(H + m)}^{m}\cdot \Delta + T_{LP}.
\end{gather*}
\end{theorem}
The proof can be found in Section \ref{proof_Delta_ILP_th}.
\begin{remark}\label{hash_table_rm}
The algorithms described in the proof of Theorem \ref{Delta_ILP_th} are using hash tables with linear expected constructions time and constant worst-case lookup time to store information dynamic tables. An example of a such hash table can be found in the book \cite{CORMEN}. So, strictly speaking, algorithms of Theorem \ref{Delta_ILP_th} are randomized.
Randomization can be removed by using any balanced search-tree, for example, $RB$-tree \cite{CORMEN}. It will lead to additional logarithmic term in the complexity bound.
\end{remark}
Applying the proximity bounds \eqref{Delta_H_bound} and \eqref{Delta1_H_bound} to the previous Theorem \ref{Delta_ILP_th}, we can obtain estimates that are independent of $H$. For example, we obtain the following corollary:
\begin{corollary}\label{Delta_ILP_cor}
The problem \ref{ILP_standard} can be solved by an algorithm with the following arithmetical complexity bound:
\begin{gather*}
n \cdot O(\log m)^{m^2} \cdot O(m)^{m^2+m} \cdot \Delta^{m+1} + T_{LP}\quad\text{ and }\\
n \cdot \Delta^{m + 1}, \quad\text{for $m$ being fixed}.
\end{gather*}
\end{corollary}
\begin{remark}\label{BKNAP_rm}
Taking $m = 1$ in the previous corollary we obtain the
$
O(n \cdot \Delta^2)
$ complexity bound for the classical bounded knapsack problem, where $\Delta$ is the maximal absolute value of item weights. Our bound is better than the previous state of the art bounds $O(n^2 \cdot \Delta^2)$ and $O(n \cdot \Delta^2 \cdot \log^2 \Delta)$ due to \cite{STEINITZILP}.
\end{remark}
Better complexity bound for searching of an exact solution can be achieved for the unbounded version of the \ref{ILP_standard} problem. More precisely, for this case, the paper \cite{CONVILP} gives the complexity bound
$$
O(\sqrt{m} \Delta)^{2m} + T_{LP}.
$$ We note that the original complexity bound from the work \cite{CONVILP} is stated with respect to the parameter $\Delta_1(A) = \|A\|_{\infty}$ instead of $\Delta(A)$ (see the next Remark \ref{Delta1_complexity_rm}), but, due to Lemma 1 of \cite{FPT18}, we can assume that $\Delta_1(A) \leq \Delta(A)$.
\begin{remark}\label{Delta1_complexity_rm}
Another interesting parameter of the considered problems \ref{main_prob} and \ref{ILP_standard} is $\Delta_1(A) = \|A\|_{\max}$. Let us denote $\Delta_1 = \Delta_1(A)$. The first exact quasipolynomial-time algorithm for \ref{ILP_standard} was constructed in the seminal work of C.~H.~Papadimitriou \cite{PAPA}. The result of \cite{PAPA} was recently improved in \cite{STEINITZILP}, where it was shown that the \ref{ILP_standard} can be solved exactly by an algorithm with the arithmetical complexity
\begin{gather}
n \cdot O(m)^{(m+1)^2} \cdot \Delta_1^{m(m+1)} \cdot \log^2(m \Delta_1) + T_{LP}\quad\text{ and }\notag\\
n \cdot \Delta_1^{m(m+1)} \cdot \log^2(\Delta_1), \quad\text{ for $m$ being fixed}.\label{Delta1_ILP_bound}
\end{gather}
Due to the results of \cite{CONVILP}, the unbounded version of the problem can be solved by an algorithm with the arithmetical complexity
\begin{equation}\label{Delta1_UILP_bound}
O(\sqrt{m} \Delta_1)^{2m} + T_{LP}.
\end{equation}
The results of our note can be easily restated to work with the $\Delta_1$ parameter. Using the inequality \eqref{Delta1_H_bound}, the arithmetical complexity bound of Corollary \ref{Delta_ILP_cor} becomes
\begin{gather*}
n \cdot O(\log m)^{m^2} \cdot O(m)^{m^2 + m} \cdot \Delta_1^{m(m+1)} + T_{LP}\quad\text{ and }\notag\\
n \cdot \Delta_1^{m(m+1)}, \quad\text{ for $m$ being fixed},
\end{gather*}
which is slightly better, than the bound \eqref{Delta1_ILP_bound} of \cite{STEINITZILP}.
Additionally, Corollary \ref{Delta_ILP_cor} gives currently best bound $O(n \cdot \Delta^2_1)$ for the classical $1$-dimensional bounded knapsack problem, see Remark \ref{BKNAP_rm}.
The analogue result can be stated for our FPTAS. Definitely, for $\gamma > 0$ and $M = \{ y = A x \colon x \in \RR_+^n,\, \|x\|_1 \leq \gamma\}$ we trivially have $|M \cap \ZZ^m| \leq (\gamma \Delta_1)^m$. Applying the algorithm from Section \ref{proof_sec} to this analogue of Corollary \ref{DP_width_cor}, it gives an algorithm with the arithmetical complexity
\begin{gather*}
O(T_{LP} \cdot (1/\varepsilon)^{m+3} \cdot m^{2m + 6} \cdot (2\Delta_1)^{m}) \quad\text{ and }\notag\\
O(n \cdot (1/\varepsilon)^{m+3} \cdot \Delta_1^{m}), \quad \text{ for $m$ being fixed}.\label{Delta1_FPTAS_bound}
\end{gather*}
For sufficiently large $\varepsilon$ the last bounds give a better dependence on $m$ and $\Delta_1$, than bounds \eqref{Delta1_ILP_bound} from \cite{STEINITZILP}.
\end{remark}
\begin{remark}[Why $\Delta$-modular ILPs could be interesting?]
It is well known that the Maximal Independent Set (shortly $\MAXIS$) problem on a simple graph $G = (V,E)$ can be formulated by the ILP
\begin{gather}
\BUnit^\top x \to \max \notag\\
\begin{cases}
A(G)\, x \leq \BUnit\\
x \in \{0,1\}^{|V|},
\end{cases}\tag{$\MAXIS$}\label{IS_ILP}
\end{gather}
\end{remark} where $A(G) \in \{0,1\}^{|E| \times |V|}$ is the edge-vertex incidence matrix of $G$. Due to the seminal work \cite{MinorsGraphs} $$\Delta(A(G)) = 2^{\nu(G)},$$ where $\nu(G)$ is the odd-cycle packing number of $G$. Hence, the existence of a polynomial-time algorithm for $\Delta$-modular ILPs will lead to the existence of a polynomial-time algorithm for the \ref{IS_ILP} problem for graphs with a fixed $\nu(G)$ value. Recently, it was shown in \cite{BIMODULAR_STRONG} that $2$-modular ILPs admit a strongly polynomial-time algorithm, and consequently, the $\MAXIS \in P$ for graphs with one independent odd-cycle. But, existence of a polynomial-time algorithms even for the $3$-modular or $4$-modular ILP problems is an interesting open question, as well as existence of a polynomial-time algorithm for the \ref{IS_ILP} problem on graphs with $\nu(G) = 2$. Finally, due to \cite{AZ}, if $\Delta(\bar A)$ is fixed, where $\bar A = \binom{\BUnit^\top}{A(G)}$ is the extended matrix of the ILP \ref{IS_ILP}, then the problem can be solved by a polynomial time algorithm. The shorter proof could be found in \cite{GRIBM17,GRIBM18}, as well as analogue results for vertex and edge Maximal Dominating Set problems. For recent progress on the \ref{IS_ILP} problem with respect to the $\nu(G)$ parameter see the papers \cite{BOCK14,STABLE_SET_GENUS,KNOWN_ODD_CYCLES}.
Additionally, we note that, due to \cite{BOCK14}, there are no polynomial-time algorithms for the \ref{IS_ILP} problem on graphs with $\nu(G) = \Omega(\log n)$ unless the ETH (the Exponential Time Hypothesis) is false. Consequently, with the same assumption, there are no algorithms for the $\Delta$-modular ILP problem with the complexity bound $\poly(s) \cdot \Delta^{O(1)}$, where $s$ is an input size. Despite the fact that algorithms with complexities $\poly(s) \cdot \Delta^{f(\Delta)}$ or $s^{f(\Delta)}$ may still exist, it is interesting to consider existence of algorithms with a polynomial dependence on $\Delta$ in their complexities for some partial cases of the $\Delta$-modular ILP problem. It is exactly what we do in the paper while fixing the number of constraints in ILP formulations of the problems \ref{main_prob} and \ref{ILP_standard}.
Due to the Hadamard's inequality, the existence of an ILP algorithm, whose complexity depends on $\Delta$ linearly, can give sufficiently better complexity bounds in terms of $\Delta_1$, than the bounds of Remark \ref{Delta1_complexity_rm}.
\begin{remark}[Some notes about lower bounds for fixed $m$.]
Unfortunately, there are not many results about lower complexity bounds for the problem \ref{ILP_standard} with fixed $m$. But, we can try to adopt some bounds based on the $\Delta_1$ parameter to our case. For example, the existence of an algorithm with the complexity bound
$$
2^{o(m)} \cdot 2^{o(\log_2 \Delta)} \cdot \poly(s)
$$ will contradict to the ETH. It is a straightforward adaptation of \cite[Theorem 3]{FOMIN}.
The Theorem 13 of \cite{CONVILP} states that for any $\delta> 0$ there is no algorithm with the arithmetical complexity bound
$$
f(m) \cdot (n^{2-\delta} + \Delta_1^{2m - \delta}),
$$ unless there exists a truly sub-quadratic algorithm for the $(min, +)$-convolution. Using Hadamard's inequality, it adopts to
$$
f(m) \cdot (n^{2-\delta} + \Delta^{2 - \delta/m}).
$$
The best known bound in terms of $m$ is given in \cite[Corollary 2]{TightComplexityLB}. More precisely, the existence of an algorithm with the complexity bound
$$
2^{o(m \log m)} \cdot \Delta_1^{f(m)} \cdot \poly(s)
$$ will contradict to the ETH. But, we does not know how to adopt it for $\Delta$-modular case at the moment.
Unfortunately, all mentioned results are originally constructed for the version of \ref{ILP_standard} with unbounded variables and it is the main reason, why their bounds are probably weak with respect to the dependence on the $\Delta$ parameter. And it would be very interesting to construct a lower bound of the form
$$
f(m) \cdot \Delta^{\Omega(m)} \cdot \poly(s)
$$ for the \ref{ILP_standard} problem. Additionally, at the moment we does not know any FPTAS lower bounds for the \ref{main_prob} problem. These questions are good directions for future research.
\end{remark}
\section{Proof of the Theorem \ref{main_th}}\label{proof_sec}
\subsection{Greedy Algorithm}
The $1/(m+1)$-approximate algorithm for the \ref{main_prob} is presented in \cite{BestPTAS} (see also \cite[p.~252]{TheMKP}) for the case $u = \BUnit$. This algorithm can be easily modified to work with a generic upper bounds vector $u$.
\begin{algorithm}[H]
\caption{The greedy algorithm}
\begin{algorithmic}[1]
\REQUIRE an instance of the \ref{main_prob} problem;
\ENSURE return $1/(m+1)$-approximate solution of the \ref{main_prob};
\STATE compute an optimal solution $x^{LP}$ of the LP relaxation of the \ref{main_prob};
\STATE $y := \lfloor x^{LP} \rfloor$ --- a rounded integer solution;
\STATE $F := \{ i \colon x^{LP}_i \notin \ZZ \}$ --- variables with fractional values;
\RETURN $\Coast_{gr} := \max\{c^\top y, \max_{i\in F}\{c_i\} \}$;
\end{algorithmic}
\end{algorithm}
Since the vector $x^{LP}$ can have at most $m$ fractional coordinates and
\begin{equation}\label{greedy_ineq}
c^\top y + \sum_{i \in F} c_i = c^\top \lceil x^{LP} \rceil \geq c^\top x^{LP} \geq \Coast_{opt},
\end{equation}
we have $\Coast_{gr} \geq \frac{1}{m+1} \Coast_{opt}$.
\subsection{Dynamic Programming by Costs}
The dynamic programming by costs is one of the main tools in many FPTASes for the $1$-BKP. Unfortunately, it probably can not be generalized to work with $m$-BKPs for greater values of $m$. However, such generalizations can exist for some partial cases such as the $\Delta$-modular \ref{main_prob}.
Suppose that we want to solve the \ref{main_prob}, and it is additionally known that $\|x\|_1 \leq \gamma$, for any feasible solution $x$ and some $\gamma > 0$. Then, to develop a dynamic program it is natural to consider only integer points $x$ that satisfy to $\|x\|_1 \leq \gamma$. The following simple lemma and corollary help to define such a program.
\begin{lemma}\label{DP_width_lm}
Let $A \in \ZZ^{m \times n}$ and $B \in \ZZ^{m \times m}$ be the non-degenerate sub-matrix of $A$. Let additionally $\gamma \in \RR_{>0}$, $\Delta = \Delta(A)$, $\delta = |\det B|$ and
$$
M = \{y = A x \colon x \in \RR^n,\, \|x\|_1 \leq \gamma \},
$$
$$
\text{then}\quad |M \cap \ZZ^m| \leq 2^m \cdot \lceil 1 + \gamma \cdot \frac{\Delta}{\delta} \rceil^m \cdot \Delta.
$$
Points of $M \cap \ZZ^m$ can be enumerated by an algorithm with the arithmetical complexity bound:
$$
O(m^2 \cdot 2^m \cdot D ),
$$ where $D = \Delta \cdot \left( \gamma \cdot \frac{\Delta}{\delta} \right)^m$.
\end{lemma}
\begin{proof}
W.l.o.g. we can assume that first $m$ columns of $A$ form the sub-matrix $B$. Consider a decomposition $A = B \bigl(I \; U\bigr)$, where $\bigl(I \; U\bigr)$ is a block-matrix, $I$ is the $m \times m$ identity matrix and the matrix $U$ is determined uniquely from this equality. Clearly, $\Delta(\bigl(I \; U\bigr)) = \frac{\Delta}{\delta}$, so $\Delta_k(U) \leq \frac{\Delta}{\delta}$ for all $k \in \intint m$. Consider the set
$$
N = \{ y = \lceil 1 + \gamma \cdot \frac{\Delta}{\delta} \rceil B x \colon x \in (-1,1)^m\}.
$$
Let us show that $M \subseteq N$. Definitely, if $y = A x$ for $\|x\|_1 \leq \gamma$, then $y = B \bigl(I \; U\bigr) x = B t$, for some $t \in [-\gamma,\gamma]^m \cdot \frac{\Delta}{\delta}$. Finally, $\frac{1}{\lceil 1 + \gamma\cdot \frac{\Delta}{\delta} \rceil} t \in (-1,1)^m$ and $y \in N$.
To estimate the value $|N \cap \ZZ^m|$ we just note that $N$ can be covered by $2^m$ parallelepipeds of the form $\{ y = Q x \colon x \in [0,1)^m\}$, where $Q \in \ZZ^{m \times m}$ and $|\det Q| = \lceil 1 + \gamma\cdot \frac{\Delta}{\delta} \rceil^m \cdot \Delta$. It is well known that the number of integer points in such parallelepipeds is equal to $|\det Q|$, see for example \cite{SEB99} or \cite[Section~16.4]{SCHR98}. Hence, $|M \cap \ZZ^m| \leq |N \cap \ZZ^m| \leq 2^m \cdot \lceil 1 + \gamma\cdot \frac{\Delta}{\delta} \rceil^m \cdot \Delta$. Points inside of the parallelipiped can be enumerated by an algorithm with arithmetical complexity
$$
O(m \cdot \min\{\log (|\det Q|), m\} \cdot |\det Q|),
$$ see for example \cite{FPT18}. Applying the last formula, we obtain the desired complexity bound to enumerate all integer points inside $N$.
\end{proof}
\begin{corollary}\label{DP_width_cor}
Let $A \in \ZZ^{m \times n}$, $\gamma \in \RR_{>0}$, $\Delta = \Delta(A)$ and
$$
M = \{y = A x \colon x \in \RR^n,\, \|x\|_1 \leq \gamma \},
$$
$$
\text{then}\quad |M \cap \ZZ^m| \leq 2^m \cdot \lceil 1 + \gamma \rceil^m \cdot \Delta.
$$
Points of $M \cap \ZZ^m$ can be enumerated by an algorithm with the arithmetical complexity bound:
$$
O(\log m)^{m^2} \cdot \Delta \cdot \gamma^m.
$$
\end{corollary}
\begin{proof}
W.l.o.g. we can assume that $\rank(A) = m$. Let us choose $B \in \ZZ^{m \times m}$, such that $|\det B| = \Delta$, then the desired $|M \cap \ZZ^m|$-bound follows from the previous Lemma \ref{DP_width_lm}. Due to \cite{SUBDET_APPROX}, we can compute a matrix $\hat B \in \ZZ^{m \times m}$ such that $\Delta = {O(\log m)}^m \cdot \delta$, where $\delta = |\det \hat B|$, by a polynomial time algorithm. Finally, we take a complexity bound of the previous Lemma \ref{DP_width_lm} with $D = \Delta \cdot \gamma^m \cdot {O(\log m)}^{m^2}$.
\end{proof}
We note that in the current section we need only first parts of these Lemma \ref{DP_width_lm} and Corollary \ref{DP_width_cor} that only estimate number of points nor enumerate them.
Assume that the goal function of the \ref{main_prob} is bounded by a constant $C$. Then, for any $c_0 \in \intint C$ and $k \in \intint n$ we denote by $DP(k,c_0)$ the set of all possible points $y \in \ZZ^m_+$ that satisfy to the system
\begin{equation*}\label{DP_relations}
\begin{cases}
c^\top_{\intint k} x = c_0\\
y = A_{\intint k} x \\
A_{\intint k} x \leq b \\
0 \leq x \leq u_{\intint k}\\
x \in \ZZ^k.
\end{cases}
\end{equation*}
In particular, the optimal value of the \ref{main_prob} can be computed by the formula
$$
c^\top x^{opt} = \max\{c_0 \in [1,C] \cap \ZZ \colon DP(n,c_0) \not= \emptyset\}.
$$
The set $DP(k,c_0)$ can be recursively computed using the following algorithm:
\begin{algorithm}[H]
\caption{An algorithm to compute $DP(k,c_0)$}
\begin{algorithmic}[1]
\FORALL{$z \in [0,\gamma] \cap [0,u_k] \cap \ZZ$}
\FORALL{$y \in DP(k-1,c_0-z c_k)$}
\IF{$y + A_k z \leq b$}
\STATE {\bf add} $y + A_k z$ {\bf into} $DP(k,c_0)$
\ENDIF
\ENDFOR
\ENDFOR
\end{algorithmic}
\end{algorithm}
By Corollary \ref{DP_width_cor}, we have $|DP(k,c_0)| \leq 2^m \cdot \lceil 1 + \gamma \rceil^m \cdot \Delta$. Consequently, to compute $DP(k,c_0)$ we need at most $O(m \cdot (2\gamma)^{m+1} \cdot \Delta)$ arithmetic operations. The total complexity bound is given by the following trivial lemma.
\begin{lemma}\label{DP_complexity_lm}
The sets $DP(k,c_0)$ for $c_0 \in \intint C$ and $k \in \intint n$ can be computed by an algorithm with the arithmetical complexity
$$
O(n \cdot C \cdot m \cdot (2\gamma)^{m+1} \cdot \Delta).
$$
\end{lemma}
\subsection{Putting Things Together}
Our algorithm is based on the scheme proposed in the seminal work \cite{IK} of O.~Ibarra and C.~Kim. Our choice of an algorithmic base is justified by the fact that it is relatively easy to generalize the approach of \cite{IK} to the $m$-dimensional case. On the other hand, more sophisticated schemes described in the papers \cite{KP1,KP2,LAW,MO} give constant improvements in the exponent or improvements in the memory usage only.
First of all, let us define two parameters $\alpha, \beta \in \QQ_{>0}$, whose purpose will be explained later. Let $\Coast^{gr}$ be the value of the greedy algorithm applied to the original $\Delta$-modular \ref{main_prob}, $x^{opt}$ be its integer optimal point and $\Coast^{opt} = c^\top x^{opt}$. As it was proposed in \cite{IK}, we split items into heavy and light: $H = \{i \colon c_i > \alpha \, \Coast^{gr} \}$ and $L = \{i \colon c_i \leq \alpha \, \Coast^{gr} \}$.
It can be shown that $\|x_H\|_1 \leq \frac{m+1}{\alpha}$ for any feasible solution $x$ of \ref{main_prob}. Definitely, if $\|x_H\|_1 > \frac{m+1}{\alpha}$, then $\Coast^{opt} = c^\top x^{opt} \geq c^\top x \geq c^\top_H x_H > \alpha \Coast^{gr} \frac{m+1}{\alpha} = (m+1) \Coast^{gr} \geq \Coast^{opt}$.
Let $s = \beta \Coast^{gr}$, we put $w = \lfloor \frac{c}{s} \rfloor$. Consider a new $\Delta$-modular $m$-BKP that consists only from heavy items of the original problem with the scaled costs $w$.
\begin{gather}
w_H^\top x \to \max\notag\\
\begin{cases}
A_H x \leq b\\
0 \leq x \leq u_H\\
x \in \ZZ^{|H|}.
\end{cases}\tag{HProb}\label{H_prob}
\end{gather}
It follows that $\|x\|_1 \leq \frac{m+1}{\alpha}$ for any feasible solution of \eqref{H_prob}. Additionally, we have $w_H^\top x \leq \frac{1}{s} c_H^\top x \leq \frac{m+1}{s} \Coast^{gr} \leq \frac{m+1}{\beta}$, for any $x$ being feasible solution of \eqref{H_prob}. Hence, we can apply Lemma \ref{DP_complexity_lm} to construct the sets $DP_H(k,c_0)$ for $k \in \intint n$ and $c_0 \in \intint \lceil \frac{m+1}{\beta} \rceil$. Due to Lemma \ref{DP_complexity_lm}, the arithmetical complexity of this computation is bounded by
\begin{equation}\label{applied_DP_complexity}
O(n \cdot \frac{m (m+1)}{\beta} \cdot (2\frac{m+1}{\alpha})^{m+1} \cdot \Delta ).
\end{equation}
To proceed further, we need to define a new notation $Pr(I, t)$. For a set of indexes $I \subseteq \intint n$ and for a vector $t \in \ZZ_+^m$, we denote by $Pr(I, t)$ the optimal value of the sub-problem, induced by variables with indexes in $I$ and by the right hand side vector $t$. Or by other words, $Pr(I, t)$ is the optimal value of the problem
\begin{gather*}
c_I^\top x \to \max\\
\begin{cases}
A_I x \leq t\\
0 \leq x \leq u_I\\
x \in \ZZ^{|I|}.
\end{cases}
\end{gather*}
After $DP_H(n,c_0)$ being computed we can construct resulting approximate solution, which will be denoted as $\Coast^{apr}$, by the following algorithm.
\begin{algorithm}[H]
\caption{An FPTAS for \ref{main_prob}}
\begin{algorithmic}[1]
\FORALL{$c_0 \in \intint \lceil \frac{m+1}{\beta} \rceil$}
\FORALL{$y \in DP_H(n,c_0)$}
\STATE compute an approximate solution $q$ of the problem $Pr(L,b-y)$
\begin{gather*}
c_L^\top x \to \max \notag\\
\begin{cases}
A_L x \leq b - y \\
0 \leq x \leq u_L \\
x \in \ZZ^{|L|}
\end{cases} \label{L_prob}
\end{gather*}
using the greedy algorithm.
\STATE $\Coast^{apr} := \max\{\Coast^{apr}, s\, c_0 + q\}$.
\ENDFOR
\ENDFOR
\end{algorithmic}
\end{algorithm}
Due to Corollary \ref{DP_width_cor}, the arithmetical complexity of the algorithm can be estimated as
\begin{equation}\label{join_complexity}
O( T_{LP}\cdot \frac{m}{\beta} \cdot (2 \frac{m+1}{\alpha})^m \cdot \Delta ).
\end{equation}
Clearly, $x^{opt} = x_H^{opt} + x_L^{opt}$. We denote $C^{opt}_H = c^\top_H x^{opt}_H$, $C^{opt}_L = c^\top_L x^{opt}_L$ and $c_0^* = w^\top_H x^{opt}_H$. The value of $c_0^*$ will arise in some evaluation of Line 1 of the proposed algorithm. Or by other words, we will have $c_0 = c_0^*$ in some evaluation of Line 1. Let $y*$ be the value of $y \in DP_H(n,c_0^*)$ such that $s\,c^*_0 + Pr(L,b-y^*)$ is maximized and $q^*$ be the approximate value of $Pr(L,b-y^*)$, given by the greedy algorithm in Line 3. Clearly, $\Coast^{apr} \geq s\,c_0^* + q^*$, so our goal is to chose parameters $\alpha,\beta$ in such a way that the inequality $s\,c_0^* + q^* \geq (1-\varepsilon) \Coast^{opt}$ will be satisfied.
Firstly, we estimate the difference $\Coast^{opt}_H - s\, c_0^*$:
\begin{multline*}
\Coast^{opt}_H - s\, c_0^* \leq (c^\top_H - s\, w^\top_H) x^{opt}_H \leq s\,\{c^\top_H/s\} x^{opt}_H \leq \\
\leq s \frac{m+1}{\alpha} = \frac{(m+1) \beta \Coast_{gr}}{\alpha } \leq \frac{(m+1) \beta \Coast_{opt}}{\alpha}
\end{multline*}
To estimate the difference $\Coast^{opt}_L - q^*$ we need to note that $Pr(L,b-y^*) \geq Pr(L, b - A_H x^{opt}_H) = \Coast^{opt}_L$. It follows from optimality of $y^*$ with respect to the developed dynamic program. Next, since $c_i \leq \alpha \Coast^{gr}$ for $i \in L$, due to the inequality \eqref{greedy_ineq}, we have
\begin{equation*}
q^* \geq Pr(L,b-y^*) - m\, \alpha \Coast^{gr} \geq Pr(L,b-y^*) - (m+1)\, \alpha \Coast^{opt}.
\end{equation*}
Finally, we have
\begin{equation*}
\Coast^{opt}_L - q^* \leq Pr(L,b-y^*) - q^* \leq (m+1)\, \alpha \Coast^{opt}.
\end{equation*}
Putting all inequalities together, we have
\begin{multline*}
\Coast^{opt} - \Coast^{apr} \leq (\Coast^{opt}_H - s\,c_0^*) + (\Coast^{opt}_L - q^*) \leq \\
\leq (m+1)(\alpha + \frac{\beta}{\alpha}) \Coast^{opt}
\end{multline*} and
\begin{equation}\label{final_error}
\Coast^{apr} \geq (1-(m+1)(\alpha + \frac{\beta}{\alpha}))\Coast^{opt}.
\end{equation}
The total arithmetical complexity can be estimated as
\begin{equation}\label{final_complexity}
O(T_{LP} \cdot \frac{1}{\beta} \cdot (2 m)^{m+3} \cdot \left(\frac{1}{\alpha}\right)^{m+1} \cdot \Delta).
\end{equation}
Finally, after the substitution $\beta = \alpha^2$ and $\alpha = \frac{\varepsilon}{2(m+1)}$ to \eqref{final_error} and \eqref{final_complexity}, we have
$$
\Coast^{apr} \geq (1 - \varepsilon) \Coast^{opt}
$$ and a complexity bound
$$
O(T_{LP} \cdot (1/\varepsilon)^{m+3} \cdot (2m)^{2m + 6} \cdot \Delta)
$$ that finishes the proof.
\section{Proof of Theorem \ref{Delta_ILP_th}}\label{proof_Delta_ILP_th}
Let $x^*$ be an optimal vertex solution of the LP relaxation of the $\Delta$-modular \ref{ILP_standard} problem. After a standard change of coordinates $x \to x - \lfloor x^* \rfloor$ the original \ref{ILP_standard} transforms to an equivalent ILP with different lower and upper bounds on variables and a different right-hand side vector $b$. For the sake of simplicity we assume that lower bounds of the new problem are equal to zero.
Any optimal vertex solution of the LP problem has at most $m$ non-zero coordinates, so we have the following bound on the $l_1$-norm of an optimal ILP solution $z^* - \lfloor x^* \rfloor$ of the new problem:
\begin{equation*}
\|z^* - \lfloor x^* \rfloor\|_1 \leq \|x^* - z^*\|_1 + \|x^* - \lfloor x^* \rfloor\|_1 \leq H + m.
\end{equation*}
\subsection{First Complexity Bound}
Consider a weighted digraph $G = (V,E)$, whose vertices are triplets $(k,h,l)$, for $k \in \intint n$, $l \in \intint[0]{(H + m)}$ and $h \in \{A x \colon \|x\|_1 \leq l \} \cap \ZZ^m$. Using Corollary \ref{DP_width_cor}, we bound the number of vertices $|V|$ by $O(n \cdot 2^m \cdot (H + m)^{m+1} \cdot \Delta )$. By definition, any vertex $(k, h, l)$ has an in-degree equal to $\min\{u_k, l\}+1$. More precisely, for any $j \in \intint[0]\min\{u_k, l\}$ there is an arc from $(k-1, h - A_k j, l - j)$ to $(k, h, l)$, this arc is weighted by $c_k j$. Note that vertex $(k-1, h - A_k j, l - j)$ exists only if $j \leq l$. Additionally, we add to $G$ a starting vertex $s$, which is connected with all vertices of the first level $(1,*,*)$, weights of this arcs correspond to solutions of $1$-dimensional sub-problems. Clearly, the number of arcs can be estimated by
$$
|E| = O(|V| \cdot (H+m)) = O(n \cdot 2^m \cdot (H + m)^{m+2} \cdot \Delta ).
$$
The \ref{ILP_standard} problem is equivalent to searching of the longest path starting from the vertex $s$ and ending at the vertex $(n, b, H + m)$ in $G$. Since the graph $G$ is acyclic, the longest path problem can be solved by an algorithm with the complexity bound $O(|V| + |E|) = O(n \cdot 2^m \cdot (H + m)^{m+2} \cdot \Delta )$.
We note that during the longest path problem solving, the graph $G$ must be evaluated on the fly. In other words, the vertices and arcs of $G$ are not known in advance, and we build them online. To make constant-time access to vertices we can use a hash-table data structure with constant-time insert and search operations (see Remark \ref{hash_table_rm}).
Finally, using the binarization trick, described in the work \cite{STEINITZILP}, we can significantly decrease the number of arcs in $G$. The idea of the trick is that any integer $j \in [0,\min\{u_k, l\}]$ can be uniquely represented using at most $O(\log^2 (\min\{u_k, l\})) = O(\log^2 (H+m))$ bits. More precisely, for any interval $[0,\min\{u_k, l\}]$ there exist at most $O(\log^2 (H+m))$ integers $s(k,i)$ such that any integer $j \in [0,\min\{u_k, l\}]$ can be uniquely represented as
\begin{gather*}
j = \sum_i s(k,i) x_i,\quad\text{ where $x_i \in \{0,1\}$, and }\\
\sum_i s(k,i) x_i \in [0, \min\{u_k,l\}], \quad \text{for any $x_i \in \{0,1\}$}.
\end{gather*}
Using this idea, we replace the part of the graph $G$ connecting vertices of the levels $(k-1,*,*)$ and $(k,*,*)$ by an auxiliary graph, whose vertices correspond to the triplets $(i,h,l)$, where $i \in \{0,1, \dots, O(\log^2 (H+m))\}$, and any triplet $(i,h,l)$ has in-degree two. More precisely, the vertex $(i, h, l)$ is connected with exactly two vertices: $(i-1, h, l)$ and $(i-1, h - s(k,i) A_k, l - s(k,i))$. The resulting graph will have at most $O(\log^2 (H + m) |V|)$ vertices and arcs, where $|V|$ corresponds to the original graph. Total arithmetical complexity can be estimated as
$$
O(n \cdot 2^{O(m)} \cdot (H + m)^{m+1} \cdot \log^2 H \cdot \Delta ).
$$
\subsection{Second Complexity Bound}
Consider a weighted digraph $G = (V, E)$, whose vertices are pairs $(k,h)$, for $k \in \intint n$ and $h \in M := \{A x \colon \|x\|_1 \leq H + m \} \cap \ZZ^m$. The edges of $G$ have the same structure as in the graph from the previous subsection. More precisely, for any $j \in \intint[0]{u_k}$ we put an arc from $(k-1, h - A_k j)$ to $(k, h)$, if such vertices exist in $V$, the arc is weighted by $c_k j$.
We compute all vertices of $G$ directly, using Corollary \ref{DP_width_cor}. Arithmetical complexity of this step is bounded by $n \cdot {O(\log m)}^{m^2} \cdot (H + m)^m \cdot \Delta$. Due to Corollary \ref{DP_width_cor}, $|V| = n \cdot |M| = O(n \cdot 2^m \cdot (H + m)^{m} \cdot \Delta)$ and $|E| = O(n \cdot 2^m \cdot (H + m)^{m+1} \cdot \Delta)$, since an in-degree of any vertex in $G$ is bounded by $H + m +1$.
Let us fix some vertex-level $(k, *)$ of $G$ for some $k \in \intint n$, and consider an auxiliary graph $F_k$, whose vertices are exactly elements $h \in M = \{A x \colon \|x\|_1 \leq H + m\} \cap \ZZ^m$. For two vertices $h_1, h_2$ of $F_k$, we put an arc from $h_1$ to $h_2$ if $h_2 - h_1 = A_k$. Since the graph $F_k$ is acyclic and since "in" and "out" degrees of any vertex in $F_k$ are at most one, the graph $F_k$ is a disjoint union of paths. This decomposition can be computed by an algorithm with complexity $O(|V(F_k)|) = O(|M|) = O(2^m \cdot (H + m)^m \cdot \Delta)$. Let $(h_1, h_2, \dots, h_t)$ be some path of the decomposition, and $longest(k,h)$ be the value of the longest path in $G$ starting at $s$ and ending at $(k,h)$. Clearly, for any $i \in \intint t$, the value of $longest(k,h_i)$ can be computed by the formula
\begin{equation}\label{longest}
longest(k,h_i) = \max\limits_{j \in \min\{u_k, i-1\}} longest(k-1,h_{i-j}) + c_k j.
\end{equation}
Consider a queue $Q$ with operations: $Enque(Q,x)$ that puts an element $x$ into the tail of $Q$, $Decue(Q)$ that removes an element $x$ from the head of $Q$, $GetMax(Q)$ that returns maximum of elements of $Q$. It is known fact that queue can be implemented such that all given operations will have amortized complexity $O(1)$. Now, we compute $longest(k,h_i)$, for $h_i \in (h_1, h_2, \dots, h_t)$ using the following algorithm:
\begin{algorithm}[H]
\caption{Compute longest path with respect to $(h_1, h_2, \dots, h_t)$}
\begin{algorithmic}[1]
\STATE Create an empty queue $Q$;
\STATE $w := \min\{u_k, t\}$;
\FOR{$j := 0$ {\bf to} $w$}
\STATE $Enque(Q, longest(k-1,h_{t-j}) + c_k j)$;
\ENDFOR
\FOR{$i := t$ {\bf down to} $1$}
\STATE $longest(k,h_i) := GetMax(Q) - c_k (t-i)$;
\STATE $Decue(Q)$;
\IF{$i \geq w+1$}
\STATE $Enque(Q, longest(k-1,h_{i-w-1}) + c_k (t - i+1))$;
\ENDIF
\ENDFOR
\end{algorithmic}
\end{algorithm}
Correctness of the algorithm follows from the formula \eqref{longest}. The algorithm's complexity is $O(t)$.
Let us estimate the total arithmetical complexity of the whole procedure. It consists from the following parts:
\begin{enumerate}
\item Enumerating of points in the set $M$. Due to Corollary \ref{DP_width_cor}, the complexity of this part is $O(\log m)^{m^2} \cdot (H+m)^m \cdot \Delta$;
\item Constructing the graphs $F_k$ for each $k \in \intint n$. The number of edges and vertices in $F_k$ can be estimated as $O(|M|)$. Hence, due to Corollary \ref{DP_width_cor}, the complexity of this part can be estimated as $O(n \cdot |M|) = O(n \cdot 2^m \cdot (H+m)^m \cdot \Delta)$.
\item For each $F_k$, compute a path decomposition of $F_k$. For each path in the decomposition, apply an Algorithm 4. The complexity of this part is clearly the same as in the previous step.
\end{enumerate}
Therefore, the total complexity bound is roughly
$$
n \cdot O(\log m)^{m^2} \cdot (H + m)^m \cdot \Delta.
$$
\section*{Conclusion}
The paper considers the $m$-dimensional bounded knapsack problem \eqref{main_prob} and the bounded ILP in the standard form \eqref{ILP_standard}. For the problem \ref{main_prob} it gives an FPTAS with the arithmetical complexity bound
$$
O(n \cdot (1/\varepsilon)^{m+3} \cdot \Delta),
$$ where $n$ is the number of variables, $m$ is the number of constraints (we assume here that $m$ is fixed) and $\Delta = \Delta(A)$ is the maximal absolute value of rank-order minors of $A$. For details see Theorem \ref{main_th} and Corollary \ref{main_cor}.
For the problem \ref{ILP_standard} it gives an exact algorithm with the complexity bound
$$
O(n \cdot \Delta^{m+1}).
$$
Taking $m = 1$ it gives
$$
O(n \cdot \Delta^2)
$$ arithmetical complexity bound for the classical bounded knapsack problem. For details see Theorem \ref{Delta_ILP_th} and Corollary \ref{Delta_ILP_cor}.
|
\section{Introduction}
Phase transitions associated with spontaneous symmetry breaking is a central topic in modern science, appearing in the most diverse situations, both in and out-of equilibrium~\cite{Sachdev,Goldenfeld}. One of the major challenges in describing the critical behaviour of a system on the verge of a transition has always been the determination of the critical exponents. Several different powerful analytical and numerical methods were elaborated for this purpose~\cite{Domb} which, however, are mostly concerned with equilibrium conditions.
Aim of this paper is to discuss a method that seems particularly suited to determine critical exponents associated to non-equilibrium phase transitions in open quantum many-body system. This question has attracted increasing attention thanks to remarkable breakthroughs in manipulating many-body systems
coupled to an external environment. Steady-state phase transitions have been observed in various experimental platforms as, for example, circuit
QED arrays \cite{fitzpatrick2017,collodo2019}, cold atomic systems in cavities~\cite{ritsch2013} or subject to losses~\cite{tomita2017}, and Rydberg
atom ensembles~\cite{ding2020}. The level of control in these different platforms has reached a stage such that they can be considered in all respects open-system quantum simulators.
An overview of the experimental and theoretical activities can be found in the recent reviews~\cite{houck2012,Sieberer,hartmann2016,noh2017,carusotto2020}.
In all the situations mentioned above, the dynamics is well described by a Lindblad master equation~\cite{breuer} for the reduced density matrix $\rho (t)$ of the system
($\hbar=1$ hereinafter)
\begin{equation}
\dot{\rho}(t)=-i[\hat{H},\rho(t)]+\sum_{\alpha}{{\cal D}_{\alpha}[\rho(t)]},
\label{ME}
\end{equation}
where $\hat{H}$ is the Hamiltonian of the systems under consideration (cavity array, optical lattice, $\dots$), and the superoperator ${\cal D}_{\alpha}[\rho(t)]=
\hat{L}_{\alpha}\rho(t)\hat{L}_{\alpha}^\dagger-\frac{1}{2}\{\hat{L}_{\alpha}^\dagger \hat{L}_{\alpha},\rho(t)\}$, with $\{\cdot,\cdot\}$ being the
anti-commutator, describes the incoherent dissipation processes induced by the coupling to a memory-less environment. The form of the Lindblad operators
$\hat{L}_{\alpha}$ depends on the process involved in the loss of coherence. The non-equilibrium phase transitions we are referring to take place in the steady state ($t \to \infty$), when the system may undergo spontaneous symmetry breaking. The steady-state phase diagram in many-body open systems governed by Eq.~(\ref{ME}) was analyzed extensively in the recent literature, showing a complexity and
a variety of phases that do not have a counterpart in equilibrium (see~\cite{diehl2010,lee2011,poletti2013,sieberer2013,lee2013,jin2013,marcuzzi2014,
finazzi2015,weimer2015,Marino2016,jin2016,schiro2016,maghrebi2016,wilson2016,zamora2017,kshetrimayum2017,biella2017,rota2019,carollo2019,young2020,
scarlatella2020} and references therein).
In driven-dissipative systems, besides the phase diagram, also the universality class of the transition may change. This question has been studied in several different
models, see e.g.~\cite{marcuzzi2014,maghrebi2016,sieberer2013,carollo2019,young2020,tauber2014,Marino2016}. Therefore it is of great significance to
investigate the exact critical exponents associated with a steady-state phase transition. In particular, it would be very important to have reliable numerical methods to
determine them. Various powerful methods have been developed to simulate quantum many-body systems~\cite{finazzi2015,kshetrimayum2017,zwolak2004,
daley2014,werner2016,vicentini2019,Nagy2019,Yoshioka2019,hartmann2019,kilda2020,keever2020,weimer2019,Paeckel2019}. It is however important to stress that, from the numerics perspective, directly
accessing the critical properties of the steady-state may be rather demanding. Furthermore, except very few exceptions~\cite{carollo2019}, these phenomena
occur in two (or higher) dimensions, where the numerics becomes considerably more complex. It is therefore desirable to have a method to extract critical exponents that may require moderate resources to reach good accuracy.
With this goal in mind, in this work we apply the so-called Coherent Anomaly Method~\cite{suzuki1986,suzuki1987,katori1987} (CAM) to the steady-state phase
transitions of open quantum many-body systems. The method is based on a combination of Cluster-Mean Field (CMF) approximation with finite-size scaling
of its outcomes. By analyzing the singularities of the CMF response functions, we are able to extract the critical points
and exponents of continuous steady-state phase transitions. The CAM has already been successfully applied in a diversity of equilibrium systems including the Ising model \cite{suzuki1986,suzuki1987,katori1987}, transverse-field Ising model \cite{Nonomura1992} and antiferromagnetic $XXZ$-chain \cite{Nonomura1993}. Moreover, the application of CAM in classical non-equilibrium model has also been reported \cite{dickman2002,park2005}. In particular, for the one-dimensional driven pair contact process with diffusion, the estimates of critical exponents through the CAM analysis are in excellent accord with simulation results \cite{park2005}. Here we explore the power of CAM in accessing the critical exponent in dissipative phase transitions of quantum systems. We tackle this problem by testing the method in two paradigmatic models. In particular, in one case there are solid theoretical predictions that guide our analysis. Additional studies, for other cases, are needed in order to test how powerful is CAM for many-body open quantum systems.
The paper is organized as follows. In the next section, we first briefly review the CMF approach to the dissipative system. We then illustrate the idea of the CAM by considering the magnetic susceptibility and show how to access the critical exponent of dissipative phase transition through a series of CMF results. In Sec. \ref{sec_models}, we introduce two models, namely the dissipative transverse-field Ising model and the dissipative XYZ model on two-dimensional square lattice, as examples to investigate the performance of CAM. The details of numerical simulation are explained as well (Sec. \ref{sec_numerical}). In Sec. \ref{sec_results}, the critical exponent and critical point that extracted from CAM for both models are presented. We compare the CAM results with the classical critical exponent and discuss the stability of the CAM results through different choices of clusters. Finally, we summarize in Sec. \ref{sec_summary}.
\begin{figure*}
\centering
\includegraphics[width=0.9\linewidth]{clusters.pdf}
\caption{(Color online) (a) Series of isotropic clusters for a systematic CMF approximations. In the CMF simulation, the global density matrix of the whole
lattice is factorized as the tensor product of density matrices of identical finite-size clusters. The interactions inside the cluster (solid lines) are treated
exactly while each boundary site is influenced by effective fields $h_j$ (arrows) from the adjacent clusters, which are determined by a self-consistent condition.
(b) Main idea of the coherent anomaly method. The susceptibilities for a set of CMF approximations diverge at the critical points $\lambda_c(n)$ ($n=1,2,...$). While all such divergences are governed by the
classical exponent $\gamma^{\text{mf}}$, the amplitudes $\chi_0(n)$ [defined in Eq.~(\ref{chi_mf})] become anomalously large as $\lambda_c(n)$ approaches the true critical point $\lambda_c^*$. The divergence of $\chi_0(n)$ gives the true critical exponent $\gamma^*$, according to Eq. (\ref{fL}).}
\label{Fig1}
\end{figure*}
\section{Coherent Anomaly Method}
\label{sec_cam}
The CAM method is based on a scaling analysis of mean-field results. To set the stage, we start with a
brief review of the CMF approximation applied to driven-dissipative systems. The CMF approach is based on the factorization of the global density matrix
$\rho(t)=\bigotimes_C{\rho_{C}(t)}$. The clusters in the factorization are assumed to be identical. As a consequence, Eq. (\ref{ME}) is decoupled into the following master
equation with respect to the cluster $C$,
\begin{equation}
\dot{\rho}_{C}(t)=-i[\hat{H}_{\text{CMF}}(\lambda,t),\rho_{C}(t)]+\sum_{j\in C}{{\cal D}_j[\rho_{C}(t)]},
\label{ME_CMF}
\end{equation}
where $\lambda$ is a driving parameter and we supposed that dissipation acts locally on the sites $j$ of the lattice. Considering $\hat H = \sum_{\alpha} \sum_{\langle j,k \rangle} \hat{H}_{jk}^{(\alpha)}$ (which includes various nearest-neighbor interactions of type $\alpha$, with factorized form $\hat{H}^{(\alpha)}_{jk}=\hat{H}^{(\alpha)}_{j}\otimes \hat{H}^{(\alpha)}_{k}$), the CMF Hamiltonian is given by $\hat{H}_{\text{CMF}}(\lambda,t)=\hat{H}_{C}(\lambda)+\hat{H}_{\partial C}(\lambda,t)$,
where $\hat{H}_C(\lambda)=\sum_\alpha\sum_{\langle j,k\rangle\in C}{\hat{H}^{(\alpha)}_{jk}}$ describes the interactions inside the cluster and $\hat{H}_{\partial C}(\lambda,t)=
\sum_{\alpha}\sum_{j\in\partial C}{h^{(\alpha)}_j(t)\hat{H}^{(\alpha)}_j(\lambda)}$ describes the interaction of sites on the boundary with the adjacent clusters. The time-dependent
effective fields are given by $h^{(\alpha)}_j(t) = \sum^\prime_{k} \text{Tr}[\hat{H}^{(\alpha)}_{k} \rho_C(t)]$, where the prime indicates sites $k\in C$ corresponding to neighbors of $j$. Due to translational invariance, the long-time limit asymptotic solution of Eq. (\ref{ME_CMF}) is determined by the following self-consistent condition
\begin{equation}
h^{(\alpha)}_j(t\rightarrow\infty)=\text{Tr}[\hat{h}^{(\alpha)}_j \rho_C(t\rightarrow\infty)],
\label{scc}
\end{equation}
where $\hat{h}^{(\alpha)}_j= \sum^\prime_{k} \hat{H}^{(\alpha)}_{k} $ are the local effective field operators. \color{black}
A sketch of the cluster mean-field approach is illustrated in Fig.~\ref{Fig1}.
From Fig.~\ref{Fig1}(a), one can see that for the standard (single-site) mean-field approximation the effective field governs the coherent evolution, while $\hat{H}_C(\lambda)$
is absent; as the size of cluster becomes larger, the effective fields acting on the boundary sites play a less dominant role in $\hat{H}_{\text{CMF}}$, and the quantum correlations
embedded in $\hat{H}_C(\lambda)$ become more significant. Through systematically enlarging the cluster size, the gradual inclusion of correlations and the self-consistent
conditions results in stronger singular behaviors of the order parameter and response function. Such coherently enhanced singularities are intimately related to the intrinsic
fluctuations embedded in the system. Specifically, the increased fluctuations in the cluster leads to the divergence of the amplitude of the response function as the size of cluster
goes to infinity. The main idea behind the CAM~\cite{suzuki1986,suzuki1987,katori1987} is to analyze how the mean-field results change with the size of the cluster, which will
provide information of the (non mean-field) exponents as well as the exact boundaries.
The choice of the series of clusters in the CAM should guarantee that, as the cluster size $L$ is systematically enlarged, the set of critical points $\lambda_c(L)$ obtained from successive
CMF calculations should asymptotically approach the true critical point $\lambda^*$, i.e.,
\begin{equation}
\lim_{L\rightarrow\infty}{\lambda_c(L)}=\lambda^*.
\label{systematic}
\end{equation}
This is a key point of CAM because some choices of
the cluster series may not satisfy Eq.~(\ref{systematic}). For example, in a two-dimensional square lattice, the CMF analysis using a series of plaquette clusters of size
$L = l_x\times l_y$ may not approach the true critical point if $l_x\gg l_y$, because of the large anisotropy; thus we will choose the square plaquette of size $L$ with
$l_x=l_y=l$ to implement the CAM.
In order to illustrate the idea of CAM~\cite{suzuki1986,suzuki1987,katori1987} we consider, for concreteness, the magnetic susceptibility, also in the light of the models we are going to
study. In the presence of a small external magnetic field $\delta$
and taking the self-consistency condition
into account, the magnetization in the $x$ direction on the corner-site of a
cluster may be written as
\begin{equation}
\langle \hat{\sigma}^x(\lambda)\rangle_{\text{ss}} = \chi_L^{\text{cl}}(\lambda)\delta + F_L(\lambda)\langle\hat{\sigma}^x(\lambda)\rangle_{\text{ss}},
\label{mag_cmf}
\end{equation}
where the two terms on right-hand side denote the magnetic responses to the external and self-consistent fields, respectively. The subscript `ss' denotes the steady-state. $\chi_L^{\text{cl}}(\lambda)$ is the magnetic susceptibility of the cluster
without the self-consistent field and $F_L(\lambda)$ is Kubo's canonical correlation. Consequently, the magnetic susceptibility as a function of $\lambda$ is given by
$\chi(\lambda)=\chi^{\text{cl}}_{L}(\lambda)/\left[ 1-F_L(\lambda)\right]$. The critical point $\lambda_c(L)$ is determined by $1-F_L(\lambda)=0$.
In the CMF approximation, the behavior of the response function near the transition point is characterized by the classical critical exponent according to the Ginzburg-Landau
theory. Thus, in the vicinity of $\lambda_c(L)$ the magnetic susceptibility can be approximated by
\begin{equation}
\chi(\lambda)\approx \chi_0(L)\left[\frac{\lambda-\lambda_c(L)}{\lambda_c(L)}\right]^{-\gamma^{\text{mf}}},
\label{chi_mf}
\end{equation}
where $\gamma^{\text{mf}}=1$ is the classical mean-field critical exponent of the susceptibility.
The amplitude in Eq. (\ref{chi_mf}) takes the form $\chi_0(L)\approx\chi_L^{\text{cl}} (\lambda_c(L))/[-\lambda_c(L)F_L'(\lambda_c(L))]$. It reflects the fluctuations stemming
from the inclusion of short-range correlations in the relevant cluster and should be anomalously large as $\lambda_c(L)$ goes to $\lambda^*$ (or as $L\rightarrow\infty$).
The dependence of $\chi_0(L)$ on $\lambda_c(L)$ is called {\it coherent anomaly}, and carries the intrinsic information of the true critical behavior through its large-$L$ asymptotic dependence~\cite{suzuki1986},
\begin{equation}
\chi_0(L) \approx c_0\left[\lambda_c(L)-\lambda^*\right]^{-(\gamma^*-\gamma^{\text{mf}})},
\label{fL}
\end{equation}
where $\gamma^*$ is the true critical exponent and $c_0$ a prefactor. In Eq.~(\ref{fL}), there are three unknown quantities, $c_0,\lambda^*,\gamma^*$, which demands at
least the results of three different cluster sizes. As a consequence, the true critical exponent can be estimated as
\begin{equation}
\gamma^*= \gamma^{\text{mf}} + \frac{\log{\left[\chi_0(j)/\chi_0(k)\right]}}{\log{\left\{\left[\lambda_c(k)-\lambda^*\right]/\left[\lambda_c(j)-\lambda^*\right]\right\}}},
\label{gammastar}
\end{equation}
with $\{\chi_0(j),\lambda_c(j)\}$ being a set of amplitudes and critical points obtained by CMF approximations, with at least three successive systematically enlarging clusters.
The CAM procedure, in practice, goes along the following steps: (i) compute the response function $\chi(\lambda)$ at various levels of CMF approximation with systematically
enlarging clusters; (ii) extract the coefficients and critical points $\{\chi_0(L),\lambda_c(L)\}$ for the various CMF approximation with Eq. (\ref{chi_mf}); (iii) fit the dependence of $\chi_0(L)$ on
$\lambda_c(L)$ with Eq.~(\ref{fL}), to access the true critical point and exponent.
We finish this section by emphasizing that the coherent anomaly relation in Eq. (\ref{fL}) is derived in a rather general way. It doesn't make any reference to the underlying nature of the system or whether the system is equilibrium/non-equilibrium or classical/quantum. The validity of the application of CAM depends only on the convergence of the properties of scaling functions. Namely, the CAM can be applied to the system that the true critical point can be reached when the cluster of system asymptotically tends to be infinite. As shown in Refs. \cite{suzuki1986,park2005}, CAM obtained good results of the phase transitions in equilibrium and non-equilibrium systems. Based on above arguments, we apply CAM to study the critical exponents in dissipative phase transitions of quantum systems.
\section{The Models}
\label{sec_models}
To investigate explicitly the performance of CAM, we will consider two models of dissipative quantum many-body systems. The first one is the
spin-1/2 dissipative TFI model on the square lattice. The Hamiltonian governing the coherent evolution is the following,
\begin{eqnarray}
\hat{H}^{\text{TFI}} &=&-\frac{V}{4}\sum_{\left\langle j,k \right\rangle}{\hat{\sigma}^x_j\hat{\sigma}^x_k} +\frac{g}{2} \sum_k{\hat{\sigma}^z_k},
\end{eqnarray}
where $\hat{\sigma}_j^\alpha$ ($\alpha=x,y,z$) denote the spin-1/2 Pauli matrices on the $j$-th site and $\langle j,k\rangle$ denotes the nearest-neighbor spins. The first term in $\hat{H}^{\text{TFI}}$ represents Ising interactions of strength $V$ between nearest-neighbor sites, while the second term accounts for the local transverse magnetic field along the $z$ direction, with amplitude $g$.
As the second example, we will investigate the two-dimensional dissipative spin-1/2 XYZ model on the square lattice. The Hamiltonian in Eq. (1) for this model is given by
\begin{equation}
\hat{H}^{\text{XYZ}}=\sum_{\langle j,k\rangle}{\left(J_x\hat{\sigma}^x_j\hat{\sigma}^x_k+J_y\hat{\sigma}^y_j\hat{\sigma}^y_k+J_z\hat{\sigma}^z_j\hat{\sigma}^z_k\right)},
\label{Hxyz}
\end{equation}
where $J_\alpha$ ($\alpha=x,y,z$) is the coupling strength.
For both models, we consider an incoherent dissipative process acting locally on each site which tends to flip the spin down along the $z$-direction, namely
$L_j=\sqrt{\Gamma}(\hat{\sigma}_j^x-i\hat{\sigma}^y_j)/2 =\sqrt{\Gamma} \hat{\sigma}_{j}^{-}$ in Eq. (\ref{ME}), where $\Gamma$ is the decay rate.
The master equation of both models presents a $\mathbb{Z}_2$ symmetry, namely the master equation is invariant under a $\pi$-rotation about the $z$-direction, $(\hat{\sigma}_{j}^x,\hat{\sigma}_{j}^y)\rightarrow(-\hat{\sigma}_{j}^x,-\hat{\sigma}_{j}^y)$. We define the order parameter as the averaged steady-state magnetization per site
${\cal O}=\langle\hat{\sigma}^{x,y}\rangle_{\text{ss}}=\frac{1}{L}\sum_{j=1}^{L}{\text{tr}\left(\rho_{\text{ss}}\hat{\sigma}^{x,y}_j\right)}$, wher $j$ labels the site in the lattice of size $L$ and $\rho_{\text{ss}}$ is the steady-state density matrix. The nonzero order parameter indicates appearance of the ordered phase.
By means of Keldysh formalism \cite{overbeck2017} and cluster mean-field
approximation \cite{jin2018}, a continuous phase transition from the disordered paramagnetic (PM) phase ($\langle\hat{\sigma}^{x}\rangle,\langle\hat{\sigma}^{y}\rangle = 0$),
to the ferromagnetic (FM) phase ($\langle\hat{\sigma}^{x}\rangle,\langle\hat{\sigma}^{y}\rangle\neq0$) has been predicted for the dissipative TFI model. In Fig. \ref{magn}(a), we show the order parameter as a function of the strength of the transverse field by means of $3\times3$ CMF simulation.
Meanwhile, in the dissipative XYZ model, a steady-state phase transition breaking the $\mathbb{Z}_2$ symmetry has been predicted \cite{finazzi2015,biella2018,kshetrimayum2017,huybrechts2019,rota2017}. For $J_x=J_y$, the magnetization along $z$-direction is a conserved quantity and there is nothing counteracts to the dissipation, thus leads all the spins down to the $z$-direction. The steady state of the system is in the disordered PM phase. However, the anisotropic coupling $J_x\ne J_y$ in the $x$-$y$ plane of the Hamiltonian may induce an effective field to each spin. The competition between the procession of the spin around the effective field and the decay leads to a phase transition from the disordered PM to ordered FM phase. Again, the steady-state phase transition can be revealed by the order parameter as shown in Fig. \ref{magn}(b). Moreover, the steady-state phase diagram has been predicted by CMF approaches \cite{jin2016}.
\begin{figure}[htp]
\includegraphics[width=1\linewidth]{magn1.pdf}
\caption{The cluster mean-field order parameter $\langle\hat{\sigma}^{x}\rangle_{\text{ss}}$ for (a) the dissipative TFI model as a function of $g/2V$ for the cases of $\Gamma/V=0.05$, $0.1$, $0.2$ and $0.25$, (b) XYZ model, whose Hamiltonian is given by Eq. (\ref{Hxyz}), as function of $J_y/\Gamma$ for $J_x / \Gamma= 0.9$ and $J_z / \Gamma = 1$. }
\label{magn}
\end{figure}
The steady-state phase transition in both two dissipative models can also be captured by the singular behavior of the magnetic susceptibility. The susceptibility is defined as the linear response of the magnetization to the small probing field. The probing field modifies the CMF Hamiltonian according to
\begin{equation}
\hat{H}_{\text{CMF}}\rightarrow\hat{H}_{\text{CMF}}+\delta_\beta\sum_k{\hat{\sigma}^\beta_k},{\forall k}
\label{Hcmf_ext}
\end{equation}
where $\delta_\beta$ is the strength of the external driving field along $\beta$-direction ($\beta = x, y$). The magnetic susceptibility is a four-component tensor. Each component characterizes the response magnetization along $\alpha$-direction to the probing field along $\beta$-direction and is given by
\begin{equation}
\chi_{\alpha\beta} =\lim_{\delta_{\beta}\rightarrow 0}{ \frac{\partial \langle\sigma^\alpha \rangle_{\text{ss}} }{\partial \delta_{\beta}}},
\label{chiab}
\end{equation}
where $\alpha,\beta=\{x,y\}$ denote the directions of external field and response magnetization, respectively.
In the following, we will concentrate on the modulus of the $\chi_{xx}$ component of dissipative TFI model while $\chi_{yx}$ for disspative XYZ model.
\begin{table*}
\centering
\caption{The coefficients, critical points and critical exponents of steady-state phase transition in the dissipative TFI model extracted from the CAM with different groups of clusters. The errors indicate $95\%$ confidence intervals, obtained from the estimate by least-squares fitting.}
\begin{tabular}{p{1.2cm}|p{1.8cm}p{1.8cm}p{1.8cm}|p{1.8cm}p{1.4cm}p{1.8cm}|p{1.8cm}p{1.4cm}p{1.8cm}cccccccccc}
\hline
\hline
&\multicolumn{3}{c|}{($1\times1$, $2\times2$, $3\times3$, $4\times4$)} & \multicolumn{3}{c|}{($1\times1$, $2\times2$, $3\times3$)} &\multicolumn{3}{c}{($2\times2$, $3\times3$, $4\times4$)} \\
\hline
$\Gamma/V$&$c_0$&$g^*/2V$&$\gamma^*$&$c_0$&$g^*/2V$&$\gamma^*$&$c_0$&$g^*/2V$&$\gamma^*$ \\
\hline
0.1& 0.059(18)& 0.788(43)& 1.33(20)& 0.058& 0.776& 1.37& 0.063& 0.794& 1.30\\
0.2& 0.119(15)& 0.787(16)& 1.335(81)& 0.117& 0.783& 1.35& 0.122& 0.790& 1.32\\
0.05& - & - & - & 0.028&0.768& 1.34& - & - & - \\
0.25& - & - & - & 0.146&0.775& 1.37 & - & - & - \\
\hline
\hline
\end{tabular}
\label{cam_stra}
\end{table*}
\subsection{Numerical simulations}
\label{sec_numerical}
In this subsection, we will show the way for accessing the steady-state magnetization and susceptibility in our specific models. In the CMF approximations, one need to solve the CMF master equation Eq. (2). For large clusters, this is demanding since the dimension of Hilbert space increases exponentially with the size increasing. In order to avoid solving the full master equation regarding the density matrix of size $2^L\times 2^L$, one can combine the idea of CMF and quantum trajectory (QT) simulation. In this paper, for $l\le3$, we evaluate the master equation~(\ref{ME_CMF}) to long times through direct Runge-Kutta integration. Instead, for $l=4$ we rely on QT simulations (with the number of trajectories being 500). The details of CMF QT simulation are discussed in Refs. \cite{jin2016,daley2014}.
The steps of accessing the steady-state susceptibility through QT are the following. We first simulate the time-evolution of the density matrix in the presence of probing field $\delta_x$. Second, we average the temporary magnetization over a time window, in which the system nearly approaches to the steady state, as the steady-state magnetization $\langle\hat{\sigma}^{x,y}\rangle_{\text{ss}}$ . Finally we determine the susceptibility by linear fitting of $\langle\hat{\sigma}^{x,y}\rangle_{\text{ss}}$ to $\delta_x$.
\begin{figure*}[htb]
\includegraphics[width=0.9\linewidth]{qt_chi.pdf}
\caption{(a) The QT time-evolutions of the magnetization $\langle\sigma_1^x(t)\rangle$, which is the magnetization of the corner site of the $3\times4$ cluster, in the dissipative TFI model for three different probe fields $\delta_x$ (in units of $\Gamma$). The dashed horizontal lines highlight the steady-state values of the magnetization, which is obtained by averaging the temporary magnetization over $t\in{\left[150,200\right]}$. (b) The steady-state magnetization of TFI model $\langle\sigma_1^x\rangle_{\text{ss}}$ for different values of the probing field $\delta_x$. The black dash line is the linear fitting for $\langle\sigma_1^x\rangle_{\text{ss}}$ to $\delta_x$. The errorbars denote the standard deviation in taking the average. The parameters are chosen as $\Gamma/V = 0.2$ and $g/2V = 0.85$. (c) The $4\times4$ quantum trajectory time-evolution of $\langle\sigma^y_1(t)\rangle$ in the presence of probing field along $x$-direction in two-dimensional dissipative XYZ model. The horizontal dashed lines highlight the steady-state magnetizations. (d) The steady-state magnetization of XYZ model $\langle\sigma^y_1\rangle_{\text{ss}}$ versus the probing field $\delta_x$. The squares denote the steady magnetization which is obtained by averaging the temporary magnetization over $t\in[300,400]$, the error bars denote the standard deviation. The black dashed line is the linear fitting. The susceptibility is given by the slope of the straight line. Other parameters are chosen as $J_x/\Gamma=0.9$, $J_y/\Gamma = 1.036$ and $J_z/\Gamma = 1$.}
\label{cam_steadychi}
\end{figure*}
In Fig. \ref{cam_steadychi}(a), we show the QT time-evolution of the magnetization along $x$-direction $\langle\hat{\sigma}_1^x(t)\rangle$ for the corner site in the $3\times4$ cluster of TFI model. The number of trajectories is 500. Although the temporary magnetization fluctuates due to the randomness inherent to the QT method, the system approaches to the steady state at $t>150$. The steady-state magnetization $\langle\hat{\sigma}_1^x\rangle_{\text{ss}}$ is then obtained by averaging over the time interval $t\in\left[150,200\right]$. The linear fits for $\langle\hat{\sigma}_1^x\rangle_{\text{ss}}$ against $\delta$ is shown in Fig. \ref{cam_steadychi}(b). Similarly, in Figs. \ref{cam_steadychi} (c) and (d), the QT time-evolution of the magnetization $\langle\hat{\sigma}_1^y(t)\rangle$ for different $\delta_x$ and the linear fits for the steady-state magnetization in the dissipative XYZ model are shown.
\begin{figure}[h]
\includegraphics[width=1\linewidth]{TFI_cam_v1.pdf}
\caption{(a) Modulus of $\chi_{xx}$, computed for various cluster sizes as a function of $g/2V$, with $\Gamma/V=0.2$. (b) Behavior of $\chi_0(L)$ as a function
of $g_c(L)/2V$ for $\Gamma/V = 0.1$ (red) and $0.2$ (blue). The solid lines are least-squares fits of Eq.~(\ref{fL}), in which $c_0$, $g^*/2V$ and $\gamma^*$ are determined
by the $l\times l$ ($l=1,2,3,4$) CMF data (filled squares). The $\chi_0(L)$ and $g_c(L)/2V$ of $1\times2$, $2\times3$ and $3\times4$ CMF data are also presented (empty squares).
}
\label{TFI_cam_v1}
\end{figure}
\section{Results}
\label{sec_results}
\subsection{Dissipative TFI model}
\label{sec_results_TFI}
In this subsection, we present the critical exponent and critical point obtained by the CAM for the dissipative TFI model. In Fig.~\ref{TFI_cam_v1}(a) we fix $\Gamma/V=0.2$ and show the CMF results for $\left|\chi_{xx}\right|$ as function of $g/2V$, with various cluster sizes.
One can see that, for each $L$, the magnetic susceptibility always exhibits a singularity at the critical point $g_c(L)/2V$. Moreover, the critical points shift to the left with increasing cluster
size. The consistent approach of $g_c(L)$ to the true critical point $g^*$ supports choosing $l\times l$ clusters
as a systematic series of approximations. The extracted critical points and corresponding amplitudes $\chi_0(L)$ are shown in Fig.~\ref{TFI_cam_v1}(b).
In order to access the critical point $g^*/2V$ and exponent $\gamma^*$ in the thermodynamic limit, we fit the $\{\chi_0(L),g_c(L)/2V\}$ data of $l\times l$ clusters ($l=1,2,3,4$)
according to Eq.~(\ref{fL}), using the least squares method. The fit is shown by the solid line in Fig.~\ref{TFI_cam_v1}(b). We also find that the CAM analysis is robust against slight anisotropies of the clusters. Although the fit of the coherent anomaly is based on the isotropic clusters (denoted by filled squares), the data of $l\times (l+1)$ clusters (denoted by empty squares) show good agreement. A similar CAM analysis is implemented for $\Gamma/V = 0.1$. The set of $\{\chi_0(L),g_c(L)/2V\}$ and the corresponding fitting curve are shown in Fig.~\ref{TFI_cam_v1}(b) as well.
For both values of $\Gamma/V$, we obtain a critical exponent $\gamma^*\approx 1.3$. It is interesting to compare this value to Maghrebi {\it et.al.} which, from field-theory arguments, argued that the steady-state phase transition belongs to the universality class of the equilibrium 2D Ising model \cite{maghrebi2016}. Remarkably, our CAM analysis gives a critical exponent significantly smaller than critical exponent of classical 2D Ising model $\gamma^{\text{Ising}}=7/4$. The difference between the two predictions can be appreciated visually in the log-log representation of coherent amplitudes versus critical points, shown in Fig.~\ref{TFI_loglog}. While the numerical data follow an approximate linear relation, the slope is quite distinct from the critical exponent of the classical 2D Ising model (dashed lines).
To test the stability of the CAM analysis, we have adopted two other possible choices of $l\times l$ clusters: $l=1,2,3$ and $l=2,3,4$. In both cases, $c_0,\lambda^*,\gamma^*$ can be extracted by a direct solution of Eq. (\ref{gammastar}). The results are listed in Table~\ref{cam_stra} and indicate a good stability of the CAM results. In particular, the averaged value of $\gamma^*$ over the three choices is $\bar{\gamma}^*\approx 1.34$ for both $\Gamma/V = 0.1$ and $0.2$, with small discrepancies $\Delta\gamma^*=|\gamma^*-\bar{\gamma}^*|/\bar{\gamma}^* \lesssim 3\%$.
On the other hand, the error associated with the limited cluster size, $l\leq 4$, is difficult to assess. Some insight is provided by the CAM analysis of the classical 2D Ising model. By only considering $l=1,2$ (with the exact critical temperature as additional input), Suzuki obtained $\gamma^{\text{Ising}}\approx 1.67$ \cite{suzuki1986}. This value is relatively close to the exact result 1.75, suggesting that CAM can yield good estimates of the critical exponents already at small cluster sizes. However, considerable efforts might be necessary to improve the accuracy. Extending the analysis of the classical 2D Ising model to clusters with up to 145 sites ($l \simeq 12$) gives $\gamma^{\text{Ising}}\approx 1.7$ \cite{suzuki1987}.
Additionally, we also extract the critical points and critical exponents for $\Gamma / V = 0.05$ and $0.25$ by CAM involving only the $1\times 1$, $2\times2$, and $3\times 3$ clusters. The results are shown in Tab. \ref{cam_stra}.
\begin{figure}[h]
\includegraphics[width=1\linewidth]{TFI_loglog.pdf}
\caption{Log-log plot of amplitudes as functions of critical points $g_c(L)$, for $\Gamma/V = 0.1$ (red) and $0.2$ (blue).
The squares are the amplitudes $\chi_0(L)$. The solid lines use the averaged critical exponents of the dissipative TFI model, $\bar{\gamma}^* = 1.34$. The dashed lines assume the exact critical exponent of the classical 2D Ising model, $\gamma^{\text{Ising}} = 1.75$.
}
\label{TFI_loglog}
\end{figure}
\subsection{The dissipative XYZ model}
\label{sec_results_XYZ}
For the dissipative XYZ model, we focus on the critical exponent of the susceptibility component $\chi_{yx}$. In Fig.\ref{XYZ_cam_v1} (a), we show the CMF results of the steady-state susceptibility with different clusters for $J_x/\Gamma = 0.9$. The amplitude $\chi_0(L)$ and the critical point $J_{y,c}(L)/\Gamma$ for various clusters are extracted according to Eq.(\ref{chi_mf}) and shown in Fig.\ref{XYZ_cam_v1}(b).
The CAM is then implemented up on the three choices of clusters. The extracted coefficient, critical exponent and critical point are listed in Tab. \ref{cam_XYZ1}. The critical point $J_{y}^*/\Gamma$ obtained from the CAM analysis is consistent with the previous results through other approaches \cite{jin2016}. The CAM analysis gives a value of $\gamma^*$ which is slightly different from the mean-field result. We cannot however benchmark our results with other methods, so it is difficult to determine if the difference is significant. Notice that although both the dissipative TFI and XYZ models break the $\mathbb{Z}_2$ symmetry through the phase transition, the critical exponents are discriminated.
\begin{figure}[htb]
\includegraphics[width=1\linewidth]{XYZ_cam.pdf}
\caption{(a) The susceptibility $\chi_{yx}$ as a function of $J_y$ for different clusters. (b)The amplitude $\chi_0(L)$ as a function of $J_{y,c}(L)$. The parameters are chosen as $J_x/\Gamma = 0.9$ and $J_z/\Gamma = 1$. From right to left the squares denote the amplitudes $\chi_0(L)$ for $1\times1$, $1\times2$, $2\times2$, $2\times3$, $3\times3$, and $4\times4$ cluster, respectively.
}
\label{XYZ_cam_v1}
\end{figure}
\begin{figure}[htb]
\includegraphics[width=1\linewidth]{camJxJy.pdf}
\caption{(a) The amplitudes $\chi_0(L)$ as functions of $\left[J_c(L)-J^*\right]/\Gamma$ for different values of (a) $J_x/\Gamma$ and (b) $J_y/\Gamma$. The squares denote the amplitudes for each of the clusters while the line denote the fittings with the critical exponents extracted from the CAM analysis. We have put the $\chi_0(L)$ for anisotropic clusters in the plots. From the left to right, the squares correspond to the $1\times1$, $1\times2$, $2\times2$, $2\times3$, and $3\times3$ clusters, respectively. The coupling is chosen as $J_z/\Gamma = 1$.
}
\label{camJxJy}
\end{figure}
\begin{table*}
\centering
\caption{The coefficients, critical points and critical exponents of steady-state phase transition in the dissipative TFI model extracted from the CAM with different series of clusters.}
\begin{tabular}{p{1.2cm}|p{1.8cm}p{1.8cm}p{1.8cm}|p{1.8cm}p{1.4cm}p{1.8cm}|p{1.8cm}p{1.4cm}p{1.8cm}cccccccccc}
\hline
\hline
&\multicolumn{3}{c|}{($1\times1$, $2\times2$, $3\times3$, $4\times4$)} & \multicolumn{3}{c|}{($1\times1$, $2\times2$, $3\times3$)} &\multicolumn{3}{c}{($2\times2$, $3\times3$, $4\times4$)} \\
\hline
$J_x/\Gamma$&$c_0$&$J_y^*/\Gamma$&$\gamma^*$&$c_0$&$J_y^*/\Gamma$&$\gamma^*$&$c_0$&$J_y^*/\Gamma$&$\gamma^*$ \\
\hline
0.9&0.121&1.041&1.033 &0.125&1.041&1.030&0.121&1.041&1.034\\
0.8& - & - & - &0.0659& 1.021 & 1.023& -& -& -\\
0.7& - & - & - & 0.0433& 1.014 & 1.026& - & - & - \\
0.6 & - & - & - & 0.0298& 1.011 & 1.042 & - & - & - \\
\hline
\hline
\end{tabular}
\label{cam_XYZ1}
\end{table*}
Additionally, we also apply CAM analysis for $J_x/\Gamma=0.6$, $0.7$ and $0.8$ as well as $J_y/\Gamma=1.1$, $1.2$, and $1.3$ by using the $1\times1$, $2\times2$, and $3\times3$ clusters. The results are listed in Tabs. \ref{cam_XYZ1} and \ref{cam_XYZ2}. In Fig. \ref{camJxJy}, we show the amplitudes $\chi_0(L)$ as functions of $[J_{c}(L)-J^*]/\Gamma$ in the log-log scale. One can see that for case of fixed $J_x$, the critical exponents are stable around $\gamma^*\approx 1.03$, while for the case of fixed $J_y$, the critical exponents change in a wide range. We notice that for large $J_y$ the system is close to the critical regime in which the correlation length increases and clusters with larger size should be involved in the CAM analysis.
\begin{table}
\centering
\caption{The coefficients, critical points and critical exponents of steady-state phase transition in the dissipative XYZ model. The involved clusters are $1\times1$, $2\times2$, and $3\times3$.}
\begin{tabular}{p{2cm}p{2cm}p{2cm}p{2cm}}
\hline
\hline
$J_y/\Gamma$ & $c_0$ & $J_x^*/\Gamma $ & $\gamma^*$ \\
\hline
1.1 & 0.112 & 0.959 & 1.062 \\
1.2 & 0.0382& 0.978 & 1.128 \\
1.3 & 0.00389& 0.977 & 1.563 \\
\hline
\hline
\end{tabular}
\label{cam_XYZ2}
\end{table}
\section{ Summary}
\label{sec_summary}
In this work we have introduced the CAM as an effective method to determine critical exponents of phase transitions in driven-dissipative systems. In particular, we have applied CAM to the dissipative TFI model and XYZ model in two dimensions, finding a critical exponent for the magnetic susceptibility which is different from the classical value. Our study indicates that CAM is powerful in accessing the critical properties, in comparison to the resources required. We notice that although the implementation of CAM in the parameter regions considered in this paper is valid, one should be careful when going outside these regions. For example, the continuous phase transition in dissipative TFI model may break down for sufficiently strong dissipation where the CAM analysis cannot be implemented \cite{overbeck2017}.
We would like to emphasize that the sizes of the involved clusters also have effects on the performance of CAM for open systems. In other words, for systems that are dominated by long spatial correlations, large size clusters should be involved in the CAM analysis, in order to reach reliable results. For example, in the one-dimensional driven pair contact process with diffusion (PCPD) the critical exponents obtained through the CAM analysis are in excellent accord with simulation results \cite{park2005}. However, for ordinary PCPD (without driving) a non-negligible discrepancy is observed up to the maximum cluster size 13. This reminds us that the results obtained by CAM should be interpreted with caution.
To further enlarge the cluster size, there are several promising approaches worth to pursue in future work, involving a combination of corner-space renormalization~\cite{finazzi2015}, neural-networks~\cite{vicentini2019,hartmann2019,Yoshioka2019,Nagy2019} and/or tensor-networks~\cite{weimer2019,Paeckel2019}. On one side, the corner-space renormalization method and the recently introduced neural-networks quantum ansatz were shown to simulate the steady states of open quantum many-body systems, benchmarked with
a few examples including small $2$D clusters. At the moment, it is not clear how large are the computational cost and convergence of the methods for: (i) larger clusters sizes; (ii) quantum states close to criticality, for which one expects highly-entangled states; (iii) their extension to time-dependent open dynamics (as required in CAM). On the other side, tensor-networks approaches are well established methods with good convergence for short-range systems and low-entangled states. The extension of the method to $2$D clusters (effectively mapped to a long-range $1$D system) turns the simulation computationally harder at the expense of an increasing memory cost and simulation time. One could still exploit the benefits of CAM in this case, recalling that it requires a large but not \textit{macroscopically} large cluster, such that high performance supercomputers with multiple nodes may hopefully lead to an efficient simulation.
\section*{ACKNOWLEDGMENTS}
We thank M. F. Maghrebi and J. Keeling for useful correspondence. J.J. acknowledges support from the National Natural Science Foundation of China (NSFC) via Grant No. 11975064. W.B.He acknowledges support from Grant No. U1930402 of the National Science Association Funds of NSFC. F.I. acknowledges the financial support of the Brazilian funding agencies National Council for Scientific and Technological Development - CNPq (Grant No. $308205/2019-7$) and FAPERJ (Grant No. E-$26/211.318/2019$). Y.-D. W. acknowledges support by the National Key R\&D Program of China under Grant No. 2017YFA0304503, and the Peng Huanwu Theoretical Physics Renovation Center under grant No. 12047503. S.C. acknowledges support from the National Key R\&D Program of China No. 2016YFA0301200 and the NSFC Grant No. 11974040. R. F. acknowledges partial financial support from the Google Quantum Research Award. R. F. research has been conducted within the framework of the Trieste Institute for Theoretical Quantum Technologies (TQT).
\noindent
|
\section*{Results}
To study their dynamic behavior and interactions, we confine swimmers to 1-dimensional tracks by exploiting their strong affinity for surfaces~\cite{Das2015,Simcchen2015,Brown2016, Ketzetzi2020, Ketzetzi2020sep}. This affinity stems from their propulsion mechanism, which is based on an asymmetric catalytic decomposition of H$_2$O$_2$ on their Pt-coated hemisphere~\cite{Golestanian2005}. We equip planar substrates with designed 3D micro-printed posts, thereby effectively creating preferred 1-dimensional environments around the posts, where the swimmers can be in close vicinity to both posts and substrates~\cite{Das2015,Simcchen2015,Brown2016}. See Figure~\ref{fig:fig1}A for one example of our experimental setups featuring circular posts. The H$_2$O$_2$ decomposition reaction here creates gradients in solute molecules which act over the swimmer surfaces, posts and substrates, inducing phoretic and osmotic flows thereby causing swimmer capture~\cite{Ketzetzi2020, Ketzetzi2020sep}. That is, when a swimmer encounters a post, it quickly gets captured into motion along it, and is retained there for very long times.
Once attached to a circular post, our swimmers with diameters of (2.00 $\pm$ 0.05)~{\textmu}m and Pt coating thicknesses of (4.7 $\pm$ 0.2) nm moved with equal probability in either the clockwise or counterclockwise direction, without switching direction. Single swimmers, as well as multiple swimmers with the same direction of motion, on a given post, orbited their posts with approximately constant speed. We highlighted these swimmers by coloring their corresponding posts in red in Figure~\ref{fig:fig1}A. These swimmers continued their orbiting motion for at least 30 min in our experiments, during which we did not see them leave. However, swimmers that moved in opposing directions on a given post, hindered each others motion and were immobilized, see Supplementary Movie 1 as an example. We indicated these immobilized clusters by coloring their adjacent posts blue in Figure~\ref{fig:fig1}A.
\begin{figure*}[ht]
\centering
\includegraphics[width=1\linewidth]{FIG2.PNG}
\caption{\textbf{Multiple microswimmers orbiting circular posts.} Red color has been added to distinguish the stationary posts from the orbiting swimmers. Scale bars are 5~{\textmu}m. Sequence of light microscopy images showing orbiting of \textbf{A, B)} two and \textbf{C)} three swimmers around a 4~{\textmu}m diameter post, and \textbf{D)}~two swimmers around an 8~{\textmu}m diameter post. See also Supplementary Movies 2-4. Each column in panels (B)-(D) shares the same timestamp, indicated in (B). \textbf{(E)} The PDFs of propulsion speed in orbit for three swimmers on a post are almost identical. \textbf{(F)} Average propulsion speed as a function of the number of swimmers in orbit for two post sizes, showing that speed increases with the number of comoving swimmers for both post sizes in a similar fashion. All measurements were taken under the same experimental conditions. The dashed line is a least squares fit with $y=\alpha x +b$, with $\alpha$ and $b$ being (1.6 $\pm$ 0.2)~{\textmu}m/s and (2.9 $\pm$ 0.4)~{\textmu}m/s, respectively.}
\label{fig:fig2}
\end{figure*}
Surprisingly, long-term capture around posts happens even when post size is comparable to the swimmer size, see for example Figure~\ref{fig:fig1}B for a swimmer orbiting a 4~{\textmu}m diameter post. This is in stark contrast with simulations on model squirmers~\cite{Spagnolie2015, Kuron2019}, which are often used to approximate synthetic swimmers such as ours. For model squirmers, capture has only been proposed for posts of sizes several times larger than the squirmer~\cite{Spagnolie2015}. Moreover, simulations for the orbiting of a squirmer around a sphere predict a relation between curvature and squirmer speed, as can be readily obtained from the data in Ref.~\cite{Kuron2019}. To test these predictions, we track the swimmers in time and extract their speed in orbit using python routines~\cite{Trackpy}. We find that post curvature does not influence the propulsion speeds of our catalytic swimmers, at least not for the range shown in Figure~\ref{fig:fig1}C. These discrepancies between our experiments and simulations imply that catalytic swimmers are different from pure squirmers, probably due to their propulsion mechanism and to the long-range solute gradients that act across the substrate and posts~\cite{Ketzetzi2020,Ketzetzi2020sep}.
Despite orbiting short tracks, we find that propulsion speed in orbit is rather stable, see Figure~\ref{fig:fig1}D which shows that speed in orbit follows a Gaussian distribution. Note that the presence of the post itself does not have a considerable effect on propulsion speed: the speed of a swimmer in orbit is only slightly reduced with respect to its ``free" speed parallel to the substrate alone in Figure~\ref{fig:fig1}E. The dashed line is a least-squares fit with $y = ax$ where $a$ is ($0.83\pm0.08$), in line with the slope obtained previously for bimetallic microrods~\cite{Takagi2014}. These findings point out that any hydrodynamic and/or phoretic coupling to the post leads to a subdominant contribution to the speed.
Intriguing effects take place when multiple swimmers orbit the same post: the swimmers move with seemingly equal speeds while also maintaining roughly constant distances, see Figure~\ref{fig:fig2} and Supplementary Movies 2-4. This constancy in speed and distance appears both for two (Figure ~\ref{fig:fig2}A, B) and three comoving swimmers (Figure~\ref{fig:fig2}C), and is independent of the post curvature (Figure~\ref{fig:fig2}D). More quantitatively, we find that swimmers that orbit around the same post in fact have almost the same speed distribution, independent of the particle number and post size. See Figure~\ref{fig:fig2}E for the speed distributions of three swimmers on a 4~{\textmu}m post and Supplemental Information (SI) for additional data.
Even more strikingly, speed increases with the number of swimmers in orbit, as shown in Figure~\ref{fig:fig2}F for posts with 4 and 8~{\textmu}m diameters. That is, two particles orbit faster than one, and in turn three particles orbit faster than two: Under otherwise fixed conditions, speed increases by $\approx$ 20\% and 60\% for two and three comoving swimmers on 4~{\textmu}m posts, respectively, and by $\approx$ 40\% for two comoving swimmers on the 8~{\textmu}m post, in comparison to single swimmers. Interestingly, for our system there is no significant speedup of a pair of particles with respect to a separated third particle moving along the post. This contrasts strongly with the result of passive, driven particles in a torroidal optical trap~\cite{Curtis2003,Reichert2004,Lutz2006}, where pair formation and breaking is observed. That is, two driven particles overtake a third, which then leads to a fracture of the triplet with the two lead particles moving off, a result that can be understood using hydrodynamic theory~\cite{Curtis2003,Reichert2004,Lutz2006}.
These findings strongly suggest \textit{cooperative motion} of the microswimmers. Chemical gradients of swimmers in close proximity interact, thereby causing multi-bound states and a collective speedup. This speedup is independent of the post size for the here considered 4 and 8~{\textmu}m posts, see Figure~\ref{fig:fig2}F. The dashed line represents a least squares fit with $y = \alpha x+b$, with $\alpha$ (1.6 $\pm$ 0.2)~{\textmu}m/s and $b$ (2.9 $\pm$ 0.4)~{\textmu}m/s, implying a linear relationship between the number of swimmers and the collective speedup. This observation of a speedup as well as the seemingly constant swimmer distance in Figure~\ref{fig:fig2}, is surprising in view of the distribution in speed that single swimmers exhibit, and implies that swimmers experience a coupling that adjusts their speed. To understand the origin of this behavior, we will now quantify the swimmer distances as well as test for correlations of the speeds of pairs of comoving swimmers.
\begin{figure*}[t]
\centering
\includegraphics[width=1\linewidth]{FIG3.PNG}
\caption{\textbf{Interactions of microswimmers comoving along circular posts: Experiments and modeling.} \textbf{A)} Representation of the arc distance. \textbf{B)} Snapshots of swimmer pairs in orbit, color-coded so that each color marks the corresponding swimmer pair in panels (C-D). \textbf{C)} PDF of the arc distance between comoving swimmers, showing that swimmers assume a relatively constant distance in orbit with minimum center-to-center distance $\approx$ 3~{\textmu}m. Measurement duration is $\approx$ 5 min. \textbf{(D)} Potential energy in terms of thermal energy, as obtained from the swimmer distances in (C) using the Boltzmann distribution. The dashed lines represent least-squares fits with $y = \frac{1}{2}k(x-x_0)^2+y_0$. All fitted parameters are listed in the SI. \textbf{E)} Scatter plot of the speeds of the two swimmers comoving along the 8~{\textmu}m post colored magenta in panels (B-D), showing that swimmer speeds are not correlated. \textbf{(F-H)} Effective separation between swimmers based on hydrodynamic and osmotic balance. \textbf{(F)} The balance between an inward osmotic flow along the wall and an outward pusher-type dipolar flow away from the swimmer leads to a curve of zero velocity depicted in red. \textbf{(G)} Two swimmers that lie head-to-tail as in panel (A) assume a fixed distance as evidenced by the x-axis intercept. The shape of the relative velocity generates an effective potential. \textbf{(H)} Swimmer separation distance as a function of the ratio between the osmotic and pusher contributions, with $\lambda$ and $\kappa$ indicating the respective strengths of the inward and outward flows.}
\label{fig:fig3}
\end{figure*}
We quantify the arc distance, $\ell$, between comoving swimmers as depicted in Figure~\ref{fig:fig3}A. We measure the distances between various swimmer pairs, on differently sized posts and with different number of attached swimmers, see Figure~\ref{fig:fig3}B. In all cases, we find that swimmers orbit the posts at a preferred distance (Figure~\ref{fig:fig3}C), in line with our expectations based on Figure~\ref{fig:fig2}. Specifically, the average distance measured over a 4~min time interval is (5.3 $\pm$ 1.4)~{\textmu}m and (4.2 $\pm$ 0.5)~{\textmu}m for two swimmers on a 4 and 8~{\textmu}m post, respectively, and (3.8 $\pm$ 0.3)~{\textmu}m and (4.3 $\pm$ 0.4)~{\textmu}m between the front-middle and middle-back swimmers of a three-swimmer system on a 4~{\textmu}m post, respectively. We notice that swimmers never approach closer than a minimum center-to-center distance of (3.0 $\pm$ 0.1)~{\textmu}m. In addition, the arc distances we find here show similar distributions, albeit with different peak values and widths.
These common features raise the question whether the constancy in the distance stems from correlations between the speeds of the comoving swimmers. However, a closer examination reveals that speeds always vary independently of one another, see the scatter plot of the speeds of two comoving swimmers along an 8~{\textmu}m post, V$_1(t)$ vs V$_2(t)$, where $t$ is time in Figure~\ref{fig:fig3}E and SI. Already, the symmetric shape of the scatter plot indicates that they are not correlated, which is further supported by a Pearson correlation coefficient of 0.2. In addition, we excluded time-delayed correlations by considering the correlation between $V_1(t)$ and $V_2(t+\tau$) see SI, with $\tau$ the time between two frames. Again, the Pearson correlation coefficient was 0.1, signifying no speed correlation. Hence, there must be an alternative mechanism that can explain why the swimmers move with the same average speed and stable distance.
The random fluctuations around an average distance is reminiscent of the thermal motion of a passive particle inside a potential well. Here, this effective potential must be induced by the active state of \textit{both} swimmers. A single active particle is not sufficient, because passive particles in the vicinity of an active particle do not get confined but are instead either dragged along the fluid flow around the active particle~\cite{Campbell2019} or attracted to the active particle site~\cite{Palacci2013}.
Seeing that our swimmers are moving together at the same average speed, we consider their activity-confined motion in the rotating frame that comoves with the swimmers. While this is a non-inertial frame of reference and an out-of-equilibrium state, it still allows us to view our system from a simplified perspective: that of random thermal motion of each particle in a potential induced by the other. To do so, we use the Boltzmann distribution, $\mathrm{PDF}(\ell) = \exp{(-U(\ell)/k_BT)}$ to extract the relative potential energy for our systems from the probability density function of their corresponding distances. This allows us to measure the probability for the system to be in a certain distance relative to the distance's energy and system temperature in Figure~\ref{fig:fig3}D, where the relative energy is $U(\ell)/k_BT = -\log (\mathrm{PDF}(\ell)) + U/k_BT$, with $U/k_BT$ an arbitrary reference state, set such that the relative energy goes to 0 at infinite distances.
\begin{figure*}[ht!]
\centering
\includegraphics[width=1\linewidth]{FIG4.PNG}
\caption{\textbf{Actively-assembled chains of microswimmers on peanut-shaped posts.}
\textbf{(A-D)} Schematic representations of the conditions determining chain mobility. Red and blue colors indicate mobile and immobilized swimmers, respectively. The darker hemispheres represent the Pt side, which drives motion. \textbf{(A)} Two swimmers with opposing motion direction become immobile. \textbf{(B)} When a third swimmer approaches the cluster, it gets in close touch and also becomes stationary. \textbf{(C)} A fourth swimmer, entering from the same side as in (B), causes the chain to move again in the direction of the majority of swimmers. \textbf{(D)} In highly curved neck regions, swimmers stop temporarily because a reorientation is required, see also the light microscopy image at the bottom. \textbf{(E)} Dynamic chain formation, see also Supplementary Movie 5: initially, self-propelled swimmers headed by the red swimmer move clockwise in a train. After being stopped by an immobile cluster at the edge, the train becomes a compact chain that moves after further addition of swimmers. \textbf{(F)} Center-to-center distance between swimmer pairs within a mobile chain that self-propels in the direction of swimmer ``1", i.e. swimmer ``1" is at the front end while swimmer "9" is at the back end, see also the light microscopy image on the right. Swimmers at the ends maintain at all times larger distances from their neighbors, in comparison to those positioned in the middle. \textbf{(G)} Posts with a more negative neck curvature enhance the formation of compact and immobile chains, because the necks act as accumulation sites, see also Supplementary Movie 6. This time-series is a continuation of that in panel (D). Same color marks the same swimmer throughout each time-series in (E, G).}
\label{fig:fig4}
\end{figure*}
The shape of the resulting effective potentials close to the minimum resembles a harmonic function, and hence we fit them with $y = \frac{1}{2}k(x-x_0)^2+y_0$ using a least-squares fit. This provides us with the depth of the potential wells $y_0$, the preferred distance $x_0$, and the interaction strength $k$, see SI for values. Interestingly, the confinement seems to weaken with increasing preferred distance of the swimmers, which would be in line with the intuitive picture of interacting solute gradients. However, the larger preferred swimmer separation and lower interaction strength might also be due to a difference in one of the participating swimmers.
Let us now turn to the origin of this activity-induced interaction. In view of our recent work~\cite{Ketzetzi2020,Ketzetzi2020sep}, we hypothesize that there is a short-ranged repulsion due to the pusher-type nature of the propulsion mechanism, and a long-ranged attraction due to osmotic flows along the substrate. To lowest order, the former interaction induces an \textit{outward} directed dipolar fluid flow along the symmetry axis of the swimmer~\cite{Campbell2019} that scales as $u_{\mathrm{dip}}(r,\theta) = \kappa \left( 3 \cos^{2} \theta - 1 \right) / ( 2 r^4 )$, while the latter induces an \textit{inward} directed flow due to osmosis along the substrate~\cite{Uspal2018} that scales as $u_{\mathrm{osm}}(r) = -\lambda/r^2$; the factors $\kappa$ and $\lambda$ indicate the respective strength of the outward and inward flows, $r$ the radial distance --- the decay accounts for the presence of a no-slip surface ---, and $\theta$ the angle with respect to the swimmer's orientation.
When the two contributions balance at a finite distance, comoving swimmers can assume a stable separation, see Figure~\ref{fig:fig3}F for a vector plot of the total velocity profile $u_{\mathrm{tot}}(r,\theta) = u_{\mathrm{osm}}(r) + u_{\mathrm{dip}}(r,\theta)$. The angular dependence shows a lemniscate zero-velocity contour, which does not account for the finite size of the swimmer. Clearly, our simple argument would allow for swimmer contact at a finite angle, which can be stabilized by imposing 1-dimensional motion. Figure 3G shows that when aligned head to tail, there is a separation $r_{0}$ that is stable, as indicated by the slope at the intercept. For swimmers comoving in the same direction we obtain a simple expression for the separation as a balance between the inward and outward flow strength: $r_{0} = \sqrt{\kappa / \lambda}$, see Figure 3H. Hence, this profile can indeed be recast into an effective potential, as in Figure~\ref{fig:fig3}D. The collective speed-up observed in Figure~\ref{fig:fig2}F may thus stem from long-ranged chemical gradients leading to an enhancement of the phoretic driving mechanism, or from a collective hydrodynamic effect, such as a drag reduction for the actively-assembled dimers and trimer. Lastly, note that according to this minimal model if one particle is immobilized, the second particle may approach much more closely, possibly even come into contact depending on the strengths of repulsion and self-propulsion; the latter should also hold for particles moving toward each other.
Intriguingly, we can test these predictions in our experiments, where we may be able to induce situations of close contact which would disturb the stable swimmer distance, by temporarily or even permanently stopping one of the swimmers. The latter we have already observed for immobilized pairs of swimmers, on the blue-colored posts in Figure~\ref{fig:fig1}A. These swimmers probably moved in opposite directions before their encounter immobilized them, see Figure~\ref{fig:fig4}A for a schematic drawing. More surprisingly, we also observed immobile clusters consisting of three particles, see Figure~\ref{fig:fig1}A. This implies that a third particle not only was able to approach two immobilized particles much closer than the previously observed 3~{\textmu}m minimum distance, in line with our model prediction, but also that it did not remobilize the cluster despite the uneven particle number and hence presumably unbalanced forces, see Figure~\ref{fig:fig4}B.
To gain better control over the location and duration of the stops and test our hypotheses, we employed micro-printed posts with a peanut shape that feature regions of alternating positive and negative curvature. Swimmers passing through negative curvature points need time to reorient themselves to be able to move on, thereby enforcing temporary stops at precisely defined locations, see Figure~\ref{fig:fig4}D. The more curved the neck regions of these posts, the longer it takes for the swimmers to escape. In addition, these larger posts allow us to study the interactions and behavior of more than three particles.
Equivalently to our observations on the circular posts, and previous work on channels~\cite{Das2015}, we verify that swimmers with opposing directions of motion hinder each other, see the bottom edge of the peanut-shaped post in Figure~\ref{fig:fig4}E ($t$=0~s). Similarly, a third swimmer is unable to disturb this configuration and joins the immobile cluster, see bottom edge of Figure~\ref{fig:fig4}E ($t$=1~s), where the black-colored swimmer joins the immobile dimer. Thus, such immobilization not only happens in the negative curvature region of the neck, but also along the positive curvature regions of the rounded peanut edges. A fourth swimmer joining the immobile trimer from the right leads to a balanced situation and an immobile tetramer cluster similar to Figure~\ref{fig:fig4}A, see the red-colored swimmer in Figure~\ref{fig:fig4}E at $t$=2~s and $t$=3~s. Based on our experiment we found that a fourth swimmer is able to remobilize an immobile trimer, whenever there are three particles pointing in the same direction, see Figure~\ref{fig:fig4}C. Generally in combining chains, we find remobilization whenever the number of swimmers pointing in one direction exceeds the number of swimmers pointing in the other direction by $\Delta n = 2$. Notice, however, that swimmers pointing in the direction opposite to the net motion do not need to reorient, they are simply pushed along.
Based on the aforementioned simple conditions, we can furthermore deduce the directionalities of the swimmers in compact moving chains, either from their position or from the dynamic information of chain formation. For example, swimmers at the ends of long immobile chains always must have directions that point towards its center. Conversely, in this manner we can also predict the dynamics of a compact chain upon addition of a swimmer.
The larger size of our peanut-shaped microprinted posts enables the attachment of multiple moving swimmers that can actively interact and dynamically self-assemble and disassemble. These posts therefore allow us to see how swimmers that move in trains along the post evolve into compact swimmer chains, see Figure~\ref{fig:fig4}E and Supplementary Movie 5. For example, four swimmers that move in the same clockwise direction form a train led by the red-colored swimmer in Figure~\ref{fig:fig4}E ($t$=0~s). Between $t$=0~s and $t$=1~s a fifth swimmer, initially swimming on the top of the post, enters the train of comoving swimmers from above, in front of the orange one. Interestingly, the swimmers following the red maintain stable distances while orbiting. That is, before being stopped by the immobilized cluster at the bottom-left corner, their in-between distances fluctuate around a preferred distance in accordance with our Figures~\ref{fig:fig2}~and~\ref{fig:fig3} findings. The immobile cluster acts as a stopping point that compacts the train of swimmers into a chain: one by one, the swimmers encounter and join the cluster in close contact, thereby creating a compact chain. Once the number of clockwise moving swimmers is at least by two greater than the number of counterclockwise moving swimmers in the immobile cluster, the entire chain sets into motion. Since the majority of swimmers are moving clockwise, the clockwise direction is imposed to the chain as a whole.
After the chain is formed and remobilized, the distance between a swimmer and its neighbor depends on its position in the chain, see Figure~\ref{fig:fig4}F. Swimmers at the chain ends are further apart from their neighbors as opposed to the ones in the middle that are almost in touch. Swimmers at both ends are positioned at a center-to-center distance of $\approx$ (2.9 $\pm$ 0.2)~{\textmu}m from their neighbors, unlike the swimmers within the chain that move at distances of $\approx$ (2.4 $\pm$ 0.3)~{\textmu}m. For comparison, our swimmers have a diameter of (2.00 $\pm$ 0.05)~{\textmu}m, implying that indeed particles in the center of the chain are almost touching. Note that the distance of the swimmer pairs at the chain ends coincides with the minimum distance found for the swimmer pairs in the circular posts in Figure~\ref{fig:fig3}B. This observation further corroborates our hypothesis of a long-range attraction being present between the swimmers, which is balanced by a short-range repulsion. Because the attraction spans more than a single swimmer, the swimmers in the middle are more compacted than those at the end.
Contrary to the dynamic collective motion we have just described, we furthermore find that highly curved necks capture swimmers for such long times, that in fact they act as semi-permanent stopping points, see Supplementary Movie 6. Although long chain formation is still possible, the self-assembled chains remain immobile due to the pinning of the swimmers by the neck, see Figure~\ref{fig:fig4}D and G: the motion of the red swimmer immediately stops at the neck at $t$=1~s. Thus, a highly curved neck not only prohibits collective motion, but also hinders single swimmer motion since it requires reorientation.
Besides an activity-induced self-assembly into compact chains, chains may also reorganize in time, see Supplementary Movie 7. In Figure~\ref{fig:fig5}A we follow a clockwise self-propelling chain consisting of ten swimmers. \begin{figure}[ht]
\centering
\includegraphics[width=1\linewidth]{FIG5.PNG}
\caption{\textbf{Dynamics of actively-assembled chains: Effect of post topology on chain mobility.} Each color marks the same swimmer within each time-series, except the red color in (A) which marks three different swimmers all escaping the chain; corresponding arrows denote directionality. \textbf{(A)} When it reaches the rounded post edge, a swimmer escapes from the end of the self-propelled compact chain. Swimmers from the middle of the chain also escape during orbiting from the neck or the rounded post edge. See also Supplementary Movie 7. \textbf{(B)} Example of chain breakup: a chain self-propels in the clockwise direction and becomes immobile around the rounded post edge. Two swimmers with opposing direction are incorporated to the chain from the right, and the chain remobilizes in the opposite direction. When reaching the rounded edge, the chain end slows down, the chain breaks, and its end immobilizes. Upon arrival of new swimmers, the end remobilizes at the reverse direction eventually leading to the reformation of the entire chain at the opposite site of the neck. See also Supplementary Movie 8.}
\label{fig:fig5} \vspace{-5pt}
\end{figure}During orbiting, swimmers may leave the chain in the following ways: (1) swimmers at the chain end may leave when they reach locations of comparatively high positive or negative curvature, in line with earlier findings for individual swimmers \cite{Das2015, Simcchen2015}, which here is also likely facilitated due to the larger distances to their neighbor. That is the case for the red swimmer at $t$=0~s, which leaves the chain when it reaches the rounded peanut edge (top left). (2) Likewise, swimmers from the middle may exit when they pass through locations where curvature varies. This situation is visible both at $t$=11.5~s and $t$=13.0~s in Figure~\ref{fig:fig5}A, where a center-chain swimmer highlighted in red escapes while passing through the negatively curved neck and positively curved corner, respectively. In both cases, the chain slowed down before the escape. We speculate that this is enhanced for swimmers with directionality that opposes the direction of motion of the chain.
In addition to facilitating swimmer escape, locations with strong curvature variations furthermore enhance chain breakup and motion reversal, see Supplementary Movie 8 and Figure~\ref{fig:fig5}B as an example. A chain that moves clockwise slows down and becomes immobile when its middle is situated on the rounded edge. The swimmer from the middle leaves the chain while the entire chain remains stationary. At $t$=12~s two swimmers with the same counterclockwise direction as the purple swimmer enter the chain from the right. The entire chain gets remobilized and starts to move in the counterclockwise direction: when reaching the rounded edge, the end of the chain is being slowed down, then the chain breaks, leaving three swimmers immobilized on the rounded edge behind. The rest of the green-swimmer-led chain continues to move counterclockwise until it stops at the other neck side. From the stationary purple chain one swimmer escapes, and gets remobilized in the clockwise direction by two swimmers joining. When the purple chain encounters the immobile green chain, it lifts the immobilization and the entire chain continues in the clockwise direction, in line with what we observed previously in Figure~\ref{fig:fig4}E.
\section*{Discussion}
In summary, catalytically self-propelled microswimmers show a plethora of striking collective effects in 1-dimensional environments. When moving in the same direction microswimmers ``cooperate". That is, they move at a greater speed the more particles join as well as at a preferred distance, larger than a minimum distance. This activity-induced interaction can be described by an effective potential of the order of few $k_BT$. Stopping the chain can overcome this minimum spacing of comoving swimmers, thereby leading to compact chains. These active chains show very rich dynamics themselves, with activity induced self-assembly, compaction, disassembly, breakup, and reformation. A simple rule seems to distinguish between mobility and immobility for these chains: mobility is achieved if the difference in the numbers of opposing swimmers is greater or equal to two. Once in contact, the notion of the potential well still persists, as distances within the self-assembled chains are not uniform, i.e. swimmers at the ends are less tightly packed. Curvature modulations of the posts allow spatial control over chain speed and reorganization, such as compaction, breakup, and immobilization, offering an exciting playground for testing efficient modes of transport. Overall, our findings unravel unprecedented behaviors of model microswimmers in 1-dimensional environments, thereby providing novel insights into activity-induced interactions. Insights into interactions and collective behavior of synthetic microswimmers are pivotal for applications that require increased swimming efficiency or directionality across different environments, as well as answering fundamental questions on the 1-dimensional phase behavior of active systems.
\section*{Methods}
\paragraph*{{\normalfont \textbf{Particles.}}} Spherical latex particles based on polystyrene (2\% cross-linked) with diameter (2.00~$\pm$~0.05)~{\textmu}m, i.e. size polydispersity 2.5\%, were purchased from Sigma Aldrich. Pt-half-coated particles, with a Pt layer thickness of (4.7~$\pm$~0.2)~nm, were produced through physical vapor deposition, as described in Ref.~\cite{Ketzetzi2020,Ketzetzi2020sep}.
\paragraph*{{\normalfont \textbf{3D printed structures.}}} Microstructures were produced with the commercially available microprinter Photonic Professional GT of Nanoscribe which uses two-photon lithography. The microprinter was equipped with a 63X oil-immersion objective (Zeiss, NA = 1.48) and used to print the 3D structures in oil mode. Microstructure designs were performed in Autodesk Inventor and processed with Describe. The microstructures were printed onto glass coverslips, pre-cleaned with isopropanol, using the commercial photoresist IP-L as a pre-polymer. After printing, the structures were developed by submersion in propylene glycol methylether acrylate for 15 min, followed by gently dipping into isopropanol three times to remove the unpolymerized photoresist. The structures were subsequently dried with gentle air flow. All procedure was done under yellow light.
\paragraph*{{\normalfont \textbf{Imaging.}}} Pt-half-coated particles were dispersed in a 10\% aqueous H$_2$O$_2$ solution. Their motion was recorded above the planar walls with a ELWD 60x objective (S Plan Fluor, NA 0.7, zoomed at 1.5x i.e.~0.1 $\mu$m/px) mounted on an inverted Nikon Eclipse Ti microscope at a frame rate of 5 and 9 fps along the circular and peanut-shaped posts, respectively, within the first hour after sample preparation.
\paragraph*{{\normalfont \textbf{Analysis.}}} Particle positions above the planar wall and along the circular posts were obtained using the Python tracking algorithm Trackpy~\cite{Trackpy}. The speed of all particles was determined using the time derivatives of spatial displacements at consecutive frames, see inset of Figure~\ref{fig:fig1}D for the speed distribution of a single particle measured in orbit for $\approx$ 4 min ($>$ 1200 frames). For swimmers on circular posts, the (arc) displacement was obtained according to Figure~\ref{fig:fig3}A. On the peanut-shaped posts, distances were obtained using the NIS-Elements Advance Research software package by Nikon.
\paragraph*{{\normalfont \textbf{Acknowledgements}}}
D.J.K. gratefully acknowledges funding from the European Research Council (ERC) under the European Union's Horizon 2020 research and innovation program (grant agreement no. 758383). J.d.G. thanks NWO for funding through Start-Up Grant 740.018.013 and through association with the EU-FET project NANOPHLOW (766972) within Horizon 2020.
\clearpage
|
\section{Introduction and Main Results}
\label{sec:introduction_and_main_results}
We will study some aspects of the thermodynamic formalism for maps related to higher rank Anosov actions. Let $M$ be a compact metric space and $f:M\rightarrow M$ an homeomorphism. For $x\in M,\epsilon>0, n\in \Nat$ we denote by $D(x,\ep)$ the open disc of center $x$ and radius $\ep$, and by $\Ben$ the open $(\epsilon,n)$-Bowen ball centered at $x$,
\[
\Ben=\{y\in M:d(f^jx,f^jy)<\epsilon,\ j=0,\ldots,n-1 \}.
\]
Recall that the topological pressure of the (continuous) potential $\varphi:M\to \Real$ is given by
\[
P_{\mathit{top}}^f(\varphi)=\Ptop(\varphi):=\lim_{\epsilon\mapsto 0}\limsup_{n\mapsto\oo}\frac{\log S(\epsilon,n)}{n}
\]
where
\[
S(\epsilon,n)=\inf\{\sum_{x\in E}e^{\SB(x)}:M=\cup_{x\in E}D(x,\epsilon,n)\}\quad \SB=\sum_{i=0}^{n-1}\varphi\circ f^i.
\]
Denote by $\PTM{f}{M}$ the set of $f$-invariant Borel probability measures on $M$. For $\nu\in \PTM{f}{M}$ the pressure of $\varphi$ with respect to $\nu$ is
\[
P_{\nu}(\varphi):=\nu(\varphi)+h_{\nu}(f).
\]
\begin{theorem*}[Variational Principle]
It holds
\[
\Ptop(\varphi)=\sup_{\mathclap{\nu\in \PTM{f}{M}}}\ P_{\nu}(\varphi)
\]
\end{theorem*}
This by now classical theorem is due to Walters \cite{WaltersPres}.
\begin{definition}
A measure $\mu\in \PTM{f}{M}$ is an equilibrium state for the potential $\varphi$ if $\Ptop(\varphi)=P_{\mu}(\varphi)$.
\end{definition}
The existence, uniqueness and general properties of equilibrium states is of utmost relevance in ergodic theory, in particular for the differentiable one. Indeed, one of the most important theorems in smooth ergodic theory is the following.
\begin{theorem*}
Let $M$ be a closed (that is, compact and without a boundary) manifold and consider a $\mathcal{C}^1$ diffeomorphism $f:M\rightarrow M$ having a hyperbolic attractor $\Lambda$ (in particular, $f|\Lambda$ is transitive). Then for every H\"older potential $\varphi:\Lambda\rightarrow \Real$ there exists a unique equilibrium state $\muf$ for $f|\Lambda$. Furthermore, is $f|\Lambda$ is topologically mixing then the system $(f,\muf)$ is metrically isomorphic to a Bernoulli shift.
\end{theorem*}
As stated, this theorem is a culmination of the work of several authors. The existence and uniqueness is essentially due to Sinai, Ruelle and Bowen \cite{SinaiMarkov},\cite{SRBattractor},\cite{Bowen1974}, while the Bernoulli property is mainly due to Ornstein and Weiss \cite{GeoBernoulli}. See also \cite{BowenBernoulli}.
In this work we investigate the thermodynamic formalism for a generalization of the above systems, namely when instead of a single hyperbolic map we have an Anosov action. Let us recall that a differentiable action of a Lie Group on a compact Riemannian manifold $\alpha:G\curvearrowright M$ is Anosov if the $G$-orbits of $\alpha$ foliate $M$ and there exists $g\in G$ an Anosov element, i.e.\@ there exist an $\alpha(g)$-invariant decomposition $TM=E^s\oplus E^c\oplus E^u$ and $C>0,0<\lambda<1$ such that $E^c_x=T_xG\cdot x$ and with respect to the metric on $M$ it holds
\begin{align*}
&v\in E^s_x\Rightarrow \forall n\geq 0\quad \norm{D_x\alpha(g)^n(v)}\leq C\lambda^n\norm{v}\\
&v\in E^u_x\Rightarrow \forall n\leq 0\quad \norm{D_x\alpha(g)^n(v)}\leq C\lambda^{-n}\norm{v}.\\
\end{align*}
In this case it is possible to modify the metric to guarantee that $\alpha(g)$ is an isometry when restricted to each $G$-orbit. Diffeomorphisms preserving a splitting with the above properties are called \emph{center isometries} and will be our main focus. See the next section for the precise definition, or the classical reference \cite{HPS} for further information about these systems.
Anosov flows (cf.\@ \cite{AnosovThesis}) constitute the most well understood example of Anosov action, where an analogous theorem as the previous one holds \cite{BowenBernoulli}. The proof relies on the existence of Markov partitions for such flows \cite{SymbHyp}; since this technology is unavailable for Anosov actions of rank $\geq 2$, we present a geometric construction of equilibrium states for such maps. At this stage however we need to impose some condition on the potential (being constant along the orbits, or being the geometrical potential associated to a regular element). Besides the fact that these cases include the most interesting types of potentials (corresponding to the entropy maximizing measure and the SRB), we also believe that the techniques developed are interesting on their own, and may be applicable to more situations. We remark that other geometrical constructions exist in the literature \cite{Climenhaga2020}, and the case of compact extensions of Anosov has been treated previously \cite{Spatzier2016}: in both cases restrictions on the potential are imposed. The manuscript \cite{Buzzi2019} contains some results similar to ours for perturbations of time-one maps of Anosov flows, for the case of the entropy maximizing measure.
For a center isometry $f$ the associated bundles $E^s,E^u,E^c$ are integrable to continuous foliations $\Fs,\Fu,\Fc$\ called the stable, unstable and center foliations. Our results require the assumption that $\Fs,\Fu$ are minimal foliations, i.e.\@ every one of their leaves is dense (this condition is implicit in the classical rank-one case cited above).
\begin{theoremA}
Let $f:M\rightarrow M$ be a center isometry of class $\mathcal{C}^2$ and $\varphi:M\rightarrow \Real$ a H\"older potential that is either
\begin{enumerate}
\item constant on center leaves, or
\item $\varphi=-\log \det Df|E^u$ (SRB case).
\end{enumerate}
Assume that the stable and unstable foliations of $f$ are minimal. Then there exist $\muf \in \PTM{f}{M}$ and families of measures $\mu^u=\{\mux\}_{x\in M},\mu^s=\{\msx\}_{x\in M},\mu^{cu}=\{\mcux\}_{x\in M},\mu^{cs}=\{\mcsx\}_{x\in M}$ satisfying the following.
\begin{enumerate}
\item The probability $\muf$ is the unique equilibrium state for the potential $\varphi$.
\item For every $x\in M$ the measure $\mu^{\ast},\ast\in \{u,s,cu,cs\}$ is a Radon measure on $W^{\ast}(x)$ which is positive on relatively open sets, and $y\in W^{\ast}(x)$ implies $\mu^{\ast}_x=\mu^{\ast}_y$.
\item It holds for every $x\in M$
\begin{enumerate}
\item $\mu^{\sigma}_{fx}=e^{\Ptop(\varphi)-\varphi}f_{\ast}\mu^{\sigma}_{x}\quad \sigma\in\{u,cu\}$.
\item $\mu^{\sigma}_{fx}=e^{\varphi-\Ptop(\varphi)}f_{\ast}\mu^{\sigma}_{x}\quad\sigma\in\{s,cs\}$.
\end{enumerate}
\item If $\xi$ is a measurable partition that refines the partition by unstable (stable) leaves then the conditionals $(\muf)^{\xi}_x$ of $\muf$ are equivalent to $\mux$ (resp. $\msx)$ for $\muf\aep(x)$.
\item For every $\epsilon>0$ sufficiently small, for every $x\in M$ the measure $\muf|D(x,\epsilon)$ has product structure with respect to the pair $\mux,\mcsx$, i.e.\@ its equivalent to $\mux\times \mcsx$.
\item There exists $K\geq 0$ only depending on $\varphi$ such that for every $\epsilon>0$ there exist some positive constants $a(\epsilon),b(\epsilon)$ satisfying, for every $x\in M,n\geq 0$
\[
a(\epsilon)e^{-Kn}\leq\frac{\muf(D(x,\epsilon,n))}{e^{\SB(x)-n\Ptop(\varphi)}}\leq e^{Kn}b(\epsilon).
\]
If the potential is constant along centers then one can take $K=0$.
\end{enumerate}
\end{theoremA}
\begin{corollary*}
Let $\alpha:G\curvearrowright M$ be an Anosov action of class $\mathcal{C}^2$. Suppose that $g\in G$ is an Anosov element such that its associated stable and unstable foliations are minimal and denote $f=\alpha(g)$. Let $\varphi:M\rightarrow \Real$ be a H\"older potential that is either constant along orbits of $\alpha$, or the geometrical potential associated to $f$. Then there exist a unique equilibrium state $\muf \in\PTM{f}{M}$ for $\varphi$ and families of measures $\mu^u=\{\mux\}_{x\in M},\mu^s=\{\msx\}_{x\in M},\mu^{cu}=\{\mcux\}_{x\in M},\mu^{cs}=\{\mcsx\}_{x\in M}$ satisfying the $1-6$ of Theorem A.
\end{corollary*}
In the rank-one case the previous Theorem (Corollary) was established first by G. Margulis \cite{TesisMarg} for the entropy maximizing measure ($\varphi\equiv 0$), and later by R. Bowen and D. Ruelle \cite{BowenRuelle} for general potentials, except for $2$ which appears to be new in this generality; the local product structure is due to N. Haydn \cite{localproductstructureHaydn} (see also \cite{Leplaideur2000}). The last item above is a generalization of the so called Gibbs-property of the equilibrium states. See \cite{EquSta} and \cite{Haydn1992}.
For the particular case of SRB measures the construction of the equilibrium state in our context also follows from the general results in \cite{GibbsSinaiPesin} while the uniqueness is proven by Dolgopyat in \cite{lecturesugibbs}, but the other parts of the theorem are new; still, we give a different proof of existence and uniqueness for SRB measures. In this case there exists the following characterization: an invariant measure $\mathrm{m}$ is an equilibrium state for $\varphi(x)=\log\det Df|E^u_x$ if and only if for every measurable partition subordinated to the unstable foliation, the conditional measures induced by $\mathrm{m}$ are absolutely continuous with respect to Lebesgue on the corresponding leaf. Here we are able to give similar characterization for equilibrium states, that seems to be new even in the classical hyperbolic case (the condition of the potential being constant on center leaves is automatic for Anosov diffeomorphisms).
\begin{theoremB}
In the same context of Theorem A, consider an $f$-invariant measure $\mathrm{m}$. The following are equivalent.
\begin{enumerate}
\item $\mu$ is an equilibrium state for $\varphi$.
\item For every $\mathrm{m}$-measurable partition $\xi$ subordinated to $\Fu$, the induced conditional measures are absolutely continuous with respect to $\{\mux\}_x$.
\item For every $\mathrm{m}$-measurable partition $\xi$ subordinated to $\Fs$, the induced conditional measures are absolutely continuous with respect to $\{\msx\}_x$.
\end{enumerate}
\end{theoremB}
The last part is somewhat surprising in the SRB case, as in general $\msx$ will not be absolutely continuous with respect to Lebesgue.
\smallskip
We are also able to study metric properties of the equilibrium state $\muf$, namely we establish the following.
\begin{theoremC}
Consider the equilibrium state $\muf$ given in Theorem A. Then the system $(f,\muf)$ is metrically isomorphic to a Bernoulli shift.
\end{theoremC}
\smallskip
It is natural to inquiry about the requirements imposed to the potentials. It turns out that in general we cannot guarantee the existence of families of measures satisfying part $2$ of Theorem A.
\begin{theoremD}
There exists an analytic center isometry $f:\Tor^3\to\Tor^3$ and a $\mathcal{C}^{\oo}$ potential $\varphi:\Tor^3\to\Real$ for which there is no family of measures $\{\zeta_x^u\}_{x\in M}$ satisfying
\begin{itemize}
\item $f^{-1}_{\ast}\zeta_{fx}^u=e^{\varphi - \Ptop(\varphi)}\zeta_x^u$;
\item the measures depend continuously on the point $x$.
\end{itemize}
\end{theoremD}
The methods developed in this article are of geometric flavor, and is the hope of the authors that they can shed on light in the still under development theory of thermodynamic formalism for partially hyperbolic systems. Evidence of their efficacy appears in a companion article \cite{ContributionsErgodictheory}, where we establish new results in the ergodic theory of the classical rank-one case.
The rest of the paper is organized as follows. In \cref{sec:measures_along_foliations} we develop general arguments to establish the existence of families of measures on foliations with prescribed invariant properties (conformality). These results are then used in \cref{sec:measures_along_invariant_foliations_of_center_isometries} to construct the equilibrium state for a center isometry as in Theorem A; in this part we also discuss uniqueness. Then in \cref{sec:a_characterization_of_equilibrium_states} we tackle the problem of the characterization of equilibrium states in terms of their conditional measures on stable/unstable leaves. \Cref{sec:finer_ergodic_properties} is devoted to the proof the Bernoulli property of the considered system, and in the last Section we include a discussion of the rank-one case, together with some questions. The proof of Theorem D is given in an Appendix, since it can be read almost independently of the rest of the work (although Theorems A and B are used). The reader is welcomed to check directly this part after \cref{sec:measures_along_invariant_foliations_of_center_isometries}.
\section{Measures along foliations}
\label{sec:measures_along_foliations}
For $r\geq 1$ we denote by $\diffM{M}{r}$ the group of $\mathcal{C}^r$ diffeomorphisms of $M$. Let $f\in \diffM{M}{r}$ and $\mu$ be an $f$-invariant measure. Recall that (cf. \cite{Rokhlin}) given a $\mu$-measurable partition $\parP$ of $M$ we can disintegrate $\mu$ in a family of conditional probabilities $\{\mu_x^{\parP}\}_{x\in M'}$ with $\mu(M')=1$ and such that for every $f\in \Lp[1](M,\mu)$ the conditional expectation of $f$ in $\widehat{P}$, the $\sigma$-algebra generated by $\parP$, is given by
\[
\ie{\mu}{f|\hat{\parP}}=\int f d\mu_x^{\parP},\quad x\in M'.
\]
Furthermore, if $\parP$ is countably generated (that is, $\parP=\bigvee_n \parP[P]_n$ with $\parP[P]_n$ finite partition for every $n\in\Nat$) then for $x,y\in M'$, if $y$ belongs to the atom of $\parP$ containing $x$ then $\mu^x=\mu^y$. In this case, denote by $M_{\parP}=M/\parP$ the space of atoms of $\parP$ and consider $\pi_{\parP}: M\rightarrow M_{\parP}$, $\pi_{\parP}(x)=\parP(x)$, the atom that contains $x$ (this map is well defined $\mu$-a.e.). We equip $M_{\parP}$ with the $\sigma$-algebra induced by $\pi_{\parP}$ and consider the \emph{transverse measure} $\mu_{\parP}=(\pi_{\parP})_{\ast}\mu$. The considerations above imply in particular that if $\parP$ is countably generated, then $\displaystyle{\eta=\parP(x)\mapsto \mu^{\eta}:=\mu_x^{\parP}}$ defines a measurable function from $M_{\parP}$ to the set of Borel probabilities on $M$, and furthermore
\[
A\in\BM\Rightarrow \mu(A)= \int_{M_{\parP}}\mu^{\eta}(A)\mu_{\parP}(d\eta).
\]
This shows that $\mu$ is completely determined provided that we know $\mu_{\parP}$ and $\{\mu^{\eta}\}_{\eta\in M_{\parP}}$.
Our interest however is constructing a particular $f$-invariant measure $\mu$ (the equilibrium state); for this we will argue backwards and construct first the transverse measure and the conditionals. Of course this presents several technical difficulties, mainly because we do not have the partition $\parP$, nor there is a well defined notion of what it means to be $\mu$-measurable.
A way to circumvent this problem is using some (continuous) invariant structure of $f$ to define the partition. In the next part we present a general construction that allow us to build families of measures satisfying certain (quasi-)invariant properties.
\subsection{Families of quasi-invariant measures}
\label{sub:families_of_quasi_invariant_measures}
Let $M$ be a locally compact metric space and $f:M\to M$ a bi-measurable map. From now on, by a measure on $M$ we mean a Radon measure, and denote the set of such measures by $\RM$. The set of continuous functions with compact support on $M$ is denoted by $\CMc[M]$. The set $\CMc[M]$ is equipped with the uniform topology, while $\RM$ is equipped with the vague one.
\smallskip
\noindent\textbf{Notation:} If $\phi$ is measurable we write $f\phi:=\phi\circ f$. If $\mu\in\RM$ the measure $f\mu=f_{\ast}\mu$ is the unique measure $\nu$ such that
\[
\forall \phi\in\CMc[M],\quad \int f\phi \der\mu =\int \phi \der\nu
\]
\smallskip
With the above it is direct to verify that for $\mu,\nu\in\RM$, $\mu=\rho\nu$ it holds $f\mu=f^{-1}\rho f\nu$. Recall that $\mu\in\RM$ is quasi-invariant if $f^{-1}\mu \sim \mu$; in this case we can write $f^{-1}\mu=\mathtt{h}_{\mu} \mu$ where $\mathtt{h}_{\mu}=\frac{d f^{-1}\mu}{d \mu}$. Similarly, $f^{-n}\mu=\mathtt{h}_{\mu}^{(n)} \mu$.
As it can be deduced directly from the uniqueness of the Radon-Nikodym derivative, the family $\{\mathtt{h}_{\mu}^{(n)}\}_{n\in \Z}$ is a multiplicative cocycle over $f$ with generator $\mathtt{h}_{\mu}$.
\paragraph*{Multiplicative cocycles} A family $\{h^{(n)}:M\to \Real_{\geq 0}\}_{n\in\Z}$ is a multiplicative cocycle over $f$ if
\[
\forall n,m \in \Z, x\in M\Rightarrow h^{(n+m)}(x)=h^{(n)}(f^mx)\cdot h^{(m)}(x).
\]
Given $h:M\to \Real_{>0}$, it generates a multiplicative cocycle over $f$ (and in this case we call $h$ the generator of the cocycle) by defining
\[
h^{(n)}(x)=\begin{dcases}
\prod_{k=0}^{n-1}h(f^kx)& k>0\\
x & k=0\\
\prod_{k=n}^{-1} h^{-1}(f^kx) &k<0.
\end{dcases}
\]
Observe that if $\{h^{(n)}\}_{n\in\Z}$ is a positive multiplicative cocycle over $f$ (with generator $h$) then $\{a^{(n)}:=\log h^{(n)}\}_{n\in\Z}$ is an additive cocycle (with generator $\log h$):
\[
\forall n,m \in \Z, x\in M\Rightarrow a^{(n+m)}(x)=a^{(n)}(f^mx)+a^{(m)}(x).
\]
\smallskip
It will be useful to introduce the following.
\begin{definition}
Let $\Lambda$ be a set. We say that $M$ is a bundle over $\Lambda$ if can be written as
\[
M=\bigsqcup_{i\in\Lambda} M_i
\]
where $M_i\subset M$ for every $i\in \Lambda$. We say that $f:M\to M$ is a bundle isomorphism if there exists a bijection $q:\Lambda\to \Lambda$ such that $f|:M_i\to M_{q(i)}$ is an homeomorphism, for every $i\in \Lambda$.
\end{definition}
For example, if $\F=\{W(x)\}_{x\in M}$ is a foliation on $M$ then we can write $M=\bigsqcup_{L\in \Lambda} L$ where
$\Lambda=M/\F$, and thus $M$ is a bundle over $\Lambda$. If furthermore $\F$ is $f$-invariant, then $f$ induces a bijection $q:\Lambda\to \Lambda, q(L)=f(L)$, and thus $f$ is a bundle isomorphism.
\smallskip
Consider the bundles over $\Lambda$ given as
\begin{align*}
&\mathrm{Con}:=\bigsqcup_{i\in\Lambda} \CMc[M_i]\\
&\mathrm{Rad}:=\bigsqcup_{i\in\Lambda} \RM[M_i]
\end{align*}
and consider also the space of sections of $\mathrm{Rad}$,
\[
\mathrm{Meas}:=\{\nu:\Lambda\rightarrow \mathrm{Rad}:\nu_i:=\nu(i)\in \RM[M_i]\forall\ i\in\Lambda\}.
\]
The natural topology of $\mathrm{Meas}$ is the weak topology determined by $\mathrm{Con}$; it is a locally convex topology. Let $\mathrm{Con}^+:=\{\phi\in\mathrm{Con},\phi\geq 0, \phi\not\equiv 0\}$.
\begin{lemma}\label{lem:Rproperty}
Suppose that $\mathcal{A}\subset \mathrm{Meas}$ satisfies
\begin{enumerate}
\item for every $\psi\in \mathrm{Con}$ there exists $c(\psi)>0$ with the property that for every $\mu\in\mathcal{A}, \mu(\psi)\leq c(\psi)$.
\item For every $\psi \in \mathrm{Con}^+$ there exists $d(\psi)>0$ with the property that for every $\mu\in\mathcal{A}, \mu(\psi)\geq d(\psi)$.
\end{enumerate}
Then $\clo{\mathcal{A}}\subset \mathrm{Meas}$ is compact, and furthermore it does not contain the zero section.
\end{lemma}
\begin{proof}
This is a direct consequence of Tychonoff's theorem.
\end{proof}
From now we assume that $f$ is a bundle isomorphism.
\begin{definition}
$\nu\in\mathrm{Meas}$ is quasi-invariant if for every $i\in\Lambda$, $f^{-1}\nu_{qi}\sim \nu_{i}$ with positive continuous Radon-Nikodym derivative $\rho_i=\frac{d f^{-1}\nu_{qi}}{d\nu_i}$. In this case $\rho:M\to \Real_{>0}$ given by $\rho|M_i=\rho_i$ will be called the Jacobian of $\nu$.
\end{definition}
One verifies directly that if $\nu$ is quasi-invariant then $\rho$ generates a (multiplicative) cocycle over $q$, and we write $\tilde{\rho}=\log \rho$ the corresponding additive cocycle. \label{construcciondeseccion}
Now suppose that $\varphi:M\to \Real$ is a continuous function and $\nu$ is quasi-invariant: we are interested in finding $\mu\in\mathrm{Meas}$ quasi-invariant with Jacobian $e^{-\varphi}$. Consider $\nu^n=e^{\SB}\cdot f^{-n}\nu $, i.e.
\begin{equation}\label{eq:nun}
\nu^n_i=e^{\SB}f^{-n}\nu_{q^n(i)}=e^{\SB+S_n\tilde{\rho}}\nu_i,
\end{equation}
and let $\mathcal{C}^{+}$ be the positive cone generated by $\{\nu^n\}_{n\geq 0}$,
\begin{equation}
\mathcal{C}^{+}=\Set{\sum_{i=1}^ka_i\nu^{n_i}:a_i\geq 0,k\in\mathbb{N}}.
\end{equation}
\begin{theorem}\label{thm:existenciafamiliamedidas}
Assume that $\{\nu^n\}_{n\geq 0}$ satisfies:
\begin{itemize}
\item if $\phi\in\mathrm{Con}^+$, then $\nu^n(\phi)>0$ for all $n$.
\item Given $\phi, \psi \in \mathrm{Con}$ with $\phi\not\equiv 0$ non-negative, there exists $e(\phi,\psi)>0$ such that for every $n\geq 0$ it holds
\[
\frac{\nu^n(\psi)}{\nu^n(\phi)}\leq e(\phi,\psi).
\]
\end{itemize}
Then there exists $P\in\Real$ and a quasi-invariant $\mu \in\cl{\mathcal{C}^{+}}$ with Jacobian $e^{P-\varphi}$.
\end{theorem}
\begin{proof}
Define $S:\mathrm{Meas}\rightarrow \mathrm{Meas}$ by
\[
S(\mu_i)=e^{\varphi}\cdot f^{-1}\mu_{q(i)}.
\]
Then $S$ is continuous and sends $\mathcal{C}^{+}$ to itself. We compute
\begin{equation}\label{eq:Snun}
S(\nu^n_i)=e^{\varphi}\cdot f^{-1}\left(e^{\SB}\nu_{q(i)}^n\right)= e^{\varphi}\cdot f^{-1}\left(e^{\SB}f^{-n}\nu_{q^{n+1}(i)}\right)=e^{\varphi+\SB\circ f}\cdot f^{-n+1}\nu_{q^{n+1}(i)}=\nu^{n+1}.
\end{equation}
Fix a non-negative, non-identically zero function $\phi_0\in\mathrm{Con}$ and let $\hat{\nu}^n:=\frac{\nu^n}{\nu^n(\phi_0)}$; consider the set
\begin{align}\label{eq:combinacionesconvexas}
\mathcal{X}&=\clo{\Set{\sum_{i=1}^ka_i\hat{\nu}^{n_i}:a_i\geq 0,k\in\mathbb{N},\sum_{i=1}^ka_i=1}}\\
&=\clo{\Set{\nu\in\mathcal{C}^+:\nu(\phi_0)=1}}.
\end{align}
By \cref{lem:Rproperty} and the assumed hypotheses, $\mathcal{X}$ is a compact convex subset of $\mathrm{Meas}$, and by \cref{eq:Snun} the normalized operator \(\tilde{S}\nu:=\frac{S\nu}{S\nu(\psi_0)}\) sends $\mathcal{X}$ to itself. Hence, by the Schauder-Tychonoff fix point theorem, there exist $\mu\in\mathcal{X}$ such that $\widetilde{S}(\mu)=\mu$. In other words, $\psi\in \mathrm{Con}$ implies
\begin{equation}\label{eq.margulis}
\mu_{q(i)}(\psi\circ f^{-1}e^{\varphi\circ f^{-1}})=e^P \cdot \mu_{i}(\psi)\quad (\Rightarrow f^{-1}\mu_{q(i)}=e^{P-\varphi}\cdot \mu_i),
\end{equation}
where $e^P=S\mu(\psi_0)>0$.
\end{proof}
Observe that in the theorem above, if $\psi\in \mathrm{Con}$ is non-negative, non-identically zero then for some constant $c(\psi)>0$ it holds $\hat{\nu}^n(\psi)>0$ for all $n\geq 0$, and thus the same is valid for $\mu$. We deduce:
\begin{corollary}
In the same hypotheses of the Theorem above, the section $\mu$ has full support on each $M_i$.
\end{corollary}
\subsection{Quasi-invariant measures along leaves of foliations: the transverse measure}
\label{sub:quasi_invariant_measures_along_leaves_of_foliations}
We will now apply the construction of the previous part to the following setting.
\begin{enumerate}
\item $M$ is a compact (closed) smooth Riemannian manifold and $\F=\{F(x)\}_{x\in M}$ is continuous foliations with smooth leaves. We consider the bundle structure of $M=\bigsqcup_{F\in \Lambda} F$ over $\Lambda=M/\F$, and denote $\mathrm{Con}(\F),\mathrm{Meas}(\F)$ the corresponding associated bundles.
\item $f$ is a diffeomorphism of $M$ of differentiability class $\mathcal{C}^2$.
\item $\F$ is $f$-invariant; the induced map by $f$ on $\Lambda$ is denoted by $q$.
\item There exists a complementary continuous foliation $\mathcal{G}=\{G(x)\}_{x\in M}$ of $\F$ with smooth leaves, and such that
\begin{enumerate}
\item $\mathcal{G}$ is minimal: each leaf $G(x)$ is dense.
\item $\mathcal{G}$ is $f$-invariant.
\item $\mathcal{G}$ is contracting under $f$: if $d_{\mathcal{G}}$ denotes the intrinsic distance (on the corresponding leaf of $\mathcal{G}$) we have
\[
\forall x,y\in M, y\in G(x)\Rightarrow d_{\mathcal{G}}(fx,fy)\leq \lambda d_{\mathcal{G}}(x,y).
\]
for some $0<\lam<1$.
\end{enumerate}
\end{enumerate}
The idea is that if $B$ is a foliated box corresponding to $\F$, we will use the (measurable) partition $\F|B$ to play the role of the quotient space where we define our transverse measure. The existence of a complementary contracting foliation is used (crucially) to ensure the compactness property necessary for \cref{lem:Rproperty}.
\smallskip
\noindent\textbf{Notation:} If $A\subset M, \ep>0$ let
\[
G(A,\ep):=\{y\in M:\exists x\in A/ d_{\mathcal{G}}(x,y)<\ep\},
\]
and if $A=\{x\}$ we write $G(x,\ep)=G(\{x\},\ep)$.
\begin{definition}Let $\delta>0$.
\begin{enumerate}
\item Two relatively compact sets $A_1 \subset F_{i_1} ,A_2 \subset F_{i_2}$ are $\delta$-equivalent if there exists an holonomy transport (Poincar\'e map) $\hol[\mathcal{G}]:F_{i_1}\to F_{i_2}$ induced by $\mathcal{G}$ satisfying:
\begin{enumerate}
\item $\hol[\mathcal{G}]|A_1$ is an homeomorphism of $A_1$ onto $A_2$.
\item For every $x\in A_1,\hol[\mathcal{G}](x)\in G(x,\delta)$.
\end{enumerate}
\item Two functions $\psi_1,\psi_2\in \mathrm{Con}(\F)$ are $\delta$-equivalent if $\supp(\psi_1)$ is $\delta$-equivalent to $\supp(\psi_2)$ and for every $x\in \supp(\psi_1)$, $\psi_2(\hol[G](x))=\psi_1(x)$. Here $\hol[G]$ is the holonomy map connecting $\supp(\psi_1)$ with $\supp(\psi_2)$.
\end{enumerate}
\end{definition}
Observe that given $\psi \in \mathrm{Con}(\F),\delta>0$ there exists $\gamma=\gamma(\psi,\delta)>0$ such that if
$x\in G(\supp(\psi),\gamma)$ then
\begin{itemize}
\item the transverse foliation defines an homeomorphism onto its image $\hol[\mathcal{G}]|:\supp(\psi)\to F(x)$.
\item $\psi_{F(x)}:=\psi\circ \hol[\mathcal{G}]$ is continuous, of compact support, and $\delta$-equivalent to $\psi$.
\end{itemize}
We seek to find a quasi-invariant $\nu\in\mathrm{Meas}(\F)$ with regular Radon-Nikodym derivative. A natural class is obtained by fixing a Riemannian metric on $M$ and considering $\Leb\in \mathrm{Meas}(\F)$ given by
\[
F\in\Lambda\Rightarrow \Leb_F=\text{ induced Lebesgue measure on the leaf }F.
\]
By the basic change of variables theorem, $\Leb$ is quasi-invariant its Jacobian is $\rho_F=\log|\det D(f|F)|$. By compactness of $M$ we have that
\[
\rho\text{ is uniformly bounded from below: }\rho(x)\geq C_{\rho}>0\ \forall x\in M,
\]
and since $f\in \diffM{M}{1+\alpha}$ we get that $\rho$ (and therefore, $\tilde{\rho}=\log\rho$) is $\alpha$-H\"older.
We will use the minimality of $\mathcal{G}$ to compare sets (functions) defined on different leaves of $\F$. It is necessary however to control how families of measures $\nu=\{\nu_F\}_{F\in\F}$ vary with the transverse holonomy given by $\mathcal{G}$.
\begin{definition}
$\nu\in\mathrm{Meas}(\F)$ is absolutely continuous if for every (locally defined) holonomy transport $\hol[\mathcal{G}]:F_{i_1}\to F_{i_2}$, $\hol[\mathcal{G}]\nu_{F_{i_1}}\sim \nu_{F_{i_2}}$. It is strongly absolutely continuous if there exists $\delta_0>0$ and $J:\{(x,y,z)\in M\times M\times M, y\in G(x,\delta_0), z\in F(y)\}\to \Real$ continuous such that if $A\subset F(x),B\subset F(y)$ are $\delta_0$-equivalent then
\[
\hol[\mathcal{G}]\nu_{F(x)}=J(x,y,\cdot)\nu_{F(y)}
\]
\end{definition}
\begin{proposition}\label{pro:lebesgueesabscont}
$\Leb$ is strongly absolutely continuous.
\end{proposition}
\begin{proof}
This result is classical by now, and relies on the contracting behavior of $\mathcal{G}$ together with the fact of $f$ being of class $\mathcal{C}^{1+\alpha}, \alpha>0$. See \cite{ErgAnAc}.
\end{proof}
For strongly absolutely continuous sections we have the following.
\begin{proposition}\label{pro:MargulisAbsCont}
Assume that $\nu$ is strongly absolutely continuous and of full support on every $F\in\F$. Then given $\ep>0$ there exists $\delta>0$ such that for every $\psi_1,\psi_2\in \mathrm{Con}(\F)$ that are $\delta$-equivalent and non-identically zero, it holds
\[
\left|\int \psi_1 d\nu_{F_{i_1}}-\int \psi_2 d\nu_{F_{i_2}}\right|<\delta \int \abs{\psi_1} d\nu_{F_{i_1}}.
\]
\end{proposition}
\begin{proof}
We start by noting that since $\nu$ is strongly absolutely continuous, given $\ep>0$ there exists $\delta$ such that if $A,B$ are $\delta$-equivalent (with non-trivial measure) then
\[
\big|\frac{\nu(A)}{\nu(B)}-1\big|<\ep.
\]
The remaining part of the argument is spelled in \cite{TesisMarg}; use the previous remark as Lemma $2.1$ and
argue as in Lemma $2.2$, Proposition $2.2$, Lemma $2.5$ and finally Lemma $2.6$, which has the same content as our Proposition.
\end{proof}
\begin{remark}
The core of our construction in this part is due to G. Margulis (above-cited) who considered the case where the potential $\varphi\equiv 0$ and $\nu=\Leb$. Parts which follow (with minimal) mutatis-mutandis will be directed to Margulis' presentation.
\end{remark}
We now complete our construction. Fix $\nu\in \mathrm{Meas}(\F)$ that satisfies:
\begin{itemize}
\item for every leaf $F\in \F$, $\nu_F$ has full support,
\item is strongly absolutely continuous,
\item is quasi-invariant with H\"older Jacobian $\rho$,
\end{itemize}
and consider $\varphi:M\to\Real$ a H\"older continuous potential. Let $\hat{\rho}=\log \rho$ and $\hat{\varphi}=\varphi+\hat{\rho}$; this is a $(C_{\hat{\varphi}},\theta)$-H\"older function. For $n\in \Nat$ take $\nu^{n}=e^{\SB}\cdot f^{-n}\nu=e^{S_n\hat{\varphi}}\nu$ (cf. \cref{eq:nun}).
\begin{lemma}\label{lem:controlnun}
There exists a constant $D_1>0$ and $\ell:\Real_{>0}\to\Real_{>0}$ with the following property. For every $\delta>0$, if $\psi_1,\psi_2\in \mathrm{Con}^+$ are $\delta$-equivalent and $n\geq0$, then it holds
\[
\nu^n(\psi_1)\leq \ell(\delta)\cdot e^{D_1\delta^{\theta}}\nu^n(\psi_2).
\]
Furthermore, $\ell(\delta)\xrightarrow[\delta\to 0]{}1$.
\end{lemma}
\begin{proof}
Define $D_1:=C_{\hat{\varphi}}\sum_{i=0}^{\infty} \lambda^{i\theta}$; since $\lambda<1$, $D<+\oo$. Assume that $\psi_1,\psi_2$ are $\delta$-equivalent, $\psi_2\circ \hol[\mathcal{G}] =\psi_1$. Observe that for every $x\in \supp(\psi_1)$, for every $n\geq 0$, it holds
\[
\abs{\SB[\hat{\varphi}](\hol[\mathcal{G}] x)-\SB[\hat{\varphi}](x)}\leq \sum_{i=0}^{\oo} C_{\hat{\varphi}}\cdot d_{\mathcal{G}}(f^i\hol[\mathcal{G}] x,f^ix)^{\theta}\leq D_1\delta^{\theta}.
\]
It then follows that
\begin{align*}
\nu^n(\psi_1)&=\int e^{S_n\hat{\varphi}} \psi_2\circ \hol[\mathcal{G}] d\nu=\int e^{S_n\hat{\varphi}-S_n\hat{\varphi}\circ \hol[\mathcal{G}]} \psi_2\circ \hol[\mathcal{G}] e^{S_n\hat{\varphi}\circ \hol[\mathcal{G}]} d\nu\\
&\leq e^{D_1\delta^{\theta}} \int \psi_2 e^{S_n\hat{\varphi}} d\hol[\mathcal{G}]\nu\leq \ell(\delta)e^{D_1\delta^{\theta}} \nu^n(\psi_2)
\end{align*}
where $\ell$ is an upper bound of $\frac{d \hol[\mathcal{G}]\nu}{d\nu}$, and therefore converges to $1$ as $\delta\to 0$.
\end{proof}
\begin{lemma}
Let $A_1\subset F_{i_1},A_2\subset F_{i_2}$ be relatively open and pre-compact. Then there exist $\hat{e}(A_1,A_2)>0$ such that for every $n\geq 0$ it holds
\[
\frac{1}{\hat{e}(A_1,A_2)}\leq \frac{\nu^n(A_1)}{\nu^n(A_2)}\leq \hat{e}(A_1,A_2).
\]
\end{lemma}
\begin{proof}
Let us start noting that since $\mathcal{G}$ is minimal (on $M$ compact) we have the following property: for any $A\subset F, F\in \F$ there exist $\delta(A),r(A)>0$ such that for every $x\in M$ one can find $B_x\subset A$ that is $\delta(A)$-equivalent to $F(x,r(A))$.
By relative compactness of $\clo{A_1}$, the closure of $A_1$ inside $F_{i_1}$, it follows that $\clo{A_1}\subset \cup_1^m F(x_j,r(A_2))$, where each plaque $F(x_j,r(A_2))$ is $\delta(A_2)$-equivalent to some $B_j\subset A_2$.
Fix two of these sets $E_1=F(x_j,r(A_2)), E_2=B_j$ and denote by $\hol[\mathcal{G}]$ the holonomy transport such that $\one_{E_1}=\one_{E_2}\circ \hol[\mathcal{G}]$. By regularity of $\nu_{F_{i_2}}$ we can find a non-decreasing sequence $(\phi_k)_k\subset \CMc[F_{i_2}]$ satisfying
\begin{enumerate}
\item $k\in\mathbb{N}\Rightarrow 0\leq \phi_k\leq 1, \supp(\phi_k)\subset \clo{A_2}$.
\item $\lim_{k\mapsto\oo}\phi_k(x)=\one_{A_2}(x)$.
\item $\nu^n(A)=\lim_k\uparrow\nu^n(\phi_k)$.
\end{enumerate}
Define $\psi_k=\phi_k\circ \hol[\mathcal{G}]$ and apply the previous lemma to deduce that for every $n$,
\begin{align*}
\frac{\nu^n(E_1)}{\nu^n(E_2)}=\frac{\lim_k \nu^n(\psi_k)}{\sup_k \nu^n(\phi_k)}\leq \sup_k \frac{\psi_k}{\phi_k}\leq \ell(\delta(A_2))e^{D_1\delta(A_2)^{\theta}}:=e_0.
\end{align*}
Therefore,
\begin{align*}
\frac{\nu^n(A_1)}{\nu^n(A_2)}\leq \frac{\nu^n(F(x_j,r(A_2))}{\nu^n(A_2)}
\leq m\cdot\max_{j} \frac{\nu^n(F(x_j,r(A_2)))}{\nu^n(B_j)}\leq m\cdot e_0.
\end{align*}
From here the conclusion follows.
\end{proof}
\begin{corollary}\label{cor:nucompacidad}
Fix $\psi\in \mathrm{Con}^+(\F)$. Then for every $\phi\in \mathrm{Con}(\F)$ there exists $\hat{e}(\phi,\psi)>0$ such that for every $n\geq 0$ it holds
\[
\frac{\nu^n(\phi)}{\nu^n(\psi)}\leq \hat{e}(\phi,\varphi).
\]
\end{corollary}
\begin{proof}
This is essentially Lemma $2.4$ of \cite{TesisMarg}. For $r>0$ let $A_r:=\psi^{-1}(r,\oo)$ and note that since $\psi\in \mathrm{Con}$ is non-negative and non-identically zero, there exists $r>0$ such that $A_r$ is relatively open and pre-compact inside the leaf of $\F$ that contains $\supp(\psi)$. Take $A$ open, relatively compact containing $\supp(\phi)$ and use the previous Lemma to deduce
\[
\frac{\nu^n(\phi)}{\nu^n(\psi)}<\frac{\norm{\phi}_{\oo}\cdot\nu^n(A)}{r\cdot \nu^n(A_r)}<\frac{\norm{\phi}_{\oo}}{r}\ \hat{e}(A,A_r).
\]
\end{proof}
We have thus shown that the in the assumed conditions, the hypotheses of \cref{thm:existenciafamiliamedidas} are satisfied, and thus we conclude the existence of $P\in\Real$ and $\mu\in \mathrm{Meas}(\F)$ of full support on each leaf of $\F$, that is quasi-invariant with Jacobian $e^{P-\varphi}$.
\paragraph{Strong absolute continuity of the quasi-invariant section}
\label{par:strong_absolute_continuity_of_the_quasi_invariant_section}
It turns out that the quasi-invariance of $\mu$ implies that is strongly absolutely continuous. We keep the hypotheses of the previous part and denote for $x\in M$, $\mu_x=\mu_{F(x)}$. By quasi-invariance we get
\begin{equation}
\forall n\in\Nat,\quad f^{-n}\mu_{f^nx}=e^{nP-\SB}\mu_x.
\end{equation}
\begin{lemma}
Given $\ep>0$ there exists $\delta>0$ such that for every $\psi_1,\psi_2\in\mathrm{Con}^+$ that are $\delta$-equivalent, then
\[
\left|\frac{\mu(\psi_1)}{\mu(\psi_2)}-1\right|<\ep.
\]
\end{lemma}
\begin{proof}
By \cref{lem:controlnun}, for every $n\in\Nat$ we have $\hat{\nu}^n(\psi_1)\leq \ell(\delta)\cdot e^{D_1\delta^{\theta}}\hat{\nu}^n(\psi_2)$; the inequality extends to convex combinations of $\{\hat{\nu}^n\}_n$ and therefore, to elements of $\mathcal{X}$ (cf.\@ \cref{eq:combinacionesconvexas}). In particular $\mu$ satisfies the claim.
\end{proof}
We now fix $x_0, y_0$ in the same leaf of $\mathcal{G}$ and consider $\hol[\mathcal{G}]:A(x_0)\subset F(x_0)\to B(y_0)\subset F(y_0)$ the corresponding Poincar\'e map. For $x\in A(x_0)$ define
\begin{equation}\label{eq:Jacobianohol}
\Jac(x)=\prod_{j=0}^{\oo}\frac{e^{\varphi\circ f^j(\hol[\mathcal{G}]x)}}{e^{\varphi\circ f^j(x)}}=\lim_n e^{\SB(\hol[\mathcal{G}]x)-\SB(x)}.
\end{equation}
Since $\varphi$ is H\"older and $\mathcal{G}$ is contracting, the previous formula defines a continuous function $\Jac:A(x_0)\rightarrow \Real$.
\begin{proposition}\label{pro:Jacobinanohol}
It holds $(\hol[\mathcal{G}])^{-1}\mu_{y_0}=\Jac \cdot \mu_{x_0}$.
\end{proposition}
\begin{proof}
Denote $h=\hol[\mathcal{G}]$, and for $n\geq 0$ let $h_n$ be the Poincar\'e map defined on a neighborhood of $f^nx$ inside $f^n(A(x_0))$. As $f$ preserves $\F,\mathcal{G}$, $h=f^{-n}\circ h_n\circ f^{n}$ and therefore
\begin{align*}
\MoveEqLeft h^{-1}\mu_{y_0}=(f^{-n}\circ h_n^{-1})e^{\SB\circ f^{-n}-nP}\mu_{f^ny_0}=f^{-n}\left(e^{\SB\circ h\circ f^{-n}-nP}h_n^{-1}\mu_{f^ny_0}\right)\\
&=f^{-n}\left(e^{\SB\circ h\circ f^{-n}-nP}\left(\mu_{f^nx_0}+\left(h_n^{-1}\mu_{f^ny_0}-\mu_{f^nx_0}\right)\right)\right)\\
&=e^{\SB\circ h-\SB}\mu_{x_0}+f^{-n}e^{\SB\circ h\circ f^{-n}-nP}\left(h_n^{-1}\mu_{f^ny_0}-\mu_{f^nx_0}\right).
\end{align*}
By the lemma above, given $\ep>0$ there exists $n_0$ such that for $n\geq n_0$ it holds: for any $\psi\in \mathrm{Con}^+$ supported in $f^n(A(x_0))$,
\[
\abs{h_n^{-1}\mu_{f^ny_0}(\psi)-\mu_{f^nx_0}(\psi)}<\ep h_n^{-1}\mu_{f^ny_0}(\psi)
\]
which in turn implies
\begin{align*}
\abs{f^{-n}e^{\SB\circ h\circ f^{-n}-nP}\left(h_n^{-1}\mu_{f^ny_0}-\mu_{f^nx_0}\right)(\psi\circ f^n)}&<\ep \mu_{f^ny_0}(\psi\circ h_n^{-1}e^{\SB\circ f^{-n}-nP})\\
&=\ep h^{-1}\mu_{y_0}(\psi\circ f^n).
\end{align*}
Thus we can write $h^{-1}\mu_{y_0}=J_n\cdot \mu_{x_0}+\upsilon_n$, where $\upsilon_n$ converges to zero in the vague topology, and $\{J_n\}_n$ is a sequence of continuous functions that converges uniformly to $\Jac$. This concludes the proof.
\end{proof}
\section{Invariant measures for center isometries}
\label{sec:measures_along_invariant_foliations_of_center_isometries}
We will use the previous construction in the case when $f$ is a center isometry. For convenience of the reader we recall the definition and some basic facts below.
\begin{definition}\label{def:centerisometry}
Let $M$\ be a closed manifold. A diffeomorphism $f\in \diffM{r}{M}, r\geq 1$ is a center isometry if there exist a continuous splitting of the tangent bundle of the form
\[TM=E^u\oplus E^c\oplus E^s
\]
where both bundles $E^s,E^u$\ are non-trivial, a (continuous) Riemannian metric $\norm{\cdot}$ on $M$ and
a constant $0<\lambda<1$ such that
\begin{enumerate}
\item All bundles $E^u,E^s,E^c$\ are $Df$-invariant.
\item For every $x\in M$, for every unit vector $v^{\ast}\in E^{\ast}_x, \ast=s,c,u$,
\begin{gather*}
\norm{D_xf(v^{c})}=1\\
\norm{D_xf^n(v^{s})},\norm{D_xf^{-1}(v^{u})}<\lambda.
\end{gather*}
\end{enumerate}
\end{definition}
Center isometries are a subclass of the so called \emph{Partially Hyperbolic Systems}. From their theory we need the following.
\begin{theorem}\label{thm:centerisometry}
If $f\in \diffM{M}{r}$\ is a center isometry then the bundles $E^{s}, E^u, E^c$, $E^{cs}=E^c\oplus E^s,E^{cu}=E^c\oplus E^u$ are (uniquely) integrable to continuous foliations $\Fs=\{W^s(m)\}_{x\in M}$, $\Fu=\{W^u(m)\}_{x\in M},\Fc=\{W^c(m)\}_{x\in M},\Fcs=\{W^{cs}(m)\}_{x\in M},\Fcu=\{W^{cu}(m)\}_{x\in M}$\ respectively called the \emph{stable},\emph{unstable},\emph{center}, \emph{center stable} and \emph{center unstable} foliations, whose leaves are $\mathcal{C}^r$\ immersed submanifolds. Moreover, leaves of $\Fcs$ are saturated by leaves of $\Fs,\Fc$, and leaves of $\Fcu$ are saturated by leaves of $\Fu,\Fc$.
\end{theorem}
\begin{proof}
Integrability of the bundles $E^s,E^u$ is consequence of the classical Stable Manifold theorem (see for example Theorem 4.1 in \cite{PesinLect}), while the rest of the assertions can be deduced from theorem 7.5 in \cite{PartSurv}.
\end{proof}
\noindent\textbf{Standing hypotheses for the rest of the section:} $f:M\rightarrow M$ is center isometry of class $\mathcal{C}^2$ and $\varphi:M\to\Real$ is a H\"older potential. The metric in $M$ is assumed to make $E^s,E^c,E^u$ mutually perpendicular.\footnote{This can always be achieved by passing to an equivalent metric.} Both foliations $\Fs,\Fu$ are minimal.
\smallskip
Applying the results of \cref{sec:measures_along_foliations} with $\F=\Fcu, \mathcal{G}=\Fs$ we get the following.
\begin{proposition}\label{pro:medidacu}
There exists a section $\mu^{cu}=\mu^{cu}_{\varphi}\in \mathrm{Meas}(\Fcu)$ and $P\in \Real$ such that:
\begin{enumerate}
\item For every $\psi\in \mathrm{Con}^+(\Fcu,)$ it holds $\mu^{cu}(\psi)>0$.
\item For every $x\in M$, $f^{-1}\mcux[fx]=e^{P-\varphi}\mcux[x]$.
\item If $\hs=\hs_{x_0,y_0}:A(x_0)\subset \Wcu{x_0}\rightarrow B(y_0)\subset\Wcu{y_0}$ is the Poincar\'{e} map determined by the stable holonomy that sends $x_0$ to $y_0$, then
\[
(\hs)^{-1}\mu^{cu}_{y_0}=\Jacs\cdot \mu^{cu}_{x_0}
\]
where
\begin{equation}
\Jacs(x)=\prod_{j=0}^{\oo}\frac{e^{\varphi\circ f^j(\hs x)}}{e^{\varphi\circ f^j(x)}}.
\end{equation}
\end{enumerate}
\end{proposition}
Note that $f^{-1}$ is also a center isometry whose stable foliation coincides with the unstable of $f$, and likewise its unstable foliation coincides with the stable of $f$. By applying the previous arguments to $f^{-1}$ and the potential $\varphi\circ f^{-1}$ we conclude the following.
\begin{proposition}\label{pro:medidacs}
There exist $\mu^{cs}=\mu^{cs}_{\varphi} \in \mathrm{Meas}(\Fcu)$ and $P'\in \Real$ such that:
\begin{enumerate}
\item For every $\psi\in \mathrm{Con}^+(\Fcs)$ it holds $\mu^{cs}(\psi)>0$.
\item For every $x\in M$, $f^{-1}\mcsx[fx]=e^{\varphi-P'}\mcsx[x]$.
\item If $\hu=\hu_{x_0,y_0}:A(x_0)\subset \Wcs{x_0}\rightarrow B(y_0)\subset\Wcs{y_0}$ is the Poincar\'{e} map determined by the stable holonomy that sends $x_0$ to $y_0$, then
\[
(\hu)^{-1}\mu^{cs}_{y_0}=\Jacu\cdot \mu^{cs}_{x_0}
\]
where
\begin{equation}
\Jacu(x)=\prod_{j=1}^{\oo}\frac{e^{\varphi\circ f^{-j}(\hu x)}}{e^{\varphi\circ f^{-j}(x)}}.
\end{equation}
\end{enumerate}
\end{proposition}
\subsection{The conditional measures: potentials constant along the center and SRBs}
\label{sub:the_conditional_measures_potentials_constant_along_the_center_and_srbs}
We are interested in constructing $\mu^u\in\mathrm{Meas}(\Fu)$ satisfying analogous properties as $\mu^{cu}$. Our previous construction however does not apply in this case since the transverse foliation to $\Fu$ is not contracting. To bypass this problem we assume some condition on $\varphi$, namely that it is either
\begin{itemize}
\item constant on center leaves ($c$-constant case), or
\item $\varphi=-\log\det Df|E^u$ (SRB case).
\end{itemize}
In the SRB case we define
\[
\mux=\text{Lebesgue measure on }\Wu{x};
\]
by the change of variables theorem $f^{-1}\mux[fx]=\det Df|E^u\cdot \mux=e^{-\varphi}\mux[x]$. For the $c$-constant case, note that by compactness of $M$ there exists some $\ep_0>0$ such that for every $x\in M$ there exists a well defined projection $\pi_x^c:\Wc{\Wu{x},\ep_0}$ by sliding along local center plaques. Given $x$ we define a measure on $\Wu{x}$ by setting
\[
\mux=\pi^c_x\mcux
\]
Note that $\mu\in \mathrm{Meas}(\Fu)$ is quasi-invariant with Jacobian $e^{P-\varphi}$; indeed by using that
$\varphi$ is constant on center leaves and that $f$ is an isometry on these, we deduce
\begin{align}\label{eq:invarianciainest}
f^{-1}\mux[fx]=f^{-1}\pi^c_{fx}\mcux[fx]=\pi^c_{x}f^{-1}\mcux[fx]=\pi^c_{x}e^{P-\varphi}\mcux[x]=e^{P-\varphi}\pi^c_x\mcux[x]=e^{P-\varphi}\mux[x].
\end{align}
An important case of $c$-constant potential is when $\varphi\equiv 0$; in this case we denote $m^{\ast}\in \mathrm{Meas}(\F^{\ast}),\ \ast\in\{s,u,cs,cu\}$ the corresponding sections.
\begin{definition}
$m^{\ast}$ are the Margulis measures on $\F^{\ast}$.
\end{definition}
It follows that
\begin{align}\label{eq:margulisinvariance}
&f^{-1}m^{cu}_{fx}=e^{h}m^{cu}_x\\
&f^{-1}m^{u}_{fx}=e^{h}m^{u}_x
\end{align}
for some $h\in\Real$.
\begin{proposition}\hfill
\begin{enumerate}
\item $m^{cu}$ ($m^{cs}$) is invariant under $\hs$ holonomy (respectively, $\hu$ holonomy).
\item The constant $h$ is positive.
\end{enumerate}
\end{proposition}
\begin{proof}
By part 3 of \cref{pro:medidacu} it holds $\Jacs\equiv1$, thus implying the first part.
Recall the construction of $m^{cu}$ on page \pageref{construcciondeseccion}, with $\nu=\Leb$ and $\nu^n=f^{-n}\nu$ and a fixed normalized $\psi_0\in\mathrm{Con}^+(\Fcu)$. Thus we have for every $n\geq 0$,
\begin{align*}
f^{-1}\nu^n(\psi_0)\geq \min \{\det Df|E^{cu}\}\nu^{n}(\psi_0)\Rightarrow f^{-1}\hat{\nu}^n(\psi_0)\geq \min \{\det Df|E^{cu}\}.
\end{align*}
This implies the same inequality for every member of $\mathcal{X}$, and in particular
\[
e^h=m^{cu}(\psi_0\circ f^{-1})\geq \min \{\det Df|E^{cu}\}>1.
\]
\end{proof}
\begin{remark}
The reader can perceive that in principle $h$ (and $P,P'$) depend on $\psi_0$. We will show in the next section that $h=h_{\mathrm{top}}(f)$ and $P=P'=P_{\mathrm{top}}(\varphi)$.
\end{remark}
We return to the general case; we will use the family $\{\mux\}_x$ to define some new measures $\{\nu^u_x\}_x$ that will be serve as conditional measures on local unstable manifolds for the equilibrium state.
For each $x\in M, y\in\Wu{x}$ let
\begin{equation}\label{eq:JacobianoDelta}
\Delta_x^u(y):=\prod_{k=1}^{\oo}\frac{e^{\varphi\circ f^{-k}(y)}}{e^{\varphi\circ f^{-k}(x)}},
\end{equation}
and note that $\Delta_x^u:\Wu{x}\to\Real$ is continuous. Define
\begin{equation}\label{eq.nux}
\nu^u_x=\Delta_x^u\ \mux.
\end{equation}
One verifies easily that $\nux$ is a Radon measure on $\Wu{x}$ which is positive on open sets, and furthermore
\begin{align}\label{eq:invariancianu}
f^{-1}\nux[fx]=\Delta_{fx}^u\circ f\cdot f^{-1}\mux[fx]=\Delta_{fx}^u\circ f \cdot e^{P-\varphi}\mux =e^{P-\varphi(x)}\nux.
\end{align}
For $y\in \Wu{x}$ we have $\nu^u_y=c(y,x)\cdot \nux$ with $c(y,x)>0$, hence $\{\nu^u_y\}_{y\in\Wu{x}}$ defines a projective class of measures $[\nu^u_x]$. In particular if $\xi$ is a $\mcux$-measurable partition subordinated to $\Fu$ we have that the probability measures
\begin{equation}\label{eq:condicionalxi}
\nu^{\xi}_x:=\nux(\ \cdot\ |\xi(x))
\end{equation}
satisfy $\nu^{\xi}_x=\nu^{\xi}_y$ for $y\in\xi(x)$.
Before moving on we establish the following lemma that will allow us to compare different unstable measures. Consider the compact space
\[
N=\{(x_0,y_0,z)\in M^3: y_0\in\clo{\Wu{x_0,\ep/2}}, z\in \clo{\Wc{x_0,\ep/2}}\}
\]
and for $(x_0,y_0,z)\in N$ define
\begin{equation}\label{eq.Jacobianouconholonomia}
T_{x_0,y_0}(z)= \prod_{k=1}^{+\oo}\frac{e^{\varphi\circ f^{-k}(\hu_{x_0,y_0})}}{e^{\varphi\circ f^{-k}(z)}}.
\end{equation}
\begin{lemma}\label{lem:JacobianoT}
It holds that:
\begin{enumerate}
\item the constant $c(y,x)$ converges uniformly to $1$ as $y\to x$, and
\item $T:N\to\Real$ is (uniformly) continuous.
\end{enumerate}
\end{lemma}
\begin{proof}
The first follows directly since $c(y,x)=\Delta_y(x)$. For the second, for ever $k\geq 0$ consider the continuous function $T_k:N\to\Real$ defined as
\[
T_k(x_0,y_0,z)= \sum_{i=1}^{k}\varphi\circ f^{-i}(\hu_{x_0,y_0}(z))-\varphi\circ f^{-i}(z)
\]
and note that by the Weirstrass' test, $T_k$ converges uniformly to
\[
\sum_{k=1}^{+\oo}\varphi\circ f^{-k}(\hu_{x_0,y_0} z)-\varphi\circ f^{-k}(z)
\]
which implies that $\log T$ is continuous, and thus $T$ is continuous.
\end{proof}
\subsection{The equilibrium state}\label{sub:theequilibrium}
It is seldom the case that for a given $\mu$ invariant measure for $f$, the partition by unstable manifolds is $\mu$- measurable, therefore we will have to work in sets $B$ where $\Fu|B$ defines a sufficiently nice partition. Since we are trying to use the family $\{\mcsx\}_{x\in M}$ as a substitute for the transverse measure of the equilibrium state, we will have to deal with the technical difficulty of having to show coherence between different choices $B,B'$; this will be bypassed using our control on the behavior of $\{\mcsx\}_{x\in M}$ under unstable (local) holonomies. We will now make some geometrical preparations to deal with this problem.
\paragraph{Dynamical boxes} To take advantage of the local product structure between $\Fcs$ and $\Fu$, we introduce some new notation. Due to the fact that the invariant bundles are mutually perpendicular there exists $\ep_0>0$ such that for any $0<\epsilon\leq\epsilon_0$ it holds for every $x,y\in M, d(x,y)<\ep$ implies
\begin{align}\label{eq:productstructure}
&\#\Ws{x,2\ep}\cap\Wcu{y,2\ep}=1\\
&\#\Wu{x,2\ep}\cap\Wcs{y,2\ep}=1
\end{align}
\begin{definition}\label{def:dynamicalboxes}
For $0<\epsilon\leq\epsilon_0$ and $x\in M$ we consider the sets
\begin{align}
\label{eq:Pcs}\Pcs{x,\epsilon}&:=\Wc{\Ws{x,\epsilon},\epsilon}\\
\label{eq:Pcu}\Pcu{x,\epsilon}&:=\Wc{\Wu{x,\epsilon},\epsilon}\\
\label{eq:dynbox}B(x,\epsilon)&:=\bigcup_{\crampedclap{y\in \Wu{x,\epsilon}}}\ \hu_{x,y}(\Pcs{x,\epsilon})
\end{align}
where $\hu_{x,y}:\Wcs{x}\rightarrow \Wcs{y}$ is the locally defined Poincar\'e map sending $x$ to $y$.
Sets $B(x,\epsilon)$ constructed in this fashion will be referred as \emph{dynamical boxes}, while $\hu_{x,y}(\Pcs{x,\epsilon})$ will be referred as the $cs$-plaques of $B(x,\ep)$.
\end{definition}
By shrinking $\ep_0$ if necessary can guarantee also that for $0<\ep\leq\ep_0$, for every $x,x'\in M$,
\begin{itemize}[leftmargin=*]
\item $\Pcu{x,\epsilon}\cap \Pcu{x',\epsilon}\neq\emptyset\Rightarrow \Pcu{x,\epsilon}\cap \Pcu{x',\epsilon}\subset \Pcu{x'',3\epsilon}$ for some $x''\in M$.
\item $B(x,\ep)\cap B(x',\epsilon)\neq\emptyset\Rightarrow B(x,\epsilon)\cap B(x',\epsilon)\subset B(x'',5\epsilon)$ for some $x''\in M$.
\end{itemize}
\smallskip
To understand better the structure of dynamical boxes we note the following.
\begin{lemma}\label{lem:holisometry}
Fix $x\in M$ and for $y\in\Ws{y}$ consider the locally defined holonomy $\hs_{x,y}:\Wcu{x}\to\Wcu{y}$. Then $\hs$ sends center leaves into center leaves and $\hs|\Wc{x}$ is an isometry. Similarly for unstable holonomies.
\end{lemma}
\begin{proof}
Let $\delta>0$ sufficiently small so that $\hs_{x,y}:A=\Wcu{x,\delta}\to B\subset\Wcu{y}$ is a well defined homeomorphism. Since $f$ preserves $\Fcu,\Fc,\Fs$ it follows \[\hs_{x,y}=f^{-n}\circ \hs_{f^nx,f^ny}\circ f^n\quad \forall n\geq 0\]
where $\hs_{f^nx,f^ny}:f^n(A)\to f^n(B)\subset \Wc{f^ny}$. Using that $f|\Fc$ is an isometry and $f|\Fu$ expands distances, it follows that $\hs$ sends $\Wc{x,\delta}$ to $\Wc{y}$ and that the differentiable map $\hs_{f^nx,f^ny}|\Wc{f^nx}$ approaches uniformly the identity for large $n$, hence the claim.
\end{proof}
We remark that in general $\hs$ does not send $\Fu$ to itself. See figure \ref{fig:pcu}.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\linewidth]{pcu}
\caption{Comparison between $\Pcu{x,\epsilon},\Pcu{y,\epsilon}$ for $y\in\Ws{x}$. The stable holonomy sends each center plaque $\Wc{x',\ep}\in \Pcu{x,\epsilon}$ to a center plaque $\Wc{y',\ep}\in \Pcu{y,\epsilon}$, but the image of unstable plaques in $\Pcu{x,\epsilon}$ are only H\"older submanifolds in $\Pcu{y,\epsilon}$. Nevertheless, $\hs_{x,y}(\Wu{x,\epsilon})\to\Wu{x,\epsilon}$ in the $\mathcal{C}^0$ topology as $y\to x$}
\label{fig:pcu}
\end{figure}
\begin{corollary}\label{cor:dynamicalboxsymmetric}
For every $0<\ep<\epsilon_0$ and $x\in M$ it holds
\[B(x,\ep)=\bigcup_{y\in \Ws{x,\epsilon}}\hs_{x,y}(\Pcu{x,\epsilon})\]
where $\hs_{x,y}:\Wcu{x}\rightarrow \Wcu{y}$ is the locally defined Poincar\'e map sending $x$ to $y$.
\end{corollary}
\begin{proof}
This is direct consequence of Lemma \ref{lem:holisometry} above plus the fact that $\Fc$ sub-foliates $\Fcu,\Fcs$.
\end{proof}
In view of this Corollary, sets $\hs_{x,y}(\Pcu{x,\epsilon})$ will be referred as the $cu$-plaques of $B(x,\ep)$.
\smallskip
\paragraph{The equilibrium state.} We will now construct the equilibrium state. Fix $B=B(x,\epsilon)$ a dynamical box with $0<\ep<\ep_0/2$. By the choice of $\epsilon$ the set $\Wu{x,B}=\Wu{x,2\epsilon}\cap B$ consists of a unique unstable plaque. Similar considerations hold for the other foliations.
If $W$ is a $cs$-plaque of $B$ and $x\in B$ denote $x_w=\Wu{x,B}\cap W$. For $U\subset B$ open sub-box, define the function $\alpha_{W,B}^U:W\rightarrow \Real$ by
\begin{equation}\label{eq.alphaWB}
\alpha_{W,B}^U(w)=\nu^u_{w}(U\cap \Wu{w,2\epsilon}).
\end{equation}
\begin{lemma}\label{lem:semicont}
$\alpha_{W,B}^U$ is upper semi-continuous, hence measurable.
\end{lemma}
\begin{proof}
The SRB case is direct, so we will consider only the case when $\varphi$ is constant along center leaves.
We use the following notation: for $\gamma>0$ and $C$ contained in a leaf of $\Fcs|B$ we write
\[
\mathrm{Col}_{\gamma}(C):=\{y:\exists y'\in C\text{ s.t. }d_{\Fcu}(y,y')<\gamma\}.
\]
Consider $y\in B$ and a sequence $(y_m)_m\subset B$, $y_m\xrightarrow[m\mapsto\oo]{}y$. Let $A=\Wu{y,B}\cap U, A_m=\Wu{y_m,B}\cap U$ and fix $\gamma>0$. By Lemma \ref{lem:JacobianoT} the function $y\mapsto \Delta_{y_w}(y)$ is positive and uniformly continuous on $B$, hence given $\tau>0$ there exists $0<\rho<\tau$ such that $y,z\in B$, \(d(y,z)<\rho\Rightarrow \left|\frac{\Delta_{y_w}(y)}{\Delta_{z_w}(z)}-1\right|<\tau\). Moreover, by Proposition \ref{pro:medidacu} we can assume that if $\phi_1,\phi_2\in \mathrm{Con}^+(\Fcu)$ are $\rho$-equivalent then $|\frac{\mu^{cu}(\phi_1)}{\mu^{cu}(\phi_2)}-1|<\tau$; this is turn implies that if $Q_1,Q_2\in\Sigma(\Fcu)$ are $\rho$-equivalent and $\phi_1,\phi_2$ as before then
\[
\left|\frac{\mu^{cu}(\phi_1\one_{Q_1})}{\mu^{cu}(\phi_2\one_{Q_2})}-1\right|<\tau.
\]
By continuity of the unstable and center unstable foliations there exists $m_0$ such that for every $m\geq m_0$ the set $\Wc{A_m,\epsilon}$ is $\rho$-equivalent to a subset $E_m$ of $\mathrm{Col}_{\gamma}(\Wc{A,\epsilon})$ and furthermore the Hausdorff distance between $A,A_m$ is less than $\rho$. Let $g_m:=\hu| : E_m\rightarrow \Wc{A_m,\epsilon}$ the corresponding holonomy and observe that for such $m$, $|\pi^c_{y_m}\circ g_m-\pi^c_y|<\rho$. Hence,
\begin{align*}
\MoveEqLeft \nu^{u}_{y_m}(A_m)=\int_{\Wc{A_m,\epsilon}}J^{u}_{y_m}\circ \pi^c_{y_m}d\mcux[y_m]\\
&=\int_{E_m}J^{u}_{y_m}\circ \pi^c_{y_m}\circ g_md\mcux[y]+\left(\int_{\Wc{A_m,\epsilon}}J^{u}_{y_m}\circ \pi^c_{y_m}d\mcux[y_m]-\int_{E_m}J^{u}_{y_m}\circ \pi^c_{y_m}\circ g_md\mcux[y]\right)\\
&<(1+\tau)\int_{E_m}J^{u}_{y_m}\circ \pi^c_{y_m}\circ g_md\mcux[y]\\
&=(1+\tau)\int_{E_m}J^{u}_y\circ \pi^c_{y}d\mcux[y]+(1+\tau)\left(\int_{E_m}J^{u}_{y_m}\circ \pi^c_{y_m}\circ g_m-J^{u}_y\circ \pi^c_{y}d\mcux[y]\right)\\
&<(1+\tau)(1+\tau)\int_{E_m}J^{u}_y\circ \pi^c_{y}d\mcux[y]\leq (1+\tau)^2\int_{Col_{\gamma}(\Wc{A,\epsilon})}J^{u}_y\circ\pi^c_yd\mcux[y].
\end{align*}
This implies that $\limsup_m \alpha_{W,B}^U(y_m)=\limsup_m \nu^{u}_{y_m}(A_m)\leq \int_{Col_{\gamma}(\Wc{A,\epsilon})}J^{u}_y\circ\pi_y^cd\mcux[y].$ Taking $\gamma\mapsto 0$ and using regularity of the measure $\mcux[y]$ we conclude $\limsup_m \alpha_{W,B}^U(y_m)\leq \alpha_{W,B}^U(y)$, as we wanted to show.
\end{proof}
Standard arguments of measure theory together with the previous Lemma permit us to define a Borel measure $\mathrm{m}^{W,B}$ on $B$ by the condition
\begin{equation*}
U\subset B \text{ open sub-box }\Rightarrow \mathrm{m}^{W,B}(U):=\int \alpha_{W,B}^U(w)\mcsx[w_0](dw)
\end{equation*}
where $w_0\in W$.
Let $W'$ be another $cs$-plaque of $\Fcs|B$ and consider $\hu_{w_0,w_0'}:W\rightarrow W'$ the corresponding Poincar\'e map. Fix $U\subset B$ open, and to simplify the notation write $U_w^u:= U\cap \Wu{w,B}$. We compute
\begin{align*}
\mathrm{m}^{W',B}(U)&=\int_{W'} \nu^u_{w'}(U^u_{w'})\mcsx[w_0'](dw')=
\int_{W} \nu^u_{\hu_{w_0,w_0'}(w)}(U^u_{\hu_{w_0,w_0'}(w)})
\Jacu[w_0,w_0'](w)\mcsx[w_0](dw)\\
&=\int_{W} \left(\nu^u_{\hu_{w_0,w_0'}(w)}(U^u_{w})\prod_{k=1}^{+\oo}\frac{e^{\varphi\circ f^{-k}\hu_{w_0,w_0'}(w)}}{e^{\varphi\circ f^{-k}(w)}}\right)\mcsx[w_0](dw)\\
&=\int_W \nu^u_w(U_w)\mcsx[w_0](dw)=\mathrm{m}^{W,B}(U),
\end{align*}
hence $\mathrm{m}^{W',B}=\mathrm{m}^{W,B}$: we write $\mathrm{m}^B=\mathrm{m}^{W,B}$ for any plaque $W\in\Fcs|B$. The following is now clear.
\begin{lemma}\label{lem:independencebox}
Consider $0<\epsilon<\epsilon'<\epsilon_0$ and $x,x'\in M$ such that $B=B(x,\epsilon)\subset B(x',\epsilon')$. Then $\mathrm{m}^B=\mathrm{m}^{B'}|B$.
\end{lemma}
Take a finite covering $\mathscr{B}=\{B_1=B(x_1,\epsilon_1),\ldots, B_R=B(x_R,\epsilon_1)\}$ of $M$ with $0<\epsilon_1<\frac{\epsilon_0}{3}$ and define the measure $\mathrm{m}$ by the condition: $A\subset B_i$ is Borel $\Rightarrow$ $\mathrm{m}(A):=\mathrm{m}^{B_i}(A)$. If $A\subset B_i\cap B_j$ then by our choice of $\epsilon_1$ there exists $x\in M$ with $A\subset B(x,\epsilon_0)$, and thus by the previous lemma $\mathrm{m}^{B_i}(A)=\mathrm{m}^{B_j}(A)$, hence $\mathrm{m}$ is a well defined Borel measure on $M$. With no loss of generality we assume further that for every $i$ there exists $x_i'$ such that $f(B_i)\subset B(x_i',\epsilon_0)$.
\begin{proposition}\label{pro:mesinvariante}
The measure $\mathrm{m}$ is $f$-invariant.
\end{proposition}
\begin{proof}
Fix $\mathrm{m}^{B_i}=\mathrm{m}^{W_i,B_i}$ and $U\subset f(B_i)$ open. Then
\begin{align*}
f_{\ast}\mathrm{m}^{W_i,B_i}(U)&=\int_W \nu^u_w(f^{-1}U_{fw}^u)\mcsx[w_0](dw)
=\int_{W} e^{\varphi(w)-P}\nu^u_{fw}(U_{fw}^u)\mcsx[w_0](dw)\quad&\text{ by }\eqref{eq:invariancianu}\\
&=e^{P'-P}\int \nu^u_{z}(U_{z}^u)\mcsx[fw_0](dz)\quad&\text{ by Prop. }\ref{pro:medidacs}\\
&=e^{P'-P}\mathrm{m}^{fW_i,fB_i}(U).
\end{align*}
Hence $f_{\ast}\mathrm{m}^{W_i,B_i}=e^{P'-P}\mathrm{m}^{fW_i,fB_i}$, and this implies $f_{\ast}\mathrm{m}=e^{P'-P}\mathrm{m}$. The measure $\mathrm{m}$ is finite on compact sets and $fM=M$, thus $P=P'$ and $\mathrm{m}$ is $f$-invariant.
\end{proof}
\begin{corollary}\label{cor:PigualP}
It holds $P=P'$.
\end{corollary}
\begin{definition}\label{def.gibbsmeasure}
We denote $\muf$ the probability measure $\frac{\mathrm{m}}{\mathrm{m}(M)}$.
\end{definition}
Interchanging $\mu^u$ by $\mu^s$ and $\mu^{cs}$ by $\mu^{cu}$ we construct another invariant measure $\tilde{\mathrm{m}}_{\scriptstyle\varphi}$ which is given locally as
\[
\tilde{\mathrm{m}}_{\scriptstyle\varphi}(U)=\int_V \nu_v^s(U_v^s)\mu^{cs}_V(dv).
\]
\smallskip
\noindent\textbf{Convention:} The family $\mathscr{B}=\{B_1,\ldots B_R\}$ together with the corresponding center stable plaques $W_1,\ldots W_R$ is considered to be fixed.
\subsection{Product structure and the Gibb's property}
\label{sub:product_structure_and_the_gibb_s_property}
We start establishing the following.
\begin{proposition}[Product structure of $\muf$]\label{pro:estructuraproducto}
There exists $0<\epsilon_2\leq\epsilon_1$ such that for every $0<\ep\leq\epsilon_2$ there exists $c(\ep)>0$ satisfying for every $x\in M$,
\[
\frac{1}{c}\leq \frac{\muf(B(x,\ep))}{\mux(\Wu{x,B(x,\ep)})\cdot\mcsx{(\Pcs{x,\epsilon}})}\leq c.
\]
\end{proposition}
The proof follows directly by the definition of $\muf$ with the following lemma.
\begin{lemma}\label{lem:comparacionentretamanhos}
There exists $0<\epsilon_2\leq\epsilon_1$ such that for $0<\ep\leq \epsilon_2$ we can find $c(\ep)>0$ such that for all $x,x'\in M$ the following holds.
\begin{align*}
&(\ast)\quad c(\ep)^{-1}\leq \frac{\mcux(\Wcu{x,\ep})}{\mcux[x'](\Wcu{x',\ep})}, \frac{\mcsx(\Wcs{x,\ep})}{\mcsx[x'](\Wcs{x',\ep})}\leq c(\ep)\\
&(\ast\ast)\quad c(\ep)^{-1}\leq \frac{\mux(\Wu{x,\ep})}{\mux[x'](\Wu{x',\ep})} \leq c(\ep)\\
&(\ast\ast\ast)\quad c(\ep)^{-1}\leq \frac{\nux(\Wu{x,\ep})}{\nux[x'](\Wu{x',\ep})} \leq c(\ep)
\end{align*}
\end{lemma}
\begin{proof}
Let $\epsilon_2$ be the Lebesgue number associated to the covering $\mathscr{B}$. We start with $(\ast)$; it suffices to consider the first set of inequalities. Since $\mu^{cu}_x$ is a Radon measure of full support, we have the following: for $\ep>0$ fixed, there exists $\ga>0$ and $c_0(x,\ep)>0$ such that
\[
x'\in\Wcu{x,\rho}\Rightarrow c_0(\ep,x)^{-1}\leq \frac{\mcux(\Wcu{x,\ep})}{\mcux[x'](\Wcu{x',\ep})}\leq c_0(\ep,x).
\]
By \cref{lem:JacobianoT} it is no loss of generality to assume that the same holds for $x'\in D(x,\ga)$, and hence due to the compactness of $M$ we get a uniform $c(\ep)$ independent of $x$.
For $(\ast\ast)$, the claim is immediate for the SRB case, so we just need to consider when $\varphi$ is constant along centrals. But this follows from as in the previous part since $\mux(A)=\mcux[x](\Wc{A,\ep_0})$.
Finally, $(\ast\ast\ast)$ is consequence of $(\ast\ast)$ together with \cref{lem:JacobianoT}.
\end{proof}
\begin{remark}\label{rem:estructuraproducto}
Arguing similarly and re-defining $c,\ep_2$ if necessary we can guarantee that for every $x\in M$,
\[
\frac{1}{c(\ep)}\leq \frac{\tilde{\mathrm{m}}_{\scriptstyle\varphi}(B(x,\ep))}{\msx(\Ws{x,B(x,\ep)})\cdot\mcux{(\Pcu{x,\epsilon}})}\leq c(\ep).
\]
\end{remark}
Next we establish a Gibbs's type property for $\muf$ (resp. $\tilde{\mathrm{m}}_{\scriptstyle\varphi}$) that will allow us to show that it is an equilibrium state for the potential $\varphi$. By reducing $\epsilon_2$ if necessary, we can assume that for every $0<\ep<\ep_2$, for every $\epsilon$-ball $D\subset M$ the distance between $x,y\in D$ is given by
\begin{align*}
d(x,y)=\inf\{&\mathrm{length}(\alpha):\alpha:[0,1]\rightarrow D \text{ piecewise }\mathcal{C}^1,\alpha(0)=x,\alpha(1)=y, \alpha'\in E^{\ast}\ \ast=s,c,u\}.
\end{align*}
With this metric and using that $f$ is a center isometry we have
\begin{align}\label{eq:discoBowen}
&\Benu=\Ben|\Wu{x}=f^{-n}\Wu{f^nx,\epsilon}\\
\nonumber &\Ben=\bigcup_{y\in \Wcs{x,\epsilon}}\Benu[y]
\end{align}
We note the following classical remark.
\begin{lemma}\label{lem:Bowenproperty}
Given $\epsilon>0$ there exists $K(\ep)>0$ with the property that for every $x,y\in M, n\geq0$ it holds
\begin{enumerate}
\item \(y\in \Benu\Rightarrow |\SB(x)-\SB(y)|<K\).
\item \(y\in \Ben\Rightarrow |\SB(x)-\SB(y)|<nK\). If $\varphi$ does not depend on centrals then \(y\in \Ben\Rightarrow |\SB(x)-\SB(y)|<K\).
\end{enumerate}
Moreover, $K(\epsilon)\mapsto 0$ as $\epsilon\mapsto0$.
\end{lemma}
\begin{proof}
We will prove only the second part, as the arguments for the first are included in the proof of this case. Let $y_s:=\Ws{y,\epsilon}\cap f^{-n}\Wcu{f^nx,\epsilon}$. Points in the same stable leaf are exponentially contracted and $\varphi$ is H\"older, thus
\[
|\SB(x)-\SB(y)|\leq |\SB(x)-\SB(y_s)|+C_1(\epsilon)
\]
for some constant $C_1(\epsilon)$ that converges to zero as $\epsilon$ tends to zero. Since $f^ny_s\in \Wcu{f^nx,\epsilon}$, the point $z_u:=\Wu{f^ny,\epsilon}\cap \Wc{f^nx,\epsilon}$ is well defined. Observe then that the distance $d(f^{-j}z_u,f^{n-j}y_s)$ is exponentially contracted under $f^{-1}$, and thus for some other similar constant $C_2(\epsilon)$ we have
\[
|\SB(x)-\SB(y)\leq |\SB(x)-\SB(f^{-n}z_u)|+C_2(\epsilon).
\]
Since $f$ is a center isometry and $\varphi$ is H\"older it follows that
\[
|\SB(x)-\SB(f^{-n}z_u)|\begin{dcases}
=0 & c\text{-constant case}\\
\leq C_{\varphi}n\ep^{\theta}:=C_{3}(\ep) & \text{ in general}.
\end{dcases}
\]
Defining $K(\ep):=C_{1}(\ep)+C_{2}(\ep)+C_{3}(\ep)$ we finish the proof.
\end{proof}
Now we can deduce:
\begin{corollary}[Gibbs-like property of $\muf$]\label{cor:gibbslineal}
Given $\epsilon>0$ there exists $d(\epsilon)>0$ such that for every $n\in\mathbb{N},x\in M$ it holds
\begin{align*}
d^{-1}(\ep)\leq &\frac{\nux(\Benu)}{e^{\SB(x)-nP}} \leq d(\ep)\\
d^{-1}(\ep)e^{-K(\ep)n}\leq &\frac{\muf(\Ben)}{e^{\SB(x)-nP}} \leq d(\ep)e^{K(\ep)n}
\end{align*}
where $K(\ep)$ is the constant given in \cref{lem:Bowenproperty}. If $\varphi$ does not depend on centrals then in the second line of inequalities we can substitute $nK(\ep)$ by $K(\ep)$.
\end{corollary}
\begin{proof}
To simplify the notation we will write $a\simeq b$ if the quotient $\frac{a}{b}$ is bounded above and below with constants only depending on $\epsilon$. The first part is direct consequence of \cref{lem:comparacionentretamanhos} and \eqref{eq:invariancianu}. For the second part, we have $\mathrm{m}^{W_i,B_i}(\Ben)=\int_{\Wcs{x,\epsilon}} \nux[y_w](\Benu[y])\mcsx(dy)$, therefore
\begin{align*}
\mathrm{m}^{W_i,B_i}(\Ben)&\simeq\int_{\Wcs{x,\epsilon}} e^{\SB\circ f^{-n}(y)-nP}\mcsx(dy)\quad &\text{(by the first part)}\\
&\simeq e^{\SB(x)-nP\pm nK}\mcsx(\Wcs{x,\epsilon})\quad&\text{(by \cref{lem:Bowenproperty})}\\
&\simeq e^{\SB(x)-nP\pm nK} \quad &\text{(by \cref{lem:comparacionentretamanhos})}.
\end{align*}
The last claim follows from the corresponding one in \cref{lem:Bowenproperty}.
\end{proof}
We are ready to show that $\muf$ is an equilibrium state.
\begin{proposition}\label{pro:PigualPresion}
It holds $P=\Ptop(\varphi)=P_{\mu_{\varphi}}(\varphi).$
\end{proposition}
\begin{proof}
Fix $0<\epsilon<\ep_2$ small and consider a set $E$ such that $\{B(x,\epsilon)\}_{x\in E}$ is pairwise disjoint, maximal with this property. Observe that $M=\cup_{x\in E} B(x,2\epsilon)$. Take a partition by Borel sets $A=\{A_x:x\in E\}$ such that for every $x\in E$, $D(x,\epsilon)\subset A_x \subset D(x,2\epsilon)$. Denote $A^n=\bigvee_{i=0}^{n-1}f^{-i}A$ and choose a set of representatives $E_n$ such that $A^n=\{A^n_x:x\in E_n\}$. It follows that for every $x\in E_n$, $D(x,\ep,n)\subset A_x^n\subset D(x,4\epsilon,n)$. Then by \cref{lem:Bowenproperty} and the previous corollary,
\begin{align*}
-\muf(A_x^n)\log\muf(A_x^n)+\int_{A_x^n} \SB d\muf&\geq -\muf(A_x^n)(\log\muf(A_x^n)-\SB(x)+nK(4\ep))\\
&\geq -\muf(A_x^n)(-nP+2nK(4\ep)+\log d(4\ep))\shortintertext{thus,}
\frac{H_{\muf}(\bigvee_{i=0}^{n-1}f^{-i}A)}{n}+\int \varphi d\mu_{\varphi}
&=\frac{1}{n}\sum_{x\in E_n}\big(-\muf(A^n_x)\log \muf(A^n_x)+\int_{A_x^n} S_nd\muf\big)\\
&\geq P-2K(4\ep)+\frac{\log d(4\ep)}{n}.
\end{align*}
Taking limit first as $n\to\oo$ and then as $\epsilon\to 0$ we obtain
\begin{align*}
P_{\muf}(\varphi)\geq h_{\muf}(f;A)+\int \varphi d\muf\geq P.
\end{align*}
On the other hand, by the same construction
\begin{align*}
S(4\epsilon,n)\leq \sum_{x\in E_n} e^{\SB(x)}\leq \sum_{x\in E_n}\muf(D(x,\ep,n)) e^{nP\pm nK(\ep)}\leq e^{nP\pm nK(\ep)}\sum_{x\in E_n}\muf(A_x^n)= e^{nP\pm nK(\ep)}
\end{align*}
and thus $\Ptop(\varphi)\leq P$. The variational principle then implies that $P=\Ptop(\varphi)= P_{\muf}(f)$ and we are done.
\end{proof}
Since $P=0$ by construction in the SRB case, we get.
\begin{corollary}\label{cor.obvioSRB}
If $\mathrm{m}_{\scriptscriptstyle SRB}$ is the measure associated to $\varphi=-\log\det Df|E^u$, then $P_{\mathrm{m}_{\scriptscriptstyle SRB}}(f)=0$.
\end{corollary}
\subsection{Uniqueness}
\label{sub:uniqueness}
As was indicated in the introduction, the uniqueness for the SRB case is proven in \cite{lecturesugibbs}. Here we present a different argument that establishes uniqueness in both the $c$-constant and the SRB case; it is based in arguments of Bowen \cite{Bowen1974} for expansive systems with the specification property. We will use that $\muf$ is ergodic; this will be established in \cref{sec:finer_ergodic_properties}.
\smallskip
The proof is by contradiction: we suppose that there exists $\mathrm{m}\neq\muf$ another equilibrium state for $\varphi$, that without loss of generality we assume to be ergodic. Since $\mathrm{m}\perp \muf$, we can find an invariant Borel set $X\subset M$ such that $\nu(X)=1, \muf(X)=0$. Fix $\ep>0$ small ($0<\ep<\ep_2$) and construct, as in the proof of \cref{pro:PigualPresion}, a finite partition by Borel sets $\mathscr{A}^n=\{A_x^n:x\in E^n\}$, with $D(x,\ep,n)\subset A_x^n\subset D(x,4\epsilon,n)$ for every $x\in E$. Recall that by \eqref{eq:discoBowen}, $\Ben=\bigcup_{y\in \Wcs{x,\epsilon}}\Benu[y]$ therefore each $A_x^n$ contains a center stable plaque $\Wcs{x,\ep}\subset P_x^n\subset \Wcs{x,4\ep}$. We will now subdivide further each $A^n_x$, and for this we will subdivide the corresponding $\Wcs{x,4\ep}$ in the form
\[
\Wcs{x,4\ep}=\bigcup_{y\in I_x} \Ws{J_y^x,4\ep}
\]
where $I_x\subset \Wc{x,4\ep}$ is finite, and $J_y^x\subset \Wc{y,(4\ep)^n}$. This way we can induce a partition on $A^n_x$, and get a finite partition $\mathscr{C}^n=\{C^n_z:z\in F^n\}$ such that
\begin{itemize}
\item for each $z\in F$ there exists $x\in E$ such that $D(x,\ep,n)\subset C_z^n\subset D(x,4\epsilon,n)$.
\item $\sup_z\diam(f^{[n/2]}C^n_z)\xrightarrow[n\to\oo]{}0$.
\end{itemize}
\begin{definition}
A set of the form $S(x,\ep,n)=\Ws{\Wc{x,\ep^n},\ep}$ will be referred as a $(\ep,n)$-slice.
\end{definition}
We need the following refinement of \cref{lem:comparacionentretamanhos}.
\begin{lemma}\label{lem:comparaslice}
For $\ep>0$ sufficiently small there exists $e(\ep)>0$ such that for every $x,x'\in M$, it holds
\[
e(\ep)^{-1}\leq\frac{\mcsx[x](S(x,\ep,n))}{\mcsx[x'](S(x',\ep,n))}\leq e(\ep).
\]
\end{lemma}
\begin{proof}
Fix $0<\ep<\ep_2$ and $\eta=\Leb\in\mathrm{Meas}(\Fcs)$. Consider in $Q(x,\ep)=\bigcup_{y\in \Ws{x,\ep}}\hs_{x,y}(\Wc{x,\ep})$ and observe that by \cref{lem:holisometry} the set $\Q(x,\ep)$ is saturated by local center leaves $\Wc{y,\ep}$. Define the transverse measure on $\Wc{x,\ep}$:
\[
\eta^c_{x,T}(A)=\frac{\eta(A)}{\eta(Q(x,\ep))}
\]
The foliation $\Fs|\Fcs$ is $\mathcal{C}^1$ \cite{HolFol}, therefore $\eta^c_{x}$ is the Lebesgue measure. The same holds for $\eta^n$ \eqref{eq:nun} due to Fubini's theorem and the fact that $f^n(Q(x,\ep))$ is a set saturated by center plaques $\Wc{y,\ep}$ (see the proof of Theorem $7.8$ in \cite{PesinLect}), and $\det Df|E^c=1$. By passing to the limit we deduce that $\mu^c_{x,T}$ is the Lebesgue measure (the one induced by the metric on $E^c$ that is invariant by $f$).
In particular we there exists $e_0(\ep)>0$ such that for every $x,x'\in M, n\in \Nat$,
\[
e_0(\ep)^{-1}\leq\frac{\mu^c_{x,T}(\Wc{x,\ep^n})}{\mu^c_{x',T}(\Wc{x',\ep^n})}\leq e_0(\ep).
\]
On the other hand by the same token, the measure $\mcsx[x]{(S(x,\ep,n))}$ is comparable, modulo constants that only depend on $\ep$, to $\frac{\mcsx[x]{(Q(x,\ep)})}{\mu^c_{x,T}(\Wc{x,\ep^n})}$ and $\mcsx[x]{(Q(x,\ep))}, \mcsx[x']{(Q(x',\ep))}$ are uniformly comparable (\cref{lem:comparacionentretamanhos}) the claim follows.
\end{proof}
\begin{corollary}
Given $\ep>0$ there exists $l(\ep)>0$ such that for every $C^n_z$,
\[
e^{\SB(z)-n\Ptop(\varphi)}\leq l(\ep)\cdot\muf(C^n_z).
\]
\end{corollary}
\begin{proof}
Arguing exactly as in the proof of \cref{cor:gibbslineal}, using that the size of $C^n_z$ decays exponentially fast with $n$, we deduce using the previous lemma and the product structure of $\muf$, \cref{pro:estructuraproducto},
\[
\muf(C^n_z)\sim \nu^u_z(D(x,\ep,n))\cdot\mcsx(S(z,\ep,n))\sim d(4\ep)e(\ep)\cdot e^{\SB(z)-n\Ptop(\varphi)}.
\]
\end{proof}
We continue with the proof of uniqueness. Observe that the partition $f^{[n/2]}\mathscr{C}^n$ generates $\BM$, and therefore (see for example lemma $2$ in \cite{BowenMaxHyp}) there exists a Borel set $X_n$ that is union of atoms of $\mathscr{C}^n$ satisfying
\begin{align*}
&\muf(f^{[n/2]}X_n\triangle X)\xrightarrow[n\to\oo]{}0\quad \nu(f^{[n/2]}X_n\triangle X)\xrightarrow[n\to\oo]{}0\shortintertext{ which, by invariance of $X$ implies}
&\muf(X_n\triangle X)\xrightarrow[n\to\oo]{}0\quad \nu(X_n\triangle X)\xrightarrow[n\to\oo]{}0\shortintertext{ therefore, }
& \muf(X_n)\xrightarrow[n\to\oo]{}0,\quad \nu(X_n)\xrightarrow[n\to\oo]{}1.
\end{align*}
Note that for any $x$ it holds $D(x,\ep,\oo):=\cap_{n=0}^{\oo} \Ben=\Wc{x,\ep}$, and since $f$ is a center isometry $\htop(f,D(x,\ep,\oo))=0$. It follows that $f$ is what is called an $h$-expansive map; from their theory we will only use lemma $1$ of \cite{BowenMaxHyp} which tells us that for $\ep$ sufficiently small,
\[
h_{\nu}(f)\leq \frac{1}{n}H_{\nu}(\mathscr{C}^n),
\]
hence
\begin{align*}
&\Ptop(\varphi)=h_{\nu}(f)+\int \varphi d\nu\leq \frac{1}{n}(-\sum_{z\in F^n}\nu(C_z^n)\log \nu(C_z^n)+\int \SB d\nu)\\
&\Rightarrow n\Ptop(\varphi)\leq \sum_{z\in F^n}(-\nu(C_z^n)\log \nu(C_z^n)+\int \SB\cdot\one_{C_z^n} d\nu)\leq \sum_{z\in F^n}\log \nu(C_z^n)+(K+\SB(z))\nu(C_z^n)
\end{align*}
where $C_z^n\subset A^n_x\subset D(x,4\ep,n)$, and $K=K(4\ep)$ is given in \cref{lem:Bowenproperty}. It follows that
\begin{equation}\label{eq:presionotramedida}
n\Ptop(\varphi)\leq K+\sum_{C_z^n\subset X_n}\nu(C_z^n)(\SB(z)-\log\nu(C_z^n))+ \sum_{C_z^n\subset X_n^c}\nu(C_z^n)(\SB(z)-\log\nu(C_z^n)).
\end{equation}
Next we use the following Calculus fact (cf. Lemma $7$ \cite{Bowen1974}).
\begin{claim} if $a_1,\cdots, a_k\in [0,1]$ are such that $S=\sum_{i=1}^k a_i\leq 1$ then for any list of real numbers $b_1,\cdots, b_k$ it holds
\[
\sum_{i=1}^ka_i(b_i-\log a_i)\leq S(\log \sum_{i=1}^k e^{b_i}-\log S).
\]
\end{claim}
We get, using the above claim in \eqref{eq:presionotramedida} with $a_i=\nu(C_z^n), b_i=\SB(z)$,
\begin{align*}
n\Ptop(\varphi)-K\leq \nu(X_n)\log \sum_{C_z^n\subset X_n}e^{\SB(z)}+\nu(X_n^c)\log \sum_{C_z^n\subset X_n^c}e^{\SB(z)}+K'
\end{align*}
where $K'=\frac{1}{2}\max\{-t\log t:t\in[0,1]\}$. Write $-n\Ptop(\varphi)= \nu(X_n)\log e^{-n\Ptop(\varphi)}+\nu(X_n^c)\log e^{-n\Ptop(\varphi)}$ and appply \cref{cor:gibbslineal} to get
\begin{align*}
-(K+K')&\leq \nu(X_n)\log \sum_{C_z^n\subset X_n}e^{\SB(z)-n\Ptop(\varphi)}+\nu(X_n^c)\log \sum_{C_z^n\subset X_n^c}e^{\SB(z)-n\Ptop(\varphi)}\\
&\leq \nu(X_n) \log \sum_{C_z^n\subset X_n} l(\ep,n)\cdot \muf(C_z^n)+ \nu(X_n^c) \log \sum_{C_z^n\subset X_n^c} l(\ep,n)\cdot \muf(C_z^n)\\
&\leq 2\log l(\ep)+\nu(X_n)\log\muf(X_n)+\nu(X_n^c)\log \muf(X_n^c)\xrightarrow[n\to\oo]{}-\oo
\end{align*}
which is a contradiction.
\section{A characterization of equilibrium states}
\label{sec:a_characterization_of_equilibrium_states}
The aim of this section is to characterize equilibrium states in terms of their conditional measures along strong unstable leaves, very much as the case of SRB measures, where they are characterized by the property of their conditionals being absolutely continuous with respect to Lebesgue measures along unstables. We will rely on the seminal work of Ledrappier-Young \cite{LedYoungI}.
\smallskip
\noindent\textbf{Hypotheses of this section:} $f:M\rightarrow M$ is a $\mathcal{C}^2$ center isometry preserving a Borel probability measure $\mathrm{m}$. We assume the existence of a family of measures $\mu^{\ast}, \ast\in\{s,u,cs,cu\}$ satisfying $2, 3$ of Theorem A.
\smallskip
Of course the existence of such families is guaranteed only when $\varphi$ is either constant along centrals or $\varphi=-\log \det Df|E^u$, but since the results only depend on the existence of these families of measures, there is value in writing this part abstractly.
Let $\mathrm{m}\in \PTM{f}{M}$ and $\nu^{\sigma}\in \mathrm{Meas}(\F^{\ast}), \ast=s,u$. We say that $\mathrm{m}$ has conditionals
equivalent to $\nu^{\ast}$ along $\F^{\sigma}$ if for every $\mathrm{m}$-measurable partition $\xi$ subordinated to the (possibly non-measurable partition) $\F^{\ast}$ it holds $\mathrm{m}^{\xi}_x\sim \nu_x^{\ast}$ for $\maex[x]$. See \cite{Rokhlin} for a discussion of measurable partitions and conditional measures.
Define,
\begin{align}
\label{eq:tienecondicionalmu} \PM[M, \mu^u]&=\{\mathrm{m}\in \PM[M]:\mathrm{m}\text{ has conditionals equivalent to } \mu^u\}\\
\label{eq:tienecondicionalmuinv} \PTM{f}{M, \mu^u}&=\PM[M, \mu^u]\cap \PTM{f}{M}.
\end{align}
We will prove the following.
\begin{theorem}\label{thm:condicionalesdeterminanee}
The set $\PTM{f}{M\, \mu^u}$ coincides with the set of equilibrium states for $\varphi$.
\end{theorem}
\paragraph{Increasing partitions} A $\mathrm{m}$-measurable partition $\xi$ is called increasing if it refines its image ($f\xi<\xi$). If $\xi$ is a partition we denote by $\mathcal{B}_{\xi}(M)$ the sub-$\sigma$ algebra of $\BM[M]$ that consists of $\xi$-saturated sets.
\begin{definition}\label{def:SLY}
We call a measurable partition $\xi$ a SLY partition if it is increasing, subordinated to $\Fu$ and furthermore for $\mathrm{m}\aep(x)\in M$ the atom $\xi(x)$ contains a neighborhood of $x$ inside $\Wu{x}$.
\end{definition}
In our restricted setting, a SLY partition automatically satisfies the following additional properties:
\begin{enumerate}
\item $\bigvee_{n\geq 0} f^{-n}\xi=\BM$.
\item $\bigwedge_{n\geq 0} f^n\xi=\mathcal{B}^u$
\end{enumerate}
where $\mathcal{B}^u$ is the $\sigma$-algebra of \Fu\ saturated sets. A key result is the following.
\begin{theorem}\label{thm:SLY}
There exist SLY partitions of arbitrarily small mesh. If $\xi$ is a SLY partition then
\[
h_{\mathrm{m}}(f)=h_{\mathrm{m}}(f,\xi)=H_{\mathrm{m}}(f^{-1}\xi|\xi).
\]
\end{theorem}
The existence part is essentially due to Sinai \cite{SinaiMarkov} (compare Section 3 in \cite{LedStrelcyn} and Lemma 2.4.2 in \cite{LedYoungI}, where the construction is done in more generality). The second part is a deep result of Ledrappier and Young - Corollary 5.3 in \cite{LedYoungI}. We remark that although in this last article the result is presented for ergodic measures, it extends easily to non - ergodic ones by using the ergodic decomposition. We have tacitly used that the strong unstable manifolds coincide with the Pesin's unstable manifolds, as consequence of the fact that $f$ does not have any positive Lyapunov exponents along center directions, independently of the measure.
\subsection{Equilibrium states have conditionals equivalent to \texorpdfstring{$\mu^u$}{mu}}
Now we fix $\xi$ a SLY partition for the measure $\mathrm{m}$, and denote by $\mathrm{m}_x=\mathrm{m}^{\xi}_x$ the conditional measure of $\mathrm{m}$ on $\xi(x)$. The main result of \cite{LedYoungI} admits the following generalization in our setting.
\begin{theorem}\label{thm:marginalesunicas}
If $\mathrm{m}$ is any equilibrium state corresponding to $\varphi$ then
\[
\mathrm{m}_x\sim\mux \text{ on }\xi(x)\quad\text{ for } \mathrm{m}\aep(x).
\]
\end{theorem}
To give a self contained presentation we spell the complete proof instead of just citing the relevant parts of Ledrappier-Young work. Assume for the moment that the conclusion of the theorem was true. Then for $\mathrm{m}\aep(x)$ we have
\[
d\mathrm{m}_x=\rho_x d\mux,
\]
where $\rho_x$ is a non-negative function such that $\int_{\xi(x)} \rho_x d\mux=1$. Using that the partition $\xi$ is measurable we can assemble all $\{\rho_x\}$ into one measurable function $\rho:M\rightarrow \Real$. Since $\mathrm{m}\in \PTM{f}{M}$ we have
\[
f\mathrm{m}_{x}^{f^{-1}\xi}=\mathrm{m}_{fx} (=\mathrm{m}_{fx}^{\xi})\quad \mathrm{m}\aep(x).
\]
On the other hand $f^{-1}\xi>\xi$, thus for any $A\in \BM[M]$ we can write
\begin{align*}
\frac{\int_{A\cap f^{-1}\xi(x)}\rho d\mux}{\mathrm{m}_x(f^{-1}\xi(x))}&=\frac{\mathrm{m}_x(A\cap f^{-1}\xi(x))}{\mathrm{m}_x(f^{-1}\xi(x))}=\mathrm{m}_{x}^{f^{-1}\xi}(A)=\mathrm{m}_{f_x}(fA\cap \xi(fx))\\
&=\int_{f(A\cap f^{-1}\xi(x))}\rho d\mu^u_{fx}=\int_{A\cap f^{-1}\xi(x)}\rho\circ f e^{P-\varphi}d\mux,
\end{align*}
hence by uniqueness of the conditional measures $\mathrm{m}\aep(x)$ it holds
\[
\frac{\rho(y)}{\mathrm{m}_x(f^{-1}\xi(x))}=\rho(fy)\cdot e^{P-\varphi(y)}\quad \text{ for }\mu_x\aep(y)\in f^{-1}\xi(x).
\]
As the denominator appearing in the previous equation is $f^{-1}\xi$- measurable, we finally conclude that
the function
\[
x\mapsto \frac{\rho(x)}{\rho(f^{-1}x)}e^{P-\varphi(f^{-1}x)}
\]
is $\xi$ - measurable (i.e.\@ constant on the atoms of $\xi$). By a simple inductive argument one is led to conclude that $\rho$ satisfies
\begin{equation}\label{eq:jacobianoloc}
y\in \xi(x)\Rightarrow \frac{\rho(y)}{\rho(x)}=\prod_{k=1}^{\oo}\frac{e^{\varphi\circ f^{-k}(y)}}{e^{\varphi\circ f^{-k}(x)}}=\Delta_x^u(y).
\end{equation}
\begin{lemma}
For every $x\in M$ the function $\Wu{x}\ni y\mapsto \log \Delta_x^u(y)$ is locally Lipschitz and hence uniformly bounded away from $0$ and $\oo$ on the atom $\xi(x)$.
\end{lemma}
The proof is straightforward using that for every $x\in M$ the set $\Wu{x}$ is an immersed submanifold and the fact that $\varphi$ is globally H\"{o}lder.
From \eqref{eq:jacobianoloc} we deduce that $\rho$ should have the form
\begin{equation}
\rho(y)=\frac{\Delta_x(y)}{L(x)},\quad y\in \xi(x)
\end{equation}
with $L(x)=\int_{\xi(x)} \Delta_x(y)d\mux(y)$.
We now define a measure $\tilde{\mathrm{m}}$ by requiring $\tilde{\mathrm{m}}=\mathrm{m}$ on $\mathcal{B}_{\xi}$ and such its conditionals on $\xi$ are given by \(d\tilde{\mathrm{m}}_{x}=\frac{\Delta_x}{L(x)}d\mu_x=\rho d\mu_x\) (cf.\@ \eqref{eq:condicionalxi}). The proof of \cref{thm:marginalesunicas} will be achieved by establishing that $\mathrm{m}=\tilde{\mathrm{m}}$, and for this it suffices to show that $\mathrm{m}=\tilde{\mathrm{m}}$ on every $\mathcal{B}_{f^{-n}\xi},n\geq 0$, since the previous $\sigma$ - algebras generate $\BM$.
\smallskip
We start by noticing the following: for $q(x)=\tilde{\mathrm{m}}_{x}(f^{-1}\xi(x))$ we have
\begin{align*}
q(x)&=\frac{1}{L(x)}\int_{f^{-1}(\xi(fx))} \Delta_x(f^{-1}fy) \frac{e^{P-\varphi(y)}}{e^{P-\varphi(f^{-1}fy)}}d\mux\\
&=\frac{1}{L(x)}\int_{\xi(fx)} J_x^u(f^{-1}z) e^{\varphi(f^{-1}z)-P}d\mu^u_{fx}(z)=\frac{L(fx)}{L(x)}e^{\varphi(x)-P}\leq 1.
\end{align*}
Thus $\frac{L(fx)}{L(x)}\leq e^{\varphi(x)-P}$ is in $\Lp[1](\mathrm{m})$. A classic lemma of measure theory (cf.\@ proposition $2.2$ in \cite{LedStrelcyn}) then implies
\[
\int \log \frac{L\circ f}{L} d\mathrm{m}=0\Rightarrow \int -\log q(x)d\mathrm{m}(x)=P-\int\varphi d\mathrm{m}.
\]
We have shown the following.
\begin{lemma}\label{lem:lognu}
It holds,
\[
P- \int\varphi d\mathrm{m} =\int -\log\tilde{\mathrm{m}}_{x}(f^{-1}\xi(x))d\mathrm{m}(x)=\int -\log\left(\int_{f^{-1}\xi(x)} \rho(y)d\mux(y)\right) d\mathrm{m}(x).
\]
\end{lemma}
By definition, the measures $\mathrm{m}$ and $\tilde{\mathrm{m}}$ coincide on $\mathcal{B}_{\eta}$; to prove that these two measures coincide it suffices to establish that $\mathrm{m}$ and $\nu$ coincide on $\mathcal{B}_{f^{-1}\xi}$ provided that $\mathrm{m}$ is an equilibrium state, and argue by induction.
Start by noting that since the partition $f^{-1}\xi$ restricted to each atom $\xi(x)$ is countable, we have for $\mathrm{m}\aep(x)$
\[
\frac{d\tilde{\mathrm{m}}}{d\mathrm{m}}\bigg|_{f^{-1}\xi}(x)=\frac{\tilde{\mathrm{m}}_{x}(f^{-1}\xi(x))}{m_x(f^{-1}\xi(x))}.
\]
To prove that $\mathrm{m}=\tilde{\mathrm{m}}$ on $\mathcal{B}_{f^{-1}\xi}$ we will show that $\frac{d\nu}{d\mathrm{m}}\bigg|_{f^{-1}\xi}(x)\equiv 1$ for $\mathrm{m}\aep(x)$. Observe that for $\mathrm{m}\aep(x)$ one has $\mathrm{m}_x(f^{-1}\xi(x))>0$, hence the function $\frac{d\nu}{d\mathrm{m}}\bigg|_{f^{-1}\xi}(x)$ is well defined for $\mathrm{m}\aep(x)$, but not necessarily for $\tilde{\mathrm{m}}\aep(x)$. This point is not emphasized in \cite{LedYoungI}, and some minor extra considerations are required to complete the argument\footnote{This was remarked some time ago by the second author in some private notes.}. For the sake of exactness we will be precise in what follows.
For an atom $\xi(x)$ we write $f^{-1}\xi|\xi(x)=\{A^x_j\}_{j=1}^{\oo}$. We compute
\begin{align}
\nonumber&\int_{\xi(x)} \frac{d\tilde{\mathrm{m}}}{d\mathrm{m}}\bigg|_{f^{-1}\xi}(y)d\mathrm{m}_x(y)=\int_{\xi(x)} \frac{\tilde{\mathrm{m}}_{y}(f^{-1}\xi(y))}{\mathrm{m}_y(f^{-1}\xi(y))}d\mathrm{m}_x(y)
=\int_{\xi(x)}\frac{\tilde{\mathrm{m}}_{x}(f^{-1}\xi(y))}{\mathrm{m}_x(f^{-1}\xi(y))}d\mathrm{m}_x(y)\\
&=\sum_{\substack{j=1\\ \mathrm{m}_x(A^x_j)>0}}^{\oo}\int_{A^x_j}\frac{\tilde{\mathrm{m}}_{x}(A^x_j)}{\mathrm{m}_x(A^x_j)}d\mathrm{m}_x(y)
=\sum_{\substack{j=1\\ \mathrm{m}_x(A^x_j)>0}}^{\oo}\nu_{x}(A^x_j)\leq \sum_{j=1}^{\oo}\nu_{x}(A^x_j)=\nu_x(\xi(x))=1 \label{eq:ineqjensen}\\
&\Rightarrow \int \frac{d\tilde{\mathrm{m}}}{d\mathrm{m}}\bigg|_{f^{-1}\xi}(x)d\mathrm{m}(x)\leq 1 \label{eq:ineqjensendos}
\end{align}
Now we use our hypothesis that $\mathrm{m}$ is an equilibrium state: by \cref{lem:lognu},
\[
\int-\log \mathrm{m}_x(f^{-1}\xi(x))dm(x)=H_\mathrm{m}(f^{-1}\xi|\xi)=P-\int \varphi d\mathrm{m}=\int -\log \tilde{\mathrm{m}}_x(f^{-1}\xi(x))d\mathrm{m}(x)
\]
and thus $\int \frac{d\tilde{\mathrm{m}}}{d\mathrm{m}}\bigg|_{f^{-1}\xi}(x) d\mathrm{m}(x)=0$. Using Jensen's inequality and \eqref{eq:ineqjensendos} we conclude
\begin{align*}
0=\int \log \frac{d\tilde{\mathrm{m}}}{d\mathrm{m}}\bigg|_{f^{-1}\xi}d\mathrm{m}\leq \int \frac{d\tilde{\mathrm{m}}}{d\mathrm{m}}\log\bigg|_{f^{-1}\xi}(x) d\mathrm{m}(x)
\leq \log 1=0.
\end{align*}
Strict convexity of the logarithm function implies then
\[
\frac{d\tilde{\mathrm{m}}}{d\mathrm{m}}\bigg|_{f^{-1}\xi}(x) =1\quad \mathrm{m}\aep
\]
We conclude that the chain of inequalities appearing are equalities, in particular for equation \eqref{eq:ineqjensen}. As a result we have $\mu_x(A_j^x)>0\Rightarrow \mathrm{m}_x(A^x_j)>0$, and this in turn imply that $\mathrm{m}\aep(x)$ it holds $\mathrm{m}_x=\tilde{\mathrm{m}}_x$ on $f^{-1}\xi|\xi(x)$, since these measures coincide on all the atoms and not only on the ones of $\mathrm{m}_x$ positive measure. Consequently, $\mathrm{m}=\tilde{\mathrm{m}}$ on $\mathcal{B}_{f^{-1}\xi}$, and this finishes the general inductive step necessary to prove \cref{thm:marginalesunicas}.
\begin{corollary}
Suppose that $\mathrm{m}\in \PTM{f}{M}$ is an equilibrium state for the potential $\varphi$ and $\xi$ is a SLY partition as above. Then for $\mathrm{m}\aep(x)$ the conditional measures of $\mathrm{m}$ with respect to $\xi$ are given by
\[
\mathrm{m}^{\xi}_x=\tilde{\mathrm{m}}^{\xi}_x=\frac{\Delta_x}{L(x)}d\mux|\xi(x)\quad L(x)=\int_{\xi(x)}\Delta_xd\mux.
\]
In particular the disintegration can be extended to every leaf of $\Fu$ and $\mathrm{m}\in \PTM{f}{M,\mu^u}$.
\end{corollary}
We remark that above we did not assume that $\mathrm{m}=\muf$, and only used the existence of an equilibrium state, which in our context follows from general arguments (since $f$ is h-expansive the function $\mathrm{m}\in \PTM{f}{M}\to h_{\mathrm{m}}(f)$ is upper semi-continuous, therefore has a maximum). Note however that these abstract method do not shield information about the equilibrium state besides its existence.
\subsection{Conditional measures determine equilibrium states}
\label{sub:conditional_mesures_determine_equilibrium_states}
We are interested now in the following converse of \cref{thm:marginalesunicas}.
\begin{theorem}\label{thm:conversemarginal}
Let $\mathrm{m}\in \PTM{f}{M}$ and assume that with respect to some SLY partition $\xi$ we have that
$\mathrm{m}_x<<\mu_x^u$ for $\mathrm{m}\aep(x)$. Then $\mathrm{m}$ is an equilibrium state.
\end{theorem}
\begin{proof}
The proof uses similar arguments to the ones above, and is based in \cite{LedStrelcyn}. Define the measure $\tilde{\mathrm{m}}$ on $M$ by requiring that for $A\subset M$ Borel,
\begin{equation}
\tilde{\mathrm{m}}(A)=\int \mux(A) dm(x).
\end{equation}
Observe that $\tilde{\mathrm{m}}$ is not necessarily finite, but only $\sigma$-finite. By hypothesis $d\mathrm{m}_x=\rho d\mu_x$, thus $\mathrm{m}<<\tilde{\mathrm{m}}$ and moreover $\mathrm{m}\aep(x)$ it holds
\begin{equation}
\frac{d\mathrm{m}}{d\tilde{\mathrm{m}}}=\rho \quad \tilde{\mathrm{m}}\aep(x)
\end{equation}
(see proposition $4.1$ in \cite{LedStrelcyn}). Note that $\mathrm{m}^{f^{-1}\xi}_x=\frac{\mathrm{m}_x(\cdot |f^{-1}\xi)}{\mathrm{m}_x(f^{-1}\xi(x)}$; by an analogous computation as the one carried in the proof of \cref{lem:lognu} we get
\[
\mathrm{m}_x(f^{-1}\xi(x))=\frac{\rho\circ f(x)}{\rho(x)}e^{\varphi(x)-P},
\]
Finally, we compute
\begin{align*}
0\leq h_{\mathrm{m}}(f)&=H(f^{-1}\xi|\xi)=\int-\log \mathrm{m}_x(f^{-1}\xi(x))d\mathrm{m}(x)\\
&=P-\int \varphi d\mathrm{m} +\int -\log\frac{\rho\circ f(x)}{\rho(x)} d\mathrm{m}(x)
\end{align*}
and since $\varphi$ is integrable (bounded), the function $-\log^{-} \frac{\rho\circ f}{\rho}$ is integrable as well. Using again Proposition $2.2$ of \cite{LedStrelcyn} we conclude that $\log \frac{\rho\circ f}{\rho}$ has zero integral, hence
\[
h_{\mathrm{m}}(f)=P-\int \varphi d\mathrm{m}
\]
and $\mathrm{m}$ is an equilibrium state. The proof is complete.
\end{proof}
Joining \cref{thm:marginalesunicas,thm:conversemarginal} we get \cref{thm:condicionalesdeterminanee}. Below we list some of its consequences.
\begin{corollary}\label{cor:condicionales}
If $\mathrm{m}$ is an equilibrium state then $\mathrm{m}$ has conditionals equivalent to $\mu^s,\mu^u$.
\end{corollary}
\begin{proof}
The set of equilibrium states for $f$ coincides with the one for $f^{-1}$.
\end{proof}
Since the measures $\mux$ are positive on open sets, proposition $4.1$ in \cite{Katok1996} implies:
\begin{corollary}\label{cor:entropos}
For every equilibrium state $\mathrm{m}$ for $\varphi$ it holds $h_{\mathrm{m}}(f)>0$
\end{corollary}
Finally, we note that the previous theorem gives another proof of the fact that $\muf$ is an equilibrium state (and $P$ is the topological pressure), as the conditionals of $\muf$ are absolutely continuous with respect to $\mu^u$.
\section{Kolmogorov and Bernoulli properties}
\label{sec:finer_ergodic_properties}
The aim of this part is to establish the Bernoulli property for equilibrium states. Our arguments are written in some more generality than the context that we have been working, with the intention that they may be applicable in other situations.
\subsection{Kolmogorov processes}
\label{sub:kolmogorov_processes}
We will establish first that $(f,\muf)$ has the Kolmogorov property, and in particular is mixing of all orders. The proof is based on the methods originally developed by Grayson, Pugh and Shub for establishing the stable ergodicity of the time-one map of the geodesic flow on manifolds of constant sectional curvature \cite{StErgGeo}. We will rely heavily on the technology developed in \cite{StableJulienne} and its subsequent refinements. As a guide, we use the general lines of Section $4$ in \cite{AccStaErg}.
\smallskip
Fix $\mathrm{m}\in \PTM{f}{M}$ and denote $\mathrm{Pin}(f,\mathrm{m})$ the Pinsker $\sigma$-algebra of the system, i.e.\@ the $\sigma$-algebra generated by
\[\{\xi \text{ countable measurable partition }: H_{\mathrm{m}}(\xi)<\oo, h_{\mathrm{m}}(f;\xi)=0 \}.
\] We use the following characterization of Kolmogorov-systems, although the reader may want to take it as the definition of the Kolmogorov property.
\begin{theorem}[Rohklin-Sinai]\label{thm:RohklinSinai}
The system $(f,\mathrm{m})$ has the Kolmogorov property if and only if $\mathrm{Pin}(f,\mathrm{m})=\{\emptyset,M\}\modo$.
\end{theorem}
We say that a subset $X\subset M$ is $s$-saturated ($u$-saturated) if it consists of whole stable (respectively, unstable) leaves. If it is both $s$- and $u$-saturated we say that it is bi-saturated. Given a point $x\in M$ its accessibility class $\mathrm{ACC}(x)$ is the largest bi-saturated set containing $x$. Analogously, we say that $X\subset M$ is $\mathrm{m}$-essentially $s$-saturated ($u$-saturated) if there exists a $s$-saturated ($u$-saturated) Borel set $X_0$ so that $\mathrm{m}(X\Delta X_0)=0$. By \cite{LedYoungI} we have:
\begin{theorem}[Ledrappier-Young]
If $X\in \mathrm{Pin}(f,\mathrm{m})$ then $X$ is both $\mathrm{m}$-essentially $s$-saturated and $\mathrm{m}$-essentially $u$-saturated.
\end{theorem}
What we will show is that given $X\in\BM[M]$ is $\mathrm{m}$-essentially $s$-saturated, then it coincides almost everywhere with a $s$-saturated set $D(X)$ which is relevant from the point of view of the measure (it consists of some sort of density points, see below), and then use the minimality to conclude that this set is either full or null for $\mathrm{m}$. As a remark for the reader versed in Pugh-Shub's program, we point out we are not assuming $\mathrm{m}$-essentially accessibility\footnote{The map $f$ is $\mathrm{m}$-essentially accessible if any bi-saturated set has measure zero or one.}, and we will not establish that if $X$ is both $\mathrm{m}$-essentially $s$ and $u$ saturated then $D(X)$ is bi-saturated. Instead, we will use the minimality of the strong foliations to spread $D(X)$ to the whole manifold.
From now on we will assume that $\mathrm{m}$ is positive on open sets.
\begin{definition}
We say that the family $\mathscr{E}=\{\{E_n(x)\}_{n\in \mathbb{N}}\}_{x\in M}$ is a regular Vitali\footnote{More general definitions of Vitali bases (not regular) are available in the literature (cf. \cite{StableJulienne}). The definition presented here is enough for our purposes.} basis if for every $x\in M$ the family $\{E_n(x)\}_{n\in \mathbb{N}}$ is a local basis of neighborhoods at $x$ such that
\begin{enumerate}
\item each $E_n(x)$ is a pre-compact open set.
\item $\diam(E_n(x))\xrightarrow[n\mapsto\oo]{} 0$.
\end{enumerate}
\end{definition}
\begin{definition}
Let $\mathscr{E}$ be a regular Vitali basis and $X\in\BM[M]$. We say that $x\in M$ is a $\mathscr{E}$-density point of $X$ if
\begin{align*}
\lim_{n\rightarrow\oo}\frac{m(X\cap E_n(x))}{m(E_n(x))}=1.
\end{align*}
We denote by $\mathscr{D}(X;\mathscr{E})$ the set of $\mathscr{E}$-density points of $X$.
\end{definition}
In particular, density points for the basis $\{D(x;\frac{1}{2^n})\}_n$ will be called $B\gui L$ density points\footnote{$B-L$ stands for Besitcovich-Lebesgue}. It is standard that for any Borel set $X$, $\mathrm{m}$-almost every point of $X$ is a $B\gui L$ density point of $X$. To prove the Kolmogorov-property in the essentially accessible setting, it suffices to establish that if $X\subset M$ is both $\mathrm{m}$-essentially $s$-saturated and $m$-essentially $u$-saturated then the set of its $B-L$ density points is bi-saturated. This is the approach followed in \cite{AccStaErg}, where the Lebesgue measure is considered.
In our case however, we do not have as much control on the properties of the $\varphi$ equilibrium state as one has with the Lebesgue measure, and in particular we do not know of to control $B\gui L$ density points. Instead, we will work with density points corresponding to some dynamical defined Vitali bases called \emph{juliennes}, and show that under suitable hypotheses
\begin{enumerate}
\item any $X\in\BM[M]$ coincides $\mathrm{m}$-almost everywhere with the set of its julienne density points.
\item If $X$ is $\mathrm{m}$-essentially $s$-saturated then the set of its julienne density points is $s$-saturated.
\end{enumerate}
\smallskip
\noindent\textbf{Convention:} From now on we specialize in the case $\mathrm{m}=\tilde{\mathrm{m}}_{\scriptstyle\varphi}$.
\smallskip
For some simplifications in the notation we choose to work with $\tilde{\mathrm{m}}_{\scriptstyle\varphi}$, but as previously noted the properties of $\tilde{\mathrm{m}}_{\scriptstyle\varphi}$ are completely analogous to the ones of $\muf$, and we will refer directly to them without further clarification.
\smallskip
Some finer control in the local holonomies are now required, specifically that they are H\"older continuous. See Theorem A of \cite{HolFol}.
\begin{theorem}[Pugh-Shub-Wilkinson]\label{thm:HolFol}
There exists $C_H>0,\vartheta\in (0,1)$ such that for every $x,y,z\in M$ with $y\in \Ws{x;\epsilon},z\in \Wu{x,\epsilon}$ the locally defined holonomy maps $g^s:\Wcu{x}\rightarrow\Wcu{y},g^u:\Wcs{x}\rightarrow \Wcs{y}$ are $(C_H,\vartheta)$-H\"older.
\end{theorem}
Choose numbers $0<\varepsilon<\epsilon,0<\sigma<1$ such that $d(x,y)<\varepsilon\Rightarrow B(y,\varepsilon)\subset B(x,\epsilon)$, $\lambda^{\vartheta}<\sigma$ and define the $u,s,c,cu,cs,ucs,scu$-juliennes as
\begin{align*}
J^u_n(x)&:=f^{-n}(\Wu{f^nx;\varepsilon})\\
J^s_n(x)&:=f^n(\Ws{f^{-n}x;\varepsilon})\\
B^c_n(x)&:=\Wc{x;\sigma^n}\\
J^{cu}_n(x)&:=\bigcup_{\mathclap{y\in B^c_n(x)}}\ J^u_n(y)\\
J^{cs}_n(x)&:=\bigcup_{\mathclap{y\in B^c_n(x)}}\ J^s_n(y)\\
J^{ucs}_n(x)&:=\bigcup_{\mathclap{y\in J_{n}^{cs}(x)}}\ J^u_n(y)\\
J^{scu}_n(x)&:=\bigcup_{\mathclap{y\in J_{n}^{cu}(x)}}\ J^s_n(y).
\end{align*}
\begin{remark}
The definition of juliennes for general partially hyperbolic maps is more involved, since no a priori knowledge of the action of $f$ on the center direction is assumed. In our case, the fact that $f$ is a center isometry simplifies all the definitions.
\end{remark}
By \cref{cor:gibbslineal} we obtain.
\begin{lemma}\label{lem:controljuliana}
There exists constants $c_1(\ep),c_2(\ep)>0$ such that for every $x\in M$ it holds
\[
c_1\leq \frac{\nsx(J_n^s(x))}{e^{\SB\circ f^{-n}(x)-nP}}\leq c_2.
\]
\end{lemma}
From now on we omit the reference of the constants respect to $\ep$, which is considered to be fixed and small with respect to $\ep_0$. We then have the following.
\begin{corollary}\label{cor:controljuliana}
There exist $L_1(\ep),L_2(\ep)>0$ such that for every $x\in M$, for every $n\geq0$ it holds that $y,z\in J^{cu}_n(x)$ implies
\[
L_1\leq \frac{\nsx(J_n^s(y))}{\nsx(J_n^s(z))}\leq L_2.
\]
\end{corollary}
\begin{proof}
Arguing as in the proof of \cref{lem:Bowenproperty} we deduce the existence of some $K'>0$ such that for every $x\in M$ and $n\geq0$,
\[
y,z\in J^{cu}_n(x)\Rightarrow |\SB\circ f^{-n}y-\SB\circ f^{-n}z|<K';
\]
this relies on the fact the central size of $J^{cu}$ is exponentially small. The claimed property follows now from the previous lemma.
\end{proof}
Using the above and the product structure of $\mathrm{m}$ (cf.\@ \cref{pro:estructuraproducto}) we deduce the following.
\begin{corollary}\label{cor:productstructurejulienne}
There exist $L_3(\ep),L_4(\ep)>0$ such that for every $x\in M$, for every $n\geq0$ it holds
\[
L_3\leq\frac{m(J^{scu}_n(x))}{\msx(J^{s}_n(x))\cdot \mcux(J^{cu}_n(x))}\leq L_4.
\]
\end{corollary}
\smallskip
The key property of juliennes is that they behave in a controlled matter under holonomies.
\begin{proposition}\label{pro:holjulienne}
There exists $k\geq 1$ such that for every $x,x'\in M,x'\in \Ws{x,\epsilon}$ the corresponding stable holonomy $\hs:\Wcu{x,\epsilon_0}\rightarrow\Wcu{x',\epsilon_0}$ satisfies
\[
J_{n+k}^{cu}(x')\subset \hs(J_{n}^{cu}(x))\subset J_{n-k}^{cu}(x)
\]
\end{proposition}
This is completely analogous to proposition $B.8$ of \cite{AccStaErg}. As our juliennes differ from the ones used in that work, we present the proof.
\begin{proof}
Take $z\in B^c_n(x), y\in J^u_n(x)$ and consider $z'=\hs(z), y'=\hs(y)$. Then $d(f^ny,f^nz)<\varepsilon$, and thus
\begin{align*}
d(f^ny',f^nz')\leq d(f^ny',f^ny)+d(f^ny,f^nz)+d(f^nz,f^nz')
\leq \lambda^n\left(d(y,y')+d(z,z')\right)+d(f^ny,f^nz)\leq 2\epsilon
\end{align*}
if $n$ is sufficiently large. Define $w'=\Wu{y',\ep_0}\cap \Wc{x',\ep_0}$. Since the bundles mutually orthogonal, and since $\epsilon_0$ is small, there exists a constant $C_a>0$ such that
\[
d(f^ny',f^nw'),d(f^nw',f^nz')\leq C_a\epsilon.
\]
This implies at once that there exists $k\geq 1$ independent of $x$ such that if $n\geq k$, $y'\in J^u_{n-k}(w')$. On the other hand, $d(y,z)\leq \lambda^n\varepsilon$ hence by \cref{thm:HolFol}
\[
d(z',w')\leq C_a\varepsilon^{\theta}\lambda^{n\vartheta}.
\]
Using \cref{lem:holisometry} we obtain that for sufficiently large $k$
\[
d(x',w')\leq d(x,z)+d(z',w')<\sigma^n+C_a\varepsilon^{\vartheta}\lambda^{n\vartheta}<\sigma^{n-k}.
\]
This shows that $y'=\hs(y)\in J^{cu}_{n-k}(x')$ and finishes the proof of the second inclusion. The first one is obtained similarly by using $(\hs)^{-1}$.
\end{proof}
Now we can establish a central proposition.
\begin{proposition}\label{pro:Xsusat}
If $X$ is $\mathrm{m}$-essentially $s$-saturated , then the set of its $J^{scu}$-density points is $s$-saturated.
\end{proposition}
For the proof, we will employ the following Lemma.
\begin{lemma}\label{lem:Xsusat}
Let $X_s$ be an s-saturated set and $x\in M$. Then
\[
\lim_{n\mapsto \oo}\frac{\mathrm{m}(X_s\cap J_n^{scu}(x))}{\mathrm{m}(J_n^{scu}(x))}=1\Leftrightarrow \lim_{n\mapsto \oo}\frac{\mcux(X_s\cap J_n^{cu}(x))}{\mcux(J_n^{cu}(x))}=1.
\]
\end{lemma}
The proof is not too hard, and is direct consequence of \cref{cor:productstructurejulienne}. We omit the proof, and refer the reader to proposition $2.7$ of \cite{ErgPH}, where a completely analogous statement is proven.
\begin{proof}[Proof of \cref{pro:Xsusat}]
Consider $X_s$ an $s$-saturated set such that $\mathrm{m}(X\Delta X_s)=0$. Let $x$ be a $J^{scu}$-density point of $X$ and let $\hs$ be a (local) stable holonomy sending $x$ to $y$. Using the previous proposition we deduce
\[
\mcux[y](\hs(J_{n-k}^{scu}(x)\cap X_s))\leq \mcux[y](J_n^{scu}(y)\cap X_s)\leq \mcux[y](\hs(J_{n+k}^{scu}(x)\cap X_s)),
\]
and since at this scale $(\hs)^{-1}\mcux[y]$ is uniformly comparable to $\mcux$, we have
\begin{align*}
1&=\lim_{n\mapsto \oo}\frac{\mathrm{m}(X_s\cap J_n^{scu}(x))}{\mathrm{m}(J_n^{scu}(x))}=\lim_{n\mapsto \oo}\frac{\mcux(X_s\cap J_n^{scu}(x))}{\mcux(J_n^{scu}(x))}=\lim_{n\mapsto \oo}\frac{\mcux[y](\hs(X_s\cap J_n^{scu}(x))}{\mcux[y](\hs(J_n^{scu}(x))}\\
&=\lim_{n\mapsto \oo}\frac{\mcux[y](X_s\cap J_n^{scu}(y))}{\mcux[y](J_n^{scu}(y))}=\lim_{n\mapsto \oo}\frac{\mathrm{m}(X_s\cap J_n^{scu}(y))}{\mathrm{m}(J_n^{scu}(y))},
\end{align*}
i.e.\@ $y$ is also a point of $J^{scu}$-density of $X_s$.
\end{proof}
We will show now that any measurable $X$ coincides $\mathrm{m}$-almost everywhere with $D(X;J^{scu})$.
\begin{definition}
Let $\mathrm{m}\in \PM[M]$ be positive on open sets. A regular Vitali basis $\mathscr{E}$ is $\mathrm{m}$-engulfing if there exists $L>0$ with the following property: for every $x\in M,n\in\Nat$ there exists a closed set $\widehat{E}_n(x)$ such that
\begin{enumerate}
\item $E_{n}(x)\subset \widehat{E}_n(x)$ and $\mathrm{m}(\widehat{E}_n(x))\leq L\cdot \mathrm{m}(E_n(x))$.
\item If $n'\geq n,y\in M$ and $E_{n'}(y)\cap E_n(x)\neq\emptyset$ then $E_{n'}(y)\subset \widehat{E}_n(x)$.
\item Given $N_0$ there exists $\delta>0$ such that $\mathrm{m}(\clo{E_n(x)})<\delta$ implies $n\geq N_0$.
\end{enumerate}
\end{definition}
\begin{theorem}
If $\mathscr{E}$ is a $\mathrm{m}$-engulfing basis then for every $X\in \BM[M]$ the set of its $\mathscr{E}$-density points coincide almost everywhere with $X$.
\end{theorem}
See theorem $3.1$ in \cite{StableJulienne}.
\begin{proposition}\label{pro:jucsisengulfing}
$J^{scu}$ is $\mathrm{m}$-engulfing.
\end{proposition}
\begin{proof}
By uniformity of the contraction-expansion rates and the definition of $ucs$-juliennes, one sees that there exists $l>0$ such that for $n\geq l$ it holds $y\in J^{scu}_{n}(x)$ then $\clo{J^{scu}_{n+l}(y)}\subset J^{scu}_{n}(x)$. Arguing as in \cref{cor:controljuliana} we deduce the existence of $D=D(l)$ such that for every $n\geq l$,
\[
\frac{\mcux[x]((J^{cu}_n(x))}{\mcux[x]((J^{cu}_{n+l}(x))}\leq D(l).
\]
Define then $\widehat{J}^{ucs}_n(x):=\clo{J^{ucs}_{n-l}(x)}$ and observe that due to \cref{cor:productstructurejulienne} and the inequality above one can deduce the existence of $\hat{D}>0$ such that
\[
\mathrm{m}(\widehat{J}^{ucs}_n(x))\leq D\mathrm{m}(J^{ucs}_{n}(x)).
\]
This implies the $\mathrm{m}$-engulfing property.
\end{proof}
The following is now clear.
\begin{corollary}\label{cor:XcoincideconsuspdLebesgue}
Every measurable set $X$ coincides $m$-almost everywhere with $D(X;J^{scu})$.
\end{corollary}
We have thus established that if $X$ is $\mathrm{m}$-essentially $s$-saturated then it coincides almost everywhere with the
$s$-saturated set $D(X;J^{scu})$; This allow us to conclude the $K$-property
\begin{theorem}\label{thm:propiedadK}
The system $(f,\muf)$ is Kolmogorov.
\end{theorem}
\begin{proof}
Take $X\in\mathrm{Pin}(f,\mathrm{m})$ of positive measure; then $D(X;J^{scu})$ has positive measure and is $s$-saturated. Assuming that $X^c$ also has positive measure, we deduce that $D(X^c;J^{scu})$ also has positive measure. But since $\Fs$ is minimal arbitrarily close to points in $D(X^c;J^{scu})$ one can find points in $D(X;J^{scu})$, which contradicts the definition of density point. Therefore $\mathrm{m}(X)=1$ and $\mathrm{Pin}(f,\mathrm{m})=\{\emptyset, M\}\modo$, and the system is Kolmogorov, by \cref{thm:RohklinSinai}.
\end{proof}
\subsection{Bernoulli processes}
Using the previous theorem, we will prove now that $(f,\muf)$ is isomorphic to a Bernoulli shift. We rely heavily on the machinery developed by D. Ornstein and B. Weiss to establish the similar statement for the case where $f$ is the time-one map of an Anosov flow and $\muf$ is the Lebesgue measure \cite{GeoBernoulli} (see also \cite{RatGibbs}). For the sake of completeness we repeat here the main concepts and arguments from that work that we will use. Nonetheless, some familiarity with the Ornstein and Weiss article is desirable to follow this section.
Throughout this part all probability spaces are standard, and all partitions of them are considered to be finite and ordered. If $\parP$ is a partition its $i$-th atom will be denoted $P_i$. For partitions $\parP,\parP[P']$ of the probability space $(X,\mathcal{B}_X,\mu)$ having the same number $k$ of atoms, its partition distance is defined as
\begin{equation*}
|\parP-\parP[P']|:=\sum_{i=1}^k \mu(P_i\vartriangle P_i').
\end{equation*}
\begin{definition} Let $\{\parP[P_j]\}_1^n,\{\parP[Q_j]\}_1^n$ be partitions of the probability spaces $(X,\mathcal{B}_X,\mu_X)$, $(Y,\mathcal{B}_{Y},\mu_Y)$ having the same number of atoms. Then its $\overline{d}$-distance is
\begin{equation*}
\overline{d}(\{\parP[P_j]\}_1^n,\{\parP[Q_j]\}_1^n):=\inf_{g_1, g_2}\frac{1}{n}\sum_1^n|g_1\parP[P_j]-g_2\parP[Q_j]|
\end{equation*}
where $g_1, g_2$ are isomorphisms between $X, Y$ and some fixed Lebesgue space $(Z,\mathcal{B}_Z,\mu_Z)$ (say, the interval).
\end{definition}
To clarify the the meaning of this definition we observe the following. Given a partition $\parP[P]$ of $X$ and $x\in X$ the $\parP$-name of $x$ is simply the atom where is contained. Similarly, for a sequence of partitions $\{\parP[P_j]\}_1^n$ having the same number of elements the $\{\parP[P_j]\}_1^n$-name of $x\in X$ is the $n$-tuple $(P_{j_1(x)},\ldots,P_{j_n(x)})$ such that $x\in \cap_{k=1}^n P_{j_k(x)}$. It is straightforward to verify that for partitions $\parP[P], \parP[P']$ of $X$ the partition distance $|\parP-\parP[P']|$ is simply (twice) the expected value of the function $g_{\parP[P],\parP[P']}:X\rightarrow \{0,1\}$ such that
\[
g_{\parP[P],\parP[P']}(x)=1\Leftrightarrow \text{the }\parP[P], \parP[P']\text{ names of }x\text{ are different}.
\]
Consequently, we have that $\overline{d}(\{\parP[P_j]\}_1^n,\{\parP[Q_j]\}_1^n)\leq r$ if and only if there is a way to re-code simultaneously $\{\parP[P_j]\}_1^n,\{\parP[Q_j]\}_1^n$ such that the average of the expected value of points having different $n$-name with respect to the encoding is less than equal to $r$. The following definition and lemma give an important tool to estimate the size of the $\overline{d}$ distance between two sequence of partitions.
\begin{definition}
A measurable map $g:(X,\mu_X)\rightarrow (Y,\mu_Y)$ is said to be $\epsilon$-measure preserving if there exists $\mathscr{Ex}\subset X$ such that $\mu_X(\mathscr{Ex})<\epsilon$ and with the property that $B\subset Y, g^{-1}B\cap \mathscr{Ex}=\emptyset$ implies
\[
\left|\frac{\mu_Y(B)}{\mu_X(g^{-1}B)}-1\right|< \epsilon.
\]
In this case $\mathscr{Ex}$ is the exceptional set of $g$.
\end{definition}
\begin{lemma}\label{lem:lemanombres}
Let $\{\parP[P_j]\}_1^n,\{\parP[Q_j]\}_1^n$ be partitions of the probability spaces $(X,\mathcal{B}_X,\mu_X)$, $(Y,\mathcal{B}_{Y},\mu_Y)$ having the same number of atoms, and assume that there exists $g:(X,\mu_X)\rightarrow (Y,\mu_Y)$\ $\epsilon$-measure preserving with exceptional set $\mathscr{Ex}$ such that
\[
x\not\in \mathscr{Ex}\Rightarrow \frac{1}{n}\#\{1\leq i\leq n: \parP[P_j]-\text{name of }x\neq \parP[Q_j]-\text{name of }gx\}\leq \epsilon.
\]
Then $\overline{d}(\{\parP[P_j]\}_1^n,\{\parP[Q_j]\}_1^n)\leq 16\epsilon$.
\end{lemma}
See lemma $1.3$ in \cite{GeoBernoulli} for the proof. We will usually apply the previous lemma with $Y\subset X$ a subspace, where $\mu_Y:=\mu_X(\cdot|Y)$ and $\parP[Q_j]=\parP[P_j]|Y$.
\smallskip
The following definition and theorem are central for what follows.
\begin{definition}
Let $f:(X,\mathcal{B}_X,\mu_X)\rightarrow (X,\mathcal{B}_X,\mu_X)$ be an automorphism and $\parP$ a finite partition of $X$. We say that $\parP$ is very weak Bernoullian (VWB)\footnote{In \cite{GeoBernoulli} the VWB condition is defined with the roles of past and future interchanged. This however is a minor difference, since if a system is Bernoulli then its inverse is Bernoulli as well.} if given $\epsilon>0$ there exists $N_0$ such that for all $N'\geq N\geq N_0$ there exist a family of atoms $\mathscr{G}\subset \bigvee_{N'}^{N} f^{-i} \parP$ satisfying
\begin{itemize}
\item $\mu_X(\mathscr{G})>1-\ep$;
\item for $n\geq 0$ it holds \(A\in \mathscr{G}\Rightarrow \overline{d}(\{f^{j}\parP\}_1^n,\{f^{j}\parP|A\}_1^n)\leq \epsilon\).
\end{itemize}
\end{definition}
\begin{theorem}[Ornstein]\label{thm:Ornstein}
Let $f:(X,\mathcal{B}_X,\mu_X)\rightarrow (X,\mathcal{B}_X,\mu_X)$ be an automorphism.
\begin{enumerate}
\item If $\parP$ is VWB then $(X,\bigvee_{-\oo}^{\oo}f^n\parP,\mu_X)$ is isomorphic to a Bernoulli shift.
\item Suppose that there exists an increasing family of $f$-invariant $\sigma$-algebras $\{\mathcal{A}_n\}_{n=1}^{\oo}$ satisfying:
\begin{enumerate}
\item $\mathcal{A}_n\nearrow \mathcal{B}_X$.
\item For every $n$, $(X,\mathcal{A}_n,\mu_X)$ is isomorphic to a Bernoulli shift.
\end{enumerate}
Then $(X,\mathcal{B}_X,\mu_X)$ is isomorphic to a Bernoulli shift.
\end{enumerate}
\end{theorem}
From now on we will specialize in our case of interest $(X,\mathcal{B}_X,\mu_X)=(M,\mathcal{B}_M,m=\tilde{\mathrm{m}}_{\scriptstyle\varphi})$. The idea is to show that there exist VWB partitions of arbitrarily small diameter and use the previous theorem to conclude that the process $(f,\tilde{\mathrm{m}}_{\scriptstyle\varphi})$ is Bernoulli. We fix a family of $cu$-boxes $\mathscr{V}=\{V_i\}_{i=1}^R$ together with their corresponding $cu$-transversals $\{W_i\}_{i=1}^R$, and define $\tilde{\mathrm{m}}_{\scriptstyle\varphi}$ using these families.
Consider a partition $\parP$ such that each atom $P\in \parP$ is a proper set with diameter smaller than the Lebesgue number of $\{V_i\}_{i=1}^R$, and furthermore its boundary $\partial P_i$ is piecewise differentiable. Clearly there exists such partitions of arbitrarily small diameter and thus it is enough to prove that $\parP$ as above is VWB.
\begin{definition}
Let $A,B\subset M$ with $B$ a $cu$-box. We say that $x\in A\cap B$ is a point of ($s-$)tubular intersection if the stable plaque of $B$ containing $x$ is completely contained in $A\cap B$. If every point of $A$ is a point of tubular intersection with $B$ we say that $A$ is a $s$-band.
\end{definition}
\begin{lemma}\label{lem:sbandas}
Let $C$ be a $cu$-box and $\delta>0$. Then there exists $N_1$ such that for all $N'>N\geq N_1$ one can find $\mathscr{H}\subset \bigvee_{N}^{N'}f^n\mathcal{P}$ with $\mathrm{m}(\bigcup_{H\in\mathscr{H}} H)\geq 1-\delta$ satisfying that if $A\in H$ there exists $E\subset A$ with the properties
\begin{enumerate}
\item $\mathrm{m}(E|A)\geq 1-\delta$.
\item $E\cap C$ is a $s$-band.
\end{enumerate}
\end{lemma}
Compare lemma $2.1$ in \cite{GeoBernoulli}. Note that in general the disintegration of $\mathrm{m}$ into local stable manifolds is much worse behaved that the case considered by Ornstein and Weiss (where $\mathrm{m}$ is Lebesgue).
\begin{proof}
Define the set
\[
\mathcal{H}:=\{x:\lim_{n\mapsto\oo}\frac{\SB(x)}{n}=\int \varphi dm\}.
\]
The system $(f,\mathrm{m})$ is ergodic and $\varphi$ is continuous, thus $\mathcal{H}$ is a full measure $s$-saturated set in $M$. By Egoroff's theorem we can find $\mathcal{H}_0\subset \mathcal{H}$ with $\mathrm{m}(\mathcal{H}_0)\geq 1-\frac{\delta^2}{4}$ where the above convergence is uniform, hence there exists $N_0$ such that for all $N\geq N_0$, for all $x\in \mathcal{H}_0$ it holds
\begin{equation*}
\SB(x)-nP=\SB(x)-n(\int \varphi d\mathrm{m} + h_{\mathrm{m}}(f))<-\frac{nh_{\mathrm{m}}(f)}{2}
\end{equation*}
(recall that $h_{\mathrm{m}}(f)>0$ by \cref{cor:entropos}). Let $\zeta:=\max\{\diam \Ws{x,C}:x\in C\}$, and for $n\geq N_1$ define
\begin{align*}
B_n&=\cup_i\{x\in C\cap f^{-n}P_i: x\text{ is not of tubular intersection}\}\\
&=\cup_i\{x\in f^{n}C\cap P_i: x\text{ is not of tubular intersection}\}.
\end{align*}
As $f^nC$ is a $cu$-box, $B_n\subset \bigcup_{i}\bigcup_{x\in \partial P_i} f^n \Ws{f^{-n}x,\zeta}$. Now by formula \eqref{eq:invariancianu}, \cref{lem:comparacionentretamanhos} and the fact that $\mathcal{H}_0$ is $s$-saturated we deduce the existence of a constant $d=d(\zeta)>0$ such that
\[
\forall x\in \partial P_i\cap \mathcal{H}_0,\quad \nsx[x_w](f^n \Ws{f^{-n}x,\zeta})\leq e^{\SB(x_w)-nP}\leq de^{-n\frac{h_{\muf}(f)}{2}}
\]
where $P_i\subset V=V(P_i)$. On the other hand, each $\partial P_i$ consists of finitely many differentiable pieces, and each piece has a well defined angle with the $cu$-plaques of $V$. By \cref{thm:marginalesunicas}, each piece intersects almost every $cu$-plaque in a set of zero measure, and since there are finitely many atoms we conclude that for some constant $d'$ it holds
\begin{equation*}
\mathrm{m}(B_n\cap \mathcal{H}_0)\leq d'e^{-n\frac{h_{m}(f)}{2}}.
\end{equation*}
Define $\mathcal{B}:=\cup_{N_1}^{\oo}\mathcal{B}_n$ where $N_1>N_0$ is such that $\mathrm{m}(\mathcal{B}\cap \mathcal{H}_0)<\frac{\delta^2}{4}$.
\smallskip
\noindent\textbf{Claim 1:} For all $N'>N\geq N_1$ there exists $\mathscr{H}_1\subset \bigvee_{N}^{N'}f^{-n}\mathcal{P}$ such that $\mathrm{m}(\cup_{A\not\in \mathscr{H}_1}A)\leq \frac{\delta}{2}$ and $A\in \mathscr{H}_1 \Rightarrow \mathrm{m}(\mathcal{H}_0^c|A)< \frac{\delta}{2}$.
Otherwise $\mathrm{m}(\mathcal{H}_0^c)\geq \frac{\delta^2}{4}$, contradicting the choice of $\mathcal{H}_0$.
\smallskip
\noindent\textbf{Claim 2:} For all $N'>N\geq N_1$ there exists $\mathscr{H}_2\subset \bigvee_{N}^{N'}f^{-n}\mathcal{P}$ such that $\mathrm{m}(\cup_{A\not\in \mathscr{H}_2}A)\leq \frac{\delta}{2}$ and $A\in \mathscr{H}_2 \Rightarrow \mathrm{m}(\mathcal{B}\cap \mathcal{H}_0|A)< \frac{\delta}{2}$.
Otherwise $\mathrm{m}(\mathcal{B}\cap \mathcal{H}_0)\geq \frac{\delta^2}{4}$, contradicting the definition of $\mathcal{B}$.
\smallskip
Let $\mathscr{H}=\bigvee_{N}^{N'}f^n\mathcal{P}\cap \mathscr{H}_1^c\cap \mathscr{H}_2$. Then $\mathrm{m}(\mathscr{H})\geq 1-\delta$ and for $A\in \mathscr{H}$ the existence of the $cu$-band $E\subset A$ satisfying the conclusion of the lemma is easily deduced from the above two claims.
\end{proof}
Since $(f,\mathrm{m})$ has the $K$-property we have the following.
\begin{lemma}\label{lem:Kconsecuencia}
Let $C$ be a $cu$-box and $\delta>0$. Then there exists $N_2\geq N_1$ such that for any $N'\geq N\geq N_2$ there exists $\mathscr{G}\subset \bigvee_{N}^{N'}f^n\parP$ satisfying
\begin{itemize}
\item $\mathrm{m}(\cup_{A\in \mathscr{G}})\geq 1-\delta$;
\item $A\in \mathscr{G}\Rightarrow |\mathrm{m}(C|A)-\mathrm{m}(C)|<\delta$.
\end{itemize}
\end{lemma}
We are ready to prove the analogue of the main lemma of \cite{GeoBernoulli}.
\begin{proposition}\label{pro:mainOW}
Given $\delta>0$ there exists $\rho>0$ such that if $C$ is a $cu$-box with diameter less than $\rho$ and $E\subset C$ is a $s$-band then there exists a one-to-one mapping $\theta:E\rightarrow C$ such that
\begin{enumerate}
\item $\theta$ is measure preserving.
\item For all $n\geq 0,x\in E$ it holds $d(f^{-n}x,f^{-n}\theta(x))\leq \delta$.
\end{enumerate}
\end{proposition}
\begin{proof}
Consider $\rho>0$ smaller than the Lebesgue number of the covering $\{V_i\}_{i=1}^R$. If $C$ is a $cu$-box with $diam C\leq \rho$ then it is contained in some $V=V_i$, and hence using \cref{lem:independencebox} it is no loss of generality to assume $C=V$.
Now observe that if $g^s=\hs_{x,y}:W_x\rightarrow W_y$ denotes the Poincar\'e transport along stables for $x,y\in V,y\in \Ws{x,B}$ we can multiply $g^s$ by a non-negative continuous function to obtain a map
$\widehat{g}^s:W_x\rightarrow W_y$ which sends $\mcux|W_x$ to $\mcux[y]|W_y$ (cf.\@ \cref{pro:Jacobinanohol}). Given $E\subset B$ an $s$-band define $\theta|E\cap W_x\rightarrow W_x$ by any measure preserving map, and $\theta| E\cap W_y:=\widehat{g}^s_{x,y}\circ \theta|E\cap W_x\circ \widehat{g}^s_{y,x} $. The map $\theta:E\rightarrow B$ is a well defined one to one map that sends $\mathrm{m}|E$ to $\mathrm{m}|B$. Moreover it preserves the $cu$-plaques of $B$, thus it suffices to take $C$ such that their $cu$-plaques have diameter smaller than $\delta$.
\end{proof}
Now we can apply lemma $2.3$ of \cite{GeoBernoulli} to obtain the following.
\begin{lemma}
Let $\varepsilon,\delta>0$ be given. Then there exists $N$ with the property that for all $N'\geq N$ one can find $\mathscr{G}\subset \bigvee_{N}^{N'}f^{-n}\parP$ satisfying:
\begin{enumerate}
\item $\mathrm{m}(\mathscr{G})\geq 1-\varepsilon$;
\item for $A\in \mathscr{G}$ there exists $E\subset A$ and a one-to-one mapping $\theta: A\rightarrow M$ such that the following is true
\begin{enumerate}
\item $\mathrm{m}(E|A)\geq 1-\varepsilon$;
\item $\theta$ is $\varepsilon$-measure preserving;
\item for all $n\geq 0,x\in E$ it holds $d(f^{-n}x,f^{-n}\theta(x))\leq \delta$.
\end{enumerate}
\end{enumerate}
\end{lemma}
For convenience of the reader we spell the proof.
\begin{proof}
Consider $0<\rho$ corresponding to $\delta$ in the previous proposition. Using Betsicovich's covering lemma we construct a partition $\{U_1,\ldots,U_l\}$ such that
\begin{enumerate}
\item each $U_j$ is a cu-box, $1\leq j\leq u$ of diameter less than $\rho$.
\item $m(U_0)\leq \frac{\varepsilon}{10}$.
\end{enumerate}
Then using \cref{lem:sbandas,lem:Kconsecuencia} we deduce that there exists $N$ such that for all $N'\geq N$ we can find $\mathscr{G}\subset \bigvee_{N}^{N'}f^{-n}\parP$ with $\mathrm{m}(\cup_{A\in\mathscr{G}}A)\geq 1-\varepsilon$ and such that for $A\in \mathscr{G}, 1 \leq j \leq l$ there exists an $s$-band $E\subset A$ of relative measure bigger than $1-\varepsilon$ with the property that if $1\leq j\leq u$ it holds
\begin{enumerate}
\item $\sum_1^l|\mathrm{m}(B_j|E)-\mathrm{m}(B_j)|<\varepsilon$.
\item $E\cap U_j$ s-band, $E\cap U_0=\emptyset$.
\end{enumerate}
Finally we can use the previous proposition in each $E\cap U_j$ to define $\theta:E\rightarrow M$ satisfying the conclusion of the lemma.
\end{proof}
We are ready to finish the proof.
\begin{theorem}\label{thm:VWB}
The partition $\mathcal{P}$ is VWB.
\end{theorem}
\begin{proof}
Let $\epsilon>0$ be given. Since for every $s$-saturated set $E$ the $f^j\parP$-name and the $f^j\parP|E$-name of a point $x$ are determined by the $\parP$-name of $f^{-j}x$, and since all atoms $A\in \parP$ have (piecewise) smooth boundaries, we can find $\delta>0$ such that the conclusion of the previous lemma implies the hypotheses of \cref{lem:lemanombres}. The argument is similar to the one employed in the proof of \cref{lem:sbandas}, as points $x,y$ such that $d(x,y)<\delta$ and $x,y$ are in different atoms are in the $\delta$-neighborhood of the union of the boundary of the atoms.
Choose $N$ sufficiently large such that the previous lemma holds for $\varepsilon$ replaced with $\frac{\epsilon}{16}$. If $N'\geq N$ and $G\subset\bigvee_{N}^{N'}f^n\parP$ is the set of atoms given above, for $A\in \mathscr{G}$ we can define $\theta:E\subset A\rightarrow E$ $\frac{\epsilon}{16}$-measure preserving, so extending $\theta$ arbitrarily to $A\setminus E$ and applying \cref{lem:lemanombres} we conclude that for all $n\geq 0$ it holds
\[
\overline{d}(\{f^j\parP\}_1^n,\{f^j\parP|A\}_1^n)\leq \epsilon.
\]
Hence $\parP$ is VWB.
\end{proof}
\Cref{thm:VWB,thm:Ornstein} imply Theorem C.
\section{Applications and concluding remarks}
\label{sec:applications_and_concluding_remarks}
In this last section we apply our results, particularly of \cref{sec:a_characterization_of_equilibrium_states}, to the rank-one case. We finish with some questions.
\subsection{Hyperbolic flows}
\label{sub:hyperbolic_flows}
Let $\Phi=(f^t)_{t\in\Real}$ be an Anosov flow with invariant splitting $TM=E^s\oplus E^c\oplus E^u$, where $E^c$ is generated by tangents to flow direction. If this flow is transitive there is a dichotomy: either
\begin{enumerate}
\item $\Phi$ is a suspension of an Anosov diffeomorphism, and therefore $M$ is a fiber bundle over $S^1$, with fibers transverse to the flow line, or
\item both $\Fs,\Fu$ are minimal foliations.
\end{enumerate}
See \cite{AnosovFlow}. In what follows we will focus in this last case.
For Anosov flows the powerful of symbolic representations is available \cite{SymbHyp,RatnerMarkov}. With this one can establish the following.
\begin{proposition}\label{pro:medidasinestablesrankone}
Given a H\"older potential $\varphi:M\to\Real$ there exists $\mu^u\in\mathrm{Meas}(\Fu)$ satisfying:
\begin{enumerate}
\item[a)] $\mu^u_x$ has full support inside $\Wu{x}$, and is non-atomic.
\item[b)] $\forall t\in\Real, x\in M$, $f^{-t}\mu^u_{f^t(x)}=e^{\Ptop(\varphi) t-\int_0^t \varphi(f^s(x))ds}\mu^u_x$.
\item[c)] The map $\mu^u:x\to \mu^u_x$ is weakly continuous in the following sense: given $y\in \Wu{x,\epsilon_0}$ and $\hs_{y,x}:\Wu{y}\to\Wu{x}$ the locally defined Poincaré map, it follows that for any $A\subset \Wu{x}$ relatively open and pre-compact it holds
\[
\mu_{y}^u(\hs_{y,x}(A))\xrightarrow[y\to x]{}\mu^u_x(A).
\]
Moreover the convergence is uniform in $x$.
\end{enumerate}
\end{proposition}
\begin{proof}
The construction is due to N. Haydn \cite{localproductstructureHaydn}, based in the original construction of Gibbs measures from Sinai \cite{Sinai_1972}. Details are given in the Appendix of the paper of M. Pollicot and C. Waldken \cite{Pollicott2001}. Part $c)$ is not explicitly stated by Haydn who only states (a stronger version of) this fact for holonomies corresponding to the strong stable foliation, although it clearly follows from this and part $b)$.
\end{proof}
Let $f=f_1$ the time-one map. Given $\varphi$ we denote $\tilde{\varphi}$ its average along flow lines $\tilde{\varphi}(x)=\int_0^1 \varphi(f^tx)dt$. Observe that by the above proposition the family of measures $\mu^u=\{\mu^u_x\}_{x\in M}$ satisfy the same conditions as the measures along unstables that we constructed in \cref{sub:the_conditional_measures_potentials_constant_along_the_center_and_srbs}, therefore all our construction goes through and we can determine an equilibrium state $\muf[\tilde{\varphi}]$ for the system $(f,\tilde{\varphi})$ having conditionals along unstables given by the family $\mu^u$. We now relate $\muf[\varphi]$ and the equilibrium states for the flow.
Denote by $\PTM{\Phi}{M}=\cap_{t\in\Real} \PTM{f_t}{M}$ the set of invariant measures for the flow. The topological pressure of the system $(\Phi,\varphi)$ is given by the variational principle (\cite{WaltersPres}, Corollary $4.12$)
\begin{equation}
\Ptop(\Phi, \varphi)= \sup_{\mathclap{\nu\in \PTM{\Phi}{M}}}\{h_{\nu}(\Phi_1)+\int \varphi d\nu\},
\end{equation}
and the definition of equilibrium state for this system is clear. In what follows we will write $\mathscr{Eq}(f,\varphi)$ for the set of equilibrium measures of $(f,\varphi)$, and analogously $\mathscr{Eq}(\Phi,\varphi)$ denotes the set of equilibrium states for $(\Phi,\varphi)$. Clearly $\mathscr{Eq}(\Phi,\varphi)\subset \mathscr{Eq}(f,\varphi)$, and although in principle these sets can be different, it is known that
\[
\Ptop(\Phi,\varphi)=\Ptop(f,\varphi).
\]
On the other hand, using the previous line it is direct to check that $\Ptop(\Phi,\varphi)=\Ptop(f,\tilde{\varphi})$, therefore it follows
\[
\mathscr{Eq}(\Phi,\varphi)=\{\mu\in \PTM{\Phi}{M}:h_{\mu}(f)+\int \tilde{\varphi} d\mu\}\subset \mathscr{Eq}(f, \tilde{\varphi}).
\]
\begin{proposition}\label{pro:unicamedidaequilibrio}
If $\Phi$ is a transitive Anosov flow, then the sets $\mathscr{Eq}(\Phi,\varphi)$ and $\mathscr{Eq}(f,\tilde{\varphi})$ coincide. Therefore $\mathscr{Eq}(f, \tilde{\varphi})=\{\muf[\tilde{\varphi}]\}$.
\end{proposition}
\begin{proof}
Let $A(t,x)=\int_0^t \varphi(\Phi_sx)ds$; then $A:M\times\Real\to\Real$ is an additive cocycle over $\Phi$ ($A(t+s,x)=A(t,\Phi_sx)+A(s,x))$ and $A(1,x)=\tilde{\varphi}$. For transitive hyperbolic flows it is known (cf. \cite{BowenRuelle})that $\mathscr{Eq}(\Phi,\varphi)=\{\mu\}$, and in particular $\mu$ is ergodic for the flow (in fact $(\Phi,\mu)$ is a Bernoulli flow). Fix any $\mathrm{m}\in \mathscr{Eq}(f,\tilde{\varphi})$ and note that since $f$ and $f^s$ commute, $f^s\mathrm{m}\in \mathscr{Eq}(f,\tilde{\varphi} \circ f^s)$.
Now the cocycle property tells us that
\begin{align*}
&A(1+s,x)=A(1,f^s(x))+A(s,x)=A(s,f(x))+A(1,x)\\
&\Rightarrow \tilde{\varphi}\circ f^s(x)-\tilde{\varphi}(x)=A(1,f^s(x))-A(1,x)=A(s,f(x))- A(s,x)
\end{align*}
and $\tilde{\varphi}\circ f^s$ is cohomologous to $\tilde{\varphi}$, i.e.\@ its difference is of the form $k(f(x))-k(x)$ for some $k$ (that in this case is H\"older). It is simple to check that cohomologous potentials have the same equilibrium states, therefore $f^s\mathrm{m}\in \mathscr{Eq}(f,\tilde{\varphi})$. The measure $\tilde{\mu}=\int_0^1 f^s\mathrm{m}$ is invariant under the flow, and a equilibrium state for $(f, \tilde{\varphi})$ hence $\tilde{\mu}=\mu$. By ergodicity of $\mu$ we then deduce that
\[
f^s\mathrm{m}=\mu\quad\forall s\in[0,1]
\]
and in particular $\mathrm{m}=\mu$.
\end{proof}
We can pair this with Theorem B and deduce the following.
\begin{theorem}\label{thm:Bforrankone}
Given a transitive Anosov flow $\Phi$ of class $\mathcal{C}^2$ that is not a suspension, there exists a family $\mu^u\in\mathrm{Meas}(\Fu)$ satisfying the following: for any $\mathrm{m}\in\PM[M]$, if
\begin{enumerate}
\item $\mathrm{m}$ is $f^t$-invariant for some $t\neq \emptyset$, and
\item $\mathrm{m}$ has conditionals absolutely continuous with respect to $\mu^u$,
\end{enumerate}
then $\mathrm{m}$ is the unique equilibrium state for the system $(\Phi,\varphi)$.
\end{theorem}
In a subsequent article \cite{ContributionsErgodictheory} we will use the previous theorem to investigate properties of the horocyclic flow. In a restricted setting we will obtain a stronger characterization of equilibrium states.
Here is the last application of these ideas.
\begin{corollary}\label{cor:uniquequasiinvariantcs}
Suppose that $\zeta^{cs}\in\mathrm{Meas}(\Fcs)$ is such that for every $x\in M$, $f^{-1}\zeta^{cs}_{fx}=e^{\tilde{\varphi}-P'}\zeta^{cs}_{x}$, where $P'\in\Real$. Then $\zeta^{cs}=\mu^{cs}$ and $P'=\Ptop(\varphi)$.
\end{corollary}
\begin{proof}
The quasi-invariance condition in the hypotheses permit us to compute the Jacobian of $\zeta^{cs}$ with respect to the unstable holonomy, and this coincides with the Jacobian of $\mu^{cs}$. We can thus use $\zeta^{cs}$ instead of $\mu^{cs}$ and construct a probability $\mathrm{m}$ that is an equilibrium state for $(f, \tilde{\varphi})$ (cf.\@ \cref{sub:theequilibrium}), and therefore it is the unique equilibrium state for $(\Phi,\varphi)$. Using the local product structure one gets that $\zeta^{cs}=\mu^{cs}$.
\end{proof}
\subsection{Epilogue}\label{subs:epilogue}
We finish our article with some questions and considerations. In this part the standing hypotheses are the ones of Theorem A.
\begin{question}
Does the existence of the families of measures $\mu^u,\mu^s$ satisfying the quasi-invariance condition imply the uniqueness of the unique equilibrium state for $f$?
\end{question}
This question has eluded us for some time, and for the time being the only cases where uniqueness is established are the ones treated in this article. Note however that the methods for the two types of potentials are different in nature, and it seems possible that an unified approach may be available. Maybe one can start from the following.
\begin{question}
Is always the case that $\muf=\tilde{\mathrm{m}}_{\scriptstyle\varphi}$?
\end{question}
Since these measures are ergodic, it would suffice to establish that they are not singular with respect to each other. Here is example of such situation.
\begin{claim}
If the center disintegration of the measures $\mu^{cu}, \mu^{cs}$ is Lebesgue, then $\muf=\tilde{\mathrm{m}}_{\scriptstyle\varphi}$.
\end{claim}
\begin{proof}
It suffices to establish that $\tilde{\mathrm{m}}_{\scriptstyle\varphi}<<\muf$. During the proof we will denote $a\sim b$ to indicate that the quotient $\frac{a}{b}$ is bounded above and below, with constants independent of $\epsilon$. Using the \cref{pro:estructuraproducto} and the remark after it (the role of $\ep_2,\ep_1$ is to guarantee the product structure of the measure) we get
\begin{align*}
&\sup_{0<\epsilon<\ep_2} \frac{\tilde{\mathrm{m}}_{\scriptstyle\varphi}(B(x,\epsilon))}{\muf(B(x,\epsilon))}\leq D\cdot \sup_{0<\epsilon<\frac{\ep_1}{4}} \frac{\msx(\Ws{x,\epsilon})\cdot \mcux(\Pcu{x,\epsilon})}{\mcsx(\Pcs{x,\epsilon})\cdot \mux(\Wu{x,\epsilon})}
\end{align*}
where $D$ does not depend on $x$. As $\mcsx$ has Lebesgue disintegration,
\[
\frac{\msx(\Ws{x,\epsilon})}{\mcsx(\Pcs{x,\epsilon})}\sim \left(\frac{\epsilon_0}{\epsilon}\right)^c
\]
and likewise,
\[
\frac{\mcux(\Pcu{x,\epsilon})}{\mux(\Wu{x,\epsilon}}\sim \left(\frac{\epsilon}{\epsilon_0}\right)^c.
\]
We conclude that there exists some constant $\widetilde{D}>0$ such that for every $x\in M$ it holds
\begin{equation}
\sup_{0<\epsilon<\ep_2} \frac{\tilde{\mathrm{m}}_{\scriptstyle\varphi}(B(x,\epsilon))}{\muf(B(x,\epsilon))}\leq \widetilde{D}.
\end{equation}
This implies that $\tilde{\mathrm{m}}_{\scriptstyle\varphi}<<\muf$ and $\frac{d\tilde{\mathrm{m}}_{\scriptstyle\varphi}}{d\muf}(x)\leq \widetilde{D}$ for $\muf\aep(x)$, as consequence of the standard Lebesgue differentiation theorem for Borel measures. To check the absolute continuity, assume that $\muf(A)=0$ and let $r>0$ be arbitrary. Take a sequence of boxes $\{B(x_k,\epsilon_k)\}_{k\in\Nat}$ satisfying
\begin{enumerate}
\item $A\subset \bigcup_{k\geq 0}B(x_k,\epsilon_k)$,
\item $\sum_{k=0}^{\oo} \muf(B(x_k,\epsilon_k)) <\frac{r}{\widetilde{D}}$.
\end{enumerate}
Then
\[
\tilde{\mathrm{m}}_{\scriptstyle\varphi}(A)\leq \sum_{k=0}^{\oo} \tilde{\mathrm{m}}_{\scriptstyle\varphi}(B(x_k,\epsilon_k))\leq \widetilde{D} \sum_{k=0}^{\oo} \muf(B(x_k,\epsilon_k))<r,
\]
and since $r>0$ is arbitrary, $\tilde{\mathrm{m}}_{\scriptstyle\varphi}(A)=0$.
\end{proof}
The properties of the center disintegration $\mu^{cs}$ and $\mu^{cu}$ along centrals seem to give important information about the system, although it is not obvious how to deduce to deduce this information. One possibility is to use the construction os \cref{sec:measures_along_foliations} as in the following example.
\begin{example}
For $x\in M$ denote by $\lambda_x$ the Lebesgue measure on $\Wc{x}$, and if $B$ is foliated box corresponding to the center foliation with $x\in B$, let $\widetilde{\lambda}_x^B$ be the corresponding probability measure induced by $\lambda_x$ on the plaque of $B$ containing $x$. Assume that $E^c$ is differentiable, and that $\varphi$ is $c$-constant.
\begin{claim}
For every $x\in M$ the conditionals measures of $\mcux$ with respect to the center foliation are given by $\{\lambda_y\}_{y\in\Wc{x}}$.
\end{claim}
That is, for $B$ as above with $x\in B$, the conditional measures of $\mcux|B$ with respect to the partition by center plaques of $B$ are given by $\{\widetilde{\lambda}_x^B\}_{y\in B}$ (in particular the disintegration is defined everywhere). Similarly for $\mcsx$.
To see this fix $x\in M, 0<\ep\leq\ep_1$ and let $B=\Wc{\Wu{x,\ep},\ep}$. For $n\geq 0$ let $B_n=f^n(B)$ and note that since $f$ is an center isometry then $\forall y\in B_n$ the map $f^{-n}|: \Wc{y}\cap B_n\to \Wc{f^{-n}y}\cap B$ is an isometry with respect to the induced metric. As $\Fc$ is differentiable, the disintegration of the Lebesgue measure $\nu|B_n=\Leb^{cu}|B_n$ along center plaques is given by $\{\widetilde{\lambda}^{B_n}_y\}_{y\in B_n}$, therefore for $\upsilon_n=f^{-n}\nu|B_n$ its disintegration is given by $\{f^{-n}\lambda^{B_n}_y\}_{y\in B_n}=\{\lambda^{B}_z\}_{z\in B}$. This in turn implies that the center disintegration of $\nu^n=e^{\SB}f^{-n}\nu$ is given by
\[
(\nu^n)^c_y=\frac{e^{\SB}}{\int e^{\SB} d(\upsilon_n)^c_y}(\upsilon_n)^c_y=(\upsilon_n)^c_y=\widetilde{\lambda}_y^c.
\]
From this it follows that every section $\nu\in\mathcal{X}$, in particular $\mu^{cu}$, has the same disintegration, proving our claim.
\end{example}
\begin{question}
What are in general the disintegrations of $\mu^{cs}$ and $\mu^{cu}$ along center leaves?
\end{question}
We had to impose some restriction in the potentials considered with the sole purpose of obtaining a family of obtaining a family $\mu^u\in\mathrm{Meas}(\Fu)$ defined everywhere (and in particular, depending measurably on the point). As explained in Theorem D, this is not an artificial restriction of the method, and in fact is a subtle point. However, the following it is still not clear and more research is needed.
\begin{question}
What are the conditions and/or obstructions for the existence of a continuous family $\mu^u\in\mathrm{Meas}(\Fu)$?
\end{question}
In particular:
\begin{question}
Assume the hypotheses of Theorem A and furthermore that $f$ is accessible (meaning that any pair of points in $M$ can be joined by a finite concatenation of curves whose derivatives are in $E^u\oplus E^s$). Does there exist a continuous family $\mu^u\in\mathrm{Meas}(\Fu)$ in this case?
\end{question}
We remark that the example given in the Appendix does not satisfy this condition, but only essential accessibility (any bi-saturated set by $\Fs,\Fu$ has Lebesgue measure either zero or one).
\smallskip
The fact that equilibrium states are determined by some families of measures on a dynamical structure (the invariant foliations) seem to indicate strong restrictions on the geometry of these measures. In particular we can ask the following.
\begin{question}
Is $\PM[M, \mu^u]=\PTM{f}{M,\mu^u}$?
\end{question}
In a shortcoming article \cite{ContributionsErgodictheory} we give an answer to this for codimension-one Anosov flows.
\section{Appendix: Proof of Theorem D}
Let $A:\Tor^2\to\Tor^2$ be a linear Anosov map with and consider $f:\Tor^2\times \Tor^2\to \Tor^4$ the group extension of $A$ given by
\[
f(x,\theta_1,\theta_2)=(A\cdot x,\theta_1+\alpha_1 k(x),\theta_2+\alpha_2 k(x))
\]
where $k:\Tor^2\to\Tor$ is analytic, not cohomologous to constant, and $(\alpha_1,\alpha_1)$ is a Liouville vector (see below). Then $f$ is a conservative analytic center isometry with center bundle parallel to the vertical fibers $\{0\}\times \Real^2$. This example was considered by Dolgopyat \cite{dolgolivsic} in other context. Due to a result of Katok, Theorem 1 in \cite{Katok1980a}, this map is a $K$-automorphism, therefore in particular its unstable foliation $\Fu$ is ergodic.
\begin{lemma}
$\Fu$ is minimal.
\end{lemma}
\begin{proof}
Let $e^u$ be a unit vector in $\Tor^2$ generating the unstable bundle of $A$ and, consider the linear flow $\psi^u_t:\Tor^2\to\Tor^2$, $\psi^u_t(x)=x+t\cdot e^u$. The unstable foliation of $A$ are the orbits of $\psi^u$, and it holds
\[
A\psi^u_t(x)=\psi^u_{\lambda e^u}(Ax)
\]
where $\lambda>1$ is the unstable eigenvalue of $A$. It is not hard to verify that the unstable foliation of $f$ is the orbit foliation of the flow $\hat{\psi}^u_t:\Tor^4\to\Tor^4$ given by
\[
\hat{\psi}^u_t(x,\theta_1,\theta_2)=(\psi^u_t(x), \theta_1+\alpha_1l_t(x),\theta+\alpha_2l_t(x))
\]
where $l_t(x)=\sum_{n=1}^{\oo} k(A^{-n}x)-k(A^{-n}\psi^u_t(x))=\sum_{n=1}^{\oo} k(A^{-n}x)-(k(A^{-n}x)+\frac{t}{\lambda^n}e^u)$; since $k$ is differentiable the previous series is convergent. In particular $\hat{\psi}^u$ is conservative.
The flow $\psi^u$ is uniquely ergodic, $\Leb$ being its invariant measure, therefore by a classical result due to Furstenberg, either $\hat{\psi}^u_t$ is uniquely ergodic (and therefore minimal, since it preserves $\mu=\Leb\times d\theta_1\times d\theta_2$), or any $\hat{\psi}^u_t$ ergodic invariant measure has atomic disintegration along the partition given by the vertical tori (i.e., the center foliation of $f$). See Theorem $4.1$ in \cite{Furstenberg1961}. The later possibility is prohibited by Katok's result, as $\hat{\psi}^u_t$ is ergodic with respect to Lebesgue.
\end{proof}
The vector $(\alpha_1,\alpha_2)$ is chosen so that $\frac{\alpha_1}{\alpha_2}\not\in\mathbb{Q}$, and that for every $n\in\Nat$ there exists $m_{n,1},m_{n,2}\in \Nat$ such that
\begin{itemize}
\item \(|\al_1m_{n,1}+\al_2m_{n,2}|<\frac{1}{m_{n,2}^n}\);
\item \(m_{n,2}\geq m_{n,1}>n\).
\end{itemize}
Denote $d_{n}(x,\theta_1,\theta_2)=\exp(2\pi i \prodi{(\theta_1,\theta_2)}{(m_{n,1},m_{n,2})})$; then $d_n:\Tor^4\to\Tor^2$ satisfies
\begin{align*}
d_n(f(x,\theta_1,\theta_2))=\exp(2\pi i \prodi{(\theta_1+\alpha_1 k(x),\theta_2+\alpha_2 k(x))}{(m_{n,1},m_{n,2})})\\
=\exp(2\pi ik(x) \prodi{(\alpha_1,\alpha_2 )}{(m_{n,1},m_{n,2})})d_n(x,\theta_1,\theta_2).
\end{align*}
Define $\varphi:M\to \Real$ by
\begin{align*}
\varphi(x,\theta_1,\theta_2)&=\sum_{n\geq 0} d_{n}(x,\theta_1,\theta_2)-d_{n}(f(x,\theta_1,\theta_2))\\
&=\sum_{n} \left(1-\exp\Big(2\pi ik(x) \prodi{(\alpha_1,\alpha_2 )}{(m_{n,1},m_{n,2})}\Big)\right) d_n(x,\theta_1,\theta_2).
\end{align*}
Observe that $\varphi$ is the $\mathcal{C}^{\oo}$ limit of coboundaries, therefore $\mathscr{Eq}(f,\varphi)=\mathscr{Eq}(f,0)$. It is clear on the other hand that the the Lebesgue measure in $\Tor^3$ maximizes entropy
Assume that there exists a family of measures $\zeta^{u}=\{\zeta^u_L: L\in\Fu\}$ satisfying the quasi-invariance condition $f^{-1}\zeta_{fL}^u=e^{\varphi - \htop}\zeta_{L}$, and depending continuously on the point. One can choose $k$ so that $\Fu$ is minimal, and then by Theorem B one would have
\[
\zeta_L=e^{\omega_L}\mathrm{m}^u_L
\]
where $\mathrm{m}^u_L$ is the corresponding Margulis measure on $L$, and $\omega_L:L\to\Real_{>0}$ is continuous. Then we get
\[
e^{\varphi-\htop+\omega_L}\mathrm{m}^u_L=f^{-1}(e^{\omega_{fL}\mathrm{m}^u_{fL}})=e^{\omega_{fL}\circ f-\htop}
\]
and $\varphi=\omega_{fL}\circ f-\omega_L$ (since the Margulis measures have full support). If $k(0)=0$ then $L_0=\Wu{0}$ is fixed, and we get that for $\omega=\omega_{L_0}$,
\[
\varphi=\omega\circ f-\omega\quad \text{on }L_0.
\]
Observe however that the family $\{\omega_{\Wu{x}}\}_x$ depends continuously on $x$ ($\mu^u$ are essentially the Lebesgue measure), and therefore $\omega$ extends continuously to $M$, thus implying that $\varphi$ is cohomologous to the zero function, with continuous transfer function $\phi$. However, by taking the Fourier expansion it is direct to check that necessarily \(\phi=\sum_n d_n\), which is not continuous. This contradiction shows that our assumption on the existence of $\zeta^{u}$ is incorrect, thus implying Theorem D.
\newpage
\printbibliography
\end{document} |
\section{Introduction}
\label{sec:Introduction}
Machine learning (ML) field of computer science that is core of artificial intelligence, in particular deep learning, showed to be a very powerful tool for many modern technologies tasks requiring feature learning and pattern recognition such as image classification , natural language processing, speech recognition , and video games development, with deep learning being particularly successful at uncovering features in structured data (feature learning and compression).\cite{Hershey,Silver,LeCun,Goodfellow,Kelleher} Success of deep learning algorithms can be explained by their connection to variational renormalization group (RG)\cite{Mehta, Chung, Koch,Janusz,Funai} which is an iterative coarse-graining scheme that allows extraction of relevant features (operators) as a physical system is studied at different length scales and one of the most important and successful techniques in theoretical physics.
Recently artificial neural networks (ANN) and ML techniques also proved to be very powerful methods for studying variety of complex many-body problems\cite{Mechta1,Carleo1,DasSarma,Jia,Carleo,Carleo2,Szabo,Shi,Borin,Wu,Choo,Vieijra,Nomura,Nomura2,Astrakhantsev,Irikura,Choo2,Westerhout,Hartmann,Schmitt,Vicentini,Nagy,Yoshioka,Liang,Gao,Deng,Torlai,Deng2,Saito,Sharir,Kaubruegger,Duric,Pilati,Deng3} and is therefore of great importance to further investigate applicability of these methods in studying challenging models for which exact solutions are not known. In this paper we study one of such models, spin-$1/2$ square lattice $J_1$-$J_2$ model with additional plaquette structure, and demonstrate suitability of ML techniques to study complex phases and phase transitions by comparing the ML results with the results obtained by other available methods.
ML algorithms are designed to deal with large and complex data sets and make predictions on data by classifying and extracting features from data and are therefore very successful in studying various complex quantum systems where complete description of quantum many-body states requires exponentially large data sets. Namely in ML algorithms machine is an ANN that can learn probability distribution over set of its inputs. These networks originally introduced as simplified models of human brain can also be used to construct compact representations of many-body quantum states where the many-body wave-function corresponds to the probability distribution that the network tries to approximate.
Boltzmann machines, in particular restricted Boltzmann machines (RBMs), widely used in machine learning community, showed very promising results so far since RBMs can represent many quantum states of interest and can also be efficiently numerically optimized using variational Monte Carlo (VMC) method.\cite{Carleo} Additionally there are strong connections between restricted Boltzmann machines and some classes of tensor network states in arbitrary dimensions, for example Jastrow wave-functions, entangled plaquette states (EPS) and string-bond states (SBS).\cite{Clark,Glasser,Chen} In ML language VMC optimization of an ANN is equivalent to reinforcement learning (RL)\cite{Sutton, Otterlo} where an agent learns from an interactive environment by trial and error using feedback from its own actions and experiences. The agent takes a suitable action to maximize reward in a particular situation and creates an action-reward feedback loop of an RL algorithm. Mathematical frameworks to describe an environment in RL are Markov decision processes and within VMC method a wave-function is learned on the basis of feedback from variational principle.
ML techniques, both in supervised and unsupervised forms, also proved to be very powerful techniques for detecting phase transitions in various systems. \cite{Mechta1,Carleo1,Carrasquilla,Carrasquilla2,Lian,Iakovlev,Broecker,Dong,Hsu,Hu,Chng,Zhang,Wang,Nieuwenburg,Broecker2,Yoshioka2,Wetzel,RodriguezNieva,Venderley,Giannetti,Zhang2,Zhang3,Berezutskii,Kharkov,Rem,Rao,Canabarro,Wetzel2,Morningstar,Zhao1,Cheng,Jadrich,Kim,Bohrdt,Rzadkowski,Yao,Che,Ohtsuki,Zlabys,Tan} ML allows automatic classification of different phases of matter without necessity to characterize phases by local order parameters, topological invariants or many-body localized phases as in conventional approaches. Such power of ML techniques to extract information of phases and phase transitions directly from many-body configurations where classification within an ANN occurs without any knowledge of the Hamiltonian or locality of interactions defined ML phases of matter as a new research frontier and a basic research tool in the field of condensed matter and statistical physics to identify poorly understood phases where the order parameter or topological description are not known a priori.
\begin{figure}[t!]
\includegraphics[width=\columnwidth]{Fig1.eps}
\caption{\label{Fig:Hamiltonian} Visualization of the spin-$1/2$ two-dimensional square lattice Heisenberg model with plaquette structure (\ref{eq:Hamiltonian}). NN and NNN antiferromagnetic couplings correspond to lattice bonds illustrated with blue and red lines, respectively. Solid lines correspond to intra-plaquette couplings ($J_1$ and $J_2$) whilst dashed lines correspond to inter-plaquette couplings ($\lambda J_1$ and $\lambda J_2$). For $\lambda=1$ the model corresponds to the standard uniform spin-$1/2$ $J_1$-$J_2$ Heisenberg antiferromagnet with NN and NNN couplings, and for $\lambda=0$ the Hamiltonian describes unconnected four-spin $J_1$-$J_2$ plaquettes. At fixed $J_2/J_1$ a quantum phase transition from paramagnetic plaquette resonating valence bond state to a state with antiferromagnetic long-range order occurs at some critical value $\lambda_c$ of the inter-plaquette coupling strength parameter.
}
\end{figure}
In an unsupervised ML algorithm only input to the algorithm is a set of data and the algorithm attempts to extract the features from the data or to arrange data into clusters.\cite{Mechta1} On the other hand, the basic idea of supervised machine learning (SML) is to train a machine (ANN) with large previously solved data set of input-output pairs and then use trained ANN model to process, characterize and make predictions for new data.\cite{Mechta1} Training procedure for an ANN is an optimization problem where a non-linear variational function is optimized with respect to a loss (cost) function by slowly adjusting the free parameters associated with connections between ANN neurons and their biases until high classification accuracy is obtained. Loss function is a function that gives information how good an ANN is for a certain task.
A typical example of SML is image recognition where a large set of labeled images is used as a training set, for example, images of cats (label $1$) and dogs (label $0$) with images as input and image names (labels) as output. After training procedure trained machine (ANN) that has learned key features of training set of labeled images is used to make predictions (perform a task), that is to recognize new images that are not in the training data set as being images of a cat or a dog. ANNs particularly successful in image recognition and classification are convolutional neural networks (CNNs), a class of deep neural networks that take into consideration additional input data set symmetries and structure (locality and translation invariance). Taking into account hierarchical pattern in data CNNs can construct more complex patterns by considering smaller and simpler patterns.
In this paper we study spin-$1/2$ square lattice $J_1$-$J_2$ model with plaquette structure using combination of RBM based RL (VMC) and CNN based SML methods. The model Hamiltonian is:
\begin{equation}\label{eq:Hamiltonian}
H=\sum_{a=0}^{1}(\delta_{0,a}+\delta_{1,a}\lambda)(J_1\sum_{\langle i,j \rangle_a}\vec{S}_i\vec{S}_j + J_2\sum_{\langle\langle i,j\rangle\rangle_a}\vec{S}_i\vec{S}_j)
\end{equation}
where $\vec{S}_i$ are spin-$1/2$ operators, $\delta_{a,b}$ is Kronecker delta, intra- and inter-plaquette interactions correspond to $a=0$ and $a=1$, and $\langle i,j\rangle$ and $\langle\langle i,j \rangle\rangle$ denote the nearest-neighbor (NN) and the next-nearest-neighbor (NNN) pairs of sites, respectively, as illustrated in FIG. \ref{Fig:Hamiltonian}. Here $J_1$ and $J_2$ are intra-plaquette whilst $\lambda J_1$ and $\lambda J_2$ are inter-plaquette NN and NNN couplings with $\lambda$ being the inter-plaquette coupling strength parameter.
The model (\ref{eq:Hamiltonian}) with plaquette structure, and in general spin-$1/2$ antiferromagnetic Heisenberg model with various interactions, has been a topic of great interest because of its close connection to cuprate superconductors\cite{Lee,Manousakis,Harland,Altman} and owing to possibility of realization of such models in experiments with cold atoms in optical lattices.\cite{Goral,Lahaye,Nascimbene,Rey,Olschlager} The model was also proposed as a possible model for (CuCl)LaNb$_2$O$_7$.\cite{Kogeyama,Ueda}Additionally models with quadrumerized or various other checkerboard patterns were suggested as models for Bi$_2$Sr$_2$CaCu$_2$O$_{8+\delta}$ and Ca$_{2-x}$Na$_x$CuO$_2$Cl$_2$.\cite{Stock,Tranquada,Shen}
In these models a quantum phase transition (QPT) from paramagnetic to antiferromagnetic state occurs by competition between antiferromagnetic bonds of different strength as in well studied dimerized Heisenberg models.\cite{Schmidt,Singh, Wenzel2,Jiang,Ma,Fritz,Leite,Merchant} As expected from quantum-to-classical mapping two-dimensional quantum coupled-dimer and coupled-plaquette magnets belong in the same universality class as the class of the three-dimensional classical Heisenberg model referred to as $O(3)$ universality class.\cite{Sachdev3} However in some cases when the dimer pattern lacks a certain symmetry, for example for staggered dimer model also called $J$-$J'$ model which corresponds to Hamiltonian (\ref{eq:Hamiltonian}) in the limit $J_2\gg J_1$ that lacks discrete lattice rotational symmetry, an additional cubic interaction of critical fluctuations not present in the classical $O(3)$ model appears in the low-energy quantum field theory.\cite{Ma,Fritz} This cubic interaction is an interesting quantum effect with no classical counterpart that causes non-monotonic finite-size scaling behavior. Similar non-monotonic scaling also appears for deconfined quantum phase transitions that are characterized by exotic fractionalized quasi-particles and emergent gauge fields.\cite{Sachdev,Sachdev2,Sentil1,Sentil2,Sentil3,Balents,Levin,Sandvik1,Sandvik2,Shao,Ma2,Zhao,You} Moreover $J$-$J'$ model is also connected with Shastry-Sutherland model that explains critical properties of SrCu$_2$(BO$_3$)$_2$.\cite{Shastry,Lauchli,Zayed,Corboz,Zhao2}
In our numerical calculations we use NetKet\cite{Carleo3} and TesorFlow\cite{Abadi} libraries and consider the cases when $J_2=0$ and $J_2=J_1$ for a range of the inter-plaquette coupling strength parameter $\lambda$ values. To identify ordered and disordered phases and quantum phase transitions we write the ground-state wave-function in the form of a RBM and optimize the RBM parameters using RL (VMC) method. For both cases we further calculate the ground-state spin-spin correlation matrices $M_{ij}=\langle \vec{S}_i\vec{S}_j\rangle$ for a range of inter-plaquette coupling strength parameter $\lambda$ values away from the QPT point and use images of these matrices as input data for a CNN based SML algorithm that detects QPT.
The unfrustrated case with the NNN couplings $J_2=0$ has so far been studied by several methods, exact diagonalization (ED),\cite{Voigt} Ising series expansion,\cite{Singh2} contractor renormalization,\cite{Albuquerque} real space RG,\cite{Fledderjohann} coupled cluster,\cite{Gotze} and quantum Monte Carlo (QMC)\cite{Albuquerque,Wenzel,Xu,Ran} methods. The results for the frustrated case have been calculated by combination of the Lanczos ED and coupled cluster method (CCM) which are both powerful general many-body methods. \cite{Gotze,Kruger} Providing accuracy of the finite-size scaling procedure QMC calculations are expected to give the most accurate results in unfrustrated regime. However in the frustrated regime QMC calculations can not be used to study properties of the system since the method in that case suffers from the minus sign problem. Namely, when partition function of a D dimensional quantum system is expanded in terms of D+1 dimensional classical configurations, weights of these configurations can be both positive and negative or complex and thus invalidate their usual Monte Carlo interpretation as probability distribution.\cite{Hirsch, Loh} Contrary to QMC method SML combined with RBM VMC is a sign problem free method that can provide accurate description of the system phases and phase transitions even in the presence of frustration.
After confirming validity of the RBM ansatz to study ground-state properties of the system for a smaller system size of $N=16$ lattice sites and with periodic boundary conditions for which ED calculations can be done for comparison, we further calculate the ground-state spin-spin correlation matrices for the system sizes of $N=L\times L$ lattice sites with $L=6$, $8$ and $10$. For each system size we calculate 4000 matrices for the values of the inter-plaquette coupling strength parameter $\lambda$ deep in pRVB and AFM phases away from the QPT point. Images are further resized to $L\times L$ pixel values and used as a training (90\% of the images) and validation (10\% of the images) datasets in a CNN based SML algorithm. Trained CNN is then used to classify images of the ground state spin-spin correlation matrices calculated for the remaining interval of $\lambda$ values that contains QPT point $\lambda_c$. SML calculations are performed for hundred random choices of 10\% of the validation data and the final probabilities for the images to belong to pRVB or AFM class calculated as an average over hundred CNN outputs. The point of QPT is identified as the point where pRVB and AFM probabilities are both $50\%$, that is as the point where the network is maximally confused.
Although consideration of much larger system sizes would be necessary to accurately extrapolate results for finite system sizes to thermodynamic limit ($L\rightarrow \infty$) our results show that the values for $\lambda_c(L)$ rapidly approach $\lambda_c(L\rightarrow \infty)$ values calculated previously by several other methods. We find the best agreement with the results obtained with real space RG method and CCM. For the unfrustrated case with NNN couplings $J_2=0$ previously obtained real space RG and CCM results are $\lambda_c^{RG}\approx 0.4822$\cite{Fledderjohann} and $\lambda_c^{CCM}\approx 0.47$,\cite{Gotze} whilst the QMC and finite-size scaling method result is $\lambda_c^{QMC}\approx 0.5485$.\cite{Albuquerque,Wenzel,Ran} For the frustrated $J_2=J_1$ case the CCM result is $\lambda_c=0.301$.\cite{Gotze} For the largest system size of $N=10\times10$ lattice sites we find the values $\lambda_c\approx 0.465$ and $\lambda_c\approx 0.355$. Whilst the QMC results suggest that the real space RG results converge non-monotonously towards the value obtained with QMC calculations when further excited states with more that one quintuplet are included in the RG calculations,\cite{Fledderjohann} our results indicate that the RG results might converge monotonously towards the value close to the value obtained with the CCM.
The paper is organized as follows. In Sec. \ref{sec:RBM} we describe the RBM ansatz for the ground-state wave-function. In Sec. \ref{sec:VMC} we explain RL scheme for RBM parameters optimization and calculation of the ground-state spin-spin correlation matrices that serve as a training data in the SML algorithm for QPT detection described in Sec. \ref{sec:SML}. We summarize our results, draw conclusions and discuss future research directions in the last section, Sec. \ref{sec:Conculsions}.
\section{Restricted Boltzmann machine representation of the ground-state wave function}
\label{sec:RBM}
To calculate the ground-state spin-spin correlation matrices used to train SML algorithm described in Sec. \ref{sec:SML} we represent the ground-state wave-function of the system with a restricted Boltzmann machine (RBM) ansatz:
\begin{equation}\label{eq:RBM1}
|\psi_{RBM}\rangle = \sum_{\{\vec{s}\}} (-1)^{n_{A\uparrow}}\psi_{RBM}(\{\vec{s}\})|\{\vec{s}\}\rangle,
\end{equation}
where
\begin{equation}\label{eq:RBM2}
\psi_{RBM}(\{\vec{s}\})=e^{\sum_{j=1}^Na_js_j}\prod_{i=1}^M\cosh(b_i+\sum_jw_{ij}s_j)
\end{equation}
and $|\{\vec{s}\}\rangle=|s_1,s_2,...,s_N\rangle$ are spin- $1/2$ configurations for $N$ lattice sites. Here we have included basis rotation corresponding to the Marshall's sign rule ( sign $(-1)^{n_{A\uparrow}}$ where $n_{A\uparrow}$ is the total number of $\uparrow$ spins on A subset of the lattice sites) which speeds up the convergence of the VMC calculations by providing a good starting sign structure particularly at low frustration.\cite{Marshall,Auerbach,Schollwock,Choo} For a bipartite lattice typical choice of the sites A corresponds to one of the two sublattices A and B in the bipartite lattice. For a square lattice A and B sublattices usually form checkerboard or collinear patterns as shown in FIG. \ref{Fig:AB_sublattices}. For the two cases that we have studied, $J_2=0$ and $J_2=J_1$ subset A corresponds to sublattice A in FIG. \ref{Fig:AB_sublattices} (a) and FIG. \ref{Fig:AB_sublattices} (b), respectively.
\begin{figure}[b!]
\includegraphics[width=\columnwidth]{Fig2.eps}
\caption{\label{Fig:AB_sublattices} Square lattice sublattices (A and B, denoted with red and blue sites) with (a) checkerboard and (b) collinear patterns.
}
\end{figure}
\begin{figure}[t!]
\includegraphics[width=\columnwidth]{Fig3.eps}
\caption{\label{Fig:RBM} RBM architecture that is used as an ansatz for the ground-state wave-function. The network consists of one visible layer with $N$ visible artificial neurons ($s_1$,...,$s_N$) and one hidden layer with $M$ hidden artificial neurons ($h_1$,...,$h_M$) with interactions only between the visible and hidden units. Network inputs are visible spin configurations $|\{\vec{s}\}\rangle\equiv|s_1,...,s_N\rangle$ and as an output the network calculates probability of a visible configuration $P(\{\vec{s}\})$ equivalent to the restricted Boltzmann machine ansatz coefficients $\psi_{RBM}(\{\vec{s}\})=P(\{\vec{s}\})$.
}
\end{figure}
RBM ansatz that encodes a many-body quantum state of spin-$1/2$ quantum system with $N$ spins is constructed of one visible layer with $N$ nodes (neurons, visible spins) that correspond to physical qubits ($s_1$, $s_2$, ... ,$s_N$), and one hidden layer with $M$ auxiliary spin variables ($h_1$, $h_2$, ... , $h_M$) as shown in FIG. \ref{Fig:RBM}. Two sets of binary units (classical spins) that correspond to visible ($s_i$, $i=1,...,N$) and hidden ($h_i$, $i=1,...,M$) neurons in Boltzmann machine architectures interact through an Ising interaction and corresponding Hamiltonian H is defined as
\begin{eqnarray}\label{eq:H_RBM}
H=\sum_{j=1}^N a_js_j+\sum_{i=1}^M b_ih_i + \sum_{i<j} c_{ij}s_is_j \\
+ \sum_{i,j}w_{ij}h_iv_j+\sum_{i<j}d_{ij}h_ih_j. \nonumber
\end{eqnarray}
Joint probability distribution over visible and hidden units is then defined as the Boltzmann weight of this Hamiltonian:
\begin{equation}\label{eq:Boltzmann_weight}
P(\{\vec{s}\},\{\vec{h}\})=\frac{1}{Z}e^{H(\{\vec{s}\},\{\vec{h}\})}
\end{equation}
where $Z$ is the partition function $Z=\sum_{\{\vec{s}\},\{\vec{h}\}}e^{H(\{\vec{s}\},\{\vec{h}\})}$. Then marginal probability of a visible configuration can be obtained by summing over all possible hidden configurations
\begin{equation}\label{eq:P_visible}
P(\{\vec{s}\})=\sum_{\{\vec{h}\}}\frac{1}{Z}e^{H(\{\vec{s}\},\{\vec{h}\})}.
\end{equation}
Here we take visible spin configurations $|s_1,...,s_N\rangle$ as network inputs and interpret the wave-function as complex probability distribution that the network tries to approximate. Therefore the Boltzmann machine ansatz wave-function coefficients are
\begin{equation}\label{eq:wf_coefficients}
\psi_{BM}(\{\vec{s}\})=P(\{\vec{s}\}).
\end{equation}
For the RBM ansatz Hamiltonian (\ref{eq:H_RBM}) includes only interactions $w_{ij}$ between the visible and hidden units ($c_{ij}=0$, $d_{ij}=0$) and biases $a_j$ and $b_j$. Sum over the hidden spins in (\ref{eq:P_visible}) can then be performed analytically and the wave-function ansatz coefficients written as
\begin{equation}\label{eq:coefficients_RBM}
\psi_{RBM}(\{\vec{s}\})=e^{\sum_{j=1}^Na_js_j}\prod_{i=1}^M\cosh(b_i+\sum_jw_{ij}s_j).
\end{equation}
ANN coefficients are in general taken to be complex in order to describe both the amplitude and the phase of the wave-function. Representability theorems\cite{Kolgomorov,Hornik1,Hornik2,LeRoux} guarantee existence of an ANN approximation for every sufficiently smooth and regular high-dimensional function, allowing ANN description of complicated many-body wave-functions. Quality of the ANN approximation can be systematically improved by increasing hidden variable density $\rho=M/N$ where $M$ is the number of hidden and $N$ the number of visible neurons (spins), respectively.
In one dimension, for well studied matrix product state (MPS) ansatz hidden variable density $\rho$ is analogous to the bond dimension, that is dimension of the matrices in the MPS ansatz, which determines the space complexity of the MPS. However, unlike MPS ansatz, ANN ansatz is well suited for description of quantum states in arbitrary dimensions due to intrinsically nonlocal correlations in space generated by hidden units. In our calculations we take $\rho=4$ which shows to be sufficient to obtain quite accurate estimates of the ground state properties of the system.
The parameters of the ANN are obtained using VMC method equivalent to a reinforcement learning scheme in which the ground-state wave-function is learned on the basis of feedback from variational principle as described in the following section.
\section{Variational Monte Carlo optimization of the restricted Boltzmann machine parameters - reinforcement learning of the ground-state wave-function based on feedback from the variational principle}
\label{sec:VMC}
Parameters of the RBM ansatz can be optimized (trained) using VMC method through minimization of the energy expectation value
\begin{equation}\label{eq:Energy}
E=\frac{\langle\psi_{RBM}|H|\psi_{RBM}\rangle}{\langle\psi_{RBM}|\psi_{RBM}\rangle}
\end{equation}
with respect to the network parameters ($a_j$,$b_j$,$w_{ij}$), $i,j=1,...,N$, with N being the number of lattice sites.
Energy $E$ can be optimized using several optimization approaches. Here we use stochastic reconfiguration (SR) method introduced by Sorella \emph{et al.}\cite{Sorella2,Becca} which can be viewed as an effective approximate imaginary time evolution in the variational subspace of the full Hilbert space. The method was initially developed to stabilize and partially solve the sign instability in the Green function Monte Carlo technique for continuous models\cite{Sorella} and then as an efficient and robust optimization scheme for lattice systems\cite{Sorella1}, small atoms\cite{Casula1} and molecules\cite{Casula2} that allows energy expectation value minimization for a many-parameter variational ansatz wave-function in an arbitrary functional form. SR method uses more information about the variational ansatz wave-function than other common simpler optimization protocols such as stochastic gradient descent (SGD) and therefore allows faster optimization of the many-body ansatz wave-function and more reliable convergence. Instead of optimizing energy directly the method at each step tries to maximize the overlap between the ansatz and the result of its imaginary time evolution.
\begin{figure}[t!]
\includegraphics[width=\columnwidth]{Fig4.eps}
\caption{\label{Fig:E0_RBM} Monte Carlo energy estimate obtained with variational optimization of the RBM ansatz with hidden unit density $\rho=4$ for the system size of $N=4\times 4 =16$ lattice sites and with periodic boundary conditions. Here the inter-plaquette coupling strength parameter $\lambda=0.5$ and figures (a) and (b) correspond to the cases $J_2/J_1=0$ and 1, respectively. Energy is calculated at each iteration step. Monte Carlo energy estimate converges to the exact ground state energy (red lines) up to a relative error $100\times(E^{RBM}-E^{Exact})/E^{Exact}$.
}
\end{figure}
For an ansatz wave-function $|\psi(\alpha_1,\alpha_2,...,\alpha_\mathcal{N})\rangle$ with a set of variational parameters $\{\alpha_1,\alpha_2,...,\alpha_\mathcal{N}\}$ the variational subspace is spanned by the wave-function and its derivatives with respect to each variational parameter $|\psi^i\rangle \equiv \frac{\partial |\psi\rangle}{\partial \alpha_i}$, $i=1,...,\mathcal{N}$. Within the SR method the energy of the ansatz wave-function is minimized by repeatedly applying imaginary time evolution operator $T=e^{-\tau H}\approx 1-\tau H$ to the wave-function where $\tau$, that corresponds to the learning rate in ML language, is taken to be a small positive number and operator $T$ expanded to the first order in $\tau$. After each application of $T$ the result is projected onto variational subspace to obtain a new function of the same form and a new set of variational parameters calculated using Monte Carlo sampling. Namely, after application of the evolution operator $T\approx 1-\tau H$ variational parameters $\{\alpha_1,\alpha_2,...,\alpha_\mathcal{N}\}$ are replaced by a new set of parameters $\{\alpha'_1,\alpha'_2,...,\alpha'_\mathcal{N}\}$ such that $|\psi(\alpha'_1,...,\alpha'_\mathcal{N})\rangle \equiv |\psi'\rangle $ is a good approximation to $e^{-\tau H}|\psi(\alpha_1,...,\alpha_\mathcal{N})\rangle \approx (1-\tau H) |\psi(\alpha_1,...,\alpha_\mathcal{N})\rangle \equiv |\psi_\tau^{'}\rangle$, that is, projection of $|\psi'_\tau\rangle$ to variational subspace.
\begin{figure}[b!]
\includegraphics[width=\columnwidth]{Fig5.eps}
\caption{\label{Fig:Rerr} Relative error $R=100\times(E_0^{RBM}-E_0^{ED})/E_0^{ED}$ for the ground-state energy of the Hamiltonian (\ref{eq:Hamiltonian}) obtained by RBM optimization with respect to the exact ground state energy (ED result) for $J_2=0$ and $J_2=J_1$ (blue and red data points, respectively) and a range of values of the inter-plaquette coupling parameter $\lambda$. Here results are for the system size of $N=4\times4=16$ lattice sites with periodic boundary conditions. RBM hidden unit density is $\rho=4$.
}
\end{figure}
\begin{figure}[t!]
\includegraphics[width=\columnwidth]{Fig6.eps}
\caption{\label{Fig:sscf} The ground-state spin-spin correlation matrices $M_{ij}=\langle \vec{S}_i\vec{S}_j\rangle$ for the inter-plaquette coupling strength parameter $\lambda=0.5$ in Hamiltonian (\ref{eq:Hamiltonian}) and for the system size of $N=4\times4$ lattice sites with periodic boundary conditions. Figures (a) and (b) correspond to the unfrustrated case $J_2=0$, whilst (c) and (d) correspond to the frustrated case with $J_2=J_1$. Matrices in figures (a) and (c) are obtained using RBM based VMC method and matrices in figures (b) and (d) using ED method. RBM hidden artificial neuron density is $\rho=4$.
}
\end{figure}
$|\psi'\rangle$ can be found by maximizing the overlap of the unnormalized wave-functions $|\psi'\rangle$ and $|\psi'_{\tau}\rangle$:
\begin{equation}\label{eq:overlap}
|D|^2=\frac{\langle\psi'_{\tau}|\psi'\rangle\langle\psi'|\psi'_{\tau}\rangle}{\langle\psi'_{\tau}|\psi'_{\tau}\rangle\langle\psi'|\psi'\rangle},
\end{equation}
that is, from the conditions:
\begin{equation}\label{eq:overlap_max}
\frac{\partial |D|^2}{\partial \alpha_i}=0 \; \; \; \forall i.
\end{equation}
Writing new variational parameters $\{\alpha'_1,\alpha'_2,...,\alpha'_\mathcal{N}\}$ as $\alpha'_{i}=\alpha_i+\delta\alpha_{i}$ ($i=1,...,\mathcal{N}$) and keeping only terms linear in $\tau$ and $\delta\alpha_i$ ($\forall i$) the conditions (\ref{eq:overlap_max}) can be rewritten as:
\begin{eqnarray}\label{eq:overlap_max2}
&&\sum_j \delta \alpha_j \left[\frac{\langle\psi^j|\psi^i\rangle}{\langle\psi|\psi\rangle}-\frac{\langle\psi^j|\psi\rangle}{\langle\psi|\psi\rangle}\frac{\langle\psi|\psi^i\rangle}{\langle\psi|\psi\rangle}\right]= \\
&& \tau\left[\frac{\langle\psi|H|\psi^i\rangle}{\langle\psi|\psi\rangle}-\frac{\langle\psi|H|\psi\rangle}{\langle\psi|\psi\rangle}\frac{\langle\psi|\psi^i\rangle}{\langle\psi|\psi\rangle}\right]\nonumber,
\end{eqnarray}
where $|\psi\rangle\equiv |\psi(\alpha_1,\alpha_2,...,\alpha_\mathcal{N})\rangle$ and $|\psi^i\rangle=\partial|\psi\rangle/\partial\alpha_i$.
To efficiently evaluate $\left\{\delta\alpha_1,...,\delta\alpha_{\mathcal{N}}\right\}$ numerically energy and all other expectation values in Eq. (\ref{eq:overlap_max2}) are written as Monte Carlo averages with respect to quantum probability distribution
\begin{equation}\label{eq:probability_distribution}
P(\left\{\vec{s}\right\})= \frac{|\langle\left\{\vec{s}\right\}|\psi\rangle|^2}{\langle\psi|\psi\rangle}
\end{equation}
by inserting resolution of identity summation of all possible spin configurations $\sum_{\{\vec{s}\}} |\{\vec{s}\}\rangle\langle\{\vec{s}\}|=\mathbbm{1}$ in the expectation values in Eq. (\ref{eq:overlap_max2}) where $|\{\vec{s}\}\rangle=|s_1,...,s_{\mathcal{N}}\rangle$ are $s_i^z$ basis states. The expectation values are then written as
\begin{equation}\label{eq:evH}
\frac{\langle\psi|H|\psi\rangle}{\langle\psi|\psi\rangle}=\sum_{\{\vec{s}\}}P(\{\vec{s}\})E(\{\vec{s}\})
\end{equation}
and
\begin{equation}\label{evDH}
\frac{\langle\psi|H|\psi^i\rangle}{\langle\psi|\psi\rangle}=\sum_{\{\vec{s}\}}P(\{\vec{s}\})E^{*}(\{\vec{s}\})O_i(\{\vec{s}\}),
\end{equation}
where
\begin{equation}\label{eq:Eloc}
E(\{\vec{s}\})=\frac{\langle\{\vec{s}\}|H|\psi\rangle}{\langle\{\vec{s}\}|\psi\rangle}
\end{equation}
is the local energy and
\begin{equation}
O_i(\{\vec{s}\})=\frac{\langle \{\vec{s}\}|\psi^i\rangle}{\langle\{\vec{s}\}|\psi\rangle}=\frac{\partial}{\partial \alpha_i}\ln \langle \{\vec{s}\}|\psi\rangle.
\end{equation}
The Eq. (\ref{eq:overlap_max2}) further becomes
\begin{equation}\label{eq:max_overlap3}
\sum_{j}S_{ij}\delta\alpha_j=-\tau F_i
\end{equation}
with the covariance matrix $S$ and generalized forces $F_i$ defined as
\begin{equation}\label{eq:Sij}
S_{ij}=\langle O_j^{*}O_i\rangle-\langle O_j^{*}\rangle\langle O_i\rangle
\end{equation}
and
\begin{equation}\label{eq:Fi}
F_i=\langle E_{loc}^{*}O_i\rangle-\langle E_{loc}^{*}\rangle\langle O_i\rangle.
\end{equation}
SR updates for coefficients $\alpha\equiv\{\alpha_1,\alpha_2,...,\alpha_\mathcal{N}\}$ at the $n$-th iteration are therefore of the form
\begin{equation}\label{eq:SR_update}
\alpha(n+1)=\alpha(n)-\tau S^{-1}(n)F(n).
\end{equation}
In the case of non-invertible covariance matrix $S^{-1}$ is replaced with Moore-Penrose pseudo-inverse or a positive constant is added to the diagonal $S$ matrix elements to obtain regularized matrix that is invertible.
To check validity of the RBM ansatz for the ground-state wave-function we first calculate the ground-state energy and spin-spin correlation matrices $M_{ij}=\langle\vec{S}_i\vec{S}_j\rangle$ for a small system size of $N=4\times4=16$ lattice sites with periodic boundary conditions and compare with the exact results obtained using ED method. ED and RBM optimization results (with $\tau=0.05$ in SR) are shown in FIG. \ref{Fig:E0_RBM} - FIG. \ref{Fig:sscf}. In our calculations relatively small hidden unit density $\rho=4$ for the RBM ansatz already provides the same accuracy of the results as obtained for other best known ansatz wave-functions, for example projected entangled-pair states (PEPS) or EPS.\cite{Carleo,Wang2,Mezzacapo,Mezzacapo2}
Monte Carlo energy estimate converges to the exact ground-state energy up to a relative error that is even in the frustrated regime smaller than $1.3\%$ for a wide range of values of the inter-plaquette coupling strength parameter $\lambda$. For the non-frustrated case the relative error is smaller than $0.2\%$.
We also note that the optimized ansatz for the frustrated case $J_2=J_1$ does not have discrete lattice reflection symmetry in the line $x=y$ defined as R: $(x,y)\rightarrow (y,x)$ whilst the ground-state eigenstate of the Hamiltonian obtained by ED method does have the reflection symmetry. Therefore, to compare ED and RBM results for $J_2=J_1$, we calculate the ground state spin-spin correlation matrix in FIG. \ref{Fig:sscf} (c) with the symmetrized RBM ansatz wave-function $|\psi^R_{RBM}\rangle=|\psi_{RBM}\rangle+\mathbb{R}|\psi_{RBM}\rangle$. The matrix elements are then $M_{ij}^R=\langle\psi_{RBM}|\vec{S}_i\vec{S}_j|\psi_{RBM}\rangle + \langle\psi_{RBM}|\vec{S}_{R(i)}\vec{S}_{R(j)}|\psi_{RBM}\rangle $.
\section{Supervised machine learning of the antiferromagnet to plaquette resonating valence bond state quantum phase transitions}
\label{sec:SML}
We further use SML approach to distinguish phases and locate QPTs for both frustrated and unfrustrated parameter regimes, in particular for $J_2=J_1$ and $J_2=0$. Here a phase transition occurs at some critical value $\lambda_c$ of the inter-plaquette coupling strength parameter $\lambda$. For $\lambda < \lambda_c$ the system is in a paramagnetic plaquette resonating valence bond (pRVB) phase and for $\lambda > \lambda_c$ the system exhibits antiferromagnetic ordering. When NNN coupling $J_2=0$ pRVB phase has s-wave symmetry and corresponding AFM phase N\'eel order (NAFM) whilst for the frustrated case with $J_2=J_1$ pRVB phase has d-wave symmetry and is separated by a phase transition from corresponding AFM phase with collinear striped long range order (CAFM). Single pRVB plaquette and schematic representation of classical AFM phase for both cases are illustrated in FIG. \ref{Fig:pRVB} and FIG. \ref{Fig:AFM}.
\begin{figure}[b!]
\includegraphics[width=\columnwidth]{Fig7.eps}
\caption{\label{Fig:pRVB} A singlet bond (c) and pRVB states with (a) s-wave and (b) d-wave symmetry. The s-wave pRVB state is symmetric with respect to an exchange of two diagonal spins, whilst d-wave pRVB state is antisymmetric.
}
\end{figure}
In our SML calculations gray-scale images of spin-spin correlation matrices $M_{ij}=\langle\vec{S}_i\vec{S}_j\rangle$ in AFM and pRVB part of the phase diagram are placed in two classes labeled by 0 and 1, respectively, and used as input data to train a CNN architecture that after training procedure predicts labels (classes) of new images not included in the training data set and identifies position of the QPT. Training procedure corresponds to the CNN parameters optimization with respect to a loss function here taken to be the binary cross-entropy loss function:
\begin{equation}\label{eq:binary_cross_entropy}
L = - \frac{1}{N_t}\sum_{i=1}^{N_t}[y_i\log P(y_i) + (1 - y_i)\log(1-P(y_i)]
\end{equation}
where $\log$ is a base $2$ logarithm, $y_i$ is assigned label ($1$ or $0$ for two classes), $P(y_i)$ is the predicted probability of the input from training data set labeled by $y_i$ being in class 1 and $N_t$ is the number of training data examples. The loss function decreases as the predicted probability approaches the value of the actual label and gives information how good the CNN is to make predictions on new images not included in the training data set.
Cross-entropy is a measure from information theory that calculates difference between two probability distributions for a given random variable or set of events by calculating total entropy between two distributions. Low probability surprising events contain more information whilst high probability unsurprising events contain less information where information is defined as $-\log(P(x))$ with $P(x)$ being the probability of an even $x$. For a random variable $x$ with $N_t$ discrete states (data set of size $N_t$) the entropy, which is defined as the number of bits required to encode and transmit an event randomly selected from a probability distribution, is then $-\sum_{x\in N_t}P(x)\log(P(x))$.
\begin{figure}[t!]
\includegraphics[width=\columnwidth]{Fig8.eps}
\caption{\label{Fig:AFM} Illustration of the classical ($s\rightarrow \infty$) phases for (a) $J_2=0$ with N\'{e}el antiferromagnetic order and (b) $J_2=J_1$ with collinear striped antiferromagnetic order.
}
\end{figure}
\begin{figure*}[t!]
\includegraphics[width=\textwidth]{Fig9.eps}
\caption{\label{Fig:CNN} A typical CNN architecture consists of convolutional layers interspersed with pooling layers and then followed by fully connected layers and an output classification layer. CNN input is an image that the network attempts to classify, convolutional layers extract features from the input image, pooling layers coarse-grain spatial information by performing sub-sampling, fully connected layers combine features obtained from previous convolutional and pooling layers and evaluate classification decision and the output layer predicts a class of the input image.
}
\end{figure*}
On the other hand cross-entropy calculates average number of bits needed to encode data from original distribution $P$ compared to another model distribution $Q$ that approximates distribution $P$. The cross-entropy then represents the number of additional bits to represent an event using $Q$ instead of $P$ and can be calculated as $-\sum_{x\in N}P(x)\log(Q(x))$. For classification tasks class labels are known and corresponding probability distribution has no information since the outcome is certain which means that the entropy of a known class label is always zero. When two distributions are identical the cross-entropy between them is equal to the entropy for the probability distribution, which means that the cross-entropy of real and predicted probability distributions for a class label is also zero. When cross-entropy loss function is calculated for a training data set the cross-entropy is averaged over all examples in the dataset and zero loss therefore indicates that
the predicted class probabilities are identical to the probabilities in the training dataset. Binary cross-entropy can thus be used as a loss function for a binary classification task with only two classes.
During the training process the loss function (\ref{eq:binary_cross_entropy}) is optimized using mini-batch SGD method. Within gradient descent (GD) optimization method values of the CNN parameters that minimize a loss function are found iteratively by updating each parameter (weight) $\alpha$ in the loss function in each iteration step $n$ according to the equation:
\begin{equation}\label{eq:GD}
\alpha^{n+1}=\alpha^n -\nu \frac{\partial L}{\partial \alpha},
\end{equation}
where $\nu$ is a step size hyper-parameter called learning rate in ML. The learning rate is one of the most important parameters that requires careful tuning when training an ANN to achieve fast convergence. This hyper-parameter controls the speed at which the ANN model learns. If the learning rate is too small optimization will be very slow and many iterations will be necessary to reach the minimum of the loss function, and if the rate is too large the training may not converge at all or can even diverge since weight changes can be so large that the optimizer overpasses the minimum causing an increase of the loss function value.
The mini-bath SGD algorithm is computationally more efficient than GD algorithm and allows more robust convergence. In GD optimization all training data set samples are passed through the network for a single update of the network weights. In mini-batch SGD learning algorithm\cite{Sra} instead of the whole data set a randomly selected subset of data is needed for each update where the true gradient is approximated by stochastic approximation of the gradient calculated from that randomly selected subset of data. The size of the subset is called batch size. One complete pass of all training samples through the network is called an epoch and training step usually requires more than a few epochs.
One of the most common problems in ML training step is over-fitting. Over-fit models fit very well training data. However such models are not able to classify well new data not included in the training data set. In other words over-fitting leads to a model that is very well trained however unable to generalize, that is unable to correctly predict the label of a new input sample. The model just memorizes the training data and the noise and has high variance and low bias. Variance reflects model dependence on training data set and bias reflects assumptions made about the training data. Contrary to over-fit models an under-fit model has low variance and high bias. In that case the model makes strong assumptions about data and fails to capture underlying pattern in data. Over-fitting and under-fitting can be identified by calculating validation and training accuracies.
A validation data set is a set of data for which the labels that an ANN tries to predict are known and which is excluded from the training data set. Depending on the size of the available data set for which the labels are known a validation set is formed from 10-30$\%$ of the data while the rest of the data is used for training. The accuracy of a ML classification algorithm is defined as
\begin{equation}\label{eq:accuracy}
A=\frac{N_c}{N_{tot}},
\end{equation}
where $N_c$ is the number of correct predictions and $N_{tot}$ is the total number of predictions. If the accuracies for the training and validation data sets are $A_t$ and $A_v$ over-fitting happens when $A_t > A_v$ and under-fitting when
$A_t < A_v$. Optimally $A_t\approx A_v\approx 1$. Whilst increasing the model complexity can help with under-fitting there are several methods to reduce over-fitting, like increasing the data set size, decreasing complexity of the model, or applying various regularization techniques.\cite{Sra} In our SML calculations we set the batch size to 32, and find that the loss function (\ref{eq:binary_cross_entropy}) approaches zero and $A_t$, $A_v\approx 1$ already after ten training epochs.
A typical CNN architecture consists of three kinds of layers: convolutional layers, pooling layers and fully connected layers as illustrated in FIG. \ref{Fig:CNN}.
An input gray-scale image can be represented as a matrix of pixel values where the value of each pixel in the matrix ranges from 0 to 255 with zero indicating black and 255 indicating white. A color image can be represented with three 2D matrices each having pixel values in range 0 to 255, one for each color - red, green or blue. A color image is said to have three channels (each channel representing certain component of an image), whilst a gray-scale image has just one channel.
In image processing a kernel or convolution matrix is a small matrix used for achieving a wide range of effects on original image, for example blurring, sharpening, embossing, and edge detection. In CNN architectures and ML such kernels are used for feature extraction, that is for determining the most important parts of an image, the process more generally referred to as convolution. Convolution is the operation of adding each element of the image to its local neighbor weighted by the kernel. The general form for convolution (*) of two matrices A and B is:
\begin{eqnarray}\label{eq:MatrixConvolution}
&&\begin{bmatrix}
A_{11} & A_{12} & . & . & . &A_{1n} \\
A_{21} & A_{22} & & & &A_{2n}\\
. & . & . & & & . \\
. & . & &. & & . \\
. & . & & & . & .\\
A_{m1} & A_{m2} & . & . & . &A_{mn}
\end{bmatrix}
*
\begin{bmatrix}
B_{11} & B_{12} & . & . & . &B_{1n} \\
B_{21} & B_{22} & & & &B_{2n}\\
. & . & . & & & . \\
. & . & &. & & . \\
. & . & & & . & .\\
B_{m1} & B_{m2} & . & . & . &B_{mn}
\end{bmatrix} \nonumber \\
&=&\sum_{i=0}^{m-1}\sum_{j=0}^{n-1}A_{(m-i)(n-j)}B_{(1+i)(1+j)}
\end{eqnarray}
Within a convolution step in CNN architecture a small kernel matrix is slid across the image and convolved with small squares of input data as shown in FIG. \ref{Fig:Convolution}. In ML algorithm CNN can learn kernel values that can extract important features. Since image features are learnt using small portions (squares) of input data convolution preserves spatial relationship between pixels. The region of the input image connected to a hidden neuron is called local receptive field for the hidden neuron. If the local receptive field is moved by S pixels at a time we say a stride length of S is used. In FIG. \ref{Fig:Convolution} stride length is $S=1$. In practice stride lengths of $S\geq 3$ are rarely used. Also, output volume spatial size can be controlled by padding the input with zeros on the border of the input volume. In our calculations kernel matrix is a $3\times 3$ matrix and the stride length is set to $S=1$.
Whilst for a two-dimensional (2D) convolutions, kernel is a $\mathcal{H}\times \mathcal{W}$ matrix, a filter is defined as a concentration of multiple kernels, each kernel assigned to a particular channel of the input. For $k$ input channels, the filter dimension is therefore $k\times \mathcal{H}\times \mathcal{W}$. A general convolutional layer in a CNN architecture consists of multiple such filters. Filters' shapes and sizes are chosen based on the dataset to create abstractions and recognize features at the proper scale.
\begin{figure}[t!]
\includegraphics[width=\columnwidth]{Fig10.eps}
\caption{\label{Fig:Convolution} Kernel (convolution matrix, feature detector) is a matrix of weights which are multiplied with the input to extract relevant features. A small kernel matrix is slid across the image and convolved with a restricted area of input data (typically of a square shape). Here input image is a gray-scale image with only one channel. The region in the input image connected to a hidden neuron is called local receptive field for the hidden neuron. The process of finding features within a certain window of data is continued until the entire image is covered and a feature map is obtained which is represented by simplified set of pixel values.
}
\end{figure}
Each hidden neuron has a bias and weights connected to its local receptive field as illustrated in FIG. \ref{Fig:Convolution}. It is important to note that within a CNN the same weights and bias are used for each hidden neuron in a given feature map, that is, for $j,k$-th hidden neuron the output is:
\begin{equation}
\sigma\left(b+\sum_{l=0}^{\mathcal{W}}\sum_{m=0}^{\mathcal{H}} w_{l,m}a_{j+l,k+m}\right)
\end{equation}
Here $w_{l,m}$ is an array of shared weights, $b$ is shared bias, $a_{x,y}$ is the input activation (value) at position $x,y$ and $\sigma$ is the neural network activation function like, for example, sigmoid function $\sigma(x)=1/(1+e^{-x})$ or rectified linear unit (ReLU) function $\sigma(x)=\max(0,x)$. In practice ReLU activation is often preferred to other activation functions because it allows faster neural network training without significant penalty to generalization accuracy. In our calculations activation function for all neurons in convolutional layers and fully connected layers is ReLU, whereas activation function of the output layer neurons is the softmax function $\sigma(\vec{o})=e^{o_i}/\sum_{i=0}^{N-1} e^{o_i}$ where $\vec{o}=\{o_0,...,o_{N-1}\}$ is the output vector of the network with $N$ output classes and $N=2$ for the problem that we have studied. The softmax function is used as an activation for the output layer to normalize total probability over predicted output classes to one.
In other words, each neuron, illustrated in FIG. \ref{Fig:Neuron}, is a mathematical function that multiplies inputs by weights, adds them together, adds a bias $b$ which can help better fit the data, and then passes the sum to a non-linear (activation) function to become the neuron's output. A non-linear function is applied by a neuron to introduce non-linear properties in the network and allow the network to capture more complex patterns.
\begin{figure}[t!]
\includegraphics[width=\columnwidth]{Fig11.eps}
\caption{\label{Fig:Neuron}Schematic representation of an artificial neuron. Neurons consist of a linear transformation that weights the importance of various inputs and a non-linear activation function $\sigma$ which allows the network consisting of these neurons to capture more complex patterns. Neurons are generally arranged in layers with the output of one layer serving as the input to the next layer.
}
\end{figure}
\begin{figure}[b!]
\includegraphics[width=\columnwidth]{Fig12.eps}
\caption{\label{Fig:Pooling} Pooling step in a CNN architecture. A pooling layer reduces dimensions of each feature map output from the previous convolutional layer and creates condensed feature maps whilst keeping the most important information. For example max pooling step for $2\times2$ clusters replaces the clusters with the maximum value from each cluster. To avoid information loss we exclude pooling layers from the convolutional neural network in our calculations.
}
\end{figure}
A parameter sharing scheme used in CNNs is based on assumption that if a feature is useful to compute in a region of input (local receptive field) at some spatial position, then it should also be useful to compute at other positions. Such parameter sharing contributes to the translation invariance of the CNN architectures.
For image recognition more than one feature map is needed and a complete convolutional layer, which is the core building block of a CNN, consists of several different feature maps where each feature map is defined by a set of shared weights and a shared bias. Depth of a convolutional layer corresponds to the number of filters in that layer. The number of feature maps necessary in a given convolutional layer depends on the number of available examples and task complexity.
To summarize, the primary purpose of convolution is to extract features from the input image so that CNN can learn the values of feature filters (analyze important features of the image) on its own during the training process. More filters correspond to more image features extracted which allows the network to better recognize patterns in unseen images.
\begin{figure}[t!]
\includegraphics[width=\columnwidth]{Fig13.eps}
\caption{\label{Fig:Pswave} Probabilities of the s-wave pRVB ($P_{pRVB}$) and NAFM phases ($P_{AFM}=1-P_{pRVB}$) for the unfrustrated case with the NNN coupling $J_2=0$ in the Hamiltonian (\ref{eq:Hamiltonian}). The probabilities are calculated as averages over hundred CNN output label predictions trained in each case with slightly different training data set. Corresponding variances of the predictions are shown in inset. Standard error for each prediction is $\sigma/\sqrt{n}$ with $n=100$. The results are calculated for the system sizes $N=L\times L$ lattice sites with $L=6,8,10$ and with periodic boundary conditions using SML algorithm with the CNN architecture consisting of three convolutional layers with 256 filters and one fully connected layer with 256 neurons. The output layer has two neurons, one for each phase class. The ground-state spin-spin correlation matrices are calculated with RBM based RL algorithm in the inter-plaquette coupling strength parameter $\lambda\in\left[0,0.2\right]$ and $\lambda\in\left[0.8,1.0\right]$ intervals (2000 matrices in each interval) that do not contain the transition point and random 10\% of data separated as a validation data set. The remaining 90$\%$ of data is then used to train the network. The network then predicts class labels for 600 matrices calculated in the interval $\lambda\in\left[0.2,0.8\right]$ and detects the QPT which corresponds to the point where the network is maximally confused ($P_{pRVB}=P_{AFM}=0.5$). The solid lines are fits to a logistic function (smooth approximation to the Heaviside step function) $f(x)=1/(1+e^{-k(x-x_0)})$ and $1-f(x)$. For the studied system sizes $\lambda_c (L) \approx 0.573$, $0.537$ and $0.465$ for $L=6$, $8$ and $10$, respectively.
}
\end{figure}
In addition to convolutional layers CNNs often also contain pooling layers, usually a pooling layer after each convolutional layer. The role of a pooling layer is to simplify the output from the convolutional layer and create condensed feature maps by applying spatial pooling (also called sub-sampling or down-sampling) to each feature map separately. Pooling attempts to reduce the dimensions of each feature map while keeping the most important information. There are several different types of pooling, for example max, average and sum pooling. In a pooling step small regions of output from a convolutional layer are replaced by maximum, average or sum of values in each region as illustrated in FIG. \ref{Fig:Pooling}. Pooling layers therefore coarse-grain convolutional layer output while maintaining spatial structure and locality. This pooling step is very similar to RG decimation step.\cite{Mehta,Janusz,Kadanoff1,Kadanoff2,Efrati,Iso,Lin}
Translational invariance built in CNN architecture, that is based on assumption that relative positions of various features are more important than exact locations of each feature, is achieved by a combination of convolutional and pooling layers. However a major issue of adding pooling layers in a CNN is that pooling can result in excess information loss. Namely, CNN with pooling layers fails to learn precise spatial correlations between different features and can thus predict inaccurate class label if all features are present even if those features are at wrong spatial positions. Therefore discarding pooling layers in CNN architecture can sometimes lead to better results. To avoid information loss in our SML calculations we thus exclude pooling layers.
Each convolutional and pooling step is a hidden layer and such hidden layers in a CNN architecture are usually followed by fully connected hidden layers and a final output layer. Whilst in a convolutional layer neurons receive input only from a subarea (receptive field) of the previous layer in a fully connected layer each neuron is connected and receives input from every neuron of the previous layer. Although high-level features obtained from convolutional and pooling layers may be sufficient for classifying an input image into different classes based on the training data set, non-linear combinations of such features might be better for the image classification. Purpose of the fully connected layer is therefore to learn non-linear feature combinations relevant for the classification task, that is prediction of the best label for an input image.
\begin{figure}[b!]
\includegraphics[width=\columnwidth]{Fig14.eps}
\caption{\label{Fig:Pdwave} The same as in FIG. \ref{Fig:Pswave} for the frustrated case with the NNN coupling $J_2=J_1$ in the Hamiltonian (\ref{eq:Hamiltonian}). Here the probabilities are for the d-wave pRVB ($P_{pRVB}$) and CAFM phases ($P_{AFM}=1-P_{pRVB}$) and $\lambda_c (L) \approx 0.413$, $0.371$ and $0.355$ for $L=6$, $8$ and $10$, respectively. The probabilities and fits are calculated following the same procedure as in FIG. \ref{Fig:Pswave}.
}
\end{figure}
\begin{figure}[t!]
\includegraphics[width=\columnwidth]{Fig15.eps}
\caption{\label{Fig:PswaveL10} Probabilities of the s-wave pRVB ($P_{pRVB}$) and NAFM phases ($P_{AFM}=1-P_{pRVB}$) for the unfrustrated case with NNN couplings $ J_2=0$ in the Hamiltonian (\ref{eq:Hamiltonian}) and the system size $N=10\times 10$ lattice sites with periodic boundary conditions for three different CNN architectures. The probabilities and fits are calculated following the same procedure as in FIG. \ref{Fig:Pswave}. All three architectures have one fully connected layer and three convolutional layers. Red and blue data points correspond to a CNN with 256 filters in each convolutional layer and 256 neurons in the fully connected layer, magenta and purple data points to a CNN with convolutional layers containing 64, 128 and 256 filters and one fully connected layer with 256 neurons, and orange and green data points to a CNN with 128 filters in each convolutional layer and 128 neurons in the fully connected layer. All three network architectures give similar value of the inter-plaquette coupling strength parameter at which the QPT occurs: $\lambda_c(L)\approx 0.465$, $0.476$ and $0.458$.
}
\end{figure}
The final layer of a CNN is the output layer where each neuron has softmax activation function that predicts a single class of $K$ mutually exclusive classes and normalizes total probability over predicted output classes to one. In our calculations the number of neurons in the output layer is two corresponding to the number of different phase labels of the input data set. Output probabilities $P_1$ and $P_2$, with $P_1+P_2=1$, correspond to the probabilities of the system to be in AFM or pRVB phase.
In our SML calculations the CNN consists of three convolutional layers with each layer having 256 filters, one fully connected layer with 256 neurons and an output layer that classifies an input image in one of the two classes corresponding either to pRVB or AFM phase. Calculations are performed for the system sizes of $N=L\times L$ lattice sites with $L=6$, $8$ and $10$ and with periodic boundary conditions where for each $L$ the input spin-spin correlation matrix images are resized to $L\times L$ pixel values (using OpenCV Python machine learning software library resize function). The ground-state spin-spin correlation matrices are calculated with the RBM based RL algorithm in the inter-plaquette coupling strength parameter $\lambda\in\left[0,0.2\right]$ and $\lambda\in\left[0.8,1.0\right]$ intervals (2000 matrices in each interval) that do not contain the transition point and random 10\% of data separated as a validation data set. The remaining 90$\%$ of data is then used to train the network. The network then predicts class labels for 600 matrices calculated in the interval $\lambda\in\left[0.2,0.8\right]$ and detects the QPT. The probabilities are calculated as averages over hundred CNN output label predictions, each corresponding to one random choice of 10$\%$ of the validation data samples and the network training with the remaining 90$\%$ of data calculated in intervals $\lambda\in\left[0,0.2\right]$ and $\lambda\in\left[0.8,1.0\right]$. Obtained probabilities are shown in FIG. \ref{Fig:Pswave} for the unfrustrated case with the NNN couplings $J_2=0$ and in FIG. \ref{Fig:Pdwave} for the frustrated case with $J_2=J_1$. The QPT point corresponds to the point where the network is maximally confused, that is to the point where $P_{pRVB}=P_{AFM}=0.5$.
To achieve accurate extrapolation of the results to the thermodynamic limit ($L\rightarrow\infty$) and find $\lambda_c(L\rightarrow \infty)$ values it would be necessary to do all presented calculations for much larger system sizes. Particularly difficult and time consuming task would be to generate large enough training data sets for even larger system sizes with described RBM based RL method. However even from the presented results for the system sizes of $N=L\times L$ lattice sites with $L=6$, $8$ and $10$ and periodic boundary conditions it is clearly visible that the obtained values for $\lambda_c(L)$ as the system size is increased rapidly approach the $\lambda_c(L\rightarrow \infty)$ values obtained previously with other methods. In particular the values obtained with the CCM and the real-space RG method are $\lambda_c^{CCM}(L\rightarrow\infty)= 0.47$\cite{Gotze} and $\lambda_c^{RG}(L\rightarrow\infty)=4822$\cite{Fledderjohann} for the unfrustrated case with $J_2=0$, and $\lambda_c^{CCM}=0.301$\cite{Gotze} for the frustrated case with $J_2=J_1$. For the largest system size of $N=10\times 10$ lattice sites with periodic boundary conditions we obtain the values $\lambda_c\approx 0.465$ and $\lambda_c\approx 0.355$ for $J_2=0$ and $J_2=J_1$, respectively, as shown in FIG. \ref{Fig:Pswave} and FIG. \ref{Fig:Pdwave}.
However, the result of the QMC and finite-size scaling method for the unfrustrated case is $\lambda_c^{QMC}\approx 0.5485$.\cite{Albuquerque,Wenzel,Ran} Our result that is still in reasonable agreement with the QMC result, however indicates slightly larger NAFM region than QMC calculations. To reconfirm our result for the unfrustrated case we perform further calculations with two additional CNN architectures. Both additional network architectures have one fully connected layer and three convolutional layers. The first additional architecture has 128 filters in each convolutional layer and 128 neurons in the fully connected layer, and the second architecture 64, 128 and 256 filters in three convolutional layers and 256 neurons in the fully connected layer. The results for the system size of $N=10\times 10$
lattice sites are shown in FIG. \ref{Fig:PswaveL10}.
As expected for the initial CNN architecture with 256 filters in each convolutional layer variances of the predictions are smaller than for the other two CNN architectures with reduced number of filters in convolutional layers, since increasing the number of filters improves the accuracy of the predictions resulting in steeper probability lines near the transition. Nonetheless all three architectures give similar values of $\lambda_c(L)$ confirming that over-fitting does not happen in the more complex CNN model training step. Whilst the QMC results suggest that the real space RG results converge non-monotonously towards the value obtained with the QMC calculations when further excited states with more that one quintuplet are included in the RG calculations,\cite{Fledderjohann} our results indicate that the RG results might converge monotonously towards the value close to the value obtained with the CCM.
\section{Conclusions}
\label{sec:Conculsions}
In conclusion we have studied the spin-$1/2$ square lattice $J_1$-$J_2$ model with additional plaquette structure, relevant for the high-temperature superconducting materials and realizable in cold atom experiments, using ML techniques. We have demonstrated the power of these techniques, in particular RBM bases RL algorithm and CNN based SML method, to study complex phases and phase transitions in challenging many-body problems even in the presence of frustration where other important techniques like QMC are limited. Our results show good agreement with previously obtained results, particularly with the CCM and the real-space RG results.
An interesting direction for the future research would be to explore the power of the presented ML techniques to study dynamical properties of similar complex many-body quantum systems, particularly when the presence of certain dynamical symmetries leads to breaking of the continuous time-translation symmetry and results in existence of a set of observables that can enter robust non-equilibrium limit cycles.
\begin{acknowledgments}
We thank Tomislav \v{S}eva, Hrvoje Buljan, Robert Pezer, and Danko Radi\'{c} for helpful discussions. This work was supported by the QuantiXLie Centre of Excellence, a project cofinanced by the Croatian Government and European Union through the European Regional Development
Fund - the Competitiveness and Cohesion Operational Programme (Grant KK.01.1.1.01.0004).
\end{acknowledgments}
|
\section*{Introduction}
The physics of quasiparticles constitutes a very prominent research field over the last decade \cite{QP_Zoo}. Especially magnons and phonons---the quanta of spin waves \cite{Dyson1956} and lattice vibrations \cite{Einstein1907}---have been extensively studied within the solid state research domain.
In the early days, these two systems were usually considered to be non-interacting.
Subsequently, the emphasis shifted to a full description of the solid-state system and the study of related subsystems \cite{Tiersten1964,Kobayashi1973I,Kobayashi1973II,Dransfeld1959,Pomerntz1961,Rezende1969,Rueckriegel2014,Kikkawa2016,Baryakhtar2017}.
The variety of interactions between these two subsystems creates a wide range of applications \cite{Bozhko2020} exploiting the fact that in magnetostrictive materials the mechanical stress affects the magnetization orientation and vice versa \cite{Olabi2008,Chumak2010,Domann2015,Kamra2015,Guerreiro2015,Kryshtal2017,An2020,Zhao2020}.
For instance, the mechanical stress produced by an acoustic wave can drive magnetization dynamics and excite magnons in such materials if the frequency of lattice vibrations matches the eigenexcitations of the spin system \cite{Geilen2020,Weiler2012}.
Furthermore, hybrid waves \cite{Kittel1958,Strauss1965,Camley1978,Camley1979} with mixed magnonic and phononic features \cite{Kamra2014,Bauer2015,Ogawa2015,Hashimoto2017,Weiler2020} also exist. The corresponding quasiparticles are called magneto-elastic bosons \cite{Bozhko2017} or magnon polarons \cite{Kikkawa2016,Sukhanov2019,Yahiro2020}.
Due to their mixed hybrid state between phonons and magnons, they are capable of carrying spin information at velocities close to those of phonons. This is currently attracting much attention to these quasiparticles as promising data carriers for future applications in spintronics \cite{Kamra2014,Ogawa2015,Flebus2017,Ruckriegel2020,Holanda2021}.
In particular, the spontaneous bottleneck accumulation of hybrid magneto-elastic quasiparticles with rather high group velocities was recently discovered in an overpopulated magnon gas in yttrium iron garnet (YIG) films \cite{Bozhko2017}.
\begin{figure} [t]
\includegraphics[width=8.6cm]{Dispersion.pdf}
\caption{
\label{dispersion}
A magnon-phonon hybrid spectrum of an in-plane magnetized 5.6\,$\mu$m-thick YIG film calculated for wavevectors $\pm q_\mathrm{x}$ and $\pm q_\mathrm{y}$ oriented along and perpendicular to the bias magnetic field, respectively. The external magnetic field is $\mu_0 H = $135\,mT. The violet arrow indicates the parallel parametric pumping process in the $\textbf{\textit{q}}_\mathrm{m} \perp \textbf{\textit{H}}$ magnon branch using microwave photons of frequency $f_\mathrm{p}$. The population of the pumped magnons at frequency $f_\mathrm{p}/2$ (purple dot) thermalizes over the spin-wave spectrum via four-magnon scattering processes. The thermalization process leads to the accumulation of magneto-elastic quasiparticles within the magnon-phonon hybridization area (dashed square) and magnon Bose-Einstein condensation at the bottom of the spin-wave spectra (blue dot). The green and red dots mark the spectral positions of the measured hybrid quasiparticles. The color map shows the thermal magnon-phonon density distribution measured by means of frequency- and wavevector-resolved Brillouin light scattering spectroscopy.
}
\end{figure}
In this paper, we report measurements of the two-dimensional transport properties of the accumulated quasiparticles using space- and time-resolved Brillouin light scattering (BLS) spectroscopy. We show that the accumulation process leads to the appearance of two---slow and fast---groups of quasiparticles, represented by the green and red dots in the black dashed square in Fig.\,\ref{dispersion}, propagating in distinctly different directions. The slow group, indicated by the red dot, has a frequency slightly higher than that of the magnon Bose-Einstein condensate (BEC) \cite{Demokritov2006,Serga2014} formed at the bottom of the spin-wave spectrum. The frequency of the fast group of hybrid bosons, shown by the red dot, is lower then the BEC frequency.
We also present an analytical model that accounts for the details of the relaxation dynamics in the presence of a nonzero BEC frequency minimum including variation of the scattering coefficient in the region of the avoided crossing between phonon and magnon branches. The model qualitatively explains the observed phenomenon of double accumulation and the spectral positions of the two quasiparticle groups.
Our analysis of the two-dimensional hybrid frequency spectra quantitatively relates the values and directions of quasiparticle velocities to the formation of caustic beams \cite{Veerakumar2006,Serga2010} in the spatial distribution of the quasiparticle density. The presented findings will potentially allow to control the characteristics of the accumulated quasiparticles and manipulate their behavior for information transfer and processing \cite{Holanda2018,Heussner2018,Hioki2020}.
The paper is organized as follows. The experimental setup is described in \Sec{ss:setup}. In \Sec{ss:results}, the experimental results on propagation velocities of the slow and fast quasiparticle packets are presented. In \Sec{s:background}, we refer to the basics of magnon-phonon hybridization theory by deriving a hybridized magnon-phonon Hamiltonian in \Sec{ss:hybrid} and presenting a statistical description of magneto-elastic modes in \Sec{ss:stat}. The model of double accumulation of hybrid bosons is presented in \Sec{s:model}, where we introduce a one-dimensional differential model in \Sec{ss:dif}, estimate the scattering rates within and between the quasiparticle spectral branches in \Sec{ss:scattering}, and solve the rate equations above and below the BEC frequency in \Sec{ss:sol-1} and \Sec{ss:sol-2}, respectively. Analysis of the experimental data based on the theory of caustic magnon transport in \Sec{s:2D} allows us to determine frequencies and wavevectors of the two groups of hybrid quasiparticles. In \Sec{s:discussion}, we discuss and summarize the obtained results.
\vspace{-3mm}
\section{Experiment}
\vspace{-3mm}
\subsection{\label{ss:setup} Setup}
\vspace{-3mm}
We studied the process of spontaneous accumulation of hybrid quasiparticles and determined their properties by means of BLS spectroscopy \cite{Sandercock}. The sample is illuminated by focused laser light with a wavelength of 532\,nm. Due to inelastic scattering processes, a photon can absorb or excite a magnon and therefore gain (anti-Stokes process) or lose energy (Stokes process). Thus, the intensities of the Stokes and anti-Stokes spectral peaks are proportional to the density of quasiparticles. The backward scattered light is collected and directed to a tandem Fabry-P\'{e}rot interferometer, where its frequency spectrum is analyzed with a resolution of about 100\,MHz. By carefully selecting the incident angle $\alpha_x$ of the laser light along the external magnetic field we achieve a wavevector selection (see Fig.\,\ref{section}) \cite{Sandweg2010,Bozhko2020_unconventional}. The linear response function of the BLS setup in the wavevector plane $(q_x, q_y)$ follows a Gaussian distribution and consequently limits the wavevector resolution with a standard deviation of $\sigma$ = 1500\,rad/cm. To freely move the sample along all three spatial dimensions it is mounted on a motorized linear stage system. Additionally, a time-resolution system with a resolution of 400\,ps is combined with the interferometer and a pulsed microwave source \cite{Buettner2000}. Therefore, the setup is able to perform \mbox{frequency-,} \mbox{wavevector-,} \mbox{time-,} and space-resolved scans \cite{THATec}.
\begin{figure}[b]
\includegraphics[width=8.6 cm]{Section.pdf}
\caption{
\label{section}
Sketch of the experimental setup. The microwave part consist of a microwave stripline, which blends into a half-loop with its plane oriented perpendicular to the sample surface and the external magnetic field $\textbf{\textit{H}}$. The sample is mounted on top of two distant holders in such a way that the center of the sample is located above the excitation loop. The sample itself consists of a YIG film of 5.6\,$\mathrm{\mu m}$ thickness grown on a GGG substrate (500\,$\mathrm{\mu m}$) and covered by a thin dielectric mirror coating ($< 1 \,\mathrm{\mu m}$). The magnons in the YIG film are detected via inelastic laser light scattering ($\lambda=532$\,nm) measured by a tandem Fabry-P\'{e}rot interferometer. The value of an incident angle $\alpha_x$ of the probing laser beam in the $(x,z)$-plane defines the measured wavevector $q_x$ oriented along the magnetic field.
}
\end{figure}
\begin{figure*}
\includegraphics[width=1\textwidth]{TimeshootsLOG_norm.pdf}
\caption{
\label{timeshoots}
Spatial distribution of the BLS-signal intensity corresponding to the distribution of the quasiparticle density for different times from the moment in time the pump pulse is turned on. The propagation of two groups of quasiparticles is visible. The magnon-phonon hybrid packets propagate away from the excitation loop in six directions. Each of the shown intensity distributions are a summation of the BLS experiment's Stokes ($+q_x$) and anti-Stokes ($-q_x$) signals.
}
\end{figure*}
The measurements are carried out at room temperature in a single-crystal Yttrium Iron Garnet (YIG: Y$_3$Fe$_5$O$_{12}$) sample. This ferrimagnetic material was chosen for the experiment because it uniquely combines extremely low damping of magnetic and elastic excitations \cite{Cherepanov1993}. The sample is a piece of 5.6\,$\mu$m-thick YIG film grown by liquid phase epitaxy on top of a 500\,${\mu}$m-thick Gadolinium Gallium Garnet (GGG: Gd$_3$Ga$_5$O$_{12}$) substrate in (111) crystallographic plane (see Fig.\,\ref{section}). Micron-thick YIG films as well as GGG substrates are transparent to green light and thus make BLS probing of magnon-phonon hybrid waves possible. The probing light beam is focused onto the YIG film through the GGG substrate. To ensure homogeneous reflection of the probing light from the YIG-film surface and thus uniform spatial sensitivity of the BLS setup, YIG is covered with a thin dielectric mirror coating created with a SiO$_2$/TiO$_2$ bi-layer structure ($<$1\,$\rm{\mu m}$). The mirror reflects over 90\,$\%$ light at 532\,nm.
Magnons are excited in the in-plane magnetized YIG sample by microwave electromagnetic pumping via the parallel parametric pumping process. In such a process \cite{Lvov1994}, one microwave photon of the pumping frequency $f_\text{p}$ splits into two magnons with opposite wavevectors $\pm \textbf{\textit{q}}_\mathrm{m}$ and each the frequency $f_\text{p}/2$.
Due to the choice of the values of the bias magnetic field $\mu_0 H = 135\,$mT and the pumping frequency $f_\text{p}=14\,$GHz, the parametrically excited magnons are injected (see violet arrow in Fig.\,\ref{dispersion}) into the transverse branch ($\textbf{\textit{q}}_\mathrm{m} \perp \textbf{\textit{H}}$) of the spin-wave frequency spectrum \cite{Serga2012} at a distance of about 2\,GHz above the bottom of the spectrum.
In this experiment, to achieve two-dimensional spatial localization of the pumping microwave field required for two-dimensional transport measurements, the microstrip pump resonator used in our previous experiments \cite{Bozhko2017,Bozhko2018} is replaced by a half-loop antenna with a diameter of 100\,$\mu$m (see Fig.\,\ref{section}). The half-loop is made of a gold wire with a diameter of 25\,$\mu$m and is shorted to the ground of the microstrip line that feeds the pumping power to it.
Since the quality factor of this antenna is unity, the pumping power required to reach a pronounced accumulation of hybrid quasiparticles and the Bose-Einstein condensation of magnons is about 400\,W.
This power is applied during a 200\,ns-long excitation pulse. To ensure that no additional heating effects influence the measurements, the excitation pulse is followed by a 250\,$\mu$s idle time, where the magnons can decay and the created heat can dissipate into the surroundings.
\begin{figure}
\includegraphics[width=8.6 cm]{Space-time_diagrams.pdf}
\caption{
\label{space-time_diagrams} Space-time diagrams along the $x$- and $y$-axes in the film plane for packets of fast and slow quasiparticles, which correspond to Stokes components of the BLS signals and thus have positive values of wavevectors $q_x$. The green and red dashed lines show a linear temporal fit of the positions of these packets.
Two packets of quasiparticles with wavevectors $\pm q_y$ propagate with the same speed in opposite directions along the $y$ axis, and two packets -- along the $x$ axis with different speeds.}
\end{figure}
\vspace{-3mm}
\subsection{\label{ss:results}Experimental results}
\vspace{-3mm}
The measurement cycle begins at $t=0$, when the pumping power is switched on and when a microwave pulse begins to inject magnons into the system. From thereon the magnon gas is formed and it thermalizes via four-magnon scattering processes \cite{NSW,Serga2014}.
The thermalization process causes an increase in the local chemical potential of the injected magnon gas in the vicinity of the frequency minimum. As soon as it reaches the minimum of the magnon frequency, a magnon BEC forms at the spot marked by the blue dot in Fig.\,\ref{dispersion}.
Another cluster of quasiparticles is formed at the intersection of the phonon and magnon branches \cite{Bozhko2017}. Quasiparticles belonging to this cluster are in hybrid states between phonons and magnons, combining the properties of both. Due to the rather small size of the hybridization region in the frequency and wavevector space, the BLS setup is not able to differentiate the population of quasiparticles within the magneto-elastic interaction region. However, time-resolved measurements of the spatial distribution of these quasiparticles \mbox{[see Figs.\,\ref{timeshoots}(a-l)]} reveal the presence of two groups of quasiparticles with significantly different propagation properties.
Figure\,\ref{timeshoots}(a) depicts the beginning of quasiparticle accumulation in real space in the vicinity of the pump loop at time $t=120\,$ns after the start of pumping. Here, the loop position is shown by the gold segment. During the pumping pulse ($t\le 200\,$ns) the number of particles in the hybridization region rises [see Figs.\,\ref{timeshoots}(a,\,b)], and two beams of quasi-particles start to travel upward and downward along the direction of the external magnetic field $\B H$ as shown in Figs.\,\ref{timeshoots}(b,\,c). The white arrows mark the Stokes data, indicating the quasiparticles propagating with a positive wavevector of $q =q\sb{fast}$, while the blue arrows for anti-Stokes data correspond to the quasi particles propagating with a negative wavevector $q=-q\sb{fast}$.
Beginning at $t=360$\,ns [see Fig.\,\ref{timeshoots}(d)], four additional magnon packets propagate obliquely to the external magnetic field as it is marked by two white and two green arrows in Figs.\,\ref{timeshoots}(e-l)]. The speed of their propagation is approximately ten times lower than the speed of propagation of quasiparticle packets along the field. Hereinafter we will refer to these two groups of quasiparticles as fast and slow ones.
To analyze the propagation process of the fast and slow packets, a center-of-mass algorithm tracks the path of each of those packets individually. The space-time diagrams for the fast and slow packets in $x$- and $y$-directions are shown in Fig.\,\ref{space-time_diagrams}(a) and Fig.\,\ref{space-time_diagrams}(b).
The packets positions $\B R(t)=\{X(t), Y(t)\}$, represented by the red and green arrows, are found by the two-dimensional linear in time fitting process $\B R(t)= \B v t + \B R_0$ with the Levenberg-Marquardt algorithm \cite{More1977}:
\begin{subequations}\label{V}
\begin{eqnarray}\label{slow}
v\sb{slow}&\approx& (368\pm 2) \mbox{m/s}, \qquad \alpha\sb{slow}\approx 59^\circ\pm 0.3^\circ\,,\\
\label{fast}
v\sb{fast}&\approx& (3070\pm 207) \mbox{m/s}, \ \ \alpha\sb{fast}\approx 0^\circ\pm 0^\circ\ .
\end{eqnarray}
\end{subequations}
We see that the velocity of the slow quasiparticle group is higher than that of pure magnons ($\approx 90$\,m/s) with wavevectors corresponding to the hybridization region, while the velocity of the fast group is lower than that of pure phonons ($\approx 3850$\,m/s in bulk YIG). This comparison of velocity values allows us to approximately localize the spectral positions of the fast and slow quasiparticles in the hybridization region--indicated by the red and green dots inside the black dashed box in Fig.\,\ref{dispersion}. It can be assumed that the slow quasiparticles are mainly of magnonic nature with a small contribution of phononic admixture. In turn, the fast quasiparticles predominantly consist of phonons with a small magnonic contribution. Particularly remarkable is the absence of the population of quasiparticle states between these two---quasi-magnon and quasi-phonon---extreme cases. Because of close spectral positions of the quasi-magnon and quasi-phonon states, such selective population is difficult to explain by some features of scattering of parametrically pumped magnons far from the bottom of the spin-wave spectrum. Thus, to understand the observed phenomenon of the double quasiparticle accumulation, a theoretical analysis of the situation near the magnon-phonon hybridization region is required. It is presented in the following sections.
\section{\label{s:background} Magnon-phonon hybridization background}
\vspace{-3mm}
The goal of this section is to recall the Hamiltonian description of interacting magnons and photons in the vicinity of cross-section of their frequency spectra\,\cite{ZLF,NSW}, required for further discussion of our experimental finding. Some details on statistical descriptions of resulting magneto-elastic modes can be found in the Appendix of \Refn{arc-1}.
\vspace{-3mm}
\subsection{\label{ss:hybrid}Hybridized magnon-phonon Hamiltonian}
\vspace{-3mm}
At high temperatures we can restrict ourselves to the classical limit and describe the system of interacting magnons and phonons in the framework of a classical Hamiltonian formalism. This approach is applicable to a wide class of weakly interacting wave systems, allowing a physically transparent and very compact description of their common properties, see, for instance, Chapter\,1 in \Refn{NSW} and in more detail in \Refn{ZLF}.
Introducing complex canonical amplitudes of magnons and phonons in the wavevector $\textbf{\textit{q}}$ representation, $a(\textbf{\textit{q}},t) \equiv a_\textbf{\textit{q}}$ and $b(\textbf{\textit{q}},t) \equiv b_\textbf{\textit{q}}$ we can write their Hamiltonian equation of motion as follows:
\begin{subequations}\label{Ham}
\begin{equation}\label{HamA}
i\, \frac{\partial a_\textbf{\textit{q}}}{\partial t}=\frac{\partial \mathcal{H}}{\partial a^*_{\textbf{\textit{q}}}}\,, \quad i\, \frac{\partial b_\textbf{\textit{q}}}{\partial t}=\frac{\partial \mathcal{H}}{\partial b^*_\textbf{\textit{q}}}\, \ .
\end{equation}
Here, the Hamiltonian function $\mathcal{H}$ (hereafter referred to as``the Hamiltonian'' for brevity) is a functional of the canonical variables $a_\textbf{\textit{q}}$, $b_\textbf{\textit{q}}$ for all $\B q$ and their complex conjugated counterparts $a^*_\textbf{\textit{q}}$, $b^*_\textbf{\textit{q}} $. We chose $\mathcal{H}$ as
\begin{eqnarray}\label{HamB}
\mathcal{H} &=& \mathcal{H}_2+ \mathcal{H}_4\,, \\
\mathcal{H}_2&=&\sum_\textbf{\textit{q}}\Big[\omega^\textrm{m}_\textbf{\textit{q}}a_\textbf{\textit{q}}a^*_\textbf{\textit{q}}+ \omega^\textrm{p}_\textbf{\textit{q}}b_\textbf{\textit{q}}b^* _\textbf{\textit{q}} +\frac{\Delta}2 \Big (a_\textbf{\textit{q}}b^*_\textbf{\textit{q}}+a^*_\textbf{\textit{q}}b_\textbf{\textit{q}} \Big ) \Big]\,, ~~\label{HamC}\\
\mathcal{H}_4&=&\frac{1}{4} \sum_{\textbf{\textit{q}}_1+ \textbf{\textit{q}}_2=\textbf{\textit{q}}_3+\textbf{\textit{q}}_4}T_{\textbf{1} \textbf{2},\textbf{3} \textbf{4}}\,a_\textbf{1}^* a^*_\textbf{2}
a_\textbf{3}a_\textbf{4}\,, \quad a_{\B j}\equiv a_{\B q_j} .~~~ \label{HamD}
\end{eqnarray}
\end{subequations}
The first two terms in $\mathcal{H}_2$ describe the free propagation of magnons and phonons with the dispersion laws $\omega^\textrm{m}_\textbf{\textit{q}}$ and $\omega^\textrm{p}_\textbf{\textit{q}}$,respectively. The last two terms in $\mathcal{H}_2$ are responsible for their linear coupling due to the magneto-elastic effect with a coupling amplitude $\Delta$. In the acoustic system nonlinearity can be safely neglected in comparison with the strongly nonlinear spin-wave (magnon) subsystem.
Without interaction, the dispersion laws $\omega^\textrm{m}_\textbf{\textit{q}}$ and $\omega^\textrm{p}_\textbf{\textit{q}}$ will cross at some hybridization wavevector $\textbf{\textit{q}}=\textbf{\textit{q}}_\times: \omega_{\textbf{\textit{q}}_\times}^{\textrm m}=\omega_{\textbf{\textit{q}}_\times}^{\textrm p}=\omega_\times$, as seen in Fig.\,\ref{dispersion}. Here, we consider a weak coupling regime with a narrow hybridization region $\delta q$ around $\textbf{\textit{q}}_\times$, such that
\begin{equation}
\Delta\equiv \delta q\frac{\partial (\omega_{\textbf{\textit{q}}_ \times}^{\textrm p}-\omega_{\textbf{\textit{q}} _\times}^{\textrm m})}{\partial q}\ll \omega_\times\ .
\end{equation}
Being interested in the system evolution in the hybridization region $\delta q$, we can restrict ourselves to the four-magnon interaction Hamiltonian $\mathcal{H}_4$, Eq.\,(\ref{HamD}) \cite{NSW} with the sum restricted by the hyper-surface $\textbf{\textit{q}}_1+ \textbf{\textit{q}}_2=\textbf{\textit{q}}_3+\textbf{\textit{q}}_4$. The interaction amplitude $T_{\textbf{1} \textbf{2},\textbf{3} \textbf{4}}$ depends on $\textbf{\textit{q}}_1$, $\textbf{\textit{q}}_2$, $\textbf{\textit{q}}_3$, $\textbf{\textit{q}}_4$.
The quadratic Hamiltonian $\mathcal{H}_2$ can be diagonalized by the linear canonical Bogoliubov $(u,v)$-transformation of the form:
%
\begin{subequations} \begin{equation}\label{u-v}
\begin{array}{ccc}
a_\textbf{\textit{q}}&=& c^{\mbox{\tiny L}}_\textbf{\textit{q}} \cos \varphi_\textbf{\textit{q}} + c^{\mbox{\tiny U}}_\textbf{\textit{q}}\sin \varphi_\textbf{\textit{q}} \,, \\
%
b_\textbf{\textit{q}}&=& - c^{\mbox{\tiny L}}_\textbf{\textit{q}} \sin \varphi_\textbf{\textit{q}} + c^{\mbox{\tiny U}}_\textbf{\textit{q}} \cos \varphi_\textbf{\textit{q}}\,,
\end{array}
\end{equation}
%
in which the coordinate system rotation angle $\varphi_\textbf{\textit{q}}$ is chosen as follows:
\begin{eqnarray} \label{3b}
~\hskip - .7 cm\cos^2 \varphi_{\B q}= \frac 12 \Big[ 1+ \frac {O_{\B q}}{ \big(1+O_{\B q}^2\big)^{1/2}} \Big] \,, \ O_q=\frac{\omega_{\B q}\sp p-\omega _{\B q}\sp m}{\Delta}\ .
\end{eqnarray} \end{subequations}
Here, $O_q$ is the dimensionless frequency distance from the crossover. This transformation leads to a new diagonal Hamiltonian $ \widetilde{\mathcal{H}}_2$ in terms of new normal canonical amplitudes of the upper (U) and lower (L) magneto-elastic modes (MEMs) $c\sp{\tiny U}_{\B q}$ end $c\sp{\tiny L}_{\B q}$ with frequencies $\Omega^{^{\rm U}}_\textbf{\textit{q}}$ and $\Omega^{^{\rm L}}_\textbf{\textit{q}}$:
\begin{subequations}\label{newH2}
\begin{eqnarray}\label{newH2A}
\widetilde{\mathcal{H}}_2&=& \sum_\textbf{\textit{q}}\Big [\Omega^{^{\rm U}}_\textbf{\textit{q}}c^{\mbox{\tiny U}}_\textbf{\textit{q}}c^{\mbox{\tiny U}*}_\textbf{\textit{q}}+ \Omega^{^{\rm L}}_\textbf{\textit{q}}c_\textbf{\textit{q}}^{\mbox{\tiny L}}c^{\mbox{\tiny L}*}_\textbf{\textit{q}}\Big ]\,,\\ \label{Omegas}
\Omega^{^{\rm U}}_\textbf{\textit{q}}&=&\frac{1}{2}\Big \{\omega^\textrm{m}_\textbf{\textit{q}} + \omega^\textrm{p}_\textbf{\textit{q}} + \sqrt{\big [\omega^\textrm{m}_\textbf{\textit{q}} - \omega^\textrm{p}_\textbf{\textit{q}} \big ]^2 + \Delta^2} \, \Big \}\ , \\
\Omega^{^{\rm L}}_\textbf{\textit{q}}&=&\frac{1}{2}\Big \{\omega^\textrm{m}_\textbf{\textit{q}} + \omega^\textrm{p}_\textbf{\textit{q}} - \sqrt{\big [\omega^\textrm{m}_\textbf{\textit{q}} - \omega^\textrm{p}_\textbf{\textit{q}} \big ]^2 + \Delta^2} \, \Big \}\ .
\end{eqnarray}
As usual, the group velocities of these modes are given by \looseness=-1
\begin{equation}\label{group}
\B v_{\B q}^{^{\rm U, L}} = \frac {d \, \Omega^{^{\rm U, L}} _{\B q} } {d {\B q}}\ .
\end{equation}
\end{subequations}
The interaction Hamiltonian\,(\ref{HamD}) in these new variables includes terms ${\mathcal{H}}^{^{\rm UU}}_4, {\mathcal{H}}^{^{\rm LL}}_4$, responsible for the interactions within the U- and L-MEMs, respectively, and a term ${\mathcal{H}}^{^{\rm LU}}_4$, that describs the interaction between them:
\begin{subequations}\label{newHam}
\begin{align}\begin{split}
{\mathcal{H}}^{^{\rm UU}}_4
&= \frac{1}{4}\sum_{\textbf{\textit{q}}_1+ \textbf{\textit{q}}_2=\textbf{\textit{q}}_3+\textbf{\textit{q}}_4} \!\!\!\!
T^{^{\rm UU}}_{\textbf{1} \textbf{2},\textbf{3} \textbf{4}}
\,c_{\textbf{1}}^{{\mbox{\tiny U}}*} c^{\mbox{\tiny U}*}_{\textbf{2}}c^{\mbox{\tiny U}}_{\textbf{3}}c^{\mbox{\tiny U}}_{\textbf{4}}\,,
\\ T^{^{\rm UU}}_{\textbf{1} \textbf{2},\textbf{3} \textbf{4}} &= T^{^{\rm UU}}_{\textbf{1} \textbf{2},\textbf{3} \textbf{4}} \sin \varphi_{\textbf{1}}\sin \varphi_{\textbf{2}}\sin \varphi_{\textbf{3}}\sin \varphi_{\textbf{4}}\,, \label{newHamA}
\end{split}\\
\begin{split}
\mathcal{H}_4^{^{\rm LL}}&= \frac{1}{4} \sum_{\textbf{\textit{q}}_1+ \textbf{\textit{q}}_2=\textbf{\textit{q}}_3+\textbf{\textit{q}}_4} \!\!\!\!
T^{^{\rm LL}}_{\textbf{1}\textbf{2},\textbf{3} \textbf{4}}
\,c_{\textbf{1}}^{{\mbox{\tiny L}}*} c^{\mbox{\tiny L}*}_{\textbf{2}}
c^{\mbox{\tiny L}}_{\textbf{3}}c^{\mbox{\tiny L}}_{\textbf{4}}\,,\\
T^{^{\rm LL}}_{\textbf{1} \textbf{2},\textbf{3} \textbf{4}} &= T^{^{\rm LL}}_{\textbf{1} \textbf{2},\textbf{3} \textbf{4}} \cos \varphi_{\textbf{1}}\cos \varphi_{\textbf{2}}\cos \varphi_{\textbf{3}}\cos \varphi_{\textbf{4}}\,, \label{newHamB} \end{split}\\
\begin{split}
\mathcal{H}_4^{^{\rm LU}}&= \frac 14\sum_{\textbf{\textit{q}}_1+ \textbf{\textit{q}}_2=\textbf{\textit{q}}_3+\textbf{\textit{q}}_4} \!\!\!\!
T^{^{\rm LU}}_{\textbf{1} \textbf{2},\textbf{3} \textbf{4}}\Big [
\,c_{\textbf{1}}^{\mbox{\tiny U}*} c^{\mbox{\tiny U}*}_{\textbf{2}}c^{\mbox{\tiny L}}_{\textbf{3}}c^{\mbox{\tiny L}}_{\textbf{4}}+ \mbox{c.c.}\Big ]\,,
%
\\ T^{^{\rm LU}}_{\textbf{1} \textbf{2},\textbf{3} \textbf{4}} &= T^{^{\rm LU}}_{\textbf{1} \textbf{2},\textbf{3} \textbf{4}} \sin \varphi_{\textbf{1}}\sin \varphi_{\textbf{2}}\cos \varphi_{\textbf{3}}\cos \varphi_{\textbf{4}}\ . \label{newHamC}
\end{split} \end{align}
\end{subequations}
%
Here, ``c.c.'' stands for complex conjugation. The energy and particle fluxes within the U- and L-MEM branches, are proportional to $|T^{^{\rm UU}}_{\textbf{1} \textbf{2},\textbf{3} \textbf{4}}|^2$ and $|T^{^{\rm LL}}_{\textbf{1} \textbf{2},\textbf{3} \textbf{4}}|^2$, the energy and particle exchange between modes are proportional to $|T^{^{\rm LU}}_{\textbf{1} \textbf{2},\textbf{3} \textbf{4}}|^2$.
To illustrate how these objects depend on the frequency near the hybridization crossover, where $\omega^\textrm{m}_\textbf{\textit{q}}=\omega^\textrm{p}_\textbf{\textit{q}}$, we plot them in Fig.\,\ref{model}(b) with the shorthand notations:
%
\begin{equation}\label{TT}
T^{^{\rm UU}}_q \equiv T^{^{\rm UU}}_{\textbf{\textit{q}},\textbf{\textit{q}};\textbf{\textit{q}},\textbf{\textit{q}} }\,, \ T^{^{\rm LL}}_q \equiv T^{^{\rm LL}}_{\textbf{\textit{q}},\textbf{\textit{q}};\textbf{\textit{q}},\textbf{\textit{q}} }\,, \ T^{^{\rm LU}}_q \equiv T^{^{\rm LU}}_{\textbf{\textit{q}},\textbf{\textit{q}};\textbf{\textit{q}},\textbf{\textit{q}} }\,,
\end{equation}
as functions of the dimensionless distance to the crossover $O_q$ defined by Eq.\,\eqref{3b}.
\vspace{-3mm}
\subsection{ \label{ss:stat} Statistical description of magneto-elastic modes}
\vspace{-3mm}
A statistical description of weakly interacting waves can be obtained \cite{ZLF} in terms of a kinetic equation, shown below for the continuous limit, when the system size $L$ is much larger than the wavelength $2\pi/k$:
%
\begin{equation}\label{KE}
\frac{\partial n^{\mbox{\tiny L }}(\textbf{\textit{q}}, t)}{\partial t} = \mbox{St}^{^{\rm L }}(\textbf{\textit{q}}, t)\,, \quad \frac{\partial n^{\mbox{\tiny U }}(\textbf{\textit{q}}, t)}{\partial t} = \mbox{St}^{^{\rm U }}( \textbf{\textit{q}}, t)\ .
\end{equation}
Here $n^{\mbox{\tiny L}}_\textbf{\textit{q}} =n^{\mbox{\tiny L}}(\textbf{\textit{q}}, t)$ and $n^{\mbox{\tiny U }}_\textbf{\textit{q}} =n^{\mbox{\tiny U}}(\textbf{\textit{q}}, t)$ are the same-time pair correlations of the L- and U-MEMs, defined by
$$
\langle c^{\mbox{\tiny L}}_\textbf{\textit{q}} c^{\mbox{\tiny L }}_{\textbf{\textit{q}}'}\rangle = \dfrac{4\pi^2}{L^2} \, \delta(\textbf{\textit{q}}-\textbf{\textit{q}}') n^{\mbox{\tiny L }}_\textbf{\textit{q}}\,, \
%
\langle c^{\mbox{\tiny L }}_\textbf{\textit{q}} c^{\mbox{\tiny L }}_{\textbf{\textit{q}}'}\rangle =\dfrac{4\pi^2}{L^2} \, \delta(\textbf{\textit{q}}-\textbf{\textit{q}}') n^{\mbox{\tiny L }}_\textbf{\textit{q}}\,,
$$
where $\langle \dots \rangle$ stands for the ensemble averaging. In the classical limit, when the occupation numbers of Bose particles $n_q (\textbf{\textit{q}}, t) \gg 1$, $n(\textbf{\textit{q}}, t)=\hbar n_q (\textbf{\textit{q}}, t)$.
In what follows we discuss only the relevant L-MEM population $n^{\mbox{\tiny L }}_{\textbf{\textit{q}}}$.
The collision integral $\mathrm{St}^{^{\rm L}}(\textbf{\textit{q}}, t)$ may be found in various ways~\cite{ZLF}, including the Golden Rule, widely used in quantum mechanics. Accounting for the 2L$\Rightarrow$2L and 2L$\Rightarrow$2U scattering events, we have
\begin{subequations}\label{St}
\begin{align}
\label{StA}
\mbox{St}^{^{\rm L}}(\textbf{\textit{q}} ,t)= & \mbox{St}^{^{\rm LL}}(\textbf{\textit{q}} ,t) + \mbox{St}^{^{\rm LU}}(\textbf{\textit{q}} ,t)\,, \\
%
\label{StB}
\begin{split}
\mbox{St}^{^{\rm LL}}(\textbf{\textit{q}} ,t)= & \frac{\pi}{4} \int d\textbf{\textit{q}}_1 d\textbf{\textit{q}}_2 d\textbf{\textit{q}}_3 \, \delta(\textbf{\textit{q}}+\textbf{\textit{q}}_1-\textbf{\textit{q}}_2-\textbf{\textit{q}}_3)\\
&\times \delta (\Omega_\textbf{\textit{q}}^{^{\rm L}} +\Omega_1^{^{\rm L}}-\Omega_2^{^{\rm L}} -\Omega_3^{^{\rm L}}) \,
|T^{^{\rm LL}}_{\textbf{\textit{q}} 1\, \textbf{2} \textbf{3}}|^2 \, \\
& \times [n_\textbf{2}^{\mbox{\tiny L}} n_\textbf{3}^{\mbox{\tiny L}} (n_\textbf{\textit{q}} ^{\mbox{\tiny L}} + n_\textbf{1}^{\mbox{\tiny L}} )- n_\textbf{\textit{q}} ^{\mbox{\tiny L}} n_\textbf{1}^{\mbox{\tiny L}} ( n_\textbf{2}^{\mbox{\tiny L}} + n_\textbf{3}^{\mbox{\tiny L}} ) ]\,,
\end{split}\\
%
\begin{split}
\label{StC}
\mbox{St}^{^{\rm LU}}(\textbf{\textit{q}} ,t)=& \frac{\pi}{4} \int d\textbf{\textit{q}}_1 d\textbf{\textit{q}}_2 d\textbf{\textit{q}}_3 \, \delta({\textbf{\textit{q}}+\textbf{\textit{q}}_1-\textbf{\textit{q}}_2-\textbf{\textit{q}}_3})\\
&\times \delta (\Omega_\textbf{\textit{q}}^{^{\rm U}} +\Omega_1^{^{\rm U}}-\Omega_2^{^{\rm L}} -\Omega_3^{^{\rm L}}) \,
|T^{^{\rm LU}}_{{\textbf{\textit{q}} 1\, \textbf{2} \textbf{3}}}|^2 \, \\%\label{StD}
&\times [n_\textbf{2}^{\mbox{\tiny U}} n_\textbf{3}^{\mbox{\tiny U}} (n_\textbf{\textit{q}}^{\mbox{\tiny L}} + n_\textbf{1}^{\mbox{\tiny L}} )- n_\textbf{\textit{q}}^{\mbox{\tiny L}} n_{\bm 1}^{\mbox{\tiny L}} ( n_\textbf{2}^{\mbox{\tiny U}} + n_\textbf{3}^{\mbox{\tiny U}} ) ]\, .
\end{split}
\end{align}
\end{subequations}
\section{\label{s:model} Double accumulation of hybrid bosons}
A full consistent theoretical description of the bottleneck accumulation process of hybrid magneto-elastic bosons, based on the kinetic equation\,(\ref{St}) for the actual YIG frequency spectra and the interaction amplitudes, is beyond the scope of this article.
In \Sec{ss:dif}, for a quantitative understanding of this problem, we restrict ourselves to the simplifying assumption of axial symmetry of the problem in the vicinities of the frequency minima at $\B q=\pm \B q_0$. Then, in \Sec{s:2D}, we turn to a two-dimensional description of the problem, considering caustics in the quasiparticle propagation which results in sharp anisotropic beams.
\subsection{\label{ss:dif} One-dimensional differential model}
Denoting $\B \kappa_\pm=\B q\mp \B q_0$ we introduce a one-dimensional approximation of the L- and U-MEM densities
\begin{subequations}
\begin{equation}\label{not}
N^{^{ \rm L}}_{\kappa_\pm}= 4\pi \kappa_\pm^2 n^{\mbox{\tiny L}}_{\kappa_\pm}\,, \ N\Sp{U}_{\kappa_\pm}= 4\pi \kappa_\pm^2 n^{\mbox{\tiny U}}_{\kappa_\pm}\ .
\end{equation}
To simplify further the appearance of the equations, we will ignore the difference between the $\pm \B q_0$ frequency minima and omit the symbol $``\pm"$. This simplification does not affect the qualitative picture of the described phenomena.
Now we can present the kinetic equation\,(\ref{KE}) as follows
\begin{align} \label{KE-L}
\frac{\partial N^{^{\rm L}}_\kappa}{ \partial t} & = 4\pi \kappa^2 \, [\mbox{St}^{^{\rm LL}}( \kappa , t)+ \mbox{St}^{^{\rm LU}}( \kappa, t)] \,, \\ \label{KE-U}
\frac{\partial N^{^{\rm U}}_ \kappa}{ \partial t} & = 4\pi \kappa^2 \, [\mbox{St}^{^{\rm UU}}( \kappa, t)- \mbox{St}^{^{\rm LU}}( \kappa, t)] \ .
\end{align}
\end{subequations}
\subsection{\label{ss:scattering} Estimates of the flux and the L$\to$U transformation rate}
The collision terms $\mbox{St}^{^{\rm LL}}(\kappa ,t)$ and $\mbox{St}^{^{\rm UU}}(\kappa ,t)$ preserve the total number of quasiparticle in the L-MEM and U-MEM, respectively. Therefore, they may be represented in a divergent form, for instance,
\begin{subequations}\label{10}\begin{equation}\label{10A}
\mbox{St}^{^{\rm LL}}( \kappa ,t)= d \mu_\kappa ^{\mbox{\tiny L}}/ d \kappa\,,
\end{equation}
where $\mu_\kappa\Sp L$ is the L-MEM particle flux towards small wavenumbers. Together with Eq.\,(\ref{10A}), this gives
\begin{equation} \label{10B}
\mu_\kappa\Sp L= 4\pi\int^\kappa \kappa^2 \mbox{St}^{^{\rm LL}}(\kappa') d\kappa'\ .
\end{equation}
Now, using Eq.\,(\ref{StB}), we obtain an estimate for the flux $\mu_q ^{\mbox{\tiny L}}$:
\begin{equation} \label{10C}
\mu_\kappa \Sp L \simeq \kappa^3_\times \big(T_\kappa \Sp{LL}\big)^2 \big(N_\kappa\Sp{L}\big)^3 \big / \delta\omega_\kappa\ .
\end{equation}
Here, $\delta \omega_\kappa \equiv \kappa^2 \big[ d^2 \omega_q/ 2 (d q)^2\big]_{q=q_0} $ and $T_\kappa\Sp{LL}= T_0 \cos^4 \varphi_\kappa$, with $T_0=T_{\B q_0\B q_0, \B q_0\B q_0}$ and $\cos \varphi_\kappa$ given by \Eq{3b} were $\B q=\B q_0+\B \kappa$.
A similar estimate of the collision integral $\mbox{St}^{^{\rm LU}}_\kappa = \mbox{St}^{^{\rm LU}}(\kappa) $, \Eq{StC}, takes the form
\begin{equation}\label{10D}
\mbox{St}^{^{\rm LU}}_\kappa \simeq \kappa ^2_\times \big(T^{^{\rm LU}}_\kappa\big)^2\, N_\kappa^{^{\rm L}} N_\kappa ^{^{\rm U}}\big ( N_\kappa^{^{\rm U}} - N_\kappa^{^{\rm L}}\big ) \big / \delta\omega_\kappa
\end{equation}\end{subequations}
with $T_\kappa\Sp{LU}= T_0 \cos^2 \varphi_\kappa\sin^2 \varphi_\kappa$. It can be shown that conservation laws of frequency and momentum fully forbid this scattering process when the frequency of the lower mode becomes smaller than the lowest possible frequency of the upper mode, i.e., the BEC frequency $\omega_0$. It means that the estimate\,\eqref{10D} is valid if $\Omega^{^{\rm L}}_{\kappa} > \min \Omega^{^{\rm U}}_{\kappa}=\omega_0$. Otherwise, $\mbox{St}^{^{\rm LU}}_\kappa =0$. Note that \Eqs{10C} and \eqref{10D} are more accurate than the preliminary estimates given in \Refn{arc-1}.
\begin{figure*}
\includegraphics[width=1\textwidth]{Model.pdf}
\caption{
\label{model}
Panel (a) shows the calculated magnon-phonon spectrum near the hybridization region. Panel (b) gives the schematic representations of the scattering and cross-scattering four-particle processes in the hybridization region. Panel (c) presents squares of interaction amplitudes $T^{^{\rm UU}}_q$, $\ T^{^{\rm LL}}_q$ and $4 T^{^{\rm LU}}_q$, normalized by $T_q$, vs the dimensionless distance from the hybridization crossover $O_q$. Panel (d): Solutions of balance equations in the cross-section area $-2<O_\kappa<2$, (solid red line) and below the BEC frequency $-2>O_\kappa$, solid blue line. Red and blue dashed lines: Particle profiles $N\Sp L_\kappa$ for constant particle fluxes $\mu\Sp L_\kappa=$const.
}
\end{figure*}
Now, in the stationary case the kinetic equation\,(\ref{KE-L}) for the lower mode can be written as follows:
\begin{equation}\label{flux}
\frac{d \mu\Sp L_\kappa}{d \kappa}+ \mbox{St}\Sp{ LU}_\kappa =0\ .
\end{equation}
\subsection{\label{ss:sol-1} Rate equation above the BEC frequency}
Using estimates\,\eqref{10C} and \eqref{10D} we can rewrite the rate \Eq{flux} as
\begin{align}
\begin{split}\label{bal-1}
& \frac{d}{d \kappa}\big [\kappa_\times^3\cos^8 \varphi_\kappa \big(\C N_\kappa\Sp L\big)^3\big]\\ =& 3\, a \big(\C N_\kappa\Sp L\big)^2 \kappa^2 \cos^4 \varphi_\kappa\sin^4 \varphi_\kappa\ .
\end{split}
\end{align}
Here, $\C N_\kappa\Sp L= N_\kappa\Sp L/N_+\Sp L$ is the dimensionless L-MEM density, normalized by the density $N_+\Sp L$ at a sufficiently large positive $O_\kappa $, say at $O_\kappa=3$. In \Eq{10D}, we assume for simplicity's sake that $N_\kappa\Sp L\gg N_\kappa\Sp U$ and prescribe $N_\kappa\Sp U$ as $\kappa^2 N_+\Sp L/\kappa_\times^2$ according to \Eq{not} with $ n_\kappa^{\mbox{\tiny U}}=$const. The dimensionless parameter $a\sim 1$ combines all uncontrolled parameters in our estimates.
The ordinary differential equation\,\ref{{bal-1}} can be solved with the boundary condition $\C N_\kappa\Sp L=1$ for $\kappa\to \infty$ giving the relative MEMs population in the hybridization region (above the BEC frequency and thus with a plus-symbol)
\begin{align}
\C N _\kappa \Sp {+,L}=\frac 1{\cos^{8/3}\varphi_\kappa }\Big [1- a \int \limits ^\infty _{\kappa/\kappa_\times }\frac {x^2 \sin^4 \varphi_x d x}{cos^{4/3}\varphi_x}\Big] \ .
\end{align}
In Fig.\,\ref{model}(d) we plot $\C N\Sp {+,L}_\kappa$ in red as a function of the dimensionless distance to the crossover $O_\kappa$ defined by \Eq{3b}. For concreteness we took the position of the frequency minimum $\kappa=0$ as $O_0=-2$. We see a sharp peak of $\C N\Sp {+,L}_O$ demonstrating the bottleneck accumulation of quasi-magnons in the hybridization region above the BEC frequency around $O\approx -1$. This peak is a result of the completion of two processes. The first one is the quasiparticle flux towards lower frequencies (negative $O$), which on its own leads to an infinite growth of $N\Sp {+,L}_O$, shown by the red dashed line in Fig.\,\ref{model}(d). This growth is caused by a decrease in the intrinsic LL nonlinearity [see plot of $T_\kappa\Sp{LL}$ in Fig.\,\ref{model}(d)], which has to be compensated by an increase in $N\Sp {+,L}_O$ to ensure a constant particle flux. This increase of $N\Sp {+,L}_O$ is limited by the second process: the intermodal L$\to$U particle flux, provided by the St$\Sp{LU}_\kappa$ collision integral\,\eqref{10D}. For frequencies of the lower mode $\Omega^{^{\rm L}}_\kappa <\omega_0$ (i.e., below the BEC frequency) St$\Sp{LU}_\kappa$ becomes zero (by the conservation laws of frequency and momentum) and growth of $N\Sp L_O$ reappears, see blue dashed line in Fig.\,\ref{model}(d). We consider this effect in \Sec{ss:sol-2}.
\subsection{\label{ss:sol-2} Rate equation below the BEC frequency}
Below the BEC frequency we have St$\Sp{LU}_\kappa=0$ (see \eq{flux}) and we have to account for another dissipation mechanism able to suppress the infinite growth of $\C N\Sp L_\kappa$. The simplest option is a small linear damping term $\gamma \C N\Sp L_\kappa$ originating from three-magnon scattering processes, magnon-phonon interaction or other processes. With this term instead of \Eq{flux} we have:
\begin{equation}\label{flux1}
\frac{d \mu^{\mbox{\tiny L}}_\kappa}{d \kappa}+ \gamma \C N\Sp{LU}_\kappa =0\ .
\end{equation}
Now, using an estimate\,\eqref{10C} we can rewrite rate \Eq{flux1} similarly to \Eq{bal-1}
\begin{align}
\begin{split}\label{bal-2}
& \frac{d}{d \kappa}\big [\kappa_\times \cos^8 \varphi_\kappa \big(\C N_\kappa\Sp L\big)^3\big] + \frac 32 b \, \C N_\kappa\Sp L=0 \ .
\end{split}
\end{align}
The dimensionless parameter $b\ll 1$ combines all the uncontrolled parameters in our estimates including a small damping term.
The solution of the ordinary differential equation\,(\ref{bal-2}) with proper boundary conditions at the BEC frequency (i.e. for $\kappa=0$) provides the relative population of the L-MEM branch for smaller frequencies (minus-symbol)
\begin{align}
\C N _\kappa \Sp{$-$,L}= \C N _0 \Sp{+,L}\frac {\cos^{8/3}\varphi_0 }{\cos^{8/3}\varphi_\kappa } \Big [1- b \int \limits ^0 _{\kappa/\kappa_\times }\frac { d x}{cos^{4/3}\varphi_x}\Big]^{1/2}\ .
\end{align}
%
In Fig.\,\ref{model}(d) we plot $\C N\Sp{$-$,L}$ by a solid blue line as a function of the dimensionless distance to the crossover $O_\kappa$ below the BEC frequency (in our case for $O<-2$). For concreteness we took $b=0.01$. We see a second sharp peak of $\C N\Sp L_O$ demonstrating the bottleneck accumulation of quasi-phonons much below the BEC frequency around $O\approx -4$.
\vspace{-3mm}
\section{\label{s:2D}Two-dimensional model: \\ caustics and anisotropic beams}
\vspace{-3mm}
\begin{figure*}
\centering
\includegraphics{Caustics.pdf}
\caption{
\label{caustics}
The graph in panel (a) shows the two-dimensional dispersion of magneto-elastic quasiparticles at the magnon-phonon crossing. The color scale symbolizes the kind of the particle. dark red stands for a pure phonon state, while blue stands for a pure magnon state. Two continuous orange lines mark the dispersion branches with the perpendicular wavevector $k_\mathrm{y} = 0$ rad/cm. The red, purple and blue lines are selected isofrequency curves. (b) This panel shows an intensity map of the lower part of the magnon-phonon dispersion shown in panel (a). The arrows indicate the directions of quasiparticle propagation in different sections of the isofrequency curves, and , by definition of the group velocity, they are always perpendicular to the isofrequency curves. The green circle symbolizes the $3\sigma$ resolution of the measurement setup in wavevector space. Panel (c) shows the quasiparticle population along the isofrequency curve in dependence of the propagation angle $\alpha$.
It becomes obvious that each of the shown isofrequency curves has preferred angles and velocities of the quasiparticles populating it. These angles increase with growing frequency. The propagation angle of the blue curve is nearly 0$^{\circ}$, while the red curve has two angles of $\pm 59^{\circ}$.}
\end{figure*}
Above, we formulated a one-dimensional model of the bottleneck accumulation of hybrid magneto-elastic quasiparticles during the thermalization of parametrically pumped magnons.
The model predicts the existence of two regions of such accumulation (slow quasi-magnons and fast quasi-phonons) with frequencies above and below the BEC frequency, $f\sb{slow}> f_0 > f\sb{fast}$, exactly as observed in the experiment.
However, we are unable to determine the spectral positions of these peaks quantitatively. The problem is that instead of integration over the wavevector space we used local estimations \eq{10C} and \eq{10D} of the collision integral~(\ref{StA}). This approximation works more or less reasonably well, in hydrodynamic turbulence theory with a large scale-invariant interval in wavevector space and smooth dependence of interaction amplitudes, as in Ref.\,\cite{Lvov-Pomyalov2018}.
However, this is not the case in our problem: the relevant interaction amplitudes demonstrate a sharp $q$-dependence in the vicinity of the crossover region. Thus, a quantitative description of the problem require an explicit solution of the rate equation with the actual form (\ref{StA}) of the collision integral and realistic boundary conditions.
Moreover, the suggested one-dimensional model leaves the question unanswered about the angular dependence of the propagation directions of the slow and fast quasiparticles, which should be the subject of a full three-dimensional theory. Such a theory must take into account the strongly anisotropic frequency spectrum of magnons and requires knowledge of the three-dimensional population of the L-MEM branch above the crossing frequency $f_\times$, and the U-MEM branch from $f_0$ to $f_\times$ and higher. Neither do we know these distributions experimentally nor theoretically. In this situation, we can only make the simplest assumptions about the mentioned distribution functions and compare the obtained results with the experiment.
Figure\,\ref{caustics}(a) shows two-dimensional frequency spectra of U- and L-MEMs in the vicinity of the magnon-phonon hybridization area. With the solid blue line we plotted here the isofrequency curve of fast quasi-phonons at $3720\,$MHz. Its shape looks like a segment of a perfect circle formed by the intersection of the isofrequency plane with the phonon dispersion cone, only slightly distorted by hybridization with the magnon spectrum. The red solid line shows the isofrequency curve of slow quasi-magnons at $3843\,$MHz. This curve deviates significantly from the circular shape due to the influence of strongly anisotropic magnon dispersion. To more clearly demonstrate the transition between these two curves, we also plotted two intermediate isofrequency curves at $3800\,$MHz and $3830\,$MHz as dashed lines. The same four curves are also represented in the $(q_x,q_y)$-plane in Fig.\,\ref{caustics}(b). They are complemented by arrows indicating the directions of group velocities. It is clear that the L-MEM quasiparticles belonging to these isofrequency curves have a different primary direction of propagation.
Obviously, the sensitivity of the BLS detection of quasiparticles $S_{\B q}$ is limited in the $(q_x,q_y)$-space. In our experiment, the center of the area of maximum sensitivity corresponds to the position of the small black circle in Fig.\,\ref{caustics}(b). When moving away from it, the sensitivity decreases approximately tantamount to a Gaussian curve with a standard deviation of $\sigma \sim 10^3\,$rad/cm. A qualitative estimate of the observational region in which we can register quasiparticles is shown as a circle-shaped green area.
We see that it includes some part of the blue line shown in Fig.\,\ref{caustics}(b). The group velocities of all corresponding quasi-phonons are almost directed along $q_x$. On the other hand, the quasi-magnons belonging to the red isofrequency curve in the green area of observation can be divided into two groups with positive and negative $q_y$, above and below the horizontal orange solid line. A substantial fraction of the upper part of the isofrequency curve around the inflection point has almost equally directed group velocities pointing upward at an angle of about $60^\circ$, causing the formation of a caustic in the propagating quasiparticle beam. Similarly, the lower part of the isofrequency curve with $q_y<0$ is responsible for the formation of the second caustic beam propagating at an angle $\alpha=-60^\circ$ with respect to the horizontal line $q_y=0$.
To qualitatively characterize the distribution of group velocities $\B v\sb{gr}\sp{\tiny L} (\B q)$ within each L-MEM quasiparticle group registered in our experiments, we must find the product $ S_{\B q} n_{\B q}\sp{\tiny L} \B v\sb{gr}\sp{\tiny L} S (\B q)$ along the corresponding part of the dispersion curve. Unfortunately, we do not know the L-MEM population $n_{\B q}\sp{\tiny L}$ and can only assume for simplicity's sake that $n_{\B q}\sp{\tiny L}=\mathrm{const}$ in the actual area. The resulting distributions of quasiparticles vs. the group velocity angle are shown in Fig.\,\ref{caustics}(c), where all plots are normalized to their maximum value for better comparison. We see three sharp peaks: a peak of fast quasi-phonons with $\alpha=0^\circ$ and two peaks of slow quasi-magnons with $\alpha\approx \pm 59^\circ$. This perfectly resembles the experimentally measured direction of propagation of slow and fast quasiparticles. By extracting the velocities for the points on the isofrequency curves matching to propagation angles $\alpha=0^\circ$ and $\pm 59^{\circ}$, it becomes possible to calculate the average propagation velocity values $v_\mathrm{gr}^\mathrm{theor}$ for both propagation angles. The results are $v_\mathrm{gr}^\mathrm{theor}(0^{\circ}, 3720\,\mathrm{MHz})$=3215\,m/s and $v_\mathrm{gr}^\mathrm{theor}(59^{\circ}, 3843\,\mathrm{MHz})$=360\,m/s, which are in good agreement with our experimental data (see Table\,\ref{t:1}).
\begin{table}[b]
\begin{tabular}{c||c|c|c|c}
& Fast & Slow & BEC & Crossing \\
& quasi-phonons & quasi-magnons & minima & point \\ \hline\hline
$ f $ (MHz) & 3720 & 3843 & 3819 & 3833 \\
$q_x \big( \frac{\mbox{rad}}{\mbox{cm}} \big)\textcolor{white}{\Big |}$ & 61238 & 67670 & 43360 & 62620 \\
$q_y \big( \frac{\mbox{rad}}{\mbox{cm}} \big)\textcolor{white}{\Big |}$ & 0 & $\pm 4600$ & 0 & 0 \\
$v_\mathrm{gr}^\mathrm{theor} \big( \frac{\mbox{m}}{\mbox{s}} \big)\textcolor{white}{\Big |}$ & 3215 & 360 & 0 & - \\
$v_\mathrm{gr}^\mathrm{exp} \big( \frac{\mbox{m}}{\mbox{s}} \big)\textcolor{white}{\Big |}$ & 3070 & 368 & 0 & - \\
\end{tabular}
\caption{\label{t:1}
Characteristic values of the problem determined from the two-dimensional model.
}
\end{table}
One should distinguish a somewhat different nature of the directed quasiparticle fluxes for angles of $59^\circ$ and $0^\circ$. In the first case, the isofrequency curve has two inflections at the points ($q_x \approx 67670 \, \mathrm{rad/cm}, q_y \approx \pm 4600 \, \mathrm{rad/cm}$); in their vicinity, canonical caustic patterns are formed \citep{Veerakumar2006}. Such caustics are protected from diffraction broadening and have a stable transverse aperture, which can be of subwavelength size \citep{Serga2010}. In the second case, the isofrequency curve has no inflection points and, thus, a conventional weakly divergent beam is formed by quasiparticles with approximately co-directed group velocities \cite{Heussner2020}.
It should be noted that any isofrequency curve lying between the blue and red curves, being populated, should form either focused or caustic beams at some propagation angles in the range from 0 to $\pm 59^{\circ}$. In our experiment, however, we do not observe quasiparticle propagation between these two boundary cases. This fact convincingly confirms the model of double accumulation of magneto-elastic bosons in the magnon-phonon hybridization region.
\vspace{-3mm}
\section{\label{s:discussion} Discussion and summary}
\vspace{-3mm}
We experimentally showed the appearance of two groups of hybrid magneto-elastic modes at the bottom of the spectrum of a parametrically overpopulated magnon gas in an in-plane magnetized magnetic film. These two groups form spatially separated beams with different group velocities. The first---``slow''---group propagates with velocity $ v\sb{slow}\approx 368\,$m/s under the angles $\alpha\sb{slow} \approx \pm 59^\circ$ with respect to the external magnetic field $\B H\| \hat {\B x}$.
The second---``fast''---group appears later and propagates with velocity $ v\sb{fast}\approx 3070\,$m/s along $\hat {\B x}$: $ \B v\sb{fast}\| \B H\| \hat {\B x}$.
We formulated a simple one-dimensional model of the bottleneck accumulation of hybrid MEMs during the process of parametrically pumped magnons evolution toward their Bose-Einstein condensation. The model predicts two accumulation areas (slow quasi-magnons and fast quasi-phonons) with the frequencies above and below the BEC frequency, $f\sb{slow}> f_0> f\sb{fast}$, exactly as observed in experiments. We consider the qualitative agreement of the experimental results with the simple analytical model as a strong evidence that our one-dimensional model grasps adequately the basic physics of the bottleneck accumulation phenomenon in the frequency domain.
To explain the reason for the observed slow and fast quasiparticle propagation in narrow angular intervals around $\alpha\sb{slow}\sp{exp}\approx \pm 59^\circ$ and $\alpha\sb{fast}\sp{exp}\approx 0^\circ$ we considered in \Sec{s:2D} two-dimensional frequency spectra and found well-defined regions with almost the same group velocities--caustics--around the inflection points in the BLS registered part of the $\B q$-space. Using only this knowledge, we theoretically found propagation angles $\alpha\sb{slow}\sp{theor}= \pm 59^\circ$ and $\alpha\sb{fast}\sp{theor}= 0^\circ$ [see \Fig{caustics}(c)], which perfectly agree with corresponding experimental values $\alpha\sb{slow}\sp{exp}\approx \pm 59^\circ$ and $\alpha\sb{fast}\sp{exp}\approx 0^\circ$.
We consider the discovered double accumulation of magnetoelastic modes with nonzero group velocities as a promising effect for applications in future magnon spintronic devices. To determine the specific field of such applications, one should clarify the degree of real spectral localization of each of these quasiparticle groups and, consequently, the degree of their coherence. We believe that the answer to this question can be obtained by interference experiments with accumulated quasiparticles.
\vspace{-3mm}
\section*{Acknowledgments}
\vspace{-3mm}
Financial support by the European Research Council within the Advanced Grant 694709 SuperMagnonics -- ``Supercurrents of Magnon Condensates for Advanced Magnonics'' as well as financial support by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) through the Collaborative Research Center ``Spin+X: Spin in its collective environment'' TRR -- 173 -- 268565370 (project B04) is gratefully acknowledged. We thank G. A. Melkov for fruitful discussion.
|
\section{Introduction}
Reading is one of the important life skills that a learner has to continuously practice and improve beginning grade school. When a prescribed reading material exceeds the reading capability of a learner, they may become frustrated, disinterested, and lose confidence in their reading abilities \cite{Cambria2010,Hasyim2018}. Readability assessment is the process used by linguists and educators to evaluate the ease or difficulty of texts with the intent of prescribing reading materials to learners that are appropriate to their abilities.
Automated readability assessment can be viewed as a supervised task where annotated data is required and thus can be approached generally in two ways: regression and classification. In regression-based approaches \cite{flor-etal-2013-lexical,Guevarra2011}, the readability level of books is provided as a value from a fixed range. This provides more fine-grained information about the magnitude of readability of texts. However, recent works have favored classification-based models \cite{chatzipanagiotidis-etal-2021-broad,weiss-meurers-2018-modeling,xia-etal-2016-text,reynolds-2016-insights,hancke-etal-2012-readability,vajjala-meurers-2012-improving} for their simplicity and ease in collecting gold-standard data as experts usually give annotations by category of difficulty \cite{deutsch-etal-2020-linguistic}.
\begin{comment}
Research works in classification-based approaches \cite{dellorletta-etal-2012-genre,ma-etal-2012-ranking,dellorletta-etal-2011-read,tanaka-ishii-etal-2010-sorting} have explored on extending readability assessment as a ranking or sorting task based on difficulty. For ranking, a model or comparator is trained using traditional machine learning algorithms such as SVM. The model is used to provide a comparable value, usually the probability for a certain class or an integer, for each text in the data and then sorted accordingly. This approach gives the relative ordering of texts and the output can be transformed to a discrete level of difficulty by looking at the readability levels of neighboring texts. In addition, this approach also allows systems to be more agnostic to various types of difficulty level categories since the majority of sets of texts will be ranked the same way, even if the differences in levels are not exactly equal.
We highlight one important work by \citet{ma-etal-2012-ranking} which frames the conventional readability assessment problem into a ranking task by using a RankSVM model \cite{joachims2002optimizing}. From the relative order output of the model,
used an approach of transforming the output of in terms of relative order of difficulty into a discrete level of difficulty using children's literature data in English.
The discrete output is obtained through a hard voting scheme by looking at the three nearest reading levels of neighboring leveled books and using the difficulty level associated with majority of the neighbors. However, we argue that resolving the readability level of an unknown text just by looking at its neighbors' discrete levels is too simplistic and can be further improved. To the best of our knowledge, no prior work has focused on transforming the relative order of difficulty to a discrete level by considering the actual contents or distribution of words in the text.
\end{comment}
In this paper, we propose the use of Word Mover's Distance (WMD) as an additional post-processing task for automatic readability assessment. WMD, a novel distance function developed by \citet{pmlr-v37-kusnerb15} using word embeddings, allows identification of dissimilarity between two texts by calculating the distance or the amount of effort to transport all words from one document to another within the word embedding space. Thus, we argue that the shorter the distance between two texts, the similar they are in terms of readability and vice versa. Likewise, the application of a post-processing technique acts as a \textit{second-tier reduction process} by minimizing the errors made of a trained classification model by considering the distribution of words among similar texts in the same category of reading level. We test the proposed methodology on top of a trained SVM model using three monolingual datasets in Filipino, English, and German. We also provide an empirical evaluation of the model's performance and limitations with other approaches in readability assessment.
\section{Preliminaries}
The Word Mover's Distance (WMD) is a word embedding-based distance function by \citet{pmlr-v37-kusnerb15} inspired by the Earth Mover's Distance \cite{rubner1998metric} transportation problem. WMD calculates the dissimilarity between texts in terms of Euclidean distance in the word embedding space. The \textit{travel cost} of texts (represented as bag-of-words vectors) is calculated as the minimized cumulative aggregation cost of \textit{moving} words from a source text $D_i$ to a destination text $D_j$. The model can be formally viewed as follows,
\begin{equation}
\min_{\mathbf{T}\geq 0} \sum_{i,j=1}^{n} = \mathbf{T}_{i,j} c(i,j)
\end{equation}
where $n$ is the total vocabulary size, $c(i,j) = \left \| x_i - x_j \right \|$ is the distance between two words in the semantic space, and $\mathbf{T}\geq 0$ denotes the converted sparse flow matrix of how much words from the source text $D_i$ travels to the destination text $D_j$.
WMD is a semantic-based approach for estimating the (dis)similarity of documents. However, we hypothesize that WMD can also capture the readability of documents as long as the documents being compared do not have the same, exact meaning. We leverage our hypothesis on the common knowledge that documents with a high level of reading difficulty will often use complex words and documents with a low level reading difficulty will often use simple, easy words regardless of genre. For example, difficult or challenging words that may be encountered in college textbooks are considered high level and most likely will not be present in low level reading materials such as children's story and picture books. Thus, if documents $D_1$ and $D_2$, where $D_2$ is a document of higher reading level than $D_1$, then the WMD distance of a target document $D_0$ with unknown reading level is shorter with respect to $D_2$ if they belong to the same category of reading difficulty and share similar words.
To ground this hypothesis, we use the OneStopEnglish corpus \cite{vajjala-lucic-2018-onestopenglish} which is composed of open-source English reading materials\footnote{Additional description found in Section 4.}. The purpose of this test is to identify if the WMD scores of documents with distinct readability levels do not coincide with documents with distinct semantics and is statistically significant in terms of difference. From the corpus, we handpicked and aggregated two groups: (a) the first group being 189 pairs of documents with distinct reading levels and (b) the second group being 189 pairs of documents belonging to the same reading level. Documents from both groups are semantically different. We obtained the WMD scores of each pair for each group and performed a two-sample Mann-Whitney U test. Results showed that the difference between WMD scores between readability ($\mu$ = 0.089) and semantics ($\mu$ = 0.041) of documents is statistically significant ($p$ < 0.05)\footnote{test $Z$ statistic = 16.214 not in the 95\% critical value accepted range: [-1.960,1.960]}, proving our initial hypothesis.
\section{Post-Processing Technique with WMD}
\label{postProcessing}
The variation of this approach from previous works \cite{ma-etal-2012-ranking} is in the method of ranking texts and the procedure used for transforming relative order to a discrete value. Our proposed post-processing technique involves three major phases: classification, ranking, and grounding.
For the \textbf{classification phase}, a model is trained using vanilla SVM (based from the commonly-used models mentioned in previous works) using hand-crafted linguistic feature sets to produce a series of probabilities for each instance of the training data per corresponding class. The target text where the readability level is needed will also be subjected to the model. Each text in the training data as well as the target text will now have a collection of probabilities for each class.
In the \textbf{ranking phase}, texts from the training data plus the target text will then be ranked in an increasing order with respect to the probability of the most difficult class in terms of readability to produce a collection of sorted texts or a \textit{bookshelf} arranged from easiest to most difficult. Each instance of the training data in this phase acts as a \textit{guide book} wherein it will preemptively help identify the readability level of the target text since it is assumed that the bookshelf is arranged according to readability.
In the \textbf{grounding phase}, the location of the target text is identified from the sorted bookshelf of texts. The distribution of words of the target text will be compared with its first three neighboring texts from left and right with WMD. To get the final readability level, a hard-voting approach will be performed to get the majority readability level from the three compared neighbors bidirectionally. In case of a tie, the neighbor with the least, normalized WMD difference (since the algorithm is distance-based) to the target text will be selected for the corrected label. Figure~\ref{fig:tie} visually describes how the post-processing technique operates in the case of a three-way tie.
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.45\textwidth]{figures/tie2.png}
\caption{Label evaluation in the case of triple tie. The class with the least, normalized distance from the target text $x$ will be selected, in this case it is 3.}
\label{fig:tie}
\end{figure}
\section{Data Description}
We test the proposed post-processing method on three datasets composed of leveled texts in English, German, and Filipino as described in Table~\ref{tab:data}. For implementation procedures, we obtained the code from the authors for the extraction of the linguistic feature sets spanning traditional, lexical, syntactic, psycholinguistic, and morphological used to train readability assessment models. We extracted 165, 155, and 54 features as done in previous works for German \cite{hancke-etal-2012-readability}, English \cite{vajjala-lucic-2018-onestopenglish}, and Filipino \cite{Imperial2019,imperial2020exploring,imperial2021application} respectively. \linebreak
\noindent\textbf{GEO--GEOlino corpus.} The GEO--GEOlino corpus is an open-source dataset containing 4,599 magazine articles in German from the work of \citet{hancke-etal-2012-readability} and \citet{weiss-meurers-2018-modeling}. These are frequently used as baseline for German-related readability assessment. The articles in the GEO corpus cover the domains of nature, culture, and science published by Gruner and Jahr\footnote{https://www.geo.de} while GEOlino contains topics on handicraft, games, and animals targeted for children age 8-14 years old. \linebreak
\noindent\textbf{OneStopEnglish (OSE) corpus.} The OSE corpus is a collection of English texts in three levels of difficulty (beginner, intermediate, and advanced) for adult ESL learners. The corpus was compiled by \citet{vajjala-lucic-2018-onestopenglish} from the OneStopEnglish learning resource website\footnote{https://www.onestopenglish.com}. A total of 567 texts were used for this study.
\linebreak
\noindent\textbf{Adarna House corpus.} The Adarna House data is a collection of Filipino story books published by Adarna House Inc.\footnote{https://adarna.com.ph/} that are commonly used in the general educational sector of the Philippines. A total of 265 reading materials distributed into three levels (beginner, intermediate, and advanced) were used for the study.
\linebreak
\begin{table}[!htbp]\small
\centering
\begin{tabular}{|l|c|c|c|}
\hline \bf Data & \bf Doc Count & \bf \# Class &\bf Vocab\\ \hline
GEO-GEOlino & 4,599 & 2 & 191,262\\ \hline
OneStopEnglish & 567 & 3 & 17,818\\ \hline
Adarna House & 265 & 3 & 16,058\\ \hline
\end{tabular}
\caption{\label{tab:data}
Data distribution for English, German, and Filipino.}
\end{table}
\section{Experiment Setup}
We trained an SVM model using the various feature sets extracted for each corresponding dataset used. We chose SVM due to its significant usage in previous works \cite{chatzipanagiotidis-etal-2021-broad,deutsch-etal-2020-linguistic,imperial2020exploring,weiss-meurers-2018-modeling,hancke-etal-2012-readability}. We used a 5-fold cross validation procedure for training the models. After training, we applied the proposed post-processing technique described in Section \ref{postProcessing} and evaluated the performance using accuracy and F1 with the test split. For the word embeddings of English, German, and Filipino needed for the technique, we downloaded the resources from the fastText website\footnote{https://fasttext.cc/docs/en/crawl-vectors.html}. The word embeddings in various languages were trained from Common Crawl and Wikipedia datasets by \citet{grave-etal-2019-training}.
\begin{table*}[!htbp]\small
\centering
\begin{tabular}{|l|c|c|c|c|c|c|}
\hline
\multicolumn{1}{|c|}{\textbf{Method}} &
\multicolumn{2}{c|}{\textbf{GEO-GEOlino}} &
\multicolumn{2}{c|}{\textbf{OneStopEnglish}} &
\multicolumn{2}{c|}{\textbf{Adarna House}} \\ \cline{2-7}
\multicolumn{1}{|c|}{} &
\multicolumn{1}{c|}{\textbf{Acc}} & \multicolumn{1}{c|}{\textbf{F1}} & \multicolumn{1}{c|}{\textbf{Acc}} & \multicolumn{1}{c|}{\textbf{F1}} & \multicolumn{1}{c|}{\textbf{Acc}} & \multicolumn{1}{c|}{\textbf{F1}} \\ \hline
Vanilla SVM & 0.813 & 0.799 & 0.711 & 0.694 & 0.413 & 0.416 \\ \hline
Binary Insertion Sort Ranking & {0.812} & {0.797} & {0.659} & {0.667} & {0.444} & {0.357} \\
\cite{tanaka-ishii-etal-2010-sorting} & & & & & & \\ \hline
RankSVM w/ 3-Neighbor Scheme & {0.867} & {0.861} & {0.658} & {0.644} & {0.365} & {0.359} \\
\cite{ma-etal-2012-ranking} & & & & & & \\ \hline
\bf SVM w/ Proposed WMD Technique & \bf 0.890 & \bf 0.892 & \bf 0.746 & \bf 0.749 & \bf 0.538 & \bf 0.538 \\ \hline
\end{tabular}
\caption{\label{experiments}
The accuracy and F1 scores of the proposed post-processing technique with WMD on top of a trained SVM model outperformed previous ranking-based approaches in three monolingual datasets.}
\end{table*}
\section{Results and Discussion}
We compared the performance of the proposed technique against three previous approaches: (a) a trained vanilla SVM model, (b) Binary Insertion Sort ranking by \citet{tanaka-ishii-etal-2010-sorting}, the first ranking method applied in readability assessment, and (c) a RankSVM model by \citet{ma-etal-2012-ranking} which also uses a bidirectional three-neighbor scheme. We especially highlight the work of \citet{ma-etal-2012-ranking} as it is the one closest to our approach for analysis; the difference in our approach is that we consider the actual text contents of the documents through WMD instead of just looking at the discrete labels.
Table~\ref{experiments} shows results of the proposed technique improving the performances of conventional classification-based model such as SVM in terms of accuracy and F1 scores in all three languages. The accuracy score provides an overview of each model's performance while the F1 values show the harmonic mean between the models' precision and recall. As can be seen from the table, the most notable result is the largest increase when the post-processing technique is applied to the Filipino dataset with $\approx$13\% in accuracy and F1. For English and German, though not as large as the performance increase in Filipino, the post-processing technique still managed to \textit{correct} a few misclassified instances with around 2.9\% and 7.7\% for accuracy and 5.5\% and 9.3\% for F1. We attribute the improvement in performance specifically to the ranking phase wherein each instance of the training data is used as a \textit{guide book} to place the target text in its correct position in the \textit{bookshelf} through sorting. From this, since the comparator for the ranking phase utilized the output of the trained SVM model, it is less likely to make mistakes because the SVM model itself was developed using the training data.
With the Filipino dataset having the largest increase in performance, this means that the post-processing technique made a substantial number of corrections of labels during the ranking and grounding phases of the technique and despite the subpar performance of the vanilla SVM model. To quantify this value, the occurrence of the post-processing technique invoking the WMD tie-breaker method of normalizing distances for each readability class are 13.3\%, 10\%, and 8.7\% for Filipino, German, and English respectively. These values observe a direct relationship with the improvements of upgrading the SVM model with the WMD-based post-processing technique.
\section{Related Work}
\label{RRL}
In automatic readability assessment, traditional machine learning models are often used rather than neural network-based models for languages except English due to limited resource of annotated data. Recent works considered modelling the hierarchical and sequential nature of documents \cite{meng2020readnet,deutsch-etal-2020-linguistic,martinc2019supervised,azpiazu-pera-2019-multiattentive}. In \citet{deutsch-etal-2020-linguistic}, the performance of models trained using SVM and Logistic Regression remains closely at par with CNN, hierarchical attention models, and Transformer-based models. Thus, using these traditional models are still practical for the field.
Our proposed post-processing technique was inspired by ranking-based approaches which are often done for low-resource languages. \citet{tanaka-ishii-etal-2010-sorting} made a workaround with the inadequate data problem in Japanese texts by using an SVM model as a comparator to determine the relative ease of texts rather than based on an indicator of the absolute difficulty. \citet{dellorletta-etal-2012-genre} used the combined cosine distance of a target text from two poles (feature vectors extracted from easy-to-read and difficult-to-read corpus) and then performed ranking based on the accumulated distance value. \citet{vajjala-meurers-2014-assessing} used a ranking-based approach on identifying the relative difficulty of sentence pairs obtained from Wikipedia and Simple Wikipedia websites.
\section{Conclusion}
Readability assessment is the process of automating how language experts evaluate the ease or difficulty of texts. In this study, we devised a post-processing algorithm using Word Mover's Distance which can be added to classification-based readability assessment models to improve their performances for the task by correcting labels through a majority-voting scheme. The approach is (a) language-agnostic as tested with the English, German, and Filipino datasets and (b) works in a multiclass setting. Results from experiments show that the addition of the post-processing technique significantly improves performances of SVM especially for low-resource languages such as Filipino where neural-based approaches are not feasible. Further directions of the study include using knowledge-enriched word embeddings as well as more explorations in other low-resource languages.
\section{Impact Statement}
We discuss the broad impact statement and ethical considerations of the study. First, the study does not introduce a new dataset nor does it involve any human subjects. It covers only a small, focused and interesting application of a new post-processing technique in improving readability assessment models of texts in various languages. In general, the intended use of readability assessment models and its related studies are to help linguists, educators, and publishers in doing away with manual and formula-based approaches that are often shallow and time-consuming. We emphasize that it is the language experts in the field who will have a final verdict on the proper readability level of texts especially in the education sector.
|
\section{Introduction}
Graphs appear in many applications
as abstraction of real-world phenomena,
where vertices represent certain objects
and edges represent their relations.
Rather than being stationary,
graph data obtained in applications
usually change with respect to
some parameter such as time.
A summary of these changes in a quantifiable
manner can help gain insight into the data.
Persistent homology~\cite{carlsson2010zigzag,edelsbrunner2000topological}
is a suitable tool for this goal because it quantifies
the life span of topological features
as the graph changes.
One drawback of using standard non-zigzag persistence~\cite{edelsbrunner2000topological}
is that it
only allows addition of vertices and edges during the change,
whereas deletion may also happen in practice.
For example,
many complex systems such as social networks, food webs,
or disease spreading
are modeled by
the so-called ``dynamic networks''~\cite{holme2012temporal,kim2017stable,skarding2020foundations},
where
vertices and edges can appear and disappear at different time.
A variant of the standard persistence called
{\it zigzag persistence}~\cite{carlsson2010zigzag} is thus a more natural tool in such scenarios because
simplices can be both added and deleted.
Given a sequence of graphs possibly with additions and deletions
(formally called a {\it zigzag filtration}),
zigzag persistence produces a set of intervals termed as {\it zigzag barcode}
in which each interval
registers the birth and death time
of a homological feature.
Figure~\ref{fig:dm_net} gives an example of a graph sequence in which
clusters may split (birth of 0-dimensional features)
or vanish/merge (death of 0-dimensional features).
Moreover, addition of edges within the clusters creates 1-dimensional cycles
and deletion of edges makes some cycles disappear.
These births and deaths
are captured by zigzag persistence.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig/dm_net}
\caption{A sequence of graphs with four prominent clusters
each colored differently. Black edges connect
different clusters and forward (resp. backward) arrows
indicate additions (resp. deletions) of vertices and edges.
From (a) to (b), two clusters
split; from (b) to (c), two clusters merge; from (c) to (d), one cluster
disappears.
}
\label{fig:dm_net}
\end{figure}
Algorithms for both zigzag and non-zigzag persistence
have a general-case time complexity of
$O(m^\omega)$~\cite{carlsson2009zigzag-realvalue,edelsbrunner2000topological,maria2014zigzag,milosavljevic2011zigzag},
where
$m$ is the length of the input filtration and $\omega<2.37286$ is the matrix multiplication exponent~\cite{alman2021refined}.
For the special
case of graph filtrations, it is well known that
non-zigzag persistence
can be computed in $O(m\,\alpha(m))$ time,
where $\alpha(m)$ is the inverse Ackermann's function
that is almost constant for all practical purposes~\cite{CLRS3rd}.
However, analogous faster algorithms for
zigzag persistence on graphs are not known.
In this paper,
we present algorithms for zigzag persistence on graphs with near-linear time complexity.
In particular, given a zigzag filtration of length $m$
for a graph with $n$ vertices and edges,
our algorithm for 0-dimension
runs in $O(m\log^2 n+m\log m)$ time,
and our algorithm for $1$-dimension
runs in $O(m\log^4 n)$ time.
Observe that the
algorithm for $0$-dimension works for arbitrary complexes
by restricting to the $1$-skeletons.
The difficulty in designing faster zigzag persistence algorithms for the special
case of graphs lies in the deletion of vertices and edges.
For example, besides merging into bigger ones,
connected components can also split into smaller ones
because of edge deletion.
Therefore,
one cannot simply kill the younger component during merging
as in standard persistence~\cite{edelsbrunner2000topological},
but rather has to pair the {\it merge}
and {\it departure} events with the {\it split} and {\it entrance}
events (see Sections~\ref{sec:0-zigzag} for details).
Similarly, in dimension one, deletion of edges may kill 1-cycles
so that one has to properly pair the creation and destruction
of 1-cycles, instead of simply treating all 1-dimensional intervals
as infinite ones.
Our solutions are as follows: in dimension zero,
we find that the $O(n\log n)$ algorithm by Agarwal et al.~\cite{agarwal2006extreme}
originally designed for pairing critical points of Morse functions on 2-manifolds
can be utilized
in our scenario. We formally prove the correctness of applying
the algorithm and use a {\it dynamic connectivity} data
structure~\cite{holm2001poly} to achieve the claimed complexity.
In dimension one, we observe that a positive and a negative edge
can be paired by finding the
{\it earliest} 1-cycle containing both edges
which resides in all intermediate graphs.
We further reduce the pairing to finding the {\it max edge-weight}
of a path in a minimum spanning forest.
Utilizing a data structure for {\it dynamic minimum spanning forest}~\cite{holm2001poly},
we achieve the claimed time complexity.
Section~\ref{sec:1-zigzag} details this algorithm.
Using Alexander duality,
we also extend the algorithm for $0$-dimension
to compute $(p-1)$-dimensional zigzag for $\mathbb{R}^p$-embedded
complexes.
The connection between these two cases
for non-zigzag persistence
is well known~\cite{edelsbrunner2012alexander,schweinhart2015statistical},
and the challenge comes in adopting
this duality to the zigzag setting while maintaining
an efficient time budget.
With the help of a {\it dual filtration}
and an observation about faster void boundary reconstruction
for {\it$(p-1)$-connected} complexes~\cite{dey2020computing},
we achieve a time complexity of
$O(m\log^2 n+m\log m+n\log n)$.
\paragraph*{Related works.}
The algorithm for computing persistent homology by
Edelsbrunner et al.~\cite{edelsbrunner2000topological}
is a cornerstone of topological data analysis. Several extensions
followed after this initial development.
De Silva et al.~\cite{de2011persistent} proposed to compute
persistent {\it cohomology} instead of homology
which gives the same barcode.
De Silva et al.~\cite{de2011dualities} then showed
that the persistent cohomology algorithm runs faster in practice
than the version that uses homology.
The {\it annotation} technique proposed by Dey et al.~\cite{dey2014computing}
implements the cohomology algorithm by maintaining a cohomology basis more
succinctly and extends to {\it towers} connected by simplicial maps.
These algorithms
run in $O(m^3)$
time.
Carlsson and de Silva~\cite{carlsson2010zigzag}
introduced zigzag persistence
as an extension of the standard persistence,
where they also presented a
decomposition algorithm for computing zigzag barcodes
on the level of vector spaces and linear maps.
This algorithm is then adapted to zigzag filtrations at simplicial level by
Carlsson et al.~\cite{carlsson2009zigzag-realvalue}
with a time complexity of $O(m^3)$.
Both algorithms~\cite{carlsson2010zigzag,carlsson2009zigzag-realvalue}
utilize a construct called {\it right filtration} and a {\it birth-time vector}.
Maria and Oudot~\cite{maria2014zigzag} proposed
an algorithm for zigzag persistence
based on some {\it diamond principles}
where an inverse non-zigzag filtration is always maintained
during the process.
The algorithm in~\cite{maria2014zigzag} is shown to run faster
in experiments than the algorithm in~\cite{carlsson2009zigzag-realvalue}
though the time complexities remain the same.
Milosavljevi\'{c} et al.~\cite{milosavljevic2011zigzag}
proposed an algorithm for zigzag persistence based on
matrix multiplication
which runs in $O(m^\omega)$ time, giving the best asymptotic bound
for computing zigzag and non-zigzag persistence
in general dimensions.
The algorithms reviewed so far
are all for general dimensions
and many of them
are based on matrix operations.
Thus, it is not surprising that
the best time bound achieved is $O(m^\omega)$
given that
computing Betti numbers
for a simplicial $2$-complex of size $m$
is as hard as computing the rank of
a $\mathbb{Z}_2$-matrix with $m$ non-zero entries
as shown by Edelsbrunner and Parsa~\cite{edelsbrunner2014computational}.
To lower the complexity,
one strategy
(which is adopted by this paper)
is to consider special cases
where matrix operations can be avoided.
The work by
Dey~\cite{dey2019computing} is probably most related to ours
in that regard,
who proposed an $O(m\log m)$ algorithm for non-zigzag persistence
induced from height functions on $\mathbb{R}^3$-embedded complexes.
\section{Preliminaries}
A {\it zigzag module} (or {\it module} for short)
is a sequence of vector spaces
\[\mathcal{M}: V_0 \leftrightarrowsp{\psi_0} V_1 \leftrightarrowsp{\psi_1}
\cdots \leftrightarrowsp{\psi_{m-1}} V_m\]
in which
each $\psi_i$ is either a forward linear map $\psi_i:V_i\toV_{i+1}$
or a backward linear map $\psi_i:V_i\leftarrowV_{i+1}$.
We assume
vector spaces are over field $\mathbb{Z}_2$ in this paper.
A module $\mathcal{S}$ of the form
\[\mathcal{S}:W_0 \leftrightarrowsp{\phi_0} W_1 \leftrightarrowsp{\phi_1}
\cdots \leftrightarrowsp{\phi_{m-1}} W_m\]
is called a {\it submodule} of $\mathcal{M}$ if each $W_i$ is a subspace of $V_i$ and
each $\phi_i$
is the restriction of $\psi_i$.
For an interval $[b,d]\subseteq[0,m]$,
$\mathcal{S}$ is called an {\it interval submodule} of $\mathcal{M}$ over $[b,d]$
if $W_i$
is one-dimensional for $i\in[b,d]$
and is trivial for $i\not\in[b,d]$,
and $\phi_i$ is an isomorphism for $i\in[b,d-1]$.
It is well known~\cite{carlsson2010zigzag} that $\mathcal{M}$ admits
an {\it interval decomposition}
$\mathcal{M}=\bigoplus_{\alpha\in\mathcal{A}}\mathcal{I}^{[b_\alpha,d_\alpha]}$
which is a direct sum of interval submodules of $\mathcal{M}$.
The (multi-)set of intervals
$\Set{[b_\alpha,d_\alpha]\,|\, \alpha\in\mathcal{A}}$
is called the {\it zigzag barcode} (or {\it barcode} for short) of $\mathcal{M}$
and is denoted as $\mathsf{Pers}(\mathcal{M})$.
Each interval in a zigzag barcode is called a {\it persistence interval}.
In this paper,
we mainly focus on a special type of zigzag modules:
\begin{definition}[Elementary zigzag module]
A zigzag module is called \textbf{elementary} if
it starts with the trivial vector space
and all linear maps in the module
are of the three forms: {\rm(}{\sf i}{\rm)} an isomorphism;
{\rm(}{\sf ii}{\rm)} an injection with rank 1 cokernel;
{\rm(}{\sf iii}{\rm)} a surjection with rank 1 kernel.
\end{definition}
A {\it zigzag filtration} (or {\it filtration} for short)
is a sequence of simplicial complexes
\[\mathcal{F}: K_0 \leftrightarrowsp{\fsimp{\mathcal{F}}{0}} K_1 \leftrightarrowsp{\fsimp{\mathcal{F}}{1}}
\cdots \leftrightarrowsp{\fsimp{\mathcal{F}}{m-1}} K_m\]
in which each
$K_i\leftrightarrowsp{\fsimp{\mathcal{F}}{i}} K_{i+1}$ is either a forward inclusion
$K_i\hookrightarrow K_{i+1}$ with a single simplex $\fsimp{\mathcal{F}}{i}$ added,
or a backward inclusion $K_i\hookleftarrow K_{i+1}$
with a single $\fsimp{\mathcal{F}}{i}$ deleted.
When the $\fsimp{\mathcal{F}}{i}$'s
are not explicitly used,
we drop them and simply denote $\mathcal{F}$ as
$\mathcal{F}: K_0 \leftrightarrow K_1 \leftrightarrow
\cdots \leftrightarrow K_m$.
For computational purposes,
we sometimes assume that a filtration starts with the empty complex,
i.e., $K_0=\emptyset$ in $\mathcal{F}$.
Throughout the paper, we also assume that each $K_i$ in $\mathcal{F}$ is a subcomplex
of a fixed complex $K$; such a $K$, when not given,
can be constructed by taking the union
of every $K_i$ in $\mathcal{F}$.
In this case, we call $\mathcal{F}$ a filtration {\it of} $K$.
Applying the $p$-th homology with $\mathbb{Z}_2$ coefficients on $\mathcal{F}$,
we derive the
{\it $p$-th zigzag module of $\mathcal{F}$}
\[\mathsf{H}_p(\mathcal{F}):
\mathsf{H}_p(K_0) \leftrightarrowsp{\morph{\mathcal{F}}{0}{p}}
\mathsf{H}_p(K_1) \leftrightarrowsp{\morph{\mathcal{F}}{1}{p}}
\cdots
\leftrightarrowsp{\morph{\mathcal{F}}{m-1}{p}} \mathsf{H}_p(K_m)\]
in which each $\morph{\mathcal{F}}{i}{p}$ is the linear map
induced by the inclusion.
In this paper,
whenever $\mathcal{F}$ is used to denote a filtration,
we use $\morph{\mathcal{F}}{i}{p}$ to denote a linear map
in the module $\mathsf{H}_p(\mathcal{F})$.
Note that $\mathsf{H}_p(\mathcal{F})$ is an elementary module
if $\mathcal{F}$ starts with an empty complex.
Specifically, we call $\mathsf{Pers}(\mathsf{H}_p(\mathcal{F}))$
the {\it $p$-th zigzag barcode} of $\mathcal{F}$.
\section{Zero-dimensional zigzag persistence}
\label{sec:0-zigzag}
We present our algorithm for 0-th zigzag
persistence\footnote{For brevity,
henceforth we call $p$-dimensional zigzag persistence
as {\it $p$-th} zigzag persistence.}
in this section.
The input is assumed to be on graphs
but note that our algorithm can be applied to any complex by restricting to its 1-skeleton.
We first define the barcode graph of
a zigzag filtration
which is a construct that our algorithm implicitly works on.
In a barcode graph,
nodes correspond to
connected components of graphs in the filtration
and edges encode the mapping between the components:
\begin{figure}
\begin{subfigure}[t]{0.9\textwidth}
\includegraphics[width=\textwidth]{fig/filt}
\vspace{-1em}
\caption{A zigzag filtration of graphs
with 0-th barcode $\Set{[2,2],[4,4],[6,8],[8,9],[7,10],[1,10]}$.}
\label{fig:filt}
\end{subfigure}
\vspace{1.5em}
\begin{subfigure}[t]{0.65\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/bcgraph}
\vspace{-1em}
\caption{The barcode graph for the filtration shown in Figure~\ref{fig:filt}.}
\label{fig:bcgraph}
\end{subfigure}
\vspace{1em}
\begin{subfigure}[t]{\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/bcforest}
\vspace{-1em}
\caption{Barcode forests constructed in Algorithm~\ref{alg:0-zigzag}
for the barcode graph in Figure~\ref{fig:bcgraph}.
For brevity, some forests are skipped.
The horizontally arranged labels indicate the levels.}
\label{fig:bcforest}
\end{subfigure}
\caption{Examples of a zigzag filtration, a barcode graph, and barcode forests.}
\label{fig:example}
\end{figure}
\begin{definition}[Barcode graph]\label{dfn:bcgraph}
For a graph $G$ and a zigzag filtration
$\mathcal{F}: G_0 \leftrightarrow G_1 \leftrightarrow
\cdots \leftrightarrow G_m$ of $G$,
the
\textbf{barcode graph} $\mathbb{G}_\mathrm{B}(\mathcal{F})$ of $\mathcal{F}$
is a graph whose vertices {\rm(}preferably called \textbf{nodes}{\rm)}
are associated with a \textbf{level}
and whose edges connect nodes only at adjacent levels.
The graph $\mathbb{G}_\mathrm{B}(\mathcal{F})$ is constructively described as follows:
\begin{itemize}
\item For each $G_i$ in $\mathcal{F}$ and each connected component of $G_i$,
there is a node in $\mathbb{G}_\mathrm{B}(\mathcal{F})$ at {level}~$i$ corresponding
to this component;
this node is also called a \textbf{level-$i$ node}.
\item For each inclusion $G_i\leftrightarrow G_{i+1}$ in $\mathcal{F}$,
if it is forward,
then there is an edge connecting a level-$i$ node $v_i$
to a level-$(i+1)$ node $v_{i+1}$ if and only if
the component of $v_i$ maps to the component of $v_{i+1}$ by the inclusion.
Similarly, if the inclusion is backward, then
$v_i$ connects to $v_{i+1}$ by an edge
iff
the component of $v_{i+1}$ maps to the component of $v_{i}$.
\end{itemize}
For two nodes at different levels in $\mathbb{G}_\mathrm{B}(\mathcal{F})$,
the node at the higher {\rm(}resp. lower{\rm)} level
is said to be \textbf{higher} {\rm(}resp. \textbf{lower}{\rm)} than the other.
\end{definition}
\begin{remark}
Note that some works~\cite{dey2019computing,kim2017stable} also have used similar notions of barcode graphs.
\end{remark}
Figure~\ref{fig:filt} and~\ref{fig:bcgraph} give an example of
a zigzag filtration and its barcode graph.
Note that a barcode graph is of size $O(mn)$,
where $m$ is the length of $\mathcal{F}$
and $n$ is the number of vertices and edges
of $G$. Although we present our algorithm
(Algorithm~\ref{alg:0-zigzag}) by first building the barcode graph,
the implementation does not do so explicitly, allowing us
to achieve the claimed time complexity;
see Section~\ref{sec:0-zz-imp}
for the implementation details.
Introducing barcode graphs
helps us justify the algorithm,
and more importantly, points to the fact
that the algorithm can be applied whenever such a barcode graph can be built.
\begin{algr}[Algorithm for 0-th zigzag persistence]
\label{alg:0-zigzag}
\begin{itemize}\item[]\end{itemize}\noindent
Given a graph $G$ and a zigzag filtration
$\mathcal{F}: \emptyset=G_0 \leftrightarrow G_1 \leftrightarrow
\cdots \leftrightarrow G_m$ of $G$,
we first build the barcode graph $\mathbb{G}_\mathrm{B}(\mathcal{F})$,
and then apply the pairing algorithm described in~\cite{agarwal2006extreme}
on $\mathbb{G}_\mathrm{B}(\mathcal{F})$ to compute $\mathsf{Pers}(\mathsf{H}_0(\mathcal{F}))$.
For a better understanding,
we rephrase this algorithm
which originally works on Reeb graphs:
The algorithm iterates for $i=0,\ldots,m-1$
and maintains a \textbf{barcode forest} $T_i$,
whose leaves have a one-to-one correspondence
to level-$i$ nodes of $\mathbb{G}_\mathrm{B}(\mathcal{F})$.
Like the barcode graph,
each tree node in a barcode forest is associated with a level and each tree edge
connects nodes at adjacent levels.
For each tree in a barcode forest,
the lowest node is the root.
Initially, $T_0$ is empty;
then, the algorithm builds $T_{i+1}$ from $T_{i}$
in the $i$-th iteration.
Intervals for $\mathsf{Pers}(\mathsf{H}_0(\mathcal{F}))$ are produced while updating the barcode forest.
{\rm(}Figure~\ref{fig:bcforest} illustrates such updates.{\rm)}
Specifically, the $i$-th iteration proceeds as follows:
first, $T_{i+1}$ is formed by copying
the level-$(i+1)$ nodes of $\mathbb{G}_\mathrm{B}(\mathcal{F})$ and their connections
to the level-$i$ nodes, into $T_{i}$;
the copying is possible because leaves of $T_i$ and
level-$i$ nodes of $\mathbb{G}_\mathrm{B}(\mathcal{F})$ have a one-to-one correspondence;
see transitions from $T_5$ to $T_6$ and from $T_9$ to $T_{10}$
in Figure~\ref{fig:bcforest}.
We further change $T_{i+1}$
under the following events:
\vspace{0.3em}\begin{description}
\item[Entrance:] One level-$(i+1)$ node in $T_{i+1}$, said to be \textbf{entering},
does not connect to any level-$i$ node.
\item[Split:] One level-$i$ node in $T_{i+1}$,
said to be \textbf{splitting},
connects to
two different level-$(i+1)$ nodes.
For the two events so far,
no changes need to be made on $T_{i+1}$.
\item[Departure:] One level-$i$ node $u$ in $T_{i+1}$,
said to be \textbf{departing},
does not connect to any level-$(i+1)$ node.
If $u$ has splitting ancestors {\rm(}i.e., ancestors which are also splitting nodes{\rm)}, add an
interval $[j+1,i]$ to $\mathsf{Pers}(\mathsf{H}_0(\mathcal{F}))$, where
$j$ is the level of the highest splitting ancestor $v$ of $u$;
otherwise, add an
interval $[j,i]$ to $\mathsf{Pers}(\mathsf{H}_0(\mathcal{F}))$, where $j$ is the level of the root $v$ of $u$.
We then delete the path from $v$ to $u$ in $T_{i+1}$.
\item[Merge:] Two different level-$i$ nodes $u_1,u_2$ in $T_{i+1}$
connect to the same level-$(i+1)$ node.
Tentatively, $T_{i+1}$ may now contain a loop
and is not a tree.
If $u_1,u_2$ are in different trees in $T_i$,
add an
interval $[j,i]$ to $\mathsf{Pers}(\mathsf{H}_0(\mathcal{F}))$,
where $j$ is the level of the higher root of $u_1,u_2$ in $T_i$;
otherwise, add an
interval $[j+1,i]$ to $\mathsf{Pers}(\mathsf{H}_0(\mathcal{F}))$, where $j$ is the level of
the highest common ancestor of $u_1,u_2$ in $T_i$.
We then glue the two paths from $u_1$ and $u_2$ to their level-$j$ ancestors
in $T_{i+1}$,
after which $T_{i+1}$ is guaranteed to be a tree.
\item[No-change:] If none of the above events happen,
no changes are made on $T_{i+1}$.
\end{description}
\vspace{-0.5em}
At the end, for each root in $T_{m}$ at a level $j$,
add an
interval $[j,m]$ to $\mathsf{Pers}(\mathsf{H}_0(\mathcal{F}))$,
and for each splitting node
in $T_{m}$ at a level $j$,
add an
interval $[j+1,m]$ to $\mathsf{Pers}(\mathsf{H}_0(\mathcal{F}))$.
\end{algr}
\begin{remark}
The justification of Algorithm~\ref{alg:0-zigzag} is given in Section~\ref{sec:0-zigzag-proof}.
\end{remark}
Figure~\ref{fig:bcforest} gives examples of barcode forests
constructed by Algorithm~\ref{alg:0-zigzag}
for the barcode graph shown in Figure~\ref{fig:bcgraph},
where $T_1$ and $T_2$ introduce entering nodes,
$T_6$ introduces a splitting node,
and $T_{10}$ introduces a departing node.
In $T_{10}$, the departure event happens and the dotted path is deleted,
producing an interval $[8,9]$.
In $T_3$ and $T_9$, the merge event happens and the dotted paths
are glued together, producing intervals $[2,2]$ and $[6,8]$.
Note that the glued level-$i$ nodes are in different trees
in $T_3$ and are in the same tree in $T_9$.
\subsection{Implementation}
\label{sec:0-zz-imp}
As mentioned, to achieve the claimed time complexity,
we do not explicitly build the barcode graph.
Instead, we differentiate the different events as follows:
inserting (resp. deleting) a vertex in $\mathcal{F}$ simply corresponds to
the entrance (resp. departure) event,
whereas inserting (resp. deleting) an edge corresponds to
the merge (resp. split) event only when connected components in the graph
merge (resp. split).
To keep track of the connectivity of vertices,
we use a {\it dynamic connectivity} data structure by Holm~et~al.~\cite{holm2001poly},
which we denote as $\mathbb{D}$.
Assuming that
$m$ is the length of $\mathcal{F}$ and
$n$ is the number of vertices and edges of $G$,
the data structure $\mathbb{D}$ supports the following operations:
\begin{itemize}
\item Return the identifier\footnote{Since $\mathbb{D}$
maintains the connectivity information by
dynamically updating the spanning forest for the current graph,
the identifier of a connected component is indeed the identifier
of a tree in the spanning forest.}
of the connected component of a vertex $v$ in $O(\log n)$ time.
We denote this subroutine as $\mathtt{find}(v)$.
\item Insert or delete an edge,
and possibly update the connectivity information, in $O(\log^2 n)$ amortized time.
\end{itemize}
We also note the following implementation details:
\begin{itemize}
\item All vertices of $G$ are added to $\mathbb{D}$ initially
and are then never deleted.
But we make sure that
edges in $\mathbb{D}$ always equal edges in $G_i$
as the algorithm proceeds
so that $\mathbb{D}$ still records the connectivity of $G_i$.
\item At each iteration $i$,
we update $T_i$
to form $T_{i+1}$ according to the changes of the connected components
from $G_i$ to $G_{i+1}$.
For this,
we maintain a key-value map $\phi$ from connected components of $\mathbb{D}$
to leaves of the barcode forest,
and $\phi$ is initially empty.
\item In a barcode forest $T_i$,
since the level of a leaf
always equals $i$,
we only record the level of a non-leaf node.
Note that at iteration $i$, a leaf in $T_{i}$ may
uniquely connect to a single leaf in $T_{i+1}$.
In this case,
we simply let the leaf in $T_{i}$ automatically become
a leaf in $T_{i+1}$; see Figure~\ref{fig:T3toT4}.
The size of a barcode forest is then $O(m)$.
\end{itemize}
\begin{figure}
\centering
\includegraphics[width=0.33\linewidth]{fig/T3toT4}
\caption{For the example in Figure~\ref{fig:example},
to form $T_4$,
our implementation only adds a level-4 entering node,
whereas the leaf in $T_3$ is not touched.
Since the level of a leaf always equals the index of the barcode forest,
the leaf at level 3 in $T_3$
automatically becomes a leaf at level 4 in $T_4$.}
\label{fig:T3toT4}
\end{figure}
Now we can present the full detail of the implementation.
Specifically,
for each addition and deletion in $\mathcal{F}$,
we do the following in each case:
\begin{description}
\item[Adding vertex $\fsimp{\mathcal{F}}{i}=v$:]
Add an isolated node to the barcode forest and let
$\phi(\mathtt{find}(v))$ equal this newly added node.
\item[Deleting vertex $\fsimp{\mathcal{F}}{i}=v$:]
Let $\ell=\phi(\mathtt{find}(v))$; then, $\ell$ is the node in the barcode
forest that is departing.
Update the barcode forest as described in Algorithm~\ref{alg:0-zigzag}.
\item[Adding edge $\fsimp{\mathcal{F}}{i}=(u,v)$:]
Let $t_1=\mathtt{find}(u)$, $t_2=\mathtt{find}(v)$,
$\ell_1=\phi(t_1)$, and $\ell_2=\phi(t_2)$.
If $t_1=t_2$, then the no-change event happens;
otherwise, the merge event happens.
We then add $(u,v)$ to $\mathbb{D}$.
For the no-change event,
do nothing after this.
For the merge event, do the following:
glue the paths from $\ell_1$ and $\ell_2$ to their ancestors
as described in Algorithm~\ref{alg:0-zigzag};
attach a new child $\ell$ to the highest
glued node; update $\phi(\mathtt{find}(u))$ to be $\ell$.
\item[Deleting edge $\fsimp{\mathcal{F}}{i}=(u,v)$:]
Let $\ell=\phi(\mathtt{find}(u))$, and then delete $(u,v)$ from $\mathbb{D}$.
If $\mathtt{find}(u)=\mathtt{find}(v)$ after this, then the no-change event happens
but we have to update $\phi(\mathtt{find}(u))$ to be $\ell$
because the identifiers of the connected components in $\mathbb{D}$ may change
after deleting the edge~\cite{holm2001poly}.
Otherwise, the split event happens: we attach two new
children $\ell_1$, $\ell_2$ to $\ell$ in the barcode forest and set
$\phi(\mathtt{find}(u))=\ell_1$, $\phi(\mathtt{find}(v))=\ell_2$.
\end{description}
\paragraph*{Mergeable trees.}
Following the idea in~\cite{agarwal2006extreme},
the barcode forest can be implemented
using the {\it mergeable trees} data structure by Georgiadis et al.~\cite{georgiadis2011data}.
Since the maximum number of nodes in a barcode forest is $O(m)$,
the data structure supports the following operations,
each of which takes $O(\log m)$ amortized time:
\begin{itemize}
\item Return the root of a node.
\item Return the nearest common ancestor of two leaves (in the same tree).
\item Glue the paths from two leaves (in the same tree) to their nearest common ancestor.
\end{itemize}
Note that
while we delete the path from the departing node to its ancestor
in the departure event,
deletions are not supported by mergeable trees.
However,
path deletions are indeed unnecessary
which are only meant for a clear exposition.
Hence,
during implementation, we only traverse each ancestor
of the departing node until an {\it unpaired}\footnote{An entering
or splitting node is initially {\it unpaired} when introduced and becomes {\it paired}
when its level is used to produce an interval.
E.g., the node $v$ becomes paired in the departure event in Algorithm~\ref{alg:0-zigzag}.}
one is found
without actual deletions.
Since each node can only
be traversed once, the traversal in the departure events
takes $O(m)$ time in total.
See \cite[Section 5]{georgiadis2011data}
for details of implementing the barcode forest and its operations
using mergeable trees.
\paragraph*{Complexity.}
The time complexity of the algorithm is $O(m\log^2 n+m\log m)$
dominated by the operations of the dynamic connectivity
and the mergeable trees data structures.
\subsection{Justification}
\label{sec:0-zigzag-proof}
In this subsection, we justify the correctness of Algorithm~\ref{alg:0-zigzag}.
For each entering node $u$ in a $T_i$ of Algorithm~\ref{alg:0-zigzag},
there must be a single
node in $\mathbb{G}_\mathrm{B}(\mathcal{F})$ at the level of $u$
with the same property.
So we also have entering nodes in $\mathbb{G}_\mathrm{B}(\mathcal{F})$.
Splitting and departing nodes
in $\mathbb{G}_\mathrm{B}(\mathcal{F})$
can be similarly defined.
We first prepare some standard notions and facts
in zigzag persistence
(Definition~\ref{dfn:rep-cls} and~\ref{dfn:pos-neg-inds},
Proposition~\ref{prop:pn-paring-w-rep})
that help with our proofs.
Some notions also appear in previous works in different forms;
see, e.g., \cite{maria2014zigzag}.
\begin{definition}[Representatives]
\label{dfn:rep-cls}
Let
$\mathcal{M}: V_0 \leftrightarrowsp{\psi_0} \cdots \leftrightarrowsp{\psi_{m-1}} V_m$
be an elementary zigzag module
and $[s,t]\subseteq[1,m]$ be an interval.
An indexed set
$\Set{\alpha_i\in V_i\,|\, i\in[s,t]}$
is called a set of \textbf{partial representatives}
for $[s,t]$ if for every $i\in[s,t-1]$,
$\alpha_i\mapsto\alpha_{i+1}$ or $\alpha_i\mapsfrom\alpha_{i+1}$ by $\psi_i$;
it is called a set of \textbf{representatives} for $[s,t]$
if the following additional conditions are satisfied:
\begin{enumerate}
\item \label{itm:dfn-rep-cls-birth}
If $\psi_{s-1}:V_{s-1}\rightarrow V_s$ is forward
with non-trivial cokernel,
then $\alpha_s$ is not in $\mathsf{img}(\psi_{s-1})$;
if $\psi_{s-1}: V_{s-1}\leftarrow V_{s}$ is backward
with non-trivial kernel,
then $\alpha_s$ is the non-zero element in $\ker(\psi_{s-1})$.
\item \label{itm:dfn-rep-cls-death}
If $t<m$ and
$\psi_{t}: V_{t}\leftarrow V_{t+1}$ is backward
with non-trivial cokernel,
then $\alpha_t$ is not in $\mathsf{img}(\psi_{t})$;
if $t<m$
and $\psi_{t}: V_t\rightarrow V_{t+1}$ is forward
with non-trivial kernel,
then $\alpha_t$ is the non-zero element in $\ker(\psi_{t})$.
\end{enumerate}
Specifically, when $\mathcal{M}:=\mathsf{H}_p(\mathcal{F})$ for a zigzag filtration $\mathcal{F}$,
we use terms
\textbf{$p$-representatives}
and \textbf{partial $p$-representatives} to emphasize the dimension $p$.
\end{definition}
\begin{remark}
Let $\mathcal{F}$ be the filtration given in Figure~\ref{fig:filt},
and let $\alpha_8$, $\alpha_9$ be the sum of the component containing
vertex 1 and the component containing vertex 2 in $G_8$ and $G_9$.
Then, $\Set{\alpha_8,\alpha_9}$ is a set of 0-representatives
for the interval $[8,9]\in\mathsf{Pers}(\mathsf{H}_0(\mathcal{F}))$.
\end{remark}
\begin{definition}[Positive/negative indices]
\label{dfn:pos-neg-inds}
Let
$\mathcal{M}: V_0 \leftrightarrowsp{\psi_0} \cdots \leftrightarrowsp{\psi_{m-1}} V_m$
be an elementary zigzag module.
The set of \textbf{positive indices} of $\mathcal{M}$,
denoted $\mathsf{P}(\mathcal{M})$,
and the set of \textbf{negative indices} of $\mathcal{M}$,
denoted $\mathsf{N}(\mathcal{M})$,
are constructed as follows:
for each forward $\psi_i:V_i\toV_{i+1}$,
if $\psi_i$ is an injection with non-trivial cokernel,
add $i+1$ to $\mathsf{P}(\mathcal{M})$;
if $\psi_i$ is a surjection with non-trivial kernel,
add $i$ to $\mathsf{N}(\mathcal{M})$.
Furthermore,
for each backward $\psi_i:V_i\leftarrowV_{i+1}$,
if $\psi_i$ is an injection with non-trivial cokernel,
add $i$ to $\mathsf{N}(\mathcal{M})$;
if $\psi_i$ is a surjection with non-trivial kernel,
add $i+1$ to $\mathsf{P}(\mathcal{M})$.
Finally, add $\mathsf{rank}\, V_m$ copies of $m$ to $\mathsf{N}(\mathcal{M})$.
\end{definition}
\begin{remark}
For each $\psi_i:V_i\leftrightarrowV_{i+1}$ in Definition~\ref{dfn:pos-neg-inds},
if $i+1\in\mathsf{P}(\mathcal{M})$, then $i\not\in\mathsf{N}(\mathcal{M})$;
similarly, if $i\in\mathsf{N}(\mathcal{M})$, then $i+1\not\in\mathsf{P}(\mathcal{M})$.
Furthermore, if $\psi_i$ is an isomorphism,
then $i\not\in\mathsf{N}(\mathcal{M})$ and $i+1\not\in\mathsf{P}(\mathcal{M})$.
\end{remark}
Note that $\mathsf{N}(\mathcal{M})$ in Definition~\ref{dfn:pos-neg-inds}
is in fact a multi-set; calling it a set
should not cause any confusion in this paper though.
Also note that $|\mathsf{P}(\mathcal{M})|=|\mathsf{N}(\mathcal{M})|$,
and every index in $\mathsf{P}(\mathcal{M})$ (resp. $\mathsf{N}(\mathcal{M})$)
is the start (resp. end) of an interval in $\mathsf{Pers}(\mathcal{M})$.
This explains why we add $\mathsf{rank}\, V_m$ copies of $m$ to $\mathsf{N}(\mathcal{M})$
because there are always $\mathsf{rank}\, V_m$ number of intervals ending with $m$
in $\mathsf{Pers}(\mathcal{M})$;
see the example in Figure~\ref{fig:filt} where $\mathsf{rank}\,\mathsf{H}_0(G_{10})=2$.
\begin{proposition}\label{prop:pn-paring-w-rep}
Let $\mathcal{M}$ be an elementary zigzag module
and $\pi:\mathsf{P}(\mathcal{M})\to\mathsf{N}(\mathcal{M})$ be a bijection.
If every $b\in\mathsf{P}(\mathcal{M})$ satisfies that $b\leq\pi(b)$ and the interval $[b,\pi(b)]$
has a set of representatives,
then $\mathsf{Pers}(\mathcal{M})=\Set{[b,\pi(b)]\,|\, b\in\mathsf{P}(\mathcal{M})}$.
\end{proposition}
\begin{proof}
For each $b\in\mathsf{P}(\mathcal{M})$,
let $\bigSet{\alpha^b_{j}\,|\, j\in[b,\pi(b)]}$ be
a set of representatives for $[b,\pi(b)]$.
Then, define $\mathcal{I}^{[b,\pi(b)]}$ as an interval submodule of $\mathcal{M}$ over $[b,\pi(b)]$
such that $\mathcal{I}^{[b,\pi(b)]}(j)$
is generated by $\alpha^b_j$
if $j\in[b,\pi(b)]$ and is trivial otherwise,
where $\mathcal{I}^{[b,\pi(b)]}(j)$ denotes the $j$-th vector space in $\mathcal{I}^{[b,\pi(b)]}$.
We claim that
$\mathcal{M}=\bigoplus_{b\in\mathsf{P}(\mathcal{M})}\mathcal{I}^{[b,\pi(b)]}$,
which implies the proposition.
To prove this,
suppose that $\mathcal{M}$ is of the form
\[\mathcal{M}: V_0 \leftrightarrowsp{\psi_0}
V_1 \leftrightarrowsp{\psi_1}
\cdots \leftrightarrowsp{\psi_{m-1}} V_m\]
Then, we only need to verify that
for every $i\in[0,m]$, the set
$\bigSet{\alpha^b_i\,|\, b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni i}$
is a basis of $V_i$.
We prove this by induction on $i$.
For $i=0$, since $V_0=0$,
$\bigSet{\alpha^b_0\,|\, b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni 0}=\emptyset$
is obviously a basis.
So we can assume that for an $i\in[0,m-1]$,
$\bigSet{\alpha^b_i\,|\, b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni i}$
is a basis of $V_i$.
We have the following cases:
\begin{description}
\item[$\psi_i$ an isomorphism:]
In this case,
$i\not\in\mathsf{N}(\mathcal{M})$ and $i+1\not\in\mathsf{P}(\mathcal{M})$.
If $\psi_i:V_i\toV_{i+1}$ is forward, then
$\bigSet{\alpha^b_{i+1}\,|\, b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni i+1}
=\bigSet{\psi_i(\alpha^b_{i})\,|\, b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni i}$.
The elements in
$\bigSet{\alpha^b_{i+1}\,|\, b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni i+1}$
must then form a basis of $V_{i+1}$
because $\psi_i$ is an isomorphism.
The verification for $\psi_i$ being backward is similar.
\item[$\psi_i:V_i\toV_{i+1}$ forward, $\mathsf{coker}(\psi_i)$ non-trivial:]
In this case,
$i\not\in\mathsf{N}(\mathcal{M})$ and $i+1\in\mathsf{P}(\mathcal{M})$.
For each $b\in\mathsf{P}(\mathcal{M})$ such that $[b,\pi(b)]\ni i$,
$[b,\pi(b)]\ni i+1$ and
$\alpha^b_i\mapsto \alpha^b_{i+1}$ by $\psi_i$.
We then have that elements in
$\bigSet{\alpha^b_{i+1}=\psi_i(\alpha^b_i)
\,|\, b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni i}$
are linearly independent
because $\psi_i$ is injective.
Since $\alpha^{i+1}_{i+1}\not\in\mathsf{img}(\psi_i)$
by Definition~\ref{dfn:rep-cls},
$\bigSet{\alpha^b_{i+1}\,|\, b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni i+1}=
\bigSet{\alpha^b_{i+1}
\,|\, b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni i}
\union\bigSet{\alpha^{i+1}_{i+1}}$ must contain linearly independent elements.
The fact that the cardinality of the set equals $\mathsf{rank}\,V_{i+1}$ implies that
it must form a basis of $V_{i+1}$.
\item[$\psi_i:V_i\toV_{i+1}$ forward, $\ker(\psi_i)$ non-trivial:]
In this case,
$i\in\mathsf{N}(\mathcal{M})$ and $i+1\not\in\mathsf{P}(\mathcal{M})$.
Let $j=\pi^{-1}(i)$.
For each $b\in\mathsf{P}(\mathcal{M})$ such that $[b,\pi(b)]\ni i$ and $b\neq j$,
$[b,\pi(b)]\ni i+1$ and
$\alpha^b_i\mapsto \alpha^b_{i+1}$ by $\psi_i$.
We then have that
$\bigSet{\alpha^b_{i+1}\,|\, b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni i+1}
=\bigSet{\psi_i\big(\alpha^b_{i}\big)\,|\,
b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni i}\setminus
\bigSet{\psi_i\big(\alpha_{i}^j\big)}$.
Since $\psi_i$ is surjective,
elements in $\bigSet{\psi_i\big(\alpha_{i}^b\big)
\,|\, b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni i}$
generate $V_{i+1}$,
in which $\psi_i\big(\alpha_{i}^j\big)=0$ by Definition~\ref{dfn:rep-cls}.
It follows that
$\bigSet{\alpha^b_{i+1}\,|\, b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni i+1}$
forms a basis of $V_{i+1}$
because it generates $V_{i+1}$
and its cardinality equals $\mathsf{rank}\,V_{i+1}$.
\item[$\psi_i:V_i\leftarrowV_{i+1}$ backward, $\mathsf{coker}(\psi_i)$ non-trivial:]
In this case,
$i\in\mathsf{N}(\mathcal{M})$ and $i+1\not\in\mathsf{P}(\mathcal{M})$.
For each $b\in\mathsf{P}(\mathcal{M})$ such that $[b,\pi(b)]\ni i$ and $\pi(b)\neq i$,
$[b,\pi(b)]\ni i+1$ and
$\alpha^b_i\mapsfrom \alpha^b_{i+1}$ by $\psi_i$.
We then have that elements in
$\bigSet{\alpha^b_{i+1}=(\psi_i)^{-1}(\alpha^b_i)
\,|\, b\in\mathsf{P}(\mathcal{M}),\,[b,\pi(b)]\ni i,\,\text{and}\,\pi(b)\neq i}$
are linearly independent
because if they are not,
then their images under $\psi_i$ are also not,
which is a contradiction.
Note that
$\bigSet{\alpha^b_{i+1}\,|\, b\in\mathsf{P}(\mathcal{M}),[b,\pi(b)]\ni i+1}=
\bigSet{\alpha^b_{i+1}
\,|\, b\in\mathsf{P}(\mathcal{M}),\,[b,\pi(b)]\ni i,\,\text{and}\,\pi(b)\neq i}$
and its cardinality equals $\mathsf{rank}\,V_{i+1}$,
so it must form a basis of $V_{i+1}$.
\item[$\psi_i:V_i\leftarrowV_{i+1}$ backward, $\ker(\psi_i)$ non-trivial:]
In this case,
$i\not\in\mathsf{N}(\mathcal{M})$ and $i+1\in\mathsf{P}(\mathcal{M})$.
For each $b\in\mathsf{P}(\mathcal{M})$ such that $[b,\pi(b)]\ni i$,
$[b,\pi(b)]\ni i+1$ and
$\alpha^b_i\mapsfrom \alpha^b_{i+1}$ by $\psi_i$.
We then have that elements in
$\bigSet{\alpha^b_{i+1}\in(\psi_i)^{-1}(\alpha^b_i)
\,|\, b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni i}$
are linearly independent
because their images under $\psi_i$ are.
We also have that there is no non-trivial linear combination of
$\bigSet{\alpha^b_{i+1}
\,|\, b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni i}$
falling in $\ker(\psi_i)$ because otherwise
their images under $\psi_i$ would not be linearly
independent.
Since $\alpha_{i+1}^{i+1}$ is the non-zero element in $\ker(\psi_i)$
by Definition~\ref{dfn:rep-cls},
we have that
$\bigSet{\alpha^b_{i+1}\,|\, b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni i+1}=
\bigSet{\alpha^b_{i+1}
\,|\, b\in\mathsf{P}(\mathcal{M})\text{ and }[b,\pi(b)]\ni i}
\union\bigSet{\alpha_{i+1}^{i+1}}$
contains linearly independent elements.
Then, it must form a basis of $V_{i+1}$
because its cardinality equals $\mathsf{rank}\,V_{i+1}$,
\qedhere
\end{description}
\end{proof}
Now we present several propositions leading to
our conclusion (Theorem~\ref{thm:0-zigzag-alg-correct}).
Specifically,
Proposition~\ref{prop:path-induce-partial-rep} states that a certain path
in $\mathbb{G}_\mathrm{B}(\mathcal{F})$ induces a set of partial 0-representatives.
Proposition~\ref{prop:0-zigzag-alg-invari} lists some invariants
of Algorithm~\ref{alg:0-zigzag}.
Proposition~\ref{prop:path-induce-partial-rep} and~\ref{prop:0-zigzag-alg-invari}
support the proof of Proposition~\ref{prop:has-rep},
which together with Proposition~\ref{prop:pn-paring-w-rep}
implies Theorem~\ref{thm:0-zigzag-alg-correct}.
From now on, $G$ and $\mathcal{F}$ always denote the input to Algorithm~\ref{alg:0-zigzag}.
Since each node in a barcode graph
represents a connected component,
we also interpret nodes in a barcode graph as 0-th homology classes
throughout the paper.
Moreover,
a path in a barcode graph from a node $v$ to a node $u$
is said to be {\it within level $j$ and $i$}
if for each node on the path, its level $\ell$
satisfies $j\leq \ell\leq i$;
we denote such a path as $\pathwl{v}{u}{j}{i}$.
\begin{proposition}\label{prop:path-induce-partial-rep}
Let $v$ be a level-$j$ node and $u$ be a level-$i$ node in $\mathbb{G}_\mathrm{B}(\mathcal{F})$
such that $j<i$ and there is a path $\pathwl{v}{u}{j}{i}$ in $\mathbb{G}_\mathrm{B}(\mathcal{F})$.
Then, there is
a set of partial 0-representatives $\Set{\alpha_k\in\mathsf{H}_0(G_k)\,|\, k\in[j,i]}$
for the interval $[j,i]$
with $\alpha_j=v$ and $\alpha_i=u$.
\end{proposition}
\begin{proof}
We can assume that $\pathwl{v}{u}{j}{i}$ is a simple path because if it were not
we could always find one.
For each $k\in[j+1,i-1]$,
let $w_1,\ldots,w_r$ be all the level-$k$ nodes on $\pathwl{v}{u}{j}{i}$
whose adjacent nodes on $\pathwl{v}{u}{j}{i}$ are at different levels.
Then, let
$\alpha_k=\sum_{\ell=1}^r w_\ell$.
Also, let $\alpha_j=v$ and $\alpha_i=u$.
It can be verified that $\Set{\alpha_k\,|\, k\in[j,i]}$
is a set of partial 0-representatives for $[j,i]$.
See Figure~\ref{fig:path} for an example of a simple path
$\pathwl{\tilde{v}_2}{u_2}{10}{13}$ (the dashed one)
in a barcode graph, where
the solid nodes
contribute to the induced partial 0-representatives
and the hollow nodes are excluded.
\end{proof}
For
an $i$ with $0\leq i\leqm$,
we define the {\it prefix} $\mathcal{F}^i$ of $\mathcal{F}$ as the filtration
$\mathcal{F}^i:G_0\leftrightarrow\cdots\leftrightarrow G_i$
and observe that $\mathbb{G}_\mathrm{B}(\mathcal{F}^i)$ is the subgraph of $\mathbb{G}_\mathrm{B}(\mathcal{F})$
induced by nodes at levels less than or equal to $i$.
We call level-$i$ nodes of $\mathbb{G}_\mathrm{B}(\mathcal{F}^i)$ as {\it leaves}
and do not distinguish
leaves in $T_i$ and $\mathbb{G}_\mathrm{B}(\mathcal{F}^i)$
because they bijectively map to each other.
It should be clear from the context though which graph or forest
a particular leaf is in.
\begin{figure}
\centering
\includegraphics[width=0.85\linewidth]{fig/path.eps}
\vspace{1em}
\caption{Illustration of invariants of Proposition~\ref{prop:0-zigzag-alg-invari}.
The top part contains a barcode graph with its filtration given
($a$, $b$, $c$, and $d$
are vertices of the complex). The bottom contains two barcode forests.}
\label{fig:path}
\end{figure}
\begin{proposition}\label{prop:0-zigzag-alg-invari}
For each $i=0,\ldots,m$,
Algorithm~\ref{alg:0-zigzag} maintains the following invariants:
\begin{enumerate}
\item\label{itm:tree-graph-bij-invari}
There is a bijection $\eta$ from
trees in $T_i$ to
connected components in $\mathbb{G}_\mathrm{B}(\mathcal{F}^i)$ containing leaves
such that a leaf $u$ is in a tree $\Upsilon$ of $T_i$ if and only if
$u$ is in $\eta(\Upsilon)$.
\item\label{itm:ances-leaf-path-invari}
For each leaf $u$ in $T_i$
and each ancestor of $u$ at a level $j$,
there is a path $\pathwl{\tilde{v}}{u}{j}{i}$
in $\mathbb{G}_\mathrm{B}(\mathcal{F})$
where $\tilde{v}$ is a level-$j$ node.
\item\label{itm:split-ances-leaf-path-invari}
For each leaf $u$ in $T_i$
and each splitting ancestor of $u$ at a level $j$,
let $\tilde{v}$ be the unique level-$j$ splitting node in $\mathbb{G}_\mathrm{B}(\mathcal{F})$.
Then, there is a path $\pathwl{\tilde{v}}{u}{j}{i}$
in $\mathbb{G}_\mathrm{B}(\mathcal{F})$.
\end{enumerate}
\end{proposition}
\begin{remark}
See Figure~\ref{fig:path} for examples of invariant~\ref{itm:ances-leaf-path-invari}
and~\ref{itm:split-ances-leaf-path-invari}.
In the figure,
$v_1$ is a level-1 non-splitting ancestor of $u_1$ in $T_7$
and $\tilde{v}_1$ is a level-1 node in the barcode graph;
$v_2$ is a level-10 splitting ancestor of $u_2$ in $T_{13}$
and $\tilde{v}_2$ is the unique level-10 splitting node in the barcode graph.
The paths $\pathwl{\tilde{v}_1}{u_1}{1}{7}$
and $\pathwl{\tilde{v}_2}{u_2}{10}{13}$ are marked with dashes.
\end{remark}
\begin{proof}
We only verify invariant~\ref{itm:split-ances-leaf-path-invari} as
the verification for invariant~\ref{itm:ances-leaf-path-invari}
is similar but easier and invariant~\ref{itm:tree-graph-bij-invari}
is straightforward.
The verification is by induction.
When $i=0$, invariant~\ref{itm:split-ances-leaf-path-invari} trivially holds.
Now suppose that invariant~\ref{itm:split-ances-leaf-path-invari} is true
for an $i\in[0,m-1]$.
For the no-change, entrance, and split event in Algorithm~\ref{alg:0-zigzag},
it is not hard to see that invariant~\ref{itm:split-ances-leaf-path-invari}
still holds for $i+1$.
For the departure event,
because we are only deleting a path from $T_i$ to form $T_{i+1}$,
invariant~\ref{itm:split-ances-leaf-path-invari}
also holds for $i+1$.
For the merge event,
let
$u$ be a leaf in $T_{i+1}$,
$v$ be a splitting ancestor of $u$ at a level $j$,
and $\tilde{v}$ be the unique splitting node in $\mathbb{G}_\mathrm{B}(\mathcal{F})$ at level $j$.
The node $v$ may correspond to one or two nodes in $T_i$,
in which only one is splitting, and let $v'$ be the splitting one.
Note that $u$'s parent
may correspond to one or two nodes in $T_i$,
and we let $W$ be the set of nodes in $T_i$
that $u$'s parent corresponds to.
If $v'$ is an ancestor of a node $w\in W$
in $T_i$, then by the assumption,
there must be a path $\pathwl{\tilde{v}}{w}{j}{i}$
in $\mathbb{G}_\mathrm{B}(\mathcal{F})$.
From this path,
we can derive a path $\pathwl{\tilde{v}}{u}{j}{i+1}$
in $\mathbb{G}_\mathrm{B}(\mathcal{F})$.
If $v'$ is not an ancestor of any node of $W$ in $T_i$,
the fact that $v$ is an ancestor of $u$'s parent in $T_{i+1}$ implies that
there must be an ancestor $v''$ of a node $w\in W$ in $T_{i}$
which $v$ corresponds to.
So we have that $v$ is a gluing of two nodes from $T_{i}$.
Note that $u$'s parent must not be a glued node in $T_{i+1}$
because otherwise $v'$ would have been an ancestor
of a node of $W$ in $T_i$;
see Figure~\ref{fig:path_exist}
where $z_1$ and $z_2$ are the two level-$i$ nodes glued together.
Let $x$ be the highest one among the nodes on the path from $v$ to $u$
that are glued in iteration~$i$.
We have that $x$ must correspond to a node $x'$ in $T_i$
which is an ancestor of $w$.
Recall that $z_1,z_2$ are the two leaves in $T_{i}$ which are glued,
and let $z_3$ be the child of the glued node of $z_1,z_2$ in $T_{i+1}$,
as shown in Figure~\ref{fig:path_exist}.
From the figure,
we have that $x'$ must be splitting because
one child of $x'$ (which is not glued) descends down to $w$
and the other child of $x'$ (which is glued) descends down to $z_1$.
The fact that $v'$ is an ancestor of $z_2$ in $T_i$ implies that
there is a path $\pathwl{\tilde{v}}{z_2}{j}{i}$
in $\mathbb{G}_\mathrm{B}(\mathcal{F})$.
Let $\tilde{x}$ be the unique splitting node in $\mathbb{G}_\mathrm{B}(\mathcal{F})$ at the same level with $x'$;
then, $z_1$ and $w$ being descendants of $x'$ in $T_i$
implies that there are
paths $\pathwl{z_1}{\tilde{x}}{j}{i}$ and $\pathwl{\tilde{x}}{w}{j}{i}$
in $\mathbb{G}_\mathrm{B}(\mathcal{F})$.
Now
we derive a path $\pathwl{\tilde{v}}{u}{j}{i+1}$
in $\mathbb{G}_\mathrm{B}(\mathcal{F})$
by concatenating the following paths and edges:
$\pathwl{\tilde{v}}{z_2}{j}{i}$,
$\bar{z_2z_3}$,
$\bar{z_3z_1}$,
$\pathwl{z_1}{\tilde{x}}{j}{i}$,
$\pathwl{\tilde{x}}{w}{j}{i}$,
$\bar{wu}$.
\end{proof}
\begin{figure}
\centering
\includegraphics[width=0.82\linewidth]{fig/path_exist.eps}
\vspace{1em}
\caption{Illustration of parts of $T_{i+1}$
for the proof of Proposition~\ref{prop:0-zigzag-alg-invari},
where the left one is before the path gluing
and the right one is after.
Note that the part between level $j$ and $i$ for the left tree
actually belongs to $T_i$. The paths with dashed marks are the glued ones
(before and after), in which $v'$ and $v''$ are identified as $v$,
and $x'$ is identified as $x$ with another node.
}
\label{fig:path_exist}
\end{figure}
\begin{proposition}\label{prop:has-rep}
Each interval produced by Algorithm~\ref{alg:0-zigzag} admits a set of 0-representatives.
\end{proposition}
\begin{proof}
Suppose that an interval is produced by the merge event at iteration~$i$.
We have the following situations:
\begin{itemize}
\item
If the nodes $u_1,u_2$ in this event (see Algorithm~\ref{alg:0-zigzag})
are in the same tree in $T_i$,
let $v$ be the highest common ancestor of $u_1,u_2$
and note that $v$ is a splitting node at level $j$.
Also note that
$u_1,u_2$ are actually leaves in $T_i$ and hence
can also be considered as level-$i$ nodes in $\mathbb{G}_\mathrm{B}(\mathcal{F})$.
Let $\tilde{v}$ be the unique level-$j$ splitting node in $\mathbb{G}_\mathrm{B}(\mathcal{F})$.
By invariant~\ref{itm:split-ances-leaf-path-invari}
of Proposition~\ref{prop:0-zigzag-alg-invari}
along with Proposition~\ref{prop:path-induce-partial-rep},
there are two sets of partial 0-representatives
$\Set{\alpha_k\,|\, k\in[j,i]},\Set{\beta_k\,|\, k\in[j,i]}$ for $[j,i]$
with $\alpha_j=\tilde{v}$, $\alpha_i=u_1$, $\beta_j=\tilde{v}$, and $\beta_i=u_2$.
We claim that $\Set{\alpha_k+\beta_k\,|\, k\in[j+1,i]}$
is a set of 0-representatives
for the interval $[j+1,i]$.
To prove this, we first note the following obvious facts:
{\sf(i)}~$\Set{\alpha_k+\beta_k\,|\, k\in[j+1,i]}$
is a set of partial 0-representatives;
{\sf(ii)}~$\alpha_{j+1}+\beta_{j+1}\in\ker(\morph{\mathcal{F}}{j}{0})$;
{\sf(iii)}~$\alpha_{i}+\beta_{i}$ is the non-zero element in $\ker(\morph{\mathcal{F}}{i}{0})$.
So we only need to show that $\alpha_{j+1}+\beta_{j+1}\neq 0$.
Let $v_1,v_2$ be the two level-$(j+1)$ nodes in $\mathbb{G}_\mathrm{B}(\mathcal{F})$
connecting to $\tilde{v}$.
Then, $\alpha_{j+1}$ equals $v_1$ or $v_2$ and the same for $\beta_{j+1}$.
To see this,
we first show that $\alpha_{j+1}$ can only contain $v_1,v_2$.
For contradiction,
suppose instead that $\alpha_{j+1}$ contains
a level-$(j+1)$ node $x$ with $x\neq v_1$, $x\neq v_2$.
Let $\pathwl{\tilde{v}}{u_1}{j}{i}$ be the simple path
that induces $\Set{\alpha_k\,|\, k\in[j,i]}$ as in
Proposition~\ref{prop:path-induce-partial-rep} and its proof.
Then, $x$ is on the path $\pathwl{\tilde{v}}{u_1}{j}{i}$ and
the two adjacent nodes of $x$ on $\pathwl{\tilde{v}}{u_1}{j}{i}$
are at level $j$ and $j+2$,
in which we let $y$ be the one at level $j$.
Note that $y\neq\tilde{v}$ because
$x$ is not equal to $v_1$ or $v_2$.
Since $\pathwl{\tilde{v}}{u_1}{j}{i}$ is within level $j$ and $i$,
$y$ must be adjacent to another level-$(j+1)$ node distinct from $x$
on $\pathwl{\tilde{v}}{u_1}{j}{i}$.
Now we have that $y$ is a level-$j$ splitting node with $y\neq\tilde{v}$,
contradicting the fact that $\mathbb{G}_\mathrm{B}(\mathcal{F})$
has only one level-$j$ splitting node.
The fact that $\alpha_{j+1}$ contains $v_1$ or $v_2$ but not both
can be similarly verified.
To see that $\alpha_{j+1}+\beta_{j+1}\neq 0$,
suppose instead that $\alpha_{j+1}+\beta_{j+1}= 0$, i.e., $\alpha_{j+1}=\beta_{j+1}$, and
without loss of generality they both equal $v_1$.
Note that we can consider $T_i$ as derived by contracting nodes of $\mathbb{G}_\mathrm{B}(\mathcal{F}^i)$
at the same level\footnote{
We should further note that
this contraction is not done on the entire $\mathbb{G}_\mathrm{B}(\mathcal{F}^i)$
but rather on connected components of $\mathbb{G}_\mathrm{B}(\mathcal{F}^i)$ containing leaves.}.
The fact that $\alpha_{j+1}=\beta_{j+1}=v_1$
implies that $u_1,u_2$ are descendants
of the same child of $v$ in $T_i$,
contradicting the fact that $v$ is the highest common ancestor of $u_1,u_2$.
So we have that
$\alpha_{j+1}+\beta_{j+1}\neq 0$.
\item
If $u_1,u_2$
are in different trees in $T_i$,
then without loss of generality
let $u_1$ be the one whose root $v_1$ is at the higher level (i.e., level $j$).
As the root of $u_1$,
the node $v_1$ must be an entering node, and
the connected component of $\mathbb{G}_\mathrm{B}(\mathcal{F}^i)$ containing $u_1$
must have a single level-$j$ node $\tilde{v}_1$.
Then, by invariant~\ref{itm:ances-leaf-path-invari}
of Proposition~\ref{prop:0-zigzag-alg-invari}
along with Proposition~\ref{prop:path-induce-partial-rep},
there are two sets of partial 0-representatives
$\Set{\alpha_k\,|\, k\in[j,i]},\Set{\beta_k\,|\, k\in[j,i]}$ for $[j,i]$
with $\alpha_j=\tilde{v}_1$, $\alpha_i=u_1$, $\beta_j=\tilde{v}_2$, and $\beta_i=u_2$,
where $\tilde{v}_2$ is a level-$j$ node.
We claim that $\Set{\alpha_k+\beta_k\,|\, k\in[j,i]}$
is a set of 0-representatives
for the interval $[j,i]$
and the verification is similar to the previous
case where $u_1$ and $u_2$ are in the same tree.
\end{itemize}
For intervals produced by the departure events and
at the end of the algorithm,
the existence of 0-representatives can be similarly argued.
\end{proof}
\begin{theorem}\label{thm:0-zigzag-alg-correct}
Algorithm~\ref{alg:0-zigzag} computes the 0-th zigzag barcode
for a given zigzag filtration.
\end{theorem}
\begin{proof}
First, we have the following facts:
every level-$j$ entering node in $\mathbb{G}_\mathrm{B}(\mathcal{F})$
introduces a $j\in\mathsf{P}(\mathsf{H}_0(\mathcal{F}))$
and uniquely corresponds to a level-$j$ root in $T_i$ for some $i$;
every level-$j$ splitting node in $\mathbb{G}_\mathrm{B}(\mathcal{F})$
introduces a $j+1\in\mathsf{P}(\mathsf{H}_0(\mathcal{F}))$
and uniquely corresponds to a level-$j$ splitting node in $T_i$ for some $i$.
Whenever an interval $[j,i]$ is produced in Algorithm~\ref{alg:0-zigzag},
$i\in\mathsf{N}(\mathsf{H}_0(\mathcal{F}))$ and the entering or splitting node in $T_i$
introducing $j$ as a positive index either becomes a {\it regular} node
(i.e., connecting to a single node on both adjacent levels)
or is deleted in $T_{i+1}$.
This means that $j$ is never the start of another interval produced.
At the end of Algorithm~\ref{alg:0-zigzag}, the number of intervals produced
which end with $m$
also matches the rank of $\mathsf{H}_0(G_m)$.
Therefore, intervals produced by the algorithm
induce a bijection $\pi:\mathsf{P}(\mathsf{H}_0(\mathcal{F}))\to\mathsf{N}(\mathsf{H}_0(\mathcal{F}))$.
By Proposition~\ref{prop:pn-paring-w-rep} and~\ref{prop:has-rep},
our conclusion follows.
\end{proof}
\section{One-dimensional zigzag persistence}
\label{sec:1-zigzag}
In this section, we present an efficient algorithm for 1-st zigzag persistence
on graphs.
We assume that
the input is a graph $G$ with a zigzag filtration
\[
\mathcal{F}:\emptyset=G_0 \leftrightarrowsp{\fsimp{\mathcal{F}}{0}}
G_1 \leftrightarrowsp{\fsimp{\mathcal{F}}{1}} \cdots
\leftrightarrowsp{\fsimp{\mathcal{F}}{m-1}} G_m
\]
of $G$.
We first describe the algorithm
without giving the full implementation details.
The key to the algorithm
is a pairing principle
for the positive and negative indices.
We then prove the correctness of the algorithm.
Finally, in Section~\ref{sec:1-zigzag-impl}, we make several observations
which reduce the index pairing to
finding the {\it max edge-weight}
of a path in a minimum spanning forest,
leading to an efficient implementation.
We notice that the following are true
for every inclusion $G_{i}\leftrightarrowsp{\fsimp{\mathcal{F}}{i}} G_{i+1}$ of $\mathcal{F}$
(recall that $\morph{\mathcal{F}}{i}{1}$ denotes
the corresponding linear map in the induced module $\mathsf{H}_1(\mathcal{F})$):
\begin{itemize}
\item If $\fsimp{\mathcal{F}}{i}$ is an edge being added
and vertices of $\fsimp{\mathcal{F}}{i}$ are connected
in $G_i$,
then $\morph{\mathcal{F}}{i}{1}$ is an injection with non-trivial cokernel,
which provides $i+1\in\mathsf{P}(\mathsf{H}_1(\mathcal{F}))$.
\item If $\fsimp{\mathcal{F}}{i}$ is an edge being deleted
and vertices of $\fsimp{\mathcal{F}}{i}$ are connected
in $G_{i+1}$,
then $\morph{\mathcal{F}}{i}{1}$ is an injection with non-trivial cokernel,
which provides $i\in\mathsf{N}(\mathsf{H}_1(\mathcal{F}))$.
\item In all the other cases,
$\morph{\mathcal{F}}{i}{1}$ is an isomorphism
and $i\not\in\mathsf{N}(\mathsf{H}_1(\mathcal{F}))$, $i+1\not\in\mathsf{P}(\mathsf{H}_1(\mathcal{F}))$.
\end{itemize}
As can be seen from Section~\ref{sec:0-zigzag},
computing $\mathsf{Pers}(\mathsf{H}_1(\mathcal{F}))$ boils down to finding a pairing of
indices of $\mathsf{P}(\mathsf{H}_1(\mathcal{F}))$ and $\mathsf{N}(\mathsf{H}_1(\mathcal{F}))$.
Our algorithm adopts this structure,
where $\Ucal_i$ denotes
the set of unpaired positive indices at the {\it beginning} of each iteration $i$:
\pagebreak
\begin{algr}[Algorithm for 1-st zigzag persistence on graphs]
\label{alg:1-zigzag-graph-abstr}
{\begin{itemize}
\item[]
\item[] $\Ucal_0:=\emptyset$
\item[] \textbf{for} $i:= 0,\ldots,m-1${\rm:}
\begin{itemize}
\item[] \textbf{if} $G_{i}\rightarrowsp{\fsimp{\mathcal{F}}{i}} G_{i+1}$
provides $i+1\in\mathsf{P}(\mathsf{H}_1(\mathcal{F}))${\rm:}
\begin{itemize}
\item[]
$\Ucal_{i+1}:=\Ucal_i\union\Set{i+1}$
\end{itemize}
\item[] \textbf{else if} $G_{i}\leftarrowsp{\fsimp{\mathcal{F}}{i}} G_{i+1}$
provides $i\in\mathsf{N}(\mathsf{H}_1(\mathcal{F}))${\rm:}
\begin{itemize}
\item[] pair $i$ with a $j_*\in\Ucal_i$ based on the Pairing Principle below
\item[] output an interval $[j_*,i]$ for $\mathsf{Pers}(\mathsf{H}_1(\mathcal{F}))$
\item[] $\Ucal_{i+1}:=\Ucal_i\setminus\Set{j_*}$
\end{itemize}
\item[] \textbf{else}{\rm:}
\begin{itemize}
\item[] $\Ucal_{i+1}:=\Ucal_i$
\end{itemize}
\end{itemize}
\item[] \textbf{for each} $j\in\Ucal_m${\rm:}
\begin{itemize}
\item[] output an interval $[j,m]$ for $\mathsf{Pers}(\mathsf{H}_1(\mathcal{F}))$
\end{itemize}
\end{itemize}}
\end{algr}
Note that in Algorithm~\ref{alg:1-zigzag-graph-abstr},
whenever a positive or negative index is produced,
$\fsimp{\mathcal{F}}{i}$ must be an edge.
One key piece missing from the algorithm
is how we choose a positive index to pair with a negative index:
\theoremstyle{theorem}
\newtheorem*{pprincile}{Pairing Principle for Algorithm~\ref{alg:1-zigzag-graph-abstr}}
\begin{pprincile}
In each iteration $i$
where $G_{i}\leftarrowsp{\fsimp{\mathcal{F}}{i}} G_{i+1}$
provides $i\in\mathsf{N}(\mathsf{H}_1(\mathcal{F}))$,
let $J_i$ consist of every $j\in\Ucal_i$ such that
there exists a 1-cycle $z$ containing both $\fsimp{\mathcal{F}}{j-1}$ and $\fsimp{\mathcal{F}}{i}$
with $z\subseteq G_k$ for every $k\in[j,i]$.
Then, $J_i\neq\emptyset$ and
Algorithm~\ref{alg:1-zigzag-graph-abstr} pairs $i$
with the {\rm smallest} index {$j_*$} in $J_i$.
\end{pprincile}
\begin{remark}
See Proposition~\ref{prop:iter-cyc-exist}
for a proof of $J_i\neq\emptyset$ claimed above.
\end{remark}
\begin{remark}
Algorithms for non-zigzag
persistence~\cite{edelsbrunner2000topological,zomorodian2005computing} always
pair a negative index with the \textit{largest} (i.e., youngest) positive index
satisfying a certain condition,
while Algorithm~\ref{alg:1-zigzag-graph-abstr} pairs with the smallest one.
This is due to the difference of zigzag and non-zigzag persistence
and our particular condition that 1-cycles can never become boundaries
in graphs.
See~\cite{carlsson2009zigzag-realvalue,maria2014zigzag}
for the pairing when assuming general zigzag filtrations.
\end{remark}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig/1cyc}
\caption{A zigzag filtration with 1-st barcode $\Set{[4,6],[2,8],[6,9],[8,9]}$.
For brevity, the addition of vertices and some edges are skipped.}
\label{fig:1cyc}
\end{figure}
Figure~\ref{fig:1cyc} gives an example of the pairing of the indices
and their corresponding edges.
In the figure, when edge $d$ is deleted from $G_6$,
there are three unpaired positive edges $a$, $b$, and $c$,
in which $b$ and $c$ admit 1-cycles as required by the Pairing Principle.
As the earlier edge, $b$ is paired with $d$
and an interval $[4,6]$ is produced.
The red cycle in $G_6$ indicates the 1-cycle containing $b$ and $d$
which exists in all the intermediate graphs.
Similar situations happen when $e$ is paired with $a$ in $G_8$,
producing the interval $[2,8]$.
For the correctness of Algorithm~\ref{alg:1-zigzag-graph-abstr},
we first provide
Proposition~\ref{prop:iter-cyc-exist}
which justifies
the Pairing Principle
and is a major step leading toward our conclusion (Theorem~\ref{thm:1-zigzag-graph-corr}):
\begin{proposition}\label{prop:iter-cyc-exist}
At the beginning of each iteration $i$ in Algorithm~\ref{alg:1-zigzag-graph-abstr},
for every $j\in\Ucal_i$,
there exists a 1-cycle $z^i_j$ containing $\fsimp{\mathcal{F}}{j-1}$
with $z^i_j\subseteq G_k$ for every $k\in[j,i]$.
Furthermore, the set $\Set{z^i_j\,|\, j\in\Ucal_i}$ forms a basis of $\mathsf{Z}_1(G_i)$.
If the iteration $i$ produces a negative index $i$,
then the above statements imply that
there is at least one $z^i_j$ containing $\fsimp{\mathcal{F}}{i}$.
This $z^i_j$ satisfies the condition that $z^i_j\subseteq G_k$ for every $k\in[j,i]$,
$\fsimp{\mathcal{F}}{j-1}\in z^i_j$, and $\fsimp{\mathcal{F}}{i}\in z^i_j$,
which implies that $J_i\neq\emptyset$ where $J_i$ is as defined in the Pairing Principle.
\end{proposition}
\begin{proof}
We prove this by induction. At the beginning of iteration $0$,
since $G_0=\emptyset$ and $\Ucal_0=\emptyset$,
the proposition is trivially true.
Suppose that the proposition is true at the beginning of an iteration $i$.
For each $j\in\Ucal_i$,
let $z^i_j$ be the 1-cycle as specified in the proposition.
If $G_{i}\leftrightarrowsp{\fsimp{\mathcal{F}}{i}} G_{i+1}$
produces neither a positive index nor a negative index,
then $\mathsf{Z}_1(G_i)=\mathsf{Z}_1(G_{i+1})$ and
$\Ucal_i=\Ucal_{i+1}$.
Let $z^{i+1}_j=z^i_j$ for each $j$;
then, $\Set{z^{i+1}_j\,|\, j\in\Ucal_{i+1}}$
serves as the 1-cycles as specified in the proposition
for iteration $i+1$.
If $G_{i}\rightarrowsp{\fsimp{\mathcal{F}}{i}} G_{i+1}$
produces a new unpaired positive index $i+1$,
let $z^{i+1}_{i+1}$ be any 1-cycle in $G_{i+1}$ containing $\fsimp{\mathcal{F}}{i}$.
Also, for each $j\in\Ucal_i$,
let $z^{i+1}_j=z^i_j$.
It can be verified that
$\Set{z^{i+1}_j\,|\, j\in\Ucal_{i+1}}$ serves as the 1-cycles as specified in the proposition
for iteration $i+1$.
If $G_{i}\leftarrowsp{\fsimp{\mathcal{F}}{i}} G_{i+1}$
produces a negative index $i$,
then there must be a 1-cycle in $G_i$ containing $\fsimp{\mathcal{F}}{i}$.
The fact that $\Set{z^{i}_j\,|\, j\in\Ucal_{i}}$ forms a basis of $\mathsf{Z}_1(G_i)$
implies that there must be at least one $z^i_j$ containing $\fsimp{\mathcal{F}}{i}$
because otherwise no combination of the $z^i_j$'s can
equal a cycle containing $\fsimp{\mathcal{F}}{i}$.
Let $\bar{j}$ be the smallest $j\in\Ucal_i$
such that $z^i_j$ contains $\fsimp{\mathcal{F}}{i}$.
We claim that $j_*=\bar{j}$, where $j_*$ is as defined in the Pairing Principle.
For contradiction, suppose instead that $j_*\neq\bar{j}$.
Note that $\bar{j}\in J_i$, where $J_i$ is as defined in the Pairing Principle.
Since $j_*$ is the smallest index in $J_i$,
we have that $j_*<\bar{j}$.
By the Pairing Principle,
there exists a 1-cycle $\zeta$ containing both
$\fsimp{\mathcal{F}}{j_*-1}$ and $\fsimp{\mathcal{F}}{i}$
with $\zeta\subseteq G_k$ for every $k\in[j_*,i]$.
Since $\Set{z^{i}_j\,|\, j\in\Ucal_i}$ forms a basis of $\mathsf{Z}_1(G_i)$
and $\zeta\subseteq G_i$,
$\zeta$ must equal a sum $\sum_{\ell=1}^s z^i_{\lambda_\ell}$,
where each $\lambda_\ell\in\Ucal_i$.
We rearrange the indices such that
$\lambda_1<\lambda_2<\cdots<\lambda_s$.
We have that $\lambda_s\geq\bar{j}$ because otherwise
each $\lambda_\ell<\bar{j}$ and
so its corresponding $z^i_{\lambda_\ell}$ does not contain $\fsimp{\mathcal{F}}{i}$.
This implies that $\zeta=\sum_{\ell=1}^s z^i_{\lambda_\ell}$
does not contain $\fsimp{\mathcal{F}}{i}$,
which is a contradiction.
For each $\ell$ such that $1\leq \ell<s$, since $\lambda_\ell\leq\lambda_s-1\leq i$,
we have that $z^i_{\lambda_\ell}\subseteq G_{\lambda_s-1}$,
which means that $\fsimp{\mathcal{F}}{\lambda_s-1}\not\in z^i_{\lambda_\ell}$
because $\fsimp{\mathcal{F}}{\lambda_s-1}\not\in G_{\lambda_s-1}$.
Since $\fsimp{\mathcal{F}}{\lambda_s-1}\in z^i_{\lambda_s}$,
it follows that $\fsimp{\mathcal{F}}{\lambda_s-1}\in\sum_{\ell=1}^s z^i_{\lambda_\ell}=\zeta$.
This implies that $\zeta\nsubseteq G_{\lambda_s-1}$
because $\fsimp{\mathcal{F}}{\lambda_s-1}\not\in G_{\lambda_s-1}$.
However, we have that $j_*\leq\lambda_s-1\leq i$ because $j_*<\bar{j}\leq\lambda_s\leq i$,
which means that $\zeta\subseteq G_{\lambda_s-1}$.
So we have reached a contradiction, meaning that $j_*=\bar{j}$.
For each $j\in\Ucal_{i+1}$,
if $z^i_j$ does not contain $\fsimp{\mathcal{F}}{i}$,
let $z^{i+1}_j=z^i_j\subseteq G_{i+1}$.
If $z^i_j$ contains $\fsimp{\mathcal{F}}{i}$,
let $z^{i+1}_j=z^i_j+z^i_{j_*}\subseteq G_{i+1}$.
Note that since $j_*=\bar{j}$,
we must have that $j_*<j$,
which means that $z^i_{j_*}\subseteq G_{k}$ for every $k\in[j,i]\subseteq[j_*,i]$.
Therefore, $z^{i+1}_j=z^i_j+z^i_{j_*}\subseteq G_{k}$ for every $k\in[j,i]$.
Also since $z^i_{j_*}\subseteq G_{j-1}$,
$z^i_{j_*}$ does not contain $\fsimp{\mathcal{F}}{j-1}$,
which means that $z^{i+1}_j$ contains $\fsimp{\mathcal{F}}{j-1}$.
Note that $\Set{z^{i+1}_j\,|\, j\in\Ucal_{i+1}}$ must still be linearly independent,
so they form a basis of $\mathsf{Z}_1(G_{i+1})$.
Now we have that
$\Set{z^{i+1}_j\,|\, j\in\Ucal_{i+1}}$ serves as the 1-cycles as specified in the proposition
for iteration $i+1$.
\end{proof}
\begin{theorem}\label{thm:1-zigzag-graph-corr}
Algorithm~\ref{alg:1-zigzag-graph-abstr} computes the 1-st zigzag barcode
for a given zigzag filtration on graphs.
\end{theorem}
\begin{proof}
The claim follows directly from Proposition~\ref{prop:pn-paring-w-rep}.
For each interval $[j_*,i]$ produced from the pairing
in Algorithm~\ref{alg:1-zigzag-graph-abstr},
by the Pairing Principle,
there exists a 1-cycle $z$ containing both
$\fsimp{\mathcal{F}}{j_*-1}$ and $\fsimp{\mathcal{F}}{i}$
with $z\subseteq G_k$ for every $k\in[j_*,i]$.
The cycle $z$ induces a set of 1-representatives for $[j_*,i]$.
For each interval produced at the end,
Proposition~\ref{prop:iter-cyc-exist} implies that such an interval
admits 1-representatives.
\end{proof}
\subsection{Efficient implementation}\label{sec:1-zigzag-impl}
For every $i$ and every $j\leq i$,
define
$\Gamma_j^i$ as the graph derived from $G_{j}$ by deleting
every edge $\fsimp{\mathcal{F}}{k}$
s.t.\ $j\leq k<i$
and $G_{k}\leftarrowsp{\fsimp{\mathcal{F}}{k}} G_{k+1}$
is backward.
For convenience, we also assume that $\Gamma_j^i$
contains all the vertices of $G$.
We can simplify the Pairing Principle
as suggested by the following proposition:
\begin{proposition}
\label{prop:pairing-rewrite}
In each iteration $i$ of Algorithm~\ref{alg:1-zigzag-graph-abstr}
where $G_{i}\leftarrowsp{\fsimp{\mathcal{F}}{i}} G_{i+1}$
provides $i\in\mathsf{N}(\mathsf{H}_1(\mathcal{F}))$,
the set $J_i$ in the Pairing Principle can be alternatively
defined as consisting of every $j\in\Ucal_i$
s.t.\ $\fsimp{\mathcal{F}}{i}\in\Gamma_j^{i}$ and
the vertices of $\fsimp{\mathcal{F}}{i}$ are connected in $\Gamma_j^{i+1}$
{\rm(}$\fsimp{\mathcal{F}}{i}\not\in\Gamma_j^{i+1}$ by definition{\rm)}.
\end{proposition}
\begin{proof
We prove an equivalent statement, which is that $J_i$
consists of every $j\in\Ucal_i$
s.t.\ there is a 1-cycle in $\Gamma_j^i$ containing $\fsimp{\mathcal{F}}{i}$.
Let $j$ be any index in $\Ucal_i$.
It is not hard to see that a 1-cycle is in $G_k$ for every $k\in[j,i]$
iff
the 1-cycle is in $\Gamma_j^i$.
So we only need to prove that there is a 1-cycle in $\Gamma_j^i$
containing both $\fsimp{\mathcal{F}}{j-1}$ and $\fsimp{\mathcal{F}}{i}$
iff
there is a 1-cycle in $\Gamma_j^i$ containing $\fsimp{\mathcal{F}}{i}$.
The forward direction is easy. So let $z$ be
a 1-cycle in $\Gamma_j^i$ containing $\fsimp{\mathcal{F}}{i}$.
If $z$ contains $\fsimp{\mathcal{F}}{j-1}$, then the proof is done.
If not, by Proposition~\ref{prop:iter-cyc-exist}
there is a 1-cycle $z'$ containing $\fsimp{\mathcal{F}}{j-1}$
with $z'\subseteq G_k$ for every $k\in[j,i]$.
So $z'$ is a 1-cycle in $\Gamma_j^i$ containing $\fsimp{\mathcal{F}}{j-1}$.
If $z'$ contains $\fsimp{\mathcal{F}}{i}$, we again finish our proof.
If not, then $z+z'$ is a 1-cycle
containing both edges.
\end{proof}
We then turn graphs in $\mathcal{F}$ into weighted ones
in the following way:
initially, $G_0=\emptyset$;
then, whenever an edge $\fsimp{\mathcal{F}}{i}$ is added from $G_{i}$ to $G_{i+1}$,
the weight $w(\fsimp{\mathcal{F}}{i})$ is set to $i$.
We have the following fact:
\begin{proposition}\label{prop:G-j-i-charac}
For every $i$ and every $j\leq i$,
the edge set of $\Gamma_j^{i}$, denoted $E(\Gamma_j^{i})$,
consists of all edges of $G_{i}$ whose weights are less than $j$.
\end{proposition}
\begin{proof}
We can prove this by induction on $i$. For $i=0$, $G_{i}=\emptyset$
and the proposition is trivially true.
Suppose that the proposition is true for $i$.
If $G_i$ and $G_{i+1}$ differ by a vertex, then the proposition
is also true for $i+1$ because the edges stay the same.
If $G_{i+1}$ is derived from $G_i$ by adding an edge $\fsimp{\mathcal{F}}{i}$,
by the assumption, $E(\Gamma_j^{i})$
consists of all edges of $G_{i}$ whose weights are less than $j$
for each $j\leq i$.
Note that $E(\Gamma_j^{i})=E(\Gamma_j^{i+1})$ because $G_i\rightarrowsp{\fsimp{\mathcal{F}}{i}}G_{i+1}$
is an addition.
So we have that $E(\Gamma_j^{i+1})$
consists of all edges of $G_{i+1}$ whose weights are less than $j$
because $w(\fsimp{\mathcal{F}}{i})=i\geq j$.
Since $E(\Gamma_{i+1}^{i+1})=E(G_{i+1})$, the claim is also true for $E(\Gamma_{i+1}^{i+1})$.
Now consider the situation that
$G_{i+1}$ is derived from $G_i$ by deleting an edge $\fsimp{\mathcal{F}}{i}$.
Then,
$\fsimp{\mathcal{F}}{i}$ must be added to the filtration previously,
and let $G_k\rightarrowsp{\fsimp{\mathcal{F}}{k}}G_{k+1}$
with $k<i$ and $\fsimp{\mathcal{F}}{k}=\fsimp{\mathcal{F}}{i}$ be the {\it latest}
such addition. Note that $w(\fsimp{\mathcal{F}}{i})=k$ in $G_i$.
For $j\leq k$,
$\fsimp{\mathcal{F}}{i}\not\in E(\Gamma_j^{i})$ because $w(\fsimp{\mathcal{F}}{i})=k\geq j$.
Since $E(\Gamma_j^{i+1})=E(\Gamma_j^{i})\setminus\Set{\fsimp{\mathcal{F}}{i}}$,
we have that $E(\Gamma_j^{i+1})=E(\Gamma_j^{i})$.
Therefore,
$E(\Gamma_j^{i+1})$
consists of all edges of $G_{i+1}$ whose weights are less than $j$
because $w(\fsimp{\mathcal{F}}{i})\geq j$ in $G_i$.
For each $j$ s.t.\ $k<j\leq i$,
we have $\fsimp{\mathcal{F}}{i}\in E(\Gamma_j^{i})$ and
$E(\Gamma_j^{i+1})=E(\Gamma_j^{i})\setminus\Set{\fsimp{\mathcal{F}}{i}}$.
Since $w(\fsimp{\mathcal{F}}{i})<j$ in $G_i$,
it is true that $E(\Gamma_j^{i+1})$
consists of all edges of $G_{i+1}$ whose weights are less than $j$,
and the proof is done.
\end{proof}
Suppose that in an iteration $i$ of Algorithm~\ref{alg:1-zigzag-graph-abstr},
$G_{i}\leftarrowsp{\fsimp{\mathcal{F}}{i}} G_{i+1}$
provides a negative index~$i$.
Let $\Ucal_i=\Set{j_1<j_2<\cdots<j_\ell}$
and the vertices of $\fsimp{\mathcal{F}}{i}$ be $u,v$.
Proposition~\ref{prop:G-j-i-charac} implies that
\begin{equation}\label{eqn:G-j-i-seq}
\Gamma_{j_1}^{i+1}\subseteq \Gamma_{j_2}^{i+1}\subseteq \cdots
\subseteq \Gamma_{j_\ell}^{i+1}\subseteq \Gamma_{i+1}^{i+1}
\end{equation}
By Proposition~\ref{prop:pairing-rewrite},
in order to find
the positive index to pair with $i$,
one only needs to find the smallest $j'_*\in\Ucal_i$
s.t.\ $u,v$ are connected in $\Gamma_{j'_*}^{i+1}$.
(Note that for a $j\in\Ucal_i$ to satisfy
$\fsimp{\mathcal{F}}{i}\in\Gamma_j^{i}$,
$j$ only needs to be greater than $w(\fsimp{\mathcal{F}}{i})$;
such a smallest $j$ is easy to derive.)
We further expand Sequence~(\ref{eqn:G-j-i-seq}) into the following finer version:
\begin{equation}\label{eqn:G-j-i-seq-finer}
\Gamma_{0}^{i+1}\subseteq \Gamma_{1}^{i+1}\subseteq \cdots
\subseteq \Gamma_{i}^{i+1}\subseteq \Gamma_{i+1}^{i+1}
\end{equation}
where each consecutive $\Gamma_{k}^{i+1},\Gamma_{k+1}^{i+1}$
are either the same or differ by only one edge.
To get $j'_*$, we instead scan Sequence~(\ref{eqn:G-j-i-seq-finer})
and find
the smallest $k_*\in\Set{0,1,\ldots,i+1}$
s.t.\ $u,v$ are connected in $\Gamma_{k_*}^{i+1}$.
Proposition~\ref{prop:k-start-in-MST}
characterize such a $k_*$:
\begin{proposition}\label{prop:k-start-in-MST}
For a path in a weighted graph, let the \textbf{max edge-weight} of the path
be the maximum weight of its edges.
Then,
the integer
$k_*-1$ equals the max edge-weight of the unique path connecting $u,v$
in the unique minimum spanning forest of $G_{i+1}$.
\end{proposition}
\begin{proof}
We first notice that, since weighted graphs considered in this paper
have distinct weights, they all have unique minimum spanning forests.
Let $T$ be the minimum spanning forest of $\Gamma_{i+1}^{i+1}$;
we prove an equivalent statement of the proposition,
which is that
$k_*-1$ equals the max edge-weight of the unique path connecting $u,v$
in $T$.
Since $k_*$ is the smallest index
s.t.\ $u,v$ are connected in $\Gamma_{k_*}^{i+1}$,
we have that $u,v$ are {\it not} connected in $\Gamma_{k_*-1}^{i+1}$.
This indicates that $\Gamma_{k_*}^{i+1}\neq\Gamma_{k_*-1}^{i+1}$.
Assume that $\Gamma_{k_*}^{i+1}$ and $\Gamma_{k_*-1}^{i+1}$ differ
by an edge $e$; then, $w(e)=k_*-1$.
Let $T'$ be the minimum spanning forest of $\Gamma_{k_*-1}^{i+1}$.
Then, $u,v$ are not connected in $T'$ and $e$ connects the connected
components of $u,v$ in $T'$. Since spanning forests have a matroid structure,
$T'\union\Set{e}$ must be a subforest of $T$ (indeed, $e$ would be the edge added to $T'$
by Kruskal's algorithm; see, e.g.,~\cite{CLRS3rd}).
Also, since there is a unique path between two vertices in a forest,
the path from $u$ to $v$ in $T'\union\Set{e}$
must be the path from $u$ to $v$ in $T$.
This path has a max edge-weight of $k_*-1$ and the proof is done.
\end{proof}
Based on Proposition~\ref{prop:k-start-in-MST},
finding $k_*$ reduces to computing
the max edge-weight of the path
connecting $u,v$ in the minimum spanning forest (MSF) of $G_{i+1}$.
For this, we utilize the {\it dynamic-MSF} data structure
proposed by Holm et al.~\cite{holm2001poly}.
Assuming that
$n$ is the number of vertices and edges of $G$,
the dynamic-MSF data structure supports the following operations:
\begin{itemize}
\item Return the identifier of a vertex's connected component in $O(\log n)$ time,
which can be used to determine whether two vertices are connected.
\item Return the max edge-weight of the path
connecting any two vertices in the MSF in $O(\log n)$ time.
\item Insert or delete an edge from the current graph (maintained by the data structure)
and possibly update the MSF in $O(\log^4 n)$ amortized time.
\end{itemize}
We now present the full details of Algorithm~\ref{alg:1-zigzag-graph-abstr}:
\pagebreak
\begin{algr*}[Algorithm~\ref{alg:1-zigzag-graph-abstr}: details]
\label{alg:1-zigzag-graph-concr}
\begin{itemize}\item[]\end{itemize}\noindent
Maintain a dynamic-MSF data structure $\mathbb{F}$,
which consists of all vertices of $G$ and no edges initially.
Also, set $\Ucal_0=\emptyset$.
Then, for each $i=0,\ldots,m-1$,
if $\fsimp{\mathcal{F}}{i}$ is a vertex, do nothing;
otherwise, do the following:
\vspace{0.3em}
\begin{description}
\item[Case $G_i\rightarrowsp{\fsimp{\mathcal{F}}{i}}G_{i+1}$:]
Check
whether vertices of $\fsimp{\mathcal{F}}{i}$ are connected in $G_i$
by querying $\mathbb{F}$, and then add $\fsimp{\mathcal{F}}{i}$ to $\mathbb{F}$.
If vertices of $\fsimp{\mathcal{F}}{i}$ are connected in $G_i$,
then
set $\Ucal_{i+1}=\Ucal_i\union\Set{i+1}$;
otherwise, set $\Ucal_{i+1}=\Ucal_i$.
\item[Case $G_i\leftarrowsp{\fsimp{\mathcal{F}}{i}}G_{i+1}$:]
Delete $\fsimp{\mathcal{F}}{i}$ from $\mathbb{F}$.
If the vertices $u,v$ of $\fsimp{\mathcal{F}}{i}$ are found to be {\rm not} connected in $G_{i+1}$
by querying $\mathbb{F}$,
then set $\Ucal_{i+1}=\Ucal_i$;
otherwise,
do the following:
\begin{itemize}
\item Find the max edge-weight $w_*$ of the path
connecting $u,v$ in
the MSF of $G_{i+1}$ by querying $\mathbb{F}$.
\item Find the smallest index $j_*$ of $\Ucal_i$
greater than $\max\Set{w_*,w(\fsimp{\mathcal{F}}{i})}$.
{\rm(}Note that we can store $\Ucal_i$ as
a red-black tree~\cite{CLRS3rd}, so that finding $j_*$
takes $O(\log n)$ time.{\rm)}
\item Output an interval $[j_*,i]$ and set $\Ucal_{i+1}=\Ucal_i\setminus\Set{j_*}$.
\end{itemize}
\end{description}
At the end, for each $j\in\Ucal_m$,
output an interval $[j,m]$.
\end{algr*}
Now we can see that Algorithm~\ref{alg:1-zigzag-graph-abstr} has time complexity $O(m\log^4 n)$,
where each iteration is dominated by the update of $\mathbb{F}$.
\section{Codimension-one zigzag persistence of embedded complexes}
\label{sec:alex-dual}
In this section,
we present an efficient algorithm for computing the $(p-1)$-th
barcode given a zigzag filtration of an $\mathbb{R}^p$-embedded complex,
by extending our algorithm for 0-dimension with the help of Alexander duality~\cite{munkres2018elements}.
Throughout this section,
$p\geq 2$, $K$ is a simplicial complex embedded in $\mathbb{R}^p$,
and $\mathcal{F}: \emptyset=K_0 \leftrightarrow \cdots \leftrightarrow K_m$
is a zigzag filtration of $K$.
We call connected components of $\mathbb{R}^p\setminus|K|$ as {\it voids} of $K$
or {\it $K$-voids},
to emphasize that
only voids of $K$ are considered in this section.
The {\it dual graph} $G$ of $K$ has the vertices corresponding to
the voids as well as the $p$-simplices of $K$,
and has the edges corresponding to the $(p-1)$-simplices of $K$.
The {\it dual filtration}
$\Ecal: G=G_0 \leftrightarrow G_1 \leftrightarrow \cdots \leftrightarrow G_m$
of $\mathcal{F}$ consists of subgraphs $G_i$ of $G$ such that:
({\sf i})~all vertices of $G$ dual to a $K$-void are in $G_i$;
({\sf ii})~a vertex of $G$ dual to a $p$-simplex is in $G_i$ iff
the dual $p$-simplex is {\it not} in $K_i$;
({\sf iii})~an edge of $G$ is in $G_i$ iff
its dual $(p-1)$-simplex is {\it not} in $K_i$.
One could verify that each $G_i$ is a well-defined subgraph of $G$.
We note the following: ({\sf i})~inclusion directions in $\Ecal$ are reversed;
({\sf ii})~$\Ecal$ is not exactly a zigzag filtration (because
an arrow may introduce no changes)
but can be easily made into one.
Figure~\ref{fig:df} gives an example
in $\mathbb{R}^2$,
in which we observe the following: whenever a $(p-1)$-cycle
(i.e., 1-cycle) is formed in the primal filtration,
a connected component in the dual filtration splits;
whenever a $(p-1)$-cycle
is killed in the primal filtration,
a connected component in the dual filtration vanishes.
Intuitively, $G_i$ encodes the connectivity
of $\mathbb{R}^p\setminus|K_i|$,
and so by Alexander duality,
we have the following proposition:
\begin{figure}
\centering
\includegraphics[width=0.75\linewidth]{fig/df}
\caption{A zigzag filtration of an $\mathbb{R}^2$-embedded complex $K$
and its dual filtration, where the dual graphs are colored red.
For brevity, changes of vertices in the primal filtration are ignored.}
\label{fig:df}
\end{figure}
\begin{proposition}\label{prop:equiv-barc}
$\mathsf{Pers}(\mathsf{H}_{p-1}(\mathcal{F}))=\mathsf{Pers}(\mathsf{\tilde{H}}_0(\Ecal))$.
\end{proposition}
We conduct the proof of Proposition~\ref{prop:equiv-barc}
by introducing two modules $\mathcal{M}_3$, $\mathcal{M}_4$ and proving the following:
\begin{itemize}
\item
$\mathsf{Pers}(\mathsf{H}_{p-1}(\mathcal{F}))=\mathsf{Pers}(\mathcal{M}_3)$, in Proposition~\ref{prop:equiv-barc-M3}.
\item
$\mathsf{Pers}(\mathcal{M}_3)=\mathsf{Pers}(\mathcal{M}_4)$, in Proposition~\ref{prop:equiv-barc-M3-M4}.
\item
$\mathsf{Pers}(\mathcal{M}_4)=\mathsf{Pers}(\mathsf{\tilde{H}}_0(\Ecal))$, in Proposition~\ref{prop:equiv-barc-M4}.
\end{itemize}
\begin{proposition}\label{prop:equiv-barc-M3}
The following zigzag modules are isomorphic:
\[
{\xymatrix{
\mathsf{H}_{p-1}(\mathcal{F}):\hspace{-5em}
& \mathsf{H}_{p-1}(K_0) \ar[d]^(.43){\rotatebox{90}{$\approx$}} \ar@{<->}[r]^(.52){\iota_*}
& \mathsf{H}_{p-1}(K_1) \ar[d]^(.43){\rotatebox{90}{$\approx$}} \ar@{<->}[r]^(.72){\iota_*}
& \cdots \ar@{<->}[r]^(.29){\iota_*}
& \mathsf{H}_{p-1}(K_m) \ar[d]^(.43){\rotatebox{90}{$\approx$}}\\
\mathcal{M}_2:\hspace{-2.8em}
& \mathsf{\tilde{H}}^{0}(\mathbb{R}^p\setminus|K_0|) \ar[d]^(.43){\rotatebox{90}{$\approx$}} \ar@{<->}[r]^(.52){\iota^*}
& \mathsf{\tilde{H}}^{0}(\mathbb{R}^p\setminus|K_1|) \ar[d]^(.43){\rotatebox{90}{$\approx$}} \ar@{<->}[r]^(.72){\iota^*}
& \cdots \ar@{<->}[r]^(.29){\iota^*}
& \mathsf{\tilde{H}}^{0}(\mathbb{R}^p\setminus|K_m|) \ar[d]^(.43){\rotatebox{90}{$\approx$}} \\
\mathcal{M}_3:\hspace{-2.8em}
& \mathsf{Hom}(\mathsf{\tilde{H}}_{0}(\mathbb{R}^p\setminus|K_0|)) \ar@{<->}[r]^(.52){(\iota_*)^*}
& \mathsf{Hom}(\mathsf{\tilde{H}}_{0}(\mathbb{R}^p\setminus|K_1|)) \ar@{<->}[r]^(.75){(\iota_*)^*}
& \cdots \ar@{<->}[r]^(.29){(\iota_*)^*}
& \mathsf{Hom}(\mathsf{\tilde{H}}_{0}(\mathbb{R}^p\setminus|K_m|)) \\
}}
\]
which implies that $\mathsf{Pers}(\mathsf{H}_{p-1}(\mathcal{F}))=\mathsf{Pers}(\mathcal{M}_3)$.
\end{proposition}
\begin{proof}
In the diagram, $\iota_*$ and $\iota^*$ are induced by inclusion,
$\mathsf{Hom}(\mathsf{\tilde{H}}_{0}(\mathbb{R}^p\setminus|K_i|))$ is the space of all linear maps
from $\mathsf{\tilde{H}}_{0}(\mathbb{R}^p\setminus|K_i|)$ to $\mathbb{Z}_2$,
and $(\iota_*)^*$ is the dual of $\iota_*$ (i.e., $(\iota_*)^*(g)=g\circ\iota_*$).
The isomorphism between $\mathsf{H}_{p-1}(\mathcal{F})$ and $\mathcal{M}_2$
is given by Alexander duality~\cite[Corollary~72.4]{munkres2018elements},
and the isomorphism between $\mathcal{M}_2$ and $\mathcal{M}_3$ is
given by the universal coefficient theorem~\cite[pg.~196,198]{hatcher2002algebraic}.
\end{proof}
\begin{proposition}\label{prop:equiv-barc-M3-M4}
Define an elementary zigzag module $\mathcal{M}_4$ as:
\[
\mathcal{M}_4:
\mathsf{\tilde{H}}_{0}(\mathbb{R}^p\setminus|K_0|) \leftrightarrowsp{\iota_*}
\mathsf{\tilde{H}}_{0}(\mathbb{R}^p\setminus|K_1|) \leftrightarrowsp{\iota_*}
\cdots \leftrightarrowsp{\iota_*}
\mathsf{\tilde{H}}_{0}(\mathbb{R}^p\setminus|K_m|)
\]
Then, one has that $\mathsf{Pers}(\mathcal{M}_3)=\mathsf{Pers}(\mathcal{M}_4)$.
\end{proposition}
\begin{proof}
First note that directions of corresponding arrows in $\mathcal{M}_3$ and $\mathcal{M}_4$
are reversed.
To prove the proposition,
first observe that for each linear map
$\iota_*$ in $\mathcal{M}_4$ and
its corresponding map $(\iota_*)^*$ in $\mathcal{M}_3$,
if $\iota_*$ is an isomorphism, then $(\iota_*)^*$ is an isomorphism;
if $\iota_*$ is injective with non-trivial cokernel,
then $(\iota_*)^*$ is surjective with non-trivial kernel;
if $\iota_*$ is surjective with non-trivial kernel,
then $(\iota_*)^*$ is injective with non-trivial cokernel.
Therefore, $\mathsf{P}(\mathcal{M}_3)=\mathsf{P}(\mathcal{M}_4)$ and $\mathsf{N}(\mathcal{M}_3)=\mathsf{N}(\mathcal{M}_4)$.
So intervals in $\mathsf{Pers}(\mathcal{M}_4)$ also induce a bijection
from $\mathsf{P}(\mathcal{M}_3)$ to $\mathsf{N}(\mathcal{M}_3)$,
by mapping the start of each interval to the end.
Based on Proposition~\ref{prop:pn-paring-w-rep},
in order to prove that $\mathsf{Pers}(\mathcal{M}_3)=\mathsf{Pers}(\mathcal{M}_4)$,
we only need to show that every interval in $\mathsf{Pers}(\mathcal{M}_4)$
admits a set of representatives in the module $\mathcal{M}_3$.
Let $\mathcal{M}_4=\bigoplus_{k\in\Lambda}\mathcal{I}^{[b_k,d_k]}$
be an interval decomposition of $\mathcal{M}_4$.
Moreover,
for each $k\in\Lambda$ and each $i\in[b_k,d_k]$,
let $\mathcal{I}^{[b_k,d_k]}(i)$ be the $i$-th vector space in $\mathcal{I}^{[b_k,d_k]}$
and let $\alpha^k_i$ be the non-zero element
in $\mathcal{I}^{[b_k,d_k]}(i)$.
Then, for each $k\in\Lambda$, define a set of representatives
$\bigSet{\beta_i\in\mathsf{Hom}(\mathsf{\tilde{H}}_{0}(\mathbb{R}^p\setminus|K_i|))
\,|\, i\in[b_k,d_k]}$ for $[b_k,d_k]$ in $\mathcal{M}_3$
as follows:
for each $i\in[b_k,d_k]$, let
$\mathcal{B}=\bigSet{\alpha^\ell_i\,|\, \ell\in\Lambda\text{ and }[b_\ell,d_\ell]\ni i}$
and note that $\mathcal{B}$ forms a basis of $\mathsf{\tilde{H}}_{0}(\mathbb{R}^p\setminus|K_i|)$;
then, $\beta_i$ maps
$\alpha^k_i$ to 1 and all the other elements in $\mathcal{B}$ to 0.
It can then be verified that $\Set{\beta_i\,|\, i\in[b_k,d_k]}$
forms a valid set of representatives for $[b_k,d_k]$ in $\mathcal{M}_3$.
\end{proof}
\begin{proposition}\label{prop:equiv-barc-M4}
The following zigzag modules are isomorphic:
\[
{\xymatrix{
\mathsf{\tilde{H}}_0(\Ecal):\hspace{-3.8em}
& \mathsf{\tilde{H}}_{0}(G_0) \ar[d]_(.44)\theta^(.43){\rotatebox{90}{$\approx$}} \ar@{<->}[r]^(.5){\iota_*}
& \mathsf{\tilde{H}}_{0}(G_1) \ar[d]_(.44)\theta^(.43){\rotatebox{90}{$\approx$}} \ar@{<->}[r]^(.68){\iota_*}
& \cdots \ar@{<->}[r]^(.33){\iota_*}
& \mathsf{\tilde{H}}_{0}(G_m) \ar[d]_(.43)\theta^(.44){\rotatebox{90}{$\approx$}} \\
\mathcal{M}_4:\hspace{-2.8em}
& \mathsf{\tilde{H}}_{0}(\mathbb{R}^p\setminus|K_0|) \ar@{<->}[r]^(.5){\iota_*}
& \mathsf{\tilde{H}}_{0}(\mathbb{R}^p\setminus|K_1|) \ar@{<->}[r]^(.68){\iota_*}
& \cdots \ar@{<->}[r]^(.33){\iota_*}
& \mathsf{\tilde{H}}_{0}(\mathbb{R}^p\setminus|K_m|) \\
}}
\]
which implies that $\mathsf{Pers}(\mathcal{M}_4)=\mathsf{Pers}(\mathsf{\tilde{H}}_0(\Ecal))$.
\end{proposition}
\begin{proof}
In the diagram,
$\theta$ is induced by a map $\tilde{\theta}$ which
maps a vertex of $G_i$ to a point in the corresponding $K$-void or $p$-simplex
contained in $\mathbb{R}^p\setminus|K_i|$,
so the commutativity of the diagram can be easily verified.
To see that $\theta$ is an isomorphism, we observe the following facts,
which can be verified by induction:
\begin{itemize}
\item Any point in $\mathbb{R}^p\setminus|K_i|$ is path connected to a point
in a $K$-void or a $p$-simplex contained in $\mathbb{R}^p\setminus|K_i|$.
\item For any two vertices $v_1,v_2$ of $G_i$, $v_1$ is path connected to $v_2$
in $G_i$ if and only if $\tilde{\theta}(v_1)$ is path connected to
$\tilde{\theta}(v_2)$ in $\mathbb{R}^p\setminus|K_i|$.
\qedhere
\end{itemize}
\end{proof}
Proposition~\ref{prop:reduced-barcode}
indicates a way to compute $\mathsf{Pers}(\mathsf{\tilde{H}}_0(\Ecal))$:
\begin{proposition}\label{prop:reduced-barcode}
$\mathsf{Pers}(\mathsf{\tilde{H}}_0(\Ecal))=\mathsf{Pers}(\mathsf{H}_0(\Ecal))\setminus\Set{[0,m]}$.
\end{proposition}
\begin{proof}
We first note that $\mathsf{H}_0(\Ecal)$ is not an elementary module
because it does not start with the trivial vector space,
whereas most of our definitions and conclusions assume elementary ones.
However, we can add a trivial vector space at the beginning of $\mathsf{H}_0(\Ecal)$
to make it elementary so that
those constructions remain valid.
We then show that $[0,m]\in\mathsf{Pers}(\mathsf{H}_0(\Ecal))$ by simulating
Algorithm~\ref{alg:0-zigzag} on $\Ecal$.
However, $\Ecal$ does not start with an empty complex
as assumed by Algorithm~\ref{alg:0-zigzag}.
Hence, the lowest node of $\mathbb{G}_\mathrm{B}(\Ecal)$ is at level $0$
instead of $1$; the latter is the case for a filtration
starting with an empty complex.
With this minor adjustment, Algorithm~\ref{alg:0-zigzag}
can still be applied on $\mathbb{G}_\mathrm{B}(\Ecal)$.
Let $T_i$ be the barcode forest in iteration $i$ of
Algorithm~\ref{alg:0-zigzag} when taking $\Ecal$ as input.
Then, it can be proved by induction that for every $i$,
$T_i$ contains
a tree $\Upsilon_i$ with a level-0 root and all the connected components
of $G_i$, which contain vertices dual to the $K$-voids,
correspond to leaves in $\Upsilon_i$.
Therefore, Algorithm~\ref{alg:0-zigzag} must produce an interval $[0,m]$
at the end.
Note that $\mathsf{P}(\mathsf{\tilde{H}}_0({\mathcal{E}}))=\mathsf{P}(\mathsf{H}_0({\mathcal{E}}))\setminus\Set{0}$
and $\mathsf{N}(\mathsf{\tilde{H}}_0(\Ecal))=\mathsf{N}(\mathsf{H}_0(\Ecal))\setminus\Set{m}$,
so $\mathsf{Pers}(\mathsf{H}_0(\Ecal))\setminus\Set{[0,m]}$ induces
a bijection from $\mathsf{P}(\mathsf{\tilde{H}}_0(\Ecal))$
to $\mathsf{N}(\mathsf{\tilde{H}}_0(\Ecal))$.
By Proposition~\ref{prop:pn-paring-w-rep}, we only need to show that
every interval in $\mathsf{Pers}(\mathsf{H}_0(\Ecal))\setminus\Set{[0,m]}$
admits representatives in the module $\mathsf{Pers}(\mathsf{\tilde{H}}_0(\Ecal))$.
Let $\Set{\alpha_i\,|\, i\in[s,t]}$ be an arbitrary set of representatives for
an interval $[s,t]\subseteq[0,m]$
in the module $\mathsf{Pers}(\mathsf{H}_0(\Ecal))$; we note the following facts:
\begin{itemize}
\item The parity of the number of connected components in
each $\alpha_i$ for $i\in[s,t]$ is the same.
\item A 0-th homology class is in the reduced homology group
if it consists of an even number of connected components.
\end{itemize}
For an interval $[b,d]$ in $\mathsf{Pers}(\mathsf{H}_0(\Ecal))\setminus\Set{[0,m]}$,
let $\Set{\beta_i\,|\, i\in[b,d]}$ be a set of representatives for $[b,d]$
in the module $\mathsf{Pers}(\mathsf{H}_0(\Ecal))$.
If $\psi^0_{b-1}$ is backward, where $\psi^0_{b-1}$ is the linear map
in $\mathsf{H}_0(\Ecal)$ connecting $\mathsf{H}_0(G_{b-1})$ and $\mathsf{H}_0(G_{b})$,
then $\psi^0_{b-1}(\beta_b)=0$ by definition.
It follows that $\beta_b$ consists of
an even number of connected components.
Therefore,
$\Set{\beta_i\,|\, i\in[b,d]}$ is also a set of representatives for $[b,d]$
in the module $\mathsf{Pers}(\mathsf{\tilde{H}}_0(\Ecal))$ by the noted facts.
Similar situations happen when $d<m$ and $\psi^0_{d}$ is forward.
Now suppose that $\psi^0_{b-1}$ is forward
and either $d=m$ or $\psi^0_{d}$ is backward.
If $\beta_b$ consists of an even number of connected components,
then $\Set{\beta_i\,|\, i\in[b,d]}$ is also a set of representatives for $[b,d]$
in the module $\mathsf{Pers}(\mathsf{\tilde{H}}_0(\Ecal))$.
If $\beta_b$ consists of an odd number of connected components,
let $\Set{\alpha_i\,|\, i\in[0,m]}$ be
a set of representatives for the interval $[0,m]$
in the module $\mathsf{Pers}(\mathsf{H}_0(\Ecal))$.
Then,
$\Set{\beta_i+\alpha_i\,|\, i\in[b,d]}$ is a set of representatives for $[b,d]$
in the module $\mathsf{Pers}(\mathsf{\tilde{H}}_0(\Ecal))$
because $\alpha_0$ consists of a single connected component.
\end{proof}
The above two propositions
suggest a naive algorithm for computing
$\mathsf{Pers}(\mathsf{H}_{p-1}(\mathcal{F}))$
using Algorithm~\ref{alg:0-zigzag}.
However,
building the dual graph $G$
requires reconstructing the void boundaries of $K$,
which is done by
a ``walking'' algorithm obtaining a set of $(p-1)$-cycles
and then by a nesting test of these $(p-1)$-cycles~\cite[Section 4.1]{dey2020computing}.
The running time of this process is $\Omega(n^2)$ where $n$
is the size of $K$.
To achieve the claimed complexity,
we first define the following~\cite{dey2020computing}:
\begin{definition}\label{dfn:q-conn}
In a simplicial complex $X$,
two $q$-simplices $\sigma,\sigma'$
are \textbf{$q$-connected}
if there is a sequence $\sigma_0,\ldots,\sigma_\ell$ of $q$-simplices of $X$
such that $\sigma_0=\sigma$, $\sigma_\ell=\sigma'$,
and every $\sigma_i,\sigma_{i+1}$ share a $(q-1)$-face.
A maximal set of $q$-connected $q$-simplices of $X$
is called a \textbf{$q$-connected component} of $X$,
and $X$ is \textbf{$q$-connected}
if it has only one $q$-connected component.
\end{definition}
Based on the fact that void boundaries
can be reconstructed without the nesting test
for $(p-1)$-connected complexes in $\mathbb{R}^p$~\cite{dey2020computing},
we restrict $\mathcal{F}$ to several $(p-1)$-connected subcomplexes of $K$
and then take the union of
the $(p-1)$-th barcodes of these restricted filtrations:
\begin{algr}[Algorithm for $(p-1)$-th zigzag persistence on $\mathbb{R}^p$-embedded complexes]
\label{alg:codim-1-zigzag}
\hspace{3em}
\begin{enumerate}
\item Compute the $(p-1)$-connected components $\comp{1},\ldots,\comp{r}$
of $K$.
\item For each $\ell=1,\ldots,r$, let
\[\compcmplx{\ell}=\cl{\comp{\ell}}\union\Set{\tau\in K\,|\,
\tau\text{ is a }p\text{-simplex whose }(p-1)\text{-faces are in }\comp{\ell}}\]
and let $\rfilt{\ell}$ be a filtration of $\compcmplx{\ell}$ defined as
\[\rfilt{\ell}:K_0\intersect\compcmplx{\ell} \leftrightarrow
K_1\intersect\compcmplx{\ell} \leftrightarrow \cdots \leftrightarrow K_m\intersect\compcmplx{\ell}\]
Then, compute $\mathsf{Pers}(\mathsf{H}_{p-1}(\rfilt{\ell}))$ for each $\ell$.
\item Return $\bigunion_{\ell=1}^r\mathsf{Pers}(\mathsf{H}_{p-1}(\rfilt{\ell}))$
as the $(p-1)$-th barcode of $\mathcal{F}$.
\end{enumerate}
\end{algr}
In Algorithm~\ref{alg:codim-1-zigzag},
$\cl{\comp{\ell}}$ denotes the closure of $\comp{\ell}$, i.e.,
the complex consisting of all faces of simplices in $\comp{\ell}$.
For each $\ell$,
let $n_\ell$ be the number of simplices in $\compcmplx{\ell}$;
then, the dual graph of $\compcmplx{\ell}$
can be constructed in $O(n_\ell\log n_\ell)$
time because $\compcmplx{\ell}$ is $(p-1)$-connected~\cite{dey2020computing}.
Using Algorithm~\ref{alg:0-zigzag}
to compute $\mathsf{Pers}(\mathsf{H}_{p-1}(\rfilt{\ell}))$
as suggested by the duality,
the running time of Algorithm~\ref{alg:codim-1-zigzag}
is $O(m\log^2 n+m\log m+n\log n)$,
where $m$ is the length of $\mathcal{F}$ and $n$ is the size of $K$.
The correctness of Algorithm~\ref{alg:codim-1-zigzag} can be seen from the following proposition:
\begin{proposition}\label{prop:codim-1-mod-sum}
The modules $\mathsf{H}_{p-1}(\mathcal{F})$ and $\bigoplus_{\ell=1}^r\mathsf{H}_{p-1}(\rfilt{\ell})$
are isomorphic
which implies that
\[\mathsf{Pers}(\mathsf{H}_{p-1}(\mathcal{F}))=\bigunion_{\ell=1}^r\mathsf{Pers}(\mathsf{H}_{p-1}(\rfilt{\ell}))\]
\end{proposition}
To prove Proposition~\ref{prop:codim-1-mod-sum},
we define the following notations:
for a simplicial complex $X$, $\DG_{q}(X)$ denotes the set of $q$-simplices
of $X$; also, for a $q$-chain $A$ of $X$
and a $k$-simplex $\sigma$ of $X$ with $k<q$,
$\mathsf{cof}_{q}(A,\sigma)$ denotes the set of $q$-cofaces of $\sigma$ belonging to $A$.
We first prove the following proposition:
\begin{proposition}\label{prop:dirsum-homo}
Let $q\geq2$ and
$X$ be a simplicial complex with subcomplexes $X_1$, $X_2$, and $X'$.
If $\DG_{q-1}(X_1)\union\DG_{q-1}(X_2)=\DG_{q-1}(X)$,
$\DG_{q}(X_1)\union\DG_{q}(X_2)=\DG_{q}(X)$,
and no $(q-1)$-simplex of $X_1$ is $(q-1)$-connected to a
$(q-1)$-simplex of $X_2$,
then the following diagram commutes:
\[
{\xymatrix{
\mathsf{H}_{q-1}(X_1\intsec X')\oplus\mathsf{H}_{q-1}(X_2\intsec X')
\ar[d]_(.44){\theta_1}^(.43){\rotatebox{90}{$\approx$}} \ar@{<->}[r]^(.585){\iota_*}
& \mathsf{H}_{q-1}(X_1)\oplus\mathsf{H}_{q-1}(X_2)
\ar[d]_(.44){\theta_2}^(.43){\rotatebox{90}{$\approx$}} \\
\mathsf{H}_{q-1}(X') \ar@{<->}[r]^(.585){\iota_*}
& \mathsf{H}_{q-1}(X) \\
}}
\]
where $\theta_1$ and $\theta_2$ are isomorphisms, the upper $\iota_*$ is the direct sum
of the linear maps induced by inclusions, and the lower $\iota_*$ is induced by
the inclusion.
\end{proposition}
\begin{remark}\label{rmk:ext2gen}
The above proposition can be easily extended to the general case with
a finite number of subcomplexes $X_1,X_2,\ldots,X_k$ of $X$ such that
$\bigunion_{\ell=1}^k\DG_{q-1}(X_\ell)=\DG_{q-1}(X)$,
$\bigunion_{\ell=1}^k\DG_{q}(X_\ell)=\DG_{q}(X)$,
and $(q-1)$-simplices of each pair of $X_1,\ldots,X_k$ are not $(q-1)$-connected.
\end{remark}
\begin{proof}
In the diagram, $\theta_2$ is defined as follows ($\theta_1$ is similarly defined):
for each $\big([z_1],[z_2]\big)\in\mathsf{H}_{q-1}(X_1)\oplus\mathsf{H}_{q-1}(X_2)$,
$\theta_2\big([z_1],[z_2]\big)=[z_1+z_2]$.
It is not hard to see that
the diagram is commutative and $\theta_1$, $\theta_2$ are linear maps.
Therefore, we only need to show that $\theta_1$, $\theta_2$ are isomorphisms.
We only do this for $\theta_2$ because the verification for $\theta_1$ is similar.
For the surjectivity of $\theta_2$, let $[z]\in\mathsf{H}_{q-1}(X)$ be arbitrary and
let $z_1=z\intsec X_1$, $z_2=z\intsec X_2$.
We claim that $\partial(z_1)=\partial(z_2)=0$, which implies that
$\theta_2\big([z_1],[z_2]\big)=[z]$ and hence the surjectivity.
We only show that $\partial(z_1)=0$ because for $\partial(z_2)$
it is similar. For contradiction, suppose instead that there is a $(q-2)$-simplex
$\sigma\in\partial(z_1)$. Then, $|\mathsf{cof}_{q-1}(z_1,\sigma)|$ is an odd number.
Note that $|\mathsf{cof}_{q-1}(z,\sigma)|$ is an even number
and $\mathsf{cof}_{q-1}(z_1,\sigma)\subseteq\mathsf{cof}_{q-1}(z,\sigma)$,
so $\mathsf{cof}_{q-1}(z,\sigma)\setminus\mathsf{cof}_{q-1}(z_1,\sigma)$ is not empty.
Let $\tau\in\mathsf{cof}_{q-1}(z,\sigma)\setminus\mathsf{cof}_{q-1}(z_1,\sigma)$;
then $\tau\not\in X_1$ because if $\tau\in X_1$,
$\tau$ must belong to $z_1$ and hence belongs to $\mathsf{cof}_{q-1}(z_1,\sigma)$.
So $\tau\in X_2$.
Note that $\tau$ must be $(q-1)$-connected to a $\tau'\in\mathsf{cof}_{q-1}(z_1,\sigma)$
which is in $X_1$ because they share a common $(q-2)$-face $\sigma$.
But this contradicts the fact that
no $(q-1)$-simplex of $X_1$ is $(q-1)$-connected to a
$(q-1)$-simplex of $X_2$,
and so $\partial(z_1)=0$.
For the injectivity of $\theta_2$, let $\big([z_1],[z_2]\big)$ be
any element of $\mathsf{H}_{q-1}(X_1)\oplus\mathsf{H}_{q-1}(X_2)$
such that $\theta_2\big([z_1],[z_2]\big)=[z_1+z_2]=0$.
Because $z_1+z_2$ is a boundary in $X$, let $A$ be the $q$-chain
in $X$ such that $\partial(A)=z_1+z_2$. Moreover,
let $A_1=A\intsec X_1$ and $A_2=A\intsec X_2$.
We claim that $z_1=\partial(A_1)$ and $z_2=\partial(A_2)$,
which implies that $\big([z_1],[z_2]\big)=(0,0)$ and hence the injectivity.
We only show that $z_1=\partial(A_1)$ because the other is similar:
\begin{description}
\item[$z_1\subseteq\partial(A_1)$:]
For any $(q-1)$-simplex $\sigma\in z_1$, because $\sigma\in X_1$
and no $(q-1)$-simplex of $X_1$ is $(q-1)$-connected to a
$(q-1)$-simplex of $X_2$, $\sigma\not\in X_2$.
So $\sigma\not\in z_2$, which means that $\sigma\in X_1+X_2=\partial(A)$.
We claim that $\mathsf{cof}_q(A,\sigma)=\mathsf{cof}_q(A_1,\sigma)$.
Then,
the fact that $|\mathsf{cof}_q(A,\sigma)|$ is an odd number implies that
$|\mathsf{cof}_q(A_1,\sigma)|$ is an odd number, and hence $\sigma\in\partial(A_1)$.
To prove that $\mathsf{cof}_q(A,\sigma)=\mathsf{cof}_q(A_1,\sigma)$,
first note that $\mathsf{cof}_q(A_1,\sigma)\subseteq\mathsf{cof}_q(A,\sigma)$.
We then show that $\mathsf{cof}_q(A,\sigma)\subseteq\mathsf{cof}_q(A_1,\sigma)$.
Let $\tau$ be any $q$-simplex in $\mathsf{cof}_q(A,\sigma)$.
We must have that $\tau\in X_1$ because otherwise $\tau\in X_2$
which implies that $\sigma\in X_2$, a contradiction.
It follows that $\tau\in A_1=A\intsec X_1$
and hence $\tau\in\mathsf{cof}_q(A_1,\sigma)$.
\item[$\partial(A_1)\subseteq z_1$:]
The proof is similar to the previous one and is omitted.
\qedhere
\end{description}
\end{proof}
Now we prove Proposition~\ref{prop:codim-1-mod-sum}.
First note that for each $i$,
$\bigunion_{\ell=1}^r\DG_{p-1}(K_{i+1}\intsec\compcmplx{\ell})
=\DG_{p-1}(K_{i+1})$
and $\bigunion_{\ell=1}^r\DG_{p}(K_{i+1}\intsec\compcmplx{\ell})
=\DG_{p}(K_{i+1})$.
We also have that $K_i\subseteq K_{i+1}$
and $(K_{i+1}\intsec\compcmplx{\ell})\intersect K_i=K_{i}\intsec\compcmplx{\ell}$
for each $\ell$.
By Proposition~\ref{prop:dirsum-homo} and Remark~\ref{rmk:ext2gen},
we have the following commutative diagram:
\[
{\xymatrix{
\bigoplus_{\ell=1}^r\mathsf{H}_{p-1}(K_i\intsec\compcmplx{\ell})
\ar[d]^(.43){\rotatebox{90}{$\approx$}} \ar@{<->}[r
& \bigoplus_{\ell=1}^r\mathsf{H}_{p-1}(K_{i+1}\intsec\compcmplx{\ell})
\ar[d]^(.43){\rotatebox{90}{$\approx$}}
\\
\mathsf{H}_{p-1}(K_i) \ar@{<->}[r]^(.5){\morph{\mathcal{F}}{i}{p-1}}
& \mathsf{H}_{p-1}(K_{i+1})
\\
}}
\]
and therefore Proposition~\ref{prop:codim-1-mod-sum} follows.
\section{Conclusions}
In this paper, we propose near-linear algorithms for computing zigzag persistence
on graphs with the help of some dynamic graph data structures.
The algorithm for $0$-dimensional homology relates the computation to
an algorithm in \cite{agarwal2006extreme} for
pairing critical points of Morse functions on $2$-manifolds,
thereby giving a correctness proof for the algorithm in~\cite{agarwal2006extreme}.
The algorithm for $1$-dimensional homology reduces the computation to
finding the earliest positive edge so that
a $1$-cycle containing both the positive and negative edges
resides in all intermediate graphs.
With the help of Alexander duality,
we also extend the algorithm for $0$-dimension
to compute the $(p-1)$-dimensional zigzag persistence for $\mathbb{R}^p$-embedded
complexes in near-linear time.
An obvious open question is
whether the time complexities can be improved.
Since the running time of our algorithms is determined in part
by the data structures we use,
one may naturally ask whether these data structure
can be improved.
Furthermore, we note that the following may be helpful to future research efforts:
\begin{enumerate}
\item In our algorithm for $0$-dimension, we build the barcode graph
and utilize the algorithm in~\cite{agarwal2006extreme} to compute
the zigzag barcode, which is also adopted by~\cite{dey2019computing}.
Is there any other scenario in persistence computing
where such a technique can be applied so that a more efficient algorithm
can be derived?
\item In our algorithm for $1$-dimension,
we utilize the dynamic-MSF
data structure~\cite{holm2001poly} for computing the max edge-weight
of the path connecting two vertices in an MSF.
The update of the
data structure which takes $O(\log^4 n)$ amortized time becomes the bottleneck of the
algorithm.
However,
for the computation,
it can be verified that we only need to know the {\it minimax}\footnote{The {\it minimax} distance
of two vertices in a graph is the minimum of the max edge-weights
of all paths connecting the two vertices.}
distance of two vertices in a graph.
Is there any faster way to compute the minimax distance in a dynamic graph?
\item Another interesting question is whether our algorithms are more efficient
practically when implemented
compared to some existing implementation~\cite{maria2014zigzag,Dionysus}
of zigzag persistence algorithms for general dimension.
\end{enumerate}
|
\section{Preliminaries and main results}
Whereas the regular Poisson process can be generalized in a straight-forward manner to a non-homogeneous Poisson process with a nicely behaved time-dependent intensity function by a mere time change using the chain rule, this trick does not work for fractional derivatives (e.g. \cite{Tarasov2016}). The original fractional Poisson process (fPp) as firstly introduced by \cite{Laskin2003} cannot be generalized using his approach. A possible way to go forth is following \cite{Scalas2017} who introduced the marginals for a fractional non-homogeneus Poisson process (fnhPp) by equation Eq.(2.18) in their paper
\begin{eqnarray}
P_{\beta}(n,t) = \int_{0}^{\infty} du~ \frac{ e^{-\Lambda(u)} \Lambda(u)^n}{n!} h_{\beta}(t,u)
\label{eq:non-hom_fPp}
\end{eqnarray}
where $0<\beta<1$ and $\Lambda(t)=\int_{0}^{t} du~ \lambda(u)$ with an intensity function $\lambda(\cdot):[0,\infty)\rightarrow [0,\infty)$ and the density of the inverse stable subordinator (e.g. (\cite{Meerschaert2013})
\begin{eqnarray}
h_{\beta}(t,x)=\frac{t}{\beta x^{1+1/\beta}}~ g_{\beta}\left( \frac{t}{x^{1/\beta}}\right)
\label{eq:h_beta}
\end{eqnarray}
and
\begin{eqnarray}
g_{\beta}(u)=\frac{1}{\pi} \sum_{j=1}^{\infty} (-1)^{j+1}\frac{\Gamma(\beta j+1)}{\Gamma(j+1)}\frac{1}{u^{\beta j +1}}\sin(\beta \pi j)
\label{eq:g_beta}
\end{eqnarray}
The authors in \cite{Scalas2017} show in their Corollary 1 that the marginals Eq.(\ref{eq:non-hom_fPp}) satisfy the following system of fractional differential-integral equations
\begin{eqnarray}
\frac{d^{\beta}}{dt^{\beta}} P_{\beta}(n,t) & = &\int_{0}^{t} du ~\lambda(u) \left( -\frac{ e^{-\Lambda(u)} \Lambda(u)^n}{n!}+\frac{ e^{-\Lambda(u)} \Lambda(u)^{n-1}}{(n-1)!}\right) h_{\beta}(t,u),
\hspace{1cm} n= 1,2,3,...
\nonumber
\\
\label{eq:nhfPp_eq}
\\
\frac{d^{\beta}}{dt^{\beta}} P_{\beta}(0,t)~ & = &~ -\int_{0}^{t} du ~\lambda(u) e^{-\Lambda(u)} h_{\beta}(t,u) \hspace{5.4cm} n=0
\nonumber
\end{eqnarray}
with initial conditions $P_{\beta}(0,0)=1$ and $P_{\beta}(n,0)=0$ for $n=1,2,...$. The fractional derivative is the fractional Caputo-Djrbashian derivative which is defined as (e.g. \cite{Podlubni1999})
\begin{eqnarray}
\frac{d^{\beta}}{dt^{\beta}} F(t) = \frac{1}{\Gamma(1-\beta)} \int_0^t \frac{ds}{(t-s)^{\beta}}~\frac{d}{ds}f(s)
\nonumber
\end{eqnarray}
As expected these equations Eq.(\ref{eq:nhfPp_eq}) reduce to the usual fractional Poisson process with Kolmogorov-Feller equations as discussed by \cite{BeghinOrsingher2009} (see also \cite{Meerschaert2011} for a general theory, and \cite{Kreer2014} with an emphasis of their relation to Smoluchowsky-type coagulation-fragmentation equations)
when $\lambda(t)$ is chosen to be just a positive constant $\lambda$,
\begin{eqnarray}
\frac{d^{\beta}}{dt^{\beta}} P_{\beta}(n,t) = \lambda \left( P_{\beta}(n-1,t)-P_{\beta}(n,t)\right)
\label{eq:fPp_eq}
\end{eqnarray}
where we have $0<\beta<1$ and $P_{\beta}(-1,t)=0$ for convenience and the same initial conditions as above. This fact proven by \cite{Scalas2017} will be demonstrated independently as we proceed with proving our main theorem.
\begin{thm} Let the function $\Lambda(\cdot)$ satisfy the following conditions
(i) $\lim_{x\rightarrow\infty} \Lambda(x) =\infty$ and
(ii) There exists some positive constant $c$ such that as $x\rightarrow \infty $
\begin{eqnarray}
\frac{\Lambda'(x)}{\Lambda(x)} = \frac{c}{x} + \mathcal{O}\left(\frac{1}{x\log{x}}\right)
\nonumber
\end{eqnarray}
Then for $0<\beta<1$ the marginal probabilities of the fnhPp as given in Eq.(\ref{eq:non-hom_fPp}) have the following limiting behaviour
\begin{eqnarray}
\lim_{n\rightarrow\infty, n=\Lambda(z_0 t^{\beta})} n P_{\beta}(n,t) =\frac{z_0}{c} M_{\beta}(z_0)
\nonumber
\end{eqnarray}
where for any positive $z_0>0$ we relate the positive $n$ to any positive $t$ by the equation $n=\Lambda(z_0 t^{\beta})$, and $M_{\beta}(\cdot)$ is the M-Wright function or Mainardi function defined by
\begin{eqnarray}
M_{\beta}(z) =\frac{1}{\pi} \sum_{j=1}^{\infty} \frac{(-z)^{j-1}}{(j-1)!}\Gamma(\beta j)\sin{(\pi \beta j)}
\nonumber
\end{eqnarray}
\label{thm2}
\end{thm}
From Theorem \ref{thm2} we obtain immeadiately a useful
\begin{cor} For any positive number $r$ let $\Lambda(x)= x^r$. Then for $0<\beta<1$ the marginal probabilities as given in Eq.(\ref{eq:non-hom_fPp}) have the following limiting behaviour
\begin{eqnarray}
\lim_{t\rightarrow\infty, n t^{-r\beta}=z_0^r } t^{r\beta} P_{\beta}(n,t) =\frac{z_0^{1-r}}{r} M_{\beta}(z_0)
\nonumber
\end{eqnarray}
for any $z_0>0$ relating $t$ and $n$ by the condition $n t^{-r\beta}=z_0^{r}$. In particular for $r=1$ we obtain the dynamical scaling result for the standard fPp satisfying the Kolmogorov-Feller equations Eq. (\ref{eq:fPp_eq}) with $\lambda=1$.
\label{cor3}
\end{cor}
Note that for $\beta=1$ and $\lambda(t)=1$, when we recover the standard Poisson process, the function $M_{\beta}(z)=0$ and our proof fails in this case.
\\
\section{Proof of Theorem \ref{thm2}}
We first show that the representation for the marginals as given in Eq.(\ref{eq:non-hom_fPp}) together with Eq.(\ref{eq:h_beta})-(\ref{eq:g_beta}) by \cite{Scalas2017} can be expressed in terms of the M-Wright function (Mainardi function). Setting in Eq.(\ref{eq:non-hom_fPp}) for convenience $u=z t^{\beta}$ we get
\begin{eqnarray}
P_{\beta}(n,t) = t^{\beta}\int_{0}^{\infty} dz~ \frac{ e^{-\Lambda(z t^{\beta})} \Lambda(z t^{\beta})^n}{n!} h_{\beta}(t,z t^{\beta})
\nonumber
\end{eqnarray}
Because of Eq.(\ref{eq:h_beta})-(\ref{eq:g_beta}) we obtain after some straight forward computations
\begin{eqnarray}
t^{\beta} h_{\beta}(t,z t^{\beta})
& = & \frac{1}{\beta z^{1+1/\beta}}~ g_{\beta}\left( \frac{1}{z^{1/\beta}}\right) \nonumber \\
& = & \frac{1}{\beta z^{1+1/\beta}}~ \frac{1}{\pi} \sum_{j=1}^{\infty} (-1)^{j+1}\frac{\Gamma(\beta j+1)}{\Gamma(j+1)} \frac{1}{\left(\frac{1}{z^{1/\beta}}\right)^{\beta j +1}}\sin(\beta \pi j)
\nonumber \\
& = & \frac{1}{\pi} \sum_{j=1}^{\infty} (-1)^{j-1}\frac{\Gamma(\beta j)}{\Gamma(j)} z^{j-1} \sin(\beta \pi j) =M_{\beta}(z)
\nonumber
\end{eqnarray}
The last equality follows from the series expansion of the M-Wright function (e.g. Eq. (10) in \cite{Mainardi2020})
Therefore we have as a backbone of our investigation the identity
\begin{eqnarray}
P_{\beta}(n,t) = \int_{0}^{\infty} dz~ \frac{ e^{-\Lambda(z t^{\beta})} \Lambda(z t^{\beta})^n}{n!} M_{\beta}(z)
\label{eq:non-hom_fPp1}
\end{eqnarray}
Note that when $\Lambda(\cdot)$ is a linear function we obtain a useful Lemma for fPp firstly stated by \cite{GorenfloMainardi2012}:
\begin{lem} (Gorenflo and Mainardi) For the fractional Poisson probabilities $P_{\beta}(n,t)$ we have
the integral representation (the subordination integral)
\begin{eqnarray}
P_{\beta}(n,t) & = &\frac{1}{n!}\int_{0}^{\infty} dt_{*} t_{*}^n ~ t^{-\beta} \exp{(-t_{*})} M_{\beta}(t_{*}t^{-\beta})
\nonumber \\
& = & \frac{1}{n!} t^{n \beta} \int_{0}^{\infty} dz ~ z^n e^{-z t^{\beta}} M_{\beta}(z)
\label{eq:int_repres}
\end{eqnarray}
where we have set for convenience $z=t_{*}t^{-\beta}$ as new integration variable.
\label{thm3}
\end{lem}
Thanks to Lemma 3 and its Eq.(\ref{eq:int_repres}), following as a special case from Eq.(\ref{eq:non-hom_fPp1}), we see immediately that without loss of generality if $\lambda(t)=1$ and consequently $\Lambda(t)=t$ is linear in $t$, the integration in the fractional differential-integral equations Eq.(\ref{eq:nhfPp_eq}) on the right-hand side can be carried out explicitly yielding the term $\left( P_{\beta}(n-1,t)-P_{\beta}(n,t)\right)$ and we have thus recovered the fractional Kolmogorov-Feller equations Eq.(\ref{eq:fPp_eq}) from \cite{BeghinOrsingher2009} as claimed. Therefore our dynamical scaling result covers also the standard fPp. We finally want to emphasize the fact that Lemma 3 is of highest practical relevance in computing the marginal probabilities of the fPp as functions of time for applications in statistics or physics.
{\em Proof of Theorem \ref{thm2}}
We write Eq.(\ref{eq:non-hom_fPp1}) as follows
\begin{eqnarray}
P_{\beta}(n,t) = \frac{1}{n!} \int_{0}^{\infty} dz~ e^{n f(z|t)} M_{\beta}(z)
\nonumber
\end{eqnarray}
Here we have defined an auxilary function by
\begin{eqnarray}
f(z|t)= \log{\Lambda(zt^{\beta})}-\frac{1}{n} \Lambda(z t^{\beta})
\label{eq:f_funct}
\end{eqnarray}
and we treat $t$ as a parameter.
We compute now the first and second derivative of $f(\cdot|t)$ with respect to the first argument, namely
\begin{eqnarray}
f'(z|t)
= \Lambda'(z t^{\beta}) t^{\beta}\left[\frac{1}{\Lambda(z t^{\beta})}-\frac{1}{n}\right]
\label{eq:f_deriv1}
\end{eqnarray}
and
\begin{eqnarray}
f''(z|t) = \Lambda''(z t^{\beta}) t^{2\beta}\left[\frac{1}{\Lambda(z t^{\beta})}-\frac{1}{n}\right]
-\left( \frac{\Lambda'(z t^{\beta})t^{\beta}}{\Lambda(z t^{\beta})} \right)^2
\label{eq:f_deriv2}
\end{eqnarray}
Due to the properties of $\Lambda(\cdot)$, condition (i), there exists for any positive $n$ and for any $z_0>0$ a positive number $t$ solving , $f'(z_0|t)=0$, yielding from Eq.(\ref{eq:f_deriv1})
\begin{eqnarray}
n = \Lambda(z_0 t^{\beta})
\label{eq:n_Lambda}
\end{eqnarray}
Note that from Eq.(\ref{eq:n_Lambda}) we have with Eq. (\ref{eq:f_funct}) that $f(z_0|t)=\log{\Lambda(z_0 t^{\beta})}-\Lambda(z_0 t^{\Lambda})/n = \log{n} -1$.
And we see from Eq.(\ref{eq:f_deriv2}) that $f''(z_0|t)<0$ and thus $z_0>0$ is a positive maximum of $f(\cdot|t)$, i.e. using condition (ii) on $\Lambda(\cdot)$ we get
\begin{eqnarray}
\sqrt{-f''(z_0|t)} = \frac{\Lambda'(z_0 t^{\beta}) t^{\beta}}{\Lambda(z_0 t^{\beta})} = \frac{c}{z_0 t^{\beta}} t^{\beta} + t^{\beta}\mathcal{O}\left(\frac{1}{t^{\beta}\log{t}}\right)
\end{eqnarray}
Because the M-Wright function $M_{\beta}(\cdot)$ is analytic on the real line (e.g. \cite{GorenfloLuchenkoMainardi1999}, \cite{Podlubni1999}) we can evaluate the integral using the method of steepest decent because the conditions for applying \cite{Small2010} Chapter 6 Proposition 2 and Proposition 3 are satisfied:
\begin{eqnarray}
\int_0^{\infty} dz ~ e^{n f(z|t) }M_{\beta}(z) & = & \frac{\exp{(n f(z_0|t))}}{\sqrt{n}} \left(M_{\beta}(z_0)~\sqrt{\frac{2\pi}{-f''(z_0|t)}} +\mathcal{O}\left(\frac{1}{n}\right)\right)
\nonumber \\
& = &
\frac{e^{n f(z_0|t)}}{\sqrt{n}}\sqrt{2\pi} \left(M_{\beta}(z_0)~\frac{z_0}{c} +\mathcal{O}\left(\frac{1}{n}\right)+\mathcal{O}\left(\frac{1}{\log{t}}\right)\right)
\nonumber
\end{eqnarray}
and with Stirling's formula
\begin{eqnarray}
n! = \sqrt{2 \pi n} \left(\frac{n}{e}\right)^n \left(1+\mathcal{O}\left(\frac{1}{n}\right)\right)
\nonumber
\end{eqnarray}
and Eq.(\ref{eq:n_Lambda}) we obtain altogether
\begin{eqnarray}
\frac{1}{n!} \int_0^{\infty} dz ~ e^{n f(z|t) }M_{\beta}(z)
& = & \frac{1}{\sqrt{2 \pi n}} \left(\frac{n}{e}\right)^{-n}\left(1+\mathcal{O}\left(\frac{1}{n}\right)\right) \cdot
\nonumber \\
& & \cdot
e^{n f(z_0|t)} \sqrt{\frac{2\pi}{n}} \left(M_{\beta}(z_0)~\frac{z_0}{c} +\mathcal{O}\left(\frac{1}{n}\right)+\mathcal{O}\left(\frac{1}{\log{t}}\right)\right)
\nonumber \\
& = & \frac{1}{n
\left(M_{\beta}(z_0)~\frac{z_0}{c} +\mathcal{O}\left(\frac{1}{n}\right)+\mathcal{O}\left(\frac{1}{\log{t}}\right)\right)
\nonumber
\end{eqnarray}
where we have used $f(z_0|t)= \log{n} -1$ in the last line. Multiply both sides with $n$ and note that for large increasing $n$ by the properties of $\Lambda(\cdot)$ also $t$ increases and $z_0>0$ remains a maximum as long as $n$ and $t$ are related by Eq. (\ref{eq:n_Lambda}). This concludes the proof.
$\blacksquare$
\section{Outlook and application}
Since \cite{ViscekFamily1984} the dynamical scaling hypothesis became a ``regular'' feature in the physical theory of phase transitions. It was originally discovered in numerical studies and not long afterwards a mathematical proof for the simple Smoluchowski coagulation equations with constant kernel was given by \cite{KreerPenrose1994}. The dynamical scaling hypothesis is expected to hold for generalized Smoluchowski equations but not proven yet. The work of \cite{Kreer2014} related the fPp to a certain class of Smoluchowski equations and thus a dynamical scaling of the marginal probabilities would come to no surprise. Our current work has answered this in the affirmative even for the more general class of fnhPp as introduced by \cite{Scalas2017}. This generalisation has the advantage that the dynamical scaling exponent of time $t$ is not limited to positive numbers between 0 and 1 (which is the range the order of the fractional derivative can take) but can be any positive number due to the fact that the the choice of the intensity function $\Lambda(\cdot)$ allows for some additional freedom.
How would one apply a fnhPp to the problem of phase transitons? In our case we may imagine a solid crystal consisting of single monomers bound in a regular lattice structure. When the crystal is heated the lattice takes up heat and the monomeres start to oscillate and finally break free. At the end, the lattice has been dissolved because all monomers have gone. The very simple description of this ``phase transition'', in which the ordered state (= crystall lattice) has undergone a transition to an unordered state (free-moving monomers corresponding to a liquid or a gas) is to count the number of monomers breaking free from the remaining crystal structure, one at a time. The reverse of this model would be monomers condensing to a newly formed crystal. In fact, scientists have used this simple counting model to describe the nucleation of lysozyme and paracetamol in aqueous droplets: \cite{Goh2010} describe the nucleation by a Poisson process and estimate its parameters.
A simple computation shows that a standard Poisson process does not possess a regular dynamical scaling limit, namely with $\beta=1$ in Eq.(\ref{eq:int_repres}) we have after some straight-forward computations
\begin{eqnarray}
\lim_{t\rightarrow\infty, t^{-1}n=z_{0}} t^{1/2} P_{1}(n,t) =
\left\{
\begin{array}{ll}
\frac{1}{\sqrt{2\pi}} \hspace{0.5cm} \textrm{for} \hspace{0.2cm} z_{0}=1\\
0 \hspace{1cm} \textrm{otherwise}
\end{array}
\right.
\nonumber
\end{eqnarray}
With respect to the dynamical scaling hypothesis for phase transitions it would be very interesting to analyse the nucleation data of \cite{Goh2010} using a fractional Poisson process as model.
\section*{Acknowledgements}
Stimulating discussions with Lukas Kreer and Theo Manoussos (both Johannes-Gutenberg Universit\"{a}t, Mainz) and Dr. Ay\c{s}e K{\i}z{\i}lers\"{u} (University of Adelaide) are gratefully acknowledged.
This research did not receive any specific grant from funding agencies in the public, commercial or not-for-profit sectors.
\bibliographystyle{unsrt}
|
\section{Introduction:}
The presence of viscosity in the fluid content introduces many interesting features in the dynamics of homogeneous cosmological models \cite{1, 2, 3, 4, 5}. The dissipative mechanisms not only modify the nature of the singularity usually occurring for a perfect fluid, but also can successfully account for the large entropy per baryon in the present universe. Misner \cite{6, 7} suggested that any anisotropy in an expanding universe would be reduced to a rather insignificant level today by neutrino viscosity. Murphy \cite{8} presented an exactly soluble cosmological model of Friedmann type in the presence of bulk viscosity alone. Later, Banerjee and Santos \cite{9} extended the calculations to more general cases such as $k \pm 1$. It was shown, however, that in all the cases where the singularity is said to appear at infinite past, the fluid did not satisfy Hawking-Penrose energy conditions.\\
Exact solutions for homogeneous anisotropic models are not much known in the literature. There are, however, a few \cite{10, 11} which utilize certain simplifying assumptions to get exact solutions at the cost of a physically reasonable equation of state. Belinski$\check{\mathrm{i}}$ and Khalatnikov \cite{4} assumed an equation of state of the form $\rho \propto p$, but did not give any exact solution. They have, however, investigated some general features of the isotropic and anisotropic homogeneous cosmological models in the presence of bulk as well as shear viscosity in asymptotic limits. We consider in this paper the Bianchi I model with a fluid characterized by both bulk and shear viscosity and having an equation of state $\rho \propto p$. The viscosity coefficients are further assumed to be power functions of the matter density as suggested by Belinski$\check{\mathrm{i}}$ and Khalatnikov \cite{4}. Exact solutions in several particular cases for stiff fluid $\rho = p$ are worked out.\\
In Sec. II we have considered Einstein's field equation for a Bianchi I cosmological model and discussed how the dynamical importance of the shear and matter change in the course of the cosmological evolution. The entropy variation is also explicitly stated. In Sec. III exact solutions are obtained and their asymptotic characteristics are studied. The general behavior in the limits is compared with that of Belinski$\check{\mathrm{i}}$ and Khalatnikov \cite{4} at some places. Explicit solutions could be obtained considering the viscous coefficients only for a few restricted power functions of the mass density. These include the special cases with constant viscosity coefficients. In many of these cases the matter density vanishes at the initial instant, then increases in the course of evolution and finally reduces to zero. In such cases, therefore, at the initial singularity the metric is determined by the free space Einstein equations. In this context Belinski$\check{\mathrm{i}}$ and Khalatnikov \cite{4} remarked that the gravitational field creates the matter in the course of evolution.
\section{Einstein's field equations and some general results}
The metric of the homogeneous Bianchi type I model is
\be\label{2.1} ds^2 = -dt^2 + e^{2\alpha} dx^2 + + e^{2\beta} dy^2 + e^{2\gamma} dz^2,\ee
where $\alpha,~\beta~ \mathrm{and}~ \gamma$ are functions of time alone. The energy momentum tensor of the viscous fluid \cite{12} is given by
\be\label{2.2} T_{ij} = (\rho+\bar{p})v_iv_j + \bar{p} g_{ij} - \eta \mu_{ij},\ee
with
\be\label{2.3} \bar{p} = p - \left(\zeta - {2\over 3}\eta\right)\theta,\ee
and
\be\label{2.4} \mu_{ij} = v_{i;j} + v_{j;i} + v_i v^{a}v_{j;a} + v_j v^{a}v_{i;a},\ee
where,
\be \theta = {v^a}_{;a}.\ee
In the above equations $\zeta ~\mathrm{and}~ \eta$ stand for the bulk and shear viscosity coefficients, $\rho$ and $p$ are the mass density and pressure, respectively, $\bar{p}$ is the effective pressure, and $v_i$ represents the four-velocity, so that
\be \label{2.5} v_i v^j = -1.\ee
Choosing units $8\pi G = C = 1$, Einstein's field equations can be written as
\be \label{2.6} {R^i}_j - {1\over 2}{\delta^i}_j R = -{T^i}_j.\ee
Using comoving coordinates, so that $v_i = {\delta_0}^i$, the explicit
forms of Eq. \eqref{2.6} are
\be \label{2.7} {9\over 2}\left({\dot R\over R}\right)^2 - {1\over 2}(\dot\alpha^2 + \dot \beta^2 + \dot\gamma^2) = \rho,\ee
\be \label{2.8}\ddot \beta + \ddot\gamma + {3\over 2}\left({\dot R\over R}\right)(\dot\beta + \dot \gamma - \dot\alpha) + {1\over 2}(\dot\alpha^2 + \dot \beta^2 + \dot\gamma^2) = -(\bar p - 2\eta\dot\alpha),\ee
\be \label{2.9}\ddot\gamma + \ddot\alpha + {3\over 2}\left({\dot R\over R}\right)(\dot \gamma + \dot\alpha - \dot\beta)
+ {1\over 2}(\dot\alpha^2 + \dot \beta^2 + \dot\gamma^2)=-(\bar p - 2\eta\dot\beta),\ee
\be \label{2.10}\ddot\alpha + \ddot\beta + {3\over 2}\left({\dot R\over R}\right)(\dot\alpha + \dot\beta - \dot \gamma )
+ {1\over 2}(\dot\alpha^2 + \dot \beta^2 + \dot\gamma^2)=-(\bar p - 2\eta\dot\gamma),\ee
where the dot indicates time differentiation and
\be\label{2.11} R^3 = \exp{(\alpha + \beta + \gamma}).\ee
The usual definitions of the dynamical scalars such as the expansion scalar $\theta$ and the shear scalar $\sigma$ are considered to be
\be\label{2.12} \theta = {v^i}_{;i} ~~~~~\mathrm{and}~~~~~\sigma^2 = {1\over 2}\sigma_{ij}\sigma^{ij},\ee
where
\be\label{2.13} \sigma_{ij}= v_{i;j} + {1\over 2}(v_{i;k} v^k v_j + v_{j;k} v^k v_i) + {1\over 3} \theta(g_{ij} + v_i v_j).\ee
For the Bianchi type I metric with comoving coordinates we have
\be\label{2.14} \theta = 3\left(\dot R\over R\right),\ee
and
\be\label{2.15} 2\sigma^2 = (\dot\alpha^2 + \dot\beta^2 + \dot\gamma^2) - {1\over 3}\theta^2.\ee
The field equations \eqref{2.7} - \eqref{2.10} now yield
\be\label{2.16} {T^4}_4 = {1\over 3} \theta^2 - \sigma^2 = \rho,\ee
and
\be\label{2.17}g_{ij} G^{ij} = 2\dot\theta + {4\over 3}\theta^2
+ 2\sigma^2 = \rho - 3(\bar p - \zeta\theta).\ee
One can further obtain from the Bianchi identity
\be\label{2.18} \dot \rho = -(\rho + p)\theta + \zeta \theta^2 + 4\eta\sigma^2.\ee
It follows directly from Eq. \eqref{2.18} that for contraction, that is, $\theta < 0$ we have $\dot\rho > 0$ so that the matter density increases or decreases depending on whether the viscous heating is greater or less than the cooling due to expansion. It may be mentioned here that for an ultrarelativistic fluid $\rho = {1\over 3}p$ and $\zeta = 0$, Stewart \cite{13} has shown that the rate of viscous heating does not exceed one-half the rate of adiabatic cooling due to expansion.\\
Now, eliminating $\sigma^2$ from Eqs. \eqref{2.17} and \eqref{2.18}, one readily obtains
\be\label{2.19} \theta = {3\over 2}(\rho - p + \zeta\theta) - \theta^2,\ee
and
\be\label{2.20} \dot \rho = - (\rho + p)\theta + \zeta\theta^2 + 4\eta \left({1\over 3}\theta^2 - \rho\right).\ee
The relation \eqref{2.19} may be written as
\be\label{2.21} {d\over dt}\left[\ln{(\theta^2 R^6)}\right] = 3\left[(\rho - p)\theta^{-1} + \zeta\right].\ee
Using Eq. \eqref{2.16}, Eq. \eqref{2.19} can also be expressed in a different
form such as
\be\label{2.22} \dot\theta = -2\sigma^2 -{1\over 3}\theta^2 - {1\over 2}\left[\rho + 3(\rho - \zeta \theta)\right].\ee
This is exactly the Raychaudhuri equation \cite{14}. Further, we have
\be\label{2.23} R_{ij} v^i v^j = -{1\over 2}[\rho + 3(p - \zeta \theta)].\ee
The Hawking-Penrose energy condition is satisfied when $R_{ij}v^i v^j \le 0$. Thus, in a contracting model, so long as the fluid density and pressure remain positive, the energy conditions are satisfied. When the bulk viscosity is insignificant the energy condition is independent of the viscosity of the fluid. Again from the Raychaudhuri equation \eqref{2.22} it is evident that with the energy condition being satisfied, $\dot\theta < 0$, so that there is no bounce from a minimum volume.\\
Now for $\theta \ne 0$, that is, for a nonstatic model, one has, in view of Eq. \eqref{2.16},
\be\label{2.24} \left({\sigma^2\over \theta^2}\right)^{\textbf{.}} = - \left({\rho\over \theta^2}\right)^{\textbf{.}}\ee
Using the expressions for $\dot\rho ~\mathrm{and}~ \dot \theta$ from Eqs. \eqref{2.19} and \eqref{2.20} in Eq. \eqref{2.24}, one obtains after simplification the following result:
\be\label{2.25} \left({\sigma^2\over \theta^2}\right)^{\textbf{.}} = - \left({\rho\over \theta^2}\right)^{\textbf{.}} = - \left({\sigma^2\over \theta^2}\right)\left[3(\rho - p)\theta^{-1} + 3\zeta + 4\eta\right].\ee
The reasonable physical properties of the fluid demand $\rho \ge p > 0,~ \zeta > 0,~ \mathrm{and} ~ \eta > 0$, so that for an expanding model $(\theta > 0)$, we have $\left({\rho\over \theta^2}\right)^{\textbf{.}} > 0$ and $\left({\sigma^2\over\theta^2}\right)^{\textbf{.}} < 0$. It is evident, therefore, that $\left({\rho\over \theta^2}\right)$ increases with time, while $\left({\sigma^2\over\theta^2}\right)$ decreases. The dynamical influence of matter, therefore, increases with expansion, whereas that of shear decreases. For contraction, however, $\theta < 0$ and nothing can be said with certainty. It is interesting to note that for a stiff fluid, that is, $\rho = p$, $\left({\rho\over\theta^2}\right)$ is greater than zero and $\left({\sigma^2\over\theta^2}\right)$ is less than zero, so long as viscosity coefficients are positive, and this behavior holds irrespective of whether the model expands or contracts. Combining Eqs. \eqref{2.21} and \eqref{2.25}, one obtains
\be\label{2.26} \left[\ln{(\sigma^2 R^6)}\right] = - 4\eta,\ee
which in turn can also be written as
\be\label{2.27} (\sigma^2)^{\textbf{.}} = -2(2\eta + \theta)\sigma^2.\ee
This is the shear propagation equation. It follows from Eq. \eqref{2.27} that for the expanding model $\theta > 0$ the shear decreases with time. The rate of work done by anisotropic stresses augments the shear dissipation. It is also evident from Eq. \eqref{2.27} that the shear dissipation depends on the expansion rate $(\theta)$, which is again affected by the presence of bulk viscosity as is evident from Eq. \eqref{2.22}. The bulk viscosity has thus a significant role in the process of the shear dissipation mechanism.\\
When $\eta$ is assumed to be a constant the relation \eqref{2.27} can be directly integrated to yield
\be\label{2.28} \sigma^2 = \left({\sigma_0^2 \over R^6}\right)e^{-4\eta t},\ee
$\sigma_0$ being the integration constant. The effect of shear viscosity is to reduce the anisotropy in the course of time in the form of an exponential factor. This purely relativistic result is due to Misner \cite{6, 7}.\\
Following Belinski$\check{\mathrm{i}}$ and Khalatnikov \cite{4} the time derivative of the entropy density in the model is given by
\be\label{2.29} {\dot\Sigma\over \Sigma} = {\dot\rho\over (\rho + p)},\ee
where $\Sigma$ is the entropy density. Defining the total entropy by $S = R^3\Sigma$, one gets from Eq. \eqref{2.20} using Eq. \eqref{2.29} the relation
\be\label{2.30} {\dot S\over S} = {(\zeta \theta^2 + 4\eta \sigma^2)\over (\rho+p)}.\ee
We now restrict ourselves to an equation of state of the form
\be\label{2.31} p = (\gamma - 1)\rho,~~~~1 \le \gamma \le 2,\ee
and assume that the viscosity coefficients are constants so that $\zeta = \zeta_0$ and $\eta = \eta_0$ The qualitative aspects of the presence of viscosity will, however, be present in such a restricted case also (see Misner \cite{6, 7} and Treciokas and Ellis \cite{15}. In this case Eq. \eqref{2.30} can be written as
\be\label{2.32} {\dot S\over S} = {\left[\zeta_0 + 4 \eta_0 \left({\sigma^2 \over \theta^2}\right)\right]\over \gamma\left({\rho\over \theta^2}\right)}.\ee
By further differentiation with respect to time and using Eqs. \eqref{2.25} and \eqref{2.32}, we find
\be\label{2.33} \begin{split}S^{-1} \ddot S =& \gamma^{-2} \left({\rho\over \theta^2}\right)^{-2}\bigg[\zeta_0^2 - \left({\sigma^2\over \theta^2}\right)\left({\rho\over \theta^2}\right) \times \left\{16\eta_0^2 + \gamma(2-\gamma)(3\zeta_0 + 4\eta_0)\theta \right\}\\&
+ 8\eta_0(\gamma-1)\left(\zeta_0 + {2\over 3}\eta_0\right)+ 3\gamma\zeta_0^2\bigg].\end{split}\ee
For an expanding model, $\left({\sigma^2\over \theta^2}\right)$ decreases with time, whereas $\left({\rho\over \theta^2}\right)$ increases, as was already discussed. The minimum of $\left({\sigma^2\over \theta^2}\right)$ is zero, when $\left({\rho\over \theta^2}\right) = {1\over 3}$ and
$S^{-1} \ddot S = \zeta_0^2\gamma^{-2} \left({\rho\over \theta^2}\right)^{-2}$, which is greater than zero. In the
course of time $\left({\rho\over \theta^2}\right)$ decreases and we consider the extreme
case when $\left({\rho\over \theta^2}\right) = 0$. At this instant $\left({\sigma^2\over \theta^2}\right) = {1\over 3}$ and from Eq. \eqref{2.33} it is evident that $\ddot S < 0$. The relation \eqref{2.32} indicates that $\dot S > 0$, that is, the total entropy always increases for nonnegative values of matter density, whereas for expansion, $\ddot S$ is initially negative and later becomes positive in the course of time. Since ${\dot S\over S} \rightarrow \infty$ as $\left({\rho\over \theta^2}\right) \rightarrow 0$, we have the $S - t$ curve intersecting the time axis. It means that $S$ reduces to zero at some finite time. The picture is more clear for a stiff fluid when $\rho = p$. We have then from Eq. \eqref{2.25}
\be {\left({\sigma^2\over \theta^2}\right)^{\textbf{.}}\over \left({\sigma^2\over \theta^2}\right)} = (3\zeta_0 + 4\eta_0),\ee
which yields on integration
\be \label{2.34} \left({\sigma^2\over \theta^2}\right) = A^2 e^{-{(3\zeta_0 + 4\eta_0)t}},\ee
where $A^2$ is the magnitude of the ratio ${\sigma^2\over \theta^2}$ at $t = 0$. It is evident from Eq. \eqref{2.34} that for both $\zeta$ and $\eta$ as constants the ratio of shear to expansion decays exponentially and the rate falls in the absence of either bulk or shear viscosity. From Eqs. \eqref{2.16} and \eqref{2.34} one gets the expression for ${\rho\over \theta^2}$ in the form
\be\label{2.35} {\rho\over \theta^2} = {1\over 3} - A^2 e^{-(3\zeta_0 + 4\eta_0)t}.\ee
So, initially, if one starts with zero mass density at some
finite time one must have ${\left(\rho\over \theta^2\right)_{max}} = {1\over 3}$ at $t \rightarrow 0$. In view of Eqs. \eqref{2.34} and \eqref{2.35}, Eq. \eqref{2.32} can now be written as
\be\label{2.36} {\dot S\over S} = {\zeta_0 + 4\eta_0 A^2 e^{-(3\zeta_0 + 4\eta_0)t}\over 2\left({1\over 3} - A^2 e^{-(3\zeta_0 + 4\eta_0)t}\right)},\ee
which in turn yields on integration
\be\label{2.37} S = S_0\left[{1\over 3}e^{3\zeta_0 t} - A^2 e^{-4\eta_0 t}\right]^{1\over 2},\ee
with $S_0$ being the constant of integration. It is evident that at
\be\label{2.38} t = \left[\ln{(3A^2)}\right](3\zeta_0 + 4\eta_0)^{-1},\ee
the total entropy $S = 0$, when we also have ${\rho\over \theta^2} = 0$. Again as $t\rightarrow \infty,~ S\rightarrow \infty$ and ${\rho\over \theta^2}$ approaches its maximum value. Combining Eqs. \eqref{2.35} and \eqref{2.37} together one can also write
\be\label{2.39} S^2 = S_0^2 \left({\rho\over \theta^2}\right)e^{3\zeta_0 t},\ee
which in turn demands that the matter density must be nonnegative in this case.
\section{Special solutions for a Bianchi I model with a viscous fluid}
In what follows we consider some special cases with restrictions on the behavior of the bulk and shear viscosity coefficients. It is true that the assumptions regarding these viscosity coefficients may not always be valid in an actual fluid throughout the entire evolution history of the cosmological models hitherto discussed: the solutions are nevertheless interesting in indicating the role of viscosity in cosmological evolution.\\
In view of Eqs. \eqref{2.18} and \eqref{2.26} we now have the relation
\be\label{3.1} {d\over dt}\left[\rho + \sigma^2) R^6\right] = \zeta \theta^2 R^6.\ee
Using Eq. \eqref{2.16} in Eq. \eqref{3.1} we further obtain
\be\label{3.2} {d\over dt}\left[\ln{(\theta^2 R^6)}\right] = 3\zeta.\ee
\noindent
\textbf{Case - 1:}\\
In this case we consider $\zeta = 0$ and $\eta = \eta_0 \rho^n$, where $\eta_0$ and $n$ are constants. Equation \eqref{3.2} can immediately be integrated in this case to yield
\be\label{3.3} R^3 = {R_0}^3 t,\ee
where $R_0$ is an arbitrary constant and the time coordinate is
chosen such that the proper volume vanishes at $t = 0$. It is,
however, not difficult to show that one will get the same
solution \eqref{3.3} if one assumes the shear and bulk viscosity to
be absent for a perfect fluid. The expansion scalar $\theta$ is given
by
\be\label{3.4} \theta = t^{-1}.\ee
Equation \eqref{2.26} now yields in view of Eq. \eqref{2.16}
\be\label{3.5} \left({\sigma^2\over \theta^2}\right)^{-1}\left({\sigma^2\over \theta^2}\right)^\textbf{.} = - 4\eta_0 \theta^{2n}\left({1\over 3} - {\sigma^2\over \theta^2}\right)^n.\ee
Writing ${\sigma^2\over \theta^2} = y$ and using the relation \eqref{3.4}, Eq. \eqref{3.5} can be written as a first-order differential equation
\be\label{3.6}{\dot y\over y}= -4\eta_0 t^{-2n}\left({1\over 3} - y\right)^n.\ee
In view of Eq. \eqref{2.16} $\rho$ is positive when $\left({1\over 3} - y\right) > 0$. For $n = 1$ the solution for $y$ is obtained by integrating Eq. \eqref{3.6} as
\be \label{3.7} y = {\sigma^2\over \theta^2} = {1\over 3}\left(1 + a^2 e^{-{4\eta_0\over 3t}}\right)^{-1},\ee
where $a^2$ is a positive constant. It follows from Eq. \eqref{3.7}, in view of Eq. \eqref{3.4}, that
\be \label{3.8} \sigma^2 = {1\over 3 t^2}\left(1 + a^2 e^{-{4\eta_0\over 3t}}\right)^{-1}.\ee
Using Eq. \eqref{2.16} we therefore obtain,
\be \label{3.9} \rho = \left({a^2\over 3 t^2}\right)e^{-{4\eta_0\over 3t}}\left(1 + a^2 e^{-{4\eta_0\over 3t}}\right)^{-1}.\ee
Now consider an expanding model, for which $\theta > 0$. In this case at $t \rightarrow 0,~ R^3 \rightarrow 0$ and both the shear and expansion scalars attain infinitely large magnitudes, while the density reduces to zero. It is an interesting behaviour as noted previously by Belinski$\check{\mathrm{i}}$ and Khalatnikov \cite{4}. The singularity at $t = 0$ in this case corresponds to vanishing proper volume $R^3 = 0$. But, unlike in the usual case of cosmological singularity, the density vanishes in the limit instead of increasing to infinity. The matter density subsequently increases and again decreases to approach zero magnitude at the final stage at $t\rightarrow \infty$, as is evident from Eq. \eqref{3.9}. In this limit, however, both the expansion $(\theta)$ and shear $(\sigma^2)$ scalars vanish and $R^3 \rightarrow \infty$. In other words the model may be said to approach asymptotically the isotropic Friedmann universe (cf. Belinski$\check{\mathrm{i}}$ and Khalatnikov \cite{4}). On the other hand, collapse may be discussed for a negative value of $t$ when $t < 0,~\theta < 0$, which represents a contracting model. As $t\rightarrow -\infty$, all the quantities such as $\theta,~\sigma^2 ~\mathrm{and}~ \rho$ vanish with infinitely large proper volume representing a Friedmann model. In the course of time as $t \rightarrow 0$, the proper volume reduces to zero whereas $\theta,~\sigma^2 ~\mathrm{and}~ \rho$ all increase indefinitely. These properties in asymptotic limits only are discussed by Belinski$\check{\mathrm{i}}$ and Khalatnikov \cite{4}.\\
For $n = {3\over 2}$ the relation \eqref{3.2} takes the following form:
\be\label{3.10} {\dot y\over y} = -\left(4\eta_0\over t^3\right)\left({1\over 3} - y\right)^{3\over 2}.\ee
Since ${\left(\sigma^2\over \theta^2\right)^\textbf{.}} < 0$ for positive $\zeta$ and $\eta$, that is, $\dot y < 0$, $t$ should assume only positive values and thus one can have only expansion allowed in this case $(\theta > 0)$. Integrating Eq. \eqref{3.10}, we have
\be\label{3.11} \ln{\left[1-(1-3y)^{1\over 2}\over 1 + (1-3y)^{1\over 2}\right]} + {2\over (1-3y)^{1\over 2}} = {2\eta_0\over 3\sqrt 3}t^{-2} + \mathrm{constant}.\ee
For $n = 2$, we find from Eq. \eqref{3.6}, the solution of $y$ is given by
\be\label{3.12} 3\ln{\left[y\over \left({1\over 3} - y\right)\right]} + {1\over \left({1\over 3} - y\right)} = {4\over 9}\eta_0 t^{-3} + \mathrm{constant}.\ee
Though the solutions \eqref{3.11} and \eqref{3.12} are not in closed form, it is not very difficult to investigate the properties of these models at limits. The analysis is done in an identical manner as for $n = 1$. The behaviour can be seen to be almost identical in the limits $t\rightarrow 0$ or $t \rightarrow \infty$.\\
For $n = {1\over 2}$, Eq. \eqref{3.6} can be expressed as
\be\label{3.13} {\dot y\over y} = -\left(4\eta_0\over t\right)\left({1\over 3} - y\right)^{1\over 2},\ee
which on integration yields
\be \ln{\left[1-(1-3y)^{1\over 2}\over 1 + (1-3y)^{1\over 2}\right]} = \left({t\over t_0}\right)^{-{4\eta_0\over \sqrt 3}}.\ee
where $t_0$ is the constant of integration and is less than $t$ as is clear from the above equation, which on further simplification gives
\be\label{3.14} y = {4\over 3}\left[{\left({t\over t_0}\right)^{-{4\eta_0\over \sqrt 3}}\over \left(1 + \left({t\over t_0}\right)^{-{4\eta_0\over \sqrt 3}}\right)^2}\right].\ee
One can now use Eq. \eqref{3.4} in Eq. \eqref{3.14} to obtain an expression for $\sigma^2$ (remembering, $y = \sigma^2/\theta^2$) as
\be\label{3.15} \sigma^2 = \left({4\over 3 t^2}\right)
\left({t\over t_0}\right)^{-{4\eta_0\over \sqrt 3}} \left(1 + \left({t\over t_0}\right)^{-{4\eta_0\over \sqrt 3}}\right)^{-2},\ee
and also
\be\label{3.16} \rho = \left({1\over 3t^2}\right)\left[1 - \left({t\over t_0}\right)^{-{4\eta_0\over \sqrt 3}}\right]^2 \left[1 + \left({t\over t_0}\right)^{-{4\eta_0\over \sqrt 3}}\right]^{-2}.\ee
For an expanding model $\theta > 0$. In this case, as $t\rightarrow t_0$, all the parameters $\theta,~R^3~\mathrm{and}~\sigma^2$ remain finite, while the matter density vanishes. On the other hand, as $t\rightarrow \infty$ the expansion $(\theta)$ and the shear $(\sigma^2)$ scalars vanish although the proper volume increases indefinitely, while the matter density $\rho$ approaches zero. Thus asymptotically the picture represents an isotropic Friedmann universe.\\
The simplest case is for $n = 0$, that is, $\eta = \eta_0$. We now have from Eq. \eqref{3.6}
\be\label{3.17} {\dot y\over y} = -4\eta_0,\ee
so that, $y = {\sigma^2\over \theta^2} \propto e^{-4\eta_0 t}$, yielding the relation
\be \label{3.18} \sigma^2 = \left({\sigma_0^2\over t^2}\right)e^{-4\eta_0 t}.\ee Further,
\be \label{3.19} \rho = {1\over 3}\theta^2 - \sigma^2 = {1\over t^2}\left[{1\over 3} - \sigma_0^2 e^{-4\eta_0 t} \right].\ee
The behavior in this case is quite different from the previous cases at least in the initial phase of expansion. Here at $t\rightarrow 0$, the physical and geometrical quantities such as $\rho,~\theta,~\sigma^2$ become infinitely large. The magnitude of the constant $\sigma_0^2$, however, cannot be greater than ${1\over 3}$ for positive values of matter density $\rho$. On the other hand, as $t\rightarrow \infty$, all the quantities $\rho,~\theta,~\sigma^2$ reduce to vanishingly small quantities. \\
It should be mentioned at this point that in view of all the models discussed so far, one can conclude that the Hawking-Penrose energy condition $(R_{ij} v^i v^j \le 0)$ is satisfied so long as $\rho \ge 0$. This is clear from the relation \eqref{2.23} and the fact that in the above models we have assumed $\zeta = 0$.\\
\noindent
\textbf{Case - 2:}\\
Let us now turn our attention to the nonvanishing values of bulk viscosity coefficients, that is, the situation where the bulk viscosity of the fluid cannot be completely ignored. We therefore assume
\be\label{3.20} \zeta = \zeta_0~~~~~\mathrm{and}~~~~~\eta = \eta_0 \rho^q,\ee
where, $\zeta_0$ and $\eta_0$ are constants. Now, integrating Eq. \eqref{3.2} one gets
\be\label{3.21} \theta = \left(\theta_0\over R^6\right) e^{{3\over 2}\zeta_0 t},\ee
$\theta_0$ being the constant of integration. Remembering that for the Bianchi I metric $\theta = 3\left(\dot R\over R\right)$ and so integrating further
Eq. \eqref{3.21}, we have the solution for $R$, which is given by
\be\label{3.22} R^3 = {\left(2\theta_0\over \zeta_0\right)}\left(e^{{3\over 2}\zeta_0 t} + D\right),\ee
where, $D$ is a constant of integration. Equation \eqref{3.20} therefore yields
\be \label{3.23} \theta = {3\over 2} \zeta_0\left[{e^{{3\over 2}\zeta_0 t}\over e^{{3\over 2}\zeta_0 t} + D}\right].\ee
Now from Eqs. \eqref{2.25} and \eqref{3.20} one gets
\be {\dot y\over y} = - \left(3\zeta_0 + 4\eta_0 \rho^q\right).\ee
which in turn can be written in view of Eq. \eqref{2.16} as
\be\label{3.24} {\dot y\over y} = -3\zeta_0 - 4\eta_0 \theta^{2q}\left({1\over 3} - y\right)^q.\ee
The special case for $D = 0$ is particularly simple and we discuss
only this case. We therefore have $\theta = {3\over 2}\zeta_0$ so that $\dot\theta = 0$. Here the expansion is steady and the rate is constant. One of
the relatively simple cases is $q = 1$. In this case $\eta = \eta_0\rho$ and
hence we obtain from Eq. \eqref{3.24}
\be\label{3.25} {\dot y\over y} = -3\zeta_0 - 4\eta_0 \theta^{2}\left({1\over 3} - y\right).\ee
Writing $3\zeta_0 = a_0$ and $9\eta_0^2 \zeta_0^2 = b_0$, the relation \eqref{3.25} may be written as
\be {\dot y\over y} = -a_0 -b_0\left({1\over 3} - y\right).\ee
where both $a_0$ and $b_0$ are greater than zero. Integrating Eq. \eqref{3.25} we further obtain
\be\label{3.26} {y\over (c_0 -y)} = e^{b_0c_0(t_0 - t)},\ee
So that one can write explicitly
\be\label{3.27} {\sigma^2\over \theta^2} = {c_0\over \left(1 + e^{b_0c_0(t - t_0)}\right)},\ee
where $c_0 = {a_0\over b_0} + {1\over 3}$, and $t_0$ is the constant of integration. Since here $\theta = {3\over 2} \zeta_0$ the expansion scalar $\theta$ is positive for the physical requirement $\zeta_0 > 0$. The above solution, therefore, represents an expanding model only. The density $\rho$ vanishes
at a finite time, when ${\sigma^2\over \theta^2} = {1\over 3}$, so that $\sigma^2$ remains finite. The proper volume represented by $R^3$ also has finite magnitude. But for $t \rightarrow \infty$, $R^3\rightarrow \infty,~\sigma^2 \rightarrow 0$, and $\rho\rightarrow {3\over 4}\zeta_0^2$. The singularity of vanishing volume $R^3 = 0$ exists at $t\rightarrow -\infty$, where the density is negative infinity. In fact prior to the instant corresponding to ${\sigma^2\over \theta^2} = {1\over 3}$, the density assumes only negative values. If one calculates $R_{ij}v^i v^j$ in this model, one finds it to be positive so that the energy condition is violated throughout.\\
Particularly, simple models can be constructed in this case, taking $q = 0$ in Eq. \eqref{3.20}, so that
\be\label{3.28} \zeta = \zeta_0,~~~~~\eta = \eta_0.\ee
The expressions for $R^3$ and $\theta$ remain unaltered from those given in Eqs. \eqref{3.22} and \eqref{3.23}, respectively. Equation \eqref{3.24}
then reduces to
\be\label{3.29} {\dot y\over y} = =(3\zeta_0 + 4\eta_0).\ee
This case is already mentioned in Eq. \eqref{2.34}. Integrating Eq. \eqref{3.29}, the solution can be obtained in the form
\be y = {\sigma^2\over \theta^2} = A^2 e^{-(3\zeta_0 + 4 \eta_0)t},\ee
so that
\be\label{3.30} \sigma^2 = {9\over 4} A^2 {\zeta_0}^2 e^{-4\eta_0 t}\left[e^{{3\over 2}\zeta_0 t} - D\right]^{-2},\ee
and the matter density $\rho$ is expressed as
\be \rho = \theta^2\left({1\over 3} - {\sigma^2\over \theta^2}\right) =
{9\over 4}\left[{{\zeta_0}^2 e^{3\zeta_0 t}\over \left(e^{{3\over 2}\zeta_0 t} + D\right)^2}\right]\left[{1\over 3} - A^2 e^{-(3\zeta_0 + 4\eta_0)t}\right].\ee
The maximum of ${\sigma^2\over \theta^2}$ is ${1\over 3}$ and this occurs at some finite
time $t$ given by Eq. \eqref{2.38}. In this limit $\rho = 0$ and $\sigma^2,~ \theta^2$ are both finite. For $D \ge 0$ the proper volume never reduces to zero magnitude as is evident from Eq. \eqref{3.22}. On the other hand, as $t\rightarrow \infty,~\theta \rightarrow {3\over 2}\zeta_0$, which is finite, $\sigma^2 \rightarrow 0$, and $\rho \rightarrow {3\over 2}{\zeta_0}^2$, but $R^3\rightarrow \infty$. We note that though the proper volume increases to large dimension and the anisotropy reduces to zero the fluid density $\rho$ does not vanish, unlike the Friedmann universe. For $D < 0$ we note that at a finite time $R^3 = 0$ and $\theta,~\sigma^2, ~\rho$ all become infinitely large. But, as $t\rightarrow \infty$, the proper volume also tends to infinity ($R^3 \rightarrow \infty$), while $\sigma^2 \rightarrow 0$ while, $\theta$ and $\rho$ both remain finite. In the limit $t\rightarrow \infty$ the behavior of the model is independent of the sign of the constant $D$.\\
For $D\ge 0$ it can be shown that the Hawking-Penrose energy conditions are violated throughout and for $D < 0$ this happens for large time $t$.
\section{Conclusion}
In summary, we have considered the Bianchi type I cosmological model with a viscous fluid, assuming the coefficients of viscosity as power functions of the matter density and considering an equation of state for a stiff fluid $(\rho = p)$. In most of the cases it has been observed that the matter density is zero at the initial singularity but then increases in the course of evolution, finally vanishing again in the asymptotic limit, which implies that the gravitational field creates matter. In the case of expanding models it is found that the dynamical importance of matter increases while that of shear decreases in the course of evolution. For a stiff fluid, in particular, this result is shown to hold also for contracting models. In addition to the role of shear viscosity in the dissipation of shear it is pointed out that the bulk viscosity can also augment the shear dissipation. For the stiff fluid with constant viscosity coefficients it is observed that the bulk viscosity can be an effective mechanism for large entropy in the asymptotic limit when the model approaches the isotropic Friedmann universe. Although the role of shear viscosity also is to increase the entropy in the course of the expansion, its effect becomes gradually less compared to that of the bulk viscosity in the asymptotic limit. The magnitude of the entropy is low at the highly anisotropic initial phase of evolution, as is observed from Eq. \eqref{2.37}, and then increases subsequently.\\
In the preceding section we have considered two different cases. The first case is $\eta = \eta_0\rho^n$ and $\zeta = 0$. Solutions for particular values of n such as $n = {1\over 2},~ 1, ~{3\over 2},~ 2$ are explicitly given. It is found that in all these cases except for $n = 1$, only expanding models are allowed. For $n = {1\over 2},~ 1, \mathrm{and} ~ 2$, the models have been found to approach the isotropic Friedmann universe asymptotically. The matter density vanishes at the initial phase of singularity, increases subsequently during the evolution, ultimately reduces to zero in the asymptotic limit. For $n = {3\over 2}$, the solution is not in the closed form and as such the behavior cannot be studied. For $n = 1$ as $t \rightarrow -\infty$, the proper volume $R^3$ tends to infinity whereas $\theta,~\sigma^2 ~\mathrm{and}~\rho $ vanish, representing a Friedmann model. In the course of time as $t \rightarrow 0$ the volume contracts and reduces to zero and the other physical scalars become infinitely large. In the case where the shear viscosity coefficient is assumed to be constant, i.e., $n = 0$, the behavior is different. Here at the initial phase of expansion $\theta,~\sigma^2 ~\mathrm{and}~\rho $ are infinitely large, but asymptotically, however, all of them reduce to vanishingly small quantities.\\
In the second case we have considered, $\eta = \eta_0\rho^q$, $\zeta = \zeta_0$· For $q = 0~ \mathrm{and}~ 1$, only expansion is found to be allowed. For $q = 1$, $\rho$ vanishes at a finite time, keeping the proper volume finite. In the limit $t\rightarrow \infty$, the model isotropizes with infinite proper volume but the matter density is non-vanishing, unlike the previous cases. For $q = 0$, either of the two cases is observed, depending on the sign of an integration constant $D$. For $D \ge 0$, the asymptotic behavior is the same as for $q = 1$. But for $D < 0$, the proper volume vanishes at a finite time while $\theta,~\sigma^2 ~\mathrm{and}~\rho $ take infinitely large magnitude at this instant. In the asymptotic limit, $t \rightarrow \infty$, the model is again identical to the case $q = 1$.
|
\section{\@startsection{section}{1}{\zeta@}{-3.5ex plus -1ex minus
\newcommand{\stackrel{(d)}{\longrightarrow}}{\stackrel{(d)}{\longrightarrow}}
\newcommand{\stackrel{\IP}{\longrightarrow}}{\stackrel{{\mathbb P}}{\longrightarrow}}
\newcommand{\stackrel{L^1}{\longrightarrow}}{\stackrel{L^1}{\longrightarrow}}
\newcommand{\stackrel{L^2}{\longrightarrow}}{\stackrel{L^2}{\longrightarrow}}
\newcommand{\stackrel{(d)}{=}}{\stackrel{(d)}{=}}
\newcommand{\stackrel{\rm{ f.d.m.}}{\longrightarrow}}{\stackrel{\rm{ f.d.m.}}{\longrightarrow}}
\newcommand{\bar \zeta}{\bar \zeta}
\newcommand{\mu^{(2)}_\8}{\mu^{(2)}_\infty}
\newcommand{\color{red}}{\color{red}}
\newcommand{\color{magenta}}{\color{magenta}}
\newcommand{{\mathbb{W}}}{{\mathbb{W}}}
\bibliographystyle{amsplain}
\AtBeginDocument
\def\MR#1{} }
\sloppy
\setcounter{tocdepth}{1}
\author{
Shuta Nakajima\footnote{Department of Mathematics and Computer Science of the University of Basel. \texttt{<EMAIL>}}
\and
Makoto Nakashima\footnote{Graduate School of Mathematics, Nagoya University. \texttt{<EMAIL>}}
}
\begin{document}
\title{Fluctuations of two-dimensional stochastic heat equation and KPZ equation in subcritical regime for general initial conditions}
\date{}
\maketitle
\begin{center}
\date{\today}
\end{center}
\begin{abstract}
The solution of Kardar-Parisi-Zhang equation (KPZ equation) is solved formally via Cole-Hopf transformation $h=\log u$, where $u$ is the solution of multiplicative stochastic heat equation(SHE). In \cite{CD20,CSZ20,G20}, they consider the solution of two dimensional KPZ equation via the solution $u_\varepsilon$ of SHE with flat initial condition and with noise which is mollified in space on scale in $\varepsilon$ and its strength is weakened as $\b_\varepsilon=\hat{\b} \sqrt{\frac{2\pi \varepsilon}{-\log \varepsilon}}$, and they prove that when $\hat{\b}\in (0,1)$, $\frac{1}{\b_\varepsilon}(\log u_\varepsilon-\mathbb E[\log u_\varepsilon])$ converges in distribution to a solution of Edward-Wilkinson model as a random field.
In this paper, we consider a stochastic heat equation $u_\varepsilon$ with general initial condition $u_0$ and its transformation $F(u_\varepsilon)$ for $F$ in a class of functions $\mathfrak{F}$, which contains $F(x)=x^p$ ($0<p\leq 1$) and $F(x)=\log x$. Then, we prove that $\frac{1}{\b_\varepsilon}(F(u_\varepsilon(t,x))-\mathbb E[F(u_\varepsilon(t,x))])$ converges in distribution to Gaussian random variables jointly in finitely many $F\in \mathfrak{F}$, $t$, and $u_0$. In particular, we obtain the fluctuations of solutions of stochastic heat equations and KPZ equations jointly converge to solutions of SPDEs which depends on $u_0$.
Our main tools are It\^o's formula, the martingale central limit theorem, and the homogenization argument as in \cite{CNN20}. To this end, we also prove the local limit theorem for the partition function of intermediate $2d$-directed polymers
\end{abstract}
\noindent \textbf{Keywords}: KPZ equation, Stochastic heat equation, Edwards-Wilkinson equation, Local limit theorem for polymers, Stochastic calculus.
\\
\noindent \textbf{AMS 2010 subject classifications}: Primary 60K37. Secondary 60F05, 60G42, 82D60.
\setcounter{tocdepth}{2}
\tableofcontents
{\footnotesize
\section{Introduction and Main result}
KPZ equation is an SPDE formally given by \begin{equation} \label{eq:formalKPZ}
\frac{\partial}{\partial t} h(t,x) = \frac12 \Delta h(t,x) + \frac12 |\nabla h(t,x) |^2+
\b \dot\xi(t,x),
\end{equation}
where $\xi$ is a time-space white noise on $[0,\infty)\times \mathbb{R}^d$. This SPDE is ill-posed due to the non-linear term $\nabla h$ which should be a generalized function.
For $d=1$, Bertini and Giacomin formulated the solution of \eqref{eq:formalKPZ} via Cole-Hopf solution $h=\log u$ \cite{BG97}, where $u$ is the solution of stochastic heat equation \begin{align}\label{eq:formalSHE}
\frac{\partial }{\partial t}u(t,x)=\frac{1}{2}\Delta u(t,x)+\b u(t,x)\dot{\xi}(t,x).
\end{align}
In dimension $d=2$, we consider a space-regularized multiplicative stochastic heat equation but we need to scale the disorder strength: \begin{align}
\frac{\partial u_\varepsilon}{\partial t}=\frac{1}{2}\Delta u_\varepsilon+\b_\varepsilon u_\varepsilon \dot\xi_\varepsilon,\quad u_\varepsilon(0,x)=u_0(x)
\end{align}
where $\beta_\varepsilon=\hat{\b}\sqrt{\frac{2\pi}{-\log \varepsilon}}$ with $\hat{\b}\geq 0$, and $\xi_\varepsilon$ is a mollification in space of $\xi$ such that $\xi_\varepsilon \Rightarrow \xi$ as $\varepsilon\to 0$, i.e.,\
\[
\dot\xi_{\varepsilon}(t,x) = (\dot\xi(t,\cdot)\star \phi_\varepsilon)(x)= \int \phi_\varepsilon(x - y) \dot{\xi}(t,y) \dd y,
\]
with $\phi_\varepsilon({ x}) = \varepsilon^{-2} \phi(\varepsilon^{-1} x)$ and $\phi$ being a smooth, non-negative, compactly supported, symmetric function on $\mathbb R ^2$, so that $\int \phi(x) \dd x =1$, and $\phi_\varepsilon$ converges in distribution to the Dirac mass $\delta_0$. Let $h_\varepsilon=\log u_\varepsilon$. Then, we find by It\^{o}'s formula that $h_\varepsilon$ satisfies the SPDE\begin{equation}\label{eq:KPZe}
\frac{\partial}{\partial t} h_{\varepsilon} = \frac12 \Delta h_{\varepsilon} + \bigg[\frac12 |\nabla h_\varepsilon |^2 - C_\varepsilon\bigg]+
\b_\varepsilon \dot\xi_{\varepsilon} \;,\quad\,\, h_{\varepsilon}(0,x) = h_0(x),
\end{equation}
where $C_\varepsilon$ is a diverging parameter:
\begin{equation}\label{C-eps}
C_\varepsilon= \frac{\b_\varepsilon^2 V(0) }{2\varepsilon^2},
\end{equation}
where $V(x)=\int_{{\mathbb{R}}^2}\phi(x-y)\phi(y)\dd y $.
Caravenna, Sun, and Zygouras proved that if the initial condition is flat, that is $h_\varepsilon(0,x)= h_0(x)\equiv 0$ and $\hat{\b}\in (0,1)$, then $\displaystyle {\b_\varepsilon^{-1}}\left(h_\varepsilon-\mathbb E[h_\varepsilon]\right)$ converges in distribution to the solution to Edward-Wilkinson equation as a random field \cite{CSZ20}. We should remark Chatterjee and Dunlap addressed the tightness of ${\b_\varepsilon^{-1}}\left(h_\varepsilon-\mathbb E[h_\varepsilon]\right)$ \cite{CD20} and Gu obtained Edward-Wilkinson limit in $\hat{\b}\in (0,\beta_0)$ for some $\beta_0\leq 1$ \cite{G20}.
On the other hand, the one-point distribution $h_\varepsilon(t,x)$ converges to a random variable as follows:
\begin{theorem}\label{thm:CSZ17b}{\cite[Theorem 2.15]{CSZ17b}}
For any $t>0$ and $x\in {\mathbb{R}}^2$, \begin{align*}
h_\varepsilon(t,x)\Rightarrow \begin{cases}
X_{\hat{\beta}}:=\sigma(\hat{\b})Z-\frac{1}{2}\sigma^2(\hat{\b}),\quad &0\leq \hat{\b}<1\\
0,&\hat{\b}\geq 1
\end{cases},
\end{align*}
where $Z$ is a random variable with standard normal distribution and $\sigma(\hat{\b})=\sqrt{\log \frac{1}{1-\hat{\b}^2}}$.
\end{theorem}
We will look at the fluctuation of $u_\varepsilon$ for general initial conditions in our main results. Let $\mathfrak{C}$ be a set of continuous functions which satisfies \begin{equation}\label{eq:u0cond}
0<\inf_{x\in \mathbb{R}^2}u_0(x)\leq \sup_{x\in\mathbb{R}^2}u_0(x)<\infty,
\end{equation}
or equivalently
\begin{align*}
\|\log u_0\|_\infty<\infty.
\end{align*}
{ Let $\mathfrak{F}$ be a set of functions $F\in C^3((0,\infty))$ such that there exists a constant $C=C_F>0$ such that for any $x\in (0,\infty)$\begin{align}
&|F'(x)|\leq C(x^{-1}+1),\quad
|F''(x)|\leq C(x^{-2}+1),\quad |F'''(x)|\leq C(x^{-3}+1).\label{eq:Fass}
\end{align}
Then, $\mathfrak{F}$ contains $x^p$ ($0<p\leq 1$), $\log x$, $\sin x$, $\cos x$, $e^{-x}$ $\dots$ In this paper, we focus on the fluctuation of $F(u_\varepsilon)$.
\begin{example}
In particular cases, we find that $u_\varepsilon^{(F)}=F(u_\varepsilon)$ satisfies SPDEs.
\begin{itemize}
\item If $F(x)=x^{p}$ $(0<p\leq 1)$, then $u_\varepsilon^{(F)}(0,x)=u_0(x)^p$ and \begin{align*}
\partial_t u_\varepsilon^{(F)}(t)=\frac{1}{2}\Delta u_\varepsilon^{(F)}-\frac{(p-1)}{2p}\frac{|\nabla u_\varepsilon^{(F)}|^2}{u_\varepsilon^{(F)}}+\frac{\b_\varepsilon^2V(0)p(p-1)u_\varepsilon^{(F)}}{2\varepsilon^2}+\b_\varepsilon u_\varepsilon^{(F)}\dot\xi_\varepsilon.
\end{align*}
\item If $F(x)=\log x$ $($KPZ equation$)$, then $u_\varepsilon^{(F)}(0,x)=\log u_0(x)$ and \begin{align*}
\partial_t u_\varepsilon^{(F)}(t)=\frac{1}{2}\Delta u_\varepsilon^{(F)}+\frac{1}{2}{|\nabla u_\varepsilon^{(F)}|^2}-\frac{\b_\varepsilon^2V(0)}{2\varepsilon^2}+\b_\varepsilon \dot\xi_\varepsilon.
\end{align*}
\end{itemize}
\end{example}
We remark that $u_\varepsilon$ is a process indexed by $u_0$ and $\hat{\b}$ so we should write $u_\varepsilon=u_\varepsilon^{(\hat{\b},u_0)}$ and $u_\varepsilon^{(F)}=u_\varepsilon^{(F,\hat{\b},u_0)}$. However, we omit $\hat{\b}$ and $u_0$ for simplicity of notation when it is clear from the context.
}
We denote by $\mathcal C^\infty_c$ the set of infinitely differentiable, compactly supported functions on $\mathbb R^2$.
\begin{theorem}\label{th:EWlimit} Suppose $u_0^{(1)},\cdots,u_0^{(n)}\in \mathfrak{C}$, $\hat{\b}_1,\cdots,\hat{\b}_n\in (0,1)$ and $F_1,\cdots,F_n\in \mathfrak{F}$.
For $t_{1}, \cdots, t_{n}\geq 0$ and $f_{1},\cdots,f_{n}\in\mathcal C^\infty_c$, the following convergence holds jointly as $\varepsilon\to 0$,
\begin{align*}
&\b_\varepsilon^{-1}\int_{{\mathbb{R}}^2} f_{i}(x) \left(u_\varepsilon^{(F_i,u_0^{(i)},\hat{\b}_i)}(t_{i},x) - \mathbb E\left[{u}_\varepsilon^{(F_i,u_0^{(i)},\hat{\b}_i)}(t_{i},x)\right] \dd x\right)\\
&\hspace{4em} \stackrel{(d)}{\longrightarrow} \mathscr{U}_{t_i}(f_i,F_i,\hat{\b}_i,u_0^{(i)}) \;,
\end{align*}
where $\displaystyle \left\{\mathscr{U}_{t_i}(f_i,F_i,\hat{\b}_i,u_0^{(i)})\right\}_{i=1}^n$ is centered Gaussian random variables
with covariance {\begin{align}
& \frac{1}{1-\hat{\b}_i\hat{\beta}_j}\int_{0}^{{t_i}\wedge t_j}\dd \sigma\int\dd x\dd yf_i(x)f_j(y)
I^{(i)}(x)I^{(j)}(y
\int\dd z\rho_\sigma(x,z)\rho_\sigma(y,z)\bar{u}^{\ssup i}(t_i-\sigma,z)\bar{u}^{\ssup j}(t_j-\sigma,z),\label{eq:CovSt}
\end{align}
with \begin{align*}
I^{(t,F,\hat{\b},u_0)}(x)=I(x)=
\mathbb E\left[
F'\left(e^{X_{\hat{\b}}}
\bar{u}(t,x)\right)
e^{X_{\hat{\b}}}\right]=
\mathbb E\left[F'\left(e^{X_{\hat{\b}}+{\sigma^2(\hat{\b})}}
\bar{u}(t,x)\right)
\right],
\end{align*}
{ $X_{\hat{\b}}$ is a Gaussian random variable defined in Theorem \ref{thm:CSZ17b}, $\rho_t(x)=(2\pi t)^{-1}e^{-\frac{|x|^2}{2t}}$ is the heat kernel and $\rho_t(x,y)=\rho_t(x-y)$, and $\bar{u}(t,x)=\int \rho_t(x,y)u_0(y)\dd y$}. We write $I^{(i)}$ for $I^{(t_i,F_i,\hat{\b}_i,u_0^{(i)})}(x)$ to make notation simple.
}
\end{theorem}
\begin{rem}
The centered Gaussian field $\displaystyle \left\{\mathscr{U}_{t}(f,F,\hat{\b},u_0):t\in [0,\infty),F\in \mathfrak{F},\hat{\b}\in [0,1),u_0\in \mathfrak{C}\right\}$ with covariance \eqref{eq:CovSt} can be constructed explicitly. Let $\left\{\mathscr{V}^{(\hat{\b},u_0)}(t,x):(t,x)\in [0,\infty)\times {\mathbb{R}}^2\right\}$ be solutions of the following SPDE: $\mathscr{V}^{(\hat{\b},u_0)}(0,x)\equiv 0$ and \begin{align*}
\partial_t \mathscr{V}^{(\hat{\b},u_0)}(t,x)=\frac{1}{2}\Delta \mathscr{V}^{(\hat{\b},u_0)}(t,x)+{\bar{u}(t,x)}\dot{\xi}_{\hat{\b}}(t,x),
\end{align*}
where $\xi_{\hat{\b}}=\displaystyle \sum_{n=0}^\infty \hat{\b}\xi^{(n)}$ with an independent sequence time-space white noises $\{\xi^{(n)}\}$ for $\hat\b\in (0,1)$. We remark that $\xi_\b$ is a time-space white noise with strength $\displaystyle \frac{1}{1-\b^2}$ and $\displaystyle \mathbb E\left[\dot{\xi}_{{\b}}(t,x)\dot{\xi}_{\b'}(t',x')\right]=\frac{1}{1-\b\b'}\delta_{t,t'}\delta_{x,x'}$.
Then, by Duhamel's principle, $\mathscr{V}^{(\hat{\b}_i,u_0^{(i)})}$ is given by
\begin{align}\label{explicit construction}
\mathscr{V}^{(\hat{\b}^{(i)},u_0^{(i)})}(t,x)=\int_0^t \int_{{\mathbb{R}}^2}\rho_{t-s}(x,y)\bar{u}^{(i)}(s,y)\xi_{\hat{\b}_i}(\dd s,\dd y)
\end{align}
and the centered Gaussian field given by\begin{align*}
\mathscr{V}_{t_i}(f_i,F_i,\hat{\b}_i,u_0^{(i)})=\int_{{\mathbb{R}}^2}\dd xI^{\left(i\right)}(x)f_i(x)F_i'(\bar{u}^{(i)}(t_i,x))\mathscr{V}^{(\hat{\b}_i,u_0^{(i)})}(t,x)
\end{align*}
has the covariance structure \eqref{eq:CovSt}.
\end{rem}
\begin{rem} When $F$ is a power function or the logarithm, the limit
$\mathscr{U}^{(F,u_0,\hat{\b})}$ is a solution of an SPDE:
\begin{itemize}
\item If $F(x)=x^p$, then $\mathscr{U}^{(F,u_0,\hat{\b})}(0,x)\equiv 0$ and \begin{align*}
\partial_t \mathscr{U}=\frac{1}{2}\Delta \mathscr{U}+\frac{p(p-1)}{2}|\nabla \log \bar{u}|^2\mathscr{U}+(1-p)\nabla \log \bar{u}\cdot \nabla\mathscr{U}+\frac{\bar{u}(t,x)}{(1-\hat{\b}^2)^{p^2-p+1}}\dot{\xi}(t, x).
\end{align*}
\item If $F(x)=\log x$, then $\mathscr{U}^{(F,u_0,\hat{\b})}(0,x)\equiv 0$ and \begin{align*}
\partial_t \mathscr{U}=\frac{1}{2}\Delta \mathscr{U}+\nabla \log \bar{u}\cdot \nabla\mathscr{U}+\frac{\bar{u}(t,x)}{(1-\hat{\b}^2)}\dot{\xi}(t, x).
\end{align*}
\end{itemize}
\end{rem}
From the above remarks, we have the following.
\begin{cor}
Suppose $\hat{\beta}<1$. As $p\to 0$, $\mathscr{U}_{t}(f,x^p,\hat{\b},u_0)$ converges to $\mathscr{U}_{t}(f,\log{x},\hat{\b},u_0)$.
\end{cor}
\begin{rem}
In \cite{DGRZ20}, they study the fluctuations of the transformation $F(u_\varepsilon)$ for higher dimensional case $d\geq 3$ with $F$, its derivative and second derivative growing at most $x^{-p}+x^p$. They proved that there exists a constant $\beta_p$ such that the Gaussian fluctuation holds for $\hat{\beta}\in (0,\b_p)$. Our assumption on $F$ is slightly different from theirs but we can show the Gaussian fluctuations for all $\hat{\b}$ up to critical point.
\end{rem}
To analyze $u_\varepsilon$, we use the Feynman-Kac representation given in \cite[Section 2]{BC95} where they considered the case $d= 1$ but it is easy to be modified for $d\geq 2$:\begin{align*}
u_\varepsilon(t,x)={\mathrm E}_x\left[\exp\left(\b_\varepsilon\int_0^t \int_{{\mathbb{R}}^2}\phi_\varepsilon(B_s-y)\dot{\xi}(t-s,\dd y)\dd s-\frac{\b_\varepsilon^2tV(0)}{2\varepsilon^2}\right)u_0\left(B_t\right)\right],
\end{align*}
where we denote by ${\mathrm P}_x$ and ${\mathrm E}_{{ x}}$ the law and the expectation with respect to two dimensional Brownian motion $B=\{B_t\}_{t\geq 0}$ starting from $x$.
Due to the time-reverse invariance and scale invariance of time-space white noise and the scaling invariance of Brownian motion, $\varepsilon B_{\varepsilon^{-2} s} \overset{d}{=} B_s$, $\{u_\varepsilon(t,x):x\in\mathbb{R}^2\}$ has the same distribution as
\begin{align}
&{\mathrm E}_x\left[\exp\left(\b_\varepsilon\int_0^t \int_{{\mathbb{R}}^2}\phi_\varepsilon(B_s-y)\dot{\xi}(s,\dd y)\dd s-\frac{\b_\varepsilon^2tV(0)}{2\varepsilon^2}\right)u_0\left(B_t\right)\right]\notag\\
&={\mathrm E}_{\frac{x}{\varepsilon}}\left[\exp\left(\b_\varepsilon\int_0^{\frac{t}{\varepsilon^2}} \int_{{\mathbb{R}}^2}\phi\left(B_s-\frac{y}{\varepsilon}\right)\dot{\xi}(s,\dd y)\dd s-\frac{\b_\varepsilon^2tV(0)}{2\varepsilon^2}\right)u_0\left({ \varepsilon}\,B_{\frac{t}{\varepsilon^2}}\right)\right],\label{eq:FKform}
\end{align}
where $(B_s)_{s\geq 0}$ is a Brownian motion path and ${\mathrm E}_x$ the expectation associated to Brownian motion started at $x\in\mathbb R^2$, $\b\geq 0$. In particular, for the flat initial condition, $u_\varepsilon$ has the same distribution as {\it partition function} ${\bf{\mathcal{Z}}}_{\frac{t}{\varepsilon^2}}\left(\frac{x}{\varepsilon}\right)$ of continuum directed polymers, where ${\bf{\mathcal{Z}}}_t(x)$ is given by \begin{align*}
{\bf{\mathcal{Z}}}_t(x)&={\mathrm E}_x\left[\Phi_t^{\b_\varepsilon}\right]\\
&={\mathrm E}_{{x}}\left[\exp\left(\b_\varepsilon\int_0^{t} \int_{{\mathbb{R}}^2}\phi\left(B_s-{y}\right)\dot{\xi}(s,\dd y)\dd s-\frac{\b_\varepsilon^2tV(0)}{2}\right)\right],
\end{align*}
where \begin{align*}
\Phi_t^\b=\Phi_t^\b(B,\xi):=\exp\left(\beta \int_0^t \int_{{\mathbb R}^2} \phi(y-B_s) \ \xi(\dd s,\dd y)-\frac{\b^2 V(0)t}{2}\right)
\end{align*}
for $t\geq 0$, $\b\in (0,\infty)$. Thus, we can reduce the problem on the laws of $u_\varepsilon$ to the {partition function} of continuum directed polymers. Such connections between SHE (and KPZ equation) and directed polymers have been already pointed out in \cite{KPZ} and used in a lot of researches on SHE and KPZ equation \cite{BC95,BG97,MSZ16,GRZ18,MU17,DGRZ20,CSZ17a,CSZ17b,CSZ19a,CSZ19b,CSZ20,CCM20,CCM19b,CNN20,LZ20}.
\begin{rem} We give our contribution to the problem shortly. Edward-Wilkinson type fluctuations for KPZ equation for $d=2$ have been obtained in \cite{CSZ20} and \cite{G20} with the flat initial condition. On the other hand, we obtain the Gaussian fluctuations for the general initial conditions and multi-dimensional parameters. Also, our proof uses a less technical method, ``{\it martingale CLT}" (see Theorem \ref{thm:JS}) { via It\^o's lemma and homogenization argument.}, with onerous calculation.
In \cite{CSZ20}, the problem was reduced to the case in the solution of SHE via approximating $\log u_\varepsilon$ by $``u_\varepsilon-1"$. In \cite{G20}, Gaussian fluctuation was obtained by Malliavin calculus and the second order Poincar\'e inequality.
\end{rem}
\begin{rem} The Gaussian fluctuations for partition functions \cite{LZ20} and solutions of SHE \cite{MSZ16,GRZ18,DGRZ18b,CNN20} and KPZ equation \cite{MU17,DGRZ20,CNN20} in $d\geq 3$ have been proved as well as two dimensional case, where the disorder strength is given by $\hat{\b}\varepsilon^{\frac{d-2}{2}}$ for $d\geq 3$. Also, the Gaussian fluctuations for a nonlinear stochastic heat equation with Gaussian multiplicative noise that is white in time and smooth in space \cite{GL20} and the counterpart for $d=2$ is stated in \cite{DG20} without detailed proof.
\end{rem}
\textbf{Note:} Throughout the paper and if clear from the context, the constant $C$ that appears in successive upper-bounds may take different values.
\textbf{Organization of the article} The main idea of Gaussian fluctuation is the same as in \cite{CNN20}. Section \ref{sec2} is devoted to proving key properties of partition functions of directed polymers, $L^2$-boundedness, boundedness of negative moments, and local limit theorem. Section \ref{sec:proofEWKPZ} is dedicated to the proof of Theorem \ref{th:EWlimit}. In subsection \ref{IdeaG}, we give a rough proof strategy and explain a heuristic idea of Gaussian fluctuation. The rigorous proof starts from subsection \ref{sub:4.2}.
\iffalse{
The stochastic heat equation (with multiplicative white noise) is formally given by the SPDE
\begin{equation} \label{eq:formalSHE}
\frac{\partial}{\partial t} u(t,x) = \frac12 \Delta u(t,x) +
\b u(t,x) \xi(t,x),
\end{equation}
for $t\geq 0$ and $x\in\mathbb R^d$ and
where $\xi$ is a space-time white noise on $\mathbb R_+ \times \mathbb R ^d$. Let us assume that $\xi$ is defined on a probability space $(\Omega,\mathcal F,{\mathbb P})$ and let $\mathbb E$ denote the expectation with respect to ${\mathbb P}$.
The equation is closely related to the KPZ equation which is formally described by
\begin{equation} \label{eq:formalKPZ}
\frac{\partial}{\partial t} h = \frac12 \Delta h + \frac12 |\nabla h |^2+
\b \xi,
\end{equation}
and the two equations are related through the so-called Hopf-Cole transform
\begin{equation}
h=\log u.
\end{equation}
In dimension $d=1$, much work has been concentrated in the past years in order to make sense of equation \eqref{eq:formalKPZ}, with the notable contributions \cite{BG97,H13}. One of the difficulties coming when trying to do so is that the derivative $\partial_x h$ fails to be a function, and so one has to make sense of the square of a distribution. When $d\geq 3$, the solution $u$ should also fail to be a function, hence giving sense to the product $u\xi$ is also an issue. In this case, both SPDEs fall into the so-called \emph{super-critical} dimensions classification and do not enter in the framework of the recent theories \cite{H14,GIP15,KM17,GP18}.
}\fi
\iffalse{
\subsection{The weak disorder and the $L^2$-region}
It is a standard result of the polymer litterature that the following $0$-$1$ law holds:
\begin{equation}
\text{either:}\quad (i)\ {\bf{\mathcal{Z}}}_T \to 0\quad {\mathbb P}\text{-a.s,} \quad \text{or:} \quad (ii) \,{\bf{\mathcal{Z}}}_T \to {\bf{\mathcal{Z}}}_\infty \text{ where } {\bf{\mathcal{Z}}}_\infty> 0\quad {\mathbb P}\text{-a.s,}
\end{equation}
and that there is a critical parameter $\b_c\geq 0$ such that
\begin{equation}
{\bf{\mathcal{Z}}}_\infty = 0 \text{ a.s. } \text{ if }\b>\b_c \quad \text{and} \quad
{\bf{\mathcal{Z}}}_\infty > 0 \text{ a.s. } \text{ if } \b<\b_c,
\end{equation}
and $\b_c>0$ when $d\geq 3$. Moreover, (ii) holds if and only if $({\bf{\mathcal{Z}}}_T)_{T\geq 0}$ is uniformly integrable, see\ \cite{MSZ16,CY06}.
When situation (i) holds, we say that the system is in the \emph{strong disorder} regime, while if (ii) holds, we say that it is in the \emph{weak disorder} regime. Localization properties for the polymer model and the solution $u_\varepsilon$ of SHE have been shown to hold inside the strong disorder region, see \cite{BC20,C17,BM19}. On the other hand, the polymer is diffusive in the full weak disorder region \cite{CY06} (see Appendix \ref{app:LLNDirac} in the continuous case).
Another region of interest is the $L^2$-region $[0,\b_{L^2})$, where
\begin{equation} \label{eq:defBeta2}
\b_{L^2} = \sup\{\b\geq 0,\, \sup\nolimits_t\mathbb E{\bf{\mathcal{Z}}}_t^2 = \mathbb E[{\bf{\mathcal{Z}}}_\infty^2] < \infty\},
\end{equation}
satisfies $\b_{L^2}\leq \b_c$. In fact, it is widely believed that $\b_{L^2} < \b_c$, this strict inequality being proved for the discrete polymer when $d\geq 3$ \cite{BGH11,BT10,BS10,BS11} (see in particular \cite[Section 1.4]{BS10} for $d=3,4$). The $L^2$-region has been first introduced in \cite{B89,IS88} to show that diffusivity could happen for the polymer for small $\b$. It has the particular interest of allowing second moment computations that are not accessible in the weak disorder region part where $\b \geq \b_{L^2}$. The region has also been recently involved in the study of the speed of convergence and fluctuations of the polymer partition function \cite{CL17,CN19}, as well as for asymptotic properties of $u_\varepsilon$ and $h_\varepsilon$ that we describe now.
}\fi
\iffalse{\subsection{Presentation of the results}
\iffalse{
It was first observed in \cite{MSZ16} that when $\b<\b_{L^2}$, the following law of large numbers holds if $u_0= u_\varepsilon(0,\cdot)$ is regular enough:
\begin{equation} \label{eq:LLN}
\int_{\mathbb R ^d} u_\varepsilon(t,x) f(x) \dd x \stackrel{\IP}{\longrightarrow} \int_{\mathbb R ^d} \bar{u}(t,x) f(x) \dd x,
\end{equation}
as $\varepsilon\to 0$ for any test function $f\in \mathcal{C}_c^\infty$, where
\begin{equation} \label{eq:Def_ubar}
\partial_t \bar{u}(t,x) = \frac{1}{2} \Delta \bar{u}(t,x),\qquad \bar{u}(0,\cdot) = u_0.
\end{equation}
The reason behind \eqref{eq:LLN} is that $u_\varepsilon(t,x)$ and $u_\varepsilon(t,y)$ become asymptotically independent for $x\neq y$ when $\varepsilon\to 0$, which is well captured by a covariance computation when $\b<\b_{L^2}$.
Our first main result (cf. Section \ref{sec:LLNstatements}) is to extend \eqref{eq:LLN} to the full \emph{weak disorder} region.
Extending \eqref{eq:LLN} to the full weak disorder region requires more sophisticated arguments than a covariance computation, because the second moments blow up when $\b\geq \b_{L^2}$. To get around this issue, we rely on both the central limit theorem for polymer paths \cite{CY06} (which was shown in the full weak disorder region) and a homogenization argument as in \cite{CN19} that builds on the uniform integrability of $({\bf{\mathcal{Z}}}_T)_{T\geq 0}$. Besides, we obtain a similar convergence to \eqref{eq:LLN} in the case of Dirac initial condition $u_0=\delta_{x_0}$ (although the limit will be purely random), see Theorem \ref{LNDirac}.
}\fi
Fluctuations in the law of large numbers \eqref{eq:LLN} have been studied in \cite{GRZ18} in the case where the noise can also be colored in time. The authors showed that for $\b$ sufficiently small (namely for $\b<\b_0$ with $\b_0<\b_{L^2}$) and for all $u_0\in\mathcal C_b(\mathbb R^d)$,
\begin{equation} \label{eq:CV_toEW_intro}
\varepsilon^{1-\frac d 2}\int_{\rd} f(x) \left(u_\varepsilon(t,x) - \bar{u}(t,x)\right) \dd x \stackrel{(d)}{\longrightarrow} \int_{\rd} f(x)\, \mathscr{U}_1(t,x) \dd x\;,
\end{equation}
with $\mathscr U_1$ solving
the stochastic heat equation with additive noise (also called the Edwards-Wilkinson (EW) equation):
\begin{equation}\label{eq:EW_GRZ}
\partial_t \mathscr{U}_1(t,x)= \frac 12 \Delta \mathscr{U}_1(t,x)+ \bar{u}(t,x) {\gamma}(\beta) \xi(t,x),\qquad \mathscr{U}_1(0,x)=0,
\end{equation}
where ${\gamma}(\beta)$ is defined in \eqref{def-sigma-beta}. To prove this convergence, the authors relied on Malliavin calculus techniques.
Our second main result (c.f.\ Section \ref{sec:ResultsSHE}) is a proof that the Edwards-Wilkinson regime described in \eqref{eq:CV_toEW_intro} can be extended to the region $\b\in (0,\b_{L^2})$ with joint in time convergence. Note that the parameter ${\gamma}(\beta)$ is finite in $(0,\b_{L^2})$ and blows up precisely at $\b_{L^2}$, indicating that \eqref{eq:CV_toEW_intro} should indeed hold up to that point (and not above), but the proof in \cite{GRZ18} does not carry up until there. To reach $\b_{L^2}$, we follow here a distinct strategy of proof that builds on stochastic calculus and the martingale central limit theorem, in the same spirit as in \cite{CN19,CCM19b,CL17,CN95}. Like in \cite{CN19}, the additional ingredients that we use to control the quadratic variations and the cross-brackets of the martingales are the local limit theorem for polymers (Theorem \ref{th:errorTermLLT} in the appendix, see also \cite{CCM20,V06,Si95}) and a homogenization argument, c.f.\ Section \ref{sec:heuristicsEW} for a quick description of the proof. We believe that in addition to leading us to $\b_{L^2}$, this approach to the problem has the advantage of being quite straightforward and simple and that it should be applicable in different contexts (for example, discrete structures). Furthermore, the method gives directly multi-dimensional in time convergence in \eqref{eq:CV_toEW_intro}, which has not been considered in \cite{GRZ18}.
Let us turn to results on the KPZ equations. In \cite{DGRZ20}, the authors have considered the \emph{flat initial condition} $h_0\equiv 0$ and proved that there is some $\b_0<\b_{L^2}$ such that for all $\b<\b_0$,
\begin{equation} \label{eq:EWKPZintro}
\varepsilon^{1-\frac d 2}\int_{\rd} f(x) \left(h_\varepsilon(t,x)-\mathbb E[h_\varepsilon(t,x)]\right) \dd x \stackrel{(d)}{\longrightarrow} \int_{\rd} f(x)\, \mathscr{U}_{1}(t,x) \dd x,
\end{equation}
with $\bar u(t,x) = 1$ in the definition of $\mathscr{U}_{1}(t,x)$. See also \cite{MU17} for another approach to the problem. Both of these proofs are restricted to small $\b$ and some additional arguments have to be brought to go up to $\b_{L^2}$. In particular, higher moments for the partition function (which do not hold when approaching $\b_{L^2}$) are technically required in the proof of \cite{DGRZ20}, while the renormalization approach in \cite{MU17} is restricted by essence to a small $\b$ region.
We prove in this paper (c.f.\ Section \ref{sec:resultsOnKPZ}) that \eqref{eq:EWKPZintro} extends to $(0,\b_{L^2})$ with joint in time convergence and general continuous initial conditions. Note that in \cite{MU17}, the authors consider initial conditions at a macroscopic level (i.e.,\ $u_\varepsilon(0,x) = u_0(\varepsilon^{-1}x)$), while ours are at a microscopic level (i.e.,\ $u_\varepsilon(0,x) = u_0(x)$) and contrary to our result (see Theorem \ref{th:EWlimitKPZ}), the dependence in the initial condition disappears in the limit in their case. Our proof strategy is based on the same martingale method as for SHE and relies on It\^o's formula. The proof requires significant additional work compared to the SHE case. One of the difficulties to overcome is to deal with ratios of partition functions. We do so by relying on a concentration inequality for ${\bf{\mathcal{Z}}}_T$ proven in \cite{CCM20} (see also \cite{CH02}) and precise estimates in the local limit theorem for polymers that we obtain in Theorem \ref{th:errorTermLLT}. Another issue to tackle is to control the It\^o correction and show that it satisfies some homogenization property by studying some 4-th moments.
As we just mentioned, we needed some precise estimates (particularly a speed of convergence) on the local limit theorem for polymers that were not accessible via the pre-existing statements \cite{CCM20,V06 Si95}. Our Theorem \ref{th:errorTermLLT} improves the previous ones on the following points: (i) it is not restricted to the diffusive scale between the starting and ending points. In particular, it shows that the error still vanishes up to a linear scale; (ii) it gives a speed of convergence with exponents that do not depend on $\b$. We think that these improvements are interesting on their own and may be useful for other purposes.
}\fi
\iffalse{
\subsection{Further results on convergence to stationary Edwards-Wikinson and to the Gaussian free field}
As additional results, we show (stationary) Edwards-Wilkinson asymptotics for the \emph{stationary solutions} of the regularized SHE and KPZ (cf.\ Theorem \ref{th:CV_stationarySHE} and Theorem \ref{th:CV_stationaryKPZ}). These considerations have a counterpart in the polymer framework, namely that the diffusively rescaled and normalized (infinite-horizon) partition function ${\bf{\mathcal{Z}}}_\infty(x)$ and log-partition function $\log {\bf{\mathcal{Z}}}_\infty(x)$ converge to a Gaussian Free Field (cf.\ Corollary \ref{th:PolymerCV} and Corollary \ref{th:LogPolymerCV}), which we believe are an interesting results on their own.
The notion of stationary solutions of the regularized equations can be defined as follows. If $\xi\to \mathcal V(\xi)$ denotes a measurable function such that $\mathcal V(\xi) = {\bf{\mathcal{Z}}}_\infty(0,\xi)$, then (similarly to \eqref{eq:uZ})
\begin{equation} \label{eq:defustat}
\quad u^{\rm stat}_\varepsilon(t,x) = \mathcal V (\xi^{\ssup{\varepsilon,t,x}}),
\end{equation}
defines a stationary solution of the SPDE in \eqref{eq:SHEintro}, where by stationary we mean that the distribution of $u^{\rm stat}_\varepsilon(t,x)$ is invariant by time-space translation (see \cite{CCM20,CCM19b,DGRZ18b}).
The function $u^{\rm stat}_\varepsilon$ plays a central role in the \emph{pointwise} asymptotic approximation of $u_{\varepsilon}$ (in contrast to \eqref{eq:LLN} which is averaged), at first order, for diffent types of initial conditions $u_0$ (continous and bounded, Dirac). For example, for any continous and bounded $u_0$, it holds that for all $\b<\b_{L^2}$, as $\varepsilon \to 0$,
\[\forall t>0,\,x\in\mathbb R^d, \quad u_\varepsilon(t,x) - \bar{u}(t,x)u^{\rm stat}_\varepsilon(t,x) \stackrel{L^2}{\longrightarrow} 0,\]
see \cite{DGRZ18b,CCM20} for further discussion on this matter.
In the KPZ case, a stationary solution to the SPDE in \eqref{eq:KPZe} is given by the Hopf-Cole transform
\begin{equation} \label{eq:defhstat}
h^{\rm stat}_\varepsilon = \log u^{\rm stat}_\varepsilon,
\end{equation}
and as for SHE, $h^{\rm stat}_\varepsilon$ is a central object in the \emph{pointwise} approximation (of first order) of $h_\varepsilon(t,x)$ as $\varepsilon\to 0$ for a variety of initial conditions, c.f.\ \cite{CCM20,DGRZ18b}. In particular, when $\b<\b_c$ and $h_0=0$,
\begin{equation} \label{eq:pointwiseApproxKPZ}
\forall t>0,\, x\in\mathbb R^d, \quad h_\varepsilon(t,x) - h^{\rm stat}_\varepsilon(t,x) \stackrel{\IP}{\longrightarrow} 0.
\end{equation}
The (pointwise) fluctuations in the convergence \eqref{eq:pointwiseApproxKPZ} have been studied in \cite{CCM19b} deep in the $L^2$-region, and the limit of the correctly rescaled fluctuations involves a convolution of the heat kernel with a Gaussian free field. By a heuristic combination of this result and of \eqref{eq:EWKPZintro}, it was conjectured therein that convergence \eqref{eq:statKPZresult} below should hold in the $L^2$-region. Our method enables us to prove this property (Theorem \ref{th:CV_stationaryKPZ}).
}\fi
\section{Some Estimates for Partition Functions}\label{sec2}
{ In this section, we discuss some properties of partition functions of directed polymers in random environment.}
Hereafter, we set
\begin{equation}\label{Def: beta gamma T}
\begin{split}
&T=T_\varepsilon:=\varepsilon^{-2}, \\
&\b=\b_\varepsilon:=\hat{\b}\sqrt{\frac{2\pi}{\log \varepsilon^{-1}}}=\hat{\b}\sqrt{\frac{4\pi}{\log T}}, \text{ and } \\
&\gamma=\gamma_\varepsilon=\hat{\gamma}\sqrt{\frac{2\pi}{\log \varepsilon^{-1}}}=\hat{\gamma}\sqrt{\frac{4\pi}{\log T}}.
\end{split}
\end{equation}
{ Throughout the paper, we write the subscript $\varepsilon$ in $T_\varepsilon$ and $\b_\varepsilon$ in each statement to emphasize its dependence but we often omit the subscript $\varepsilon$ in the proofs for simplicity.}
\subsection{$L^p$-bound of partition functions}
First, we remark that for $x,y\in {\mathbb{R}}^2$
\begin{align}
\mathbb E \left[{\mathrm E}_x[\Phi^\b_{tT}]{\mathrm E}_y[\Phi^\gamma_{tT}]\right
=1+\sum_{n=1}^\infty \b^{n}\gamma^n\int_{0<s_1<\cdots<s_n<tT}\int_{\left({\mathbb{R}}^2\right)^n}\prod_{i=1}^n\left(V(\sqrt{2}x_i)\rho_{s_i-s_{i-1}}(x_{i-1},x_{i})\right)\dd \mathbf{s} \dd\mathbf{x}\label{eq:L2decomn}
\end{align}
where $\mathbb E$ and ${\mathbb P}$ denote the expectations and probability with respect to the white noise $\xi$ and we set $x_0=\frac{x-y}{\sqrt{2}}$, $s_0=0$ and $\dd\mathbf{s}=\dd s_1\cdots \dd s_n$, $\dd \mathbf{x}=\dd x_1\cdots \dd x_n$.
This representation is obtained from the general property of the white noise:
\begin{align}
\mathbb E\left[ \exp{\left(\int_{0}^t \int_{{\mathbb{R}}^2} f(t,x)\xi(\dd s, \dd x) \right)}\right]= \exp{\left(\frac{1}{2}\int_{0}^t\int_{{\mathbb{R}}^2} f(t,x)^2 \dd s \dd x \right)}.\label{eq:L2white}
\end{align}
Indeed, we have
\aln{
\mathbb E \left[{\mathrm E}_x[\Phi^\b_{tT}]{\mathrm E}_y[\Phi^\gamma_{tT}]\right]&= {\mathrm E}_{x}\otimes {\mathrm E}_{y}\left[\exp\left(\int_0^t ( \b\phi(B_s-y)+\gamma\phi(\tilde{B}_s-y))^2 \dd s \dd y - \frac{(\b^2+\gamma^2) V(0)t}{2}\right)\right] \notag\\
&={\mathrm E}_x\otimes {\mathrm E}_y\left[\exp\left(\b\gamma\int_0^t V(B_s-\tilde{B}_s)\dd s\right)\right]\notag\\
&= {\mathrm E}_{\frac{x-y}{\sqrt{2}}}\left[\exp\left(\b \gamma\int_0^t V(\sqrt{2} B_s)\dd s\right)\right],\notag
}
and the Taylor expansion $e^x=\sum_{n=0}^\infty \frac{x^n}{n!}$ gives \eqref{eq:L2decomn
\begin{comment}
\al{
&\quad 1+\sum_{n=1}^\infty \frac{(\b\gamma)^{n}}{n!}{\mathrm E}_{\frac{x-y}{\sqrt{2}}}\left[\left(\int_0^T V(\sqrt{2}B_s)\dd s\right)^n\right]\\
&=1+\sum_{n=1}^\infty \b^{n} \gamma^n{\mathrm E}_{x_0}\left[\int_{0<s_1<\cdots<s_n<T} V(\sqrt{2}B_s)\dd s_1\cdots \dd s_n\right]\\
&=1+\sum_{n=1}^\infty \b^{n}\gamma^n\int_{0<s_1<\cdots<s_n<tT}\int_{\left({\mathbb{R}}^2\right)^n}\prod_{i=1}^n\left(V(\sqrt{2}x_i)\rho_{s_i-s_{i-1}}(x_{i-1},x_{i})\right)\dd \mathbf{s} \dd\mathbf{x},
}
where $\mathbb E$ and ${\mathbb P}$ denote the expectations and probability with respect to the white noise $\xi$ and we set $x_0=\frac{x-y}{\sqrt{2}}$ and $\dd\mathbf{s}=\dd s_1\cdots \dd s_n$, $\dd \mathbf{x}=\dd x_1\cdots \dd x_n$.
\end{comment}
\begin{lemma}\label{lem:L2bdd}
Suppose $\hat{\b},\hat{\gamma}\in (0,1)$ and fix $t>0$. Then,
\begin{align}
&\lim_{{ \varepsilon\to 0}}\mathbb E \left[{\mathrm E}[\Phi_{tT_\varepsilon}^{{ \b_\varepsilon}}]{\mathrm E}[\Phi_{tT_\varepsilon}^{{\gamma_\varepsilon}}]\right]=\frac{1}{1-\hat{\b}\hat{\gamma}},\label{eq:L2bdd}\\
&\sup_{\varepsilon\leq 1}\sup_{s\leq t T_\varepsilon}\mathbb E \left[{\mathrm E}_{0,0}^{s,0}\left[\Phi^{{ \b_\varepsilon}}_s\right]^2\right]<\infty,\label{eq:L2ppbdd}
\end{align}
where ${\mathrm P}_{0,x}^{t,y}$ and ${\mathrm E}_{0,x}^{t,y},$ denote the probability measure and expectation of the Brownian bridge from $(0,x)$ to $(t,y)$ in ${\mathbb{R}}^2$.
\end{lemma}
\begin{rem}
The proof of Theorem 2.15 in \cite{CSZ17b}, where \eqref{eq:L2bdd} for $\b_\varepsilon=\gamma_\varepsilon$ was proved by reducing the problem to discrete directed polymers in random environment, can be mollified for $\b_\varepsilon\not=\gamma_\varepsilon$, but we will give a direct proof in this paper.
\end{rem}
\begin{rem}
We call \begin{align*}
{\bf{\mathcal{Z}}}_{0,x}^{t,y}={{\bf{\mathcal{Z}}}_{0,x;t,y}^{(\b)}:=}{\mathrm E}_{0,x}^{t,y}\left[\Phi_t^{\b}\right]
\end{align*}
the point-to-point partition function of continuum directed polymers.
\end{rem}
\begin{proof}[Proof of \eqref{eq:L2bdd}]
We have from
\eqref{eq:L2decomn}
\begin{align}
&\mathbb E \left[{\mathrm E}\left[\Phi_{tT}^\b\right]{\mathrm E}\Big[\Phi_{tT}^\gamma\Big]\right]= 1+\sum_{n=1}^\infty \b^{n}\gamma^n\int_{0<s_1<\cdots<s_n<tT}\int_{\left({\mathbb{R}}^2\right)^n}\prod_{i=1}^n\left(V(\sqrt{2}x_i)\rho_{s_i-s_{i-1}}(x_{i-1},x_{i})\right)\dd \mathbf{s} \dd\mathbf{x},\label{white noise exp}
\end{align}
with $x_0=0$.
We first consider the upper bound: $$\varlimsup_{\varepsilon\to 0}\mathbb E \left[{\mathrm E}[\Phi_{tT_\varepsilon}^\b]{\mathrm E}[\Phi_{tT}^\gamma]\right]\leq \frac{1}{1-\hat{\b}\hat{\gamma}}.$$
Let us consider the function
\ben{\label{Def: rs}
r_s=\sup_{x\in {\mathbb{R}}^2}\int_{{\mathbb{R}}^2}V(\sqrt{2}y)\rho_{s}(x,y)\dd y\geq 0.
}Since $\int_{\mathbb{R}} \rho_s(x,y)\dd y=1$, $\sup_{s> 0} |r_s|\leq \|V\|_{\infty}$. Moreover, using $\int V(x) \dd x=1$, we obtain
\begin{align}
r_s&= \frac{1}{4\pi s } \sup_{x\in {\mathbb{R}}^2}\int_{{\mathbb{R}}^2}V(y) e^{-\frac{|x-y|^2}{4s}}\dd y\leq \frac{1}{4\pi s } \sup_{x\in {\mathbb{R}}^2} \int_{{\mathbb{R}}} V(y)\dd y= \frac{1}{4\pi s }.\label{eq:rorder}
\end{align}
}
Hence, \eqref{white noise exp} is bounded from above by
\begin{align*}
1+\sum_{n=1}^\infty \b^{n}\gamma^n \left(\int_0^{tT}r_s\dd s\right)^{n}\\
&\leq 1+\sum_{n=1}^\infty \b^{n}\gamma^n \left(\frac{\log (tT)}{4\pi }+\|V\|_{\infty}\right)^{n}=1+\sum_{n=1}^\infty \hat{\b}^{n}\hat{\gamma}^n \left(1+\frac{\log{t}+4\pi \|V\|_{\infty}}{\log{T} }\right)^{n}\notag\\
& \to \sum_{n=0}^\infty (\hat{\b}\hat{\gamma})^n= \frac{1}{1-\hat{\beta}\hat{\gamma}},
\end{align*}
as $\varepsilon\to 0$, where the convergence is absolute since $\hat{\b}\hat{\gamma}<1$.
Next, we consider the lower bound. Let
$$\mathbf{T}_n=\{0<s_1<\cdots<s_n<tT,\, s_i-s_{i-1}>1,\,\forall i\in\{1,\cdots, n\} \}.$$ Then, since each term is non-negative, it is enough to show for fixed $L\in{\mathbb{N}}$,
\begin{align}
\varliminf_{\varepsilon\to 0} \left(1+\sum_{n=1}^L \b^{n}\gamma^n\int_{\mathbf{T}_n}\int_{\left({\mathbb{R}}^2\right)^n}\prod_{i=1}^n\left(V(\sqrt{2}x_i)\rho_{s_i-s_{i-1}}(x_{i-1},x_{i})\right)\dd \mathbf{s} \dd\mathbf{x}\right)\geq \sum_{n=0}^L (\hat{\b}\hat{\gamma})^n.\label{eq:l2tildeT}
\end{align}
For fixed $n\in{\mathbb{N}}$, we can find that
\begin{align}
&\int_{\mathbf{T}_n}\int_{\left({\mathbb{R}}^2\right)^n}\prod_{i=1}^n\left(V(\sqrt{2}x_i)\rho_{s_{i}-s_{i-1}}(x_{i-1},x_{i})\right)\dd \mathbf{s} \dd\mathbf{x}\notag\\
&=\int_{\mathbf{T}_n}\int_{\left({\mathbb{R}}^2\right)^n}\prod_{i=1}^n\left(\frac{1}{2\pi (s_i-s_{i-1})}V(\sqrt{2}x_i)-\bar{r}_{s_i-s_{i-1}}(x_{i-1},x_{i})\right)\dd \mathbf{s}\dd \mathbf{x}\notag\\
&=\int_{\mathbf{T}_n}\prod_{i=1}^n \frac{1}{4\pi (s_i-s_{i-1})}\dd \mathbf{s}+A_\varepsilon^n,\label{poppoppo}
\end{align}
where
\begin{align*}
\bar{r}_s(x,y)&=\frac{1}{2\pi s}V(\sqrt{2}y)-V(\sqrt{2}y)\rho_{s}(x,y)\\
&=\frac{1}{2\pi s }\left(1-\exp\left(-\frac{|y-x|^2}{2s}\right)\right)V(\sqrt{2}y)\geq 0,
\end{align*}
\begin{align*}
A_\varepsilon^n&= \sum_{k=1}^n (-1)^k \sum_{j_1<j_2<\cdots<j_k}\int_{\mathbf{T}_n}\int_{\left({\mathbb{R}}^2\right)^n}\prod_{i\not=j_1,\cdots,j_k}\frac{V(\sqrt{2}x_i)}{4\pi (s_i-s_{i-1})} \prod_{j=j_1,\cdots,j_k}\bar{r}_{s_{j}-s_{j-1}}(x_{j-1},x_j)\dd \mathbf{s} \dd \mathbf{x}.
\end{align*}
Let $D_V=\overline{\{x\in {\mathbb{R}}^d:~V(\sqrt{2}x)\neq 0\}}$, which is compact. Note that
\al{
\sup_{x\in D_V}\int_{{\mathbb{R}}^2}\bar{r}_s(x,y)\dd y&= (2\pi s)^{-1} \sup_{x\in D_V}\int_{{\mathbb{R}}^2} \left(1-\exp\left(-\frac{|y-x|^2}{2s}\right)\right)V(\sqrt{2}y) \dd y\\
&\leq (2\pi s)^{-1} |D_V| \sup_{x,y \in D_V} \left(1-\exp\left(-\frac{|y-x|^2}{2s}\right)\right) \\
&\leq C(s^{-2}\wedge 1),
}
with some $C=C(V)\geq 1\lor \|V\|_\infty$. In particular, we have \begin{align*}
|A_\varepsilon^n|
&\leq C^{n+1}\sum_{k=1}^n \sum_{j_1<j_2<\cdots<j_k}\int_{\mathbf{T}_n}\int_{\left({\mathbb{R}}^2\right)^n}\prod_{i\not=j_1,\cdots,j_k}\frac{\dd s_i}{s_i-s_{i-1}} \\
&\leq C^{n+1} \sum_{k=1}^n n^k \left(\log{( { tT})}\right)^{n-k}\\
&\leq (Cn)^{n+1}(\log{( { tT})})^{n-1}.
\end{align*}
Thus, we have for any fixed $L>0$\begin{align}
\sum_{n=1}^L \beta^{n}\gamma^nA_\varepsilon^n= \sum_{n=1}^L \hat{\beta}^{n}\hat{\gamma}^n \left(\frac{4\pi}{\log{T}}\right)^n A_\varepsilon^n \to 0,\label{eq:L2bdderror}
\end{align}
as $\varepsilon \to 0$. Also,
\begin{align*}
\int_{\mathbf{T}_n}\prod_{i=1}^n \frac{1}{4\pi (s_i-s_{i-1})}\dd \mathbf{s}\geq \left(\int_{1}^{ \frac{tT}{n}}\frac{1}{4\pi s}\dd s\right)^n=\left(\frac{\log{(tT/n)}}{4\pi}\right)^n,
\end{align*}
and hence we have \begin{align}
\varliminf_{\varepsilon\to 0} \sum_{n=1}^L \b^{n}\gamma^n\int_{\mathbf{T}_n}\prod_{i=1}^n \frac{1}{4\pi (s_i-s_{i-1})}\dd \mathbf{s}\geq \sum_{n=1}^L \hat{\b}^n\hat{\gamma}^n.\label{eq:L2bddlead}
\end{align}
Then, \eqref{poppoppo}, \eqref{eq:L2bdderror} and \eqref{eq:L2bddlead} yield \eqref{eq:l2tildeT}.
\end{proof}
\begin{proof}[Proof of \eqref{eq:L2ppbdd}]
We obtain by the same manner as \eqref{eq:L2decomn} that
\begin{align}
&\mathbb E \left[{\mathrm E}_{0,0}^{tT,0}\left[\Phi^\b_{tT}\right]^2\right]= {\mathrm E}_{0,0}^{tT,0}\left[ \exp{\left( \b^2\int_0^{t T} V(\sqrt 2 B_u) \dd u\right)}\right] \notag\\
&=1+\sum_{n=1}^\infty \b^{2n}\int_{0<s_1<\cdots<s_n<s}\int_{\left({\mathbb{R}}^2\right)^n}\left(\prod_{i=1}^nV(\sqrt{2}x_i)\rho_{s_{i}-s_{i-1}}(x_{i-1},x_{i})\right)\frac{\rho_{s-s_n}(x_n)}{\rho_{tT}(0)}\dd \mathbf{s}\dd \mathbf{x},\label{eq:L2ppdecom}
\end{align}
where we use the orthogonal transformation invariance of Brownian bridges. We have for $s,t>0$, by Markov property of Brownian motions,
\al{
\int_{{\mathbb{R}}}\rho_{s}(x,y) \rho_t(y) V(\sqrt{2} y)\dd y&\leq \|V\|_\infty \int_{{\mathbb{R}}}\rho_{s}(x,y) \rho_t(y) \dd y\leq \|V\|_{\infty} \rho_{s+t}(x),
}
and by $\int_{\mathbb{R}} V(\sqrt{2} y)\dd y=1/2$,
\aln{
\int_{{\mathbb{R}}}\rho_{s}(x,y) \rho_t(y) V(\sqrt{2} y)\dd y&= \frac{1}{4\pi s t}\int_{\mathbb{R}} V(\sqrt{2} y) \exp{\left(-\frac{|x-y|^2}{2s}-\frac{|y|^2}{2t}\right)}\notag\\
&=\rho_{s+t}(x)\,\frac{s+t}{2\pi s t} \int_{\mathbb{R}} V(\sqrt{2} y) \exp{\left(-\frac{(s+t)|y+*|^2}{2st}\right)}\dd y\notag\\
&\leq \rho_{s+t}(x)\,\frac{s+t}{2\pi s t} \int_{\mathbb{R}} V(\sqrt{2} y) \dd y=\frac{s+t}{4\pi st} \rho_{s+t}(x).\label{pp recursive inequality}
}
Putting things together with $C=4\pi \|V\|_\infty$, we have
\al{
\int_{{\mathbb{R}}}\rho_{s}(x,y) \rho_t(y) V(\sqrt{2} y)\dd y\leq \frac{1}{4\pi }\left(C \land \frac{s+t}{st} \right) \rho_{s+t}(x).
}
Using this successively, we can bound each term of \eqref{eq:L2ppdecom} as
\aln{
& \b^{2n}\int_{0<s_1<\cdots<s_n<{tT}}\int_{\left({\mathbb{R}}^2\right)^n}\left(\prod_{i=1}^nV(\sqrt{2}x_i)\rho_{s_{i}-s_{i-1}}(x_{i-1},x_{i})\right)\frac{\rho_{s-s_n}(x_n)}{\rho_{tT}(0)}\dd \mathbf{s}\dd \mathbf{x}\notag\\
&\leq \left(\frac{\b^{2}}{4\pi}\right)^n \int_{0<s_1<\cdots<s_n<{tT}}\prod_{i=1}^n\left(C \land \frac{{tT}-s_{i-1}}{(s_i-s_{i-1})({tT}-s_{i})}\right) \dd \mathbf{s}\notag\\
&= \left(\frac{\b^{2}}{4\pi}\right)^n \int_{0<s_1<\cdots<s_n<{tT}} \prod_{i=1}^n\left(C\land \left(\frac{1}{s_i-s_{i-1}}+\frac{1}{{tT}-s_{i}}\right)\right) \dd \mathbf{s},\label{PP partition estimate}
}
where we set $s_0=0$ and $s_{n+1}={tT}$ and we have used $\frac{1}{{tT}-s_i}+\frac{1}{s_i-s_{i-1}}=\frac{{tT}-s_{i-1}}{(s_i-s_{i-1})({tT}-s_{i})}$ in the last line.
We write $\log_+(x)=\log{x}\lor 0$ and $C_1=2C$. We use the following integral estimate: for $s<{tT}$ and $k\geq 0$,
\al{
& \int_{s}^{tT} (C_1+\log_+{({tT}-t)})^k \left(C\land \left(\frac{1}{t-s}+\frac{1}{{tT}-t}\right)\right) \dd t\\
&\leq 2 C(C_1+\log_+{({tT}-s)})^k+ \int_{s+1}^{{tT}-1} (C_1+\log{({tT}-t)})^k \left(\frac{1}{t-s}+\frac{1}{{tT}-t}\right) \dd t\\
&\leq C_1(C_1+\log_+{({tT}-s)})^k+ (C_1+\log_+{({tT}-s)})^k \int_{s+1}^{{tT}-1} \frac{1}{t-s} \dd t - (k+1)^{-1}\left[ (C_1+\log_+{({tT}-t))^{k+1}}\right]_{s+1}^{{tT}-1}\\
&\leq C_1(C_1+\log_+{({tT}-s)})^k+ (C_1+\log_+{({tT}-s)})^k \log_+{({tT}-s)}+ (k+1)^{-1} (C_1+\log_+{({tT}-s))^{k+1}}\\
&= \frac{k+2}{k+1}(C_1+\log_+{({tT}-s)})^{k+1}.
}
Using this, \eqref{PP partition estimate} can be successively bounded from above as
\al{
&\quad \int_{0<s_1<\cdots<s_n<{tT}} \prod_{i=1}^n\left(C\land \left(\frac{1}{s_i-s_{i-1}}+\frac{1}{{tT}-s_{i}}\right)\right) \dd \mathbf{s}\leq (n+1)(C_1+\log{{tT}})^n.
}
Together with \eqref{eq:L2ppdecom} and \eqref{PP partition estimate}, using $\b=\hat{\beta}\sqrt{\frac{4\pi}{\log{T}}}$ with $\hat{\beta}<1$, we have
\al{
\varlimsup_{\varepsilon\to 0} \mathbb E \left[{\mathrm E}_{0,0}^{{tT},0}\left[\Phi^\b_{tT}\right]^2\right]& \leq \varlimsup_{\varepsilon\to 0} \sum_{n=0}^\infty \left(\frac{\beta^2}{4\pi}\right)^n (n+1)(C_1+\log{{tT}})^n= \sum_{n=0}^\infty (n+1)\hat{\beta}^{2n}<\infty.
}
\end{proof}
\begin{lemma}\cite[(5.11)]{CSZ20}\label{lem:lp}
Fix $\hat{\b}\in (0,1)$. Then, there exists $p_{\hat{\b}}>2$ such that for any $2\leq p<p_{\hat{\b}}$ and for $t\geq 0$\begin{align*}
\varlimsup_{\varepsilon\to 0}\mathbb E\left[{\mathrm E}_x\left[\Phi_{tT}^{{ \b_\varepsilon}}\right]^p\right]<\infty.
\end{align*}
\end{lemma}
\begin{lemma}\label{lem:p2pge}
Suppose $\hat{\b}\in (0,1)$ and fix $t>0$. Then, \begin{align}
&\sup_{\varepsilon \leq 1}\sup_{x\in {\mathbb{R}}^2}\sup_{s\leq tT_\varepsilon}{\mathrm E}_{0,0}^{s,x}\left[\exp\left(\beta_\varepsilon^2\int_0^s V(\sqrt{2}B_u)\dd u\right)\right]<\infty.\label{eq:p2pge}
\end{align}
\end{lemma}
\begin{proof}
By \eqref{eq:L2ppbdd}, \eqref{eq:L2ppdecom}, for $s\leq t T$,
\al{
{\mathrm E}_{0,0}^{s,z}\left[ e^{\b^2\int_0^{s} V(\sqrt 2 B_u) \dd u} \right]&= \mathbb E[{\mathrm E}_{0,0}^{s,z} [\Phi^\b_{s}]{\mathrm E}_{0,0}^{s,0} [\Phi^\b_{s}]]\\
&\leq \mathbb E[{\mathrm E}_{0,0}^{s,0} [\Phi^\b_{s}]^2]= {\mathrm E}_{0,0}^{s,0}\left[ e^{\b^2\int_0^{s} V(\sqrt 2 B_u) \dd u} \right]\\
&\leq \sup_{\varepsilon \leq 1}\left[\exp\left(\beta_\varepsilon^2\int_0^{tT_\varepsilon} V(\sqrt{2}B_u)\dd u\right)\right]<\infty,
}
where we have used the remark below.
\end{proof}
\begin{rem}
By the shear invariance of environment, we have that \begin{align*}
{\mathrm E}_{0,x}^{t,y}\left[\Phi_t^\b\right]\stackrel{(d)}{=} {\mathrm E}_{0,0}^{t,0}\left[\Phi_t^\b\right]
\end{align*}
for any $t>0$ and $x,y\in{\mathbb{R}}^2$.
\end{rem}
We end this subsection by
presenting the boundedness of negative moments of partition functions:
\iffalse{
\begin{align}
\overline{{\bf{\mathcal{Z}}}}_{t,r}(z)&={ \overline{{\bf{\mathcal{Z}}}}_{t,r}^{\b}(z)}={\mathrm E}_{z}\left[\Phi_{t}^\b:\mathtt{F}_{t,r}(B,z) \right],\label{eq:barZ}
\end{align}
where $\mathtt{F}_{t,r}(B,z)$ is the event that Brownian motion $B$ does not escape from the open ball $B(z,r)=\{x\in {\mathbb{R}}^2:|x-y|<r\}$ up to times $t$, \begin{align*}
\mathtt{F}_{t,r}(B,z)=\{B_s\in B(z,r)\ \text{for any }s\in [0,t]\}.
\end{align*}
}\fi
\begin{lemma}\label{lem:negativemoment}\cite[(5.12), (5.13), (5.14)]{CSZ20}
Let $\hat{\b}\in (0,1)$ and fix $t>0$. For any $p\geq 0$ and $x\in{\mathbb{R}}^2$,
\begin{align*}
&\sup_{s\in [0,t]}\mathbb E\left[\left({{\bf{\mathcal{Z}}}}_{sT_\varepsilon}^{{\beta_\varepsilon}}(x)\right)^{-p}\right]<\infty
\end{align*}
\end{lemma}
\subsection{Local limit theorem}
In this subsection, we give an estimate of local limit theorem for partition functions.
To describe the statement, we introduce the time-reversed partition function of time horizon $\ell$, $\overleftarrow{{\bf{\mathcal{Z}}}}^\b_{T,\ell}(z)$ :
\begin{equation
\overleftarrow{{\bf{\mathcal{Z}}}}^\b_{t,\ell}(z) = {\mathrm E}_{z} \left[\exp\left\{\b\int_{t-\ell}^t \int_{\mathbb R^2} \phi(B_{t-s}-y) \xi(\dd s,\dd y) -\frac{\b^2 V(0) \ell}{2}\right\} \right].\notag
\end{equation}
\begin{theorem}[Local limit theorem for polymers] \label{th:errorTermLLT}
{ Fix $t>0$.} Let $0<\ell_{{T_\varepsilon}}^a<\ell_{{T_\varepsilon}}^b<{ L({ T_\varepsilon})}\leq t{T_\varepsilon}$ be functions with $\displaystyle \lim_{\varepsilon\to0 }\ell_{{ T_\varepsilon}}^a\to \infty$, $\displaystyle \lim_{\varepsilon\to 0}\frac{\ell_{T_{{ T_\varepsilon}}}^b}{L(T_{{ T_\varepsilon}})}=0$, { $\displaystyle \lim_{\varepsilon\to 0}\frac{\log L({ T_\varepsilon})}{\log {T_\varepsilon}}=1$}. Then, for all ${\hat{\b}}<1$ there exists $C=C(\hat{\b})$ such that for all positive $\ell>0$ verifying $\ell_{{ T_\varepsilon}}^a\leq \ell\leq \ell_{{ T_\varepsilon}}^b$ and for all $x,y\in\mathbb R^d$,
\begin{align*}
&\mathbb E\left({\mathrm E}_{0, 0}^{{ L({ T_\varepsilon})},x} [\Phi_{{ L({ T_\varepsilon})}}^{{ \b_\varepsilon}}] - {\bf{\mathcal{Z}}}_\ell^{{\b_\varepsilon}}( 0) \overleftarrow{{\bf{\mathcal{Z}}}}^{{ \b_\varepsilon}}_{{ L({ T_\varepsilon})},\ell}( x)\right)^2 \\
&\leq \begin{cases}
C\frac{\ell}{{ L({T_\varepsilon})}}+C\b_\varepsilon^2\left( \log \frac{{ L({T_\varepsilon})}}{\ell}+\frac{|x|\log \ell}{{ L(T)}}+\frac{|x|^2 \ell}{{L(T)}^2}\right)\quad &|x|\leq \sqrt{{ L({T_\varepsilon})}\log { L({T_\varepsilon})}}\\
C&|x|\geq \sqrt{{ L({T_\varepsilon})}\log { L({ T_\varepsilon})}}
\end{cases}.
\end{align*}
\end{theorem}
\begin{rem}
Theorem states that the point-to-point partition function from $(0,x)$ to $({ L(T)},y)$ is approximated by the product of partition function from $(0,x)$ with length $\ell$ and time-reversed partition function from $({ L(T)},y)$ with length $\ell$ in $L^2$-sense. For $d\geq 3$, the reader may refer to \cite{CNN20,Si95,V06}.
\end{rem}
\begin{notation}
Fix $R_V>0$ such that ${\rm supp}\,V\subset B(0,R_V)$.
\end{notation}
The proof is composed of three steps.
\begin{lemma}[Step 1]\label{lem:LLT1} Fix $t>0$. There exists a constant $C=C(\tilde{\b})>0$ such that for $\ell_{{ T_\varepsilon}}^a\leq \ell\leq \ell_{{ T_\varepsilon}}^b$,
\[
\sup_{x\in\mathbb R^2} \mathbb E\left[\left({\mathrm E}_{0,0}^{{ L({ T_\varepsilon})},x} [\Phi^\b_{{ L({ T_\varepsilon})}}] - {\mathrm E}_{0,0}^{{ L({ T_\varepsilon})}, x}[\Phi^\b_{\ell}\Phi^\b_{{ L({ T_\varepsilon})}-\ell,{ L({ T_\varepsilon})}}]\right)^2\right] \leq C\b_\varepsilon^2{\log \frac{{L({ T_\varepsilon})}}{\ell}},
\]
where \begin{align*}
\Phi_{s,t}^\b=\exp\left(\b\int_s^t\int_{{\mathbb{R}}^2}\phi(y-B_u)\xi(\dd u,\dd y)-\frac{\b^2 V(0)(t-s)}{2}\right).
\end{align*}
\end{lemma}
\begin{proof}
Since $B_s^{(1)}-B_s^{(2)} \stackrel{(d)}{=} \sqrt 2 B_s$ for two independent Brownian motions, by $1-e^{-x}\leq x$ for $x\geq 0$, we have
\begin{align*}
&\mathbb E\bigg[\bigg({\mathrm E}_{0,0}^{{ L(T)}, x} [\Phi^\b_{L(T)} - \Phi^\b_{\ell}\Phi^\b_{{L(T)}-\ell,{ L(T)}}]\bigg)^2\bigg] \nonumber\\
& = {\mathrm E}_{0,0}^{{ L(T)},0} \left[\mathrm e^{\b^2 \int_0^{{ L(T)}} V(\sqrt 2 B_s)ds} -\mathrm e^{\b^2 \int_0^{\ell} V(\sqrt 2 B_s)ds}\mathrm e^{\b^2 \int_{{ L(T)}-\ell}^{{ L(T)}} V(\sqrt 2 B_s)ds}\right]\nonumber\\
& \leq {\mathrm E}_{0,0}^{{ L(T)},0}\left[ e^{\b^2\int_0^{{ L(T)}} V(\sqrt 2 B_s)ds}\,\b^2\int_\ell^{{ L(T)}-\ell} V(\sqrt 2 B_s)\dd s\right].
\end{align*}
The last expectation equals
\begin{equation*
\begin{split}
& \b^2\int_{[\ell,L(T)-\ell]\times \mathbb R ^2} V(\sqrt 2 z) \frac{\rho_s(z)\rho_{L(T)-s}(z)}{\rho_{L(T)}(0)} {\mathrm E}_{0,0}^{s,z} \left[e^{\b^2\int_0^{s} V(\sqrt 2 B_u)\dd u} \right] {\mathrm E}_{0,z}^{{ L(T)}-s,0} \left[ e^{\b^2\int_0^{tT-s} V(\sqrt 2 B_u)\dd u}\right] \dd z \dd s\\
& \leq \b^2\left( \sup_{s\leq { L(T)}} \sup_{|z|\leq R_V} {\mathrm E}_{0,0}^{s,z}\left[ e^{\b^2\int_0^{s} V(\sqrt 2 B_u) \dd u} \right]\right)^2 \int_\ell^{{ L(T)}-\ell} \int_{\mathbb R^2} V(\sqrt 2 z) \frac{\rho_s(z)\rho_{{L(T)}-s}(z)}{\rho_{{ L(T)}}(0)} \dd s \dd z,
\end{split}
\end{equation*}
where the supremum on the last line is finite by Lemma~\ref{lem:p2pge}.
Finally, by \eqref{pp recursive inequality},{
\begin{align*}
\int_\ell^{{L(T)}-\ell} \int_{\mathbb R^2} \frac{\rho_{s}(z)\rho_{{ L(T)}-s}(z)}{\rho_{{ L(T)}}(0)} V\left(\sqrt 2 z\right)\dd s \dd z
& \leq \frac{1}{4\pi} \int_\ell^{{ L(T)}-\ell} \frac{{L(T)}}{s({ L(T)}-s)}\dd s \\
& \leq \frac{1}{2\pi} \log \frac{{ L(T)}-\ell}{\ell}\leq \log \frac{{ L(T)}}{\ell},
\end{align*}}
and the statement of the lemma follows.
\end{proof}
By translation invariance of Brownian bridge, Brownian motion and noise, we have
\begin{align*}
&{\mathrm E}_{0,0}^{{ L(T)}, x}\left[\Phi^\b_{\ell}\Phi^\b_{{ L(T)}-\ell,{ L(T)}}\right]- {\bf{\mathcal{Z}}}_\ell^\b(0) \overleftarrow{{\bf{\mathcal{Z}}}}^\b_{{L(T)},\ell}( x)\\
&\stackrel{(d)}{=} {\mathrm E}_{0,0}^{{L(T)},0}\left[\Phi^\b_{\ell}\Phi^\b_{{L(T)}-\ell,{L(T)}}\right]\\
&-{\mathrm E}_0\left[\Phi^\b_\ell\left(B_\cdot+\frac{ x}{{{L(T)}}}\cdot\right)\right]{\mathrm E}_0\left[\exp\left(\b\int_{{ L(T)}-\ell}^{{L(T)}}\phi\left(B_{{L(T)}-s}+\frac{({L(T)}-s)x}{{{L(T)}}}-y\right)\xi(\dd s,\dd y)-\frac{\b^2 V(0)\ell}{2}\right)\right],
\end{align*}
where $\displaystyle B_\cdot+\frac{ x}{{{L(T)}}}\cdot$ is a Brownian motion with drift $\displaystyle \frac{x}{{ L(T)}}$.
Define
\begin{align*}
&A_{{ L(T)},\ell} := {\mathrm E}_{0,0}^{{L(T)},0}\left[\Phi^\b_{\ell}\Phi^\b_{{ L(T)}-\ell,{ L(T)}}\right] - {\mathrm E}_0\left[\Phi^\b_{\ell}\right] \overleftarrow{{\bf{\mathcal{Z}}}}^\b_{{L(T)},\ell}(0),\\
&B_{{ L(T)},\ell,x}:={\mathrm E}_0\left[\Phi^\b_{\ell}\right]-{\mathrm E}_0\left[\Phi^\b_\ell\left(B_\cdot+\frac{ x}{{L(T)}}\cdot\right)\right].
\end{align*}
\begin{lemma}[Step 2]
There exists a constant $C=C(\hat{\b})$ such that for all positive $\ell>0$ with $\ell_{{ T_\varepsilon}}^a\leq \ell \leq \ell_{{T_\varepsilon}}^b$,
\begin{align*}
\mathbb E\left[A_{{ L({ T_\varepsilon})},\ell}^2\right]\leq C\frac{\ell}{{L({ T_\varepsilon})}}.
\end{align*}
\end{lemma}
\begin{proof}
\eqref{eq:L2white} yields that
\begin{align*}
&\mathbb E\left[{\mathrm E}_{0,0}^{{ L(T)},0}\left[\Phi^\b_{\ell}\Phi^\b_{{ L(T)}-\ell,{L(T)}}\right]^2\right]={\mathrm E}_{0,0}^{{L(T)},0}\otimes {\mathrm E}_{0,0}^{L(T),0}\left[\exp\left(\b^2\int_0^\ell V(B_s-\widetilde{B}_s)ds+\b^2\int_{L(T)-\ell}^{L(T)}V(B_s-\widetilde{B}_s)ds\right)\right]\\
&=\int_{{\mathbb{R}}^2\times {\mathbb{R}}^2}\dd x\dd y\, \rho_{\ell}(x)\rho_\ell(y)\frac{\rho_{{L(T)}-2\ell}(y-x)}{\rho_{{L(T)}}(0)}\int_{{\mathbb{R}}^2\times {\mathbb{R}}^2}\dd z\dd w\rho_{\ell}(z)\rho_\ell(w)\frac{\rho_{{L(T)}-2\ell}(z-w)}{\rho_{{L(T)}}(0)}\\
&\hspace{2.5em}\times {\mathrm E}_{0,0}^{\ell,x}\otimes {\mathrm E}_{0,0}^{\ell,z}\left[\exp\left(\b^2\int_0^\ell V(B_s-\widetilde{B}_s)ds\right)\right]
{\mathrm E}_{0,0}^{\ell,y}\otimes {\mathrm E}_{0,0}^{\ell,w}
\left[\exp\left(\b^2\int_0^\ell V(B_s-\widetilde{B}_s)ds\right)\right],
\intertext{and}
&\mathbb E\left[{\mathrm E}_{0,0}^{{ L(T)},0}\left[\Phi^\b_{\ell}\Phi_{{L(T)}-\ell,{L(T)}}\right]{\mathrm E}_0\left[\Phi^\b_{\ell}\right]\overleftarrow{{\bf{\mathcal{Z}}}}_{{L(T)},\ell}(0)\right]\\
&=\int_{{\mathbb{R}}^2\times {\mathbb{R}}^2}\dd x\dd y\, \rho_{\ell}(x)\rho_\ell(y)\frac{\rho_{{ L(T)}-2\ell}(y-x)}{\rho_{{L(T)}}(0)}\int_{{\mathbb{R}}^2\times {\mathbb{R}}^2}\dd z\dd w \rho_\ell(z)\rho_\ell(w)\\
&\hspace{2.5em}\times {\mathrm E}_{0,0}^{\ell,x}\otimes {\mathrm E}_{0,0}^{\ell,z}\left[\exp\left(\b^2\int_0^\ell V(B_s-\widetilde{B}_s)ds\right)\right]{\mathrm E}_{0,0}^{\ell,y}\otimes {\mathrm E}_{0,0}^{\ell,w}\left[\exp\left(\b^2\int_0^\ell V(B_s-\widetilde{B}_s)ds\right)\right],
\end{align*}
where $B=\{B_s^{B}:0\leq s\leq \ell\}$ and $\widetilde{B}=\{\widetilde{B}_s:0\leq s\leq \ell\}$ are independent Brownian bridges with the law ${\mathrm P}_{0,0}^{\ell,u}$ ($u=x,y,z,w$).
Then, it is easy to see that \begin{align*}
&\mathbb E\left[{\mathrm E}_{0,0}^{{L(T)},0}\left[\Phi^\b_{\ell}\Phi^\b_{{L(T)}-\ell,{L(T)}}\right]^2\right]-\mathbb E\left[{\mathrm E}_{0,0}^{{L(T)},0}\left[\Phi^\b_{\ell}\Phi_{{L(T)}-\ell,{L(T)}}\right]{\mathrm E}_0\left[\Phi^\b_{\ell}\right]\overleftarrow{{\bf{\mathcal{Z}}}}_{{ L(T)},\ell}(0)\right]\\
&=\int_{{\mathbb{R}}^2\times {\mathbb{R}}^2}\dd x\dd y\, \rho_{\ell}(x)\rho_\ell(y)\frac{\rho_{{ L(T)}-2\ell}(y-x)}{\rho_{{ L(T)}}(0)}\int_{{\mathbb{R}}^2\times {\mathbb{R}}^2}\dd z\dd w \rho_\ell(z)\rho_\ell(w)\left(\frac{\rho_{{ L(T)}-2\ell}(z-w)}{\rho_{{ L(T)}}(0)}-1\right)\\
&\hspace{2.5em}\times{\mathrm E}_{0,0}^{\ell,x}\otimes {\mathrm E}_{0,0}^{\ell,z}\left[\exp\left(\b^2\int_0^\ell V(B_s-\widetilde{B}_s)ds\right)\right]{\mathrm E}_{0,0}^{\ell,y}\otimes {\mathrm E}_{0,0}^{\ell,w}\left[\exp\left(\b^2\int_0^\ell V(B_s-\widetilde{B}_s)ds\right)\right].
\end{align*}
Combining with \eqref{eq:p2pge}, \begin{align*}
\left|\mathbb E\left[{\mathrm E}_{0,0}^{{ L(T)},0}\left[\Phi^\b_{\ell}\Phi^\b_{{ L(T)}-\ell,{L(T)}}\right]^2\right]-\mathbb E\left[{\mathrm E}_{0,0}^{{ L(T)},0}\left[\Phi^\b_{\ell}\Phi_{{ L(T)}-\ell,{ L(T)}}\right]{\mathrm E}_0\left[\Phi^\b_{\ell}\right]\overleftarrow{{\bf{\mathcal{Z}}}}_{{ L(T)},\ell}(0)\right]\right|\leq C\frac{\ell}{{L(T)}}.
\end{align*}
Also, the same argument holds for $\mathbb E\left[{\mathrm E}_0\left[\Phi^\b_{\ell}\right]^2 \overleftarrow{{\bf{\mathcal{Z}}}}_{{ L(T)},\ell}(0)^2\right]$.
\end{proof}
\begin{lemma}[Step 3]\label{lem: estimate B}
Fix $t>0$. There exists a positive constant $C$ such that for all ${ \hat{\b}}< 1$ there exists a positive constant $C=C(\hat{\b})$ such that for all positive $\ell>0$ with $\ell_{{T_\varepsilon}}^a\leq \ell\leq \ell_{{ T_\varepsilon}}^b$ and all $x\in\mathbb R^d$,
\[
\mathbb E\left[B_{{ L({ T_\varepsilon})},\ell,x}^2\right]\leq \begin{cases}
C\b_\varepsilon^2\left( \frac{|x|\log \ell}{{ L(T)}}+\frac{|x|^2 \ell}{{ L(T)}^2}\right)
&|x|\leq \sqrt{{ L({ T_\varepsilon})}\log { L({ T_\varepsilon})}}\\
C&|x|\geq \sqrt{{ L({ T_\varepsilon})}\log { L({ T_\varepsilon})}}
\end{cases}.
\]
\end{lemma}
\begin{proof}
For $|x|\geq \sqrt{{L(T)}\log { L(T)}}$, it is trivial from \eqref{eq:L2bdd}.
Combining \eqref{eq:L2white} and transformation of Brownian motions yield that
\begin{align*}
\mathbb E\left[B_{{L(T)},\ell,x}^2\right]&={ 2}E\left[\exp\left(\b^2\int_0^\ell V(\sqrt{2}B_s)ds\right)-\exp\left(\b^2\int_0^\ell V\left(\sqrt{2}B_s+\frac{xs}{{ L(T)}}\right)ds\right)\right]\\
&={ 2}\sum_{n=1}^\infty \b^{2n}\int_{0<t_1<\cdots<t_n<\ell}\int_{{{\mathbb{R}}^2}^n}\dd \mathbf{s}\dd \mathbf{x}\prod_{i=1}^n V(\sqrt{2}x_i)\\
&\hspace{12mm}\times \left(\prod_{i=1}^n\rho_{s_i-s_{i-1}}(x_{i-1},x_{i})-\prod_{i=1}^n \rho_{s_i-s_{i-1}}\left(x_i-x_{i-1}+\frac{x(s_i-s_{i-1})}{{ L(T)}}\right)\right),
\end{align*}
{ where we set $x_0=0$.} When we use the relation \begin{align*}
&\prod_{i=1}^na_i-\prod_{i=1}^nb_i=\sum_{j=1}^n \left(\prod_{i=1}^{j-1} b_i\right)(a_j-b_j)\left(\prod_{k=j+1}^n a_k\right),
\end{align*}
and recall the notation $r_s$ from \eqref{Def: rs}, we have \begin{align}
& \mathbb E\left[B_{{ L(T)},\ell,x}^2\right]\notag\\
&\leq { 2}\sup_{z\in B(0,R_V)}\b^2 \int_0^\ell \int_{{\mathbb{R}}^2} V(\sqrt{2}y)\left|\rho_s(y-z)-\rho_s(y-z+\frac{xs}{{ L(T)}})\right| \dd y \dd s \notag\\
&\hspace{12mm} \times \sum_{n=1}^\infty \sum_{k=1}^{ n}
\b^{2n-2} \left(\sup_{z\in {\mathbb{R}}}\int_{0}^{\ell} \int_{{\mathbb{R}}^2} V(\sqrt{2}y)\rho_s(y-z)\dd y \dd s\right)^{n-1}\notag\\
&\leq { 2}\sum_{n=1}^\infty { n
\b^{2(n-1)} \left(\int_{0}^{\ell}r_s \dd s\right)^{n-1}} \sup_{z\in B(0,R_V)}\b^2 \int_0^\ell \int_{{\mathbb{R}}^2} V(\sqrt{2}y)\left|\rho_s(y-z)-\rho_s\left(y-z+\frac{xs}{{ L(T)}}\right)\right| \dd y \dd s\notag\\
&\leq C\b^2 \sup_{z\in B(0,R_V)} \int_0^\ell \int_{{\mathbb{R}}^2} V(\sqrt{2}y)\left|\rho_s(y-z)-\rho_s\left(y-z+\frac{xs}{{ L(T)}}\right)\right| \dd y \dd s,\label{eq:Best}
\end{align}
where we have used the estimate \eqref{eq:rorder} in the last line. Also, we have that for $y,z\in B(0,R_V)$, and for $s>0$\begin{align*}
&\left|\rho_s(y-z)-\rho_s\left(y-z+\frac{xs}{{{ L(T)}}}\right)\right|\\
&=\rho_{s}(y-z)\left|1-\exp\left(-\frac{\langle y-z,x\rangle}{{L(T)}}-\frac{|x|^2s}{2{L(T)}^2}\right)\right|\\
&\leq \rho_s(y-z)\left(\frac{R_V|x|}{{L(T)}}\exp\left(\frac{{2}R_V|x|}{{L(T)}}-\frac{|x|^2s}{2{L(T)}^2}\right)+\frac{R_V|x|}{{L(T)}}+\frac{|x|^2s}{2{L(T)}^2}\right)\\
&\leq C \rho_s(y-z)\left( \frac{|x|}{L(T)}+\frac{|x|^2 s}{L(T)^2}\right)
\end{align*}
where we denote by $\langle x,y\rangle$ the inner product of $x$ and $y\in {\mathbb{R}}^2$ and we use $e^x-1\leq xe^x$ if $x\geq 0$ and $1-e^x\leq -x$ if $x<0$ in the last line.
For $|x|\leq \sqrt{{ L(T)}\log{ L(T)}}$, \begin{align*}
\mathbb E\left[B_{{L(T)},\ell,x}^2\right]\leq&C\b^2\left(\frac{|x|\log \ell}{{ L(T)}}+\frac{|x|^2 \ell}{{L(T)}^2}\right).
\end{align*}
\end{proof}
{ Putting things together, we conclude the proof of Theorem \ref{th:errorTermLLT}.}\\
We also use the following lemma later.
\begin{lemma}\label{lem:covpart}
For fixed $t>0$ and $\hat{\b}\in (0,1)$, there exists a constant $C=C_{\hat{\b},t}$ such that for $x\in {\mathbb{R}}^2$ and for $1\leq \ell \leq t{ T_\varepsilon}$ with $t>0$
\begin{align}
\mathbb E\left[\left({\mathrm E}_x\left[\Phi_\ell^{{ \b_\varepsilon}}\right]-{\mathrm E}_0\left[\Phi_\ell^{{ \b_\varepsilon}}\right]\right)^2\right]\leq \begin{cases}
\displaystyle C\b_\varepsilon^2(1+|x|^2)\quad &|x|\leq \sqrt{\log { \ell}}\\
C&|x|>\sqrt{{\color{red} }\log { \ell}}.
\end{cases}
\label{eq:covpart}
\end{align}
\end{lemma}
\begin{proof}
For $|x|\geq \sqrt{ \log \ell}$, it is trivial from \eqref{eq:L2bdd}. We suppose $|x|< \sqrt{\log \ell}$.
Using the same argument as in \eqref{eq:Best}, with the convention $x_0=0$,
\begin{align*}
\mathbb E\left[\left({\mathrm E}_x\left[\Phi_\ell^{{ \b_\varepsilon}}\right]-{\mathrm E}\left[\Phi_\ell^{{\b_\varepsilon}}\right]\right)^2\right]&=E\left[\exp\left(\b^2\int_0^\ell V(\sqrt{2}B_s)ds\right)-\exp\left(\b^2\int_0^\ell V\left(x+\sqrt{2}B_s\right)ds\right)\right]\\
&=\sum_{n=1}^\infty \b^{2n}\int_{0<t_1<\cdots<t_n<\ell}\int_{{{\mathbb{R}}^2}^n}\dd \mathbf{s}\dd \mathbf{x}\prod_{i=1}^n V(\sqrt{2}x_i)\\
&\hspace{12mm}\times \left(\prod_{i=1}^n\rho_{s_i-s_{i-1}}(x_{i}-x_{i-1})-\rho_{s_i-s_{i-1}}\left(x_1-x\right) \prod_{i=2}^n \rho_{s_i-s_{i-1}}\left(x_i-x_{i-1}\right)\right)\\
&\leq C\b^2\left(1+\sup_{z\in B(0,R_V)} \int_1^\ell \int_{{\mathbb{R}}^2} V(\sqrt{2}y)\left|\rho_s(y-z)-\rho_s(y-z+x)\right| \dd y \dd s\right).
\end{align*}
Also, we have that for $y,z\in B(0,R_V)$ and $s\geq 1$, \begin{align*}
\left|\rho_s(y-z)-\rho_s(y-z+x)\right|&=\rho_{s}(y-z)\left|1-\exp\left(-\frac{2\langle y-z,x\rangle+|x|^2}{2s}\right)\right|\\
&\leq \frac{C}{s}\rho_s(y-z)\left(\exp\left(\frac{R_V^2}{s}\right)+{|x|+|x|^2}\right)\\
&\leq \frac{C}{s}\rho_s(y-z)\left(1+|x|^2\right).
\end{align*}
Thus, we have
\begin{align*}
\mathbb E\left[\left({\mathrm E}_x\left[\Phi_\ell^{{ \b_\varepsilon}}\right]-{\mathrm E}_0\left[\Phi_\ell^{{ \b_\varepsilon}}\right]\right)^2\right]\leq&C \b^2\left(1+|x|^2\right)
\end{align*}
\end{proof}
\section{Proofs of Theorem \ref{th:EWlimit}}\label{sec:proofEWKPZ}
For fixed $t>0$ and for $u_0\in \mathfrak{C}$, let us define the martingale
\begin{equation*}
s \to \mathcal{W}_{s}(x)= \mathcal{W}_{s}^{(t,T,\hat{\b},u_0)}(x)={\mathrm E}_x\left[ \Phi_{s}^\b(B) \, u_0\left(\frac{B_{t T}}{\sqrt{T}}\right)\right]
\end{equation*}
with respect to the filtration $\{\mathcal{F}_s:0\leq s\leq tT\}$ associated to the white noise $\xi$. Then, it follows from Feynman-Kac formula (see \eqref{eq:FKform}) that for each $(t,x)\in [0,\infty)\times {\mathbb{R}}^2$
\begin{equation} \label{eq:hepsWtT}
u_{\varepsilon}^{(T,\hat{\b},u_0)}(t,x)\stackrel{(d)}{=} {\mathcal{W}_{t T}^{(t,T,\hat{\b},u_0)}(\sqrt{T}x)}.
\end{equation}
We omit some superscripts $t$, $T$, $\hat{\b}$, and $u_0$ to make notation simple for several notations when it is easily understood from the context.
Since both $\|u_0^{-1}\|_{\infty}$ and $\|u_0\|_{\infty}$ are finite,
$$\|u_0^{-1}\|^{-1}_{\infty} {\bf{\mathcal{Z}}}_{s}(x)\leq \mathcal{W}_{s}(x)\leq \|u_0\|_{\infty} {\bf{\mathcal{Z}}}_{s}(x).$$
Hereafter, we use this without any comment.
\begin{comment}
To prove Proposition \ref{prop:mainpropKPZ}, we rely again on the CLT for martingales. This time, because of the presence of ${\bf{\mathcal{Z}}}_\tau$ in the denominator of \eqref{eq:diffOfH}, the bracket of $H$ is slightly more delicate to deal with compared to the previous sections and our strategy is to replace $H$ in \eqref{eq:CLTmartH} by another martingale with no denominator.
\end{comment}
It\^o's formula yields that for each $x\in {\mathbb{R}}^2$
\begin{align}
\mathcal{W}_s^{(t,T,\hat{\b},u_0)}(x)&=\bar{u}(t,x)+\int_0^s \dd \mathcal{W}^{(t,T,\hat{\b},u_0)}_u(x)\label{eq:Itozdeco}\\
\mathcal{W}_s^{(t,T,\hat{\gamma},v_0)}(x)&=\bar{v}(t,x)+\int_0^s \dd \mathcal{W}^{(t,T,\hat{\gamma},v_0)}_u(x)\label{eq:Itozdeco2}
\end{align}
with \begin{align}
\langle \mathcal{W}^{(\hat{\b},u_0)}(x),\mathcal{W}^{(\hat{\gamma},v_0)}(y)\rangle_s=\int_0^s \b\gamma{\mathrm E}_{x}{\otimes }{\mathrm E}_y\left[V(B_u-\widetilde{B}_u)\Phi^\b_u(B)\Phi^\gamma_u(\widetilde{B})u_0\left(\frac{B_{t T}}{\sqrt{T}}\right)v_0\left(\frac{\widetilde{B}_{t T}}{\sqrt{T}}\right)\right]\dd u \label{eq:QuadVarW}
\end{align}
for each $x,y\in{\mathbb{R}}^2$, where ${\mathrm E}_{x}{\otimes }{\mathrm E}_y$ denotes the expectation in two independent Brownian motions $B$ and $\widetilde{B}$ starting from $x$ and $y$.
Then, we find by It\^{o}'s formula that for ${F}\in \mathfrak{F}$, $F(\mathcal{W}_s(x))$ has the following semimartingale representation \begin{align}
F(\mathcal{W}^{(\hat{\b},u_0)}_s(x))&=F(\bar{u}(t,x))+\int_0^s F'(\mathcal{W}^{(\hat{\b},u_0)}_u(x))\dd \mathcal{W}^{(\hat{\b},u_0)}_u(x)\notag\\
&\hspace{6em}+\frac{1}{2}\int_0^s F''(\mathcal{W}^{(\hat{\b},u_0)}_u(x))\dd\langle \mathcal{W}^{(\hat{\b},u_0)}(x)\rangle_u\label{eq:ItoDecoF}
\end{align}
and we denote by \begin{align*}
&G_s^{(t,T,F,\hat{\b},u_0)}(x)=G_s(x)=\int_0^s F'({\mathcal{W}_u^{(\hat{\b},u_0)}(x)})\dd \mathcal{W}_u^{(\hat{\b},u_0)}(x)\\
&H^{(t,T,F,\hat{\b},u_0)}_s(x)=H_s(x)=\int_0^s F''(\mathcal{W}^{(\hat{\b},u_0)}_u(x))\dd\langle \mathcal{W}^{(\hat{\b},u_0)}(x)\rangle_u.
\end{align*}
First, we will prove the fluctuations of martingale parts converge to centered Gaussian random variables.
\begin{proposition} \label{prop:mainpropKPZ} Suppose $u_0^{(1)},\cdots,u_0^{(n)}\in \mathfrak{C}$, $\hat{\b}^{(1)},\cdots,\hat{\b}^{(n)}\in (0,1)$ and $F_{1},\cdots,F_{n}\in \mathfrak{F}$.
For any test function $f_1,\cdots,f_n\in C_c^\infty({\mathbb{R}}^2)$, as $T\to \infty$
\begin{equation} \label{eq:CLTmartH}
\left\{\frac{1}{\b^{(i)}_\varepsilon}\int_{\mathbb{R}^2} f_i(x) G^{\left(F_i,\b^{(i)},u_0^{(i)}\right)}_{{T_\varepsilon} t} (\sqrt {T_\varepsilon} x) \dd x\right\}_{i=1,\cdots,n} \stackrel{(d)}{\longrightarrow} \left\{\mathscr U(t,f_i,F_i,\hat{\b}^{(i)},u_0^{(i)})\right\}_{i=1,\cdots,n},
\end{equation}
where $\left\{\mathscr U(t,f_i,F_i,\hat{\b}^{(i)},u_0^{(i)})\right\}_{i=1,\cdots,n}$ are Gaussian random variables with zero means and covariance
\begin{align*}
&\mathrm{Cov}\left(\mathscr U(t,f_i,F_i,\hat{\b}^{(i)},u_0^{(i)}),\mathscr U(t,f_j,F_j,\hat{\b}^{(j)},u^{(j)}_0)\right)\\
&={\frac{1}{1-\hat{\b}^{(i)}\hat{\beta}^{(j)}}\int_{0}^t\dd s\int_{({\mathbb{R}}^2)^2}\dd x\dd yf_i(x)f_j(y)
I^{\left(t,F_i,\hat{\b}^{(i)},u_0^{(i)}\right)}(x)I^{\left(t,F_j,\hat{\b}^{(j)},u^{(j)}_0\right)}(y)}
\\
&\hspace{14em}{ \times \int_{{\mathbb{R}}^2}\dd z\rho_\sigma(x,z)\rho_\sigma(y,z)\bar{u}^{(i)}(t-\sigma,z)\bar{u}^{(j)}(t-\sigma,z)}.
\end{align*}
\end{proposition}
Then, we will prove that the It\^o correction term can be neglected in the limit:
\begin{proposition}\label{prop:VanishBracket}
For any $t>0$, $\hat{\b}\in(0,1)$, $u_0\in \mathfrak{C}$ and $F\in \mathfrak{F}$, as $\varepsilon\to 0$,
\begin{equation}
\frac{1}{\b_\varepsilon} \int_{\mathbb{R}^2} f(x)\,\left( H_{{T_\varepsilon}t}^{(F)}(\sqrt {T_\varepsilon} x) - \mathbb E\left[ H_{ {T_\varepsilon}t}^{(F)}(\sqrt {T_\varepsilon} x) \right]\right) \stackrel{L^1}{\longrightarrow} 0.
\end{equation}
\end{proposition}
Proposition \ref{prop:mainpropKPZ} and Proposition \ref{prop:VanishBracket} combined {with \eqref{eq:hepsWtT} and \eqref{eq:ItoDecoF}} imply Theorem~\ref{th:EWlimit} for $1$-dimensional in time. Thus, Gaussian limit comes from the martingale part of $\displaystyle \int f(x)F(\mathcal{W}_s(\sqrt{ { T_\varepsilon}}x))\dd x$.
\subsection{Proof of Proposition \ref{prop:mainpropKPZ} and heuristics}\label{IdeaG}
In the following, we give a heuristic idea of the proof of Proposition \ref{prop:mainpropKPZ}.
First, we introduce the key theorem to prove the convergence of martingale to Gaussian process in this paper:
\begin{theorem}\label{thm:JS}{\cite[Theorem 3.11 in Chap.\,8]{JS87}, \cite[Theorem 1.4 in Chap.\,7]{EK86}}
For each $n\geq 1$, let $\mathcal{F}^{n}=\{\mathcal{F}_t^n:t\geq 0\}$ be a filtration and let $X^{(n)}=(X_{t}^{(n,d)},\dots,X_{t}^{(n,d)})$ be an ${\mathbb{R}}^d$-valued continuous $\mathcal{F}^n$-martingale with $X_0^n=0$. Suppose that there exists a $d\times d$ positive definite matrix-valued continuous function $c=\{c_{ij}(t)\}_{i,j=1}^d$ such that for each $t\geq 0$, $\langle X^{(n,i)},X^{(n,j)}\rangle_t\to c_{ij}(t)$ in probability. Then, $X^{(n)}\stackrel{(d)}{\longrightarrow} X$, where $X=(X_t^{(1)},\cdots,X_t^{(d)})$ is an ${\mathbb{R}}^d$-valued Gaussian process with $\langle X^{(i)},X^{(j)}\rangle_t=c_{ij}(t)$.
\end{theorem}
\begin{rem}
Theorem \ref{thm:JS} is simplified from the original one for our convenience.
\end{rem}
Thus, we will focus our analysis on the cross-variation of martingales.
By the local limit theorem (Theorem \ref{th:errorTermLLT}), we may expect that for large $s$
\begin{align}
\mathcal{W}_s(x)&= {\mathrm E}_x\left[ \Phi_s^\b \, u_0\left(\frac{B_{t T}}{\sqrt{T}}\right)\right]\notag\\
&=\int_{{\mathbb{R}}^2}\rho_s(z-x) {\mathrm E}_{0,x}^{s,z}[\Phi^\b_s] {\mathrm E}_z\left[ u_0\left(\frac{B_{t T-s}}{\sqrt{T}}\right)\right] \dd z \label{eq:Wexpan}\\
&\approx \int_{{\mathbb{R}}^2}\rho_s(z-x) {\bf{\mathcal{Z}}}^\b_{s{\ell(T)}}(x) \overleftarrow{{\bf{\mathcal{Z}}}}^\b_{s,{ s\ell(T)}}(z) {\mathrm E}_z\left[ u_0\left(\frac{B_{t T-s}}{\sqrt{T}}\right)\right] \dd z\notag,
\end{align}
{ where for fixed $\delta\in (0,\frac{1}{100})$, we set
\begin{align}\label{def: ell(T)}
\ell(T)=\exp\left(-\left(\log T\right)^{\frac{1}{2}-\delta}\right).
\end{align}}
{ Moreover, we may expect that the last term is approximated in some sense by }
\begin{align}
& {\bf{\mathcal{Z}}}^\b_{{ s\ell(T)}}(x) \int_{{\mathbb{R}}^2}\rho_s(z-x)\mathbb E\left[\overleftarrow{{\bf{\mathcal{Z}}}}^\b_{s,{s\ell(T)}}(z)\right] {\mathrm E}_z\left[ u_0\left(\frac{B_{t T-s}}{\sqrt{T}}\right)\right] \dd z= {\bf{\mathcal{Z}}}^\b_{{ s\ell(T)}}(x) \bar{u}(t,T^{-\frac{1}{2}}\,x)\notag
\end{align}
since Lemma \ref{lem:covpart} may imply that $(\overleftarrow{{\bf{\mathcal{Z}}}}_{s,{ s\ell(T)}}(x))_{x\in{\mathbb{R}}^2}$ are asymptotically independent and homogenization occurs.
Therefore, one may observe for $F\in \mathfrak{F}$ that for large $s$
\begin{align*}
F'(\mathcal{W}_s(x)){\dd \mathcal{W}_s(x)} & = {\b}F'({\mathcal{W}_s(x)}) \int_{\mathbb R^2} \xi(\dd s,\dd b) {\mathrm E}_x\left[ \phi(B_s-b) \Phi_s^\b { \, u_0\left(\frac{B_{t T}}{\sqrt{T}}\right)} \right]\\
& = {\b}F'({\mathcal{W}_s(x)}) \int_{\mathbb R^d} \xi(\dd s,\dd b) \int_{\mathbb R^2} \rho_s(z-x) \phi(z-b) {\mathrm E}_{0,x}^{ s,z}\left[\Phi^\b_s\right] { \, {\mathrm E}_z\left[ u_0\left(\frac{B_{t T-s}}{\sqrt{T}}\right)\right]} \dd z\\
&\approx {\b}F'({\bf{\mathcal{Z}}}^\b_{{ s\ell(T)}}(x)\bar{u}(t,T^{-\frac{1}{2}}\,x)) {\bf{\mathcal{Z}}}^\b_{{ s\ell(T)}}(x)\\
&\hspace{4em}\times \int_{{\mathbb{R}}^2} \xi(\dd s, \dd b) \int_{{\mathbb{R}}^2}\rho_s(z-x) \phi(z-b) \overleftarrow{{\bf{\mathcal{Z}}}}^\b_{s , { s\ell(T)}}(z)
\,
{\mathrm E}_z\left[ u_0\left(\frac{B_{t T-s}}{\sqrt{T}}\right)\right]\dd z
\end{align*}
where we have used the local limit theorem in the third line.
We denote by \begin{align*}
\displaystyle I^{(T)}_s(x)=I_{s}^{(t,T,F,\hat{\b},u_0)}(x)=F'\left({\bf{\mathcal{Z}}}_{{s\ell(T)}}^{\b}(\sqrt{T}x)\bar{u}(t,x)\right){\bf{\mathcal{Z}}}_{{ s\ell(T)}}^{\b}(\sqrt{T}x).
\end{align*}
Also, we have for $F_1,F_2\in \mathfrak{F}$ that for $s=T\sigma$ and $x=\sqrt{T}x'$ and $y=\sqrt{T}y'$, \begin{align*}
&F_1'(\mathcal{W}^{(\hat{\b},u_0)}_s(x))F_2'(\mathcal{W}^{(\hat{\gamma},v_0)}_s(y)){\dd \langle \mathcal{W}^{(\hat{\b},u_0)}(x),\mathcal{W}^{(\hat{\gamma},v_0)}(y)\rangle}_s \\
&=\dd s\b\gamma F'_1(\mathcal{W}_s^\b(x))
F'_2(\mathcal{W}_s^\gamma(y))
\int_{({\mathbb{R}}^2)^2}\dd z_1\dd z_2 \rho_s(x,z_1)\rho_s(y,{z_2})V(z_1-z_2)
{\mathrm E}_{0,x}^{s,z_1}\left[\Phi_s^\b\right]
{\mathrm E}_{0,x}^{s,z_2}\Big[\Phi_s^\gamma\Big]\\
&\hspace{15em}\times
{\mathrm E}_{z_1}\left[ u_0\left(\frac{B_{t T-s}}{\sqrt{T}}\right)\right]{\mathrm E}_{z_2}\left[ v_0\left(\frac{\widetilde{B}_{t T-s}}{\sqrt{T}}\right)\right]\\
&\approx \dd s{\b\gamma}
F'_1\left({\bf{\mathcal{Z}}}_{{s\ell(T)}}^\b(x)\bar{u}(t,x')\right)
F'_2\left({\bf{\mathcal{Z}}}_{{ s\ell(T)}}^\gamma(y)\bar{v}(t,y')\right)
{\bf{\mathcal{Z}}}_{{ s\ell(T)}}^\b(x){\bf{\mathcal{Z}}}_{{s\ell(T)}}^\gamma(y)\\
&\times \int_{({\mathbb{R}}^2)^2}\dd z_1\dd z_2 \rho_s(x,z_1)\rho_s(y,{z_2})V(z_1-z_2)
\overleftarrow{{\bf{\mathcal{Z}}}}^\b_{s ,{ s\ell(T)}}(z_1)
\overleftarrow{{\bf{\mathcal{Z}}}}^\gamma_{s , { s\ell(T)}}(z_2)
{\mathrm E}_{z_1}\left[ u_0\left(\frac{B_{t T-s}}{\sqrt{T}}\right)\right]{\mathrm E}_{z_2}\left[ v_0\left(\frac{\widetilde{B}_{t T-s}}{\sqrt{T}}\right)\right]\\
&\approx \dd \sigma{\b\gamma}
I_s^{(t,T,F_1,\hat{\b},u_0)}(x')I_s^{(t,T,F_2,\hat{\gamma},v_0)}(y')\\
&\times \int_{({\mathbb{R}}^2)^2}\dd z\dd w \rho_\sigma(x',z)\rho_\sigma(y',z-T^{-\frac{1}{2}}w)V(w)
\overleftarrow{{\bf{\mathcal{Z}}}}^\b_{T\sigma ,{ T\sigma\ell(T)}}(z)
\overleftarrow{{\bf{\mathcal{Z}}}}^\gamma_{T\sigma , { T\sigma\ell(T)}}(z-T^{-\frac{1}{2}}w)\\
&\hspace{5em}\times \bar{u}(t-\sigma,z)
\bar{v}(t-\sigma,z-T^{-\frac{1}{2}}w)
\end{align*}
and thus by homogenization, $\overleftarrow{{\bf{\mathcal{Z}}}}^\b\overleftarrow{{\bf{\mathcal{Z}}}}^\gamma$ would be replaced by $\mathbb E\left[\overleftarrow{{\bf{\mathcal{Z}}}}^\b\overleftarrow{{\bf{\mathcal{Z}}}}^\gamma\right]$, and $F'({\bf{\mathcal{Z}}} \bar{u}) {\bf{\mathcal{Z}}}$ terms would be replaceed by its expectation so that the cross variation would be approximated by
\begin{align*}
&\dd \sigma{\b\gamma}
\mathbb E\left[I_s^{(t,T,F_1,\hat{\b},u_0)}(x')\right]
\mathbb E\left[I_s^{(t,T,F_2,\hat{\gamma},v_0)}(y')\right]
\\
&\times \int_{{\mathbb{R}}^2}\dd z\rho_\sigma(x',z)\rho_\sigma(y',z)
\mathbb E\left[\overleftarrow{{\bf{\mathcal{Z}}}}^\b_{T\sigma ,{ T\sigma\ell(T)}}(z)
\overleftarrow{{\bf{\mathcal{Z}}}}^\gamma_{T\sigma , { T\sigma\ell(T)}}(z) \right]\bar{u}(t-\sigma,z)
\bar{v}(t-\sigma,z).
\end{align*}
Due to Theorem \ref{thm:CSZ17b} and \eqref{eq:L2bdd}, we have\begin{align}
&\int_{0}^{sT}\int_{({\mathbb{R}}^2)^2}f(x)g(y)F_1'(\mathcal{W}^\b_u(\sqrt{T}x))F_2'(\mathcal{W}^\gamma_u(\sqrt{T}y)){\dd \langle \mathcal{W}^\b(\sqrt{T}x),\mathcal{W}^\gamma(\sqrt{T}y)\rangle}_{u} \notag\\
&\approx \frac{1}{1-\hat{\b}\hat{\gamma}}\mathbb E\left[F'(e^{X_{\hat{\b}}-\frac{1}{2}\sigma^2(\hat{\b})}\bar{u}(t,x))e^{X_{\hat{\b}}-\frac{1}{2}\sigma^2(\hat{\b})}\right]
\mathbb E\left[F'(e^{X_{\hat{\gamma}}-\frac{1}{2}\sigma^2(\hat{\gamma})}\bar{v}(t,x))e^{X_{\hat{\gamma}}-\frac{1}{2}\sigma^2(\hat{\gamma})}\right]\notag\\
&\times \int_0^s\dd \sigma\int_{({\mathbb{R}}^2)^2}\dd x \dd y f(x)g(y)\int_{{\mathbb{R}}^2}\dd z\rho_{\sigma}(x,z)\rho_{\sigma}(y,z)\bar{u}(t-\sigma,z)\bar{v}(t-\sigma,z)\label{eq:limitcov}
\end{align}
and Theorem \ref{thm:JS} implies that the limit process is the Gaussian process with covariance function \eqref{eq:limitcov}.
For simplicity of notations in the proof, we will focus on the quadratic variation of $\displaystyle \int_{\mathbb{R}^2} f(x) G^{\left(F,\b,u_0\right)}_{{ T_\varepsilon} t} (\sqrt {T_\varepsilon} x) \dd x$. The reader can easily recover the proof for the cross-bracket from the above argument.
\vspace{1em}
To make this rough idea rigorous, we introduce a martingale increment $\dd\mathcal{M}^{(t,T,F,\hat{\b},u_0)}_s(x)$ for fixed $t>0$, $x\in {\mathbb{R}}^2$, $\hat{\b}\in (0,1)$, $F\in \mathfrak{F}$, and $u_0\in \mathfrak{C}$ as
\begin{align}
&\dd \mathcal{M}_s(x)=\dd \mathcal{M}_s^{(t,T,F,\beta,u_0)}\nonumber \\
&={\b}F'({\bf{\mathcal{Z}}}^\b_{{ s\ell(T)}}(x)\bar{u}(t,T^{-\frac{1}{2}}\,x)) {\bf{\mathcal{Z}}}^\b_{{ s\ell(T)}}(x)\notag\\
&\hspace{4em}\times \int_{{\mathbb{R}}^2} \xi(\dd s, \dd b) \int_{{\mathbb{R}}^2}\rho_s(z-x) \phi(z-b) \overleftarrow{{\bf{\mathcal{Z}}}}^\b_{s, { s\ell(T)}}(z) \, {\mathrm E}_z\left[ u_0\left(\frac{B_{t T-s}}{\sqrt{T}}\right)\right]\dd z,\label{eq:defdMbartau}
\end{align}
and set \[
\mathcal{M}_s(x)=\mathcal{M}_{s}^{(t,T,F,\b,u_0)}(x) := \begin{cases}
\displaystyle \int_{t { T_\varepsilon}m( { T_\varepsilon})}^{s} \dd \mathcal{M}_u(x) \quad &s\geq t { T_\varepsilon}m( { T_\varepsilon})\\
0&0\leq s\leq t { T_\varepsilon}m( { T_\varepsilon}),
\end{cases}
\]
where { \begin{align*}
m( { T_\varepsilon})=\exp\left(-\left(\log { T_\varepsilon}\right)^{\frac{1}{2}-\delta}\right).
\end{align*}}
The following proposition computes the covariances of $\mathcal{M}_s$:
\begin{proposition} \label{prop:MwidetildeIsGaussianFlat}
Suppose $u_0^{(1)},\cdots,u_0^{(n)}\in \mathfrak{C}$, $\hat{\b}^{(1)},\cdots,\hat{\b}^{(n)}\in (0,1)$ and $F_1,\cdots,F_n\in \mathfrak{F}$.
For any test function $f_1,\cdots,f_n\in C_c^\infty({\mathbb{R}}^2)$, as $\varepsilon\to 0$
\begin{equation} \label{eq:cvMtoU3}
\frac{1}{\b_\varepsilon^{(i)}}\int_{\mathbb R^2} f_i(x) \, \mathcal{M}^{(i)}_{ { T_\varepsilon}t}(\sqrt { T_\varepsilon} x) \dd x \stackrel{(d)}{\longrightarrow} \mathscr U(t,f_i,F_i,\hat{\b}^{(i)},u_0^{(i)}).
\end{equation}
\end{proposition}
The following proposition states that $\dd G_s(x)$ can be replaced by $\dd \mathcal{M}_s(x)$, which concludes the proof of Proposition \ref{prop:mainpropKPZ}:
\begin{proposition} \label{prop:replaceByM}
For any test function $f$ and $s>0$,
\al{
\frac{1}{\b_\varepsilon}\mathbb E\left[ \left|\int_{{\mathbb{R}}^2} f(x) \left(G_{s { T_\varepsilon}}^{(t,T,F,\b,u_0)}(\sqrt { T_\varepsilon} x)-\mathcal{M}^{(t,T,F,\b,u_0)}_{s { T_\varepsilon}}(\sqrt{ { T_\varepsilon}}x)\right)\dd x\right| \right] \to 0
}
as $\varepsilon \to 0$.
\end{proposition}
The proof of Proposition \ref{prop:MwidetildeIsGaussianFlat} is given in the following subsection and the proof of Proposition \ref{prop:replaceByM} is given in subsection \ref{sub:4.3}.
\subsection{Proof of Proposition \ref{prop:MwidetildeIsGaussianFlat}}\label{sub:4.2}
We will focus on only the quadratic variation of $G_{sT}(x\sqrt{T})$ to make the argument simple. Readers can easily replace the quadratic variation by the cross variation.
To prove Proposition \ref{prop:MwidetildeIsGaussianFlat}, we will show the following two lemmas:
\begin{lemma}\label{lem:CLTmainpart}
Let $0<\tau_0\leq \tau\leq t$. Then, as $\varepsilon\to 0$,
\begin{align} \label{eq:CVbracketMt}
&\frac{1}{\b_\varepsilon^2} \int_{ { T_\varepsilon}\tau_0}^{ { T_\varepsilon}\tau} \int_{\left({\mathbb{R}}^2\right)^2}\dd x\dd yf(x)f(y)\dd \left\langle \mathcal{M}(x_{ { T_\varepsilon}}), \mathcal{M}(y_{ { T_\varepsilon}}) \right\rangle_{s}\dd s\notag\\
& \stackrel{L^1}{\longrightarrow} \frac{1}{1-\hat{\b}^2}\int_{\tau_0}^\tau\dd s\int_{({\mathbb{R}}^2)^2}\dd x\dd yf(x)f(y)
I(x)I(y
\int_{{\mathbb{R}}^2}\dd z\rho_\sigma(x-z)\rho_\sigma(y-z)\bar{u}(t-\sigma,z)^2,
\end{align}
where we set $x_{{T_\varepsilon}}=x\sqrt{ { T_\varepsilon}}$ for $x\in {\mathbb{R}}^2$ and
\begin{align*}\displaystyle I(x)=I^{(t,F,\hat{\b},u_0)}(x):=
\mathbb E\left[
F'\left(e^{X_{\hat{\b}}}
\bar{u}(t,x)\right)
e^{X_{\hat{\b}}}\right]=
\mathbb E\left[F'\left(e^{X_{\hat{\b}}+{\sigma^2(\hat{\b})}}
\bar{u}(t,x)\right)
\right]
\end{align*}
for $t>0$, $x\in{\mathbb{R}}^2$, $F\in \mathfrak{F}$, $\hat{\b}\in (0,1)$, and $u_0\in \mathfrak{C}$.
\end{lemma}
Lemma \ref{lem:CLTmainpart} with Theorem \ref{thm:JS} implies that the centered martingale $\displaystyle \left(\int_{{\mathbb{R}}^2}\frac{1}{\b}f(x)\left(\mathcal{M}_{T \tau} (x_T)- \mathcal{M}_{T \tau_0} (x_T)\right)\dd x\right)_{\tau_0\leq \tau\leq t}$ converges in distribution to a Gaussian process with covariance given by the RHS of \eqref{eq:CVbracketMt}.
\begin{lemma}\label{lem:MtildeTau0negl}
\begin{equation} \label{eq:MtildeTau0negl}
\lim_{\tau_0\to 0}\varlimsup_{\varepsilon\to 0}\mathbb E\left[\frac{1}{\b_\varepsilon^2}\left(\int f(x) \mathcal{M}_{ { T_\varepsilon}\tau_0}(x_{ { T_\varepsilon}})\dd x\right)^2\right]=0.
\end{equation}
\end{lemma}
Thus, letting $\tau_0\to 0$ and $\tau=t$, the RHS of \eqref{eq:CVbracketMt} is exactly the covariance function of the Gaussian process { $\mathscr{U}_t(f,F,\hat{\b},u_0)$.}
\subsubsection{Proof of Lemma \ref{lem:CLTmainpart} and Lemma \ref{lem:MtildeTau0negl}}
The proof of Lemma \ref{lem:CLTmainpart} is divided into several steps.
Recall that $x_T=\sqrt{T}\,x$. First of all, we can easily find by Markov property and \eqref{eq:QuadVarW} that
\begin{align*}
&\frac{1}{\b^2} \int_{T\tau_0}^{T\tau} \int_{\left({\mathbb{R}}^2\right)^2}\dd x\dd yf(x)f(y)\dd \left\langle \mathcal{M}(x_T), \mathcal{M}(y_T) \right\rangle_{s}\dd x\dd y\dd s\notag\\
& = \int_{T\tau_0}^{T\tau}\dd s \int_{({\mathbb{R}}^2)^2}
f(x)f(y)
I_s^{(T)}(x)
I_s^{(T)}(y)\\
&\hspace{4em}\int_{({\mathbb{R}}^2)^2}\dd z_1\dd z_2\rho_s(z_1-x_T)\rho_s(z_2-y_T)V(z_1-z_2)\notag\\
&{\hspace{7em} \times \overleftarrow{{\bf{\mathcal{Z}}}}^{\b}_{s,{ s\ell(T)}}(z_1)
\overleftarrow{{\bf{\mathcal{Z}}}}^{\b}_{s,{s\ell(T)}}(z_2)
{\mathrm E}_{z_1}\left[u_0\left(\frac{B_{tT-s}}{\sqrt{T}}\right)\right]
{\mathrm E}_{z_2}\left[u_0\left(\frac{B_{tT-s}}{\sqrt{T}}\right)\right]}\\
&=T \int_{\tau_0}^{\tau}\dd \sigma \int_{({\mathbb{R}}^2)^2}\dd x\dd y
f(x)f(y)
I_{T\sigma}^{(T)}(x)
I_{T\sigma }^{(T)}(y)\\
&\hspace{4em}\int_{({\mathbb{R}}^2)^2}\dd z\dd w \rho_\sigma(z-x)\rho_\sigma(w-y)V(z_T-w_T)\notag\\
&{\hspace{7em} \times
\overleftarrow{{\bf{\mathcal{Z}}}}^{\b}_{T\sigma,{T\sigma\ell(T)}}(z_T)
\overleftarrow{{\bf{\mathcal{Z}}}}^{\b}_{T\sigma,{T\sigma\ell(T)}}(w_T)
\bar{u}(t-\sigma,z)\bar{u}(t-\sigma,w)
}.
\end{align*}
We define for $x,y\in {\mathbb{R}}^2$ and $\tau_0\leq \sigma \leq \tau$\begin{align*}
\Psi_{{ \sigma}}^T(x,y)
&=T\int_{({\mathbb{R}}^2)^2}\dd z\dd w\rho_\sigma(z-x)\rho_\sigma(w-y)V(z_T-w_T)\notag\\
&{\hspace{4em}\times
\overleftarrow{{\bf{\mathcal{Z}}}}^{\b}_{T\sigma,{T\sigma\ell(T)}}(z_T)
\overleftarrow{{\bf{\mathcal{Z}}}}^{\b}_{T\sigma,{T\sigma\ell(T)}}(w_T)
\bar{u}(t-\sigma,z)\bar{u}(t-\sigma,w)
}.
\end{align*}
\begin{lemma}\label{lem:approxPsi}
For each $x,y\in{\mathbb{R}}^2$, $\sigma \in [\tau_0,\tau]$
\begin{align*}
\lim_{T\to\infty}\mathbb E\left[\left|\Psi^T_\sigma(x,y)-\Psi_\sigma(x,y)\right|\right]=0,
\end{align*}
where \begin{align*}
\Psi_{{\sigma}}(x,y)=\frac{1}{1-\hat{\b}^2}\int_{{\mathbb{R}}^2}\dd w\rho_\sigma(w-x)\rho_\sigma(w-y)\bar{u}(t-\sigma,z)^2.
\end{align*}
\end{lemma}
{ Combining this with Lemma \ref{lem:negativemoment} and \eqref{eq:Fass}, it is easy to see by the dominated convergence theorem that \begin{align*}
&\mathbb E\left[\int_{\tau_0}^{\tau}\dd \sigma \int_{({\mathbb{R}}^2)^2}
\left|f(x)f(y)
I_{\sigma T}^{(T)}(x)
I_{\sigma T}^{(T)}(y)\right|\left|
\Psi_\sigma^T(x,y)-\Psi_\sigma(x,y)\right|\right]\\
&=\int_{\tau_0}^{\tau}\dd \sigma \int_{({\mathbb{R}}^2)^2}
\left|f(x)f(y)\right|
\mathbb E\left[\left|I_{\sigma T}^{(T)}(x)
I_{\sigma T}^{(T)}(y)\right|\right]\mathbb E\left[\left|
\Psi_\sigma^T(x,y)-\Psi_\sigma(x,y)\right|\right]\to 0.
\end{align*}
}
\begin{lemma}\label{lem:approxFF}
For any test function $f\in C_c^\infty({\mathbb{R}}^2)$ \begin{align*}
{ \int_{\tau_0}^\tau \dd \sigma}\int_{({\mathbb{R}}^2)^2}\dd x\dd y f(x)f(y)I_{\sigma T}^{(T)}(x)I_{\sigma T}^{(T)}(y) \Psi_\sigma(x,y)
&\approx_{L^1} { \int_{\tau_0}^\tau\dd \sigma }\int_{({\mathbb{R}}^2)^2}\dd x\dd y f(x)f(y)
I_{\sigma T}(x)I_{\sigma T}(y)
\Psi(x,y)
\end{align*}
as $T\to\infty$, where the $\approx_{L^1}$ sign means that the difference between the left and right sides goes to $0$ in $L^1$-sense.
\end{lemma}
\begin{proof}[Proof of Lemma \ref{lem:approxPsi}]
(Step 1) Letting $z=w+\frac{v}{\sqrt{T}}$,
\begin{align*}
\Psi^T_{ \sigma }(x,y)&=\int_{({\mathbb{R}}^2)^2}\dd w\dd v\, \rho_\sigma(w-y)V(v) \overleftarrow{{\bf{\mathcal{Z}}}}^{\b}_{T\sigma,{T\sigma\ell(T)}}(w_T) \bar{u}(t-\sigma,w) \notag\\
&{\hspace{4em}\times
\rho_\sigma(w+\frac{v}{\sqrt{T}}-x)\overleftarrow{{\bf{\mathcal{Z}}}}^{\b}_{T\sigma,{T\sigma\ell(T)}}(w_T+v)\bar{u}\left(t-\sigma,w+\frac{v}{\sqrt{T}}\right)
}.
\end{align*}
{ We note that
\al{
&\rho_\sigma\left(w+\frac{v}{\sqrt{T}}-x\right)
\overleftarrow{{\bf{\mathcal{Z}}}}^{\b}_{T\sigma,{T\sigma\ell(T)}}(w_T+v)
\bar{u}\left(t-\sigma,w+\frac{v}{\sqrt{T}}\right)
- \rho_\sigma(w-x)\overleftarrow{{\bf{\mathcal{Z}}}}^{\b}_{T\sigma,{ T\sigma\ell(T)}}(w_T)\bar{u}(t-\sigma,w) \\
&\leq \left|\rho_\sigma\left(w+\frac{v}{\sqrt{T}}-x\right)-\rho_\sigma(w-x)\right|
\overleftarrow{{\bf{\mathcal{Z}}}}^{\b}_{T\sigma,{T\sigma\ell(T)}}(w_T+v)
\bar{u}\left(t-\sigma,w+\frac{v}{\sqrt{T}}\right)\\
&\qquad+ \rho_\sigma(w-x)\left|\overleftarrow{{\bf{\mathcal{Z}}}}^{\b}_{T\sigma,{T\sigma\ell(T)}}(w_T+v)-\overleftarrow{{\bf{\mathcal{Z}}}}^{\b}_{T\sigma,{ T\sigma\ell(T)}}(w_T)\right|\bar{u}(t-\sigma,w)\\
&\qquad\qquad+\rho_\sigma(w-x)\overleftarrow{{\bf{\mathcal{Z}}}}^{\b}_{T\sigma,{ T\sigma\ell(T)}}(w_T)\left|\bar{u}\left(t-\sigma,w+\frac{v}{\sqrt{T}}\right)-\bar{u}(t-\sigma,w)\right|,
}
all of which converge to $0$ as $T\to\infty$ by Lemma~\ref{lem:covpart}. Since $\bar{u}$ and $\rho_\sigma(x)$ is bounded for $\sigma\in[\tau_0,\tau]$ and $x\in{\mathbb{R}}^2$, using \eqref{eq:L2bdd} and $\int V(v) \dd v=1$, we have by the dominated convergence theorem,
\begin{align*}
\Psi_\sigma^T(x,y)&\approx_{L^1}
\int_{{\mathbb{R}}^2}\dd w \dd v\,\rho_\sigma(w-x)\rho_\sigma(w-y)V(v)
\overleftarrow{{\bf{\mathcal{Z}}}}^{\b}_{T\sigma,\ell(T\sigma)}(w_T)^2
\bar{u}(t-\sigma,w)^2\\&=
\int_{{\mathbb{R}}^2}\dd w\rho_\sigma(w-x)\rho_\sigma(w-y
\overleftarrow{{\bf{\mathcal{Z}}}}^{\b}_{T\sigma,\ell(T\sigma)}(w_T)^2
\bar{u}(t-\sigma,w)^2.
\end{align*}
}
(Step 2) Since we have \begin{align*}
\overleftarrow{{\bf{\mathcal{Z}}}}^\b_{s,{ s\ell(T)}}(w)\stackrel{(d)}{=} {{\bf{\mathcal{Z}}}}^\b_{{ s\ell(T)}}(w)
\end{align*}
for each $w\in {\mathbb{R}}^2$ and $s>0$, it is enough from \eqref{eq:L2bdd} to show that for each $\sigma\in [\tau_0,\tau]$ and $x,y\in {\mathbb{R}}^2$\begin{align}
&\int_{{\mathbb{R}}^2}\dd w\rho_\sigma(w-x)\rho_\sigma(w-y)
{{\bf{\mathcal{Z}}}}^{\b}_{{ T\sigma\ell(T)}}(w_T)^2
\bar{u}(t-\sigma,w)^2\notag\\
&\approx_{L^1} \int_{{\mathbb{R}}^2}\dd w\rho_\sigma(w-x)\rho_\sigma(w-y)
\mathbb E\left[{{\bf{\mathcal{Z}}}}^{\b}_{{ T\sigma\ell(T)}}(w_T)^2\right]
\bar{u}(t-\sigma,w)^2.\label{eq:approxex}
\end{align}
It follows from the approximations: \begin{align}
&\int_{{\mathbb{R}}^2}\dd w\rho_\sigma(w-x)\rho_\sigma(w-y)\left({{\bf{\mathcal{Z}}}}^\b_{{ T\sigma\ell(T)}}(w_T) \right)^2\bar{u}(t-\sigma,w)^2\notag\\
&\approx_{L^1} \int_{{\mathbb{R}}^2}\dd w\rho_\sigma(w-x)\rho_\sigma(w-y)\left(\left(\widetilde{{\bf{\mathcal{Z}}}}^\b_{{ T\sigma\ell(T)},{ \ell'(\sigma,T)}}(w_T) \right)^2\wedge \left({{ \ell(T)}}\right)^{-\frac{1}{2}}\right)\bar{u}(t-\sigma,w)^2\label{eq:approxT}\\
&\approx_{L^1}\int_{{\mathbb{R}}^2}\dd w\rho_\sigma(w-x)\rho_\sigma(w-y)
\mathbb E\left[\left(\widetilde{{\bf{\mathcal{Z}}}}^\b_{{ T\sigma\ell(T)},{ \ell'(\sigma,T)}}(w_T) \right)^2\wedge \left({{ \ell(T)}}\right)^{-\frac{1}{2}}\right]\bar{u}(t-\sigma,w)^2,\label{eq:approxTL}
\end{align}
{ where we define for $t\geq 0$, $r>0$, $z\in {\mathbb{R}}^2$ \begin{align}
\widetilde{{\bf{\mathcal{Z}}}}_{t,r}(z)&={\widetilde{{\bf{\mathcal{Z}}}}_{t,r}^{\b}(z)}={\mathrm E}_{z}\left[\Phi_{t}^\b(B):\mathtt{F}_{t,r}(B,z) \right],\label{eq:barZ}
\end{align}
and $\mathtt{F}_{t,r}(B,z)$ is the event that Brownian motion $B$ does not escape from the open ball $B(z,r)=\{x\in {\mathbb{R}}^2:|x-y|<r\}$ up to times $t$: \begin{align*}
\mathtt{F}_{t,r}(B,z)=\{B_s\in B(z,r)\ \text{for any }s\in [0,t]\}
\end{align*}
and }we set { \begin{align*}
\ell'(\sigma,T)=\sqrt{T\sigma}\ell(T)^{\frac{1}{4}}
\end{align*}}
and we denote by $\mathcal{V}_{{ T,\sigma}}^\b(w)
=\left(\widetilde{{\bf{\mathcal{Z}}}}^\b_{({ T\sigma\ell(T)},{ \ell'(\sigma,T)})}(w_T) \right)^2\wedge \left({\ell(T)}\right)^{-\frac{1}{2}}$ for simplicity.
The following argument yields \eqref{eq:approxT}: We find
\begin{align*}
&\mathbb E\left[\left({{\bf{\mathcal{Z}}}}^\b_{T\sigma \ell(T)}(w_T) \right)^2-\left(\left(\widetilde{{\bf{\mathcal{Z}}}}^\b_{{ T\sigma\ell(T)},{ \ell'(\sigma,T)}}(w_T) \right)^2\wedge \left({{ \ell(T)}}\right)^{-\frac{1}{2}}\right)\right]\\
&\leq \mathbb E\left[\left({{\bf{\mathcal{Z}}}}^\b_{T\sigma\ell(T)}(w_T) \right)^2-\left(\widetilde{{\bf{\mathcal{Z}}}}^\b_{{ T\sigma\ell(T)},{ \ell'(\sigma,T)}}(w_T) \right)^2\right]\\
&+\mathbb E\left[\left(\widetilde{{\bf{\mathcal{Z}}}}^\b_{{ T\sigma\ell(T)},{ \ell'(\sigma,T)}}(w_T) \right)^2: \left(\widetilde{{\bf{\mathcal{Z}}}}^\b_{{ T\sigma\ell(T)},{ \ell'(\sigma,T)}}(w_T) \right)^2\geq \left({{ \ell(T)}}\right)^{-\frac{1}{2}}\right]
\end{align*}
and the last term tends to $0$ as $T\to\infty$ by Lemma \ref{lem:lp}. Furthermore,
\begin{align*}
&\mathbb E\left[\left({{\bf{\mathcal{Z}}}}^\b_{T\sigma\ell(T)}(w_T) \right)^2-\left(\widetilde{{\bf{\mathcal{Z}}}}^\b_{{T\sigma\ell(T)},{ \ell'(\sigma,T)}}(w_T) \right)^2\right]\\
&\leq 4\mathbb E\left[\left({{\bf{\mathcal{Z}}}}^\b_{{T\sigma\ell(T)}}(w_T) \right)^2\right]
\mathbb E\left[{\mathrm E}_{w_T}\left[\Phi_{{ T\sigma \ell(T)}}(B):\mathtt{F}_{{T\sigma\ell(T)},{ \ell'(\sigma,T)}}(B,w_T)^c\right]^2\right]\\
&\leq C{\mathrm E}_{w_T}\otimes {\mathrm E}_{w_T}\left[\exp\left(\b^2\int_0^{{ T\sigma \ell(T)}}V(B_s-B_s')\dd s\right): \mathtt{F}_{{ T\sigma\ell(T)},{ \ell'(\sigma,T)}}(B,w_T)^c\right]\\
&\leq C{\mathrm E}_{w_T}\otimes {\mathrm E}_{w_T}\left[\exp\left(\b^2p\int_0^{
{ T\sigma\ell(T)}}V(B_s-B_s')\dd s\right)\right]^{1/p}{\mathrm P}_{w_T}\left(\mathtt{F}_{{T\sigma\ell(T)},{ \ell'(\sigma,T)}}(B,w_T)^c\right)^{1/q}\\
&\leq C{\mathrm P}_{w_T}\left(\mathtt{F}_{{ T\sigma\ell(T)},{ \ell'(\sigma,T)}}(B,w_T)^c\right)^{1/q}\to 0
\end{align*}
where $B$ and $B'$ are independent Brownian motions starting from $w_T$ and we have used the Cauchy-Schwarz inequality in the first line and the H\"{o}lder inequality for $p,q>1$ with $\frac{1}{p}+\frac{1}{q}=1$ and the fact that there exists a { constant} $p>1$ such that \begin{align}
\varlimsup_{T\to\infty}{\mathrm E}_{x}\otimes {\mathrm E}_x\left[\exp\left(\b^2p\int_0^{{ T\sigma \ell(T)}}V(B_s-B_s')\dd s\right)\right]<\infty. \label{eq:collisionLp}
\end{align}
(Step 3)
We end the proof by showing \eqref{eq:approxTL}. First, we remark that if $|w_T-w_T'|>2({ \ell'(\sigma,T)}+R_\phi)$, then
\begin{align*}
\textrm{Cov}\left(\mathcal{V}_{{ T,\sigma}}^\b(w_T),\mathcal{V}_{{ T,\sigma}}^\b(w'_T)\right)=0,
\end{align*}
where $R_\phi$ is a constant with $\textrm{supp}\phi\subset B(0,R_\phi)$.
Therefore, \begin{align*}
&\mathbb E\left[\left(\int_{{\mathbb{R}}^2}\dd w\rho_\sigma(w-x)\rho_\sigma(w-y)\left(\mathcal{V}_{{ T,\sigma}}^\b(w_T)-\mathbb E\left[\mathcal{V}_{{ T,\sigma}}^\b(w_T)\right]\right)\bar{u}(t-\sigma,w)^2\right)^2\right]\\
&= \int_{({\mathbb{R}}^2)^2}\dd w\dd w'\rho_\sigma(w-x)\rho_\sigma(w-y)\rho_\sigma(w'-x)\rho_\sigma(w'-y)\\
&\hspace{3em}\times \textrm{Cov}\left(\mathcal{V}_{{ T,\sigma}}^\b(w_T),\mathcal{V}_{{ T,\sigma}}^\b(w'_T)\right)
\bar{u}(t-\sigma,w)^2\bar{u}(t-\sigma,w')^2\\
&\leq \int_{|w_T-w_T'|\leq 2(\ell'(\sigma,T)+R_V)}\dd w\dd w'\rho_\sigma(w-x)\rho_\sigma(w-y)\rho_\sigma(w'-x)\rho_\sigma(w'-y)\\
&\hspace{3em}\times \mathbb E\left[\mathcal{V}_{{ T,\sigma}}^\b(0)^2\right]
\bar{u}(t-\sigma,w)^2\bar{u}(t-\sigma,w')^2\\
&\leq C{ {\ell(T)}}^{\frac{1}{2}} \mathbb E\left[\mathcal{V}_{{ T,\sigma}}^\b(0)^2\right].
\end{align*}
Thus, it is enough to show that
\begin{align*}
\lim_{T\to \infty}\ell(T)^{\frac{1}{2}}\mathbb E\left[\mathcal{V}_{{T,\sigma}}(0)^2\right]=0,
\end{align*}
which follows from Lemma \ref{lem:lp} and the following:
\begin{lemma}{\cite[Lemma 3.3]{CN19}}
Let $(X_k)_{k\in{\mathbb{N}}}$ be a non-negative, uniformly integrable family of random variables. Then, for any sequence $a_k\to\infty$, $a_k^{-1} \mathbb E[(X_k\land a_k)^2]\to 0$ as $k\to\infty$.
\end{lemma}
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:approxFF}]
The proof is essentially the same as in Lemma \ref{lem:approxPsi}.
Indeed, we can approximate $I_{\sigma T}(x)$ by \begin{align*}
\left(F'\left({\widetilde{{\bf{\mathcal{Z}}}}_{{ T\sigma\ell(T)},{ \ell'(\sigma,T)}}(x_T)}{\bar{u}(t,x)}\right){\widetilde{{\bf{\mathcal{Z}}}}_{{T\sigma\ell(T)},{ \ell'(\sigma,T)}}(x_T)}\right)\wedge {\ell(T)}^{-\frac{1}{2}}
\end{align*}
due to the same argument as (Step 2) and (Step 3) in the proof of Lemma \ref{lem:approxPsi}. In particular, we remark that its expectation converges to $I(x)$ due to Theorem \ref{thm:CSZ17b} and assumption of $F'$. We omit the detail.
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:MtildeTau0negl}]
{By \eqref{eq:Fass}, for $s\leq t T_\varepsilon,$
$$\mathbb E \left[(I_s^{(T)}(x))^2\right] \leq C \mathbb E\left[\left(\left|\log {\bf{\mathcal{Z}}}_{{s\ell (T)}}^{\b}(x_T)\right|+{\bf{\mathcal{Z}}}_{{s\ell(T)}}^{\b}(x_T)\right)^2\right]\leq C_t,$$
with some constants $C_t>0$ independent of $x,s,\varepsilon$. Hence, we have:}
\begin{align*}
&\mathbb E\left[\frac{1}{\b^2}\left(\int f(x) \mathcal{M}_{T\tau_0}(x_T)\dd x\right)^2\right]\\
&= \int_{t{m(T)}}^{\tau_0}\dd \sigma \int_{({\mathbb{R}}^2)^2}\dd x\dd y
f(x)f(y)
\mathbb E\left[I_s^{(T)}(x)I_s^{(T)}(y)
\right
\int_{{\mathbb{R}}^2}\dd w\int _{|v|\leq R_V}\dd v \rho_\sigma\left(w+\frac{v}{\sqrt{T}}-x\right)\rho_\sigma(w-y)V(v)\notag\\
&{\hspace{6em} \times \mathbb E\left[\overleftarrow{{\bf{\mathcal{Z}}}}^\b_{T\sigma,{ T\sigma\ell(T)}}(w_T+v)\overleftarrow{{\bf{\mathcal{Z}}}}^\b_{T\sigma,{ T\sigma\ell(T)}}(w_T) \right]
\bar{u}\left(t-\sigma,w+\frac{v}{\sqrt{T}}\right)\bar{u}(t-\sigma,w)
}\\
&\leq C \int_{t{m(T)}}^{\tau_0}\dd \sigma \int_{({\mathbb{R}}^2)^2}\dd x\dd y
|f(x)f(y)|
\int_{{\mathbb{R}}^2}\dd w \int_{|v|\leq R_V}\dd v\rho_\sigma\left(w+\frac{v}{\sqrt{T}}-x\right)\rho_\sigma(w-y)\notag\\
&\leq C\left(\tau_0-t{m(T)}\right),
\end{align*}
where we have used { $\overleftarrow{{\bf{\mathcal{Z}}}}^\b_{T\sigma,{ T\sigma\ell(T)}}(x)\stackrel{(d)}{=} {{\bf{\mathcal{Z}}}}^\b_{T\sigma,{ T\sigma\ell(T)}}(x)$ with \eqref{eq:L2bdd
and $\|u_0\|_\infty<\infty$ in the second line, and
\begin{align*}
& \int_{({\mathbb{R}}^2)^2}\dd x\dd y
|f(x)f(y)|
\int_{{\mathbb{R}}^2}\dd w \int_{|v|\leq R_V}\dd v\rho_\sigma\left(w+\frac{v}{\sqrt{T}}-x\right)\rho_\sigma(w-y)\\
&=\int_{({\mathbb{R}}^2)^2}\dd x\dd y |f(x)f(y)|\int_{|v|\leq R_V}\dd v \rho_{2\sigma}\left(y-x+\frac{v}{\sqrt{T}}\right)\leq \|f\|_1^2
\end{align*}
in the last line.
}
\end{proof}
\subsection{Proof of Proposition \ref{prop:replaceByM}}\label{sub:4.3}
First, we will show that the fluctuation of martingale term is negligible at short time regime.
\begin{lemma} \label{lem:vanishingSmalltSurgery}
\begin{equation*
\lim_{\epsilon\to 0} \frac{1}{\b_\varepsilon^2}\mathbb E\left[\left(\int f(x) G_{{ tT_\varepsilon m(T_\varepsilon)}}(x_{T_\varepsilon})\dd x \right)^2\right]=0.
\end{equation*}
\end{lemma}
Then, we will prove that the remainder of martingale can be comparable to $\mathcal{M}$ in the sense:
\begin{lemma}\label{lem:G-M}
\begin{align*}
\lim_{\varepsilon\to 0}\frac{1}{\b_{\varepsilon}^2}\mathbb E\left[\left(\int f(x) \left(\left(G_{tT_\varepsilon}(x_{T_\varepsilon})-G_{{ tT_\varepsilon m(T_\varepsilon)}}(x_{T_\varepsilon})\right)-\mathcal{M}_{tT_\varepsilon}(x_{T_\varepsilon})\right)\dd x \right)^2\right]=0.
\end{align*}
\end{lemma}
Lemma \ref{lem:vanishingSmalltSurgery} and Lemma \ref{lem:G-M} comclude Proposition \ref{prop:replaceByM}.
\subsubsection{Proof of Lemma \ref{lem:vanishingSmalltSurgery}}
To prove Lemma \ref{lem:vanishingSmalltSurgery}, we will introduce a new martingale: Let \begin{align}
&n(T_\epsilon)=\exp\left(-(\log T_\epsilon)^{\frac{1}{2}-\frac{\delta}{2}}\right)\notag
\intertext{ and}
&\widetilde{\mathcal{W}}_s(x)={\mathrm E}_x\left[\Phi_s(B)u_0\left(\frac{B_{tT}}{\sqrt{T}}\right):\mathtt{F}_{{ tTm(T)},{ \sqrt{tTn(T)}}}(B,x)\right].\label{eq:tildeWdef}
\end{align}
Lemma \ref{lem:vanishingSmalltSurgery} is concluded by the following two lemmas.
\begin{lemma}\label{lem:GGtilde}
For any $f\in C_c^\infty({\mathbb{R}}^2)$, $t>0$, $x\in{\mathbb{R}}^2$, and $\hat{\b}\in (0,1)$, \begin{align*}
\lim_{\varepsilon\to 0}\frac{1}{\b_\varepsilon}\mathbb E\left[\left|G_{{ tT_\varepsilon m(T_\varepsilon)}}(x)-\int_0^{{ tT_\varepsilon m(T_\varepsilon)}}F'
{ {\mathcal{W}}_u(x))}\dd \widetilde{\mathcal{W}}_u(x) \right|\right]=0.
\end{align*}
\end{lemma}
\begin{lemma}\label{lem:tildeG}
For any $f\in C_c^\infty({\mathbb{R}}^2)$, $t>0$, and $\hat{\b}\in (0,1)$, \begin{align*}
\lim_{\varepsilon\to 0}\frac{1}{\b_\varepsilon}\mathbb E\left[\left|\int_{{\mathbb{R}}^2}\dd xf(x)\int_0^{{ tT_\varepsilon m(T_\varepsilon)}}F'
{ {\mathcal{W}}_u(x_{T_\varepsilon}))}\dd \widetilde{\mathcal{W}}_u(x_{T_\varepsilon}) \right|\right]=0.
\end{align*}
\end{lemma}
\begin{proof}[Proof of Lemma \ref{lem:GGtilde}]
We have \begin{align*}
&\mathbb E\left[\left|G_{ tTm(T)}(x)-\int_0^{{ tTm(T)}}F'({\mathcal{W}}_u(x))\dd \widetilde{\mathcal{W}}_u(x) \right|^2\right]\\
&=\mathbb E\left[\left(\int_0^{tTm(T)}F'\left(\mathcal{W}_u(x)\right)\dd \mathcal{W}_u(x)-\int_0^{tTm(T)}F'\left(\mathcal{W}_u(x)\right)\dd \widetilde{\mathcal{W}}_u(x)\right)^2\right]\\
&=\mathbb E\left[\int_0^{tTm(T)}F'\left(\mathcal{W}_u(x)\right)^2\dd \left\langle \mathcal{W}(x)-\widetilde{\mathcal{W}}(x)\right\rangle_u\right]
\end{align*}
Then, the last expectation is written by
\begin{align*}
&\mathbb E\left[\int_0^{tTm(T)}\dd u F'\left(\mathcal{W}_u(x)\right)^
{\mathrm E}_{x}\otimes {\mathrm E}_{x}\left[V(B_u-\widetilde{B}_u)\Phi_u(B_u)\Phi_s(\widetilde{B}_u)u_0\left(\frac{B_{tT}}{\sqrt{T}}\right)u_0\left(\frac{\widetilde{B}_{tT}}{\sqrt{T}}\right):A_T(B,\widetilde{B},x
\right]\right],
\end{align*}
where we set \begin{align*}
A_T(B,\widetilde{B},x):=\mathtt{F}_{{ tTm(T)},{ \sqrt{tTn(T)}}}(B,x)^c\cap \mathtt{F}_{{ tTm(T)},{ \sqrt{tTn(T)}}}(\widetilde{B},x)^c.
\end{align*}
{By using $|V(B_s-\widetilde{B}_s)|\leq \|V\|_\infty$ and \eqref{eq:Fass}, it is estimated by }\begin{align}
&C\mathbb E\left[\int_0^{tTm(T)}\dd uF'\left(\mathcal{W}_u(x)\right)^2(\mathcal{W}_u(x)-\widetilde{\mathcal{W}}_u(x))^2\right]\notag\\
&\leq C\mathbb E\left[\int_0^{tTm(T)}\dd u\left(\frac{1}{\mathcal{W}_u(x)}+1\right)^2(\mathcal{W}_u(x)-\widetilde{\mathcal{W}}_u(x))^2\right]\notag\\%.
&\leq C\int_0^{tTm(T)}\dd
\mathbb E\left[\frac{\mathcal{W}_u(x)-\widetilde{\mathcal{W}}_u(x)}{\mathcal{W}_u(x)}+\mathcal{W}_u(x)-\widetilde{\mathcal{W}}_u(x)+\left(\mathcal{W}_u(x)-\widetilde{\mathcal{W}}_u(x)\right)^2\right].\label{eq:FWtWbdd}
\end{align}
It is easy to see that \begin{align}
\mathbb E\left[\left(\mathcal{W}_u(x)-\widetilde{\mathcal{W}}_u(x)\right)^2\right]&={\mathrm E}_{x}\otimes {\mathrm E}_{x}\left[\exp\left(\int_0^u \b^2V(B_s-\widetilde{B}_s)\dd s\right)u_0\left(\frac{B_{tT}}{\sqrt{T}}\right)u_0\left(\frac{\widetilde{B}_{tT}}{\sqrt{T}}\right):A_T(B,\widetilde{B},x
\right]\notag\\
&\leq C{\mathrm E}_{x}\otimes {\mathrm E}_{x}\left[\exp\left(\int_0^u \b^2V(B_s-\widetilde{B}_s)\dd s\right):A_T(B,\widetilde{B},x
\right].\label{eq:FWtW}
\end{align}
Then, H\"older's inequality yields that \begin{align}
&{\mathrm E}_{x}\otimes {\mathrm E}_{x}\left[\exp\left(\int_0^u \b^2V(B_s-\widetilde{B}_s)\dd s\right):A_T(B,\widetilde{B},x)\right]\notag\\
&\leq {\mathrm E}_{x}\otimes {\mathrm E}_{x}\left[\exp\left(\int_0^u p\b^2V(B_s-\widetilde{B}_s)\dd s\right)\right]^{\frac{1}{p}}{\mathrm P}_x\left(\mathtt{F}_{{ tTm(T)},{ \sqrt{tTn(T)}}}(B,x)^c\right)^{\frac{2}{q}},\label{eq:LW}
\end{align}
where $p,q>1$ with $\frac{1}{p}+\frac{1}{q}=1$ are chosen such that \begin{align*}
\varlimsup_{T\to\infty}{\mathrm E}_{x}\otimes {\mathrm E}_{x}\left[\exp\left(p\b^2\int_0^{Tt} V(B_u-\widetilde{B}_u)\dd u\right)\right]<\infty.
\end{align*}
\eqref{eq:FWtW} tends to $0$ as $T\to\infty$ since we know \begin{align}
{\mathrm P}_x\left(F_{{ tTm(T)},\sqrt{tTn(T)}}(B,x)^c\right)\leq C\exp\left(-\frac{n(T)}{4m(T)}\right),\label{eq:exit}
\end{align}
which decays faster than any polynomial of $T$.
By using the Cauchy-Schwarz inequality with Lemma \ref{lem:negativemoment}, we find that \eqref{eq:FWtWbdd} converges to $0$ as $T\to \infty$.
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:tildeG}]
It is enough to show that \begin{align*}
&\frac{1}{\b^2}\mathbb E\left[\left(\int_{{\mathbb{R}}^2}\dd xf(x)\int_0^{{{ tTm(T)}}}F'({\mathcal{W}}_s(x_T))\dd \widetilde{\mathcal{W}}_s(x_T)\right)^2\right]\\
&=\frac{1}{\b^2} \int_{{\mathbb{R}}^2\times {\mathbb{R}}^2}\dd x\dd yf(x)f(y)\mathbb E\left[\int_0^{{ tTm(T)}}F'({\mathcal{W}}_s(x_T))F'({\mathcal{W}}_s(y_T))\dd \langle \widetilde{\mathcal{W}}(x_T),\widetilde{\mathcal{W}}(y_T)\rangle_s\right] \to 0,
\end{align*}
as $T\to \infty$.
Since $\widetilde{\mathcal{W}}_s(x)$ and $\widetilde{\mathcal{W}}_s(y)$ is independent and hence~$\dd \langle \widetilde{\mathcal{W}}(x),\widetilde{\mathcal{W}}(y)\rangle_u=0 $ if $|x-y|>2(\sqrt{tTn(T)}+R_\phi)$, we have \begin{align}
\mathbb E\left[\left(\int_{{\mathbb{R}}^2}\dd xf(x)\int_0^{{ tTm(T)}}F'({\mathcal{W}}_s(x_T))\dd \widetilde{\mathcal{W}}_s(x_T)\right)^2\right]
&\leq Cn(T)\mathbb E\left[\int_0^{{ tTm(T)}}F'({\mathcal{W}}_s(x))^2\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right]\notag\\
&\leq Cn(T)\mathbb E\left[\int_0^{{ tTm(T)}}\left(\frac{1}{{\mathcal{W}}_s(x)^2}+1\right)\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right].\label{eq:L2error1}
\end{align}
By construction of $\widetilde{W}_s(x)$, we have \begin{align*}
&Cn(T)\mathbb E\left[\int_0^{{ tTm(T)}}\left(\frac{1}{{\mathcal{W}}_s(x)^2}+1\right)\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right]\leq Cn(T)\mathbb E\left[\int_0^{{ tTm(T)}}\left(\frac{1}{{\mathcal{W}}_s(x)^2}+1\right)\dd \langle {\mathcal{W}}(x)\rangle_s\right]
\intertext{and furthermore, by applying It\^{o}'s lemma to $\log {\mathcal{W}}(x)$ and ${\mathcal{W}}(x)^2$, it is bounded by}
& Cn(T)\mathbb E\left[\log \bar{u}(t,x)-\log {\mathcal{W}}_{{ tTm(T)}}+{\mathcal{W}}_{{ tTm(T)}}(x)^2 \right]\leq Cn(T).
\end{align*}
Thus, Lemma \ref{lem:tildeG} is concluded.
\iffalse{ We have that \begin{align*}
&\frac{1}{\b^2}\mathbb E\left[\left(\int f(x) G_{m(T)}(x_T)\dd x \right)^2\right]\\
&=\int_0^{m(T)}\dd s\int_{({\mathbb{R}}^2)^2}\dd x\dd yf(x)f(y)\\
& \hspace{2em}\times \mathbb E\left[F'(\mathcal{W}_s(x_T))F'(\mathcal{W}_s(y_T)){\mathrm E}_{x_T,y_T}^{\otimes 2}\left[V(B_s-\widetilde{B}_s)\Phi_s(B)\Phi_s(\widetilde{B})u_0\left(\frac{B_{t T}}{\sqrt{T}}\right)u_0\left(\frac{\widetilde{B}_{t T}}{\sqrt{T}}\right)\right]\right]\\
&\leq C\int_0^{m(T)}\dd s\int_{({\mathbb{R}}^2)^2}\dd x\dd y|f(x)f(y)|\\
& \hspace{2em}\times \mathbb E\left[\left(\frac{1}{{\bf{\mathcal{Z}}}_s(x_T)}+1\right)\left(\frac{1}{{\bf{\mathcal{Z}}}_s(y_T)}+1\right){\mathrm E}_{x_T,y_T}^{\otimes 2}\left[V(B_s-\widetilde{B}_s)\Phi_s(B)\Phi_s(\widetilde{B})\right]\right].
\end{align*}
{ We can use the concentration inequality for ${\bf{\mathcal{Z}}}_s$: \begin{align*}
{\mathbb P}\left( {\bf{\mathcal{Z}}}_s(0)\leq L(T) \right)\leq e^{-C(\log L(T))^{2}}.
\end{align*}
Define the event $A_s(x)=\{{\bf{\mathcal{Z}}}_s(x)\leq L(T)\}$. Then, we have that \begin{align}
\frac{{\mathrm E}_{x_T,y_T}^{\otimes 2}\left[V(B_s-\widetilde{B}_s)\Phi_s(B)\Phi_s(\widetilde{B})\right]}{{\bf{\mathcal{Z}}}_s(x_T){\bf{\mathcal{Z}}}_s(y_T)}&\leq \|V\|_\infty\label{eq:denobdd}\\
\frac{{\mathrm E}_{x_T,y_T}^{\otimes 2}\left[V(B_s-\widetilde{B}_s)\Phi_s(B)\Phi_s(\widetilde{B})\right]}{{\bf{\mathcal{Z}}}_s(x_T)}&\leq \|V\|_\infty {\bf{\mathcal{Z}}}_s(y_T)\notag\\
\frac{{\mathrm E}_{x_T,y_T}^{\otimes 2}\left[V(B_s-\widetilde{B}_s)\Phi_s(B)\Phi_s(\widetilde{B})\right]}{{\bf{\mathcal{Z}}}_s(y_T)}&\leq \|V\|_\infty {\bf{\mathcal{Z}}}_s(x_T).\notag
\end{align}
Thus, we have \begin{align*}
&\int_0^{m(T)}\dd s\int_{({\mathbb{R}}^2)^2}\dd x\dd y|f(x)||f(y)|\mathbb E\left[ \frac{{\mathrm E}_{x_T,y_T}^{\otimes 2}\left[V(B_s-\widetilde{B}_s)\Phi_s(B)\Phi_s(\widetilde{B})\right]}{{\bf{\mathcal{Z}}}_s(x_T){\bf{\mathcal{Z}}}_s(y_T)} \right]\\
&\leq \frac{1}{L(T)^2}\int_0^{m(T)}\dd s\int_{({\mathbb{R}}^2)^2}\dd x\dd y|f(x)f(y)|\mathbb E\left[{\mathrm E}_{x_T,y_T}^{\otimes 2}\left[V(B_s-\widetilde{B}_s)\Phi_s(B)\Phi_s(\widetilde{B})\right]\right]\\
&+\int_0^{m(T)}\dd s\int_{({\mathbb{R}}^2)^2}\dd x\dd y|f(x)f(y)|\|V\|_\infty{\mathbb P}\left(A_s(x_T)^c\cup A_s(y_T)^c\right),
\intertext{and}
&\int_0^{m(T)}\dd s\int_{({\mathbb{R}}^2)^2}\dd x\dd y|f(x)||f(y)|\mathbb E\left[ \frac{{\mathrm E}_{x_T,y_T}^{\otimes 2}\left[V(B_s-\widetilde{B}_s)\Phi_s(B)\Phi_s(\widetilde{B})\right]}{{\bf{\mathcal{Z}}}_s(x_T)} \right]\\
&\leq \frac{1}{L(T)}\int_0^{m(T)}\dd s\int_{({\mathbb{R}}^2)^2}\dd x\dd y|f(x)f(y)|\mathbb E\left[{\mathrm E}_{x_T,y_T}^{\otimes 2}\left[V(B_s-\widetilde{B}_s)\Phi_s(B)\Phi_s(\widetilde{B})\right]\right]\\
&+\int_0^{m(T)}\dd s\int_{({\mathbb{R}}^2)^2}\dd x\dd y|f(x)f(y)|\|V\|_\infty\mathbb E\left[{\bf{\mathcal{Z}}}_{s}(y_T):A_s(x_T)^c\right].
\end{align*}
Here, we remark that \begin{align*}
&\int_0^{m(T)}\dd s\int_{({\mathbb{R}}^2)^2}\dd x\dd y|f(x)f(y)|\mathbb E\left[{\mathrm E}_{x_T,y_T}^{\otimes 2}\left[V(B_s-\widetilde{B}_s)\Phi_s(B)\Phi_s(\widetilde{B})\right]\right]\\
&=\mathbb E\left[\left(\int_{{\mathbb{R}}^2}|f(x)|({\bf{\mathcal{Z}}}_{m(T)}(x_T)-1)\dd x\right)^2\right]\\
&=\sum_{n=1}^\infty \b^{2n}\int_{0<s_1<\cdots<s_n<m(T)}\int_{\left({\mathbb{R}}^2\right)^2}\dd x\dd y|f(x)f(y)|\\
&\hspace{2em}\times \int_{({\mathbb{R}}^2)^n}\dd \mathbf{s} \dd\mathbf{x}p_{2s}\left(x_T-y_T-\sqrt{2}x_1\right)V(\sqrt{2}x_1)\times \prod_{i=2}^n\left(V(\sqrt{2}x_i)p_{s_i-s_{i-1}}(x_i-x_{i-1})\right)\\
&\leq \sum_{n=1}^{\infty} \b^{2n} C\left(\frac{\log m(T)}{4\pi}\right)^{n-1}\\
&\leq C\b^2\to 0,
\end{align*}
as $T\to \infty$, where we have used the fact \begin{align*}
&\sup_{|x_1|\leq R_V}\int_{0<s<m(T)}\int_{\left({\mathbb{R}}^2\right)^2}\dd x\dd y|f(x)f(y)|p_{2s}\left(x_T-y_T-\sqrt{2}x_1\right)V(\sqrt{2}x_1)\\
&\leq \|V\|_\infty \sup_{|x_1|\leq R_V}\int_{0<s<m(T)}\int_{\left({\mathbb{R}}^2\right)^2}\dd x\dd y|f(x)f(y)|p_{2s}\left(x_T-y_T-\sqrt{2}x_1\right)\\
&\leq \|V\|_\infty \sup_{|x_1|\leq R_V}\int_0^{\frac{m(T)}{T}}\dd \sigma\int_{({\mathbb{R}}^2)^2}\dd x\dd y|f(x)||f(y)|\frac{1}{2\pi \sigma}\exp\left(-\frac{\left|x-y-\frac{\sqrt{2}x_1}{\sqrt{T}}\right|^2}{2\sigma}\right)\\
&\leq \|V\|_\infty \|f\|_1\|f\|_\infty \frac{m(T)}{T}.
\end{align*}
Also, we have \begin{align*}
\mathbb E\left[{\bf{\mathcal{Z}}}_{s}(y_T):A_s(x_T)^c\right]\leq \mathbb E\left[{\bf{\mathcal{Z}}}_s(y_T)^2\right]^{\frac{1}{2}}{\mathbb P}\left(A_s(x_T)^c\right)^\frac{1}{2}\to 0.
\end{align*}
}
}\fi
\end{proof}
\iffalse{
We begin by making a few observations that we will use repeatedly in the following. For $\varepsilon>0$, define the event
$$A_s(x)=\{ {\bf{\mathcal{Z}}}_s(\sqrt{T}x)^{-1} \leq T^{\frac{\epsilon}{8}} \}.$$
By Theorem 1.3 in \cite{CCM20}{, we know that ${\bf{\mathcal{Z}}}_\infty$ admits all negative moments. Hence by Doob's inequality applied to the submartingale $({\bf{\mathcal{Z}}}^{-1}_s)_{s\geq 0}$, we find using} shift invariance that for all $\varepsilon>0$, for sufficiently large $T$,
\begin{equation} \label{eq:negligibleEvent}
\sup_{x\in \mathbb R^d} \sup_{s\geq 0} {\mathbb P}\left(A_s(x)^c \right)\leq T^{-2d}.
\end{equation}
Moreover, observe that for all $x,y$, almost-surely
\begin{equation}\label{uniform-estFLAT}
\begin{split}
&\quad \left(\mathcal{W}_s(\sqrt{T}x)\mathcal{W}_s(\sqrt{T}y)\right)^{-1
{\left\langle \mathcal{W}(\sqrt{T}x),\mathcal{W}(\sqrt{T}y) \right\rangle_s'}\\
&= \frac{{\mathrm E}_{\sqrt{T}x,\,\sqrt{T} y}^{\otimes 2} \left[\Phi_s^{(1)} \Phi_s^{(2)} V(B^{(1)}_s-B^{(2)}_s) u_0\left(T^{-\frac{1}{2}}\,B_{t T}^{(1)}\right)u_0\left(T^{-\frac{1}{2}}\,B_{t T}^{(2)}\right) \right]}{{\mathrm E}_{\sqrt{T}x,\,\sqrt{T} y}^{\otimes 2} \left[\Phi_s^{(1)} \Phi_s^{(2)} u_0\left(T^{-\frac{1}{2}}\,B_{t T}^{(1)}\right)u_0\left(T^{-\frac{1}{2}}\,B_{t T}^{(2)}\right) \right]}\leq \Vert V \Vert_\infty,
\end{split}
\end{equation}
while expanding as in \eqref{eq:bracketExpansion_intro} and thanks to \eqref{eq:unifboundForP2P}, we have for all $x,y\in \mathbb R^d$ and $s\geq 0$,
\begin{align}
\mathbb E
{\langle \mathcal{W}(\sqrt T x), \mathcal{W}(\sqrt T y) \rangle_{s}'} & \leq C T^{-d/2} \int_{(\mathbb{R}^d)^2} V(Z) \rho_\frac{s}{T}(z -x) \rho_{\frac{s}{T}}(z + T^{-\frac{1}{2}} Z - y) \dd z \dd Z\nonumber\\
&{=C T^{-d/2} \int_{\mathbb{R}^d} V(Z) \rho_{\frac{2s}{T}}(x + T^{-\frac{1}{2}} Z - y) \dd Z,}
\label{eq:bracketExpansion_W}
\end{align}
for some finite $C=C(\b,\Vert u_0 \Vert_\infty)$.
Now, from the last expression, we see that there is $C=C(\b, \|f\|_\infty,\|f\|_1,\|u_0\|_\infty,\|V\|_1)$ which is finite and such that
\begin{equation} \label{eq:boundOnBracketAt0}
\mathbb E\left[\int_0^{T^{1-\epsilon}}\int f(x)f(y) \dd \langle \mathcal{W}(\sqrt{T}x),\mathcal{W}(\sqrt{T}y) \rangle_s\dd x \dd y\right]\leq C T^{-\frac{d-2}{2}-\epsilon},
\end{equation}
(in fact this is \eqref{eq:L2upperBoundZTf} with $ \tau_0=T^{-\epsilon}$). Therefore, the expectation in \eqref{eq:vanishingSmalltSurgery} equals and satisfies
\al{
&\mathbb E\left[\int_0^{T^{1-\epsilon}}\int f(x)f(y) \frac{\dd \langle \mathcal{W}(\sqrt{T}x),\mathcal{W}(\sqrt{T}y) \rangle_s}{\mathcal{W}_s(\sqrt{T}x)\mathcal{W}_s(\sqrt{T}y)}\dd x \dd y\right]\\
&\leq T^{\frac{\epsilon}{4}}\, \int^{T^{1-\epsilon}}_0\int f(x)f(y) \mathbb E\left[
{\langle \mathcal{W}(\sqrt{T}x),\mathcal{W}(\sqrt{T}y) \rangle_s'} \right] \dd x \dd y
\\
&\hspace{8mm}+ \|V\|_\infty \int^{T^{1-\epsilon}}_0\int f(x)f(y){ {\mathbb P}\left(A_s(x)^c\cup A_s(y)^c\right)} \dd x \dd y \dd s\\
&\leq C T^{\frac{\epsilon}{4}}\, T^{-\frac{d-2}{2}-\epsilon}+ C T^{-2d+1}=o(T^{-\frac{d-2}{2}}).
}\fi
\begin{comment}
\begin{lemma}
For sufficietly large $r>0$,
\al{ \lim_{T\to\infty}T^{\frac{d-2}{2}}\mathbb E\left(\int^T_{T^{1-\epsilon}} \int f(x) \frac{\dd \mathcal{W}_\sigma(\sqrt{T}x)}{\mathcal{W}_\sigma(\sqrt{T}x)}\mathbf{1}(A_\sigma(x))\dd x \right)^2 =0.}
\end{lemma}
\begin{proof}
By \eqref{uniform-estFlat},
\al{
&\quad \mathbb E\left(\int^T_{T^{1-\epsilon}} \int f(x) \frac{\dd \mathcal{W}_\sigma(\sqrt{T}x)}{\mathcal{W}_\sigma(\sqrt{T}x)}\mathbf{1}(A_\sigma(x))\dd x \right)^2\\
&\leq \int^{T}_{T^{1-\epsilon}} \int f(x) f(y) \mathbb E\left[\left( \frac{ \langle \mathcal{W}_\sigma(\sqrt{T}x),\mathcal{W}_\sigma(\sqrt{T}y) \rangle^\prime}{\mathcal{W}_\sigma(\sqrt{T}x)\mathcal{W}_\sigma(\sqrt{T}y)}\right)\mathbf{1}(A_\sigma(x))\right]\dd x \dd y \dd \sigma\\
&\leq |V|_{\infty}\int^{T}_{T^{1-\epsilon}} \int f(x) f(y) {\mathbb P}\left(A_\sigma(x) \right)\dd x \dd y \dd \sigma\\
&\leq C T\, T^{-2d}=o(T^{\frac{d-2}{2}})
}
\end{proof}
\end{comment}
\subsubsection{Proof of Lemma \ref{lem:G-M}}
Define
\begin{align*}
&\dd \mathcal{L}_s(x)={\b_\varepsilon} {\bf{\mathcal{Z}}}^\b_{{ s\ell(T)}}(x)\int_{{\mathbb{R}}^d}\dd z \xi(\dd s, \dd b) \int_{{\mathbb{R}}^d}\rho_s(z-x) \phi(z-b) \overleftarrow{{\bf{\mathcal{Z}}}}^\b_{s , { s\ell(T)}}(z) \, {\mathrm E}_z\left[ u_0\left(\frac{B_{t T-s}}{\sqrt{T}}\right)\right].
\end{align*}
Then, we remark that \begin{align}
\dd \mathcal{M}_s(x_T)=F'\left({\bf{\mathcal{Z}}}^\b_{{ s\ell(T)}}\bar{u}(t,x)\right)\dd \mathcal{L}_s(x_T)\label{eq:relLM}
\end{align}
for $s\geq tTm(T)$ and $x\in{\mathbb{R}}^2$.
Lemma \ref{lem:G-M} follows when the next two lemmas are proved.
\begin{lemma}\label{Diff W and L}
For all $t>0$,
\al{
\lim_{\varepsilon\to 0}\frac{1}{\b_\varepsilon}\mathbb E\left[\left| \int \dd xf(x) \left(\int_{{ tT_\varepsilon m(T_\varepsilon)}}^{tT_\varepsilon}F'(\mathcal{W}_s(x_{T_\varepsilon}))\dd \mathcal{W}_{s}(x_{T_\varepsilon})-\int_{{tT_\varepsilon m(T_\varepsilon)}}^{tT_\varepsilon}F'(\mathcal{W}_s(x_{T_\varepsilon}))\dd \mathcal{L}_s(x_{T_\varepsilon})\right) \right|\right] =0.
}
\end{lemma}
\begin{lemma}\label{LAST LEMMA}
For $t>0$,
\al{ \lim_{\varepsilon\to 0}\frac{1}{\b_\varepsilon}\mathbb E\left[\left| \int f(x) \left(\int_{{ tT_\varepsilon m(T_\varepsilon)}}^{tT_\varepsilon}F'(\mathcal{W}_s(x_{T_\varepsilon}))\dd \mathcal{L}_s(x_{T_\varepsilon})- \mathcal{M}_{tT_\varepsilon}(x_{T_\varepsilon})\right) \dd x \right|\right] =0.}
\end{lemma}
\begin{proof}[Proof of Lemma \ref{Diff W and L}] By the Burkholder-Davis-Gundy inequality, we have
\al{
&\frac{1}{\b}\mathbb E\left[\left| \int \dd xf(x) \left(\int_{{ tTm(T)}}^{tT}F'(\mathcal{W}_s(x_T))\dd \mathcal{W}_{s}(x_T)-\int_{{ tTm(T)}}^{tT}F'(\mathcal{W}_s(x_T))\dd \mathcal{L}_s(x_T)\right) \right|\right] \\
&\leq \frac{C}{\b}\int \dd x|f(x)|\mathbb E\left[\left(\int_{{ tTm(T)}}^{tT} F'(\mathcal{W}_s(x_T))^2\dd \left\langle\mathcal{W}(x_T)-\mathcal{L}(x_T)\right\rangle_s\right)^{\frac{1}{2}}\right]\\
&\leq \frac{C}{\b}\int \dd x|f(x)|\mathbb E\left[\sup_{{{ tTm(T)}} \leq s\leq T}{{\bf{\mathcal{Z}}}_s(0)^
{-2}}+1\right]^{\frac{1}{2}}\mathbb E\left[\int_{{ tTm(T)}}^{tT} \dd \left\langle\mathcal{W}(x_T)-\mathcal{L}(x_T)\right\rangle_s\right]^{\frac{1}{2}}\\
&\leq \frac{C}{\b}\int \dd x |f(x)|\mathbb E\left[\int_{{ tTm(T)}}^{tT} \dd \left\langle\mathcal{W}(x_T)-\mathcal{L}(x_T)\right\rangle_s\right]^{\frac{1}{2}},
}
where we have used Doob's inequality and Lemma \ref{lem:negativemoment} in the last inequality.
By Cauchy-Schwarz inequality and boundedness of $\bar{u}$, we have from definition of $\mathcal{W}$ and $\mathcal{L}$ that for $x\in {\mathbb{R}}^2$\begin{align*}
&\frac{1}{\b_\varepsilon^2}\mathbb E\left[\int_{{ tTm(T)}}^{tT} \dd \left\langle\mathcal{W}(x_T)-\mathcal{L}(x_T)\right\rangle_s\right]\\
&= \int_{{tTm(T)}}^{tT}\dd s\int_{({\mathbb{R}}^2)^2}\dd z_1\dd z_2 \rho_s(z_1)\rho_s(z_2)V(z_1-z_2)\bar{u}(tT-s,z_1)\bar{u}(tT-s,z_2)\\
&\hspace{2em}\times \mathbb E\left[\left({\mathrm E}_{0,0}^{s,z_1}{\left[\Phi_s^\b\right]}-{\bf{\mathcal{Z}}}_{{ s\ell(T)}}(0)\overleftarrow{{\bf{\mathcal{Z}}}}_{s,{ s\ell(T)}}(z_1)\right)\left({\mathrm E}_{0,0}^{s,z_2}\left[\Phi_s^\b\right]-{\bf{\mathcal{Z}}}_{{ s\ell(T)}}(0)\overleftarrow{{\bf{\mathcal{Z}}}}_{s,{ s\ell(T)}}(z_2)\right)\right]\\
&\leq C\int_{{ tTm(T)}}^{tT}\dd s\int_{({\mathbb{R}}^2)^2}\dd z_1\dd z_2 \rho_s(z_1)\rho_s(z_2)V(z_1-z_2)\\
&\hspace{2em}\times \mathbb E\left[\left|{\mathrm E}_{0,0}^{s,z_1}\left[\Phi_s^\b\right]-{\bf{\mathcal{Z}}}_{{ s\ell(T)}}(0)\overleftarrow{{\bf{\mathcal{Z}}}}_{s,{ s\ell(T)}}(z_1)\right|\left|{\mathrm E}_{0,0}^{s,z_2}\left[\Phi_s^\b\right]-{\bf{\mathcal{Z}}}_{{ s\ell(T)}}(0)\overleftarrow{{\bf{\mathcal{Z}}}}_{s,{ s\ell(T)}}(z_2)\right|\right]\\
&\leq C\int_{{{{ tm(T)}}}}^t\dd \sigma\int_{({\mathbb{R}}^2)^2}\dd w\dd v
\rho_{\sigma}(w)\rho_\sigma(w+\frac{v}{\sqrt{T}})V(v)\\
&\hspace{2em}\times \mathbb E\left[\left({\mathrm E}_{0,0}^{T\sigma,w_T}\left[\Phi_{T\sigma}^\b\right]-{\bf{\mathcal{Z}}}_{{ s\ell(T)}}(0)\overleftarrow{{\bf{\mathcal{Z}}}}_{T\sigma,{ T\sigma\ell(T)}}(w_T)\right)^2\right]^{\frac{1}{2}}\\
&\hspace{2em}\times \mathbb E\left[\left({\mathrm E}_{0,0}^{T\sigma,w_T+v}\left[\Phi_{T\sigma}^\b\right]-{\bf{\mathcal{Z}}}_{{ T\sigma\ell(T)}}(0)\overleftarrow{{\bf{\mathcal{Z}}}}_{T\sigma,{ T\sigma\ell(T)}}(w_T+v)\right)^2\right]^\frac{1}{2}\\
&\leq C\int_{{{{ tm(T)}}}}^t\dd \sigma\int_{{\mathbb{R}}^2}\dd w
\rho_{\sigma}(w
\mathbb E\left[\left({\mathrm E}_{0,0}^{T\sigma,w_T}\left[\Phi_{T\sigma}^\b\right]-{\bf{\mathcal{Z}}}_{{ s\ell(T)}}(0)\overleftarrow{{\bf{\mathcal{Z}}}}_{T\sigma,{ T\sigma\ell(T)}}(w_T)\right)^2\right]^{\frac{1}{2}
\intertext{and furthermore Lemma \ref{th:errorTermLLT} allows us to bound it from above by }
&C\int_{tm(T)}^t\dd \sigma \int_{|w_T|\leq \sqrt{\sigma T\log (\sigma T)}}\dd w\rho_\sigma(w)\left(\ell(T)-\frac{\log \ell(T)}{\log T}+\frac{\sqrt{\log (\sigma T)}\log( \sigma T\log \ell(T))}{\sqrt{\sigma T}\log T}+\frac{\ell(T)\log (\sigma T)}{\log T}\right)\\
&\hspace{3em}+C\int_{tm(T)}^t\dd \sigma \int_{|w_T|\geq \sqrt{\sigma T\log( \sigma T) }}\dd w\rho_{\sigma}(w)\\
&\leq C\left(\ell(T)-\frac{\log \ell(T)}{\log T}+\frac{\sqrt{\log (\sigma T)}\log( \sigma T\log \ell(T))}{\sqrt{\sigma T}\log T}+\frac{\ell(T)\log (\sigma T)}{\log T}\right)+C\int_{t{m(T)}}^t \frac{1}{\sqrt{\sigma T}}\dd \sigma.
\end{align*}
Since both terms in the last line tend to $0$ as $T\to \infty$ from {definition of $m(T)$ and $\ell(T)$}, Lemma {\ref{Diff W and L}} is concluded.
\end{proof}
\begin{proof}[Proof of Lemma \ref{LAST LEMMA}]
We have from \eqref{eq:relLM}
\begin{align}
&\frac{1}{\b}\mathbb E\left[\left| \int f(x) \left(\int_{{ tTm(T)}}^{tT}F'(\mathcal{W}_s(x_T))\dd \mathcal{L}_s(x_T)- \mathcal{M}_{tT}(x_T)\right) \dd x \right|\right]\notag\\
&\leq \frac{1}{\b}\int |f(x)| \mathbb E\left[\left| \int_{{ tTm(T)}}^{tT}F'(\mathcal{W}_s(x_T))\dd \mathcal{L}_s(x_T)- \mathcal{M}_{tT}(x_T)\right|\right] \dd x \notag\\
&\leq \frac{1}{\b}\int \dd x|f(x)|\mathbb E\left[\int_{{ tTm(T)}}^{tT}\left(F'(\mathcal{W}_s(x_T))-F'({\bf{\mathcal{Z}}}_{{ s\ell(T)}}(x_T)\bar{u}(t,x))\right)^2\dd \langle \mathcal{L}(x_T)\rangle_s\right]^\frac{1}{2}\notag
\end{align}
and \begin{align*}
&\mathbb E\left[\int_{tTm(T)}^{tT}\left(F'(\mathcal{W}_s(x_T))-F'({\bf{\mathcal{Z}}}_{{ s\ell(T)}}(x_T)\bar{u}(t,x))\right)^2\dd \langle \mathcal{L}(x_T)\rangle_s\right]\\
&= \b^2\int_{tm(T)}^{t}\dd \sigma \int_{({\mathbb{R}}^2)^2}\dd z\dd v\rho_\sigma (z)\rho_\sigma(z+\frac{v}{\sqrt{T}})V(v){\mathrm E}_{z_T}\left[u_0\left(\frac{B_{tT-\sigma T}}{\sqrt{T}}\right)\right]{\mathrm E}_{z_T+v}\left[u_0\left(\frac{B_{tT-\sigma T}}{\sqrt{T}}\right)\right]\\
&\hspace{5em}\times \mathbb E\left[\left(F'(\mathcal{W}_{T\sigma }(x_T))-F'({\bf{\mathcal{Z}}}_{{ T\sigma \ell(T)}}(x_T)\bar{u}(t,x))\right)^2{\bf{\mathcal{Z}}}_{{ T\sigma \ell(T)}}(x_T)^2\overleftarrow{{\bf{\mathcal{Z}}}}_{T\sigma,T\sigma \ell(T)}(z_T)\overleftarrow{{\bf{\mathcal{Z}}}}_{T\sigma,T\sigma \ell(T)}(z_T+v)\right]
\end{align*}
By \eqref{eq:Fass} and Lemma \ref{lem:negativemoment}, we may choose $p,q>1$ such that $\frac{1}{p}+\frac{1}{q}=1$ and there exists a constant $C$ such that\begin{align*}
&\mathbb E\left[\left|F'(\mathcal{W}_{T\sigma }(x_T))-F'({\bf{\mathcal{Z}}}_{{ T\sigma \ell(T)}}(x_T)\bar{u}(t,x))\right|^{2q}\right]^{\frac{1}{q}}\mathbb E\left[{\bf{\mathcal{Z}}}_{{ T\sigma \ell(T)}}(x_T)^{2p}\overleftarrow{{\bf{\mathcal{Z}}}}_{T\sigma,T\sigma \ell(T)}(z_T)^p\overleftarrow{{\bf{\mathcal{Z}}}}_{T\sigma,T\sigma \ell(T)}(z_T+v)^p\right]^{\frac{1}{p}}\\
&\leq C
\end{align*}
uniformly in $tm(T)\leq \sigma \leq t$, $x\in {\mathbb{R}}^2$ and in $0<\varepsilon<\frac{1}{2}$.
Also, \eqref{eq:Fass} and $\displaystyle \left|F'(\mathcal{W}_{T\sigma }(x_T))-F'({\bf{\mathcal{Z}}}_{{ T\sigma \ell(T)}}(x_T)\bar{u}(t,x))\right|=\left|\int_{{\bf{\mathcal{Z}}}_{T\sigma \ell(T)}(x_T)}^{\mathcal{W}_{T\sigma}(x_T)}F''(w)\dd w\right|$ yield
\begin{align*}
&\mathbb E\left[\left|F'(\mathcal{W}_{T\sigma }(x_T))-F'({\bf{\mathcal{Z}}}_{{ T\sigma \ell(T)}}(x_T)\bar{u}(t,x))\right|^{2q}\right]\\
&\leq C\mathbb E\left[\left|\frac{1}{\mathcal{W}_{T\sigma }(x_T)}+\frac{1}{{\bf{\mathcal{Z}}}_{{ T\sigma \ell(T)}}(x_T)\bar{u}(t,x)}+1\right|^{2q-1}\left|F'(\mathcal{W}_{T\sigma }(x_T))-F'({\bf{\mathcal{Z}}}_{{ T\sigma \ell(T)}}(x_T)\bar{u}(t,x))\right|\right]\\
&\leq C\mathbb E\left[\sup_{tm(T)\leq \sigma\leq t}\left\{\left|\frac{1}{{\bf{\mathcal{Z}}}_{T\sigma }(x_T)}+\frac{1}{{\bf{\mathcal{Z}}}_{{ T\sigma \ell(T)}}(x_T)}+1\right|^{2q-1}\left|\frac{1}{{\bf{\mathcal{Z}}}_{T\sigma }(x_T)^2}+\frac{1}{{\bf{\mathcal{Z}}}_{{ T\sigma \ell(T)}}(x_T)^2}+1\right|\right\}\right.\\
&\hspace{5em}\times \left|\mathcal{W}_{T\sigma }(x_T)-{\bf{\mathcal{Z}}}_{{ T\sigma \ell(T)}}(x_T)\bar{u}(t,x) \right| \Bigg].
\end{align*}
Then, H\"older's inequality and Doob's inequality guarantee with Lemma \ref{lem:negativemoment} and Theorem \ref{th:errorTermLLT} that the last term converges to $0$ as $T\to \infty$ and thus Lemma \ref{LAST LEMMA} follows by the dominated convergence theorem.
\iffalse{
We have from \eqref{eq:Wexpan} and local limit theorem that
\begin{align*}
&\mathbb E\left[\left(\mathcal{W}_s(x_T)-{\bf{\mathcal{Z}}}_{ s\ell(T)}^\b(x_T)\bar{u}(t,x)\right)^2\right]\\
&= \mathbb E\left[\left(\int_{{\mathbb{R}}^d}\rho_s(z-x_T) {\mathrm E}_{0,x_T}^{s,z}[\Phi^\b_s] {\mathrm E}_z\left[ u_0\left(\frac{B_{t T-s}}{\sqrt{T}}\right)\right] \dd z -{\bf{\mathcal{Z}}}_{s\ell(T)}^\b(x_T)\bar{u}(t,x)\right)^2\right]\\
&\leq 2\mathbb E\left[\left(\int_{{\mathbb{R}}^d}\rho_s(z-x_T) \left({\mathrm E}_{0,x_T}^{s,z}[\Phi^\b_s] - {\bf{\mathcal{Z}}}_{s\ell(T)}^\b(x_T)\overleftarrow{{\bf{\mathcal{Z}}}}_{s,s\ell(T)}^\b(z) \right){\mathrm E}_z\left[ u_0\left(\frac{B_{t T-s}}{\sqrt{T}}\right)\right] \dd z \right)^2\right]\\
&+2\mathbb E\left[\left(\int_{{\mathbb{R}}^d}\rho_s(z-x_T) {\bf{\mathcal{Z}}}_{s\ell(T)}^\b(x_T)\overleftarrow{{\bf{\mathcal{Z}}}}_{s,s\ell(T)}^\b(z) {\mathrm E}_z\left[ u_0\left(\frac{B_{t T-s}}{\sqrt{T}}\right)\right] \dd z-{\bf{\mathcal{Z}}}_{s\ell(T)}^\b(x_T)\bar{u}(t,x)\right)^2\right].
\end{align*}
Applying the same argument in the proof pf Lemma \ref{lem:approxPsi} and local limit theorem (Theorem \ref{th:errorTermLLT}), both terms tend to $0$ as $T\to \infty$. Thus, Lemma \ref{LAST LEMMA} follows from Lemma \ref{lower-tail}.
}\fi
\iffalse{ where we have used Burkh\"older-Davis-Gundy in the last line. For ease of notation, we simply write $\mathcal{W}_s=\mathcal{W}_s(\sqrt{T}x)$ and ${\bf{\mathcal{Z}}}_s={\bf{\mathcal{Z}}}_{s}(\sqrt{T}x)$ in the proof. By Cauchy-Schwarz inequality, the expectation inside the integral of \eqref{PPPP} is further bounded from above by
\aln{
&\quad \mathbb E\left[\left(\sup_{\ell(T) \leq s\leq T t}\left|\frac{{\bf{\mathcal{Z}}}_{\ell(T)}\, \bar{u}(t,x)}{\mathcal{W}_s}-1\right| \right) \left( \int^{T t}_{T^{1-\epsilon}}\dd \langle \overline{M}(\sqrt{T} x)\rangle_s \right)^{\frac{1}{2}}\right]\nonumber\\
&\leq \mathbb E\left[ \sup_{\ell(T) \leq s\leq T t}\left(\frac{{\bf{\mathcal{Z}}}_{\ell(T)}\, \bar{u}(t,x)}{\mathcal{W}_s}-1\right)^2 \right]^{\frac{1}{2}} \mathbb E\left[ \int^{T t}_{T^{1-\epsilon}}\dd \langle \overline{M}(\sqrt{T} x)\rangle_s \right]^{\frac{1}{2}}.\label{conclusion}
}
Since
$$\frac{{\bf{\mathcal{Z}}}_{\ell(T)}\, \bar{u}(t,x)}{\mathcal{W}_s}-1=\mathcal{W}_s^{-1}({\bf{\mathcal{Z}}}_{\ell(T)}\, \bar{u}(t,x)-\mathcal{W}_{T t}+(\mathcal{W}_{tT}-\mathcal{W}_s)),$$
the first term of \eqref{conclusion} can be bounded from above via Cauchy-Schwarz inequality as
\aln{
&\quad \mathbb E\left[ \sup_{\ell(T) \leq s\leq T t}\left(\frac{{\bf{\mathcal{Z}}}_{\ell(T)}\, \bar{u}(t,x) }{\mathcal{W}_s}-1\right)^2 \right]\notag\\
&\leq \mathbb E\left[ \sup_{\ell(T) \leq s\leq T t} \left(\frac{{\bf{\mathcal{Z}}}_{\ell(T)}\, \bar{u}(t,x)-\mathcal{W}_s}{\mathcal{W}_s^{2}}\right)^2 \right]^{\frac{1}{2}} \mathbb E\left[ \sup_{\ell(T) \leq s\leq T t}\left( {\bf{\mathcal{Z}}}_{\ell(T)}\, \bar{u}(t,x)-\mathcal{W}_{T t}+(\mathcal{W}_{tT}-\mathcal{W}_s) \right)^2 \right]^{\frac{1}{2}}\notag \\
&\leq C \left(\mathbb E\left[\left({\bf{\mathcal{Z}}}_{\ell(T)}(\sqrt{T}x)\, \bar{u}(t,x) -\mathcal{W}_{T t}(\sqrt{T}x)\right)^2\right]+\mathbb E\left[ \sup_{\ell(T) \leq s\leq T t}\left(\mathcal{W}_{T t} -\mathcal{W}_s\right)^2 \right]\right)^{\frac{1}{2}}, \label{star777
}
where we have used
\al{
\mathbb E\left[ \sup_{\ell(T) \leq s\leq T t} \left(\frac{{\bf{\mathcal{Z}}}_{\ell(T)}\, \bar{u}(t,x)-\mathcal{W}_s}{\mathcal{W}_s^{2}}\right)^2 \right]&\leq \mathbb E\left[ \sup_{\ell(T) \leq s\leq T t} \mathcal{W}^{-4}\right]^{1/2}\,\mathbb E\left[ \sup_{\ell(T) \leq s\leq T t} \left(\frac{{\bf{\mathcal{Z}}}_{\ell(T)}\, \bar{u}(t,x)-\mathcal{W}_s}{\mathcal{W}_s}\right)^4 \right]^{1/2}\\
&\leq C \mathbb E\left[ \sup_{\ell \leq s} \left(\frac{{\bf{\mathcal{Z}}}_{\ell}}{{\bf{\mathcal{Z}}}_s}+1\right)^4 \right]^{1/2}<\infty\\
}
by Lemma~\ref{lower-tail} below. With the $\overline{{\bf{\mathcal{Z}}}}$ and $\Gamma(s,x,y)$ notations from \eqref{eq:barZ} and \eqref{eq:DefGamma}, the first term of \eqref{star777} can be bounded from above as
\al{
&\quad \mathbb E\left[\left({\bf{\mathcal{Z}}}_{\ell(T)}(\sqrt{T}x)\, \bar{u}(t,x) -\mathcal{W}_{T t}(\sqrt{T}x)\right)^2\right]\\
&\leq 2\mathbb E\left[\left({\bf{\mathcal{Z}}}_{\ell(T)}(\sqrt{T}x)\, \bar{u}(t,x) -{\bf{\mathcal{Z}}}_{\ell(T)}(\sqrt{T}x) {\mathrm E}_{\sqrt{T}x}\left[ \overset{\leftarrow}{{\bf{\mathcal{Z}}}}_{T t,\ell(T)}(B_{T t})u_0\left(\frac{B_{T t}}{\sqrt{T}}\right)\right]\right)^2\right]\\
&\hspace{8mm} + 2\mathbb E\left[\left(\mathcal{W}_{T t}(\sqrt{T}x)-{\bf{\mathcal{Z}}}_{\ell(T)}(\sqrt{T}x) {\mathrm E}_{\sqrt{T}x}\left[ \overset{\leftarrow}{{\bf{\mathcal{Z}}}}_{T t,\ell(T)}(B_{T t})u_0\left(\frac{B_{T t}}{\sqrt{T}}\right)\right]\right)^2\right]\\
&= 2\mathbb E[{\bf{\mathcal{Z}}}_{\ell(T)}^2]\mathbb E\left[\left({\mathrm E}_{\sqrt{T}x}\left[ \left({\bf{\mathcal{Z}}}_{\ell(T)}(B_{T t}) -1 \right) u_0\left(\frac{B_{T t}}{\sqrt{T}}\right) \right]\right)^2 \right]\\
&\hspace{8mm} + 2\mathbb E\left[ {\mathrm E}_{\sqrt{T}x}\left[ \Gamma(T t,\sqrt{T}x,B_{T t}) u_0\left(\frac{B_{T t}}{\sqrt{T}}\right)\right]^2\right]\\
&= 2\mathbb E[{\bf{\mathcal{Z}}}_{\ell(T)}]^2 \mathbb E\left[\left({\mathrm E}_{\sqrt{T}x}\left[ \left(\overline{{\bf{\mathcal{Z}}}}_{\ell(T)}(B_{T t}) -\mathbb E\left[\overline{{\bf{\mathcal{Z}}}}_{\ell(T)}(B_{T t})\right] \right) u_0\left(\frac{B_{T t}}{\sqrt{T}}\right) \right]\right)^2 \right] +O(T^{-\alpha})\\
&\leq C {\mathrm P}^{\otimes 2}_{0}\left(|B_{T t}^{(1)}-B_{T t}^{(2)}|\leq \ell(t)\right)+ O(T^{-\alpha})=O(T^{-\alpha}),\\
}
with some $\alpha>0$ where we have used the local limit theorem for polymers \eqref{Application LLT KPZ} and \eqref{Difference of sZ and bar-sZ}.
For the second term of \eqref{star777}, {we use that by \eqref{eq:bracketExpansion_W} there is a constant $C=C(\b)$ such that for all $x\in{\mathbb{R}}^d$, $\sigma > 0$,
\begin{equation} \label{eq:bracketExpansion_Wbis}
\mathbb E\left[\langle \mathcal{W}(\sqrt T x) \rangle_{T\sigma }' \right] \leq C T^{-d/2} \sigma^{-d/2},
\end{equation}
and hence}
\al{
\mathbb E\left[ \sup_{\ell(T) \leq s\leq T t}\left(\mathcal{W}_{T t}-\mathcal{W}_s\right)^2 \right] &\leq C \mathbb E \left[ \int_{\ell(T)}^{T t} \dd {\langle \mathcal{W}\rangle_s} \right]\\
&{\leq C T^{-\frac{d-2}{2}} \int_{\frac{\ell (T)}{T}}^{t}\int_{{\mathbb{R}}^d}V(Z)\sigma^{-d/2}\dd Z\dd \sigma}\\
&= O\left(T^{-\frac{(d-2)a}{2}}\right).
}
Thus, the first term of \eqref{conclusion} is bounded by $C\,T^{-\alpha'}$ with some $\alpha'>0$ that depends on $a$. If $\varepsilon$ is sufficiently small depending on $a$, \eqref{conclusion} can be further bounded from above by
\al{
C T^{-\alpha'} \mathbb E\left[ \int^{T t}_{T^{1-\epsilon}} \dd \langle \overline{M}(\sqrt{T} x)\rangle_s \right]^{\frac{1}{2}}=o(T^{\frac{d-2}{4}}),
}
which we obtain thanks to {\eqref{eq:bracketExpansion_Wbis} and the fact that}
there is a constant $C=C(\b)$ such that for $s\in [\ell(T),Tt]$,
\[
\mathbb E\left[\langle \overline{M}(\sqrt T x) \rangle_{s}' \right]\leq C'\mathbb E\left[\langle \mathcal{W}(\sqrt T x) \rangle_{s}' \right].
\]
}\fi
\end{proof}
\begin{comment}
\begin{proof}
We define
\al{
\Phi_{s,t}=\exp{\left(\beta\int_s^t \int_{{\mathbb R}^2} \phi(z- B_r) \ \xi({\rm d} r,{\rm d} z)-\frac{\b^2 (t-s) V(0)}{2}\right)}.
}
Then, we introduce a polymer measure by
$$\dd \mathtt{P}_{x}^{\b,s}=\frac{1}{{\bf{\mathcal{Z}}}_{s}}\rho_{s}(0,x) {\mathrm E}_{s,x}[\Phi_{s,t}] \dd x.$$
Since $\dd \mathtt{P}_{x}^{\b,s}$ is a probability measure, by Jensen's inequality,
\al{
\left(\frac{{\bf{\mathcal{Z}}}^\b_{s}}{{\bf{\mathcal{Z}}}^\b_t}\right)^k
&= \left(\int_{{\mathbb{R}}^2}\dd \mathtt{P}_{x}^{\b,s} {\mathrm E}_{s,x}[\Phi_{s,t}]\right)^{-k}\\
&\leq \int_{{\mathbb{R}}^2}\dd \mathtt{P}_{x}^{\b,s} \left({\mathrm E}_{s,x}[\Phi_{s,t}]\right)^{-k}.
}
Thus,
\al{
\mathbb E\left[ \left(\sup_{sT\leq u\leq tT}\frac{{\bf{\mathcal{Z}}}^\b_{sT}}{{\bf{\mathcal{Z}}}_u^\b}\right)^k\right]&\leq \mathbb E\left[\int_{{\mathbb{R}}^2}\dd \mathtt{P}_{x}^{\b,sT}\sup_{sT\leq u\leq tT} \left({\mathrm E}_{s,x}[\Phi_{sT,u}]\right)^{-k}\right]\\
&=\mathbb E\left[\int_{{\mathbb{R}}^2}\dd \mathtt{P}_{x}^{\b,sT}\right] \mathbb E\left[\sup_{sT\leq u\leq tT}\left({\bf{\mathcal{Z}}}^\b_{u-sT}\right)^{-k}\right]\\
& = \mathbb E\left[\sup_{t\geq 0}\left({\bf{\mathcal{Z}}}^\b_{t}\right)^{-k}\right].
}
By Doob's submartingale inequality, we further see that $\displaystyle \sup_{T}\mathbb E\left[\left(\sup_{sT\leq u\leq tT}{\bf{\mathcal{Z}}}_{u}^{\b}\right)^{-k}\right]<\infty$ since $t\to \left({\bf{\mathcal{Z}}}^\b_{t}\right)^{-k}$ is a submartingale with \eqref{eq:negativemoment}.
\end{proof}
\end{comment}
\subsection{Proof of Proposition \ref{prop:VanishBracket}}
Our goal is to prove that:
\begin{equation} \label{eq:goalOfReduction}
\frac{1}{\b_\varepsilon} \int_{{\mathbb{R}}^2}\dd xf(x) \left( \int^{tT_\varepsilon}_0F''(\mathcal{W}_s(x_{T_\varepsilon})){\dd \langle \mathcal{W}(x_{T_\varepsilon}) \rangle_s}-\mathbb E \left[ \int^{tT_\varepsilon}_0F''(\mathcal{W}_s(x_{T_\varepsilon})){\dd \langle \mathcal{W}(x_{T_\varepsilon}) \rangle_s}\right]\right) \stackrel{L^1}{\longrightarrow} 0.
\end{equation}
For simplicity of notation, we set $t=1$ hereafter.
The proof is composed of four steps.
In the first step, we will investigate that the influence at large time is negligible in the following sense:
\iffalse{
\begin{lemma}[Step 1]\label{lem:firstStep}
Fix $t>0$. Then,
\aln{
\lim_{T\to\infty} \frac{1}{\b_\varepsilon} \mathbb E\left[\int_{{\mathbb{R}}^2} \dd xf(x) \int^{tT}_{ tTm(T)} {F''(\mathcal{W}_s(x_T)){\dd \langle \mathcal{W} (x_T)\rangle_s}}\right]=0.
}
\end{lemma}
\begin{proof}
It is enough to see that \begin{align*}
\lim_{T\to\infty} \frac{1}{\b_\varepsilon} \mathbb E\left[\int_{{\mathbb{R}}^2} \dd xf(x) \int^{tT}_{ Tm(T)} \left(\frac{1}{\mathcal{W}_{s}(x_T)^2}+1\right){\dd \langle \mathcal{W} (x_T)\rangle_s}\right]=0.
\end{align*}
Recall that
\[
A_s(x)=\{ {\bf{\mathcal{Z}}}_s(x) \leq L(T) \}.
\]
Then, by {\eqref{eq:denobdd}}
\al{
\frac{1}{\b_\varepsilon} \mathbb E\left[\int^{T}_{Tm(T)} \frac{\dd \langle \mathcal{W}(x_T) \rangle_s}{\mathcal{W}_s(x_T)^2} \right]&\leq \frac{C\b_\varepsilon}{L(T)^2}\int^{T}_{Tm(T)} {\mathrm E}_{x,x}^{\otimes 2}\left[V(B_s-\widetilde{B}_s)\Phi_s(B)\Phi_s(\widetilde{B})\right]\dd s \\
&+\|V\|_{\infty} \b_\varepsilon\int ^{T}_{Tm(T)} {\mathbb P}({ A_s(x_T)}^c)\dd s \\
&\leq \frac{C\b_\varepsilon}{L(T)^2}\int_{Tm(T)}^{T}{\mathrm E}_0\left[V(\sqrt{2}B_s)\exp\left(\b^2\int_0^s V(\sqrt{2}B_u)\dd u\right)\right]\dd s\\
&+\|V\|_\infty \b_\varepsilon T \sup_{s\leq T}{\mathbb P}({ A_s(0)}^c)\\
&\leq \frac{C\b_\varepsilon}{L(T)^2}\int_{Tm(T)}^{T}s^{-1}\dd s++\|V\|_\infty \b_\varepsilon tT \sup_{s\leq tT}{\mathbb P}({ A_s(0)}^c).
}
Indeed, we know \begin{align*}
&{\mathrm E}_0\left[V(\sqrt{2}B_s)\exp\left(\b^2\int_0^s V(\sqrt{2}B_u)\dd u\right)\right]\\
&=\int_{|x|\leq R_V}\dd xp_{s}(x){\mathrm E}_{0,0}^{s,x}\left[\exp\left(\b^2\int_0^s V(\sqrt{2}B_u)\dd u\right)V(\sqrt{2}B_s)\right]\\
&=\int_{|x|\leq R_V}\dd xp_s(x)V(\sqrt{2}x)\\
&+\sum_{n=1}^\infty \int_{|x|\leq R_V}\dd xp_s(x)\int_{0<s_1<\cdots<s_n<s}\int_{({\mathbb{R}}^2)^n}\dd \mathbf{s}\dd \mathbf{x}\\
&\hspace{3em}\left(\prod_{i=1}^n\left(V(\sqrt{2}x_i)p_{s_i-s_{i-1}}(x_i-x_{i-1})\right)V(\sqrt{2}x)\frac{p_{s-s_n}(x-x_n)}{p_{s}(x)}\right)
\end{align*}
and the argument in Lemma \ref{lem:L2bdd} can be applied to bound the last term.
\end{proof}
We introduce a new martingale \begin{align*}
\widetilde{\mathcal{W}}_s(x)={\mathrm E}_x\left[\Phi_s^\b(B)u_0\left(\frac{tT}{\sqrt{T}}\right):
\mathtt{F}_{tTm(T),\sqrt{Tn(T)}}(B,x)^c
\right]
\end{align*}
for $s\leq tTm(T)$.
\begin{lemma}(Step 2)
For any $x\in{\mathbb{R}}^2$, \begin{align*}
\lim_{T\to\infty}\frac{1}{\b_\varepsilon}\mathbb E\left[\left|\int_0^{tTm(T)}F''(\mathcal{W}_s(x)){\dd \langle \mathcal{W}(x)\rangle_s}-\int_0^{tTm(T)}F''(\widetilde{\mathcal{W}}_s(x)){\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s}\right|\right]=0.
\end{align*}
\end{lemma}
\begin{proof}
We have that \begin{align}
&\left|\int_0^{tTm(T)}F''(\mathcal{W}_s(x)){\dd \langle \mathcal{W}(x)\rangle_s}-\int_0^{tTm(T)}F''({\widetilde{\mathcal{W}}_s(x)}){\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s}\right|\notag\\
&\leq \left|\int_0^{tTm(T)} F''(\mathcal{W}_s(x))\dd \langle \mathcal{W}(x)\rangle_s-\int_0^{tTm(T)} F''(\mathcal{W}_s(x))\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right|\notag\\
&+\left|\int_0^{tTm(T)} \left(F''(\mathcal{W}_s(x))-F''(\widetilde{\mathcal{W}}_s(x))\right)\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right|\notag\\
&\leq C\b_\varepsilon^2\int_0^{tTm(T)}\dd s
\left(\frac{1}{{\bf{\mathcal{Z}}}_s^\b(x)^2}+1\right)\notag\\
&\hspace{2em}\times {{\mathrm E}_{x,x}^{\otimes 2}\left[V(B_s-B_s')\Phi_s^\b(B)\Phi_s^\b(B'):
\mathtt{F}_{tTm(T),\sqrt{Tn(T)}}(B,x)^c\cup
\mathtt{F}_{tTm(T),\sqrt{Tn(T)}}(B',x)^c
\right]
\\
&+\b_\varepsilon^2\int_0^{tTm(T)}\dd s
\left|F''(\mathcal{W}_s(x))-F''(\widetilde{\mathcal{W}}_s(x))\right|\notag\\%\left({\mathcal{W}_s(x)-\widetilde{\mathcal{W}}_s(x)}\right)
&\hspace{2em}\times {{\mathrm E}_{x,x}^{\otimes 2}\left[V(B_s-B_s')\Phi_s^\b(B)\Phi_s^\b(B'):
\mathtt{F}_{tTm(T),\sqrt{Tn(T)}}(B,x)\cap
\mathtt{F}_{tTm(T),\sqrt{Tn(T)}}(B',x).\label{eq:F''}
\right]}
\end{align}
Then, we have from Lemma \ref{lem:L2bdd} and Lemma \ref{lem:negativemoment} with H\"older's inequality that there exists a $p>2$ such that \begin{align*}
&\mathbb E\left[\int_0^{tTm(T)}\dd s\frac{1}{{\bf{\mathcal{Z}}}^\b_s(x)^2}{{\mathrm E}_{x,x}^{\otimes 2}\left[V(B_s-B_s')\Phi_s^\b(B)\Phi_s^\b(B'):
\mathtt{F}_{tTm(T),\sqrt{Tn(T)}}(B,x)^c\cup
\mathtt{F}_{tTm(T),\sqrt{Tn(T)}}(B',x)^c
\right]}\right]\\
&\leq \|V\|_\infty \int_0^{tTm(T)}\dd s\mathbb E\left[\frac{{{\mathrm E}_{x}\left[\Phi_s^\b(B):
\mathtt{F}_{tTm(T),\sqrt{Tn(T)}}(B,x)^c\right]}}{{\bf{\mathcal{Z}}}_s^\b(x)}\right]\\
&\leq \|V\|_\infty C \int_0^{tTm(T)}{\mathrm P}_x\left(\mathtt{F}_{tTm(T),\sqrt{Tn(T)}}(B,x)^c\right)^{\frac{1}{p}}.
\end{align*}
In a similar way, we have that \begin{align*}
&\int_0^{tTm(T)}\dd s \mathbb E\left[{{\mathrm E}_{x,x}^{\otimes 2}\left[V(B_s-B_s')\Phi_s^\b(B)\Phi_s^\b(B'):
\mathtt{F}_{tTm(T),\sqrt{Tn(T)}}(B,x)^c\cup
\mathtt{F}_{tTm(T),\sqrt{Tn(T)}}(B',x)^c
\right]}\right]\\
&\leq \|V\|_\infty \int_0^{tTm(T)}{\mathrm P}_x\left(\mathtt{F}_{tTm(T),\sqrt{Tn(T)}}(B,x)^c\right)^{\frac{1}{p}}.
\end{align*}
Applying H\"older's inequality, the last term in \eqref{eq:F''} is bounded from above by
\begin{align*}
&C\|V\|_\infty\b^2 \left(\int_0^{tTm(T)}\dd s\mathbb E\left[\left|F''(\mathcal{W}_s(x))-F''(\widetilde{\mathcal{W}}_s(x))\right|\right]\right)^{\frac{1}{p}}
\left(\int_0^{tTm(T)}\dd s\mathbb E\left[{\bf{\mathcal{Z}}}_s^\b(x)^{2q}\right]\right)^{\frac{1}{q}}\\
&\leq C\b^2\left(\int_0^{tTm(T)} \dd s\mathbb E\left[\left|F''(\mathcal{W}_s(x))-F''(\widetilde{\mathcal{W}}_s(x))\right|\right]\right)^{\frac{1}{p}}\\
&\leq C\b^2 \left(\int_0^{tTm(T)}\dd s\mathbb E\left[\left(\mathcal{W}_s(x)-\widetilde{\mathcal{W}}_s(x)\right)^2\right]\right)^{\frac{1}{p}}tTm(T)^{\frac{1}{q}}\\
&\leq C\b^2tTm(T){\mathrm P}_x\left(\mathtt{F}_{tTm(T),\sqrt{Tn(T)}}(B,x)^c\right)^{\frac{1}{pp'}}
\end{align*}
\end{proof}
Then, it is easy to see that \begin{align*}
\log \widetilde{\mathcal{W}}_s(x)=\int_0^s \frac{1}{\widetilde{\mathcal{W}}_u(x)}\dd \widetilde{\mathcal{W}}_u(x)-\frac{1}{2}\int_0^s \frac{1}{\widetilde{\mathcal{W}}_u(x)^2}\dd \langle \widetilde{\mathcal{W}}(x)\rangle_u.
\end{align*}
Thus, the proof of Proposition \ref{prop:VanishBracket} is completed when we show the following lemma.
\begin{lemma}
\begin{align*}
\lim_{T\to\infty}\frac{1}{\b_\varepsilon^2}\mathbb E\left[\left(\int_{{\mathbb{R}}^2}\dd xf(x)\left(\int_0^{{tTm(T)}}\frac{\dd \langle \widetilde{\mathcal{W}}(x_T)\rangle_u}{\widetilde{\mathcal{W}}_u(x_T)^2}-\mathbb E\left[\int_0^{{tTm(T)}}\frac{\dd \langle \widetilde{\mathcal{W}}(x_T)\rangle_u}{\widetilde{\mathcal{W}}_u(x_T)^2}\right)\right] \right)^2\right]=0.
\end{align*}
\end{lemma}
\begin{proof}
We remark that when we apply It\^{o}'s lemma to $\log \widetilde{\mathcal{W}}_s(x)$, we have \begin{align*}
\log \widetilde{\mathcal{W}}_s(x)&=\log \widetilde{u}(t,x)+\int_0^s \frac{\dd \widetilde{\mathcal{W}}_u(x)}{\widetilde{\mathcal{W}}_u(x)}-\frac{1}{2}\int_0^s\frac{\dd \langle \widetilde{\mathcal{W}}(x)\rangle_u}{\widetilde{\mathcal{W}}_u(x)^2}\\
&:=\log\widetilde{u}(t,x)+\widetilde{G}_s(x)-\frac{1}{2}\widetilde{H}_s(x),
\end{align*}
where $\widetilde{u}(t,x)=\displaystyle {\mathrm E}_x\left[u_0\left(\frac{B_{tT}}{\sqrt{T}}\right): {\color{magenta} F(B)}\right]$
and \begin{align*}
\left\langle \widetilde{G}(x)\right\rangle_s=\widetilde{H}_s(x).
\end{align*}
In particular, we have\begin{align*}
\mathbb E\left[\widetilde{H}_s(x)^2\right]&\leq 12\left(\log\widetilde{u}(t,x)\right)^2+12\mathbb E\left[\widetilde{H}_s(x)\right]+12\mathbb E\left[\left(\log \widetilde{\mathcal{W}}_s(x)\right)^2\right]\\
&=12\left(\log\widetilde{u}(t,x)\right)^2+12 \mathbb E\left[(\log \widetilde{\mathcal{W}}_s)^2\right]+24\mathbb E\left[\log \widetilde{u}(t,x)-\log \widetilde{\mathcal{W}}_s(x)\right]\\
&\leq C
\end{align*}
for a some constant $C>0$.
Also, we remark that for $|x-y|\geq 3\sqrt{Tn(T)}$\begin{align*}
\textrm{Cov}\left(\widetilde{H}_{tTm(T)}(x),\widetilde{H}_{tTm(T)}(y)\right)=0
\end{align*}
so that \begin{align*}
&\mathbb E\left[\left(\int_{{\mathbb{R}}^2}\dd xf(x)\left([H_{tTm(T)}(x_T)-\mathbb E\left[H_{tTm(T)}(x_T)\right]\right) \right)^2\right]\\
&=\int_{|x-y|\leq 3\sqrt{n(T)}}\dd x\dd yf(x)f(y)\textrm{Cov}\left(H_{tTm(T)}(x_T),H_{tTm(T)}(y_T)\right)\\
&\leq \int_{|x-y|\leq 3\sqrt{n(T)}}\dd x\dd y|f(x)f(y)|\textrm{Var}\left(H_{tTm(T)}(x_T)\right)^\frac{1}{2}\textrm{Var}\left(H_{tTm(T)}(y_T)\right)^\frac{1}{2}\\
&\leq Cn(T).
\end{align*}
\end{proof}
}\fi
\begin{lemma}[Step 1]\label{lem:firstStep}
\aln{
\lim_{\varepsilon\to 0} \frac{1}{\b_\varepsilon} \mathbb E\left[\int_{{\mathbb{R}}^2} \dd xf(x) \int^{T_\varepsilon}_{T_\varepsilon m(T_\varepsilon)} F''(\mathcal{W}_s(x_{T_\varepsilon})) \dd \langle \mathcal{W} (x_{T_\varepsilon})\rangle_s\right]=0.
}
\end{lemma}
Before going to the step 2, we introduce a stopping time\begin{align*}
\tau_{T_\varepsilon}=\tau_{T_\varepsilon}(x):=\inf\left\{s\geq 0:W_s(x)+\widetilde{\mathcal{W}}_s(x)^{-1}>\frac{1}{m(T_\varepsilon)}\right\}\land T_\varepsilon m(T_\varepsilon).
\end{align*}
Let us define the event:
$$A_{T_\varepsilon}(x)=\{\tau_{T_\varepsilon}(x)= T_\varepsilon m(T_\varepsilon)\}=\left\{\mathcal{W}_s(x)+\widetilde{\mathcal{W}}_s(x)^{-1} \leq m(T_\varepsilon)^{-1}\text{ for all }s\leq T_\varepsilon m(T_\varepsilon), \right\}.$$
In the step 2, we will find that the contribution of ``large" $\mathcal{W}(x)$ and ``small" $\widetilde{\mathcal{W}}(x)$ can be negligible.
\begin{lemma}[Step 2]\label{lem:step2}
\begin{align*}
\lim_{\varepsilon\to 0}\frac{1}{\b_\varepsilon}\mathbb E\left[\int_{\tau_{T_{\varepsilon}}(x)}^{{ T_\varepsilon m(T_\varepsilon})} |F''(\mathcal{W}_s(x))| \dd \langle \mathcal{W}(x)\rangle_s\right]=0.
\end{align*}
\end{lemma}
In the step 3, we will prove that the contributions by $\mathcal{W}(x)$ and $\widetilde{\mathcal{W}}(x)$ are asymptotically identified.
\begin{lemma}[Step 3]\label{lem:step2.5}
For any $x\in{\mathbb{R}}^2$, \begin{align*}
\lim_{\varepsilon\to 0}\frac{1}{\b_\varepsilon}\mathbb E\left[\left|\int_0^{\tau_{T_\varepsilon}(x)} F''(\mathcal{W}_s(x)){\dd \langle \mathcal{W}(x)\rangle_s}-\int_0^{\tau_{T_\varepsilon}(x)} F''(\widetilde{\mathcal{W}}_s(x)){\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s}\right| \right]=0.
\end{align*}
\end{lemma}
In the last, we will prove the remainder is also negligible.
\begin{lemma}[Step 4]\label{lem:step3}
\begin{align*}
\lim_{\varepsilon\to 0}\frac{1}{\b_\varepsilon^2}\mathbb E\left[\left(\int_{{\mathbb{R}}^2}\dd x f(x)\left(\int_0^{{{\tau_{T_\varepsilon}}}} F''(\widetilde{\mathcal{W}}_u(x_{T_\varepsilon})) \dd \langle \widetilde{\mathcal{W}}(x_{T_\varepsilon})\rangle_u-\mathbb E\left[\int_0^{\tau_{T_\varepsilon}} F''(\widetilde{\mathcal{W}}_u(x_{T_\varepsilon})) \dd \langle \widetilde{\mathcal{W}}(x_{T_\varepsilon})\rangle_u\right)\right] \right)^2\right]=0.
\end{align*}
\end{lemma}
Putting these lemmas together, Proposition \ref{prop:VanishBracket} is concluded.
\begin{proof}[Proof of Lemma \ref{lem:firstStep}]
With $\delta $ from \eqref{def: ell(T)}, for $C>(\|u_0\|_\infty+\|u_0^{-1}\|_\infty)^4$,
\begin{align*}
&\mathbb E\left[\int^{T}_{Tm(T)} \left|F''(\mathcal{W}_s(x_{T_\varepsilon}))\right| \dd \langle \mathcal{W} (x_T)\rangle_s \right]\\
&\leq C \mathbb E\left[\sup_{{ Tm(T)}\leq s\leq T}\left(1+{\bf{\mathcal{Z}}}_s(x)^{-2}\right)\int^{T}_{{ Tm(T)}} \dd \langle {\bf{\mathcal{Z}}} (x_T)\rangle_s \right] \\
&\leq 2C\mathbb E\left[\sup_{{ Tm(T)}\leq s\leq T}{\bf{\mathcal{Z}}}_s(x)^{-2}\mathbf{1}\left\{\sup_{{ Tm(T)}\leq s\leq T}{\bf{\mathcal{Z}}}_s(x)^{-1}> (\log T)^{\delta/4}\right\}\int^{T}_{{ Tm(T)}} \dd \langle {\bf{\mathcal{Z}}} (x_T)\rangle_s \right]\\
&\qquad\qquad+2C(\log{T})^{\delta/2}\mathbb E\left[\mathbf{1}\left\{\sup_{{ Tm(T)}\leq s\leq T}{\bf{\mathcal{Z}}}_s(x)^{-1}\leq (\log T)^{\delta/4}\right\} \int^{T}_{{ Tm(T)}}\dd \langle {\bf{\mathcal{Z}}} (x_T)\rangle_s \right].
\end{align*}
By the Burkholder-Davis-Gundy inequality, Doob's inequality and H\"older's inequality,
the first expectation is bounded from above by
\begin{align*}
&{\mathbb P}\left(\sup_{{ Tm(T)}\leq s\leq T}{\bf{\mathcal{Z}}}_s(x)^{-1}> (\log T)^{\delta/4}\right)^{\frac{1}{2p}} \mathbb E\left[\frac{1}{{\bf{\mathcal{Z}}}_{T}(x)^{4p}}\right]^{\frac{1}{2p}}\mathbb E\left[\left(\int^{T}_{{ Tm(T)}} \dd \langle {\bf{\mathcal{Z}}} (x_T)\rangle_s\right)^{q}\right]^{\frac{1}{q}}\\
&{ \leq \frac{1}{(\log T)^{2}}\mathbb E\left[{\bf{\mathcal{Z}}}_{T}(x)^{-\frac{16 p}{\delta}}\right]^{\frac{1}{2p}}\mathbb E\left[\frac{1}{{\bf{\mathcal{Z}}}_{T}(x)^{4p}}\right]^{\frac{1}{2p}}\mathbb E\left[{\bf{\mathcal{Z}}}_{T}(x)^{2q}\right]^{\frac{1}{q}}}\\
&\leq (\log{T})^{-2},
\end{align*}
where $p,q>1$ with $\frac{1}{p}+\frac{1}{q}=1$ and $\sup_{\varepsilon<1} \mathbb E\left[{\bf{\mathcal{Z}}}_{T}(x)^{2q}\right]<\infty$ from Lemma~\ref{lem:lp}. On the other hand, the second expectation can be bounded from above by
\begin{align*}
(\log{T})^{\delta/2}\mathbb E\left[\int^{T}_{{ Tm(T)}} \dd \langle {\bf{\mathcal{Z}}} (x_T)\rangle_s \right]
&=(\log T)^{\delta/2}\int_{Tm(T)}^{T}\beta^2 {\mathrm E}_{0}\left[V(\sqrt{2}B_s)\exp\left(\int_0^s\beta^2 V(\sqrt{2}B_u)\dd u\right)\right]\dd s\\
&=(\log T)^{\delta/2}\int_{Tm(T)}^{T}\beta^2 \int_{{\mathbb{R}}^2}\dd x V(\sqrt{2}x)\rho_s(x){\mathrm E}_{0}^x\left[\exp\left(\int_0^s\beta^2 V(\sqrt{2}B_u)\dd u\right)\right]\dd s\\
&\leq \b^2(\log{T})^{\delta/2} (\log (T)-\log ( { Tm(T)}))\\
&\leq -\b^2 (\log{T})^{\delta/2} \log m(T),
\end{align*}
where we have used Lemma \ref{lem:p2pge} in the third line and $\beta^2 (\log{T})^{\delta/2} \log m(T)\to 0$ as $\varepsilon\to 0$ as desired.
\iffalse{
By hypercontractivity, we have that
Recall that
\[
A_s(x)=\{ {\bf{\mathcal{Z}}}_s(x) \leq L(T) \}.
\]
Then, by {\eqref{eq:denobdd}}
\al{
\frac{1}{\b_\varepsilon} \mathbb E\left[\int^{T}_{Tm(T)} \frac{\dd \langle \mathcal{W}(x_T) \rangle_s}{\mathcal{W}_s(x_T)^2} \right]&\leq \frac{C\b_\varepsilon}{L(T)^2}\int^{T }_{Tm(T)} {\mathrm E}_{x,x}^{\otimes 2}\left[V(B_s-\widetilde{B}_s)\Phi_s(B)\Phi_s(\widetilde{B})\right]\dd s \\
&+\|V\|_{\infty} \b_\varepsilon\int ^{T}_{Tm(T)} {\mathbb P}({ A_s(x_T)}^c)\dd s \\
&\leq \frac{C\b_\varepsilon}{L(T)^2}\int_{Tm(T)}^{T}{\mathrm E}_0\left[V(\sqrt{2}B_s)\exp\left(\b^2\int_0^s V(\sqrt{2}B_u)\dd u\right)\right]\dd s\\
&+\|V\|_\infty \b_\varepsilon T \sup_{s\leq T}{\mathbb P}({ A_s(0)}^c)\\
&\leq \frac{C\b_\varepsilon}{L(T)^2}\int_{Tm(T)}^{t}s^{-1}\dd s++\|V\|_\infty \b_\varepsilon T \sup_{s\leq T}{\mathbb P}({ A_s(0)}^c).
}
Indeed, we know \begin{align*}
&{\mathrm E}_0\left[V(\sqrt{2}B_s)\exp\left(\b^2\int_0^s V(\sqrt{2}B_u)\dd u\right)\right]\\
&=\int_{|x|\leq R_V}\dd xp_{s}(x){\mathrm E}_{0,0}^{s,x}\left[\exp\left(\b^2\int_0^s V(\sqrt{2}B_u)\dd u\right)V(\sqrt{2}B_s)\right]\\
&=\int_{|x|\leq R_V}\dd xp_s(x)V(\sqrt{2}x)\\
&+\sum_{n=1}^\infty \b^{2n}\int_{|x|\leq R_V}\dd xp_s(x)\int_{0<s_1<\cdots<s_n<s}\int_{({\mathbb{R}}^2)^n}\dd \mathbf{s}\dd \mathbf{x}\\
&\hspace{3em}\left(\prod_{i=1}^n\left(V(\sqrt{2}x_i)p_{s_i-s_{i-1}}(x_i-x_{i-1})\right)V(\sqrt{2}x)\frac{p_{s-s_n}(x-x_n)}{p_{s}(x)}\right)
\end{align*}
and the argument in Lemma \ref{lem:L2bdd} can be applied to bound the last term.
}\fi
\end{proof}
\iffalse{
Next, we will approximate the main part contributing to $H$ by using $\widetilde{W}$.
\iffalse{\begin{lemma}
For any $x\in {\mathbb{R}}^2$,
\begin{align*}
\lim_{T\to\infty}\frac{1}{\b_\varepsilon}\mathbb E\left[\log {\mathcal{W}}_{s}(x)-\log\widetilde{\mathcal{W}}_s(x)\right]=0.
\end{align*}
\end{lemma}
\begin{proof}
It is easy to see that \begin{align*}
\log {\mathcal{W}}_{s}(x)-\log\widetilde{\mathcal{W}}_s(x)=\int_{\widetilde{\mathcal{W}}_s(x)}^{\mathcal{W}_s(x)}\frac{1}{t}\dd t\leq \frac{{\mathcal{W}}_s(x)-\widetilde{\mathcal{W}}_s(x)}{\widetilde{\mathcal{W}}_s(x)}
\end{align*}
and \begin{align*}
&\mathbb E\left[\frac{{\mathcal{W}}_s(x)-\widetilde{\mathcal{W}}_s(x)}{\widetilde{\mathcal{W}}_s(x)}\right]\\
&\leq \mathbb E\left[\left({{\mathcal{W}}_s(x)-\widetilde{\mathcal{W}}_s(x)}\right)^2\right]^{\frac{1}{2}}\mathbb E\left[{\widetilde{\mathcal{W}}_s(x)}^{-2}\right]^\frac{1}{2}\\
&\leq C{\mathrm E}_{x,x}^{\otimes 2}\left[\exp\left(\b^2\int_0^s V(B_u-B_u')\dd u\right): \{B[0,s]\not\subset \mathT{R}_{s,\sqrt{Tn(T)}}(x)\}\cup\{B'[0,s]\subset \mathT{R}_{s,\sqrt{Tn(T)}}(x)\}\right]^{\frac{1}{2}}\\
&\leq C{\mathrm E}_0\left[\exp\left(p\int_0^s V(\sqrt{2}B_u)\dd u\right)\right]^{\frac{1}{2p}}{\mathrm P}_0\left(\{B[0,s]\not\subset \mathT{R}_{s,\sqrt{Tn(T)}}(x)\right)^{\frac{1}{2q}}.
\end{align*}
\end{proof}
}\fi
\begin{lemma}\label{lem:step2} [Step 2]
For any $x\in{\mathbb{R}}^2$, \begin{align*}
\lim_{\varepsilon \to 0}\frac{1}{\b_\varepsilon}\mathbb E\left[\left|\int_0^{T_\varepsilon m(T_\varepsilon)}F''(\mathcal{W}_s(x))\dd \langle \mathcal{W}(x)\rangle_s -\int_0^{T_\varepsilon m(T_\varepsilon)} F''(\widetilde{\mathcal{W}}_s(x)) \dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right|\right]=0.
\end{align*}
\end{lemma}
\begin{proof}
We have that
\begin{align*}
&\left|\int_0^{{ Tm(T)}} F''(\mathcal{W}_s(x)) \dd \langle \mathcal{W}(x)\rangle_s-\int_0^{{ Tm(T)}}F''(\widetilde{\mathcal{W}}_s(x)) \dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right|\\
&\leq \left|\int_0^{{ Tm(T)}} F''(\mathcal{W}_s(x)) \dd \langle \mathcal{W}(x)\rangle_s-\int_0^{{ Tm(T)}} F''(\mathcal{W}_s(x)) \dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right|\\
&\qquad\qquad+\left|\int_0^{{ Tm(T)}}(F''(\mathcal{W}_s(x))-F''(\widetilde{\mathcal{W}}_s(x)))\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right|.
\end{align*}
For the first term,
\begin{align*}
&\quad \mathbb E\left|\int_0^{{ Tm(T)}} F''(\mathcal{W}_s(x)) \dd \langle \mathcal{W}(x)\rangle_s-\int_0^{{ Tm(T)}} F''(\mathcal{W}_s(x)) \dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right|\\
&\leq C\b^2\int_0^{{ Tm(T)}} \mathbb E\left[ (1+{\bf{\mathcal{Z}}}_s(x)^{-2}) \right.\\
&\qquad\qquad \times \left.{\mathrm E}_{0,0}^{\otimes 2}\left[V(B_s-B_s')\Phi_s^\b(B)\Phi_s^\b(B'): \mathtt{F}_{{ Tm(T)},\sqrt{Tn(T)}}(B,x)^c\cup \mathtt{F}_{{ Tm(T)},\sqrt{Tn(T)}}(B',x)^c\right]\right]\dd s\\
&\leq C\|V\|_\infty \b^2 \int_0^{{Tm(T)}} \mathbb E[(1+{\bf{\mathcal{Z}}}(x)^{-2})({\bf{\mathcal{Z}}}_s(x)^2-\widetilde{{\bf{\mathcal{Z}}}}_s(x)^2)]\dd s\\
&\leq 2C\|V\|_\infty \b^2 \int_0^{{Tm(T)}} \mathbb E[({\bf{\mathcal{Z}}}(x)+{\bf{\mathcal{Z}}}(x)^{-1})({\bf{\mathcal{Z}}}_s(x)-\widetilde{{\bf{\mathcal{Z}}}}_s(x))]\dd s\\
&\leq C'\|V\|_\infty\b^2\int_0^{{ Tm(T)}} \mathbb E[({\bf{\mathcal{Z}}}_s(x)-\widetilde{{\bf{\mathcal{Z}}}}_s(x))^2] \dd s,
\end{align*}
where we have used $({\bf{\mathcal{Z}}}_s(x)-\widetilde{{\bf{\mathcal{Z}}}}_s(x))^2\leq {\bf{\mathcal{Z}}}_s(x)({\bf{\mathcal{Z}}}_s(x)-\widetilde{{\bf{\mathcal{Z}}}}_s(x))$ in the second inequality and the Cauchy-Schwarz inequality and the fact $\mathbb E[(({\bf{\mathcal{Z}}}(x)+{\bf{\mathcal{Z}}}(x)^{-1}))^2] $ is uniformly bounded in the last line.
Let us define the event:
$$A^k_s(x)=\{\widetilde{\mathcal{W}}_s(x)+\widetilde{\mathcal{W}}_s(x)^{-1} \leq T^k, \}.$$
By ..., we have that for any $k\in{\mathbb{N}}$ and $T$ large enough, for any $s\leq T$,
$${\mathbb P}(^c A^k_s(x))\leq T^{-k}.$$
For the second term, using $$|F''(\mathcal{W}_s(x))-F''(\widetilde{\mathcal{W}}_s(x))|=\left|\int^{\mathcal{W}_s(x)}_{\widetilde{\mathcal{W}}_s(x)}F'''(r)\dd r\right|\leq (\mathcal{W}_s(x)-\widetilde{\mathcal{W}}_s(x))(1+\widetilde{\mathcal{W}}_s(x)^{-3}),$$ and $\langle \widetilde{\mathcal{W}}(x)\rangle'_s\leq \beta^2 |V|_\infty \widetilde{\mathcal{W}}(x)^2 $, if we take $k$ sufficiently large, then
\al{
&\qquad \mathbb E \left|\int_0^{{ Tm(T)}}(F''(\mathcal{W}_s(x))-F''(\widetilde{\mathcal{W}}_s(x)))\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right|\\
&\leq \int_0^{{ Tm(T)}}\mathbb E\left[\left|F''(\mathcal{W}_s(x))-F''(\widetilde{\mathcal{W}}_s(x))\right|\langle \widetilde{\mathcal{W}}(x)\rangle'_s ; A^k_s(x) \right] \dd s\\
&\qquad\qquad +\int_0^{Tm(T)}
\mathbb E\left[\left|F''(\mathcal{W}_s(x))-F''(\widetilde{\mathcal{W}}_s(x))\right| \langle \widetilde{\mathcal{W}}(x)\rangle'_s ;(A^k_s(x))^c \right] \dd s\\
&\leq C T^{3k} \int_0^{{ Tm(T)}} \mathbb E\left[\left|\mathcal{W}_s(x)-\widetilde{\mathcal{W}}_s(x)\right|\right] \dd s+C \int_0^{Tm(T)} \mathbb E\left[1+\widetilde{\mathcal{W}}_s(x)^2 ; (A^k_s(x))^c \right] \dd s\\
&\leq C T^{3k} \int_0^{{ Tm(T)}} \mathbb E\left[\left(\mathcal{W}_s(x)-\widetilde{\mathcal{W}}_s(x)\right)^2\right]^{1/2} \dd s +C \int_0^{Tm(T)} {\mathbb P}((A^k_s(x))^c)+\mathbb E\left[\mathcal{W}_s(x)^{2q}\right]^{1/q} {\mathbb P}(A^k_s(x)^c )^{\frac{1}{p}} \dd s\\
&\leq C T^{3k} \int_0^{{ Tm(T)}} \mathbb E\left[\left(\mathcal{W}_s(x)-\widetilde{\mathcal{W}}_s(x)\right)^2\right]^{1/2} \dd s+T^{-1}.
}
Thus, we obtain that \begin{align*}
&\mathbb E\left[\left|\int_0^{{ Tm(T)}}\frac{\dd \langle \mathcal{W}(x)\rangle_s}{\mathcal{W}_s(x)^2}-\int_0^{{ Tm(T)}}\frac{\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s}{\widetilde{\mathcal{W}}_s(x)^2}\right|\right]\\
&\leq C T^{3k} \int_0^{{ Tm(T)}}\dd s \mathbb E\left[\left(\mathcal{W}_s(x)-\widetilde{\mathcal{W}}_s(x)\right)^2\right]^{\frac{1}{2}}+T^{-1}\\
&\leq C T^{3k} \int_0^{{ Tm(T)}}\dd s{\mathrm E}_{x,x}^{\otimes 2}\left[\exp\left(\b^2\int_0^sV(B_u-B_u')\dd s\right): \mathtt{F}_{{ Tm(T)},\sqrt{Tn(T)}}(B,x)^c\right]^{\frac{1}{2}}+T^{-1}\\
&\leq C T^{3k} { Tm(T)}
{\mathrm E}_{x,x}^{\otimes 2}\left[\exp\left(p\b^2\int_0^sV(B_u-B_u')\dd s\right)\right]^{\frac{1}{p}}
{\mathrm P}_x\left(\mathtt{F}_{{ Tm(T)},\sqrt{Tn(T)}}(B,x)^c\right)^{\frac{1}{2q}}+T^{-1}\\
&\leq C T^{3k+2}{\mathrm P}_x\left(\mathtt{F}_{{ Tm(T)},\sqrt{Tn(T)}}(B,x)^c\right)^{\frac{1}{2q}}+T^{-1},
\end{align*}
where $p,q>1$ with $\frac{1}{p}+\frac{1}{q}=1$ is chosen as \eqref{eq:collisionLp} is satisfied. It is easy to see that there exists a universal constant $c>0$ such that
\begin{align*}
{\mathrm P}_x\left(\mathtt{F}_{{ Tm(T)},\sqrt{Tn(T)}}(B,x)^c\right)=\exp\left(-\frac{c n(T)}{2{ m(T)}}\right),
\end{align*}
which decays faster than any polynomial of $T$ and therefore the statement holds.
\end{proof}
Thus, the proof of Proposition \ref{prop:VanishBracket} is completed when we show the following lemma.
\begin{lemma}\label{lem:step3} (Step 3).
\begin{align*}
\lim_{\varepsilon\to 0}\frac{1}{\b_\varepsilon^2}\mathbb E\left[\left(\int_{{\mathbb{R}}^2}\dd xf(x)\left(\int_0^{{{ T_\varepsilon m(T_\varepsilon)}}}\frac{\dd \langle \widetilde{\mathcal{W}}(x_{T_\varepsilon})\rangle_u}{\widetilde{\mathcal{W}}_u(x_{T_\varepsilon})^2}-\mathbb E\left[\int_0^{{{ T_\varepsilon m(T_\varepsilon)}}}\frac{\dd \langle \widetilde{\mathcal{W}}(x_{T_\varepsilon})\rangle_u}{\widetilde{\mathcal{W}}_u(x_{T_\varepsilon})^2}\right)\right] \right)^2\right]=0.
\end{align*}
\end{lemma}
\begin{proof}
We remark that when we apply It\^{o}'s lemma to $\log \widetilde{\mathcal{W}}_s(x)$, we have \begin{align*}
\log \widetilde{\mathcal{W}}_s(x)&=\log \widetilde{u}(t,x)+\int_0^s \frac{\dd \widetilde{\mathcal{W}}_u(x)}{\widetilde{\mathcal{W}}_u(x)}-\frac{1}{2}\int_0^s\frac{\dd \langle \widetilde{\mathcal{W}}(x)\rangle_u}{\widetilde{\mathcal{W}}_u(x)^2}\\
&:=\log\widetilde{u}(t,x)+\widetilde{G}_s(x)-\frac{1}{2}\widetilde{H}_s(x),
\end{align*}
where $\widetilde{u}(t,x)=\displaystyle {\mathrm E}_x\left[u_0\left(\frac{B_{T}}{\sqrt{T}}\right):\mathtt{F}_{{ Tm(T)},\sqrt{Tn(T)}}(B,x)\right]$
and \begin{align*}
\left\langle \widetilde{G}(x)\right\rangle_s=\widetilde{H}_s(x).
\end{align*}
In particular, we have\begin{align*}
\mathbb E\left[\widetilde{H}_s(x)^2\right]&\leq 12\left(\log\widetilde{u}(t,x)\right)^2+12\mathbb E\left[\widetilde{H}_s(x)\right]+12\mathbb E\left[\left(\log \widetilde{\mathcal{W}}_s(x)\right)^2\right]\\
&=12\left(\log\widetilde{u}(t,x)\right)^2+24\mathbb E\left[\log \widetilde{u}(t,x)-\log \widetilde{\mathcal{W}}_s(x)\right]+12 \mathbb E\left[(\log \widetilde{\mathcal{W}}_s)^2\right]\\
&\leq C
\end{align*}
for a some constant $C>0$.
Also, we remark that for $|x-y|\geq 3\sqrt{Tn(T)}$\begin{align*}
\textrm{Cov}\left(\widetilde{H}_{{ Tm(T)}}(x),\widetilde{H}_{{ Tm(T)}}(y)\right)=0
\end{align*}
so that \begin{align*}
&\mathbb E\left[\left(\int_{{\mathbb{R}}^2}\dd xf(x)\left([\widetilde{H}_{{ Tm(T)}}(x_T)-\mathbb E\left[\widetilde{H}_{{ Tm(T)}}(x_T)\right]\right) \right)^2\right]\\
&=\int_{|x-y|\leq 3\sqrt{tn(T)}}\dd x\dd yf(x)f(y)\textrm{Cov}\left(\widetilde{H}_{{ Tm(T)}}(x_T),\widetilde{H}_{{ Tm(T)}}(y_T)\right)\\
&\leq \int_{|x-y|\leq 3\sqrt{tn(T)}}\dd x\dd y|f(x)f(y)|\textrm{Var}\left(\widetilde{H}_{{ Tm(T)}}(x_T)\right)^\frac{1}{2}\textrm{Var}\left(\widetilde{H}_{{ Tm(T)}}(y_T)\right)^\frac{1}{2}\\
&\leq Ctn(T).
\end{align*}
\end{proof}
\subsubsection{General case $F\in \mathfrak{F}$}
For $F\in \mathfrak{F}$, the proof of \eqref{eq:goalOfReduction} is a modification of the one for $F(x)=\log x$.
The proof of Lemma \ref{lem:firstStep} goes well for $F\in \mathfrak{F}$ without any change.
To modify the proof of (Step 2), it is enough to prove that \begin{align}
&\lim_{T\to\infty}\frac{1}{\b}\mathbb E\left[\int_{0}^{{ Tm(T)}}\left|F''(\mathcal{W}_s(x))-F''(\widetilde{\mathcal{W}}_s(x))\right|\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right]=0\label{eq:fW1}\\
&\lim_{T\to\infty}\frac{1}{\b}\mathbb E\left[\left|\int_{0}^{{ Tm(T)}}F''(\mathcal{W}_s(x))\dd \langle \mathcal{W}(x)\rangle_s-\int_0^{{ Tm(T)}}F''({\mathcal{W}}_s(x))\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right|\right]=0.\label{eq:fW2}
\end{align}
The same argument as Lemma \ref{lem:step2} shows \eqref{eq:fW2}.
For \eqref{eq:fW1}, we first remark that \begin{align*}
\frac{\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s}{\dd s}\leq C\b^2\|V\|_\infty \widetilde{\mathcal{W}}_s(x)^2,
\end{align*}
and hence
\begin{align*}
\int_{0}^{{ Tm(T)}}\left|F''(\mathcal{W}_s(x))-F''(\widetilde{\mathcal{W}}_s(x))\right|\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s
&\leq \int_0^{{ Tm(T)}}\left(\int_{\widetilde{\mathcal{W}}_s(x)}^{\mathcal{W}_s(x)}|F'''(y)|\dd y\right)\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\\
&\leq C\b^2\|V\|_\infty \int_0^{{ Tm(T)}}\frac{|\mathcal{W}_s(x)-\widetilde{\mathcal{W}}_s(x)|}{\widetilde{\mathcal{W}}_s(x)}\dd s.
\end{align*}
The remainder follows from the same argument in Lemma \ref{lem:step2}.
Before proceeding (Step 3), we will prepare (Step 2.5).
We introduce a stopping time\begin{align*}
\tau_{m}^{(T)}(x)=\tau_m:=\inf\left\{s\geq 0:\widetilde{\mathcal{W}}_s(x)>\frac{1}{m(T)}\right\}.
\end{align*}
\begin{lemma}\label{lem:step2.5} (Step 2.5)
\begin{align*}
\lim_{\varepsilon\to 0}\frac{1}{\b_\varepsilon}\mathbb E\left[\int_{{ T_\varepsilon m(T_\varepsilon)}\wedge \tau_{m}(x)}^{{ T_\varepsilon m(T_\varepsilon)}} F''(\widetilde{\mathcal{W}}_s(x))\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right]=0.
\end{align*}
\end{lemma}
\begin{proof}
We have by H\"older's inequality that
\begin{align*}
&\mathbb E\left[\int_{{Tm(T)}\wedge \tau_{m}(x)}^{{ Tm(T)}}\left(\frac{1}{\widetilde{\mathcal{W}}_s(x)^2}+1\right)\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right]\\
&\leq \mathbb E\left[\int_{0}^{{ Tm(T)}}\left(\frac{1}{\widetilde{\mathcal{W}}_s(x)^2}+1\right)\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s: \tau_{m}(x)\leq { Tm(T)}\right]\\
&= \mathbb E\left[\left(\widetilde{\mathcal{W}}_{{ Tm(T)}}(x)-\widetilde{\mathcal{W}}_{0}(x)\right)^2 +2\left(\log \widetilde{\mathcal{W}}_{{ Tm(T)}}(x)-\log \widetilde{\mathcal{W}}_0(x)\right):\tau_{m}(x)\leq { Tm(T)}\right]\\
&\leq C\left(\mathbb E\left[\left|\widetilde{\mathcal{W}}_{{ Tm(T)}}(x)-\mathcal{W}_{0}(x)\right|^{2p}\right]^{\frac{1}{p}}+\mathbb E\left[\left|\log \widetilde{\mathcal{W}}_{{ Tm(T)}}(x)-\log \widetilde{\mathcal{W}}_0(x)\right|^{p}\right]^{\frac{1}{p}}\right){\mathbb P}\left(\tau_{m}(x)\leq { Tm(T)}\right)^{\frac{1}{q}},
\end{align*}
where $p,q>1$ with $\frac{1}{p}+\frac{1}{q}=1$ are constant with $2p<p_{\hat{\b}}$. Doob's inequality implies that \begin{align*}
{\mathbb P}\left(\tau_{m}(x)\leq { Tm(T)}\right)={\mathbb P}\left(\sup_{s\leq { Tm(T)}}\widetilde{\mathcal{W}}_s(x)\geq \frac{1}{m(T)}\right)\leq {\widetilde{u}(t,x)}{m(T)}.
\end{align*}
\end{proof}
Finally, (Step 3) is easily modified as follows:
\begin{align*}
&\mathbb E\left[\left(\int_0^{{ Tm(T)}\wedge \tau_{m}(x_T)}F''(\mathcal{W}_s(x_T))\dd \langle \mathcal{W}(x_T)\rangle_s\right)^2\right]\\
&\leq C\mathbb E\left[\left(\int_0^{{ Tm(T)}\wedge \tau_{m}(x_T)}\frac{\dd \langle \mathcal{W}(x_T)\rangle_s}{\widetilde{\mathcal{W}}_s(x)^2}\right)^2\right]+C\mathbb E\left[\left(\int_0^{{ Tm(T)}\wedge \tau_{\widetilde{n}}(x_T)}\dd \langle \mathcal{W}(x_T)\rangle_s\right)^2\right]
\end{align*}
The first term is bounded due to the argument in Lemma \ref{lem:step3}. On the other hand,
by Burkholder-Davis-Gundy inequality, there exists $C>0$ such that
\begin{align*}
\mathbb E\left[\left(\int_0^{{ Tm(T)}\wedge \tau_{m}(x_T)}\dd \langle \widetilde{\mathcal{W}}(x_T)\rangle_s\right)^2\right]&\leq C\mathbb E\left[\left(\widetilde{\mathcal{W}}_{{ Tm(T)}\wedge \tau_{m}(x_T)}(x_T)-\widetilde{u}(t,x)\right)^4\right]\\
&\leq Cm(T)^{-4}.
\end{align*}
Thus, we have
\begin{align*}
\textrm{Var}\left(\int_0^{{ Tm(T)}\wedge \tau_{m}(x_T)}F''(\widetilde{\mathcal{W}}_s(x_T))\dd \langle \widetilde{\mathcal{W}}(x_T)\rangle_s\right)\leq Cm(T)^{-4}
\end{align*}
and therefore we have \begin{align*}
&\mathbb E\left[\left(\int_{{\mathbb{R}}^2}\dd xf(x)\left(\int_0^{{ Tm(T)}\wedge \tau_{\widetilde{n}}(x_T)}F''(\widetilde{\mathcal{W}}_s(x_T))\dd \langle \widetilde{\mathcal{W}}(x_T)\rangle_s-\mathbb E\left[\int_0^{{ Tm(T)}\wedge \tau_{\widetilde{n}}(x_T)}F''(\widetilde{\mathcal{W}}_s(x_T))\dd \langle \widetilde{\mathcal{W}}(x_T)\rangle_s\right]\right) \right)^2\right]\\
&\leq C\frac{tn(T)}{m(T)^4}.
\end{align*}
}\fi
Before the proof of Lemma \ref{lem:step2}, we give an estimate of the probability of $A_T^c(x)$.
\begin{lemma}
There exists a constant $C>0$ such that for $T_\varepsilon>0$ and { for $x\in{\mathbb{R}}^2$},
\al{
{\mathbb P}\left(A_{T_\varepsilon}(x)^c\right)\leq C m(T_\varepsilon).
}
\end{lemma}
\begin{proof}
We have
\al{
{\mathbb P}\left( A_T(x)^c\right)\leq \P(\mathcal{W}_s(x) > (2m(T))^{-1}\text{ for some } s\in [0, Tm(T)])+\P(\widetilde{\mathcal{W}}_s(x) < 2m(T)\text{ for some }s\in [0, Tm(T)]).
}
The first term is bounded from above by $2\|u_0\|_\infty m(T)$ using Doob's inequality and $\mathbb E \left[\mathcal{W}_{Tm(T)}(x)\right]\leq \|u_0\|_\infty$. Using the fact that $B<x$ implies $A<2x$ or $A-B>x$ for $A\geq B>0$ and $x>0$, by Doob's inequality with (sub-)martingales $\mathcal{W}_s(x)^{-1},\,\mathcal{W}_s(x)-\widetilde{\mathcal{W}}_s(x)$ the second term is bounded from above by
\al{
& \P(\mathcal{W}_s(x) < 4m(T)\text{ for some } s\in [0, Tm(T)])+\P(\mathcal{W}_s(x)-\widetilde{\mathcal{W}}_s(x) > 2m(T)\text{ for some } s\in [0, Tm(T)])\\
&\leq 4m(T) \mathbb E[\mathcal{W}_{Tm(T)}(x)^{-1}]+ m(T)^{-1} \mathbb E[\mathcal{W}_{Tm(T)}(x)-\widetilde{\mathcal{W}}_{Tm(T)}(x)]\leq Cm(T).
}
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:step2}]
By H\"older's inequality and Minkowski's inequality, the expectation is bounded from above by
\begin{align*}
&C \mathbb E\left[\int_{\tau_T}^{{ Tm( T)}}\left(\frac{1}{\mathcal{W}_s(x)^2}+1\right)\dd \langle \mathcal{W}(x)\rangle_s\right]\\
&\leq C \mathbb E\left[\int_{0}^{{ Tm(T)}}\left(\frac{1}{\mathcal{W}_s(x)^2}+1\right)\dd \langle \mathcal{W}(x)\rangle_s;~A_T(x)^c\right]\\
&\leq C' \mathbb E\left[\left(\int_{0}^{{ Tm(T)}}\left(\frac{1}{\mathcal{W}_s(x)^2}+1\right)\dd \langle \mathcal{W}(x)\rangle_s \right)^p\right]^{\frac{1}{p}} {\mathbb P}\left(A_T(x)^c\right)^{\frac{1}{q}}\\
&\leq C' \left(\mathbb E\left[\left(\int_{0}^{{ Tm(T)}}\frac{\dd \langle \mathcal{W}(x)\rangle_s}{\mathcal{W}_s(x)^2} \right)^p\right]^{\frac{1}{p}}+\mathbb E\left[\left(\int_{0}^{{ Tm(T)}}\dd \langle \mathcal{W}(x)\rangle_s \right)^p\right]^{\frac{1}{p}} \right){\mathbb P}\left(A_T(x)^c\right)^{\frac{1}{q}},
\end{align*}
where $p,q>1$ with $\frac{1}{p}+\frac{1}{q}=1$ are constant with $2p<p_{\hat{\b}}$.
Then, by applying the Burkholder-Davis-Gundy inequality to the martingales $\displaystyle \int_0^{s} \dd \mathcal{W}_u(x)=\mathcal{W}_s(x)-\bar{u}(1,x)$,
we obtain that
\begin{align*}
\mathbb E\left[\left(\int_0^{Tm(T)}\dd \langle \mathcal{W}(x)\rangle_s\right)^p\right]\leq C\mathbb E\left[\left(\mathcal{W}_{Tm(T)}(x)-\bar{u}(1,x)\right)^{2p}\right]\leq C.
\end{align*}
We remark that when we apply It\^{o}'s lemma to $\log {\mathcal{W}}_s(x)$, we have \begin{align*}
\log {\mathcal{W}}_s(x)&=\log \bar{u}(t,x)+\int_0^s \frac{\dd {\mathcal{W}}_u(x)}{{\mathcal{W}}_u(x)}-\frac{1}{2}\int_0^s\frac{\dd \langle {\mathcal{W}}(x)\rangle_u}{{\mathcal{W}}_u(x)^2}\\
&:=\log \bar{u}(t,x)+{G}'_s(x)-\frac{1}{2}{H}'_s(x),
\end{align*}
with \begin{align*}
\left\langle {G}'(x)\right\rangle_s={H}'_s(x).
\end{align*}
In particular, we have\begin{align*}
\mathbb E\left[{H}'_s(x)^2\right]&\leq 12\left(\log\bar{u}(t,x)\right)^2+12\mathbb E\left[{G}'_s(x)^2\right]+12\mathbb E\left[\left(\log {\mathcal{W}}_s(x)\right)^2\right]\\
&=12\left(\log{u}(t,x)\right)^2+24\mathbb E\left[H'(x)\right]+12 \mathbb E\left[(\log {\mathcal{W}}_s)^2\right]\\
&\leq C
\end{align*}
for some constant $C>0$.
Putting things together with \eqref{eq:exit}, we have\begin{align*}
\frac{1}{\b_\varepsilon}\mathbb E\left[\int_{\tau_T(x)}^{{ Tm(T)}} |F''(\mathcal{W}_s(x))| \dd \langle \mathcal{W}(x)\rangle_s\right]\leq \frac{C}{\b}{\mathbb P}\left(A_T(x)^c\right)^{\frac{1}{q}}\to 0.
\end{align*}
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:step2.5}]
Since $\mathcal{W}_s(x)+\widetilde{\mathcal{W}}_s(x)^{-1}\leq m(T)^{-1}$ for $s\leq \tau_T(x)$, we have that
\begin{align}
&\left|\int_0^{{\tau_T}} |F''(\mathcal{W}_s(x))| {\dd \langle \mathcal{W}(x)\rangle_s}-\int_0^{{\tau_T}}F''({\widetilde{\mathcal{W}}_s(x)}){\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s}\right|\notag\\
&\leq \left|\int_0^{{\tau_T}} F''(\mathcal{W}_s(x))\dd \langle \mathcal{W}(x)\rangle_s-\int_0^{{\tau_T}} F''(\mathcal{W}_s(x))\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right|\notag\\
&+\left|\int_0^{{\tau_T}} \left(F''(\mathcal{W}_s(x))-F''(\widetilde{\mathcal{W}}_s(x))\right)\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right|\notag\\
&\leq Cm(T)^2 \int_0^{{\tau_T}}\dd s
{\mathrm E}_{x}{\otimes }{\mathrm E}_x\left[V(B_s-\widetilde{B}_s)\Phi_s^\b(B)\Phi_s^\b(\widetilde{B}): \mathtt{F}_{Tm(T),\sqrt{Tn(T)}}(B,x)^c\cup \mathtt{F}_{Tm(T),\sqrt{Tn(T)}}(\widetilde{B},x)^c
\right
\notag\\
&+\b_\varepsilon^2\int_0^{{\tau_T}}\dd s
\left|F''(\mathcal{W}_s(x))-F''(\widetilde{\mathcal{W}}_s(x))\right|
{\mathrm E}_{x}{\otimes }{\mathrm E}_x\left[V(B_s-\widetilde{B}_s)\Phi_s^\b(B)\Phi_s^\b(\widetilde{B})\right]. \nota
\end{align}
Using H\"older's inequality, there exists a $p>2$ such that the first term is bounded from above by
\begin{align*}
&\mathbb E\left[\int_0^{Tm(T)}\dd s {{\mathrm E}_{x}{\otimes }{\mathrm E}_x\left[V(B_s-\widetilde{B}_s)\Phi_s^\b(B)\Phi_s^\b(\widetilde{B}):
\mathtt{F}_{{Tm(T)},\sqrt{Tn(T)}}(B,x)^c\cup
\mathtt{F}_{{Tm(T)},\sqrt{Tn(T)}}(\widetilde{B},x)^c
\right]}\right]\\
&\leq \|V\|_\infty \int_0^{{Tm(T)}}\dd s\mathbb E\left[{{\mathrm E}_{x}\left[\Phi_s^\b(B)\Phi_s^\b(\widetilde{B}):
\mathtt{F}_{{Tm(T)},\sqrt{Tn(T)}}(B,x)^c\right]}\right]\\
&\leq \|V\|_\infty C \int_0^{{Tm(T)}} \dd s {\mathrm P}_x\left(\mathtt{F}_{{Tm(T)},\sqrt{Tn(T)}}(B,x)^c\right)^{\frac{1}{p}}.
\end{align*}
For the second term, we first note that for each $s\leq \tau_T$,
$$|F''(\mathcal{W}_s(x))-F''(\widetilde{\mathcal{W}}_s(x))|=\left|\int^{\mathcal{W}_s(x)}_{\widetilde{\mathcal{W}}_s(x)}F'''(r)\dd r\right|\leq C(1+m(T)^{-3}) (\mathcal{W}_s(x)-\widetilde{\mathcal{W}}_s(x)),$$ and $\displaystyle \frac{\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s}{ \dd s} \leq \beta^2 \|V\|_\infty \widetilde{\mathcal{W}}_s(x)^2 \leq \|V\|_{\infty} m(T)^{-2}$. Hence,
\al{
&\qquad \mathbb E\left[ \left|\int_0^{{ \tau_T}}|F''(\mathcal{W}_s(x))-F''(\widetilde{\mathcal{W}}_s(x))|\dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\right|\right]\\
&\leq C\|V\|_\infty (1+m(T)^{-3}) m(T)^{-2}\int_0^{{Tm(T)}} \mathbb E\left[\left|\mathcal{W}_s(x)-\widetilde{\mathcal{W}}_s(x)\right|\right] \dd s\\
& = C\|V\|_\infty (1+m(T)^{-3}) m(T)^{-2}\int_0^{{ Tm(T)}} {\mathrm P}_x\left(\mathtt{F}_{{Tm(T)},\sqrt{Tn(T)}}(B,x)^c\right) \dd s.
}
By \eqref{eq:exit}, the statement holds.
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:step3}]
We define
\al{
\widetilde{H}_s(x)=\int_0^{s} F''(\widetilde{\mathcal{W}}_s(x)) \dd \langle \widetilde{\mathcal{W}}(x)\rangle_s.
}
We remark that for $|x-y|\geq 3\sqrt{Tn(T)}$\begin{align*}
\textrm{Cov}\left(\widetilde{H}_{{ Tm(T)}}(x),\widetilde{H}_{{ Tm(T)}}(y)\right)=0
\end{align*}
so that \begin{align*}
&\mathbb E\left[\left(\int_{{\mathbb{R}}^2}\dd xf(x)\left([\widetilde{H}_{{ \tau_T}}(x_T)-\mathbb E\left[\widetilde{H}_{{ \tau_T}}(x_T)\right]\right) \right)^2\right]\\
&=\int_{|x-y|\leq 3\sqrt{n(T)}}\dd x\dd yf(x)f(y)\textrm{Cov}\left(\widetilde{H}_{{ \tau_T}}(x_T),\widetilde{H}_{{ \tau_T}}(y_T)\right)\\
&\leq \int_{|x-y|\leq 3\sqrt{n(T)}}\dd x\dd y|f(x)f(y)|\,\mathbb E\left[\widetilde{H}_{{ \tau_T}}(x_T)^2\right]^{\frac{1}{2}}\mathbb E\left[\widetilde{H}_{{ \tau_T}}(y_T)^2\right]^{\frac{1}{2}}.
\end{align*}
Since
\al{
|\widetilde{H}_{\tau_T}(x)|\leq C \int^{\tau_T}_0 (1+\widetilde{\mathcal{W}}_{s}(x)^{-1})^2 \dd \langle \widetilde{\mathcal{W}}(x)\rangle_s\leq C(1+m(T)^{-2})\int^{\tau_T}_0 \dd \langle \widetilde{\mathcal{W}}(x)\rangle_{s},
}
by the Burkholder-Davis-Gundy inequality, we have
\al{
\mathbb E \left[\widetilde{H}_{\tau_T}(x)^2\right]&\leq C (1+m(T)^{-2})^2 \mathbb E\left[\sup_{0\leq s\leq \tau_T}\left(\mathcal{W}_{s}(x)-\bar{u}(1,x)\right)^4\right]\leq C (1+m(T)^{-2})^2 m(T)^{-4}.
}
Putting things together
we have \begin{align*}
&\frac{1}{\b^2}\mathbb E\left[\left(\int_{{\mathbb{R}}^2}\dd xf(x)\left(\int_0^{\tau_T} F''(\widetilde{\mathcal{W}}_s(x_T))\dd \langle \widetilde{\mathcal{W}}(x_T)\rangle_s-\mathbb E\left[\int_0^{\tau_{T}} F''(\widetilde{\mathcal{W}}_s(x_T))\dd \langle \widetilde{\mathcal{W}}(x_T)\rangle_s\right]\right) \right)^2\right]\\
&\leq C\frac{n(T)}{\b^2m(T)^8}.
\end{align*}
\end{proof}
\subsection{Multidimensional convergence in the EW limits}
\label{sec:KPZmultidim}
To ease the presentation, we restrict ourselves to the case where $F(x)=x$, and $\hat{\b}\in (0,1)$ is fixed, although a repetition of the argument would lead to the result for the general initial conditions and the function $F$ that we have been considering.
Also, we note that for all $0\leq t_1\leq \cdots\leq t_n= t$, $u_0^{(1)},\cdots,u_0^{(n)}\in C_b({\mathbb{R}}^2)$, and $f_1,\cdots,f_n\in C_c^\infty({\mathbb{R}}^2)$\begin{align*}
&(u_\varepsilon(t_1,u_0^{(1)},f_1),\cdots, u_\varepsilon(t_n,u_0^{(n)},f_n))\\
&\stackrel{(d)}{=} \left(\mathcal{W}^{\left(t,T,u_0^{(1)}\right)}_{T(t-t_1)}(Tt,f_1),\cdots,\mathcal{W}^{\left(t,T,u_0^{(n-1)}\right)}_{T(t-t_{n-1})}(Tt,f_{n-1}),\mathcal{W}^{\left(t,T,u_0^{(n)}\right)}_{0}(Tt,f_n)\right),
\end{align*}
where we define for fixed $t>0$ that for $ u, s\geq 0$ and $x\in{\mathbb{R}}^2$ \begin{align*}
&\mathcal{W}_{u}(s,x)=\mathcal{W}_{u}^{(t,T,u_0)}(s,x)=\begin{cases}
\displaystyle {\mathrm E}_x\left[\Phi_{u,s}(B)u_0\left(\frac{B_{Tt-u}}{\sqrt{T}}\right)\right],\quad &0\leq u\leq s\\
u_0(x),&0\leq s\leq u.
\end{cases}
\intertext{and}
&\mathcal{W}_{u}(s,f)=\mathcal{W}_{u}^{(t,T,u_0)}(s,f)=\int_{{\mathbb{R}}^2}f(x)\mathcal{W}_{u}^{(t,T,u_0)}(s,x_T)\dd x.
\end{align*}
Thus, it suffices to show is that jointly for finitely many $u\in[0,t]$, $u_0\in C_b({\mathbb{R}}^2)$, and $f\in\mathcal C^\infty_c$, as $\varepsilon\to 0$,
\begin{equation}
\frac{1}{\b_\varepsilon} \int f(x) \left( \mathcal{W}_{T u,T t}^{\left(t,T,u_0\right)}(x_T)- \bar{u}(t-u,x) \right)\dd x
\stackrel{(d)}{\longrightarrow} \mathscr{U}_{u}^{(t,u_0)}(t,f)
\label{desired convergence}
\end{equation}
where $\displaystyle \left\{\mathscr{U}_{u}^{(t,u_0)}(s,f):f\in C_c^\infty({\mathbb{R}}^2),u_0\in C_b({\mathbb{R}}), 0\leq u\leq s\leq t\right\}$ is centered Gaussian field with covariance \begin{align*}
&\mathrm{Cov}\left(\mathscr{U}_{u}^{(t,u_0)}(s,f),\mathscr{U}_{u'}^{(t,u_0')}(s,f')\right)\\
&=\frac{1}{1-\hat{\b}^2}\int_{u\vee u'}^s \dd \sigma \int \dd x\dd y f(x)f'(y)\int \dd z\rho_{\sigma-u}(x,z)\rho_{\sigma-u'}(y,z) \bar{u}(t-\sigma,z)\bar{u}'(t-\sigma,z).
\end{align*}
Following the same strategy as in Subsection \ref{IdeaG}, we are reduced to showing that
\begin{equation}\label{eq:finalGoalKPZmuli}
\frac{1}{\b_\varepsilon} \mathcal{M}^{\ssup {t,T,u_0}}_{u}(\tau,f)\stackrel{(d)}{\longrightarrow}
\mathscr{U}_{u}^{(t,u_0)}(\tau,f)
\quad \text{jointly in } u\in[0,\tau], f\in\mathcal C^\infty_c,
\end{equation}
where (see \eqref{eq:defdMbartau})
\begin{align*}
&\mathcal{M}^{\ssup {t,T,u_0}}_{u}(\tau,f) := \begin{cases}
\displaystyle \int_{{\mathbb{R}}^2} f(x) \int_{Tu+T(t-u)m(T)}^{T \tau} \dd \mathcal{M}_{u}^{\ssup{t,T,u_0}}(s,x_T)\dd x,\quad &\tau\geq Tu+T(t-u)m(T)\\
0,\quad &\tau \leq Tu+T(t-u)m(T)
\end{cases}
\intertext{and }
&\dd \mathcal{M}_{u}^{\ssup{t,T,u_0}}(s,x)
\\
&:
\b_\varepsilon {\bf{\mathcal{Z}}}_{Tu,Tu+(s-Tu)\ell(T)}(x)\int \xi(\dd s,\dd b)\int \rho_{s-Tu}(x,z)\phi(z-b)\overleftarrow {{\bf{\mathcal{Z}}}}_{s,(s-Tu)\ell(T)}(z){\mathrm E}_z\left[u_0\left(\frac{B_{Tt-s}}{\sqrt{T}}\right)\right]\dd z.
\end{align*}
Then, for all $u\geq 0$ and $f\in\mathcal C^\infty_c$, $\tau \to \mathcal{M}^{(t,T)}_{u}(\tau,f)$ is a continuous martingale. In view of the desired convergence \eqref{desired convergence}, we have again in mind the functional CLT for martingales Theorem \ref{thm:JS}, so we are interested in the limit of the cross-bracket $\langle \mathcal{M}^{(t,T)}_{u_1}(\cdot,f_{{ 1}}),\mathcal{M}_{u_2}^{\ssup {t,T}}(\cdot,f_{{ 2}})\rangle_\tau$.
We have:
\begin{proposition}\label{lm:CVLone}
For all test functions $f$ and $f'$ in $\mathcal{C}_c^\infty$, $u_0,u_0'\in C_b({\mathbb{R}}^2)$, and $0\leq u_2\leq u_1\leq t$, for all $\tau \geq u_1$,
\begin{align}
&
\frac{1}{\beta_\varepsilon^2} \langle \mathcal{M}^{(t,T,u_0)}_{u_1}(\cdot,f_{{ 1}}),\mathcal{M}_{u_2}^{\ssup {t,T,u'_0}}(\cdot,f_{{ 2}})\rangle_\tau \nonumber\\
&\overset{L^1}{\longrightarrow}\frac{1}{1-\hat{\b}^2}\int_{u_1}^\tau \dd \sigma \int \dd x\dd y f_1(x)f_2(y)\int \dd z\rho_{\sigma-u_1}(x,z)\rho_{\sigma-u_2}(y,z) \bar{u}(t-\sigma,z)\bar{u}'(t-\sigma,z), \label{eq:LimcovStruc}
\end{align}
as $\varepsilon\to 0$.
\end{proposition}
\begin{proof}
For all $\tau \geq u_1+(t-u_1)m(T)$,
\begin{align*}
&\frac{1}{\b^2}\langle \mathcal{M}_{u_1}^{(t,T,u_0)}(\cdot,f_{{ 1}}),\mathcal{M}_{u_2}^{\ssup {t,T,u_0'}}(\cdot,f_{{ 2}})\rangle_\tau\\
&=
\int_{\mathbb{R}^2\times \mathbb R^2} f_1(x)f_2(y) \dd x \dd y \int_{Tu_1+T(t-u_1)m(T)}^{T\tau}\dd s {\bf{\mathcal{Z}}}_{Tu_1,Tu_1+(s-Tu_1)\ell(T)}(x_T) {\bf{\mathcal{Z}}}_{Tu_2,Tu_2+(s-Tu_2)\ell(T)}(y_T)\\
&\hspace{3em}\times \int_{{\mathbb{R}}^2\times {\mathbb{R}}^2}\dd z_1\dd z_2\rho_{s-Tu_1}(z_1-x_T)\rho_{s-Tu_2}(z_2-y_T)V(z_1-z_2)\overleftarrow{{\bf{\mathcal{Z}}}}_{s,(s-Tu_1)\ell(T)}(z_1)\overleftarrow{{\bf{\mathcal{Z}}}}_{s,(s-Tu_2)\ell(T)}(z_2)\\
&\hspace{5em}\times {\mathrm E}_{z_1}\left[u_0\left(\frac{B_{Tt-s}}{\sqrt{T}}\right)\right]{\mathrm E}_{z_2}\left[u_0'\left(\frac{B_{Tt-s}}{\sqrt{T}}\right)\right]
\end{align*}
By a repetition of the arguments that lead to \eqref{eq:CVbracketMt},
we find that
\begin{align*}
&\frac{1}{\b^2}\langle \mathcal{M}_{u_1}^{(t,T,u_0)}(\cdot,f_{{ 1}}),\mathcal{M}_{u_2}^{\ssup {t,T,u_0'}}(\cdot,f_{{ 2}})\rangle_\tau\\
&\approx_{L^1}\int f_1(x)f_2(y) \dd x \dd y \int_{u_1+(t-u_1)m(T)}^{\tau} \dd \sigma\mathbb E\left[{\bf{\mathcal{Z}}}_{Tu,Tu+T(\sigma-u)\ell(T)}(x_T)\right] \mathbb E\left[ {\bf{\mathcal{Z}}}_{Tu_2,Tu_2+(s-Tu_2)\ell(T)}(y_T)\right] \Theta_T(x,y),
\end{align*}
where
\begin{align*}
\Theta_T(x,y) & = \int \dd z\dd v\rho_{\sigma-u_1}(z-x)\rho_{\sigma -u_2}(z-\frac{v}{\sqrt{T}}-y)V(v)\mathbb E\left[\overleftarrow{{\bf{\mathcal{Z}}}}_{\sigma,(\sigma-u_1)\ell(T)}(z_T)\overleftarrow{{\bf{\mathcal{Z}}}}_{\sigma,(\sigma-u_2)\ell(T)}(z_T+v)\right]\\
&\hspace{5em}\times {\mathrm E}_{z_T}\left[u_0\left(\frac{B_{Tt-T\sigma}}{\sqrt{T}}\right)\right]{\mathrm E}_{z_T+v}\left[u'_0\left(\frac{B_{Tt-T\sigma}}{\sqrt{T}}\right)\right]\\%\\
& \to \frac{1}{1-\hat{\b}^2} \int \dd z\rho_{\sigma -u_1}(x-z)\rho_{\sigma-u_2}(y-z)\bar{u}(t-\sigma,z)\bar{u}'(t-\sigma,z).
\end{align*}
\end{proof}
\begin{comment}
\subsubsection{Concluding lines for the multidimensional case} \label{sec:conclusionOfProofOfSHE}
We show here the desired convergence \eqref{desired convergence}, which also writes
\begin{equation}\label{eq:finalGoal}
T\to\infty, \quad \overline{M}_t^{\ssup T}(u,f)\stackrel{(d)}{\longrightarrow}\int_{\mathbb{R}^d} f(x)\mathscr U_1(t-u,x)\dd x \quad \text{jointly in } u\in[0,t], f\in \mathcal C^\infty_c,
\end{equation}
We first observe that \eqref{eq:LimcovStruc} holds for all $0\leq u_2\leq u_1\leq t$, while if $0\leq \tau <u_1$, the cross-bracket $\langle \overline{M}^{\ssup T}_{\tau}(u_1,f_1), \overline{M}^{\ssup T}_{\tau}(u_2,f_2)\rangle_\tau $ is null and converges trivially to zero. Then, we define the quantity $C_\tau(u_1,u_2,f_1,f_2)$ to be equal to the RHS of \eqref{eq:LimcovStruc} when $\tau \geq u_1$ and set it to $0$ when $\tau < u_1$. We also define $C_\tau(u_1,u_2,f_1,f_2)$ when $u_1>u_2$ to be $C_\tau(u_2,u_1,f_1,f_2)$.
Therefore, by the multidimensional functional central limit for martingales (\cite[Theorem 3.11]{JS87}), we obtain that the family of continuous martingales $( \overline{M}^{\ssup T}_{\tau}(u,f))_{u}$ converges towards the family of Gaussian continuous martingales $(\tau\to(G_\tau(u,f))_{u}$, whose covariance structure is given by $C_\tau(u_1,u_2,f_1,f_2)$, where this convergence holds in terms of continuous processes in $\tau$ and jointly for finitely many $u\in[0,t]$, $f\in \mathcal C^\infty_c$.
Hence, \eqref{eq:finalGoal} follows from letting $\tau=t$ in this convergence and by observing that there is equality in law between the centered Gaussian families $G_t(u,f))_{u\in[0,t],f}$ and $(\int f(x)\mathscr U_1(t-u,x)\dd x)_{u\in[0,t],f}$ since they have the same covariance structure (recall that we set $u_0(\cdot)\equiv 1$).
\end{comment}
|
\section{Introduction}
\label{sec:intro}
Over the years, data acquisition has become more redundant, more complete, faster, and denser~\textendash~spatially and temporally.
Sensors such as cameras, LiDAR scanners, and RaDAR sensors guarantee multi-modal capturing of our world.
Depending on the application and the desired mapping scale, the respective sensors are mounted on platforms such as satellites, airplanes, \acp{UAV}, or autonomous vehicles.
In the domain of photogrammetry and remote sensing, particularly for urban mapping, data acquisition via imagery and \ac{ALS} is common.
Currently, data capture in urban areas at \acp{GSD} down to a few centimeters is becoming state of the art.
Traditionally, the airplane has been the platform of choice.
However, more flexible and lightweight \acp{UAV} have grown in popularity in the past decade \citep{Haala2020}.
\begin{figure}[htbp]
\centering
\begin{tabular}{lll}
\includegraphics{_figures/figure1_left.png}
&
\includegraphics{_figures/figure1_center.png}
&
\includegraphics{_figures/figure1_right.png}
\end{tabular}
\caption{
Visualization of the proposed multi-modal data fusion by means of the enabled label transfer. The figure depicts the transferred annotations to the mesh (\textit{center}) and an oblique image (\textit{right}) as transferred from the respective manually labeled point cloud (a subset of Hessigheim~3D).
Faces that cannot be linked to points are shown in textured fashion. Background or non-associated pixels respectively are colored in reddish. Pixels that are linked to an unlabeled face are depicted in black. The label scheme is given in~\autoref{fig:data_sets}.
}
\label{fig:juggler_alternative}
\end{figure}
Imagery is the fundamental photogrammetric data representation providing (multi-)spectral information.
Images project 3D real-world objects into 2D image space.
By nature of the projection into grid-like pixel space, images suffer from occlusions, distortions, discretization, and the loss of the third dimension.
However, with the help of automatic aerial triangulation, the intrinsic defects can be rectified and 3D reconstruction is possible.
As a precondition to proper reconstruction, images have to provide unambiguous texture and capture each object point at least twice.
Derived 3D products of the \ac{MVS} pipeline are colored \acp{PC} and/or textured meshes, both mapping the surface of the captured region.
In contrast, due to the polar measurement principle, 3D \acp{PC} are the immediate \ac{ALS} output.
In comparison to \ac{MVS}, LiDAR scanning provides multi-target capability, and hence penetrates semi-transparent objects such as vegetation.
Moreover, the polar measurement principle requires only a single measurement to map a 3D point.
On the other hand, bare LiDAR points do not carry color/texture information like \acp{PC}/meshes as derived from imagery.
The accuracy of \ac{ALS} points depends on the accuracy of the trajectory.
On the contrary, the accuracy of \ac{MVS} points is correlated with the \ac{GSD} which, theoretically, can be scaled arbitrarily.
For a detailed comparison of these two capturing methodologies, we refer to \cite{Mandlburger2017}.
Initially, photogrammetry and laser scanning have been competitive systems with individual processing pipelines.
However, at present, they are seen as complementary systems whose fusion results in more complete and better products.
Nowadays, joint acquisition of photogrammetric and \ac{ALS} data is state of the art for airborne systems and starts to emerge even for \ac{UAV}-based systems \citep{Mandlburger2017,cramer2018monitoring}.
Recently, \cite{glira2019} proposed the hybrid orientation of \ac{ALS} \acp{PC} and aerial imagery, which improves the georeferencing accuracy of \ac{ALS} data by integrating stabilizing image block geometry into the strip adjustment.
As a side product, the hybrid orientation enables a precise co-registration of imagery and LiDAR data.
Concerning the recent hybridization trend, from our point of view, enhancing 3D \acp{PC} to textured meshes may replace unstructured \acp{PC} as default representation for urban scenes in the future.
Intrinsically, meshes facilitate multi-modal data fusion by utilizing LiDAR points and \ac{MVS} points for the geometric reconstruction while leveraging high-resolution imagery for texturing (hybrid data storage).
Therefore, meshes are realistic-looking 3D maps of our real world and are easily understandable~\textendash~even for non-experts.
Besides benefits for visualization, textured meshes have other favorable characteristics.
Whereas \acp{PC} are unordered sets of points, meshes are graphs consisting of vertices, edges, and faces that provide explicit adjacency information.
Meshes are less memory-consuming than \acp{PC} since meshing algorithms try to minimize the number of entities while reconstructing the maximum of detail.
Before the meshing, \acp{PC} will be filtered in such a way that only geometrically relevant points are kept.
This embraces noise filtering and filtering of points that can be approximated by the same face (e.g. points on planar surfaces).
Furthermore, there will be geometric simplifications based on the desired level of detail and, as the case may be, due to 2.5D mesh geometry.
By definition, meshes are surface descriptions that cannot handle multi-target capability like LiDAR \acp{PC}.
This inevitably leads to a drop in entities to be stored.
Moreover, the high-resolution texture information is stored in texture atlases avoiding redundant image content.
Therefore, textured meshes provide geometric and textural information in a lightweight fashion.
Aside from these structural differences, georeferencing issues of imagery and LiDAR data will cause discrepancies between imagery, \acp{PC}, and meshes, too.
Being a hybrid data storage, we believe that the mesh modality is ideally suited to foster multi-modal semantic analysis.
To this end, we chose the mesh to be the core of the proposed multi-modal linking and transferring pipeline.
The aims and objectives of the paper are postulated in~\autoref{subsec:aims_and_objectives}.
In \autoref{sec:methodology}, we describe the entire methodology including the explicit entity linking and subsequent information transfer by deep-diving its building blocks.
Subsection~\ref{subsec:pcmeshassociation} describes the association of LiDAR data and the mesh in 3D space; \autoref{subsec:imgmeshassociation} outlines the association of imagery and the mesh.
Both association mechanisms operate face-centered.
Eventually, the combination of both links 3D points and pixels (cf.~\autoref{subsec:pcimgassociation}).
Subsection~\ref{subsec:preconditions_limitations} discusses in detail prerequisites and particular challenges due to the mentioned (structural) discrepancies between the mesh, the \ac{PC}, and imagery.
In~\autoref{sec:results_discussion}, we demonstrate the proper working of the presented association mechanism on two real-world data sets \citep{V2D,cramer2018monitoring}.
Since \ac{GT} is not available across all modalities, a quantitative analysis of the proposed method is difficult.
Therefore, we use the proposed label transfer to verify and showcase the proposed methodology.
Moreover, we report the best performing parameters for the association mechanism concerning the used imagery and \ac{ALS} data.
We briefly present the used data and the key parameters in~\autoref{sec:data}.
Both data sets provide significantly different resolution and co-registration quality.
\subsection{Aims and Objectives}
\label{subsec:aims_and_objectives}
Our key contribution is the explicit linking of pixels, points, and faces to jointly leverage information from available data sources aiming at multi-modal semantics (cf.~\autoref{sec:methodology}).
Each face will be linked to several points and several pixels.
In turn, points and pixels are linked while checking the visibility via the mesh.
To the best of our knowledge, there is no other holistic approach that explicitly joins imagery, mesh, and LiDAR data.
The explicitly established connections on the entity-level are used to share information across modalities.
Depending on the entity relationship, the information is aggregated prior to the transfer.
The aggregation of features is achieved by calculating the median; label aggregation is achieved by majority vote (cf.~\autoref{tab:association_options}).
In this study, we seem to focus on the label transfer since the effectiveness of the method can be shown better with labels than with features.
Furthermore, annotated \ac{GT} is of great importance for training supervised classifiers, particularly \ac{DL} approaches.
However, \ac{GT} generation is tedious, time-consuming, and expensive work wherefore real-world \ac{GT} availability is a rarity.
In particular, pixel-wise \ac{GT} generation is labor-intensive.
Generally, there is a lack of \ac{GT} data sets that jointly provide \acp{PC} and oriented imagery (and in this way textured meshes).
Therefore, available annotations are limited to a single representation and prevent exploiting the potential of multi-modal training.
This fact further motivates the necessity for the proposed approach and emphasizes its utility.
Our linking and transferring methodology facilitates the consistent labeling of various representations, given a manually annotated representation initially.
Hence, our method may help to overcome the imbalance of labeled entities among modalities.
For instance, 3D data (\ac{PC} or mesh) can be projected into various images at a stroke (cf.~\autoref{fig:results_imagery}).
Therefore, it minimizes the manual effort for \ac{GT} generation and helps considerably to foster modality-wise training of algorithms.
Our methodology is designed to process real-world data handling structural differences and co-registration issues.
To deal with the huge amount of redundant multi-modal data, the association operates in a tiled and parallelized fashion while aiming at a low memory footprint.
For the sake of good scientific practice, we critically reflect the preconditions and explore where the proposed approach might be limited (cf.~\autoref{subsec:preconditions_limitations}).
To summarize, the proposed mesh-centered multi-modal entity linking serves as the backbone to share features and labels across entities.
Representation-specific features and (manually generated) annotations can be shared at a stroke.
Thus, the methodology allows the \textit{juggling with modalities} and injects great flexibility and versatility.
The method enables the generation of multi-modal feature vectors and consistent annotation across modalities.
By these means, the proposed association mechanism fosters joint semantic analysis and consequently contributes to the completion of the hybrid processing pipeline.
\section{Related Work}
\label{sec:relatedwork}
The semantic segmentation of 3D data has become a standard task in the domain of photogrammetry and remote sensing.
The increasing availability of simultaneously acquired airborne data with different acquisition methods calls for multi-modal fusion and scene analysis (cf.~subsections~\ref{subsec:relatedwork_datafusion} and~\ref{subsec:relatedwork_semseg}).
Generally, the semantic analysis deals with various representations such as imagery, voxels, \acp{PC}, and meshes.
Regardless of modality, state-of-the-art \ac{ML} methods rely on a large amount of \ac{GT} data.
We briefly review available \ac{GT} of geospatial data in~\autoref{subsec:relatedwork_gt}.
\subsection{Multi-Modal Data Fusion}
\label{subsec:relatedwork_datafusion}
Due to complementary acquisition methods, multi-modal data acquisition has the potential to generate more complete and more detailed mapping products.
Thereby, multi-modal products feature improved geometric reconstruction and semantic analysis.
However, to the best of our knowledge, multi-modality is kept at a minimum and hence scratches only the surface.
For instance, the fusion of imagery and \ac{ALS} data on the point-level is commonly confined to the colorization of \ac{ALS} points \citep{V2D,cramer2018monitoring}.
To the best of our information, the explicit fusion of \ac{ALS} points and \ac{MVS} points and its contribution to semantics has not yet been investigated.
Possible reasons might be structural and georeferencing discrepancies across modalities as discussed in~\autoref{sec:intro} and the huge memory footprint as caused by redundant multi-modal capturing.
\cite{glira2019} propose a methodology to jointly orientate imagery and \ac{ALS} data which simplifies the fusion of the derived \acp{PC} as a side effect.
Recently, there are software solutions that enable data fusion of multi-modal \acp{PC} and refine the fusion on the mesh-level.
For instance, software SURE by nFrames \citep{rothermel2012sure} produces meshes as generated from LiDAR and \ac{MVS} dealing with orientation discrepancies of few \acp{GSD}.
As outlined in \autoref{subsec:relatedwork_semseg}, the majority of works for semantic interpretation involves only one modality in the narrow sense.
In most cases, multi-modality is a means to an end that allows abusing annotated data and well-performing classifiers of another modality.
To give an example, the well-established and fast semantic segmentation of images is mostly abused as a proxy to 3D scene analysis \citep{Boulch2017,Lawin2017,He2013,Su2015,Kalogerakis_labelMeshes}.
Theoretically, any quantity can be projected into image space adding another channel to the image.
In practice, the curse of dimensionality prevents the projection of an arbitrary number of quantities.
\cite{Peters2019} highlight issues of associating \acp{PC} and imagery, particularly time-shifts and occlusions.
To by-pass the occlusion problem, they approximate the 3D surface by voxelization of the \ac{PC}.
Our work differs from existing works since it explicitly aims at a holistic multi-modal data fusion of imagery, \acp{PC} and meshes.
Thereby, the mesh acts as core modality to solve the occlusion problem.
The subsequent information transfer shares features and labels with all modalities.
The association of an \ac{ALS} \ac{PC} and a challenging 2.5D mesh is already described in \cite{Laupheimer2020_ISPRS}.
In the current work, we improve the implementation to cope with 3D meshes with a significantly larger memory footprint than 2.5D meshes.
Moreover, we extend the association mechanism to image space (cf.~\autoref{subsec:imgmeshassociation} and~\autoref{subsec:pcimgassociation}) and enable information transfer in arbitrary directions (cf.~\autoref{tab:association_options}).
\subsection{Semantic Segmentation of 3D Data}
\label{subsec:relatedwork_semseg}
\ac{DL} methods, particularly \acp{CNN}, are state of the art for semantic segmentation in image space \citep{Garcia-Garcia2017,Minaee2020}.
Therefore, it seems reasonable to apply well-established \ac{DL} methods of the image space to \acp{PC}.
However, the unstructured nature of 3D \acp{PC} prevents to apply \acp{CNN} directly to them.
To overcome the non-Euclidean design, \acp{PC} are commonly structured into grid-like 3D or 2D representations by voxelization or multi-view rendering respectively.
Several works voxelize the \ac{PC} and train a supervised classifier. The predicted labels for the voxels will be transferred to all contained points \citep{hackel_pc_segementation,Huang2016}.
Voxelization comes along with memory overhead.
Therefore, much effort is put into networks that use sparse 3D convolutions \citep{Graham2018}.
This approach has been successfully applied to urban \acp{PC} \citep{Schmohl2019}.
Detouring via image space, multi-view approaches leverage well-performing semantic image segmentation methods. The per-pixel predictions are back-projected to 3D space \citep{Boulch2017,Lawin2017}.
To give an example, \cite{He2013} segment stereo images semantically, create the \ac{MVS} \ac{PC}, and back-project the 2D semantic segmentation results to the \ac{PC}.
The grid-like proxy enables the use of \acp{CNN} but comes along with information loss due to discretization, occlusions, and projection.
The rise of PointNet and its hierarchical successor PointNet++ constitutes a milestone in semantic \ac{PC} segmentation since they operate directly on unstructured 3D \acp{PC} \citep{qi2017pointnet, qi2017pointnet++}. \cite{Winiwarter2019} successfully applied PointNet++ to geospatial \acp{PC}.
The gist of PointNet is to use a symmetric function during encoding to be independent of set permutation.
The entire \ac{PC} is encoded by a global feature vector, which is attached to each encoded per-point feature vector.
Operating only on a global scale, PointNet misses local context.
Its extension PointNet++ hierarchically applies PointNets to the iteratively subsampled \ac{PC} and, hence, operates on several scales.
This procedure mimics hierarchical feature learning with increased contextual information similar to \acp{CNN} in image space.
Likewise, \cite{Boulch2019} introduces continuous convolutional kernels that can be applied directly to \acp{PC}.
\cite{Griffiths2019a} review the current state-of-the-art \ac{DL} architectures for processing 3D data.
\cite{Xie2020} review semantic \ac{PC} segmentation comparing \ac{DL} and traditional \ac{ML} approaches.
While \ac{DL} approaches do not require handcrafted features, they rely on a large amount of training data.
In contrast, traditional \ac{ML} depends on handcrafted features and therefore provides better interpretability.
\cite{WeinmannEtAl2015} calculate and select features based on various vicinities and subsequently perform a semantic segmentation with \ac{RF}.
To avoid noisy predictions, \cite{Landrieu2017b} extend the previous pipeline by structured regularization, a graph-based contextual strategy.
Likewise, \cite{NIEMEYER2014} avoid noisy results utilizing CRF-based methods as statistical context models.
\cite{Vosselman2017} first segment the data and subsequentially perform the semantic segmentation.
\cite{Ahmed2018} show advances of \ac{DL} on different 3D data representations.
They discuss representation-specific challenges and highlight differences between Euclidean and non-Euclidean data.
The emerging field of geometric \ac{DL} extends basic \ac{DL} operations to non-Euclidean domains such as graphs and manifolds in order to use topological information \citep{Bronstein2016}.
\acp{PC} do not provide topological information per se.
Therefore, \cite{Landrieu2017} organize \acp{PC} in \acp{SPG}.
\cite{AliKhan2020} transform \acp{PC} to an undirected symmetrically weighted graph encoding the spatial neighborhood and apply a Graph Convolutional Network.
\cite{Chang2018} propose the \ac{SACNN} that uses generalized filters, which aggregate local inputs of different learnable topological structures.
By that, \glspl{SACNN} work with both Euclidean and non-Euclidean data.
To summarize, the adaption of (geometric) \ac{DL} methods contributed to substantial progress in the field of semantic \ac{PC} segmentation in the last decade.
On the contrary, mesh interpretation has hardly been explored by the community of photogrammetry and remote sensing although recent years show increasing interest in meshed 3D models - particularly, for applications like smart city models \citep{Boussaha2018}.
In comparison, meshes are a default data representation in the domain of computer vision.
However, that community typically deals with small-scale (indoor) data sets \citep{Kalogerakis_labelMeshes}.
In contrast to photogrammetric meshes, texture is not an inherent characteristic of these meshed models.
By analogy to semantic \ac{PC} segmentation, common approaches for semantic mesh segmentation make a circuit to 2D image space to take advantage of image-based \ac{DL} methods.
Those approaches render 2D views of the 3D scene, learn the segmentation for different views, and finally, back-project the segmented 2D images onto the 3D surface \citep{Su2015,Kalogerakis_labelMeshes}.
\cite{wu20153d} voxelize the mesh and apply a convolutional deep belief network.
\cite{qiao2019laplaciannet} propose a geometric \ac{DL} approach that encodes the mesh connectivity using Laplacian spectral analysis and aggregates global information via mesh pooling blocks.
MeshCNN mimics traditional \ac{CNN} convolution and pooling operations \citep{Hanocka2019}.
The specialized layers operate on the edges and leverage the intrinsic topological information.
\cite{Schult20CVPR} propose the DualConvMesh-Net that combines geodesic and Euclidean convolutions on 3D meshes.
Geodesic convolutions utilize the underlying mesh structure and help to separate spatially adjacent but disconnected surfaces.
In contrast, Euclidean convolutions establish connections between nearby disconnected surfaces.
Notwithstanding, semantic segmentation of real-world large-scale meshes is a mostly overlooked topic.
\cite{Rouhani2017} gather faces of a \ac{MVS} mesh into so-called superfacets and train a \ac{RF} using geometric and photometric features.
\cite{tutzauer2019} utilize a \ac{DL} approach by training a multi-branch 1D \ac{CNN} with contextual features and compare the achieved results to a \ac{RF}.
They show that color information is beneficial for semantic mesh segmentation.
More precisely, \cite{Laupheimer2020_DGPF} attest that per-face color information (i.e. texture) outperforms per-vertex color information (e.g. colored \ac{PC}) by evaluation of several radiometric feature qualities.
However, they also show the inherent limitations of texture due to occlusions, absence of imagery, and the quality of the geometric reconstruction.
\subsection{Ground Truth Availability}
\label{subsec:relatedwork_gt}
\cite{Garcia-Garcia2017} and \cite{Minaee2020} review available \ac{GT} data in image space, 2.5D and 3D space.
Annotated imagery often aims at the pure semantic segmentation, wherefore orientation information is not provided \citep{Lambert2020}.
The ISPRS 2D Semantic Labeling Contest provides manually annotated orthophotos of Vaihingen and Potsdam \citep{V2D}.
\cite{Griffiths2019a} list available \ac{GT} data sets for RGB\hbox{-}D, multi-view, volumetric, and fully end-to-end architecture designs as acquired by various platforms.
\cite{Xie2020} review publicly available annotated \acp{PC} and discuss their shortcomings.
The computer vision community provides annotated mesh data for indoor scenes \citep{Armeni2017,Hua2016, Dai2017} or for single objects \citep{shilane2004princeton}.
However, to the best of our knowledge, there are no labeled meshed models that cover urban scenes.
In contrast, there are many available labeled urban data sets for 3D \acp{PC} provided by the community of photogrammetry and remote sensing \citep{Zolanvari2019, Wichmann2018, NIEMEYER2014,Hackel2017}.
The rise of data-hungry \ac{DL} methods demands efficient strategies for \ac{GT} generation.
Synthetically generated \ac{GT} such as provided by \cite{Griffiths2019b} boost the generation process per se.
However, purely synthetic data is limited by its diversity.
\cite{Koelle2020} exploit crowdsourcing and active learning to minimize manual labeling effort.
\cite{ramirez2019shooting} present a virtual reality tool that gamifies the manual labeling of meshes and \acp{PC}.
Our proposed methodology is able to derive consistently labeled meshes and imagery from publicly available annotated real-world \ac{PC} data and vice versa (provided that the necessary data is available and oriented, cf.~\autoref{sec:results_discussion}).
To the best of our knowledge, yet, there is no data set that provides consistently labeled modalities.
The proposed labeling tool has the potential to accelerate multi-modal \ac{GT} generation and consequently multi-modal semantic analysis.
\section{Data}
\label{sec:data}
To demonstrate the effectiveness of our association mechanism, we utilize the publicly available ISPRS benchmark data set \ac{V3D} and a proprietary data set which will be made publicly available in mid 2021 \citep{Cramer2010,cramer2018monitoring}.
The original purpose of the proprietary data set aims at the deformation monitoring of the ship lock and its surrounding in Hessigheim, Germany.
Thus, challenging water surfaces are part of the acquired data.
We refer to this data set as \ac{H3D}.
Although being already captured in 2008, \ac{V3D} may still be representative of large-scale country-wide mapping.
On the contrary, \ac{H3D} is an example of small-scale mapping applications with high-resolution imagery and LiDAR data.
In both cases, imagery and \ac{ALS} data have been acquired from airborne platforms.
Whereas \ac{V3D} data is captured from airplane, \ac{H3D} data is captured from \acp{UAV}.
\ac{V3D} data has been acquired asynchronously.
The time-shift between nadir imagery (GSD\,=\,\SI{8}{cm}) and \ac{ALS} acquisition (\SIrange[range-units=single, range-phrase=--]{4}{8}{points/m^2}) is several weeks.
\ac{H3D} provides two sets of oriented imagery: oblique and nadir.
Oblique imagery (GSD\,=\,\SI{2.5}{cm}) has been acquired simultaneously along with \ac{ALS} data (\SIrange[range-units=single, range-phrase=--]{400}{800}{points/m^2}) from the same \ac{UAV}.
Nadir images (GSD\,=\,\SI{3.7}{mm}) have been acquired from another \ac{UAV} with a time-shift of several hours to one day.
Accordingly, the number of entities and the memory footprint is higher for \ac{H3D}.
\autoref{tab:data} lists key parameters of both data sets relevant for the underlying study.
\begin{table}[htbp]
\centering
\footnotesize
\begin{tabular}{P{2.5cm}P{3cm}P{2.5cm}P{2.5cm}}
\toprule
\normalsize Data Set &
\normalsize Imagery &
\normalsize PC &
\normalsize Mesh
\\
\toprule
\minitab{
\ac{V3D} \\
\SI{870}{m}~$\times$~\SI{700}{m}
}
&
\minitab{
$\text{GSD}_{\text{nadir}} = \SI{8}{cm}$ \\
15 images \\
@ \SI{14430}{px} $\times$ \SI{9420}{px}
}
&
\minitab{
\SIrange[range-units=single, range-phrase=--]{4}{8}{points/m^2} \\
\SI{1.2}{M} points
}
&
\minitab{
16 tiles \\
\SI{3.3}{M} faces \\
source: \ac{MVS}$_{\text{nadir}}$
}
\\
\midrule
\minitab{
\ac{H3D} \\
\SI{670}{m}~$\times$~\SI{255}{m}
}
&
\minitab{
$\text{GSD}_{\text{oblique}} = \SI{2.5}{cm}$ \\
1979 images \\
@ \SI{6000}{px} $\times$ \SI{4000}{px} \\
\cline{1-1}$\text{GSD}_{\text{nadir}} = \SI{3.7}{mm}$ \\
524 images \\
@ \SI{11608}{px} $\times$ \SI{8708}{px}
}
&
\minitab{
\SIrange[range-units=single, range-phrase=--]{400}{800}{points/m^2} \\
$\SI{138.1}{M}$ points
}
&
\minitab{
94 tiles \\
$\SI{24.6}{M}$ faces\\
source: \\
\ac{MVS}$_{\text{oblique}}$ $+$ \ac{ALS}
}
\\
\bottomrule
\end{tabular}
\caption{\ac{V3D} and~\ac{H3D} properties. The project area is given by the mesh tiles that intersect with the labeled LiDAR cloud. The face count is adapted to the overlapping LiDAR cloud.
}
\label{tab:data}
\end{table}
For both data sets, we generate textured and tiled meshes with SURE 4.0.2 from nFrames.
We set tile sizes to $\SI{175}{m} \times \SI{175}{m}$ (\ac{V3D}) and $\SI{50}{m} \times \SI{50}{m}$ (\ac{H3D}) respectively.
The chosen tile sizes empirically showed to be a good compromise between fast tile-wise processing and small tile count.
For H3D, we generate a hybrid textured mesh by fusing the simultaneously acquired \ac{ALS} data and oblique imagery.
Utilizing oblique imagery ensures proper texturing of vertical faces such as facades.
In contrast, we generate a purely photogrammetric mesh for \ac{V3D} since the time-shift of imagery and \ac{ALS} data is roughly one month.
For this reason, the geometric and radiometric quality of the \ac{H3D} mesh outperforms the \ac{V3D} mesh.
However, since the \ac{V3D} mesh is purely photogrammetric, the relative orientation of imagery and mesh fits perfectly.
We determine the shifts between mesh and \ac{PC} data with the \ac{ICP} algorithm.
We do not apply the determined shifts to prove the effectiveness of our association methodology.
Moreover, \ac{ICP} does not solve the co-registration problem entirely.
The \ac{V3D} \ac{ALS} data is shifted against the \ac{MVS} mesh by
\mbox{$\Delta X=\SI{31}{cm}$},
\mbox{$\Delta Y=\SI{-9}{cm}$},
\mbox{$\Delta Z=\SI{9}{cm}$}.
Rephrased, the co-registration of imagery and \ac{ALS} data differs significantly.
The \ac{H3D} \ac{ALS} data is shifted against the mesh (as generated from LiDAR and \ac{MVS} points) by
\mbox{$\Delta X=\SI{8}{mm}$},
\mbox{$\Delta Y=\SI{10}{mm}$},
\mbox{$\Delta Z=\SI{6}{mm}$}.
The significantly different data sets featuring co-registration issues in 3D space are adequate to showcase the robustness and flexibility of our implementation.
Both data sets carry manual annotations for the LiDAR cloud \citep{NIEMEYER2014,KoelleEtAl2019}.
The label scheme of \ac{H3D} is oriented towards the \ac{V3D} label scheme but is more fine-grained.
Furthermore, it has been manually enhanced by class \textit{Chimney/Antenna} \citep{Laupheimer2020_DGPF}.
\autoref{fig:data_sets} shows the union of textured mesh tiles that overlap with the labeled LiDAR cloud for both data sets.
The label schemes and respective color-codings are given in the figure caption.
\begin{figure}[htbp]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.568\columnwidth]{_figures/Figure2_left.png} &
\includegraphics[width=0.285\columnwidth]{_figures/Figure2_right.png}
\end{tabular}
\caption{
Top views of \ac{V3D} (\textit{left}) and \ac{H3D} (\textit{right}) depicting the annotated LiDAR \ac{PC} and the respective overlapping mesh tiles in textured fashion.
The annotated \ac{ALS} data is color-coded utilizing the following label schemes.
\newline
\ac{V3D}: \newline
\textit{Power Line} (black), \textit{Low Vegetation} (light green), \textit{Impervious Surface} (gray), \textit{Car} (blue), \textit{Fence/Hedge} (yellow), \textit{Roof} (red), \textit{Facade} (white), \textit{Shrub} (dark green), and \textit{Tree} (green).
\newline
\ac{H3D}:\newline
\textit{Power Line} (black), \textit{Low Vegetation} (light green), \textit{Impervious Surface} (gray), \textit{Vehicle} (blue), \textit{Urban Furniture} (lilac), \textit{Roof} (red), \textit{Facade} (white), \textit{Shrub/Hedge} (orange), \textit{Tree} (green), \textit{Open Soil/Gravel} (brown), \textit{Vertical Face} (yellow), \textit{Chimney/Antenna} (magenta).
}
\label{fig:data_sets}
\end{figure}
\section{Methodology}
\label{sec:methodology}
We aim for a holistic explicit linking of the common data representations in the domain of photogrammetry and remote sensing: imagery, \ac{PC}, and mesh.
The backbone of the proposed association methodology consists of two geometry-driven parts:
(a)~\ac{PCMA} which links faces and points (cf.~\autoref{subsec:pcmeshassociation}) and (b)~\ac{ImgMA} which links faces and pixels across images (cf.~\autoref{subsec:imgmeshassociation}).
Coupling both association mechanisms yields to (c)~\ac{PCImgA} (cf.~\autoref{subsec:pcimgassociation}).
The \ac{PCImgA} establishes a connection between points and imagery via the mesh as a mediator.
Point visibility is implicitly given through the mesh.
\autoref{tab:association_options} illustrates the total association mechanism with iconic pictograms.
\begin{table}[htbp]
\resizebox{\textwidth}{!}
\centering
\footnotesize
\begin{tabular}{cccc}
\toprule
\multicolumn{4}{c}{} \\
\multicolumn{4}{c}{
\centering
\includegraphics[width=1.0\linewidth]{_figures/Table2_top.eps}
\put (-299.5, 80) {Point Cloud Mesh Association}
\put (-150, 80) {Image Mesh Association}
\put (-229.5,-10) {Point Cloud Image Association}
\put (-255, 48) {$n_\mathit{pts}:1$}
\put (-127, 48) {$1:n_\mathit{px}:n_\mathit{img}$}
\vspace{0.5cm}
} \\
\toprule
\toprule
\multicolumn{4}{l}{(a) \acrfull{PCMA}}\\
\bottomrule
&
Mesh $\mapsto$ PC ($1:n_\mathit{pts}$) &
PC $\mapsto$ Mesh ($n_\mathit{pts}:1$) &
\multirow{4}{*}{
\centering
\includegraphics[width=0.10\linewidth]{_figures/Table2_right_top.eps}
}\\
\cmidrule[0.5pt](l{0.5em} r{0.5em}){2-2}
\cmidrule[0.5pt](l{0.5em} r{0.5em}){3-3}
Feature Transfer &
Copy Value &
Median Aggregation \\
Label Transfer &
Copy Value &
Majority Vote & \\
& & \\
\toprule
\multicolumn{4}{l}{(b) \acrfull{ImgMA}}\\
\bottomrule
&
Mesh $\mapsto$ Img ($1:n_\mathit{px}:n_\mathit{img}$) &
Img $\mapsto$ Mesh ($n_\mathit{img}:n_\mathit{px}:1$) &
\multirow{4}{*}{
\centering
\includegraphics[width=0.25\linewidth]{_figures/Table2_right_center.eps}
}\\
\cmidrule[0.5pt](l{0.5em} r{0.5em}){2-2}
\cmidrule[0.5pt](l{0.5em} r{0.5em}){3-3}
Feature Transfer &
Copy Value &
Median Aggregation &\\
Label Transfer &
Copy Value &
Majority Vote & \\
& & &\\
\toprule
\multicolumn{4}{l}{(c) \acrfull{PCImgA}}\\
\bottomrule
&
PC $\mapsto$ Img &
Img $\mapsto$ PC &
\multirow{4}{*}{
\centering
\includegraphics[width=0.25\linewidth]{_figures/Table2_right_bottom.eps}
}\\
\cmidrule[0.5pt](l{0.5em} r{0.5em}){2-2}
\cmidrule[0.5pt](l{0.5em} r{0.5em}){3-3}
Feature Transfer &
Median Aggregation &
Median Aggregation & \\
Label Transfer &
Majority Vote &
Majority Vote & \\
& & & \\
\bottomrule
\end{tabular}
}
\caption{Overview of the proposed method. For each association mechanism, the transfer operations are given in dependence of the information type (feature or label) and the transfer direction.
The pictograms on the right depict the linking of the respective entities.
\ac{PCImgA} provides two association modes: implicit and explicit linking (cf.~\autoref{subsec:pcimgassociation}).
(a), (b) and the implicit version of (c) are face-centered.
The relationship of implicit \ac{PCImgA} is described by $n_\mathit{pts}:1:n_\mathit{px}:n_\mathit{img}$.
The explicit version is pixel-centered (PC~$\mapsto$~Img: $n_\mathit{pts}:1:n_\mathit{img}$) or point-centered (Img~$\mapsto$~PC: $n_\mathit{img}:n_\mathit{px}:1$).
}
\label{tab:association_options}
\end{table}
The established connections between the entities across the distinct representations enable an information transfer that allows features and labels to be shared arbitrarily.
\autoref{tab:association_options} compactly lists the information transfer operations depending on information type (feature or label) and transfer direction for each part of the entire association mechanism.
Concerning the scalability of the proposed multi-modal association approach, we process data tile-wise in a parallelized fashion while keeping the memory footprint low.
\subsection{Point Cloud Mesh Association (PCMA)}
\label{subsec:pcmeshassociation}
The \ac{PCMA} explicitly links faces and points in a face-centered geometry-driven approach.
Each face (represented by its \ac{COG}) is assigned with $n_\mathit{pts}$~points that represent the same surface by following three steps:
(i) clipping of the \ac{PC} to a spherical vicinity of the \ac{COG},
(ii) filtering of \textit{out-of-face points},
and (iii) filtering of \textit{off-the-face points} (\autoref{fig:methodology_association_pc_mesh}).
\textit{Out-of-face points} are not enclosed by the face borders when projected orthogonally onto the face plane.
\textit{Off-the-face points} do not coincide with the face plane, i.e. they are below or above the face surface.
A manually set threshold~$\theta$ decides whether a point coincides with a face or not.
Both point types are not mutually exclusive and exist due to the simplification during the meshing, the representation type differences as discussed in~\autoref{sec:intro}, and geometry differences (e.g. in case of 2.5D mesh geometry or due to asynchronous data acquisition).
\begin{figure}[htbp]
\centering
\includegraphics[width=0.675\linewidth]{_figures/Figure3.pdf}
\put (-122.5, 43) {(i)}
\put (-191.5, -4) {(ii)}
\put (-58, -4) {(iii)}
\caption{
Steps (i) - (iii) of the \ac{PCMA}.
\textit{(i)}: Clipping of the \ac{PC} (black dots) to the vicinity (blue sphere) of the considered face. Its \ac{COG} is marked with a black cross.
The mesh surface and its vertices are depicted in green.
\textit{(ii)}: Filtering of \textit{out-of-face points} based on the clipping result (orthogonal view concerning the face surface).
\textit{(iii)}: Filtering of \textit{off-the-face points} (side view with respect to the face).
The face is depicted as a black line.
The threshold band is marked in gray.}
\label{fig:methodology_association_pc_mesh}
\end{figure}
At first, we roughly reduce the search space for each face in order to accelerate the association.
To this end, we build a kD tree for the \ac{PC} (tile) and query the built tree with \acp{COG} of all faces.
Thereby, we detect all points within distance~$r$ for each face (ball query).
The query parameter~$r$ is set in dependence of the manually set association threshold~$\theta$ and the maximum distance~$t_{\max}$ of the \ac{COG} to the respective face vertices.
Geometrically, $r$~is set to the length of the hypotenuse of the triangle as defined by $\theta$ and~$t_{\max}$.
In simple terms, the query parameter~$r$ is set to the minimum distance that guarantees the manually set threshold~$\theta$ to be effective for the entire face while enclosing the entire face (cf.~\autoref{fig:sphere_radius}).
Hence, $r$ prevents prefiltering of points by a too small spherical vicinity.
The ball query delivers a subset of points, which may contain \textit{off-the-face points} and \textit{out-of-face points}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.125\linewidth]{_figures/Figure4_left.eps}
\hspace{3.5cm}
\includegraphics[width=0.125\linewidth]{_figures/Figure4_right.eps}
\put (-173, 15) { \tiny $t_{\max}$ }
\put (-153.5, 24.5) { \tiny $\theta$ }
\put (-168, 27) { \tiny $r$ }
\put (-132, 20) { \footnotesize $r = \sqrt{ t^2_{\max} + \theta^2 }$}
\put (-28, 15) { \tiny $t_{\max}$ }
\put (-12, 28) { \tiny $\theta$ }
\put (-24, 31) { \tiny $r$ }
\put (-188.5, -4) { \tiny $t_{\max} \geq \theta$ }
\put (-39, -4) { \tiny $t_{\max} < \theta$ }
\caption{Definition of radius~$r$ (\textit{blue}) for the \ac{PC} clipping (step (i) of \ac{PCMA}) shown as side view with respect to the face (\textit{dashed black line}).
Radius~$r$ guarantees to enclose the entire face by enclosing the maximum distance $t_{\max}$ (\textit{dashed horizontal orange line}) of the \ac{COG} (\textit{black cross}) to face vertices. Besides, $r$ avoids to prefilter points by enclosing threshold~$\theta$ (\textit{dashed vertical orange line}) across the entire face.
The threshold band is marked in gray.
}
\label{fig:sphere_radius}
\end{figure}
Second, we filter \textit{out-of-face points} by neglecting subset points whose orthogonal projections on the face plane are not enclosed by the face outline.
For details, we refer the interested reader to \cite{Laupheimer2020_ISPRS}.
Visually, the result of (ii) is the intersection of the spherical subset with radius~$r$ and the infinite triangular prism as defined by the face and its normal vector. We refer to this as the \textit{association prism}.
Finally, we filter the remaining \textit{off-the-face points}.
For this purpose, we calculate the orthogonal distance for each remaining point to the face plane.
If the distance exceeds a chosen association threshold~$\theta$, the point is not associated with the face.
Since we have to compensate several discrepancies between \ac{PC} and mesh, we use a more sophisticated adaptive thresholding with an arbitrary user-defined number of filter levels~$n_l$.
Each level~$l$ consists of two independent thresholds~$\theta_l^+$ and~$\theta_{l}^-$ limiting the association prism in the normal direction or the opposite direction respectively.
The absolute threshold values increase with ascending level.
Starting from level~1, the algorithm tries to associate points with the respective thresholds~$\theta_1^+$ and~$\theta_{1}^-$.
If points have been linked, the association stops.
Otherwise, the next level~$l$ is activated.
This adaptive thresholding accelerates the association process.
On the other hand, by nature of our approach, not all points might be associated.
Here, our reasoning is to favor near-surface points at the cost of missing to link a few points (fast small margin association).
The association information is stored as a per-point attribute.
For each associated point, the respective face index is attached to its attributes.
Non-associated points are marked with~$-1$.
The stored indices trivialize the transfer of features and labels from the mesh to the \ac{PC} (Mesh~$\mapsto$~PC in~\autoref{tab:association_options}).
In this case, we copy the desired values to the \ac{PC} at a stroke due to the one-to-many relationship.
Reversely, the stored face indices can also be used to transfer features and labels from the \ac{PC} to the mesh (PC~$\mapsto$~Mesh in~\autoref{tab:association_options}).
However, to speed up the transfer, we directly couple the information transfer with the association mechanism.
The many-to-one relationship calls for information aggregation.
For each face, we derive robust median features as gathered from the \ac{PC}.
Features may embrace sensor-intrinsic and handcrafted features such as pulse characteristics and derived quantities \citep{Eitel2016}.
Analogously, majority votes determine the per-face labels as transferred from the associated points.
Therefore, the association inherently is a label transfer tool and feature calculation tool (median features).
Non-associated faces are marked with~$-1$ and receive zeroed median features \citep{Laupheimer2020_ISPRS}.
\subsection{Image Mesh Association (ImgMA)}
\label{subsec:imgmeshassociation}
The \ac{ImgMA} explicitly links faces and pixels across various images in a geometry-driven approach.
To accelerate the \ac{ImgMA} we make use of the given mesh tiling.
Each pixel is assigned with the visible face as detected by the following three steps:
(I) preselection of visible mesh tiles per image utilizing \acp{MBB} of the tiles,
(II) ray casting per image and tile (image-tile-pair),
and (III) fusion of ray casting results per image via depth filtering across tiles.
We end up with $n_\mathit{px}$ associated pixels across $n_\mathit{img}$ images for each face.
\autoref{fig:methodology_association_img_mesh} shows the workflow by means of an oblique example image and two vertically separated tiles.
\begin{figure}[htbp]
\centering
\begin{tabular}{cc}
\multicolumn{2}{c}{
\includegraphics[width=0.4\linewidth]{_figures/Figure5_top.eps} \vspace{-0.25cm}}
\\
\multicolumn{2}{c}{(I)}
\vspace{0.25cm}
\\
\includegraphics[width=0.3\linewidth]{_figures/Figure5_bottom_left.png} \hspace{0.5cm} &
\hspace{0.5cm} \includegraphics[width=0.3\linewidth]{_figures/Figure5_bottom_right.png}
\vspace{-0.25cm}
\\
(II) \hspace{0.5cm} & \hspace{0.5cm} (III)
\end{tabular}
\caption{
Steps (I) - (III) of the \ac{ImgMA} shown in accordance with the information transfer Mesh~$\mapsto$~Img (cf.~\autoref{tab:association_options}).
The depicted example shows the association of an image (\textit{lower right}) with two vertically split mesh tiles of \ac{H3D}.
\textit{(I)}: Preselection of visible mesh tiles per image shown schematically in isometric (\textit{left}) and top view (\textit{right}).
The stretched camera pyramid (\textit{green}) intersects with some \acp{MBB} (\textit{blue)}. Non-intersecting \acp{MBB} are marked in \textit{reddish}. Dashed lines indicate non-visible parts.
\textit{(II)}: Ray casting per image-tile-pair.
The \textit{reddish} area shows where ray casting fails due to missing intersections of image rays and considered mesh tiles.
\textit{Black} indicates intersection with unlabeled faces.
\textit{(III)}: Final result after fusing ray casting results per image via depth filtering across tiles.
}
\label{fig:methodology_association_img_mesh}
\end{figure}
For each image, we first detect visible tiles and perform the subsequent ray casting procedure in a parallelized fashion for the subset of tiles only.
We define a tile to be visible if its \ac{MBB} intersects with the stretched camera pyramid (\ac{MBB} visibility check).
The stretched camera pyramid is defined by the projection center and the projection rays crossing the corner pixels of the respective image.
The lowest of all \ac{MBB} faces limits the stretched camera pyramid.
Since the \acp{MBB} are not fully occupied by the enclosed mesh tiles, some detected tiles might not contain any visible faces.
Nonetheless, this approach significantly reduces the number of tiles that have to be processed by the ray casting procedure for each image.
We detect intersections of the camera pyramid and a \ac{MBB} by a three-stage check starting with the most likely and fastest test.
At first, we check for each tile if any corner point of the respective \ac{MBB} is inside the camera pyramid (point in polyhedron test).
The second and third stage perform edge face intersections, checking intersections of pyramid edges starting from the projection center and any \ac{MBB} face or intersections of \ac{MBB} edges and any pyramid face.
Once a test succeeds, the respective enclosed tile is marked as visible and the residual checks are omitted (check omission).
As a result of (I), we receive a list of visible tiles for each image, i.e. each image~$\mathit{img}$ is linked to $n_\mathit{tiles}^\mathit{img}$~tiles.
Hence, there are $n_\mathit{tiles}^\mathit{img}$~image-tile-pairs for each~$\mathit{img}$.
Vice versa, a list of visible images for each tile is stored.
For each image-tile-pair, visible faces are determined via ray casting.
For this purpose, for each pixel, a 3D projection ray is created and intersected with the mesh faces of all linked tiles.
The intersected faces are candidates for the final association result.
Concerning a single image-tile-pair, all candidate faces are truly visible.
However, an image probably covers multiple tiles, and consequently, some faces might be occluded by faces of another tile (cf.~\autoref{fig:methodology_association_img_mesh}).
For the final result, we fuse the $n_\mathit{tiles}^\mathit{img}$ ray casting results across the visible tiles into one final ray casting result per image.
Implementationally, the fusion across the $n_\mathit{tiles}^\mathit{img}$ image-tile-pairs is done implicitly by depth updates.
Initially, each pixel is associated with a near-infinite depth value.
The association information is updated whenever a candidate face reduces the depth value.
Hence, the final association information is steered by faces of minimum depth that are truly visible (i.e. faces that mark the first intersection along the respective ray).
To speed up the implementation, steps (II) and (III) are parallelized with respect to images.
Each process handles $n_\mathit{tiles}^\mathit{img}$~image-tile-pairs per image.
The implicit fusion reduces the memory footprint of the algorithm since only the final ray casting result has to be stored.
To minimize the memory footprint, we avoid storing the association information channel-wise due to the curse of dimensionality.
Particularly for oblique images, only a small part of the image may be associated due to the limited reconstruction area of the mesh.
Instead, we store the final association information as a \textit{sparse pixel cloud} per image consisting only of pixels that have been linked with a face.
The sparse pixel cloud contains tuples of associated pixel positions, the depth, the tile-dependent face index, and optionally, other attributes (e.g. labels) as transferred from the mesh.
The stored face indices trivialize the transfer of features and labels from the mesh to the images (Mesh~$\mapsto$~Img in~\autoref{tab:association_options}).
We copy the desired quantities to the linked pixels of the associated images at a stroke due to the one-to-many-to-many relationship.
Reversely, the many-to-many-to-one relationship calls for information aggregation (Img~$\mapsto$~Mesh in~\autoref{tab:association_options}).
For each face, we derive robust median features as gathered from the associated pixels across the respective images.
Features may embrace sensor-intrinsic multi-spectral information, handcrafted features, and features as derived by \ac{DL} pipelines.
Analogously, majority votes determine the per-face labels as transferred from the linked pixels.
\subsection{Point Cloud Image Association (PCImgA)}
\label{subsec:pcimgassociation}
The \ac{PCImgA} aims for the linking of pixel locations and 3D points.
Theoretically, the collinearity equations establish an explicit relationship between 3D points and pixels.
Each point can be projected into the image space given the exterior and interior orientation of the respective image.
However, the bare projection cannot check for visibility, and hence, links visible and non-visible points with imagery.
Therefore, point visibility has to be checked prior to the linking of pixels and 3D points.
To this end, we leverage the mesh representation by combining mechanisms \ac{PCMA} and \ac{ImgMA} implicitly and explicitly.
The implicit linking is face-centered whereas the explicit linking is point-centered or pixel-centered (dependent on the transfer direction).
Making a detour via the mesh largely solves the visibility problem for 3D points in image space.
The implicit linking couples the association mechanisms \ac{PCMA} and \ac{ImgMA} by simply executing them sequentially.
Specifically, $n_\mathit{px}$~pixels of $n_\mathit{img}$~images and $n_\mathit{pts}$~points are exclusively linked to the respective face.
For the information transfer, the information from the starting representation is gathered per face and transduced to the target representation.
Therefore, the joint face apparently establishes a linking of points and pixels.
The per-face label and features as derived from the starting representation are determined via majority voting and median aggregation respectively (cf.~\autoref{tab:association_options}).
Subsequentially, the per-face aggregations are copied to the target modality (one-to-many relationship).
On the contrary, the explicit linking couples the association mechanisms \ac{PCMA} and \ac{ImgMA} by leveraging the stored association information and utilizing the collinearity equations.
As a result of \ac{ImgMA}, visible faces for each image are known.
At the same time, mechanism \ac{PCMA} delivers the associated points for each face.
Consequently, associated points of visible faces are marked as visible.
Therefore, the combination of \ac{PCMA} and \ac{ImgMA} results in a visible subset of the \ac{PC} per image.
The collinearity equations explicitly link the visible points and the pixel locations across all imagery.
Therefore, explicit linking truly associates points and pixel locations.
For each point, there is an unambiguous pixel location in each image.
However, the reverse situation is ambiguous.
Depending on \ac{GSD} and point density, each pixel of each image may enclose several visible points.
Therefore, transferring information from the \ac{PC} to imagery (\ac{PC}~$\mapsto$~Img in \autoref{tab:association_options}) demands a pixel-wise aggregation (many-to-one relationship per pixel and image).
The features and labels are aggregated by median aggregation and majority voting respectively.
To accelerate the process, we approximate the pixel-wise aggregation by transferring only information of the point of minimum depth.
This approximation simplifies the association to a one-to-one relationship per pixel and image.
Likewise, a one-to-one relationship holds if the \ac{GSD} is smaller than the point distance.
Since each point is covered by several images, the information transfer Img~$\mapsto$~\ac{PC} requires a point-wise aggregation across images.
\subsection{Preconditions and Limitations}
\label{subsec:preconditions_limitations}
The proposed method connects 3D \acp{PC}, photographic imagery (following the central perspective), and textured meshes.
By nature of the algorithm, it merely depends on the pure existence of those three modalities.
Therefore, it is a generic approach that works with any photographic image and \ac{PC} regardless of the acquisition platform (aerial, terrestrial, mobile), the image type (panchromatic, RGB, multi-spectral), and \ac{PC} type (\ac{MVS} cloud, LiDAR cloud, persistent scatterer cloud).
However, we focus on the linking of aerial RGB imagery and \ac{ALS} \acp{PC} along with the respective textured 3D mesh (cf.~\autoref{sec:results_discussion}).
Furthermore, photogrammetric meshes, LiDAR meshes, or hybrid meshes can be processed.
The linking is not constraint to a specific mesh generation algorithm or mesh geometry (2.5D or 3D).
However, the entire association benefits from good geometric reconstruction.
We are aware of the fact that proper meshing of our complex world is a hard task and still subject to research.
A proper reconstruction ensures an appropriate entity linking and information transfer.
As a general rule, the better the mesh represents the true 3D structure of the real world, the better works the proposed association mechanism and the subsequent information transfer.
Obviously, as a precondition to the association, the considered data representations have to cover the same area and have to be oriented in the same coordinate system.
Assuming a proper co-registration, entities across representations can only be associated when underlying real-world objects are captured or reconstructed in all representations.
In other words, each face should at least enclose one point or one pixel respectively.
Reversely, each point or pixel should be mapped onto a corresponding face.
However, these relationships do not always exist due to inter-representation differences such as object penetration (cf.~\autoref{sec:intro}, \citeauthor{Laupheimer2020_ISPRS}, \citeyear{Laupheimer2020_ISPRS}).
For instance, due to occlusion during data acquisition, a facade may not be captured in the \ac{ALS} data, and hence the respective faces cannot link any points.
Reversely, due to mesh simplification, a thin-structured object may not have been reconstructed entirely in the mesh, but the object is fully captured in the \ac{PC} and imagery.
Additionally, there might be inter-representation discrepancies due to asynchronous data acquisition.
\autoref{fig:discrepancies_interrepresentation} shows both situations for the \ac{H3D} data set.
\begin{figure}[htbp]
\centering
\begin{tabular}{c c}
\multicolumn{2}{c}{\includegraphics[width=0.48\columnwidth]{_figures/Figure6_top.jpg}} \\
\includegraphics[width=0.28\columnwidth]{_figures/Figure6_bottom_left.png}
\hspace{0.5cm}
&
\hspace{0.5cm}
\includegraphics[width=0.28\columnwidth]{_figures/Figure6_bottom_right.png}
\end{tabular}
\caption{Discrepancies and data gaps in 3D models of \ac{H3D}.
\textit{Top}: Urban area. Overlay of textured mesh and \ac{ALS} \ac{PC} (height-coded).
Facades are reconstructed entirely in the mesh but are captured only partially by the \ac{PC}.
\textit{Bottom}: Ship lock area. The mesh (\textit{left}) partially reconstructs the river but misses to fully reconstruct thin structures like light poles and power lines, which are captured entirely in the \ac{PC} (\textit{right}). The asynchronous data acquisition of images and \ac{ALS} data cause inconsistencies between mesh and \ac{PC} (e.g. for cars).
}
\label{fig:discrepancies_interrepresentation}
\end{figure}
Misalignment among modalities is a decisive issue for multi-modality, wherefore proper co-registration is subject to current research.
Ideally, imagery and \ac{PC} data are co-registered simultaneously in a joint adjustment.
Consequently, the derived mesh is aligned with both data sources and co-registration issues are obsolete.
Nonetheless, reality shows that co-registration discrepancies are an important and real issue (cf.~\autoref{sec:data}).
A good relative orientation of 3D data and imagery is beneficial to the linking of pixels with points and faces.
We are aware of the fact that \ac{PCImgA} and \ac{ImgMA} depend on the quality of the co-registration.
However, the proper co-registration of both data sources is not the focus of this work.
\autoref{fig:img_labels} depicts the influence of the co-registration quality of imagery and 3D data.
Inherently, \ac{MVS} meshes and \ac{MVS} \acp{PC} are perfectly aligned with imagery.
Hence, incorrect or missing associations between the representations are only due to the reconstruction quality and data gaps (as for~\ac{V3D}).
\begin{figure}[htbp]
\centering
\includegraphics[width=0.75\columnwidth]{_figures/Figure7.jpg}
\caption{Influence of the co-registration quality as visualized by the label transfer (\textit{left/right}: good/bad relative orientation of 3D data and imagery). The yellow lines depict "epipolar lines" crossing roof corners in the RGB image (\textit{center}). The orientation parameters as achieved by bundle adjustment have been slightly falsified artificially for the visualization on the right.}
\label{fig:img_labels}
\end{figure}
Since the relationship of 3D space and image space is strictly defined by the collinearity equations, we highlight the discussion of \ac{PCMA}.
\autoref{fig:issue_pcmeshassociation_schematic} and \autoref{tab:points_unassociated} sketch discrepancies of \ac{PC} and mesh despite representing the same real-world objects.
These discrepancies are largely covered by the adaptive thresholding.
Despite and due to this technique, not all points are associated with faces.
There are three groups of unassociated points (cf.~\autoref{tab:points_unassociated}).
\begin{figure}[htbp]
\centering
\includegraphics[width=0.25\columnwidth]{_figures/Figure8_left.eps}
\hspace{0.5cm}
\includegraphics[width=0.25\columnwidth]{_figures/Figure8_center.eps}
\hspace{0.5cm}
\includegraphics[width=0.25\columnwidth]{_figures/Figure8_right.eps}
\caption{Discrepancies between \acp{PC} and the mesh (\textit{black line}) caught by adaptive thresholding.
\textit{Left}: \textit{Black arrows} indicate the discrepancy between the 2.5D mesh and the annotated 3D \ac{PC}.
\textit{Center}: The noisy \ac{PC} (\textit{blue}) oscillates about the reconstructed mesh.
\textit{Right}: There might be misalignment between \acp{PC} (\textit{blue}: \ac{MVS}, \textit{orange}: LiDAR) and the mesh as generated of a single source or complementary sources.}
\label{fig:issue_pcmeshassociation_schematic}
\end{figure}
\begin{table}[htbp]
\resizebox{\textwidth}{!}
\centering
\footnotesize
\begin{tabular}{c c}
\toprule
\makebox[8cm][l]{\normalsize A) points outside the threshold range} & \\
\midrule
\includegraphics[width=0.5\columnwidth]{_figures/Table3_top_left.eps}
&
\includegraphics[width=0.5\columnwidth]{_figures/Table3_top_right.eps} \\
A1: points outside the association prism &
A2: points in different threshold bands ("early stopping") \\
\midrule
\makebox[8cm][l]{\normalsize B) points outside the association prisms} & \\
\midrule
\includegraphics[width=0.35\columnwidth]{_figures/Table3_center_left.eps}
&
\includegraphics[width=0.5\columnwidth]{_figures/Table3_center_right.eps} \\
B1: noisy measurements and noisy reconstruction &
B2: misalignment \\
\midrule
\makebox[8cm][l]{\normalsize C) points on the association prism (optional)} & \\
\midrule
\includegraphics[width=0.25\columnwidth]{_figures/Table3_bottom_left.eps}
&
\includegraphics[width=0.25\columnwidth]{_figures/Table3_bottom_right.eps} \\
C1: points coincide with face vertices &
C2: points coincide with association prism boundary\\
\bottomrule
\end{tabular}
}
\caption{The schematic drawings illustrate the three cases where faces (\textit{black} lines, separated by \textit{black} strokes) and points are not linked (side view with respect to the face).
Non-associated points are marked in \textit{red}, associated points in \textit{green}.
The association prisms are marked in \textit{blue}.
Adaptive thresholds are omitted - except for~A2.
A2~depicts the increasing thresholds with increasing blueness.
Diverging association prisms create \textit{dead zones} like depicted in~B (\textit{hatched in red}).
B1~mimics a perfect planar surface as \textit{dashed black} line.}
\label{tab:points_unassociated}
\end{table}
Visually, the association is done utilizing the per-face association prism which is limited by the range of manually set thresholds~$\theta_l^+$ and~$\theta_{l}^-$.
The threshold-based approach filters all points whose orthogonal projections to the face plane exceed a specific value (cf.~\autoref{tab:points_unassociated}, A1).
This prevents the linking of points and faces that most likely represent different surfaces.
Furthermore, the adaptive thresholding breaks the association once a point is associated at any threshold level~$l$ ("early stopping").
Hence, points that are closer to the face than the maximum threshold may not be linked (cf.~\autoref{tab:points_unassociated}, A2).
Levels~$l>1$ can be seen as a fall-back for scenarios where a proper association is not possible.
The adaptive thresholding favors near-surface points by ensuring the association of points fulfilling the smallest threshold.
The adaptive thresholding facilitates a varying degree of freedom by the set thresholds and their inter-level spacing.
Eventually, it balances the strictness of the point-face linking.
Small-valued thresholds along with small-spaced levels enforce a tight coupling where only points close to the mesh surface are associated.
Large values along with large level spacing loosen the coupling.
Moreover, the asymmetric two-fold thresholding per level enables non-symmetric filtering improving flexibility and adaptiveness.
Therefore, the presented association mechanism is agnostic to the geometric structure of mesh geometry: 2.5D or 3D meshes can be processed.
The asymmetric adaptive thresholding allows associating faces with points where 2.5D and 3D geometry differ significantly while favoring the association of near-surface points (e.g. facades or tree stems, cf.~\autoref{fig:issue_pcmeshassociation_schematic} on the left).
\cite{Laupheimer2020_ISPRS} discuss in detail the particular challenges for the \ac{PCMA} using a 2.5D mesh.
The set of association prisms does not enclose all points of the \ac{PC}.
Particularly, points above and below the mesh surface may fall into \textit{dead zones} not covered by the association prisms.
The prisms of adjacent faces diverge when they form convex or concave surfaces, i.e. their normal vectors are not parallel.
Non-perfect reconstructions of planar surfaces artificially introduce convex or concave structures (cf.~\autoref{tab:points_unassociated}, B1).
Naturally, points above truly convex surfaces or below concave surfaces cannot be linked.
Typically, points above the reconstructed roof ridge cannot be associated (cf.~\autoref{tab:points_unassociated}, B2).
For this reason, co-registration discrepancies increase the number of non-associations.
The previously described missed associations are owed to the nature of the problem itself and the implementation aiming for a good trade-off of memory and speed (by adaptive thresholding).
Besides, we declare points that are projected on the face edges or vertices to be \textit{out-of-face points} (cf.~\autoref{tab:points_unassociated}, C).
By these means, we avoid their linking by choice.
Technically, these points belong to two adjacent faces~A and~B.
Hence, it is hard to decide whether to assign them to face~A or its adjacent face~B.
Therefore, such points cannot be linked unambiguously and may cause ambiguity in the information transfer.
Here, our reasoning is to link not all, but unambiguous points.
As a side-effect, neglecting these points accelerates the association process.
However, if desired, our implementation allows us to link these points, too \citep{Laupheimer2020_ISPRS}.
To the best of our knowledge, meshing algorithms depend fully on geometry and do not incorporate semantics.
Therefore, reconstructed faces do not necessarily represent semantic borders.
For instance, consider the transition of a planar \textit{impervious surface} to \textit{green space} in the real world.
The mesh representation may simplify this scenario to a single large face.
On the contrary, the same scenario is captured properly in the \ac{PC} and imagery.
Consequently, too large triangles at class borders will associate points and pixels of different classes (cf.~\autoref{fig:inconsistency_overview}).
For this reason, semantically incorrect associations are unavoidable due to the meshing.
Naturally, the \ac{PCImgA} inherits the limitations of its sub pipelines.
Besides, there are specific issues regarding the \ac{PC} visibility as derived via the mesh as a proxy.
Faces are marked as visible once a pixel is associated with the face.
Points, in turn, are marked as visible if they are associated with a visible face.
However, a face marked as visible does not have to be entirely visible.
Moreover, a face marked as visible does not have to be associated with points that are truly visible as well.
The adaptive thresholding links points and faces along the normal directions whereas the line-of-sight is relevant for the point visibility.
In other words, taken individually, the made associations by the sub pipelines are correct, but their composition does not guarantee a correct visibility check for each point-pixel relationship (using explicit \ac{PCImgA}).
On the contrary, the implicit \ac{PCImgA} makes use of the correct point-face and face-pixel linking.
Rephrased, the implicit linking overcomes georeference issues utilizing the adaptive thresholding and uses the correct visibility checks for faces.
\autoref{fig:association_pc_img_visibility} depicts point visibility and showcases apparently visible points.
It is unlikely that all truly non-visible points that are marked as visible are occluded by truly visible points.
For this reason, depth filtering helps only when truly non-visible and truly visible points are on the same projection ray.
Otherwise, truly non-visible points are linked with pixels through collinearity equations causing incorrect information transfer (for the explicit linking of the \ac{PCImgA}).
Generally, smaller faces better represent the geometry and reduce the impact of truly non-visible points.
However, large triangles are beneficial concerning processing time.
\begin{figure}[htbp]
\centering
\begin{tabular}{c c}
\includegraphics[width=0.425\columnwidth]{_figures/Figure9_left.eps} \hspace{0.7cm} &
\hspace{0.7cm} \includegraphics[width=0.2\columnwidth]{_figures/Figure9_right.eps} \\
(a) \hspace{0.7cm} & \hspace{0.7cm} (b)
\put (-245.5, 13.5) {\small opaque}
\put (-186.5, 13.5) {\small transparent}
\end{tabular}
\caption{
The sketches indicate truly (\textit{green}) and apparently visible points (\textit{red}) as detected by \ac{PCImgA} utilizing the mesh (\textit{black wireframe}) as a proxy.
(a) showcases an apparently visible point due to being linked to a truly but non-fully visible face (\textit{blue}).
Fully visible faces are depicted in \textit{green}.
The dashed cubes on the right mimic the situation of the opaque cubes in "transparent" mode.
(b) showcases apparently visible points due to the divergence of normal direction and line-of-sight.
Points are associated along the normal direction of faces. Hence, visible faces might be linked to truly non-visible points.
}
\label{fig:association_pc_img_visibility}
\end{figure}
\section{Results and Analysis}
\label{sec:results_discussion}
We demonstrate the capability of our methodology, its flexibility, and adaptiveness to underlying data by deploying \ac{V3D} and \ac{H3D} (cf.~\autoref{sec:data}).
To quantitatively analyze the linking methodology, \ac{GT} data is necessary for each modality.
However, to the best of our knowledge, there is no real-world data set that provides annotations for \ac{PC}, mesh, and imagery at the same time (cf.~\autoref{subsec:relatedwork_gt}).
For this reason, we qualitatively verify the effectiveness of the explicit entity linking visualizing the label transfer.
We visualize the achieved explicit entity linking by transferring labels from the manually annotated \acp{PC} to the mesh, and therefrom, to image space.
However, we want to emphasize that the transferred information is not limited to labels.
Features can be transferred to other modalities, too.
\autoref{fig:juggler_alternative} exemplarily shows the annotated modalities for a dedicated tile from \ac{H3D} as achieved by the proposed methodology.
\autoref{fig:results_imagery} shows a selection of automatically annotated images of various \acp{GSD} as derived via the implicit \ac{PCImgA}.
We opt for the implicit linking to create densely labeled images since faces are projected to image space instead of single points.
Furthermore, for \ac{V3D}, implicit \ac{PCImgA} makes use of the perfect co-registration of the \ac{MVS} mesh and imagery.
At the same time, the enclosed \ac{PCMA} is able to dampen the co-registration discrepancy in 3D space by leveraging the adaptive thresholding.
Hence, the implicit linking avoids inconsistent point-pixel-pairs.
Both figures visually verify that the transfer operates reasonably and smoothly on both data sets featuring different scales and resolutions.
Besides, \autoref{fig:results_imagery} reveals the dependence of synchronous acquisition and mesh quality.
The picture on the center-left shows different positions of a car due to asynchronous capturing of nadir imagery and \ac{ALS} data (cf.~\autoref{sec:data}).
The picture on the center-right depicts a ship (class \textit{vehicle}) in the ship lock surrounded by water.
Since the mesh does not properly reconstruct the ship, the transferred labels do not cover the entire ship in image space.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.95\linewidth]{_figures/Figure10.jpg}
\caption{
Automatically annotated images with labels as transferred from the \ac{PC} to the mesh and, therefrom, to image space (implicit \ac{PCImgA}) for \ac{H3D} (\textit{top}: oblique, \textit{center}: nadir) and \ac{V3D} (\textit{bottom}).
Non-associated pixels are depicted with RGB values. Original RGB images are shown on the left except for the nadir image of \ac{V3D} which covers the entire labeled area.
}
\label{fig:results_imagery}
\end{figure}
Despite the absence of \ac{GT} for all modalities, we try to quantify the entity linking by a proxy analysis: forward and backward passing of labels.
The relationship of image space and 3D space is well-known and strictly defined by collinearity equations and hence, does not have to be validated.
For this reason, we highlight the label transfer from the \ac{PC} to the mesh (forward pass) and therefrom back to the \ac{PC} (backward pass).
During forward passing, we aggregate labels on the face-level via majority vote.
The backward pass is a straightforward copy operation utilizing the stored association information.
The comparison of back-transferred and manual annotations allows us to validate the effectiveness of~\ac{PCMA} (label consistency check).
For the used data sets,
we found in an empirical process the association to perform best with thresholds~$\theta_1=\pm\SI{30}{cm}$, $\theta_2=\pm\SI{60}{cm}$, $\theta_3=\pm\SI{120}{cm}$ (\ac{V3D}) and~$\theta_1=\pm\SI{5}{cm}$, $\theta_2=\pm\SI{10}{cm}$, $\theta_3=\pm\SI{15}{cm}$~(\ac{H3D}).
The chosen thresholds are guided by the shift between mesh and \ac{ALS} data (cf.~\autoref{sec:data}).
In particular, thresholds are fine-tuned to maximize the number of associations while keeping mismatches of back-transferred and manual labels at a minimum.
\autoref{fig:inconsistency_fringe} shows the fringe of the \ac{MVS} mesh (\textit{bottom}) and the respective \ac{ALS} \ac{PC} (\textit{top}) for \ac{V3D}.
The height-coding indicates reconstruction errors in the leftmost part of the \ac{MVS} mesh:
the building and tree are not reconstructed.
The adaptive thresholding guarantees a correct linking of faces and points where geometry is reconstructed correctly.
Likewise, false reconstructed faces remain unlabeled and hence avoid label inconsistencies (after the backward pass).
\begin{figure}[htbp]
\centering
\begin{tabular}{cc}
\includegraphics[width=0.45\columnwidth]{_figures/Figure11_top_left.png}
\hspace{-0.215cm} & \hspace{-0.215cm}
\includegraphics[width=0.45\columnwidth]{_figures/Figure11_top_right.png}
\\
\includegraphics[width=0.45\columnwidth]{_figures/Figure11_bottom_left.png}
\hspace{-0.215cm} & \hspace{-0.215cm}
\includegraphics[width=0.45\columnwidth]{_figures/Figure11_bottom_right.png}
\end{tabular}
\caption{Fringe area of \ac{V3D} represented by the \ac{PC} (\textit{top}) and the \ac{MVS} mesh (\textit{bottom}). The left column shows 3D data in height-coded fashion (\textit{blue}: low, \textit{red}: high). The top right shows the manually annotated \ac{PC}; the bottom right shows the automatically annotated mesh as wireframe. Faces that do not map the real geometry are not linked to the \ac{PC}, and hence, remain unlabeled (cf. holes in the leftmost part for tree and building).}
\label{fig:inconsistency_fringe}
\end{figure}
For \ac{V3D}, the adaptive thresholding associates 40.9\% of faces covering 53.8\% of the surface area with 75.6\% of LiDAR points.
The proxy analysis reveals that 98.9\% of associated points show consistency in manual and back-transferred labels.
2.0\% of associated faces are linked to points of different classes causing label inconsistencies for 1.1\% of associated points.
The achieved weighted average precision of the label consistency check is 98.9\%.
For \ac{H3D}, the adaptive thresholding associates 67.3\% of faces covering 71.2\% of the surface area with 55.9\% of LiDAR points.
99.6\% of points pass the label consistency check.
Vice versa, 0.9\% of associated faces are linked to points of different classes.
The achieved weighted average precision is 99.9\%.
Structural differences among 3D modalities prevent full association of points and faces for both data sets (cf.~\autoref{fig:discrepancies_interrepresentation} and facades in~\autoref{fig:inconsistency_fringe}).
Particularly semi-transparent objects reduce the association rates on the point-level since sub-surface LiDAR points are not linked to the mesh surface.
The majority of non-associated points belong to vegetational classes (\ac{V3D}: 68\%, \ac{H3D}: 74\%).
In this regard, the high LiDAR density of \ac{H3D} causes a low association rate on the point-level.
In contrast, the comparatively high association rate on the face-level indicates proper co-registration and high-quality mesh.
The majority of non-associated faces build the water surface where no LiDAR points are captured.
For \ac{V3D}, 15.5\% of non-linked points belong to \textit{facade} and \textit{roof} pointing out the minor \ac{MVS} mesh quality.
To summarize, the association rates indicate the impact of the mesh quality and the point density.
The proxy analysis reveals that the majority of established point-face connections is correct for both data sets (98.9\%/99.6\%).
Likewise, the forward-backward-pass shows that common meshing does not incorporate semantic borders.
Hence, faces may be linked to points of different classes. In this case, the back-transferred majority vote causes inconsistencies.
\autoref{fig:inconsistency_overview} shows inconsistently labeled points for the entire \ac{H3D} data set and as close-up.
The overview at the top exhibits 0.4\% of points failing the label consistency check.
These points represent semantic borders and indicate an improper mesh reconstruction.
We are aware of the fact that co-registration discrepancies in 3D space increase this effect.
However, due to the high mesh quality and small co-registration discrepancy, \ac{H3D} is less affected than \ac{V3D}.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.8\columnwidth]{_figures/Figure12_top.png}
\vspace{0.1cm}
\begin{tabular}{cc}
\includegraphics[width=0.22\columnwidth]{_figures/Figure12_bottom_left.png}
\put (-56.25, -7.5) {\footnotesize \ac{GT} labels}
&
\includegraphics[width=0.22\columnwidth]{_figures/Figure12_bottom_right.png}
\put (-80.25, -7.5) {\footnotesize back-transferred labels}
\end{tabular}
\caption{Overview of \ac{H3D} points that show inconsistencies in manual annotations and back-transferred labels (\textit{top}).
The close-up at the bottom depicts inconsistently labeled points marked by the manually annotated \ac{GT} on the textured mesh (\textit{left}) and the back-transferred labels on the wireframe (\textit{right}).
}
\label{fig:inconsistency_overview}
\end{figure}
Furthermore, the proxy analysis helps to detect label noise.
\autoref{fig:inconsistency_building} depicts a building from V3D as \ac{PC} (\textit{left}) and wireframe mesh (\textit{right}).
A few of the transferred labels to the mesh (\textit{lower right}) seem not to match the given \ac{GT} on the \ac{PC} (\textit{upper left}).
For instance, some faces on the facade are marked as a roof.
Consequently, the back-transferred labels to the \ac{PC} (\textit{lower left}) do not match the initial annotation.
Here, the appearance of false labels hints at label noise, since the inconsistencies cannot be explained by georeference issues.
\autoref{fig:inconsistency_building_label_noise} shows the \ac{GT} along with its class-wise \ac{GT} for classes \textit{facade} and \textit{roof}.
The figure discloses that few points erroneously carry labels of both classes.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.85\columnwidth]{_figures/Figure13.png}
\caption{\ac{PC} (\textit{left}) and mesh representation (\textit{right}) of a building in \ac{V3D}.
The top row shows the manually annotated \ac{GT} and the textured mesh overlaid with its wireframe.
The bottom row shows the automatically labeled mesh (PC~$\mapsto$~Mesh) and the respectively labeled \ac{PC} with back-transferred labels from the mesh (Mesh~$\mapsto$~PC).}
\label{fig:inconsistency_building}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=0.55\columnwidth]{_figures/Figure14.png}
\caption{Label noise in form of duplicates in \ac{V3D} for building of \autoref{fig:inconsistency_building}. The right side shows \ac{GT} separated by class roof (\textit{top}) and facade (\textit{bottom})
The label noise causes wrong label transfer and thus inconsistencies in the forward-backward-pass.
}
\label{fig:inconsistency_building_label_noise}
\end{figure}
\section{Conclusion and Future Work}
\label{sec:futurework}
To jointly leverage imagery and \ac{ALS} data for semantic scene analysis, we propose a novel holistic methodology that explicitly integrates imagery and \ac{PC} data via the mesh as the core representation.
The multi-modal data fusion establishes explicit connections of points, faces, and pixels and enables the subsequent sharing of arbitrary information across modalities.
The information transfer incorporates the established one-to-many relationships by aggregation.
Hence, representation-specific features and (manual) annotations can be shared at a stroke across all modalities (cf.~\autoref{tab:association_options}).
Therefore, the proposed association mechanism can be seen as an integrator that functions as a labeling tool and a feature sharing tool.
By these means, the novel method serves as a powerful integrative backbone boosting multi-modal learning.
In particular, the method underlines its utility for pixel-wise \ac{GT} generation.
Any labeled 3D data (\ac{PC} or mesh) can be projected into image space to annotate multiple images at once while performing the visibility check.
Hence, it minimizes the manual labeling effort.
Consequently, the versatile applicability of the information transfer fosters modality-specific and multi-modal semantic segmentation.
The linking mechanism is designed to surpass imperfections of real-world data by adaptive thresholding.
The tile-wise parallel processing aims for a trade-off of memory and speed.
We qualitatively and quantitatively demonstrate its effectiveness and adaptiveness to underlying data by deploying two airborne data sets of different resolutions and scales.
\ac{V3D} is typical for large-scale country-wide mapping with moderate \ac{GSD} of some centimeters and a considerable time shift between \ac{ALS} and image data collection.
\ac{H3D} provides extremely high-resolution data with mainly synchronous data capture from a hybrid sensor system and is representative of data collection at small-scale complex built-up areas.
Due to the absence of multi-modal \ac{GT}, a strict quantitative analysis of the proposed method is difficult.
As an alternative, we analyze the label consistency on the \ac{PC} by forward-backward-passes of labels across entities of different modalities.
The quantitative analysis shows that nearly 100\% of the established connections are consistent.
However, points of different classes that are linked to a common face might be useful for a subsequent semantically driven remeshing.
Due to structural discrepancies, full association across entities is not possible.
We discuss preconditions and limitations in detail highlighting the benefits of high-quality co-registration and high-quality reconstruction.
The strength of our method is its simplicity and flexibility that immediately profits from advances in data acquisition, co-registration, and meshing.
In the future, we aim for pixel-accurate co-registration of \ac{ALS} data and imagery leveraging the hybrid strip adjustment \citep{glira2019}.
We claim that improved co-registration improves both geometric reconstruction and semantic analysis.
To prove our assumption, we plan an ablation study for multi-modal features on different representations by analyzing the performance of a \ac{ML} classifier.
\section{Acknowledgements}
\ac{V3D} is provided by the German Society for Photogrammetry, Remote Sensing and Geoinformation (DGPF) \citep{Cramer2010}.
\ac{H3D} data originates from a research project in collaboration with the German Federal Institute of Hydrology~(BfG) in~Koblenz.
We thank all our colleagues
for insightful discussions.
In particular, we thank our students Fangwen Shu, Mohamad Hakam Shams Eddin, and Vishal Pani, who assisted the implementation.
Furthermore, we thank the whole nFrames team for their support regarding the mesh generation.
Special thanks are directed to Carmen Kaspar for proofreading.
|
\section{Introduction} \label{introduction}
Recently, for the first time, the BESIII Collaboration has reported a structure $Z_{cs}(3985)^-$ in the $K^+$ recoil-mass spectrum
near the $D_s^-D^{*0}$ and $D_s^{*-}D^0$ mass thresholds in the process of $e^+e^- \rightarrow K^+(D_s^-D^{*0}+D_s^{*-}D^0)$ at the
center-of-mass energy $\sqrt{s}=4.681$ GeV, with the mass and narrow decay width~\cite{Ablikim:2020hsk},
\begin{eqnarray}
M_{Z_{cs}}&=& (3982_{-2.6}^{+1.8}\pm2.1)\quad\rm{MeV}, \nonumber\\
\Gamma_{Z_{cs}}&=&(12.8_{-4.4}^{+5.3}\pm3.0)\quad\rm{MeV},
\end{eqnarray}
and the significance was estimated to be 5.3$\sigma$. Soon afterwards, a significant state $Z_{cs}(4000)^+$, with a mass of
$4003\pm6^{+4}_{-14}$ MeV, a width of $131\pm15\pm26$ MeV, and spin-parity $J^P = 1^+$, was observed by LHCb Collaboration,
including another exotic state $Z_{cs}(4220)^+$ \cite{Aaij:2021ivw}. The discovery of the charged heavy quarkonium-like structures
with strangeness could shed light on the properties of the charged exotic $Z$ states reported before~\cite{Zcfamily}.
Besides $Z_{cs}$ states, two excited $B_s^0$ states are observed in the $B^+K^-$ mass spectrum in a sample of proton-proton collisions at
centre-of-mass energies of 7, 8, and 13 TeV very recently by LHCb Collaboration~\cite{Aaij:2020hcw}. The masses and widths of the two states
are determined to be
\begin{eqnarray}
M_1&=& 6063.5\pm1.2(\rm{stat})\pm0.8(\rm{syst})~\rm{MeV}, \nonumber\\
\Gamma_1&=&26\pm4(\rm{stat})\pm4(\rm{syst})~\rm{MeV}, \\
M_2&=& 6114.5\pm3(\rm{stat})\pm5(\rm{syst})~\rm{MeV}, \nonumber\\
\Gamma_2&=&66\pm18(\rm{stat})\pm21(\rm{syst})~\rm{MeV}.
\end{eqnarray}
For $Z_{cs}$, it is classified into the exotic state as the strange partner of $Z_c(3900)$ and has intensively attracted more attentions and
investigations theoretically within a very short time
\cite{Meng:2020ihj,Wang:2020htx,Azizi:2020zyq,Sungu:2020zvk,Wang:2020rcx,Chen:2020yvq,Yang:2020nrt,Jin:2020yjn,Ferretti:2020ewe,Wang:2020kej}.
These explanations basically cover various exotic hadron configurations. One feature of the $Z_{cs}$ is that its mass is on the verge of the
$\bar{D}_sD^*$ or $\bar{D}_s^*D$ threshold, so a molecular resonance is suggested. For example, Lu Meng $\emph{et al.}$ obtained the mass and
width of $Z_{cs}$ in good agreement with the experimental results by considering the coupled-channel effect and strongly supported the $Z_{cs}$
states as the $U/V$-spin partner states of the charged $Z_c(3900)$~\cite{Meng:2020ihj}. In chiral effective field theory up to the
next-to-leading order, $Z_{cs}$ also was regarded as the partner of the $Z_c(3900)$ in the SU(3) flavor symmetry and the
$\bar{D}_sD^*/D_s^*D$ molecular resonance~\cite{Wang:2020htx}. In the QCD sum rule, $Z_{cs}$ can be well defined as a diquark-antidiquark
candidate with quark content $\bar{c}cu\bar{s}$~\cite{Sungu:2020zvk}.
On the contrary, J. Ferretti pointed out that the meson-meson molecular model could not be used to describe heavy-light tetraquarks with
non-null strangeness content, and in the case of $cs\bar{c}\bar{n}~(n = u~{\rm or}~d$) configurations, the compact tetraquark ground-state
is about 200 MeV below the lowest energy hadro-charmonium state, $\eta_cK$~\cite{Ferretti:2020ewe}. Another explanation is that $Z_{cs}$
can be naturally regarded as a reflection structure from a charmed-strange meson $D_{s2}^*(2573)$ by Lanzhou group~\cite{Wang:2020kej}.
By adopting a one-boson-exchange model and considering the coupled channel effect, Ref.~\cite{Chen:2020yvq} excluded $Z_{cs}$ as a
$D^{*0}D_s^-/D^0D_s^{*-}/D^{*0}D_s^{*-}$ resonance. As more and more exotic states named as $XYZ$ have been observed in different experiments,
their structures are still inexplicable and controversial theoretically. Investigating for charged charmonium-like states can extend our
knowledge of hadrons and our understanding of the nature of strong interaction. So we believe that the study of the newly observed $Z_{cs}$
states in the chiral quark model can provide some useful information on exotic hadrons. In present work, for $Z_{cs}$, the quantum number is
assigned as $I(J^P)=\frac{1}{2}(1^+)$ and the quark composition is $c\bar{c}s\bar{u}$.
Now let's turn to the newly observed excited $B_s^0$ states. In the past few years, many experiment collaborations such as CDF, D0, and LHCb
have made contributions to find the radial and orbital excitations of the bottom and bottom-strange meson families. More and more higher
excitations emerged in experiments~\cite{Aaltonen:2013atp,Aaltonen:2007ah,Abazov:2007af,Aaij:2012uva}. Therewith many theoretical studies of
the bottom and bottom-strange mesons follow close on another~\cite{Sun:2014wea,Xiao:2014ura,Liu:2015lka,Godfrey:2016nwn,Lu:2016bbk,Ferretti:2015rsa}.
The mass spectrum and strong decay patterns are studied most in conventional 2-body quark-antiquark system, which can describe the ground states
very well, but has poor understanding for higher excitations of bottom and bottom-strange mesons. Studying the $B$ and $B_s$ mesons will help us
not only understand of excited mesons, but also put the discovered excited charm and chram-strange mesons into the larger context, since
the present situation of experimental exploration of bottom and bottom-strange states is very similar to that of $D$ and $D_s$ states in 2003
\cite{Aubert:2003fg,Besson:2003cp,Krokovny:2003zq,Evdokimov:2004iy,Aubert:2006mh,Aubert:2009ah,delAmoSanchez:2010vq}.
With newly observation of the excited $B_s^0$ states by LHCb Collaboration~\cite{Aaij:2020hcw}, now it is a good time to carry out a
comprehensive theoretical study on higher bottom-strange mesons. In this work, all possible quantum numbers with $I(J^P)=0(0^+), 0(1^+),
0(2^+)$ are studied for $B_s^0$ states. Considering the possible limitation of quark-antidiquark system in conventional quark model
in describing the higher excitations and the possible production of quark-antiquark pair in the vacuum, we obtain the masses of $B_s^0$ states
in 2-body quark-antidiquark system and 4-body $b\bar{s}q\bar{q}~(q=u~\rm{or}~d)$ system, respectively. $b\bar{s}s\bar{s}$ is not included here
because of its high energy.
Each tetraquark calculation takes into account the mixing of structures, such as meson-meson and diquark-antidiquark structure, along with
all possible color, spin configurations. In the meantime, in order to find possible stable resonance states, high precision computing method
Gaussian expansion method (GEM) \cite{GEM} and an useful stabilization real scaling method are both employed~\cite{rs1,rs2} in our calculations.
The paper is arranged as follows. Theoretical framework including the chiral quark model, the wave functions of $Z_{cs}$ and $B_s^0$,
along with GEM are introduced in Section \ref{ModleandGEM}. In Section \ref{results}, the numerical results and discussion are presented.
A short summary is given in Section \ref{epilogue}.
\section{Theoretical framework}
\label{ModleandGEM}
\emph{\textbf{Chiral quark model:}} the review of the chiral quark model and GEM has been introduced in
Refs.~\cite{Chen:2018hts,Chen:2017mug,Chen:2016npt}, and here they will be introduced briefly and we mainly focus on the relevant features of
$Z_{cs}$ and $B_s^0$ states.
The Hamiltonian of the chiral quark model can be written as follows for 4-body system,
\begin{align}
H & = \sum_{i=1}^4 m_i +\frac{p_{12}^2}{2\mu_{12}}+\frac{p_{34}^2}{2\mu_{34}}
+\frac{p_{1234}^2}{2\mu_{1234}} \quad \nonumber \\
& + \sum_{i<j=1}^4 \left[ V_{ij}^{C}+V_{ij}^{G}+\sum_{\chi=\pi,K,\eta} V_{ij}^{\chi}
+V_{ij}^{\sigma}\right].
\end{align}
The potential energy: $V_{ij}^{C, G, \chi, \sigma}$ represents the confinement, one-gluon-exchange, Goldston boson exchange and $\sigma$
exchange, respectively. The detailed forms can be referred to Eq. (13) in Ref.~\cite{Chen:2017mug}, which are omit here for space
saving. All the model parameters are determined by fitting the meson spectrum, from light to heavy; and the resulting values are listed in
Table~\ref{modelparameters}.
\begin{table}[!t]
\begin{center}
\caption{ \label{modelparameters}Model parameters, determined by fitting the meson spectrum.}
\begin{tabular}{llr}
\hline\noalign{\smallskip}
Quark masses &$m_u=m_d$ &313 \\
(MeV) &$m_s$ &536 \\
&$m_c$ &1728 \\
&$m_b$ &5112 \\
\hline
Goldstone bosons &$m_{\pi}$ &0.70 \\
(fm$^{-1} \sim 200\,$MeV ) &$m_{\sigma}$ &3.42 \\
&$m_{\eta}$ &2.77 \\
&$m_{K}$ &2.51 \\
&$\Lambda_{\pi}=\Lambda_{\sigma}$ &4.2 \\
&$\Lambda_{\eta}=\Lambda_{K}$ &5.2 \\
\cline{2-3}
&$g_{ch}^2/(4\pi)$ &0.54 \\
&$\theta_p(^\circ)$ &-15 \\
\hline
Confinement &$a_c$ (MeV fm$^{-2}$) &101 \\
&$\Delta$ (MeV) &-78.3 \\
\hline
OGE & $\alpha_0$ &3.67 \\
&$\Lambda_0({\rm fm}^{-1})$ &0.033 \\
&$\mu_0$(MeV) &36.98 \\
&$s_0$(MeV) &28.17 \\
\hline
\end{tabular}
\end{center}
\end{table}
It is to be noted that, only $V^{\chi=\eta}$ of Goldston boson exchange plays a role between $u$ and $s$ quark, and for $u$ and $\bar{u}$
interacting quark-pair, not only $V^{\chi=\pi,\eta}$, but also $V^{\sigma}$ works; for other quark pairs such as, $(Q,u)$, $(Q,s)$, $(Q,Q)$
($Q=c,b$), only $V_{ij}^{C, G}$ is considered without Goldstone bosons and $\sigma$ exchange.
\emph{\textbf{Wave functions:}} There are three quark configurations for $Z_{cs}$ and $b\bar{s}q\bar{q}$ system , two meson-meosn structures and one diquark-antidiquark structure, which are shown in Fig.\ref{structure}.
For spin part, the wave functions for 2-body system are
\begin{align}
&\chi_{11}=\alpha\alpha,~~
\chi_{10}=\frac{1}{\sqrt{2}}(\alpha\beta+\beta\alpha),~~
\chi_{1-1}=\beta\beta,\nonumber \\
&\chi_{00}=\frac{1}{\sqrt{2}}(\alpha\beta-\beta\alpha),
\end{align}
then total six wave functions of 4-body system are obtained easily, which are shown in Table \ref{wavefunctions}(the first column). The subscripts $SM_S$ of $\chi$ represents the total spin and the third projection of total spin of four-quark system, with $S=0,1,2$, and only one component ($M_S=S$) is shown for a given total spin $S$.
For flavor part, the flavor wave functions (two meson-meson structures plus one diquark-antidiquark structure) are also tabulated in Table \ref{wavefunctions}(the second column). The wave functions $\chi^{f1}$, $\chi^{f2}$, $\chi^{f3}$ correspond to the picture $(a),(b),(c)$ in Fig \ref{structure} for $Z_{cs}$, respectively, and the $\chi^{f4}$, $\chi^{f5}$, $\chi^{f6}$ are the wave functions of pictures $(a^{\prime}),(b^{\prime}),(c^{\prime})$ for $b\bar{s}q\bar{q}$ system.
For color part, there are four wave functions in total(the third column in Table \ref{wavefunctions}), $\chi^{c1}$ color singlet-singlet $(1\otimes1)$ and $\chi^{c2}$
color octet-octet $(8\otimes8)$ for meson-meson structure and $\chi^{c3}$ color antitriplet-triplet $(\bar{3}\otimes3)$ and $\chi^{c4}$
sextet-antisextet $(6\otimes\bar{6})$ for diquark-antidiquark structure.
\begin{figure}
\resizebox{0.50\textwidth}{!}{\includegraphics{structure.eps}}
\caption{\label{structure} Structures of $Z_{cs}$ and $b\bar{s}q\bar{q}$ system, $(a)/(a^{\prime})$ and $(b)/(b^{\prime})$ represent
two meson-meson structures, and $(c)/(c^{\prime})$ represents diquark-antidiquark structure.}
\end{figure}
\linespread{1.5}
\begin{table*}[!t]
\begin{center}
\caption{ \label{wavefunctions}The wave functions of spin, flavor, color part for $Z_{cs}$ and $b\bar{s}q\bar{q}$ system by considering all
kinds of quark structures.}
\begin{tabular}{lll}
\hline\hline\noalign{\smallskip}
\quad \quad \quad Spin & Flavor &\quad\quad \quad \quad \quad \quad \quad Color \\
\hline
$\chi_{00}^{\sigma1}=\chi_{00}\chi_{00}$
&$\chi^{f1}=c\bar{u}s\bar{c}$
&\quad \quad \quad$\chi^{c1}=\frac{1}{3}(\bar{r}r+\bar{g}g+\bar{b}b)(\bar{r}r+\bar{g}g+\bar{b}b)$ \\
$\chi_{00}^{\sigma2}=\sqrt{\frac{1}{3}}(\chi_{11}\chi_{1-1}-\chi_{10}\chi_{10}+\chi_{1-1}\chi_{11})$
&$\chi^{f2}=c\bar{c}s\bar{u}$
&\quad \quad \quad\makecell[l]{$\chi^{c2}=\frac{\sqrt{2}}{12}(3\bar{b}r\bar{r}b+3\bar{g}r\bar{r}g+3\bar{b}g\bar{g}b+3\bar{g}b\bar{b}g+3\bar{r}g\bar{g}r$\\
$\quad \quad \quad +3\bar{r}b\bar{b}r+2\bar{r}r\bar{r}r+2\bar{g}g\bar{g}g+2\bar{b}b\bar{b}b-\bar{r}r\bar{g}g$ \\
$\quad \quad \quad -\bar{g}g\bar{r}r-\bar{b}b\bar{g}g-\bar{b}b\bar{r}r-\bar{g}g\bar{b}b-\bar{r}r\bar{b}b)$}\\
$\chi_{11}^{\sigma3}=\chi_{00}\chi_{11}$
&$\chi^{f3}=cs\bar{u}\bar{c}$
&\quad \quad \quad\makecell[l]{$\chi^{c3}=\frac{\sqrt{3}}{6}(rg\bar{r}\bar{g}-rg\bar{g}\bar{r}+gr\bar{g}\bar{r}-gr\bar{r}\bar{g}$ \\
$\quad \quad \quad+rb\bar{r}\bar{b}-rb\bar{b}\bar{r}+br\bar{b}\bar{r}-br\bar{r}\bar{b}$ \\
$\quad \quad \quad+gb\bar{g}\bar{b}-gb\bar{b}\bar{g}+bg\bar{b}\bar{g}-bg\bar{g}\bar{b})$} \\
$\chi_{11}^{\sigma4}=\chi_{11}\chi_{00}$
&$\chi^{f4}=\frac{1}{2}(b\bar{d}d\bar{s}+b\bar{u}u\bar{s})$
&\quad \quad \quad\makecell[l]{$\chi^{c4}=\frac{\sqrt{6}}{12}(2rr\bar{r}\bar{r}+2gg\bar{g}\bar{g}+2bb\bar{b}\bar{b}+rg\bar{r}\bar{g}+rg\bar{g}\bar{r}$\\
$\quad \quad \quad+gr\bar{g}\bar{r}+gr\bar{r}\bar{g}+rb\bar{r}\bar{b}+rb\bar{b}\bar{r}+br\bar{b}\bar{r}$ \\
$\quad \quad \quad+br\bar{r}\bar{b}+gb\bar{g}\bar{b}+gb\bar{b}\bar{g}+bg\bar{b}\bar{g}+bg\bar{g}\bar{b})$}\\
$\chi_{11}^{\sigma5}=\frac{1}{\sqrt{2}}(\chi_{11}\chi_{10}-\chi_{10}\chi_{11})$
&$\chi^{f5}=-\frac{1}{2}(b\bar{s}u\bar{u}+b\bar{s}d\bar{d})$
&\makecell[l]{\quad\\\quad} \\
$\chi_{22}^{\sigma6}=\chi_{11}\chi_{11}$
&$\chi^{f6}=-\frac{1}{2}(bu\bar{u}\bar{s}+bd\bar{d}\bar{s})$
& \\
\hline\hline
\end{tabular}
\end{center}
\end{table*}
\linespread{1.5}
\begin{table*}[!t]
\begin{center}
\renewcommand\tabcolsep{6.0pt}
\caption{ \label{channels} Allowed channels for $Z_{cs}$ and $b\bar{s}q\bar{q}$ system, for saving context space, we give abbreviations for channels, \emph{e.g.}, for meson-meson structure (picture $(a)$) of $Z_{cs}$, $'3 1 1'$ and $'3 1 2'$ represents $\chi_{11}^{\sigma3}\chi^{f1}\chi^{c1}$ and $\chi_{11}^{\sigma3}\chi^{f1}\chi^{c2}$, severally. And $'4 1 1'$ and $'4 1 2'$ represents $\chi_{11}^{\sigma4}\chi^{f1}\chi^{c1}$ and $\chi_{11}^{\sigma4}\chi^{f1}\chi^{c2}$, respectively. The rest channels can be read in the same manner. The last row gives the total numbers of channels by considering meson-meson structures, diquark-antidiquark structure, along with all kinds of color spin configurations for $Z_{cs}$ and $b\bar{s}q\bar{q}$ system with quantum numbers $0(0^+)$, $0(1^+)$, $0(2^+)$.}
\begin{tabular}{ccccccccccccc}
\hline\hline\noalign{\smallskip}
system &\multicolumn{3}{c}{$Z_{cs}$} &\multicolumn{9}{c}{$b\bar{s}q\bar{q}$} \\
\hline
$I(J^P)$
&\multicolumn{3}{c}{$\frac{1}{2}(1^+)$} &\multicolumn{3}{c}{$0(0^+)$} &\multicolumn{3}{c}{$0(1^+)$} &\multicolumn{3}{c}{$0(2^+)$} \\
Structure
&$(a)$ &$(b)$ &$(c)$ &$(a^{'})$ &$(b^{'})$ &$(c^{'})$ &$(a^{'})$ &$(b^{'})$ &$(c^{'})$ &$(a^{'})$ &$(b^{'})$ &$(c^{'})$ \\
\hline
\makecell[c]{channel \\ (spin $\cdot$ flavor $\cdot$ color)}
&\makecell[c]{3 1 1\\3 1 2\\4 1 1\\4 1 2\\5 1 1\\5 1 2}
&\makecell[c]{3 2 1\\3 2 2\\4 2 1\\4 2 2\\5 2 1\\5 2 2}
&\makecell[c]{3 3 3\\3 3 4\\4 3 3\\4 3 4\\5 3 3\\5 3 4}
&\makecell[c]{1 4 1\\1 4 2\\2 4 1\\2 4 2}
&\makecell[c]{1 5 1\\1 5 2\\2 5 1\\2 5 2}
&\makecell[c]{1 6 3\\1 6 4\\2 6 3\\2 6 4}
&\makecell[c]{3 4 1\\3 4 2\\4 4 1\\4 4 2\\5 4 1\\5 4 2}
&\makecell[c]{3 5 1\\3 5 2\\4 5 1\\4 5 2\\5 5 1\\5 5 2}
&\makecell[c]{3 6 3\\3 6 4\\4 6 3\\4 6 4\\5 6 3\\5 6 4}
&\makecell[c]{6 4 1\\ 6 4 2}
&\makecell[c]{6 5 1\\ 6 5 2}
&\makecell[c]{6 6 3\\ 6 6 4} \\
\hline
number of channels&6 &6 &6 &4 &4 &4 &6 &6 &6 &2 &2 &2\\
&\multicolumn{3}{c}{total 18}
&\multicolumn{3}{c}{total 12}
&\multicolumn{3}{c}{total 18}
&\multicolumn{3}{c}{total 6} \\
\hline\hline
\end{tabular}
\end{center}
\end{table*}
So we can get all allowed spin $\otimes$ flavor $\otimes$ color channels of $Z_{cs}$ and $b\bar{s}q\bar{q}$ system by taking meson-meson structures, diquark-antidiquark structure, along with all kinds of color spin configurations into account, which are shown in Table \ref{channels}.
Next, let's discuss the orbital wave functions for 4-body system. They can be obtained by coupling the orbital wave function for each
relative motion of the system,
\begin{equation}\label{spatialwavefunctions}
\Psi_{L}^{M_{L}}=\left[[\Psi_{l_1}({\bf r}_{12})\Psi_{l_2}({\bf
r}_{34})]_{l_{12}}\Psi_{L_r}({\bf r}_{1234}) \right]_{L}^{M_{L}},
\end{equation}
where $l_1$ and $l_2$ is the angular momentum of two sub-clusters, respectively. $\Psi_{L_r}(\mathbf{r}_{1234})$ is the wave function of
the relative motion between two sub-clusters with orbital angular momentum $L_r$. $L$ is the total orbital angular momentum of four-quark states.
Because of the positive parity ($P=(-1)^{l_1+l_2+L_r}$=+) for $Z_{cs}$ and $b\bar{s}q\bar{q}$, it is natural to assume that all the orbital angular
momenta are zeros. With the help of Gaussian expansion method (GEM), the spatial wave functions are expanded in series of Gaussian basis functions.
\begin{subequations}
\label{radialpart}
\begin{align}
\Psi_{l}^{m}(\mathbf{r}) & = \sum_{n=1}^{n_{\rm max}} c_{n}\psi^G_{nlm}(\mathbf{r}),\\
\psi^G_{nlm}(\mathbf{r}) & = N_{nl}r^{l}
e^{-\nu_{n}r^2}Y_{lm}(\hat{\mathbf{r}}),
\end{align}
\end{subequations}
where $N_{nl}$ are normalization constants,
\begin{align}
N_{nl}=\left[\frac{2^{l+2}(2\nu_{n})^{l+\frac{3}{2}}}{\sqrt{\pi}(2l+1)}
\right]^\frac{1}{2}.
\end{align}
$c_n$ are the variational parameters, which are determined dynamically. The Gaussian size
parameters are chosen according to the following geometric progression
\begin{equation}\label{gaussiansize}
\nu_{n}=\frac{1}{r^2_n}, \quad r_n=r_1a^{n-1}, \quad
a=\left(\frac{r_{n_{\rm max}}}{r_1}\right)^{\frac{1}{n_{\rm
max}-1}}.
\end{equation}
This procedure enables optimization of the expansion using just a small numbers of Gaussians. Finally, the complete channel wave function
$\Psi^{\,M_IM_J}_{IJ}$ for four-quark system is obtained by coupling the orbital and spin, flavor, color wave functions get in
Table \ref{channels}. At last, the eigenvalues of four-quark system are obtained by solving the Schr\"{o}dinger equation
\begin{equation}
H \, \Psi^{\,M_IM_J}_{IJ}=E^{IJ} \Psi^{\,M_IM_J}_{IJ}.
\end{equation}
To obtain stable results in our work, the Gaussian width and Gaussian number of each inner cluster takes, $r_1=0.1~\rm{fm}$,
$r_n=2~ \rm{fm}$, $n=12$. For the relative motion between two sub-clusters $r_1=0.1~\rm{fm}$, $r_n=6~\rm{fm}$, $n=7$.
\section{Calculations and analysis}
\label{results}
In the present work, we calculated the mass spectrum of newly observed $Z_{cs}$ and excited $B_s^0$ states in the chiral quark model.
For the excited $B_s^0$ states, we firstly treat them as ordinary quark-antiquark states. Using the model parameters given in
Table \ref{modelparameters}, the convergent results of $b\bar{s}$ mass spectrum up to the second $D$-wave states in the chiral
quark model are obtained and shown in Table \ref{bs_results}, where the experimental data are also listed for comparison. Until now,
there has been only very limited experimental values on the low-lying bottom-strange mesons, which are called $B_s^0(5366)$,
$B_s^*(5415)$, $B_{s1}(5830)^0$, $B_{s2}^*(5840)^0$ \cite{PDG}. In Table \ref{bs_results},
we can see that the $2S$ and $1D$ states have masses between 6000 MeV and 6200 MeV, so it is possible that the newly observed excited $B_s^0$ are $2S$ or $1D$ states of $b\bar{s}$. However, for the excitation energy as high as 700 MeV, the excitation of the light quark-antiquark pair from vaccuum is highly favored. So considering the excited $B_s^0$ states as four-quark $b\bar{s}q\bar{q}~(q = u~\rm{or}~d )$ system is also necessary. In the following the four-quark $b\bar{s}q\bar{q}~(q = u~\rm{or}~d )$ system with quantum numbers $I(J^P)=0(0^+), 0(1^+), 0(2^+)$ is investigated.
\begin{table}[!t]
\begin{center}
\renewcommand\tabcolsep{6.0pt}
\caption{ \label{bs_results} The mass spectrum of $b\bar{s}$ meson families in the chiral quark model in comparison with reference \cite{Sun:2014wea} and experimental data \cite{PDG} (unit: MeV).}
\begin{tabular}{ccccc}
\hline\hline\noalign{\smallskip}
$n^{2S+1}L_J$ &This work &Ref. \cite{Sun:2014wea} &Expt \cite{PDG}\\
\hline
$1^1S_0$ &5367.4 &5390 &5366.84$\pm$0.15\\
$1^3S_1$ &5410.2 &5447 &5415.8$\pm$1.5\\
$2^1S_0$ &6017.3 &5985 &\\
$2^3S_1$ &6057.2 &6013 &\\
$1^3P_0$ &5749.2 &5830 &\\
$1^3P_1$ &5779.3 &5859 &5828.65$\pm$0.24 \\
$1^3P_2$ &5812.0 &5875 &5839.92$\pm$0.14 \\
$1^1P_1$ &5797.6 &5858 &\\
$2^3P_0$ &6345.9 &6279 &\\
$2^3P_1$ &6381.9 &6291 &\\
$2^3P_2$ &6422.9 &6295 &\\
$2^1P_1$ &6403.9 &6284 &\\
$1^3D_1$ &6179.3 &6181 &\\
$1^3D_2$ &6145.3 &6185 &\\
$1^3D_3$ &6094.2 &6178 &\\
$1^1D_2$ &6128.2 &6180 &\\
$2^3D_1$ &6778.1 &6542 &\\
$2^3D_2$ &6743.9 &6542 &\\
$2^3D_3$ &6692.9 &6534 &\\
$2^1D_2$ &6726.8 &6536 &\\
\hline\hline
\end{tabular}
\end{center}
\end{table}
For $Z_{cs}$, the minimal quark component should be $c\bar{c}s\bar{u}$ rather than a pure $c\bar{c}$ since it is observed as a charged particle
with strangeness. Both of $Z_{cs}$ and $b\bar{s}q\bar{q}$ states have two kinds of meson-meson structures and one diquark-antidiquark structure,
which are shown in Fig. \ref{structure}. Along with all possible color and spin configurations, we take all kinds of structures into account.
Table \ref{mesonmass} gives the masses of some relevant quark-antiquark mesons in present work in the chiral quark model.
From the table, we can see that the chiral quark model is very successful in describing the meson spectra. And the mass spectra of
$Z_{cs}$ and $b\bar{s}q\bar{q}$ system are demonstrated in Table \ref{4q-results}.
\begin{table}[!t]
\begin{center}
\caption{ \label{mesonmass}The masses of some relevant mesons in present work in the chiral quark model, compared with the experimental data (unit: MeV).}
\begin{tabular}{cccccccc}
\hline\hline\noalign{\smallskip}
Meson &$D^0$ &$D^{*0}$ &$D_s^*$ &$D_s^-$ &$\eta_c$ &$J/\psi$ &$K^{\pm}$ \\
\hline
$M_{cal}$ &1862.4&1980.5 &2079.9 &1952.6 &3102.2 &3161.2 &493.9 \\
$M_{exp}$ &1864.8&2006.9 &2112.2 &1968.3 &2983.6 &3096.9 &493.6\\
\hline
Meson &$B^-$ &$B^{*-}$ &$\bar{B_s}^0$ &$\bar{B_s}^*$ &$\omega$ &$\eta$ &$K^{*\pm}$\\
\hline
$M_{cal}$ &5280.9 &5319.6 &5367.9 &5410.2 &701.5 &669.2 &913.6\\
$M_{exp}$ &5279.3 &5325.2 &5366.7 &5415.4 &782.6 &547.8 &891.6\\
\hline\hline
\end{tabular}
\end{center}
\end{table}
\begin{table*}[!t]
\begin{center}
\renewcommand\tabcolsep{5.5pt}
\caption{ \label{4q-results}The mass spectrum of $Z_{cs}$ and $b\bar{s}q\bar{q}$ system with allowed quantum numbers. $E(a)$, $E(b)$, $E(c)$ represents the energies in pure meson-meson structures and pure diquark-antidiquark structure for $Z_{cs}$, corresponding to figure $(a)$, $(b)$, and $(c)$ in Fig. \ref{structure}. $E(a)\otimes E(c)$, $E(b)\otimes E(c)$, $E(a)\otimes E(b)$ are the energies considering the mixture of one meson-meson structure and one diquark-antidiquark structure, or two meson-meson structures, severally. $E_{cc}$ represents ground state energy for each state after considering the coupling of all
possible quark structures, color and spin channels (refer to Table \ref{channels}). It is the same with $b\bar{s}q\bar{q}$. The last column gives the theoretical lowest thresholds (unit: MeV).}
\begin{tabular}{cccccccccc}
\hline\hline\noalign{\smallskip}
& &$E(a)$ &$E(b)$ &$E(c)$ &$E(a)\otimes E(c)$ &$E(b)\otimes E(c)$ &$E(a)\otimes E(b)$ &$E_{cc}$ &the lowest thresholds \\
\hline
$Z_{cs}$ &$\frac{1}{2}(1^+)$
&3934.5 &3656.3 &4247.1 &3934.5 &3655.4 &3934.5 &3656.2 &$3655.1(J/\psi K^-)$\\
& &$E(a^{\prime})$ &$E(b^{\prime})$ &$E(c^{\prime})$ &$E(a^{\prime})\otimes E(c^{\prime})$ &$E(b^{\prime})\otimes E(c^{\prime})$ &$E(a^{\prime})\otimes E(b^{\prime})$ &$E_{cc}$ &the lowest thresholds \\
$b\bar{s}q\bar{q}$ &$0(0^+)$ &5776.6 &6040.2 &6283.2 &5776.5&6040.2&5776.6 &5775.3 &$5774.8(B^-K^+)$\\
&$0(1^+)$ &5815.4 &6072.6 &6316.9 &5815.3&6072.6&5815.4 &5814.1 &$5813.5(B^{*-}K^+)$ \\
&$0(2^+)$ &6234.8 &6114.9 &6483.7 &6234.7&6114.9&6234.8 &6114.3 &$6111.7(\bar{B_s}^*\omega)$\\
\hline\hline
\end{tabular}
\end{center}
\end{table*}
\begin{figure}
\center{\includegraphics[width=7.0cm]{avoidcrossing.eps}}
\caption{\label{avoid-crossing} Stabilization graph for the resonance.}
\end{figure}
\begin{figure}
\resizebox{0.50\textwidth}{!}{\includegraphics{Zcs1.eps}}
\caption{\label{Zcs1} The stabilization plots of the energies (3600 MeV $\sim$ 3900 MeV) of $c\bar{c}s\bar{u}$ states for $I(J^P)=\frac{1}{2}(1^+)$ with respect to the scaling factor $\alpha$.}
\end{figure}
\begin{figure}
\resizebox{0.50\textwidth}{!}{\includegraphics{Zcs2.eps}}
\caption{\label{Zcs2} The stabilization plots of the energies (3900 MeV $\sim$ 4000 MeV) of $c\bar{c}s\bar{u}$ states for $I(J^P)=\frac{1}{2}(1^+)$ with respect to the scaling factor $\alpha$.}
\end{figure}
\begin{figure}
\resizebox{0.50\textwidth}{!}{\includegraphics{Zcs3.eps}}
\caption{\label{Zcs3} The stabilization plots of the energies (4000 MeV $\sim$ 4100 MeV) of $c\bar{c}s\bar{u}$ states for $I(J^P)=\frac{1}{2}(1^+)$ with respect to the scaling factor $\alpha$.}
\end{figure}
\begin{figure}
\resizebox{0.50\textwidth}{!}{\includegraphics{bs0-1.eps}}
\caption{\label{bs01} The stabilization plots of the energies (5700 MeV $\sim$ 6000 MeV) of $b\bar{s}q\bar{q}$ states for $I(J^P)=0(0^+)$ with respect to the scaling factor $\alpha$.}
\end{figure}
\begin{figure}
\resizebox{0.50\textwidth}{!}{\includegraphics{bs0-2.eps}}
\caption{\label{bs02} The stabilization plots of the energies (6000 MeV $\sim$ 6200 MeV) of $b\bar{s}q\bar{q}$ states for $I(J^P)=0(0^+)$ with respect to the scaling factor $\alpha$.}
\end{figure}
\begin{figure}
\resizebox{0.50\textwidth}{!}{\includegraphics{bs0-3.eps}}
\caption{\label{bs03}The stabilization plots of the energies (6200 MeV $\sim$ 6300 MeV) of $b\bar{s}q\bar{q}$ states for $I(J^P)=0(0^+)$ with respect to the scaling factor $\alpha$.}
\end{figure}
\begin{figure}
\resizebox{0.50\textwidth}{!}{\includegraphics{bs1.eps}}
\caption{\label{bs1} The stabilization plots of the energies of $b\bar{s}q\bar{q}$ states for $I(J^P)=0(1^+)$ with respect to the scaling factor $\alpha$.}
\end{figure}
\begin{figure}
\resizebox{0.50\textwidth}{!}{\includegraphics{bs2.eps}}
\caption{\label{bs2} The stabilization plots of the energies of $b\bar{s}q\bar{q}$ states for $I(J^P)=0(2^+)$ with respect to the scaling factor $\alpha$.}
\end{figure}
From the table, both for $Z_{cs}$ and $b\bar{s}q\bar{q}$ system, we can easily found that the low-lying energies in diquark-antidiquark
are all much larger than those in meson-meson structures. All of them are higher than the lowest theoretical thresholds. Besides, the effects of
the structures mixing seem to be tiny for the ground state energy. The coupling energies $E_{cc}$ are a little higher than the relevant
thresholds. So we cannot find the bound states of $c\bar{c}s\bar{u}$ and $b\bar{s}q\bar{q}$ tetraquark in the chiral quark model.
Because the colorful clusters cannot fall apart directly, there may exist resonances even with the higher energies. Using the
stabilization method (real scaling method), we try to find possible resonance for $c\bar{c}s\bar{u}$ and $b\bar{s}q\bar{q}$ system.
To realize the real scaling method here, we multiply the Gaussian size parameters $r_n$ in Eq. (\ref{gaussiansize}) by a factor
$\alpha$, $r_n \rightarrow \alpha r_n$ only for the meson-meson structure with color singlet-singlet configuration. Then we can locate the
resonances of $c\bar{c}s\bar{u}$ and $b\bar{s}q\bar{q}$ system with respect to the scaling factor $\alpha$, which takes the values from
1.0 to 3.5. With the variation of $\alpha$, the scattering states will level off to corresponding thresholds, but a resonance will appear as a avoid-crossing structure, which is illustrated in Fig. \ref{avoid-crossing} \cite{rs1}. The above line represents a scattering state,
and it will fall down to the threshold. The down line is the resonance state, which try to keep stable. The resonance state will interact with the scattering state, which can bring about a avoid-crossing point in Fig. \ref{avoid-crossing}. With the increasing of the scaling
factor $\alpha$, if we can observe repeated avoid-crossing point, it will be a resonance \cite{rs1}.
To make it clear for the reader, we illustrated the stabilization plots of the energies from 3600 MeV to 4100 MeV for $Z_{cs}$ states,
respectively in Fig. \ref{Zcs1}, Fig. \ref{Zcs2}, Fig. \ref{Zcs3}. In Fig. \ref{Zcs1}, we see the first green horizontal line,
which represents the lowest threshold $J/\psi K (1 \otimes 0 \rightarrow 1)$. In the energy region 3900 MeV to 4000 MeV (Fig. \ref{Zcs2}),
there are two thresholds $D^{*0}D_s^- (1 \otimes 0 \rightarrow 1)$ and $D^0D_s^* (0 \otimes 1 \rightarrow 1)$. In higher energy range
4000 MeV to 4100 MeV in Fig. \ref{Zcs3}, two thresholds $\eta_c K^* (0 \otimes 1 \rightarrow 1)$ and $D^{*0}D_s^* (1 \otimes 1 \rightarrow 1)$
appear. Meanwhile, in the figure, we can clearly see the repeated avoid-crossing points which are marked with red circles and the red horizontal
lines are on behalf of two genuine resonance states, with the energy 4023 MeV and 4042 MeV. The energies of the resonances are not far from
the experimental values of $Z_{cs}(3985)^-$ observed by BESIII and $Z_{cs}(4000)^+$ observed by LHCb.
For $b\bar{s}q\bar{q}$ system, we show the results with all possible quantum numbers $I(J^P)=0(0^+)$, $0(1^+)$ and $0(2^+)$ in
Figs. (\ref{bs01}-\ref{bs2}). Figs. (\ref{bs01})-(\ref{bs03}) represent the $b\bar{s}q\bar{q}$ system for $0(0^+)$. In the energy range
5700 MeV to 6000 MeV (Fig. \ref{bs01}), there is one threshold $BK$, and no resonance is found. Above 6000 MeV in Figs. \ref{bs02} and
\ref{bs03}, we find several resonances, such as 6050 MeV, 6078 MeV, 6140 MeV, 6155 MeV and 6241 MeV. From Fig. \ref{bs02}, we can found
that resonance states with energies 6140 MeV, 6155 MeV have the same resonant line. To identify which state is the real resonance state,
we calculate the proportions of total 12 channels for these two eigen-states. And we find that at the avoid-crossing point for state
with 6140 MeV, the color singlet channels $B_s^0\eta$ and $B_s^*\omega$ play a major role. But for state with 6155 MeV, the hidden-color
channels occupy an important role. So we abandon the state with energy 6140 MeV and the state with energy 6155 MeV is the real resonant state
in our calculation. For $0(1^+)$ and $0(2^+)$ states in Fig. \ref{bs1} and Fig. \ref{bs2}, we cannot find any resonance states in our work.
Besides, we calculated the decay widths of these resonance states using the formula taken from reference \cite{rs1},
\begin{align}
\Gamma=4|V(\alpha)|\frac{\sqrt{|S_r||S_c|}}{|S_c-S_r|},
\end{align}
where, $V(\alpha)$ is the difference between the two energies at the avoid-crossing point with the same value $\alpha$. $S_r$ and $S_c$
are the slopes of scattering line and resonance line, respectively. For each resonance, we get the decay width at the first and the second
avoid-crossing point, and we finally give the average decay width of these two values. The results are shown in Table \ref{decaywidth}.
For $c\bar{c}s\bar{u}(4042)$ state, the decay width of 13.7 MeV is very consistent with the experimental values of $Z_{cs}(3985)^-$, with decay width of 12.8 MeV. Besides, we can see that the mass and decay width of $b\bar{s}q\bar{q}~(6078)$ state are relatively close to the experimental values $M=6063$ MeV and $\Gamma=26$ MeV by LHCb Collobations \cite{Aaij:2020hcw}. Combining with the results of $b\bar{s}$ system, it is possible that the newly observed excited $B^0_s$ states are mixing states of $b\bar{s}$ and $b\bar{s}q\bar{q}~(q=u,d)$. The unquenched quark model should be invoked to study the highly excited mesons.
\begin{table}[!t]
\begin{center}
\renewcommand\tabcolsep{6.0pt}
\caption{ \label{decaywidth} The decay widths of resonances of $c\bar{c}s\bar{u}$ and $b\bar{s}q\bar{q}$ system. (unit: MeV).}
\begin{tabular}{cccc}
\hline\hline\noalign{\smallskip}
Resonance State &$\Gamma$ &Resonance State &$\Gamma$ \\
\hline
$c\bar{c}s\bar{u}$(4023) &3.1 &$c\bar{c}s\bar{u}$(4042) &13.7\\
$b\bar{s}q\bar{q}$(6050) &7.8 &$b\bar{s}q\bar{q}$(6078) &44.1 \\
$b\bar{s}q\bar{q}$(6155) &8.7 &$b\bar{s}q\bar{q}$(6241) &4.1 \\
\hline\hline
\end{tabular}
\end{center}
\end{table}
\section{Summary}
\label{epilogue}
Motivated by the recent experimental information from BESIII and LHCb Collaboration, we calculated the mass spectrum of the $Z_{cs}$
with $I(J^P)=\frac{1}{2}(1^+)$ and $B_s^0$ states with $I(J^P)=0(0^+), 0(1^+), 0(2^+)$ in the framework of the chiral quark model using
the Gaussian expansion method. Meson-meson and diquark antidiquark structures, and the coupling of them are considered.
For $Z_{cs}$ state with quark component $c\bar{c}s\bar{u}$, we found that the low-lying eigenvalues are all higher than the corresponding
thresholds in either structure, leaving no space for a bound state. But we found two resonances with mass 4023 MeV and 4042 MeV for
$c\bar{c}s\bar{u}$ system with the help of the real scaling method, and the decay width is 3.1 MeV and 13.7 MeV, respectively.
The state $c\bar{c}s\bar{u}(4042)$ has a consistent mass with the recent observed state $Z_{cs}(3985)^-$ and $Z_{cs}(4000)^+$, but
the decay width is close to the experimental value of $Z_{cs}(3985)^-$ and far narrower than the experimental value $Z_{cs}(4000)^+$.
To find the excited $B_s^0$ state observed by LHCb Collaboration, we give the mass spectrum both in 2-body $b\bar{s}$ system and 4-body
$b\bar{s}q\bar{q}(q = u~\rm{or}~d)$ system by considering the possible production of quark-antiquark pair in the vacuum. For quark-antiquark
system, the $2S$ and $1D$ states have masses close to the newly observed $B_s^0$, so the chiral quark model can accommodate these excited
$B_s^0$ states. For four-quark system, no bound state is found. However several resonances are emerged. They have energies, 6050 MeV,
6078 MeV, 6155 MeV, and 6241 MeV. The decay width are all relatively narrow, with 7.8 MeV, 44.1 MeV, 8.7 MeV and 4.1 MeV, respectively.
Comparing with the experimental data, we found that it is also possible to interpret the observed $B_s^0$ states as four-quark states. Therefore the better way to investigate the highly excited states is to invoke the unquenched quark model~\cite{Chen:2017mug}, which is our future work.
These possible resonant states should be tested in more precise experimental data in the future and we need more experimental studies
on the dominant decay channels of $Z_{cs}$ and $B_s$ to figure out their inner configurations.
|
\section{Introduction}
This work proposes improvements in osteoplasty by image analysis using deep learning. Spinal vertebra compression fractures (VCFs) are a painful and debilitating injury to the skeleton. The main reasons for such fractures are osteoporosis, trauma or tumors, as many cancers metastasize into the vertebral column. Patients with osteoporotic fractures can be relieved from their pain and regain mobility by stabilizing the affected vertebrae.
Osteoplasty, namely kyphoplasty or vertebroplasty, is an operative procedure, during which holes are drilled or hammered from the back of the patient through the pedicles of a vertebra to the vertebral body~\cite{filippiadis2017percutaneous}.
Subsequently, the vertebral body is filled with a bone cement, which stabilizes the vertebra and relieves the patient from the pain.
Yet, it is difficult to determine the correct amount of cement to be injected~\cite{janssen2017risk}. In particular, if it is too much bone cement a leakage may put pressure on the spinal cord and can even lead to pulmonary cement embolisms~\cite{sorensen2019vertebroplasty}.
The goal of this work is to provide a virtual vertebra reconstruction in a personalized manner, taking into account patient anatomy and the vertebra type. The only data readily available comes from CT imaging, which is solely used to derive the shape of a healthy-looking vertebra, matching the patient's spine, and an upper bound for the bone cement to be injected. To extract all the information necessary we propose an automated framework based on deep learning techniques with the following steps:
\begin{enumerate}
\item \emph{Vertebrae Segmentation}: Every individual vertebra in the input CT image of a patient is detected, labeled with its anatomic denomination, and segmented. The segmentation masks serve as an input to the next step.
\item \emph{Virtual Spine Straightening}: This step simulates the restoration of the vertebral column to a healthy state after stabilizing the fractured vertebra. Usually, only the post-fracture CT image is available for a patient. To estimate the size of the restored fractured vertebra, the healthy vertebrae are compared to a spine atlas, which is scaled to match the patient. The label of the fractured vertebra is required as an input. The healthy vertebrae of the patient are registered to the atlas in a vertebra-wise rigid and deformable approach, providing a physiologically healthy vertebral column.
\item \emph{Vertebra Inpainting}: The shape of the vertebra before the fracture is estimated and its volume measured. A generative adversarial neural network (GAN) produces realistic 3D shapes of vertebrae, which are put in place of the fractured vertebra. The upper bound on the amount of cement can be estimated from the difference in volume between the inpainted and the fractured vertebra.
\end{enumerate}
A visualization of the workflow is given in Figs.~\ref{fig:workflow} and \ref{fig:workflow_visual}.
The goal of this work is to provide a virtual vertebra reconstruction in a personalized manner, taking into account patient anatomy and the type of vertebra. The only data readily available comes from CT imaging, which is solely used to derive the shape of a healthy-looking vertebra matching the spine of the patient and an upper bound for the bone cement to be injected. To extract all the information necessary we propose an automated framework based on deep learning techniques following the workflow of Figs.~\ref{fig:workflow} and \ref{fig:workflow_visual}.
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{Figures/workflow.png}
\caption{A visual example of the inputs and outputs of each stage of the framework.}
\label{fig:workflow}
\end{figure*}
\begin{figure*}
\centering
\input{workflow_short_new.tex}
\caption{Proposed framework: The CT image is first used to generate a segmentation mask, which is then fed into the spine straightening block, together with the CT image and the fractured vertebra label (e.g. "L2"). The straightened spine and corresponding segmentation are
then sent to the inpainting block to generate the final virtually healthy CT
and the maximum cement needed.}
\label{fig:workflow_visual}
\end{figure*}
The following is a summary of the main contributions in this paper:
\begin{itemize}[leftmargin=*]
\item To our knowledge, this is the first framework to do an end-to-end spine straightening and vertebra reconstruction to estimate the upper bound of bone cement for osteoplasty. Further, the proposed pipeline is fully automatic.
\item We introduce a robust method for virtually straightening a fractured spine solely using a CT image and the label of the damaged vertebra and validate it on clinical data.
\item We propose a patient-tailored 2.5D inpainting method to generate a healthy-looking version of a fractured vertebra and its segmentation mask.
\end{itemize}
\section{Methodology}
Each step of the pipeline (Figs.~\ref{fig:workflow} and \ref{fig:workflow_visual}) is detailed in the following subsections. The models were implemented in PyTorch and their weights, along with code for running the pipeline will become publicly available upon acceptance.
\paragraph{Spine Segmentation.}
The goal of this step is to automatically generate voxel-level masks of the vertebrae. Several methods have been proposed for segmenting the spine. For instance, in the VerSe’19 challenge, eleven fully automated algorithms were benchmarked for this task~\cite{sekuboyina2020verse}. Here, we reimplement the approach of Sekuboyina et al. splitting the task into three modules: spine detection, vertebrae labeling, and vertebrae segmentation~\cite{sekuboyina2020verse,btrfly-sekuboyina}.
For detecting the spine, we employ a variant of the U-Net~\cite{ronneberger2015u,roy2018concurrent}, to regress a coarse 3D Gaussian heatmap of the spinal centerline and a 3D bounding box around it to localize the spine. For the labeling module, we use a self-implemented modified version of the BtrflyNet~\cite{btrfly-sekuboyina}, which works on 2D sagittal and coronal maximum intensity projections (MIP). We extract the MIPs from the localized spine region from the previous stage, thereby removing occlusions from ribs and pelvic bones. The network then predicts refined Gaussian heatmaps centered at each vertebra. Finally, once the vertebrae are labeled, their segmentation is modeled as a binary segmentation task, using again a modified U-Net. The spine image is first cropped to a 3D patch around each vertebral centroid previously detected. Similarly, the predicted heatmap image of the same centroid is also cropped. Both cropped vertebral image and vertebral heatmap are then fed into the network in order to segment only the vertebra of interest.
\paragraph{Virtual Spine Straightening.}
For this task, Forsberg et al. proposed splitting registered sub-volumes of the patient's spine and atlas with non-rigid transformations, which evidently ignored the rigid nature of the vertebrae~\cite{forsberg2015atlas}. Drobny et al. proposed registering the spine with a poly-rigid transformation model thus ensuring a rigid transformation of the vertebrae, but evaluating their approach only on synthetic data~\cite{drobny2020towards}. We follow a similar approach, however tackle the issue of varying patient sizes and use a simpler method for deforming non-vertebra voxels. From the previous step, we obtain the vertebrae centroids and segmentation mask. First, we scale a spine atlas to be the same height as the patient, where the scaling factor is computed by comparing the sum of the distances between the centroids of all visible vertebrae on the image excluding the fracture(s). The label(s) of the fracture(s) is taken as an input. We calculate the distance using the first two principal components of the centroid's coordinates to mitigate the deviation resulting from the patient's position during the CT acquisition. A set of displacement fields is generated from a per-vertebra rigid registration on the segmentation masks. Next, a distance map of each vertebra is calculated, where the voxel value represents the physical distance to the nearest vertebra. The displacement fields are then compounded with Alg.~\ref{alg:displacement}, based on our biomechanic assumption that the soft tissue are transformed inversely proportional to their distance from vertebrae, similar to linear blend skinning~\cite{lewis2000}. Finally, we generate the straightened spine by resampling using the combined displacement field.
\begin{algorithm}
\caption{Algorithm used for combining the vertebrae displacement fields.}\label{alg:displacement}
\begin{algorithmic}[1]
\Require{Vertebrae set $\mathbf{V}=\{1,2,...,24\}$ excluding the fractured vertebra(e), set of displacement fields $\{\mathbf{F_i}\}_{i\in\mathbf{V}}$, $\mathbf{F_i}\in\mathbb{R}^{h\times w\times d}$, set of distance maps, $\{\mathbf{D_i}\}_{i\in\mathbf{V}}$ , $\mathbf{D_i}\in \mathbb{R}^{h\times w\times d}$, where h, w, d are height, width and depth}{}
\For{every voxel p in scan}
\If{voxel p is in vertebrae i}
\State{$\widetilde{\mathbf{F}}(p)\gets\mathbf{F_i}(p)$}
\Else{}
\State{$\widetilde{\mathbf{F}}(p)\gets \frac{\sum_j \mathbf{D_j}^{-1}(p)*\mathbf{F_j}(p)}{\sum_j\mathbf{D_j}^{-1}(p)}$}
\EndIf
\EndFor
\Return{combined displacement field $\widetilde{\mathbf{F}}$}
\end{algorithmic}
\end{algorithm}
\paragraph{Vertebra Inpainting.}
After spine straightening, the next task in the framework is to replace the fractured vertebra with its healthy equivalent. This is done using inpainting, which has previously been used in medicine for predicting missing information~\cite{armanious2019adversarial,torrado2021inpainting}, removing lesions~\cite{zhang2020robust} or correcting limited-angle acquisitions~\cite{zhao2018unsupervised,li2019promising,wang2020effective}. None of these works tackles a similar problem to ours. Here, though we are dealing with 3D volumes, when focusing on vertebral bodies, the sagittal and coronal views are expected to provide sufficient information for reconstruction. We, therefore, chose to apply two 2D models to a volume, one trained on sagittal slices, the other on coronal, and fuse the outputs. While we are mainly interested in the segmentation mask for the framework pipeline, we tackle the inpainting as a multi-task learning problem, and train a model to generate the inpainted CT image as well as the segmentation mask. This enforces the model to better learn the training data distribution.
We remove the fractured vertebra from the CT and segmentation mask slices by applying a binary mask, which is generated from the segmentation and fractured label. These are the inputs to the inpainting network, whose architecture extends that of Yu et al.~\cite{yu-context-att}. We adjust their architecture here to receive two inputs, the corrupted CT slice, and corresponding segmentation. Similarly to the original authors' network, the input first passes through a coarse generator for a first estimate of the inpainted result and next through a refinement network, both of which have a U-like architecture. The last layer of the refinement network has been adjusted to two final parallel layers, which output the inpainted image and the inpainted mask. Four discriminators follow the generator; two, for image and mask, local for the patch region and two global, for the entire image space. We use the same losses as presented in \cite{yu-context-att}, extending the generator loss by adding dice losses for the patch region and the background of the mask and adjusting the total discriminator loss to accommodate all four sub-networks. To obtain the final volumes, we apply a simple averaging of the intensities in the two CT volumes, while for the segmentation, the softmax outputs of the segmentation layer are averaged before acquiring the final predictions. From this we then compute the virtually healthy vertebra volume, and by substracting from this the original fracture vertebra volume, compute an upper bound for the cement.
\section{Experimental Setup}
For training and evaluating the individual steps of our framework, as well as evaluating the overall pipeline we use three separate datasets.
\paragraph{Segmentation Training Dataset.}
For the vertebra segmentation pipeline, we used the public dataset VerSe’20~\cite{loffler2020vertebral,sekuboyina2020labeling,sekuboyina2020verse}, which consists of 100 patients with ground-truth vertebrae centroid annotations and segmentation masks. This data comes from multiple multi-detector CT scanners, and includes a wide range of fields-of-view, scan settings, and certain cases with vertebral fractures, metallic implants, and foreign materials. We randomly split this dataset into a training set of 70 patients, and a validation and testing set of 15 patients, respectively.
\paragraph{Inpainting Training Dataset.}
Since we want our model to learn to inpaint the region of interest with a healthy vertebra, we use a healthy dataset for training. We created two 2D datasets, one of coronal and one of sagittal views, from a total of 110 volumes (95 confirmed healthy spines from our institution and 15 from the CSI challenge\footnote{http://csi-workshop.weebly.com/challenges.html} which we verified to be healthy). From these volumes ten were set aside for testing and ten for validation of both networks. Depending on the size and resolutions of the scans, the number of slices and spine regions taken from each volume varied, making up a total of 3557 sagittal and 1358 coronal images, each consisting of five vertebra.
\paragraph{Dataset for Validation of Virtual Spine Straightening and Overall Pipeline.}
For the evaluation of the straightening and the overall workflow, we obtained a list of 316 patients who underwent kyphoplasty between 2014 and 2019 at our institution. We filtered these to include only patients who (1) have the pre-fracture, pre-operative and post-operative CT scans, and (2) have CT images with voxel sizes smaller than 1 mm in the x and y direction and smaller than 3 mm on z axis. After applying these criteria, our dataset included ten patients. For the sake of completeness, some of these ten patients have more than one fractured vertebra, which enabled us to evaluate the pipeline on a total of 15 vertebrae. The CT scans included in the dataset have resolution between 0.28 and 0.97 mm in the x and y direction, and between 0.7 and 3.0 mm in the z direction.
\paragraph{Segmentation Pipeline Training.}
We trained each module of the segmentation pipeline independently, following the same dataset split for every task. For the detection stage, we first built the ground-truth spine heatmap by combining the individual heatmaps of the vertebrae landmarks. We then trained the 3D U-Net variant with the L2 loss, using the Adam optimizer, a batch size of two and a learning rate of 1e-3 until convergence (77 epochs). Next, we extracted the bounding box around the predicted spine heatmap. For the labelling stage we trained the modified BtrflyNet on the MIPs extracted from the previous stage bounding box. Here, we used the same hyperparameters and L2 loss, since this task is also modeled as a heatmap regression, and trained until convergence (280 epochs). Finally, we used the vertebral patches and heatmaps from the previous stage to train the modified 3D U-Net. We used the Dice Loss and trained for 25 epochs, again with the same hyperparameters.
\paragraph{Inpainting Model Training.}
We trained the two models on the lateral and coronal datasets, where each slice contains four visible vertebrae and one digitally erased. Thus, each image was sent five times through the networks, which were trained using a batch size of 16, with Adam for the optimization and initial learning rates of 0.001 and 1e-4 for the generator and discriminator. A grid search was applied to find optimal loss weights and learning rates, in the space around the original authors' choice. We implement early stopping considering the segmentation metrics, since for our pipeline the segmentation result is of higher relevance. The best performing models on the validation sets were chosen.
\section{Results}
\paragraph{Evaluation of Segmentation.}
For assessing the performance of the vertebrae segmentation framework, we employed the dice coefficient metric. We computed this metric at a vertebra level over all the vertebrae annotated in the ground truth from the VerSe2020 mask labels. The average dice score over all patients in the test set is 87.06$\pm$9.54\%, a value well in the range of the reported values from the literature (83.06 to 93.01\%, Tab.~\ref{tab:inpaiting}, e.g., \cite{sekuboyina2020verse}).
\begin{table}
\centering
\caption{Evaluation of Vertebra Segmentation as compared to state-of-the-art methods}
\label{tab:inpaiting}
\begin{tabular}{|l|l|}
\hline
Team & Dice \\
\hline
Payer C. & 90.90 \\
Lessmann N. & 85.08 \\
Sekuboyina A. & 83.06 \\
Chen M. & 93.011 \\
Hu Y. & 84.07 \\
\textbf{Ours} & 87.06 \\
\hline
\end{tabular}
\end{table}
\paragraph{Evaluation of Virtual Spine Straightening.}
We evaluated the spine straightening algorithm by utilizing what we call the \emph{fracture distance}. This is the physical distance between the vertebrae above and below the fracture, which should increase to provide enough space for the inpainting of a healthy vertebra. In order to make the distance comparable, the pre-fracture spine CT was also registered to the atlas space using the straightening algorithm. We therefore compared the fracture distance of straightened pre-fracture, raw post-fracture and straightened post-fracture spine of the patients (Figs.~\ref{fig:straightening_example} and \ref{fig:straightening_and_overall}). The average error on the \emph{fracture distance} yielded 0.23$\pm$2.68 mm, or equivalently, 0.50$\pm$3.95\%. The per-vertebra results are tabulated in Tab.~\ref{tab:straightening}.
\begin{table*}[]
\caption{Results of the straightening of the spine. The values in the table is the distance between the two vertebrae specified in the second column in the pre-fractured, the fractured and straightened spine CT. The pre-fractured and and straightened distances are computed on the corresponding scaled CT scans, but the fractured distance is computed on the original (therefore not scaled) CT scan. The error (between the straightened and pre-fractured distances) and the relative error are also reported.\\}
\label{tab:straightening}
\centering
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
Patient & Vertebrae&Pre-fractured &Fractured &Straightened &Error&RE\\
&&distance [mm]&distance [mm]& distance [mm]&[mm]&[\%]\\
\hline
1&T12-L2&71.78&57.28&68.57&-3.21&-4.54\\
&L2-L4&75.00&71.31&75.01&0.02&0.02\\
&L3-L5&76.59&70.45&76.76&0.17&0.23\\
2&T8-T10&48.69&48.48&50.16&1.47&3.02\\
3&T11-L1&56.15&50.25&59.89&3.74&6.66\\
4&L1-L3&64.07&59.23&59.88&-4.19&-6.39\\
&L2-L4&67.25&62.53&67.80&0.55&0.82\\
5&L1-L3&66.99&69.07&68.42&1.44&2.02\\
6&L1-L3&68.27&60.31&69.21&0.94&1.37\\
&L2-L4&69.80&64.19&72.68&2.88&4.21\\
7&L1-L3&72.95&71.68&71.96&-1.00&-1.33\\
&L3-L5&77.37&73.27&72.13&-5.24&-6.86\\
8&L3-L5&74.80&76.27&78.33&3.53&4.77\\
9&L1-L3&70.17&68.54&72.44&2.27&3.30\\
10&T12-L2&63.37&64.64&63.53&0.16&0.25\\
\hline
\textbf{Average}&&&&&\textbf{0.23}&\textbf{0.50}\\
\textbf{STD}&&&&&\textbf{2.68}&\textbf{3.95}\\
\hline
\end{tabular}
\end{table*}
\begin{figure}
\centering
\includegraphics[width=.45\textwidth]{Figures/str_result.png}
\caption{Visual results of the straightening step. From left to right: MIPs of the pre-fractured spine, post-fractured and straightened spine, with the distance (in mm) between the vertebrae above and below the fractured one.}
\label{fig:straightening_example}
\end{figure}
\paragraph{Evaluation of Inpainting.}
To evaluate the results of the inpainting model we computed the structural similarity index metric (SSIM) and peak signal-to-noise Ratio (PSNR) of the output CT, as well as the intersection over union (IoU) of the segmentation in the patch region of interest and the mean relative volume error (MRE) of the said vertebra. Fusing the coronal and sagittal outputs improves the individual results, giving our final model an SSIM of 0.82 and PSNR of 26.45 dB for the image, an IoU of 0.76 for the mask and an MRE of 19\% for the vertebra volume (see Fig.~\ref{fig:inpaiting_example} and Tab.~\ref{tab:inpaiting}).
\begin{table}
\centering
\caption{Evaluation of inpainting per view. The metrics used for the CT inpainting are structural similarity index metric (SSIM) and peak signal-to-noise Ratio (PSNR), while for the segmentation masks we \\}
\label{tab:inpaiting}
\begin{tabular}{|l|c|c|c|c|}
\hline
Method & Image SSIM & Image PSNR & Mask MRE & Mask IoU\\
\hline
Sagittal & 0.771 & 25.380 & 0.228 & 0.734 \\
Coronal & \textbf{0.830} & 26.448 & 0.316 & 0.691 \\
Fusion & 0.823 &\textbf{28.16} & \textbf{0.190} & \textbf{0.757} \\
\hline
\end{tabular}
\end{table}
\begin{figure*}
\centering
\subfloat
{\includegraphics[width=.45\textwidth]{Figures/example_inpainting_lateral.png}} \quad \quad \quad
\subfloat
{\includegraphics[width=.45\textwidth]{Figures/example_inpainting_coronal.png}}
\caption{Visual results of the inpainting step: From left to right: ground-truth, input and output of the sagittal (left pane) and coronal (right pane) GANs. The first row shows the CTs, while the second depicts the segmentation masks.}
\label{fig:inpaiting_example}
\end{figure*}
\paragraph{Evaluation of Overall Pipeline.}
We evaluated the performance of our framework in an end-to-end manner by using ten patients. We note again here, that for these patients a CT scan before and after the VCF were fortuitously available. To evaluate the effectiveness of our pipeline, we compare the volume of the vertebrae before the VCF (pre-fractured) with the output of our framework, i.e., the virtually straightened inpainted post-fracture scan. Overall, our method has an error of 2.61$\pm$5.07 mL, which translates to a relative error of 3.08$\pm$7.63 \%. Fig.~\ref{fig:straightening_and_overall} shows the correlation between the pre-fracture and straightened inpainted vertebra volumes. The table with the results at a per-vertebra level is included in the supplementary material.
\begin{figure}
\centering
\subfloat[\emph{Inpainting}]
{\includegraphics[width=.4\textwidth]{Figures/Inpainting.png}} \quad \quad \quad
\subfloat[\emph{Straightening}]
{\includegraphics[width=.4\textwidth]{Figures/Straightening.png}}
\caption{Quantitative results of the inpainting and straightening steps.}
\label{fig:straightening_and_overall}
\end{figure}
\begin{table*}[]
\caption{Framework analysis}
\label{tab:volume_evaluation}
\centering
\begin{tabular}{|c|c|c|c|c|p{10mm}<{\centering}|p{10mm}<{\centering}|}
\hline
Patient & Vertebra & Pre-fractured & Fractured & Inpainted & Error & RE \\
&&volume [mL]&volume [mL]&volume [mL]&[mL]&[\%]\\
\hline
1&L1&71.61&62.41&69.30& 2.31& 3.23\\
&L3&86.65&79.89&86.01& 0.64& 0.74\\
&L4&82.61&79.65&86.95&-4.34&-5.25\\
2&T9&38.18&37.69&39.69&-1.51&-3.96\\
3&T12&49.31&38.99&42.88&6.43&13.05\\
4&L2&63.23&58.90&69.35&-6.12&-9.67\\
&L3&67.01&66.21&68.57&-1.57&-2.34\\
5&L2&76.45&72.05&76.29&0.16&0.21\\
6&L2&55.26&44.02&46.94&8.32&15.05\\
6&L3&60.33&55.35&55.81&4.52&7.50\\
7&L2&75.22&69.32&68.30&6.92&9.20\\
&L4&80.30&73.08&70.22&10.09&12.56\\
8&L4&89.86&86.41&80.30&9.56&10.64\\
9&L2&67.53&64.84&68.54&-1.01&-1.49\\
10&L1&53.89&51.66&49.19&4.70&8.72\\
\hline
\textbf{Average}&&&&&\textbf{2.61}&\textbf{3.88}\\
\textbf{STD} &&&&&\textbf{5.07}&\textbf{7.63}\\
\hline
\end{tabular}
\end{table*}
\paragraph{Ablation study.}
A final relevant evaluation is that of validating the significance of the spine straightening to the inpainting results. For this we compute the volume of the inpainted vertebra directly using the fractured spine, without using the straightening algorithm for the same ten patients (15 vertebrae). We compare here the volume of the inpainted vertebrae with and without straightening (Tab.~\ref{tab:ablation}).
\begin{table*}[]
\centering
\caption{Results of the ablation study in which the straightening of the spine was omitted. The values summarize the effect of the spine straightening in the pipeline, by reporting the error (between the inpainted and pre-fractured distances) and the relative error.}
\label{tab:ablation}
\centering
\begin{tabular}{|c|c|c|c|p{10mm}<{\centering}|p{10mm}<{\centering}|c|c|p{10mm}<{\centering}|p{10mm}<{\centering}|}
\hline
& & Pre-fractured & \multicolumn{3}{c|}{Straightening} & \multicolumn{3}{c|}{W/o straightening} \\
\cline{4-9}
Patient & Vertebra&volume&Inpainted & Error & RE & Inpainted & Error & RE\\
&&[mL]&volume [mL]&[mL]&[\%]&volume [mL]&[mL]&[\%]\\
\hline
1&L1&71.61&69.30&2.31&3.23&55.462&16.14&22.55\\
&L3&86.65&86.01&0.64&0.74&74.14&12.51&14.44\\
&L4&82.61&86.95&-4.34&-5.25&71.11&11.51&13.93\\
2&T9&38.18&39.69&-1.51&-3.96&33.76&4.42&11.58\\
3&T12&49.31&42.88&6.43&13.05&31.44&17.87&36.25\\
4&L2&63.23&69.35&-6.12&-9.67&55.02&8.21&12.99\\
&L3&67.01&68.57&-1.57&-2.34&59.02&7.99&11.92\\
5&L2&76.45&76.29&0.16&0.21&61.74&14.70&19.23\\
6&L2&55.26&46.94&8.32&15.05&42.64&12.62&22.83\\
6&L3&60.33&55.81&4.52&7.50&51.96&8.37&13.87\\
7&L2&75.22&68.30&6.92&9.20&59.45&15.77&20.96\\
&L4&80.30&70.22&10.09&12.56&66.10&14.20&17.69\\
8&L4&89.86&80.30&9.56&10.64&79.43&10.43&11.61\\
9&L2&67.53&68.54&-1.01&-1.49&65.04&2.49&3.69\\
10&L1&53.89&49.19&4.70&8.72&47.47&6.42&11.91\\
\hline
\textbf{Average}&&&&\textbf{2.61}&\textbf{3.88}&&\textbf{10.91}&\textbf{16.08}\\
\textbf{STD}&&&&\textbf{5.07}&\textbf{7.63}&&\textbf{4.52}&\textbf{7.46}\\
\hline
\end{tabular}
\end{table*}
\section{Discussion}
In this work, we have presented an integrated framework to estimate a realistic healthy state for fractured vertebrae. Using only a patient's post-traumatic CT image, a healthy vertebra shape replaces the fractured one, virtually restoring the spine. Every single stage in this pipeline is necessarily required to obtain a genuine result. As a clinical application of our work, we envision that the volume of the estimated vertebra can be used to estimate the amount of bone cement needed to stabilize the vertebra without leakage, albeit we are not performing this very analysis. Therefore, we view the estimated volume as an upper bound on the injection of material.
As the primary output of the workflow is a 2.5D reconstructed image generated by a GAN trained on healthy spines, it is necessary that the healthy vertebrae in the image to be inpainted are resembling a healthy spine and provide accurate space for the inpainting. The virtual spine straightening step solves this problem.
We have also validated the distance between the vertebrae neighbouring the fracture for those patients, where pre-traumatic imaging is available.
We validate the results of the inpainting on the same set of patients showing that it tallies very well with their pre-fractured state at a relative error of below 4\%. In general, the method is slightly underestimating the volume of the vertebrae, which can be considered conservative given the necessity for an upper bound. Admittedly, this was not engineered and will be investigated further.
In its current state, our framework demonstrates its high potential as an automatic method for reconstruction of healthy-looking medical images, requiring minimal diagnostic input. The reported ability to derive quantitative results proves the usefulness of deep learning approaches for planning interventions.
\section*{Acknowledgements}
The authors would like to thank the support of Magdalini Paschali, Stefan Walke, Sebastian Lutz and the rest of the team at the Interdisciplinary Research Lab at Klinikum rechts der Isar.
\bibliographystyle{splncs04}
|
\section{Introduction}
Mechanisms for heating the upper solar atmosphere and corona, be they magnetohydrodynamical (MHD) waves or magnetic reconnection, generally rely on the transverse motion of magnetic footpoints in the photosphere. Excitations of kink tube waves or compressive fast modes are often considered to arise from ``granular buffeting,'' by which photospheric magnetic flux concentrations are rapidly and stochastically moved in a sideward direction. Another option is given by turbulent motions in the convection zone, which generate a whole spectrum of transversal excitations. On the other hand, heating by magnetic reconnection is commonly considered to be induced by transverse motions of magnetic footpoints. These can lead to the twisting and braiding of magnetic flux tubes forming dissipative electric current sheets.
Different from transverse excitations, vortical
motions of the magnetic footpoints have been less in the focus of research, mainly
for lack of observational support. In the past, vortical motions and torsional Alfv\'en waves were of rather hypothetical nature, often serving theoretical models for the driving of spicules
\citep[e.g.,][]{1975SoPh...42...79S,
1981SoPh...70...25H,
2011ApJ...740L..46F}.
This state of affairs has drastically changed over the past decade.
In the photosphere, swirls of granular scale were detected by tracking magnetic bright points (BPs) in the intergranular space \citep{2008ApJ...687L.131B,
2010A&A...513L...6B,
2011A&A...531L...9M}
and by the means of local correlation tracking (LCT) \citep{2010ApJ...723L.139B,
2011MNRAS.416..148V,
2017ApJS..229...14R,
2018A&A...610A..84R}.
In the chromosphere, vortical motions with a typical size of about $2\,\mathrm{arcsec}$ were observed by \citet{2009A&A...507L...9W}
and \citet{2012Natur.486..505W}
using \ion{Ca}{ii} 854\,nm narrowband filtergrams obtained with the CRisp Imaging SpectroPolarimeter (CRISP) instrument of the Swedish 1-$\mathrm{m}$ Solar Telescope (SST). Moreover, these chromospheric swirls were colocated with magnetic BPs in the photosphere, suggesting a magnetic origin.
Higher up, \citet{2012Natur.486..505W}
detected imprints of chromospheric swirls in the transition region and the corona in the spectral lines of \ion{He}{ii} $30.4\,\mathrm{nm}$ and \ion{Fe}{ix} $17.1\,\mathrm{nm}$, respectively, using recordings of the Atmospheric Imaging Assembly (AIA) instrument of NASA's space-based Solar Dynamics Observatory (SDO). These observations suggested that chromospheric swirls are the observational signatures of rotating coherent magnetic structures, which reach from the convection zone to the outer layers of the atmosphere.
\citet{2013ApJ...768...17M}
found the chromospheric counterpart of photospheric swirls in the form of a quasi-periodic torsional motion, in time series of H$\alpha$ narrowband filtergrams obtained with the ROSA (Rapid Oscillations in the Solar Atmosphere) instrument attached to the Dunn Solar Telescope.
\cite{2016A&A...586A..25P}
detected swirls consisting of spiral arms in H$\alpha$ filtergrams taken with CRISP at SST, which coexisted with strong upflows in the upper chromosphere as seen from \ion{Mg}{ii} k line Dopplergrams obtained with the space-based Interface Region Imaging Spectrograph (IRIS).
More recently, a $1.7\,\mathrm{h}$ persistent vortex flow of $6\,\mathrm{arcsec}$ diameter was observed by \citet{2018A&A...618A..51T, 2019A&A...623A.160T}
with CRISP at SST in the cores of the H$\alpha$ and \ion{Ca}{ii} $854.2\,\mathrm{nm}$ lines, while no BPs were observed in the line wings of H$\alpha$ and \ion{Ca}{ii}.
Finally, \cite{2019ApJ...881...83S}
found, from spectral imaging in the lines of H$\alpha$ and \ion{Ca}{ii} $854\,{\rm nm}$ along with polarimetry in \ion{Fe}{i} $630.2\,{\rm nm}$, that the rotation of magnetic flux concentrations in the photosphere matches the chromospheric swirl. However, they reported that they could not determine whether a swirl is a gradual response to the photospheric motion or an actual propagating Alfvénic wave.
Alfv\'en waves were theoretically predicted by \citet{1942Natur.150..405A}, but their detection in the Sun, which provides preferential conditions for their existence, proved to be a difficult task. Nevertheless, over the past two decades, multiple observations have revealed their presence and propagation in the solar atmosphere and corona
\citep[e.g.,][]{2007Sci...317.1192T,
2011ApJ...736L..24O,
2007Sci...318.1577O}.
\citet{2009Sci...323.1582J}
found signatures of propagating torsional Alfv\'en waves in the photosphere in oscillatory phenomena associated with a conglomeration of magnetic BPs.
\citet{2012ApJ...752L..12D}
reported rotational motion in spicules, which they interpreted to be torsional Alfv\'en waves, and \citet{2017NatSR...743147S}
detected ubiquitous high frequency (${\sim}12\,\text{-}\,42\,\mathrm{mHz}$) torsional motions in thin spicular-type structures in the chromosphere that resemble torsional Alfv\'en waves.
\citet{2019NatCo..10.3504L}
called attention to ubiquitous torsional Alfv\'en pulses by correlating photospheric and chromospheric swirls.
Moreover, \citet{2018NatPh..14..480G} provided observational evidence of Alfv\'en waves heating chromospheric plasma in a sunspot umbra through the formation of shocks.
Vortex flows and their connection with Alfv\'en waves have also been extensively studied with MHD numerical simulations. Small-scale swirls appear regularly in simulations of the solar surface convection and atmosphere \citep[see, e.g.,][]{
1998ApJ...499..914S,
2004RvMA...17...69V,
2011A&A...526A...5S,
2011A&A...533A.126M,
2012A&A...541A..68M,
2012ASPC..456....3S,
2013ApJ...770...37K,
2016A&A...596A..43C,
2019A&A...632A..97L,
2020ApJ...894L..17Y}.
\citet{2013ApJ...776L...4S}
identified apparent vortex-like motions in magnetic flux tubes of the photosphere as torsional Alfv\'en waves by analyzing time-distance diagrams and plotting particle tracks.
\citet{2014PASJ...66S..10W}
put forward a model consisting of two vortex systems, where the upper chromospheric and photospheric swirling plasma is tightly coupled to the ``frozen-in'' magnetic field lines, which have their footpoints within the lower intergranular vortex flow.
This system of vortical structures was dubbed a ``magnetic tornado'' \citep{2012Natur.486..505W}.
\citet{2011AnGeo..29.1029F},
\citet{2012ApJ...755...18V},
and \citet{2015A&A...577A.126M}
studied the generation, propagation, and energy transfer of torsional Alfv\'en waves in modeled solar magnetic flux tubes, and \citet{2019NatCo..10.3504L} showed that azimuthal perturbations of the magnetic field in the upper photosphere can generate Alfv\'en pulses, which carry the information of the rotational motion into the chromosphere. However, the exact mechanism at the origin of the swirls in realistic numerical simulations is still unclear.
A particularly interesting aspect of these swirls is that they possibly generate torsional Alfv\'en waves that propagate upward and provide an efficient mechanism for chromospheric and coronal heating. From numerical simulations, \citet{2012Natur.486..505W}
estimated a net vertical Poynting flux associated with swirling motions of $440\,\mathrm{W\, m^{-2}}$ at the interface between the chromosphere and the corona, while \citet{2019NatCo..10.3504L}
estimate a minimal nonthermal energy flux of $33\,\text{-}\,131\,\mathrm{W\,m^{-2}}$ in the middle chromosphere.
In this paper, we study in detail ``magnetic swirls,'' that is, the relation between swirls, magnetic field perturbations, and Alfv\'en waves.
In particular, we investigate the MHD processes that lead to the propagation of vortical motions from the photosphere to the chromosphere, where they appear as chromospheric swirls. For that purpose, we analyze realistic radiative MHD numerical simulations carried out with the CO$5$BOLD code
\citep{2012JCoPh.231..919F}
by using the swirling strength criterion and its evolution equation
\citep{1999JFM...387..353Z,
2020A&A...639A.118C}.
The paper is organized as follows: Section \,\ref{sec:theory} reviews a number of theoretical concepts that are used for the analysis. Section\,\ref{sec:numerics} gives details on the numerical simulations and Sect.\,\ref{sec:results} presents the analysis and discusses the results.
A summary and conclusions are given in Sect.\,\ref{sec:conclusions}.
\section{Theoretical background}\label{sec:theory}
This section reviews a few theoretical concepts that are used for the analysis of the simulations. These concern properties of Alfv\'en waves, the computation of energy fluxes and the concept of swirling strength. Moreover, we introduce a new quantity, the ``magnetic swirling strength,'' which is used for detecting twists in magnetic fields.
%
%
\subsection{Alfvén waves \label{subsec:AlfvénWaves}}
Alfvén waves are perturbations in the plasma with the magnetic tension acting as a restoring force. The perturbations are transverse to the propagation direction and magnetic field lines; therefore, one speaks of shear or torsional Alfv\'en waves.
When the perturbation propagates across the magnetic field, causing compression and rarefaction of the magnetic field and plasma, one speaks of compressional Alfv\'en waves or fast mode waves \citep[see, e.g.,][]{2014masu.book.....P}.
For the purposes of this paper, we focus solely on the former.
We consider with Fig.\,\ref{fig:AlfvenWavesScheme} a static, incompressible plasma in the ideal MHD approximation, with constant density $\rho$ and uniform and vertical magnetic field $\boldsymbol{B}_0 = B_0 \boldsymbol{e}_z$. Moreover, we suppose that the magnetic field dominates the equilibrium so that the hydrodynamical pressure and gravity can be neglected. Although these conditions are not satisfied in the solar atmosphere, this abstraction still serves to describe some basic behavior of the Alfv\'en wave propagation in the solar photosphere and chromosphere.
Introducing the magnetic field perturbation with $\boldsymbol{B} = \boldsymbol{B}_0 + \boldsymbol{b}$, where $|\boldsymbol{b}| \ll |\boldsymbol{B}_0|$, the velocity perturbation $\boldsymbol{v}$, and linearizing the system of MHD equations using incompressibility
\citep[see, e.g.,][Chap. 4]{2014masu.book.....P},
yields two important characteristics of Alfv\'en wave propagation.
The first one relates the velocity perturbation to the magnetic field perturbation
\begin{equation} \label{eqn:PerturbationRelation}
\boldsymbol{v} = -\dfrac{\omega\boldsymbol{b}}{\boldsymbol{k} \cdot \boldsymbol{B}_0} = \left\{\begin{array}{r@{\;\mathrm{\quad if}\;}l}
-(v_A/B_0)\,\boldsymbol{b} & 0\le \vartheta < \pi/2\,, \\[1.0ex]
(v_A/B_0)\,\boldsymbol{b} & \pi/2 < \vartheta \le \pi\,,
\end{array}\right.
\end{equation}
where $\omega$ is the angular frequency of the plane wave, $\boldsymbol{k}$ the wave-vector indicating the propagation direction,
$v_{\rm A} = B_0 / \sqrt{4 \pi \rho}$ is the Alfvén speed, and $\vartheta$ is the angle between $\boldsymbol{e}_z$ and $\boldsymbol{k}$. We note that, if $\vartheta < \pi/2$, the perturbations $\boldsymbol{v}$ and $\boldsymbol{b}$ are anti-parallel, while they are parallel if $\vartheta > \pi/2$.
In the following, for simplicity, we only consider the case in which the wave travels in the direction of the magnetic field, that is, $0 \le\vartheta < \pi/2$. Furthermore, one can prove that both the magnetic field and the velocity perturbation are azimuthal with respect to the static magnetic field, $\boldsymbol{B}_0$, and the wave-vector $\boldsymbol{k}$ can point in any direction as depicted in Fig.\,\ref{fig:AlfvenWavesScheme}. This implies that the linearized version of the magnetic pressure is null because $p_{\rm m} = \boldsymbol{b}\cdot\boldsymbol{B}_0/8\pi = 0$. Since plasma pressure and gravity are not considered, the only driving force of torsional Alfv\'en waves is the magnetic tension.
%
\begin{figure}
\centering
\includegraphics[width=0.6\hsize]{Figures/AlfvenWave_Fig.pdf}
\caption{Torsional Alfv\'en wave propagating in the direction $\boldsymbol{k}$.
Perturbations in the magnetic field, $\boldsymbol{b}$, and in the plasma velocity field, $\boldsymbol{v}$, are normal to the plane spanned by $\boldsymbol{k}$ and $\boldsymbol{B}_0$.
}
\label{fig:AlfvenWavesScheme}
\end{figure}
%
The second result is the dispersion relation
%
\begin{equation}
\omega^2 = v_A^2 (\boldsymbol{k}\cdot\boldsymbol{e}_z)^2\,. \nonumber
\end{equation}
%
From this equation, it is possible to obtain the group velocity of an Alfvén wave packet, which corresponds to the velocity at which the energy is transmitted,
%
\begin{equation}
\boldsymbol{v}_{\text{group}} = \dfrac{\partial \omega}{\partial \boldsymbol{k}} = v_A \boldsymbol{e}_z
\quad\mathrm{if}\; 0\le \vartheta < \pi/2\,.
\end{equation}
%
It is equivalent to the Alfvén speed and energy is carried along the equilibrium magnetic field.
%
%
\subsection{Energy fluxes \label{subsec:energyflux}}
Alfvén waves can contribute to the energy transport in the solar atmosphere. To study this aspect, the MHD Poynting flux vector is employed,
%
\begin{equation}
\boldsymbol{S} = \dfrac{1}{4\pi} \boldsymbol{B} \times (\boldsymbol{v} \times \boldsymbol{B}) = \dfrac{1}{4\pi} (\boldsymbol{B} \cdot \boldsymbol{B})\,\boldsymbol{v} - \dfrac{1}{4\pi} (\boldsymbol{B} \cdot \boldsymbol{v})\, \boldsymbol{B}\,. \label{eq:poytingvector}
\end{equation}
%
Since we are interested in the vertical energy flux, the $z$-component of Eq.\,(\ref{eq:poytingvector}) is expanded following
\citet{2012ApJ...753L..22S},
%
\begin{equation}
S_z = \underbrace{\dfrac{1}{4\pi} v_z (B_x^2 + B_y^2)}_{S_z^{\rm v}}\, - \underbrace{\dfrac{1}{4\pi} B_z (v_x B_x + v_y B_y)}_{S_z^{\rm h}}, \label{eqn:z-Poynting}
\end{equation}
%
where the first term, $S_z^{\rm v}$, is related to vertical motions of horizontal magnetic field and the second term, $S_z^{\rm h}$, corresponds to the vertical flux generated by horizontal motions of magnetized plasma.
The contribution of the swirls to the mean Poynting flux over a given field of view of area $A_{\rm FOV}$ is then given by
%
\begin{equation}
\bar{S}_z = S_z \bar{N}_{\rm s} \frac{\pi\bar{r}_{\rm s}^2}{A_{\rm FOV}}\,, \label{eq:averagesflux}
\end{equation}
%
where $S_z$ is the mean contribution from a single swirl, $\bar{N}_{\rm s}$ is the average number of swirls that can be observed at any time in the given field of view, be it of a simulation or an observation, and $\bar{r}_{\rm s}$ is the average swirl radius.
A swirl can also produce a mechanical energy flux, which is given by
%
\begin{equation}
F_{z}^{\rm r} = \frac{1}{A_{\rm s}}\iint_{A_{\rm s}}\mathrm{d}x\mathrm{d}y\, \frac{1}{2} \rho v_{\rm r}^2 v_{\rm p} \,, \label{eq:mechflux_vh}
\end{equation}
%
where the integration is taken over the assumed circular area $A_{\rm s}$ associated with the swirl, $\rho = \rho(x,y)$ is the density of the plasma, and $v_{\rm r} = v_{\rm r}(x,y)$ is the rotational velocity. Furthermore, $v_{\rm p} = v_{\rm p}(x,y)$ is the propagation speed of the swirl in the vertical direction. For simplicity, we assume $A_{\rm s}$ to lie in a horizontal plane and $v_{\rm r}$ to be the velocity projected into this plane. Assuming that the flow rotates around a common axis, we can express the rotational velocity in terms of the vertical component of the swirling strength vector as $v_{\rm r} = \frac{1}{2}r \lambda_z$ (see Sect.\,\ref{subsec:vortex_detection}), hence
%
\begin{equation}
F_{z}^{\lambda} = \frac{1}{8 A_{\rm s}}\iint_{A_{\rm s}} \mathrm{d}x\mathrm{d}y\, \rho r^2 \lambda_z^2 v_{\rm p} \,, \label{eq:mechflux_vlambda}
\end{equation}
where $r = \sqrt{x^2 + y^2}$ is the distance from the center of the swirl and $\lambda_z = \lambda_z(x,y)$ is the associated vertical component of the swirling strength vector.
The contribution of swirling motions to the mean mechanical energy flux $\bar{F}_z$ is then
%
\begin{equation}
\bar{F}_z = F_z^{{\rm r},\lambda} \bar{N}_{\rm s} \frac{\pi\bar{r}_{\rm s}^2}{A_{\rm FOV}}\,,
\label{eq:averagemflux}
\end{equation}
%
where $F^{{\rm r},\lambda}_z$
is the mean mechanical flux density of a single swirl computed with either Eq.\,(\ref{eq:mechflux_vh}) or Eq.\,(\ref{eq:mechflux_vlambda}).
We notice that $S_z$, $F_z$, $\bar{F}_z$ and $\bar{S}_z$ are energy flux densities.
%
%
\subsection{Swirling strength} \label{subsec:vortex_detection}
A vortex or swirl can be intuitively described as the rotation of fluid parcels around a common axis. Despite this simple concept, a rigorous mathematical definition is still an open issue in fluid mechanics \citep[see, e.g.,][]{Kolar2007,2019JHyDy..31..205L}.
The vorticity is the classical quantity to describe rotational flows and it has been adopted to investigate vortex flows in numerical simulations of the solar convection zone \citep[see, e.g.,][]{1997A&A...328..229N}
and atmosphere \citep[see, e.g.,][]{2011A&A...526A...5S, 2013ApJ...776L...4S}.
However, \citet{1995JFM...285...69J}
showed that the vorticity is not a suitable tool for vortex identification since it cannot distinguish between non-rotational shear flows and actual vortices. This can lead to misidentifications in a dynamical and turbulent system such as the solar atmosphere. Therefore, we decided to employ the swirling strength criterion, which is not affected by shears \citep{1999JFM...387..353Z}.
The swirling strength is based on the eigenanalysis of the velocity gradient tensor $\mathcal{U}_{ij} \coloneqq \partial_j v_i$, that is, the Jacobian matrix of the local velocity field. If a vortex is present in the flow, the velocity gradient tensor can be locally diagonalized and it will exhibit one real and a pair of complex conjugated eigenvalues
%
\begin{eqnarray}
\mathcal{U} =
\underbrace{\vphantom{\begin{bmatrix}
\lambda_{\rm r} & 0 & 0\\
0 & \lambda_{\rm +}& 0 \\
0 & 0 & \lambda_{\rm -}
\end{bmatrix}}
\left[ \boldsymbol{u}_{\rm r}, \boldsymbol{u}_{\rm +}, \boldsymbol{u}_{\rm -}\right]
}_{\textstyle\mathcal{P}}
\underbrace{%
\begin{bmatrix}
\lambda_{\rm r} & 0 & 0\\
0 & \lambda_{\rm +}& 0 \\
0 & 0 & \lambda_{\rm -}
\end{bmatrix}}_{\textstyle\Lambda}
\underbrace{\vphantom{\begin{bmatrix}
\lambda_{\rm r} & 0 & 0\\
0 & \lambda_{\rm +}& 0 \\
0 & 0 & \lambda_{\rm -}
\end{bmatrix}}
\left[ \boldsymbol{u}_{\rm r}, \boldsymbol{u}_{\rm +}, \boldsymbol{u}_{\rm -}\right]^{-1}
}_{\textstyle\mathcal{P}^{-1}}\,, \label{eq:U_decomposition}
\end{eqnarray}
%
where $\lambda_{\pm} = \lambda_{\rm cr} \pm \mathrm{i}\lambda_{\rm ci}$ and $\lambda_{\rm r} \in \mathbb{R}$.
The swirling strength $\lambda$ is then defined through the imaginary part of the complex eigenvalues, $\lambda_{\rm ci}$. In this paper, we adopt the same convention as in \citet{2020A&A...639A.118C}
and define the swirling strength as $\lambda \coloneqq 2\lambda_{\rm ci}$. For a rigid body rotation, the period of revolution $T$ of the flow is then computed as $T = 4\pi/\lambda$ and the vorticity $\omega \coloneqq |\boldsymbol{\nabla}\times\boldsymbol{v}|$ equals the swirling strength $\lambda$.
Furthermore, the rotation axis and the orientation of the vortex can be inferred from the normalized eigenvector associated with the real eigenvalue, $\boldsymbol{u}_{\rm r}$. Hence, it is possible to define a swirling vector, $\boldsymbol{\lambda} \coloneqq \lambda \boldsymbol{u}_{\rm r}$, which carries the necessary information to characterize the vortex in three dimensions. For a detailed review of the swirling strength, the reader can refer to \citet{2020A&A...639A.118C}.
The swirling strength is invariant under Galilean transformations and can be used in the context of compressible \mbox{(magneto-)hydrodynamics} \citep{2009AIAAJ..47..473K}.
This criterion has already been successfully applied in studies regarding vortex flows in the solar atmosphere \citep[see][]{2011A&A...533A.126M, 2017A&A...601A.135K, 2020ApJ...894L..17Y}.
Furthermore, \citet{2020A&A...639A.118C}
derived the evolution equation for the swirling strength. A similar equation was already known for the vorticity \citep[see, e.g.,][]{2011A&A...526A...5S}
but, in this case, it can also be biased by shear flows. Therefore, the swirling equation is a valuable tool for the investigation of the physical processes responsible for the dynamics of vortices. It can be derived from a general momentum equation of \mbox{(magneto-)hydrodynamics}: For the specific case of ideal MHD, it can be formulated as
%
\begin{align}
\frac{{\rm D} \lambda }{{\rm D}t} =&
\vphantom{\left\{ \mathcal{P}^{-1} \left[ \nabla \bigg( \frac{1}{\rho}\nabla p_{\rm g} \bigg) \right] \mathcal{P} \right\}}
- 2\lambda\lambda_{\rm cr} & T^1 & \nonumber \\
%
& - 2\rm{Im}\left\{ \mathcal{P}^{-1} \left[ \nabla \otimes \bigg( \frac{1}{\rho}\nabla \textit{p}_{\rm g} \bigg) \right] \mathcal{P} \right\}_{22} & T^2 & \nonumber \\
%
& - 2\rm{Im}\left\{ \mathcal{P}^{-1} \left[ \nabla \otimes \bigg( \frac{1}{\rho} \nabla \textit{p}_{\rm m} \bigg) - \bigg(\nabla\frac{1}{\rho} \bigg)\otimes \frac{ (\boldsymbol{B}\cdot\nabla) \boldsymbol{B}}{4\pi} \right] \mathcal{P} \right\}_{22} & T^3 & \nonumber \\
%
& + 2\rm{Im}\left\{ \mathcal{P}^{-1} \left[ \frac{1}{\rho} \nabla \otimes \frac{( \boldsymbol{B} \cdot \nabla ) \boldsymbol{B}}{4\pi} \right] \mathcal{P} \right\}_{22} & T^4 & \nonumber \\
%
& \vphantom{\left\{ \mathcal{P}^{-1} \left[ \nabla \bigg( \frac{1}{\rho}\nabla p_{\rm g} \bigg) \right] \mathcal{P} \right\}}
- 2\rm{Im}\left\{ \mathcal{P}^{-1} \bigg[ \nabla \otimes \Big( \nabla \Phi \Big) \bigg] \mathcal{P} \right\}_{22}\,, & T^5
& \label{eq:swirling_eq}
\end{align}
%
where $\lambda_{\rm cr}$ is the real component of the complex eigenvalues of $\mathcal{U}$, $\mathcal{P}$ and $\mathcal{P}^{-1}$ are, respectively, the matrix composed of the eigenvectors of $\mathcal{U}$ and its inverse as shown in Eq.\,(\ref{eq:U_decomposition}), $p_{\rm g}$ is the atmospheric gas pressure, $p_{\rm m} = B^2/8\pi$ represents the pressure owing to the magnetic field, and $\Phi$ corresponds to a potential of conservative forces. We notice that the terms between curly brackets are matrices since the symbol $\otimes$ denotes the tensor product between two vectors. We are only interested in the $(2,2)$ component of the resulting matricial multiplication.
Finally, the material derivative on the left-hand side of Eq.\,(\ref{eq:swirling_eq}) is defined as
%
\begin{equation}
\frac{{\rm D} \lambda }{{\rm D}t} = \frac{\partial \lambda}{\partial t} + \left(\boldsymbol{v}\cdot\nabla\right) \lambda\,, \nonumber
\end{equation}
%
where $\partial \lambda/\partial t$ defines the local production of swirling strength, while $\left(\boldsymbol{v}\cdot\nabla\right) \lambda$ accounts for the swirling strength advected with the plasma.
Following \citet{2020A&A...639A.118C},
we give a physical interpretation for each one of the terms appearing in Eq.\,(\ref{eq:swirling_eq}). The first term, $T^1$, is a stretching term, while $T^2$ and $T^3$ are related to the baroclinic effects: The former is a pure hydrodynamical process and the latter is due to magnetic fields. Then, $T^4$ is associated with magnetic tension effects and $T^5$ describes the generation of swirling strength by conservative forces.
For the statistical analyses, plots in Sect.~\ref{sec:results}, and the appendices, a swirling strength threshold is applied to ease the interpretation and reduce noise \citep{1999JFM...387..353Z}.
It is also motivated by the request that swirls perform a significant fraction of a full rotation over the time span of the simulation.
Therefore, swirls having a period longer than $10\,\mathrm{min}$, that is, $\lambda < 2.09\,\times\,10^{-2}\,\mathrm{Hz}$, are discarded.
%
%
%
\begin{figure*}
\resizebox{\hsize}{!}{\includegraphics{Figures/BoxRendering_SS-beta1-tau1_v50_t5900_y7000_new_crop-eps-converted-to.pdf}}
\caption{
Time instant of the swirling strength $\lambda$ in a slice of dimension $9.6\times2.6\times2.8\,\mathrm{Mm}^3$ of the whole physical domain of the numerical simulation. The red contour indicates the surface of optical depth $\tau_{500} = 1$. The yellow contour corresponds to the isosurface of plasma-$\beta = 1$. Swirls with a period larger than $10\,\mathrm{min}$ ($\lambda < 2.09\,\times\,10^{-2}\,\mathrm{Hz}$) are not shown and the values $\lambda > 1\,\mathrm{Hz}$ are saturated. This rendering has been produced with ParaView \citep{Ahrens+al2005}.
}
\label{fig:SwirlsDistribution_TauBeta}
\end{figure*}
\subsection{Magnetic swirling strength}
\label{subsec:magneticswirlingstrength}
Torsional Alfvén waves are characterized by azimuthal perturbations in the vertically directed magnetic field lines. In order to identify this torsion in numerical simulations, we define the criterion of magnetic swirling strength, $\lambda^{B}$. This quantity is derived in the same way as the swirling strength, but the matrix to be diagonalized is the magnetic gradient tensor, $\mathcal{B}_{ij} \coloneqq \partial_j B_i$. Hence, the magnetic swirling strength is defined through the imaginary component of the complex eigenvalues of $\mathcal{B}$, that is, $\lambda^{B} \coloneqq 2\lambda^{B}_{\rm ci}$. Following the same reasoning as for the swirling strength, the magnetic swirling vector $\boldsymbol{\lambda}^{B}=\lambda^{B} \boldsymbol{u}^{B}_{\rm r}$ can be defined, which describes the three-dimensional twisting of the magnetic field.
It is worth noticing that the units of the magnetic swirling strength are $[\mathrm{G\,cm^{-1}}]$ and that therefore this quantity does not describe a rotational flow of the magnetic field lines, but rather their twisting around the axis represented by the real eigenvector, $\boldsymbol{u}_{\rm r}^{B}$. Thus, a magnetic field line rotating together with the plasma flow in a rigid body fashion will produce no magnetic swirling strength, while a magnetic field line with an helical structure will present a finite value of magnetic swirling strength. Consequently, this criterion is well suited to detect azimuthal torsions in the magnetic field and, therefore, imprints of torsional Alfvén waves.
For torsional Alfvén waves in a static, incompressible plasma dominated by the magnetic field, we can find a simple relation between the swirling strength $\lambda$ and the magnetic swirling strength $\lambda^{B}$. By taking the tensor product between $\nabla$ and Eq.\,(\ref{eqn:PerturbationRelation}) it is possible to obtain
%
\begin{equation}
\mathcal{U} = -\dfrac{1}{\sqrt{4\pi\rho}}\,\mathcal{B}
\quad\mathrm{if}\;0\le \vartheta < \pi/2\,,
\end{equation}
%
which can be diagonalized on both sides since $1/\sqrt{4\pi\rho}$ is a real scalar.
Given the definitions of $\lambda$ and $\lambda^{B}$, we get
%
\begin{equation}
\lambda = -\dfrac{1}{\sqrt{4\pi\rho}}\,\lambda^{B} \quad\mathrm{if}\; 0\le \vartheta < \pi/2\,,
\label{eq:relationlambdas}
\end{equation}
%
which simply states that, for a torsional Alfv\'en wave propagating in direction of the magnetic field ($\vartheta < \pi/2$), the swirling strength and the magnetic swirling strength are proportional to each other and have opposite sign (see also Fig.\,\ref{fig:AlfvenWavesScheme}).
\section{Numerical simulation}\label{sec:numerics}
The high-resolution and high-cadence simulations analyzed in this study have been carried out with the radiative MHD code CO$5$BOLD \citep{2012JCoPh.231..919F},
which solves the coupled system of compressible MHD equations in an external gravitational field and includes nonlocal, frequency-dependent radiative transfer.
The simulations are performed on a three-dimensional Cartesian grid of size $9.6\times9.6\times2.8\,\mathrm{Mm^3}$, with a grid cell size of $10\,\mathrm{km}$ in each spatial dimension. The vertical component ranges from about $1300\,\mathrm{km}$ below the optical surface $\tau_{500} = 1$ to $1500\,\mathrm{km}$ above it. Hence, the simulation domain encompasses a small volume near the solar surface, ranging from the surface layers of the convection zone to the middle chromosphere. The gravitational field is uniform and vertical with a value of $\log{(g)} = 4.44$.
The initial condition of the simulation is given by a previously relaxed hydrodynamical model, to which a uniform and vertical magnetic field of $50\,\mathrm{G}$ was superimposed. The system is then advanced with the MHD module of CO$5$BOLD \citep{2005ESASP.596E..65S}
until relaxation. The lateral boundary conditions are periodic for both plasma and magnetic fields, while the top and bottom ones are open under the condition that the net mass flux at the bottom is zero. An entropy inflow is supplied to maintain an average surface effective temperature of $T_{\rm eff} = 5770\,\mathrm{K}$. The magnetic field is constrained to be vertical at the top and bottom boundaries. More details on the simulation setup are given in
\citet[][Sect.\,2]{doi:10.13097/archive-ouverte/unige:115257},
in particular run \texttt{d3gt57g44v50fc}.
For the present study, a time series of $441$ three-dimensional data cubes with a cadence of $2\,\mathrm{s}$, amounting to about $5\,{\rm TB}$ of data, is analyzed. It starts from $t\!=\!5520\,{\rm s}$ of
\texttt{d3gt57g44v50fc}.
This relatively high-cadence has proven to be necessary to be able to capture the evolution and the dynamics of the detected vortices. The series spans a period of approximately $15\,{\rm min}$.
\begin{figure*}
\centering
\resizebox{\hsize}{!}{
\begin{minipage}[b]{.487\linewidth}
\includegraphics[width=\hsize]{Figures/Correlation_Spearman_TimeAverage_shaded_a.pdf}
\label{fig:SpearmanR}
\end{minipage}
\hspace{0.3cm}
\begin{minipage}[b]{.518\linewidth}
\includegraphics[width=\hsize]{Figures/Correlation_meanBandSS_shaded_b.pdf}
\label{fig:Mean_BandSS}
\end{minipage}
}
\caption{Statistical properties of swirling motions and mean physical quantities as a function of height $z$. (\emph{a}) Spearman's correlation coefficient $r$ between the swirling strength $\lambda$ and the magnetic field strength $|\boldsymbol{B}|$, $r(\lambda, |\boldsymbol{B}|)$ in blue, and between the swirling strength $\lambda$ and the magnetic swirling strength $\lambda^{B}$, $r(\lambda, \lambda^{B})$ in red. The green curve represents $r(\lambda, |\boldsymbol{B}|)$ calculated in regions in which $\beta < 1$ and above $z=0$ only. The curves are means over eleven different time instants with a cadence of one minute and the shaded areas correspond to the standard deviations obtained from the temporal variations. Data points with $\lambda < 2.09\,\times\,10^{-2}\,\mathrm{Hz}$ are excluded. (\emph{b}) Mean magnetic field strength $\langle|\boldsymbol{B}|\rangle$ (red) and mean swirling strength $\langle\lambda\rangle$. The dark blue and red curves refer to the same MHD time instants as in (\emph{a}), while the light blue curve refers to $14$ time instants of a purely hydrodynamical simulation.
}
\label{fig:Correlation_Spearman}
\end{figure*}
\section{Results and discussion}\label{sec:results}
This section starts by taking a global view on the simulation, evidencing the correlation between vortices and magnetic field. It then concentrates on a single swirl event, which is shown to be a propagating torsional Alfv\'en pulse, and concludes with an estimate of the energy carried by the pulse and by swirling motions in the entire simulation domain. Both global and local analyses are supplemented by appendices.
%
%
\subsection{Relation between swirls and magnetic fields} \label{subsec:DistSSinSim}
Figure\,\ref{fig:SwirlsDistribution_TauBeta} shows a volume rendering of the distribution of swirling strength seen across a vertical section through the simulation domain at an arbitrary location and time instant. Three regions can be distinguished. The first one extends from the bottom of the box to $z\approx\!0\,\mathrm{km}$, the height that roughly corresponds to the surface of optical depth $\tau_{500}=1$ indicated in red color. This region coincides with the convection zone and it is characterized by an almost isotropic distribution of swirls, mostly produced by the convective and turbulent motions of the plasma.
At the photospheric and chromospheric level, there exist two distinct regions: The first one shows tubes of intense swirling strength aligned with the magnetic field, while the other one is characterized by a severe depletion of swirls. The yellow, funnel-shaped surface, which sharply separates these two regions, is given by the isosurface of plasma-$\beta = 1$, where $\beta = p_{\rm g}/p_{\rm m}$ is the ratio between gas pressure $p_{\rm g}$ and magnetic pressure $p_{\rm m}$. This clear separation shows evidence of strong coupling between swirling motions and magnetic fields. Above the yellow surface and within the funnels $\beta < 1$, therefore, the dynamics are governed by the magnetic fields, while the region of depleted swirling strength is characterized by $\beta > 1$, meaning that the gas pressure dominates.
We notice, however, that the region in between the magnetic flux concentrations in the photosphere is not completely void of swirls. There exist arches of swirling strength that are low lying and arches that reach farther out, as well as vertically directed swirls, which rise above the solar surface as it can be seen for example in the ranges $3000\,\mathrm{km} < x < 5000\,\mathrm{km}$ and $7000\,\mathrm{km} < x < 9000\,\mathrm{km}$ in Fig.\,\ref{fig:SwirlsDistribution_TauBeta}. This kind of swirls and swirl arches were also noted and described by
\citet{2010NewA...15..460M} and
\citet{2011A&A...533A.126M}
and can probably to a great extent be identified with horizontal vortex tubes that form at the edges of granules
\citep{2010ApJ...723L.180S}.
Appendix\,\ref{app:comparison} compares the time instant of Fig.\,\ref{fig:SwirlsDistribution_TauBeta} with a magnetic field-free simulation and with a simulation of initial magnetic field strength of $200\,{\rm G}$. It also quantifies the orientation of swirls as a function of height.
\begin{sidewaysfigure*}[p]
\centering
\includegraphics[width=23cm]{Figures/AlphaMorphology.pdf}
\caption{
Time sequence of a single swirl event from $t=5760\,\mathrm{s}$ to $t=6160\,\mathrm{s}$. \emph{From top to bottom row}: Vertical component of the magnetic field $B_z$ at $z=0\,\mathrm{km}$, continuum intensity $I$, vertical velocity $v_z$ at $z=700\,\mathrm{km}$, vertical component of the swirling vector $\lambda_z$ at $z=700\,\mathrm{km}$, vertical component of the magnetic swirling vector $\lambda^{B}_z$ at $z=700\,\mathrm{km}$, and the bin-5 intensity $I_5$. Maps of $\lambda_z$ and $\lambda_z^{B}$ also show the streamlines of the velocity field and the magnetic field projected into the horizontal plane at $z=700\,\mathrm{km}$, respectively.
}
\label{fig:SingleSwirlEvent}
\end{sidewaysfigure*}
%
Next, a statistical study is performed, with the purpose of evaluating the degree of correlation between the swirling strength $\lambda$, the magnetic field strength $|\boldsymbol{B}|$, and the magnetic swirling strength $\lambda^{B}$.
To do so, Spearman's $r$ rank coefficient is employed, which assesses how well two quantities are monotonically correlated: $r=1$ indicates perfect rank correlation, $r=-1$ stands for rank anti-correlation, while $r=0$ means that the two quantities are uncorrelated. Figure\,\ref{fig:Correlation_Spearman}a shows this coefficient, as a function of height $z$, for the correlation between $\lambda$ and $|\boldsymbol{B}|$, $r(\lambda,|\boldsymbol{B}|)$, and between $\lambda$ and $\lambda^{B}$, $r(\lambda,\lambda^{B})$.
From Fig.\,\ref{fig:Correlation_Spearman}a, we see that a strong rank correlation exists between swirling strength and magnetic swirling strength, $r(\lambda,\lambda^{B})$, in particular in the near surface convection zone and in the photosphere ($r \approx\!0.8$). Into the chromosphere, the correlation coefficient decreases to $r\approx\!0.6$. Therefore, we conclude that a strong correlation exists between $\lambda$ and $\lambda^{B}$ in the solar atmosphere and, consequently, also between plasma vortices and torsional perturbations in the magnetic field. The lower values in the chromosphere can be explained by the upper boundary condition of vanishing horizontal magnetic field component, which prevents torsional perturbations at the very top of the computational domain.
Given Eq.\,(\ref{eq:relationlambdas}), one can tentatively link this high correlation to torsional Alfvén waves, which is the topic of Sect.\,\ref{subsec:alfvén_pulses}.
Concerning the correlation between swirling strength and magnetic field strength, $r(\lambda,|\boldsymbol{B}|)$, Spearman's coefficient is $r\approx\!0.35$ in the convection zone, drops to $r\approx\!0.3$ in the photosphere and rises again to $r\approx\!0.4$ around the classical temperature minimum near $z=500$\,km before decreasing to $r\approx\!0.0$ in the middle chromosphere. The behavior of the curve is similar to $r(\lambda,\lambda^{B})$, but the overall value is always smaller and the variation is more pronounced. The low value in the convection zone can be explained by the turbulent motion of the plasma. Indeed, the source of vortices is not related to the strength of the magnetic field, which is almost everywhere much smaller than the kinetic equipartition value in the convection zone. Swirls that are carried from the convection zone to the photosphere rapidly lose angular velocity owing to the steep decrease in mass density and correspondingly strong expansion of the plasma \citep{1997A&A...328..229N}.
This explains the rapid decrease in the average swirling strength above $z=0$\,km shown in Fig.\,\ref{fig:Correlation_Spearman}b. Together with the many nonmagnetic or weakly magnetized swirl arches in between the magnetic flux concentrations, this reduces the correlation between swirling strength and magnetic field in the photosphere.
However, the correlation stays high for swirls within the magnetic funnels ($\beta < 1$), which is demonstrated by the green curve in Fig.\,\ref{fig:Correlation_Spearman}a. While weak-field swirl-arches become less abundant with height, the correlation rises throughout the photosphere, reaching a maximum of $0.43$ at $z=500$\,km. Higher up in the atmosphere, the correlation $r(\lambda,|\boldsymbol{B}|)$ rapidly drops because the magnetic field becomes increasingly homogeneous and its strength drops, while the swirling strength steeply increases and remains highly structured as it can be seen from Fig.\,\ref{fig:Correlation_Spearman}b and Fig.\,\ref{fig:SwirlsDistribution_TauBeta}.\footnote{Despite the close to homogeneous magnetic field in the upper layers of the simulation domain, the swirling strength remains highly structured because the magnetic field dominates the dynamics everywhere and can produce and host swirls on small-scales; a finding recently also established by \citet{2020ApJ...894L..17Y} with MURaM simulations and by \citet[][Fig.\,5]{2020A&A...639A.118C} with CO$5$BOLD simulations.} This behavior is distinctly different from a magnetic field-free simulation in which the swirling strength does not show this steep increase (light blue curve in Fig.\,\ref{fig:Correlation_Spearman}b).
%
%
\subsection{Evidence of Alfvén pulses}\label{subsec:alfvén_pulses}
\begin{figure*}
\resizebox{\hsize}{!}
{\includegraphics{Figures/Alpha_SSDistr_Upward_cbar_arrows.pdf}}
\caption{
Time sequence of the swirling strength distribution. Each panel represents a bi-dimensional histogram normalized to the maximum density fraction at each height level $z$. The bin sizes are $\Delta z = 10$\,km and $\Delta\lambda = 3.62\,\times10^{-3}\,{\rm Hz}$. These histograms refer to stacks of quadratic, plane-parallel cross sections of $1000\,\mathrm{km}$ side length centered
on the swirl event of Fig.\,\ref{fig:SingleSwirlEvent}. Data points with $\lambda < 2.09\,\times\,10^{-2}\,\mathrm{Hz}$ are excluded. The black arrows point to the upward propagation of a local peak of swirling strength.
}
\label{fig:SSDIstrUpwardPropagation}
\end{figure*}
%
\begin{figure}
\resizebox{\hsize}{!}{
\includegraphics[width=\hsize]{Figures/Alpha_TimeDist_SS-MS_abc.pdf}
}
\caption{
Time-distance diagrams of (\emph{a}) the swirling strength $\lambda$, (\emph{b}) the vertical component of the swirling vector $\lambda_z$, and (\emph{c}) the vertical component of the magnetic swirling vector $\lambda^{B}_z$ for the swirl event shown in Fig.\,\ref{fig:SingleSwirlEvent}. Values at each time step and height are averaged over a finite horizontal plane of $150\,\mathrm{km}$ side length. The paths of two test particles moving with Alfv\'en plus bulk speed (green) and with sound plus bulk speed (red) along the swirl are shown in panel (b).
}
\label{fig:SingleSwirlEvent_TDdiagrams}
\end{figure}
%
In this subsection, we study in detail one of the swirl events that occurred in the course of the simulation with the $50\,{\rm G}$ initial field strength. A list of eight supplementary events is presented in Appendix\,\ref{app:supplementary_swirls}. The six rows of Fig.\,\ref{fig:SingleSwirlEvent} show close-ups on various observable quantities that characterize the event. The cadence of the close-ups is $50\,\mathrm{s}$. The top row corresponds to a proxy of a magnetogram. It exhibits a positive polarity (upwardly pointing) magnetic flux concentration with $B_z\approx\!1500$\,G at $z=0\,\mathrm{km}$ and around $t = 5910\,{\rm s}$, while it has merely half that strength before and after this maximum. The flux concentration is responsible for the bright knot located within the integranular vertex visible in the continuum-intensity maps of the second row. Next to this flux concentration is a weaker one of inverse polarity.
The third, fourth, and fifth rows show, respectively, the vertical component of the velocity field, $v_z$, the vertical component of the swirling vector, $\lambda_z$, and the vertical component of the magnetic swirling vector, $\lambda_z^{B}$, all in a cross section at a height of $z = 700\,\mathrm{km}$. We see the development of a clockwise rotating swirl (negative $\lambda_z$) from the start of the time series, best visible from $t \approx\!5860\,\mathrm{s}$ onward in both $v_z$ and $\lambda_z$ (third and fourth rows). The positive polarity magnetic flux concentration in the top row and the BP in the second row appear to be rotating clockwise too, that is, in the same direction as the plasma does. The rotation is unidirectional (clockwise) over the full time period from $5760\,\rm{s}$ to $6160\,\rm{s}$, that is, over $400\,\rm{s}$. The fifth row exhibits that there also exists an azimuthal perturbation in the magnetic field but opposite to the velocity field. This aspect will be discussed in the paragraph after next.
Finally, the last row shows the maps of the synthetic bin-5 intensity, $I_5$. It corresponds to the fifth opacity band of the non-gray radiative transfer used in the simulation and represents an average of intensities from strong spectral lines of large opacities \citep{Ludwig1992}. The $\tau_{500} = 1$ level computed with this opacity bin is located in the upper photosphere to lower chromosphere, and, therefore, $I_5$ can be taken as a proxy for chromospheric line core intensities.
In these maps, we recognize a chromospheric swirl with a maximal diameter of $\approx\! 1.4\,\mathrm{arcsec}$, which is in the range of sizes of the ones reported by \citet{2009A&A...507L...9W}.
From Fig.\,\ref{fig:SingleSwirlEvent} we can deduce an important aspect of this event: The vertical component of the swirling vector $\lambda_z$ and of the magnetic swirling vector $\lambda_z^{B}$ have opposite signs, that is, opposite orientations. In fact, it can be seen from the streamlines in the corresponding panels that the plasma is rotating in a clockwise fashion (negative swirling strength), while the magnetic field lines are counterclockwise twisted (positive magnetic swirling strength). This is a characteristic of upwardly propagating torsional Alfvén waves in a positive polarity (upwardly directed) magnetic field, as was pointed out in Sect.\,\ref{sec:theory}. Therefore, the fourth and fifth rows of Fig.\,\ref{fig:SingleSwirlEvent} represent the first piece of evidence of a perturbation similar to a torsional Alfvén wave in this swirl event.
The time sequence of Fig.\,\ref{fig:SingleSwirlEvent} shows the evolution of the swirl in a horizontal section at $z=700\,\mathrm{km}$ only. In order to gain a perspective of its evolution in the vertical direction, Fig.\,\ref{fig:SSDIstrUpwardPropagation} gives the temporal evolution of the histograms of the swirling strength as a function of $z$. The histograms refer to a region of interest of $1.0 \times 1.0\,\mathrm{Mm}^2$ centered on the swirl and are taken with a cadence of $10\,\mathrm{s}$. This region of interest is almost as large as the individual maps of Fig.\,\ref{fig:SingleSwirlEvent}: It is large enough to ensure that the swirl stays within this region at all heights $0 < z < 1500\,{\rm km}$ and small enough to avoid disturbances from neighboring swirling motions.
From Fig.\,\ref{fig:SSDIstrUpwardPropagation} one recognizes a decline of swirls (both in density and in strength) at photospheric levels due to reasons already mentioned in Sect.\,\ref{subsec:DistSSinSim}.
However, there is a small bump of large swirling strengths appearing at $(t,z) = (5880\,\mathrm{s}, 400\,\mathrm{km})$, marked with an arrow. It grows and moves upward in time, proceeding $\approx\! 100\,\mathrm{km}$ every $10\,\mathrm{s}$. The bump that started in the photosphere then seems to merge with other local over-densities at $t=5930\,\mathrm{s}$ to form a large peak around $z=1000\,\mathrm{km}$, which continues its ascent after $t=5940\,\mathrm{s}$. We interpret the ascension of this over-density of large swirling strength to be the signature of an upwardly propagating swirl, which starts in the photosphere at $z \approx\!400\,\mathrm{km}$ and becomes manifest at chromospheric levels at around $t=5910\,\mathrm{s}$, as was already visible in Fig.\,\ref{fig:SingleSwirlEvent}.
To further investigate the upward propagation of the swirling motion, Fig.\,\ref{fig:SingleSwirlEvent_TDdiagrams} presents three time-distance diagrams. In these diagrams, the quantities at each height and instant (time-distance point) are averages over a horizontal plane cross-section of side length $150\,\mathrm{km}$ centered on the magnetic swirl event shown in Fig.\,\ref{fig:SingleSwirlEvent}. This side length corresponds to approximately $0.2\,{\rm arcsec}$, which is distinctly smaller than the side length of the maps shown in Fig.\,\ref{fig:SingleSwirlEvent}. Panel (a) shows the evolution of the swirling strength $\lambda$ and (b) the vertical component of the swirling vector $\lambda_z$.
One easily recognizes a crest of large swirling strength in both panels, starting at $(t,z)=(5830\,{\rm s}, 0\,{\rm km})$ and continuing up to at least $(t,z)=(5970\,{\rm s}, 1050\,{\rm km})$. The similarity between these two diagrams tells us that the swirl axis is to a large degree vertically directed. The crest in $\lambda_z$ fades above $z\approx\! 900\,{\rm km}$, while that of $\lambda$ can be followed further up to $z\approx\! 950\,{\rm km}$.
A more careful analysis reveals that this disappearance is not due to a fading of the swirl itself but to the bending of the swirl out of the strictly vertical detection slit of $150\times150\,\mathrm{km}^2$ cross section used to establish these time-distance plots. In fact, from
Fig.\,\ref{fig:SSDIstrUpwardPropagation}, which is based on a wider cross-section, it can be readily seen that the swirl propagates up to the top of the computational domain, showing still a peak in swirling strength at $(t,z)=(5960\,{\rm s}, 1200\,{\rm km})$.
From panels (a) and (b) of Fig.\,\ref{fig:SingleSwirlEvent_TDdiagrams} it can firstly be seen that this swirl is not a rigidly rotating magnetic structure but it propagates wave-like throughout the atmosphere in the upward direction. However, at any time, there is swirling motion distributed over a large height range and only the local maximum (the crest) of swirling strength is confined to a narrower height range. Second, we notice from panel (b) that the swirling strength is always negative, thus the rotation of the swirl is clockwise. Only in chromospheric layers above $800\,{\rm km}$ there are traces of counterclockwise motion, but these do not arise from a precursory wave train starting from the photosphere. Thus, we do not deal with an oscillatory wave but with an unidirectional pulse. In fact, there exists a sequence of equally directed pulses as it can be best seen from panel (b) at around $z=600\,{\rm km}$ with a sequence of local maxima in swirling strength at times $t=5920\,{\rm s}$, $5970\,{\rm s}$, and $6020\,{\rm s}$.
Panel (c) of Fig.\,\ref{fig:SingleSwirlEvent_TDdiagrams} shows the evolution of the vertical component of the magnetic swirling vector $\lambda_z^{B}$. From it, it is possible to observe that the propagation of negative $\lambda_z$ occurs in parallel with the propagation of positive $\lambda_z^{B}$. This behavior confirms that the vortex flow in the plasma is linked to a twist of opposite orientation in the magnetic field, which is a characteristic of torsional Alfvén waves propagating in direction of the magnetic field, as was already pointed out in the discussion of Fig.\,\ref{fig:SingleSwirlEvent}.
The value of $\lambda_z^{B}$ is large in the photosphere and then fades away in the chromosphere for three reasons. First, the magnetic field is weaker and therefore the value of $\lambda_z^{B}$ is smaller higher up in the atmosphere than in the photosphere. Second, the magnetic swirl becomes curved in the chromosphere bending outside of the detection slit used for establishing this time-distance diagram. Finally, the boundary conditions impose the magnetic field to be vertically directed at the top boundary, hence twists in the upper layers are suppressed. We also note that the twist angle of the magnetic field (inclination with respect to the vertical direction) does not have to be large but can be imperceptibly small where $\beta \ll 1$.
Next, the propagation speed of the swirl is investigated. The green dotted curve of Fig.\,\ref{fig:SingleSwirlEvent_TDdiagrams}b shows the path of an imaginary test particle moving at speed $v(z) = \langle v_{\rm A}(z) \rangle \cdot \cos{(\langle \theta \rangle)} + \langle v_z(z) \rangle$, that is, with the mean local Alfvén speed plus the mean local, vertical plasma bulk-speed, starting in the photosphere at $z\!\approx\! 100\,{\rm km}$ from the ridge that leads highest up. Here, the mean is taken over the horizontal plane cross-section of side length $150\,\mathrm{km}$ of the detection slit and $\langle\theta\rangle$ is the mean inclination of the swirl, that is, the angle of the swirling strength vector with respect to the $z$-axis. The red particle moves with speed $v(z) = \langle c_{\rm s}(z) \rangle \cdot \cos{(\langle \theta \rangle)} + \langle v_z(z) \rangle$, which corresponds to the mean sound plus plasma bulk-speed. It starts at the same position as the green curve.
We notice how the green trajectory perfectly follows the propagation of the vertical component of the swirling vector, which means that the swirl is moving with the local Alfvén plus bulk speed. Different from that, the red curve strongly deviates from the ridge of maximal vertical swirling strength starting from approximately $z=400\,{\rm km}$, indicating that the swirl does not propagate with sound speed. This is the second piece of evidence of the Alfvénic nature of the swirl event.
It is essential to notice that the vertical component of both the swirling vector and the magnetic swirling vector do not change orientation in time. This means that the swirling motion, as well as the perturbation of the magnetic field, are unidirectional. Therefore, one must think of a torsional Alfvén pulse instead of a generic, oscillatory, torsional Alfvén wave. The existence of such torsional pulses in the solar atmosphere has been suggested by \citet{2019NatCo..10.3504L}
using numerical simulations with a torsional perturbative driver acting on a preexisting magnetic flux tube. Here, it has been demonstrated that such torsional Alfvén pulses also emerge from a fully self-consistent, realistic, numerical simulation of the solar atmosphere.
The present section focused on a single, isolated swirl event. More often, simulations show swirling motions in the chromosphere to appear as a result of the superposition of two or more individual swirls. Examples of the superposition of swirls are displayed and discussed in Appendices\,\ref{app:superposition_of_swirls} and \ref{app:supplementary_swirls}.
%
%
\subsection{Dynamics of the pulse}
\begin{figure*}[h!]
\resizebox{\hsize}{!}
{\includegraphics{Figures/Alpha_TimeDist_SingleLOS_Row1_abcde.pdf}} \\[1em]
\resizebox{\hsize}{!}
{\includegraphics{Figures/Alpha_TimeDist_SingleLOS_Row2_fghij.pdf}}
\caption{
Time-distance diagram of (\emph{a}) the swirling strength $\lambda$ and of the vertical component of the following quantities: (\emph{b}) swirling vector $\lambda_z$, (\emph{c}) magnetic swirling vector $\lambda_z^{B}$, (\emph{d}) partial derivative of the swirling vector ${\rm \partial}\lambda_z/{\rm \partial}t$, (\emph{e}) advection term $T_z^{\rm adv}=[(\boldsymbol{v}\cdot\nabla)\boldsymbol{\lambda}]_z$, (\emph{f}) stretching term $T_z^1$, (\emph{g}) hydrodynamical baroclinic term $T_z^2$, (\emph{h}) magnetic baroclinic term $T_z^3$, (\emph{i}) sum of baroclinic terms $T_z^{\rm bar} = T_z^2 + T_z^3$, and (\emph{j}) magnetic tension term $T_z^4$. The black dashed contours correspond to the $\lambda_z = -0.1\,{\rm Hz}$ contours, which highlight the propagation of the Alfvén pulse. All plots are relative to the swirl event shown in Fig.\,\ref{fig:SingleSwirlEvent} and are derived from a fixed, one pixel wide vertical slit through the center of that swirl.
}
\label{fig:SingleSwirlEvent_TDdiagrams_SingleLOS}
\end{figure*}
%
This section focuses on the dynamics of the magnetic swirl shown in Fig.\,\ref{fig:SingleSwirlEvent}, in particular, on the physical processes responsible for the production and propagation of the swirling motion and Alfvén pulse. For this purpose, we employ the swirling equation, Eq.\,(\ref{eq:swirling_eq}), to reveal the local production of the vertical component of the swirling vector by the various source terms. Each one of these source terms is related to a different physical mechanism, as described in Sect.\,\ref{subsec:vortex_detection}.
Figure\,\ref{fig:SingleSwirlEvent_TDdiagrams_SingleLOS} shows time-distance diagrams of various quantities. Different from Fig.\,\ref{fig:SingleSwirlEvent_TDdiagrams}, the time-distance slit in this case is given by a single, vertical column of computational cells (single line-of-sight), located in the center of the magnetic swirl. Panel (a) shows the time-distance plot for the swirling strength $\lambda$, (b) the vertical component of the swirling vector $\lambda_z$, and (c) the vertical component of the magnetic swirling vector $\lambda_z^{B}$, similar to Fig.\,\ref{fig:SingleSwirlEvent_TDdiagrams}. However, with the single line-of-sight slit, we see two separate pulses of swirling strength, both with the same sense of rotation. They propagate upward at local Alfvén speed: the first starting at $(t,z) = (5800\,{\rm s}, 0\,{\rm km})$, the second at $(t,z) = (5850\,{\rm s}, -50\,{\rm km})$. These pulses are well paired with perturbations in the magnetic field, as can be seen from panel (c). These plots reproduce in more detail what was already observed from Fig.\,\ref{fig:SingleSwirlEvent_TDdiagrams}.
Figures \ref{fig:SingleSwirlEvent_TDdiagrams_SingleLOS}d and \ref{fig:SingleSwirlEvent_TDdiagrams_SingleLOS}e show the time-distance diagrams of the local production and the advection of vertical component of the swirling vector, $\partial \lambda_z / \partial t$ and $[(\boldsymbol{v}\cdot\nabla)\boldsymbol{\lambda}]_z$,
respectively. Together, these two quantities constitute the material derivative that appears on the left-hand side of Eq.\,(\ref{eq:swirling_eq}). We notice that the local production of negative swirling strength between $-200\,\mathrm{km} \lesssim z \lesssim 200\,\mathrm{km}$ and $5760\,\mathrm{s} \lesssim t \lesssim 5860\,\mathrm{s}$ in panel (d) corresponds to the appearance of negative swirling strength in panel (b). The black dashed lines are contours of $\lambda_z = -0.1\,{\rm Hz}$. They are intended to visualize the boundaries of the two pulses and guide the eye in identifying the relevant source terms. Also from panel (d), we notice that the two upwardly propagating pulses are characterized by the production of negative swirling strength on the left (leading) boundary of the two stripes, which is compensated for, approximately $20\,{\rm s}$ later, by the production of positive swirling strength on the right (trailing) boundary. We conclude that these two sources cause the upward evolution of the swirl.
Contributions from the advection term to the local production of swirling strength are mostly subdominant in the photosphere and into the low chromosphere, as one can see from panel (e). Therefore, we conclude that the swirling strength is mainly locally produced in these regions. Only in the upper part of the chromospheric layers, the advection term produces some non-negligible positive vertical swirling strength\footnote{To be precise, the advection term only advects swirling strength in or out of the time-distance slit; it does not produce it. Therefore, a positive value of $[(\boldsymbol{v}\cdot\nabla)\boldsymbol{\lambda}]_z$ can indicate either the advection of positive swirling strength into the slit, or negative swirling strength out of it.}. This production can only contribute to the reduction of the strong pulse in the height range $600\,{\rm km} < z < 800 \,{\rm km}$ around $t=5920\,{\rm s}$ and to the fading of the tip of swirling strength around $(t,z) = (5900\,{\rm s}, 800\,{\rm km})$, but it is not the main cause.
Therefore, if the dynamics of the swirling strength alone cannot explain the fading of the swirl beyond 850\,km, we conclude that it must be the tilted swirl axis to cause it, as already hypothesized in Sect.\,\ref{subsec:alfvén_pulses}. This tilt also leads to the reduction of the vertical propagation speed of the swirl, seen to set in already beyond $z\approx 600$\,km.
The second row of Fig.\,\ref{fig:SingleSwirlEvent_TDdiagrams_SingleLOS} shows the time-distance diagrams of the various terms of Eq.\,(\ref{eq:swirling_eq}). It serves to infer which physical processes are responsible for the production of negative swirling strength at the photospheric level and for the propagation of the pulses.
The gravitational term $T^5_z$ is omitted because the simulations employ a constant gravitational field and thus $T^5_z=0$.
First of all, analyzing the different source terms, we notice that the stretching term, $T^1_z$, is much weaker than the other terms. Second, the two baroclinic terms, $T^2_z$ and $T^3_z$, have similar configurations but are of opposite sign. Therefore, panel (i) shows the sum $T^2_z + T^3_z$ from which it can be seen that most of their contributions cancel out.
This result is in accordance with the findings of \citet{2020A&A...639A.118C}.
The counter-balancing of the two baroclinic effects inside and across the boundary of a stationary magnetic flux tube can be explained by the required equilibrium between magnetic pressure, magnetic tension forces, and gas pressure. However, the magnetic flux tube hosting the swirl is not stationary as can be seen from Fig.\,\ref{fig:SingleSwirlEvent}. The dynamics of the magnetic structure and of the surrounding plasma can break the balance between the pressures, which can result in a net production of swirling strength, as visible from panel (i). Thus, it becomes clear that the local production of negative swirling strength seen in panel (d) between $-200\,\mathrm{km} \lesssim z \lesssim 100\,\mathrm{km}$ and $5780\,\mathrm{s} \lesssim t \lesssim 5860\,\mathrm{s}$ is caused by the imbalance between the two baroclinic terms. More specifically, around $z=0\,{\rm km}$ and between $5790\,{\rm s}\,\lesssim t \lesssim 5850\,{\rm s}$ the baroclinic magnetic term, $T^3_z$, which is producing negative swirling strength, overcomes the hydrodynamical one, $T^2_z$, which instead produces positive swirling strength.
Because the time-distance slit is located in the center of the magnetic flux concentration, where the $\beta = 1$ surface dips below $z=0\,{\rm km}$ into the convection zone (see Fig.\,\ref{fig:SwirlsDistribution_TauBeta}), the magnetic field dominates the gas pressure and the dynamics in this location. This suggests that the origin of the swirl is due to magnetic effects alone that could arise from an MHD instability or from magnetic annihilation, possibly owing to nearby inverse polarity fields such as in the case of quiet Sun Ellerman bombs \citep[QSEBs,][]{2020A&A...641L...5J}.
Alternatively, the dominance of the baroclinic magnetic term may also arise from a weakening of the hydrodynamic baroclinic term caused by, for example, a sudden low pressure region or the bath tub effect \citep[see][]{2012Natur.486..505W}
in the convective motion surrounding the magnetic flux concentration. Revealing the true origin of the swirls calls for a multidimensional analysis of the different source terms of swirling strength, which goes beyond the scope of the present paper but shall become subject of a subsequent study.
Having dealt with the triggers of such an event, we next pay attention to the propagation of the swirl. As already mentioned further above, the upward propagation is caused by the production of negative swirling strength and its successive destruction by the production of a positive counter-part, approximately 20\,s later. The inspection of the source-term diagrams of Fig.\,\ref{fig:SingleSwirlEvent_TDdiagrams_SingleLOS} reveals that the magnetic tension term, $T^4_z$, is the main responsible for the production and destruction of $\lambda_z$ along the two pulses of swirling strength, as can be seen from Fig.\,\ref{fig:SingleSwirlEvent_TDdiagrams_SingleLOS}j. The stretching term, $T^1_z$, plays no role in the propagation of the swirl, the hydrodynamical baroclinic term $T^2_z$ is overcome by its magnetic counterpart $T^3_z$, which in turn only partially contributes to the production of the first pulse of swirling strength and to the destruction of the second.
Therefore, we can conclude that magnetic tension forces are responsible for the propagation of the swirl in the photosphere and low chromosphere. This analysis provides further evidence of the Alfvénic nature of the vortex flow since the magnetic tension is the driving force of Alfvén waves.
%
%
\subsection{Energetic considerations}\label{subsec:energetics}
%
\begin{figure}
\resizebox{\hsize}{!}
{\includegraphics{Figures/Alpha_TimeDist_Poynting_abc.pdf}}
\caption{
Time-distance diagrams of the vertical component of the Poynting flux vector $S_z$ (\emph{a}) and of the two terms that compose it, $S_z^{\rm v}$ (\emph{b}) and $S_z^{\rm h}$ (\emph{c}), for the swirl event shown in Fig.\,\ref{fig:SingleSwirlEvent}. Values at each time step and height level are averages over a finite horizontal plane of $150\,\mathrm{km}$ side length.
}
\label{fig:SingleSwirlEvent_TDdiagrams_Poynting}
\end{figure}
%
The purpose of the present section is to investigate how much energy swirling events and associated Alfvén pulses can channel up from the photosphere to the upper atmospheric layers. To this end, we analyze the vertical component of the Poynting flux vector and the mechanical energy flux relative to the swirling motions, as given in Sect.\,\ref{subsec:energyflux}. The first subsection concentrates on the single swirl event of Fig.\,\ref{fig:SingleSwirlEvent} and the second one evaluates the mean Poynting flux over the full computational domain.
%
\subsubsection{Energy transported by the swirl of Fig.\,\ref{fig:SingleSwirlEvent}}\label{susubbsec:energetics1}
%
Figure\,\ref{fig:SingleSwirlEvent_TDdiagrams_Poynting}a shows the time-distance diagram
of the vertical component of the Poynting flux, $S_z$, associated with the swirl event of Fig.\,\ref{fig:SingleSwirlEvent}. Like in
Fig.\,\ref{fig:SingleSwirlEvent_TDdiagrams}, the distance slit stretches from $z = -200$\,km to $z = 1200$\,km. It has a finite horizontal, quadratic extension of $150\,\mathrm{km}$ side length, and $S_z$ is taken to be the average over this finite cross section. At the location of origin of the Alfv\'en pulse at around $z=0$
(as was seen from Fig.\,\ref{fig:SingleSwirlEvent_TDdiagrams_SingleLOS}), the upwardly directed Poynting flux
is on the order of $10^6\,\mathrm{W}\,\mathrm{m}^{-2}$.
The flux decreases with height but it is still on the order of $10^5\,\mathrm{W}\,\mathrm{m}^{-2}$ at chromospheric levels. Panels (b) and (c) of Fig.\,\ref{fig:SingleSwirlEvent_TDdiagrams_Poynting} show the corresponding time-distance diagrams of the terms $S_z^{\rm v}$ and $S_z^{\rm h}$ respectively, which are defined in Eq.\,(\ref{eqn:z-Poynting}). We notice that the upward flux is entirely due to the term $S_z^{\rm h}$, which means that the energy is carried by the horizontal motions of the magnetized plasma. Hence, the magnetic swirl and associated Alfvén pulse are responsible for the upwardly directed energy transfer.
%
\begin{figure}
\resizebox{\hsize}{!}
{\includegraphics{Figures/Alpha_zPoynting_vh.pdf}}
\caption{
Vertical component of the Poynting flux vector $S_z$ of the swirl event presented in Fig.\,\ref{fig:SingleSwirlEvent} in a horizontal section at $z=700\,\mathrm{km}$. Arrows indicate the velocity vector field projected into the horizontal plane. Their length is proportional to the magnitude of the horizontal flow.
}
\label{fig:SingleSwirl_HorizontalSection_Poynting}
\end{figure}
%
Figure \ref{fig:SingleSwirl_HorizontalSection_Poynting} reveals the spiral pattern outlined by the vertical component of the Poynting flux, $S_z$, in a horizontal section centered on the swirl and at $(t,z) = (5980\,{\rm s},\,700\,\mathrm{km})$. It roughly follows that of the swirling plasma flow, as it can be seen from the overlaid velocity field. This reinforces the idea that the swirl is carrying magnetic energy through the solar atmosphere. The maximum positive Poynting flux of the swirl is $S_{z,{\rm max}} = 3.5\times10^5\,\mathrm{W}\mathrm{m}^{-2}$, while the mean vertical Poynting flux is $S_z = 34.5\,\mathrm{kW}\mathrm{m}^{-2}$,
where the mean was taken over a circular area of diameter $1.2\,{\rm arcsec}$
centered on the swirl and over the time period from $t = 5760\,{\rm s}$ to $t = 6050\,{\rm s}$.
From these numbers, the contribution of the swirl alone to the mean vertical Poynting flux through the cross section of the full computational domain of $9.6\times 9.6\,{\rm Mm}^2$ is $236\,{\rm W}\,{\rm m}^{-2}$. This value is smaller but on the order of the $440\,{\rm W}\,{\rm m}^{-2}$ reported by \citet{2012Natur.486..505W}
for the contribution to a cross section of $8\times 8$\,Mm$^2$, however, at a height of 2000\,km, that is, at the base of the transition zone.
We attribute this weak Poynting flux of the swirl of Fig.\,\ref{fig:SingleSwirlEvent} to its relatively weak magnetic field and to its bending above $z\approx 700\,{\rm km}$.
\citet{2019NatCo..10.3504L} count from observations with the \ion{Ca}{ii} H Broadband Filter Imager (BFI) of the Solar Optical Telescope (SOT) of Hinode $\bar{N}_{\rm s} = 48.2$ swirls at any time in a field of view of $A = 800\,\mathrm{Mm}^2$. Assuming that these swirls are similar to the one of Fig.\,\ref{fig:SingleSwirlEvent} with a mean Poynting flux of $S_z = 34.5\,{\rm kW}\,{\rm m}^{-2}$ over a circular cross section of diameter $1.2\,{\rm arcsec}$, we obtain from Eq.\,(\ref{eq:averagesflux}) an average energy flux of
$\bar{S}_z = 1.2\,{\rm kW}\,{\rm m}^{-2}$ at a height of $z=700\,{\rm km}$. This flux is not enough to compensate the radiative losses of approximately
$4.3\,{\rm kW}\,{\rm m}^{-2}$ (excluding Lyman $\alpha$) in the
semiempirical model of the chromosphere by \citet{1981ApJS...45..635V}
but could still account for an important source of energy in the upper chromosphere and the corona.
In the same way as for the Poynting flux, it is possible to estimate the mean mechanical energy flux owing to swirls. We take again the single swirl of Fig.\,\ref{fig:SingleSwirlEvent} as a model and average again over a fixed circular area of diameter $1.2\,{\rm arcsec}$ and the time period from $t=5760\,{\rm s}$ to $t=6050\,{\rm s}$. Furthermore, we assume as propagation speed the local Alfv\'en plus bulk speed in the vertical direction, $v_{\rm p} = v_{\rm A}\cos{(\theta)} + v_z$, as done in Sect.\,\ref{subsec:alfvén_pulses}, to compute the mechanical energy fluxes according to Eqs.\,(\ref{eq:mechflux_vh}) and (\ref{eq:mechflux_vlambda}).
Using Eq.\,(\ref{eq:mechflux_vh}), the estimate for the mean vertical mechanical flux owing to horizontal velocities is $F_z^{\rm r} = 6.1\,{\rm kW}\,{\rm m}^{-2}$, while Eq.\,(\ref{eq:mechflux_vlambda}) yields $F_z^{\rm \lambda} = 15.2\,{\rm kW}\,{\rm m}^{-2}$.
These formulas give only an approximate upper limit of the true mechanical energy flux because of a radial velocity component that was not removed in the computation of $F_z^{\rm r}$, while $F_z^{\lambda}$ is overestimated because of contribution from peripheral swirling strength not pertaining to the swirl under consideration. Nevertheless, the computed mechanical flux is only a fraction of the Poynting flux, which reflects the circumstances that the plasma-$\beta \lesssim 1$ in the considered region.
Assuming, once again, that the swirl shown in Fig.\,\ref{fig:SingleSwirlEvent} is representative of the $\bar{N}_{\rm s} = 48.2$ swirls detected at any time in a field-of-view of $A = 800\,{\rm Mm}^2$, and using Eq.\,(\ref{eq:averagemflux}) to compute the upper limit of the average mechanical energy flux owing to swirls, we find, at $z=700\,{\rm km}$, $\bar{F}_z^{\rm r} = 218\,{\rm W}\,{\rm m}^{-2}$ and $\bar{F}_z^{\rm \lambda} = 544\,{\rm W}\,{\rm m}^{-2}$, using Eq.\,(\ref{eq:mechflux_vh}) and Eq.\,(\ref{eq:mechflux_vlambda}), respectively. For comparison, \citet{2019NatCo..10.3504L} derive, from observations alone, a lower limit of the average mechanical energy flux from swirls of $33$ to $131\,{\rm W}\,{\rm m}^{-2}$ at a height of approximately $z=1000\,{\rm km}$.
%
\subsubsection{Mean energy transport}\label{susubbsec:energetics2}
%
\begin{figure}
\resizebox{\hsize}{!}
{\includegraphics{Figures/z700_t5800_nostream.pdf}}
\caption{
Vertical component of the Poynting flux vector $S_z$ over the entire computational domain at $z=700\,\mathrm{km}$ and $t = 5800\,{\rm s}$. Black circles indicate the location of some of the swirl events listed in Table\,\ref{tab:ListSupplementarySwirls} also shown in Fig.\,\ref{fig:Movie_IcoBin5}.
}
\label{fig:z700_t5860}
\end{figure}
%
\begin{figure}
\resizebox{\hsize}{!}
{\includegraphics{Figures/Av_PoyntingFlux.pdf}}
\caption{
Mean vertical component of the Poynting flux vector $S_z$ (blue) and of the terms $S_z^{\rm h}$ (green) and $S_z^{\rm v}$ (orange) as a function of height $z$. The spatial mean is obtained by averaging over the horizontal cross sections of the full computational domain of $9.6\times 9.6\,{\rm Mm}^2$, while the temporal mean is computed from $21$ different time instants spread over the time period of the high-cadence series of $15\,{\rm min}$. The shaded areas correspond to the standard deviation obtained from the temporal variation.
}
\label{fig:Av_PoyntingFlux}
\end{figure}
%
Looking at the mean net Poynting flux across the full computational domain of $9.6\times 9.6\,{\rm Mm}^2$, it turns out that this flux is much larger than the extrapolation from the single, relatively weak, isolated swirl event of Fig.\,\ref{fig:SingleSwirlEvent}, and that the bulk of energy transport is delivered by the more complex events of superposition of swirls (see Appendix\,\ref{app:superposition_of_swirls}).
This situation is illustrated with Fig.\,\ref{fig:z700_t5860}, which shows the vertical component of the Poynting flux on the horizontal cross section at $z=700\,{\rm km}$ for the same time instant as Fig.\,\ref{fig:Movie_IcoBin5}, that is,
$t = 5800\,{\rm s}$. In this figure, the largest single contribution to the mean vertical Poynting flux of $18.4 \, {\rm kW}\,{\rm m}^{-2}$ comes from event (b) near the center of the field of view with a peak Poynting flux of $2123.7\,{\rm kW}\,{\rm m}^{-2}$. Comparing it with Fig.\,\ref{fig:Movie_IcoBin5}, it can be seen that this event starts from a relatively large magnetic footpoint patch with a complex but clearly swirling motion in the bin-5 intensity (see the animation of Fig.\,\ref{fig:Movie_IcoBin5}). For comparison, event (f) corresponds to the single, isolated swirl of Fig.\,\ref{fig:SingleSwirlEvent}.
Figure \,\ref{fig:Av_PoyntingFlux} shows the temporal and spatial mean of the vertical component of the Poynting flux $S_{z}$ together with the terms $S_z^{\rm v}$ and $S_z^{\rm h}$ as a function of height $z$. The shaded area surrounding each curve indicates the standard deviation of the temporal variation over the time period of $15\,{\rm min}$ of the high-cadence simulation\footnote{We have also created an equivalent plot for the duration of $2\,{\rm h}$ for which we had snapshots available with a cadence of $4\,{\rm min}$. The result is very similar to Fig.\,\ref{fig:Av_PoyntingFlux}.}. First we observe that all curves converge to zero flux at the upper boundary, which is a consequence of the top boundary condition for the magnetic field. Therefore, magnetic disturbances are reflected at the top boundary with the consequence that the net vertical Poynting flux vanishes in the upper part of the atmosphere\footnote{This limitation could be lifted in a future simulation run with a more relaxed top boundary condition for the magnetic field; for example using the condition ${\rm d}\boldsymbol{B}/{\rm d}z = 0$, or matching to a potential field configuration.
Varying boundary conditions would possibly lead to deviations in the energy fluxes surpassing the standard deviation shown in Fig.~\ref{fig:Av_PoyntingFlux}.}.
Second, the contribution due to horizontal motions (including swirls) is always positive, while vertical motions contribute negatively to the Poynting flux. Above $z \approx 1000\,{\rm km}$, the negative contribution is even slightly dominating, possibly due to the boundary condition of vertical magnetic field at the top, which favors vertical motions in the upper part of the atmosphere. In this height range, $\beta < 0.1$ in regions of significant Poynting flux so that the top boundary condition strongly influences the magnetic field down to $z \approx 1000\,{\rm km}$ because of the stiffness of this field. However, at the base of the chromosphere, which we take here again to be located at $z=700\,{\rm km}$, the mean flux is $12.8\pm6.5\,{\rm kW}\,{\rm m}^{-2}$, which, for a comparison, would be
sufficient to compensate the radiative losses of $4.3\,{\rm kW}\,{\rm m}^{-2}$ in the chromosphere.
We also notice that at this height almost all the vertical Poynting flux is carried by the horizontal motions of the magnetized plasma ($S_z \approx S_z^{\rm h}$), that is, largely by swirling motions. This result agrees with the findings reported by \citet{2020ApJ...894L..17Y}.
Moreover, they demonstrate that the largest contribution to the Poynting flux comes from small-scale vortices that cluster to form the larger, observable swirls. This is consistent with our finding that superposition of swirls can carry large amounts of magnetic energy as in the case of event (b) of Fig.\,\ref{fig:z700_t5860}.
This large magnetic energy flux is tantalizing, considering the findings of
\citet{2005Natur.435..919F}
that high-frequency acoustic waves are not sufficient to heat the solar chromosphere and that the emission from the middle and upper chromosphere must be balanced by processes related to the magnetic field: a problem that is still unsolved
\citep{2019ARA&A..57..189C}.
\citet{2012A&A...541A..68M}
showed with the help of simulations similar to the present ones that Ohmic dissipation of magnetized swirling motions induce substantial local heating, although mostly in the photosphere.
However, a word of caution is in order here. The initial condition of a homogeneous vertical magnetic field together with the periodic lateral boundary conditions, of which the simulation keeps a memory for all times, favor the formation of vertically extended vortices. These simulations may adequately represent regions with a predominantly vertically directed magnetic field such as magnetic network patches or plage regions. With a more heterogeneous magnetic field, such as that generated by a turbulent dynamo in very quiet regions, magnetically induced swirls are probably less numerous and hence the Poynting flux less vigorous.
\section{Summary and conclusions}\label{sec:conclusions}
We carried out numerical radiation MHD simulations of the solar atmosphere with the purpose of studying the origin and propagation of vortical motions that occur in conjunction with magnetic fields and bear great resemblance to observed chromospheric swirls. The simulations are of fairly high spatial and temporal resolution and reach from the convectively unstable subsurface layers to the chromosphere. For the analysis, we mostly use a run that started with a homogeneous vertical magnetic field of a strength of $50\,{\rm G}$, but use for comparison also runs that started with a magnetic field of $200\,{\rm G}$ and without magnetic field.
Looking first at the distribution of swirling strength, which is a measure for the vigor of swirling motions, we find a drastic difference between the simulations with and those without a magnetic field. The simulations with magnetic field, which show the usual funnel-shaped magnetic flux concentrations in the photosphere, have the swirling motion strongly concentrated within these low $\beta$ funnels and aligned with the magnetic field. Different from that, the simulation without magnetic field possesses more arch-shaped vortex tubes in the photosphere and almost isotropically distributed vortices in the chromosphere. These results agree with and confirm earlier findings of
\citet{2011A&A...533A.126M}
and \citet{2012ASPC..456....3S}.
Beyond this relation between magnetic field and swirling motion, we find a tight relationship between swirling motions and perturbations in the magnetic field of intense magnetic flux tubes. In particular, we show with the help of a statistical analysis that the swirling strength and the magnetic swirling strength are highly correlated in the solar atmosphere. While swirling strength indicates a vortex in the plasma, the magnetic swirling strength indicates a twist in the magnetic lines of force. The latter typically occurs in the presence of torsional Alfvén waves. Therefore, the tight correlation between these two quantities is a first indication of the relationship between swirling motions and torsional Alfvén waves in the simulated solar atmosphere.
We then analyze one particular swirl event that was identified in the numerical simulations (see Fig.\,\ref{fig:SingleSwirlEvent}). The well developed chromospheric swirl is colocated with a positive polarity magnetic BP in an intergranular lane of the deep photosphere and presents a twist in the magnetic field. Therefore, it can be identified as a magnetic swirl. Using time-distance diagrams to study its evolution in time and altitude, we find three pieces of evidence of the Alfv\'enic nature of this magnetic swirl.
First, the vortical motions and the magnetic perturbations, in the upwardly directed (positive polarity) magnetic field, have opposite orientation; second, the swirl propagates upward with the local Alfv\'en speed and, third, the driving force that is responsible for such propagation is the magnetic tension. These are all characteristics of torsional Alfv\'en waves. However, we do not observe an oscillatory wave, but instead pulses of unidirectional vortical motions and twisted magnetic field lines. From all that, we conclude that the swirl subsists in an upwardly propagating torsional Alfv\'en pulse. The other eight swirls listed in Table\,\ref{tab:ListSupplementarySwirls} also share these characteristics and are therefore also identified as torsional Alfvén pulses.
A similar conclusion was also reached by
\citet{2019NatCo..10.3504L}
from observations and modeling. The novelty of the present work is that these Alfv\'en pulses are not deliberately excited but naturally arise from a self-consistent numerical solution of the MHD equations and radiative transfer.
Regarding the origin of the Alfv\'en pulse, we cannot say with certainty if it is predominantly hydrodynamic or magnetic. From the analysis based on the swirling equation, we find that magnetic baroclinicity in the top convection zone and the deep photosphere is responsible for the creation of the swirl. Hence, magnetic dynamics alone would launch the pulse.
However, the present time-distance analysis is based on a single, vertical slit through the center of the swirl, which does not follow the temporal evolution of the swirl nor its geometrical shape. This can lead to misinterpretations of the diagrams, especially in the chromosphere where the plasma flows are fast.
Therefore, it is necessary to consider the full three-dimensional structure of the individual terms of the swirling equation as a function of time in order to make progress. Such a three-dimensional analysis was beyond the scope of the present paper, but it can be expected to reveal whether the magnetic baroclinicity in the center of the swirl is caused by magnetic effects alone (e.g., magnetic diffusion or an MHD instability such as that responsible for quiet Sun Ellerman bombs) or by hydrodynamic effects in the surroundings of the magnetic flux concentration (e.g., by the bathtub effect or the granular buffeting).
The upwardly propagating torsional Alfv\'en pulses carry a substantial amount of magnetic and mechanical energy from the photosphere to the outer layers of the atmosphere. At the top boundary of the computational domain, the Poynting flux vanishes because of the adopted boundary condition that forces the field to become strictly vertical. Therefore, we limited our analysis to the base of the chromosphere for which we took here the level $z=700$\,km.
For the swirl of Fig.\,\ref{fig:SingleSwirlEvent}, we evaluate at this height level a mean Poynting flux in the upward direction of $S_z = 34.5\,\mathrm{kW}\mathrm{m}^{-2}$. This flux is generated by the horizontal motions of the magnetized plasma alone and is therefore an effect of the horizontally swirling motion. Taking the observed frequency of chromospheric swirls from
\citet{2019NatCo..10.3504L}
and assuming they were all similar to the swirl of Fig.\,\ref{fig:SingleSwirlEvent}, we obtain a mean Poynting flux of $\bar{S}_z = 1.2\,{\rm kW}\,{\rm m}^{-2}$ and for the corresponding mean mechanical energy flux we find an upper limit of $\bar{F}_{z} = 218 - 544 \,{\rm W}\,{\rm m}^{-2}$. This energy flux could source the energy budget of the higher layers, but it is not enough to compensate alone for radiative losses at chromospheric level.
However, the swirl of Fig.\,\ref{fig:SingleSwirlEvent} is relatively weak; when evaluating the average Poynting flux across the full computational domain, still at a height of $z=700\,{\rm km}$, we obtain
$12.8\pm6.5\,{\rm kW}\,{\rm m}^{-2}$. This large amount mainly stems from more complex events of superposition of swirls, which originate in relatively large, complex, and intense magnetic footpoints.
Assessing these numbers one should bear in mind that the initial condition of a homogeneous, vertical magnetic field favors the formation of vertically extended vortices and may be rather applicable to regions with a predominantly vertically directed magnetic field such as network and plage regions than to quiet Sun regions.
Would it be possible to directly observe the Alfv\'enic nature of chromospheric swirls? We believe that it should be feasible with high resolution and high sensitivity spectropolarimetry in two spectral lines, preferably sampling the photosphere and the chromosphere. The challenge would consist in measuring the transverse magnetic field accurately enough to prove the twist of the magnetic field relative to the vortex motion that would be detected with the help of LCT. The two height levels would inform about the propagation speed of the swirl. High photon flux is needed for accurately measuring the transverse Zeeman effect, which requires a large aperture solar telescope such as the D.\,K.\,Innouye Solar Telescope (DKIST).
\begin{acknowledgements}
AFB wishes to acknowledge support by the Swiss National Science Foundation (SNF) through grant 200021\_189180 (Solar Orbiter STIX), JRCC support by SNF under grant ID 200020\_182094, and FC support through the CHROMATIC project (2016.0019) of the Knut and Alice Wallenberg foundation. Special thanks are addressed to L.\,Harra who directed AFB's Master's thesis from which the present work derives. This work has profited from discussions with the team of K.\,Tziotiou and E.\,Scullion (conveners) ``The Nature and Physics of Vortex Flows in Solar Plasma'' at the International Space Science Institute (ISSI) and with the Waves in the Lower Solar Atmosphere (WaLSA; www.WaLSA.team) team (S.\,Jafarzadeh, convener), which is supported by the Research Council of Norway (project number 262622).
Simulations were carried out at the Swiss National Supercomputing Centre (CSCS) under project ID s560 with support from SNF under grant ID 200020\_157103. Thanks are also extended for valuable input by G.\,Vigeesh and P.\,Rajaguru, and for very helpful and encouraging comments by an anonymous referee.
\end{acknowledgements}
\bibliographystyle{aa}
|
\section{Introduction}
This paper concerns the eigenvalues of the Brownian sheet matrix ${\boldsymbol X}= \{\boldsymbol X(s,t),0\le s,t<\infty\} $, which is
a symmetric-matrix-valued process with entries $X_{ij}$ for $1\le i, j\le d$ given by
\begin{equation}\label{e:X}
X_{ij}(s,t)=\begin{cases}
b_{ij} (s,t), &i < j,\\
\sqrt{2} b_{ii}(s,t), & i=j,
\end{cases}
\end{equation}
where $b = \left\{ b_{ij}(s,t), 0\le s, t<\infty\right \}_{1\le i \le j\le d }$ is a family of independent Brownian sheets.
After the fundamental work \cite{Wigner55} which established the celebrated Wigner's semicircle law, Brownian motion as a
one-parameter stochastic process was introduced into random matrix theory by Dyson \cite{Dyson1962}. Since then, there
has been fruitful literature on the Dyson Brownian motion which is the system of eigenvalues of symmetric Brownian matrix
(see, e.g. \cite{Cepa1997, Chan1992, Rogers1993, anderson2010, EY17} and the references therein), in which It\^o's calculus
has played a key role. By studying the high-dimensional limit of the empirical measures of the Dyson Brownian motion, one can
provide a dynamical proof for Wigner's semicircle law (see, e.g., \cite{anderson2010}). The Dyson Brownian motion is also closely
related to interacting particle systems, and the equation (known as the McKean-Vlasov equation) satisfied by its limiting empirical measure
appears naturally in the study of propagation of chaos for large systems of interacting particles (see, e.g., \cite{BO19, JW18, Serfaty20}).
Multiparameter stochastic processes (or random fields) are a natural extension of one-parameter processes, they arise naturally in statistical
mechanics (e.g. Brownian sheet appears in the Ising model \cite{KM86} and interacting particle systems \cite{KT88}), and systematic theories
have been developed (see, e.g., \cite{CW75,K02} and the references therein). Motivated by the close connection between random matrix
theory and interacting particle systems, it is natural to develop theories for random matrix with entries being random fields. Recently,
the problem on the collision of eigenvalues of symmetric (Hermitian) matrix whose entries are independent Gaussian fields was investigated
in \cite{Jaramillo2018, SXY20}, which to our best knowledge are the only literature on random matrix whose entries are random fields.
Another motivation for studying the Brownian sheet matrix $\boldsymbol X$ is from free probability theory. As shown in \cite{Voi91, VDN92}, many theorems and
concepts in free probability have classical probability analogs, and furthermore free probability is closely connected with random matrix theory. In
particular, free Brownian motion can be viewed as the high-dimensional limit of rescaled Brownian motion matrix which is define by \eqref{e:X} with
$b$ being a family of independent Brownian motions. Stochastic calculus for free Brownian motion was developed in \cite{BS98}. Free fractional
Brownian motion arose naturally in \cite{NT14} when studying the central limit theorem for long-range dependence time series in free probability,
and the stochastic calculus was developed in \cite{DS19}. It was shown in \cite{Pardo2016} that free fractional Brownian motion is the
high-dimensional limit of empirical measures of the eigenvalues of rescaled fractional Brownian motion matrices.
We remark that the free Brownian motion and the free fractional Brownian motion in \cite{BS98,NT14,Pardo2016,DS19} are one-parameter
stochastic processes, and we believe that our study of the Brownian sheet matrix in this paper will provide a useful building block for constructing
free random fields.
In the present paper we shall derive a system of stochastic partial differential equations \eqref{eq-spde''} for the eigenvalue processes
of the Brownian sheet matrix $\boldsymbol X$ given by \eqref{e:X}, obtain the tightness of the spectral empirical measures (Theorem \ref{Thm-tightness}),
and show that the limit measure satisfies a McKean-Vlasov equation \eqref{e:MV} and a Burgers' equation \eqref{e:Burgers'}.
We briefly explain the structure of the paper below.
Though the Brownian sheet is a simple multivariable extension
of standard Brownian motion,
the stochastic calculus for the Brownian sheet that one needs for deriving the stochastic partial differential equations for
the eigenvalues of the Brownian sheet matrix turns out to be highly non-trivial and cannot be adapted directly from the
classical It\^o calculus. In Section \ref{sec:stochastic-calculus}, we follow the approach of Cairoli and Walsh in \cite{CW75}
and develop stochastic calculus tools for the multi-dimensional Brownian sheet on the plane for our purpose. The main results in
this section are Theorems \ref{Thm-Green formula} and \ref{Coro-Green formula} which are multi-dimensional Green's formulas.
In Section \ref{sec:SPDE}, by applying classical It\^o's formula together with Green's formulas (Theorems \ref{Thm-Green formula}
and \ref{Coro-Green formula} ),
we derive the system of stochastic partial differential equations \eqref{eq-spde''} for the eigenvalues of the Brownian sheet matrix
$\boldsymbol X$. Compared with the following system of SDEs for the classical Dyson Brownian motion: for $1\le i\le d$,
\begin{equation}\label{Eq:Dyson}
d\lambda_i(t) = \sqrt 2 dW_i(t) + \sum_{j\neq i}\frac1{\lambda_i(t)-\lambda_j (t)}dt,
\end{equation}
where $W=(W_1, \dots, W_d)$ is a standard $d$-dimensional Brownian motion, we remark that
eq. \eqref{eq-spde''} bears some resemblance to \eqref{Eq:Dyson} but has several extra high-order terms.
In Section \ref{sec:limit}, we study the high-dimensional limit of empirical distributions for the eigenvalue processes of $\boldsymbol X$.
In Section \ref{sec:tightness}, we establish the tightness of the set of empirical spectral
measures which are viewed as $C([0,S]\times[0,T], \mathcal P(\mathbb{R}))$-valued random elements (see Theorem
\ref{Thm-tightness}). This guarantees that every sequence of the empirical spectral measures has a subsequence
which converges weakly. The tightness together with the classical Wigner's semicircle law implies the existence and
uniqueness of the high-dimensional limit of the empirical spectral measures (see Theorem \ref{Thm-limit measure}).
In Section \ref{sec:PDE}, we derive partial differential equations \eqref{e:MV} and \eqref{e:Burgers'} that are satisfied
by the limiting measure, by using the property of the semicircle distribution.
Finally, in Appendix \ref{sec:appendix} we provide some results in matrix analysis which are needed in our analysis.
\section{Stochastic calculus for the Brownian sheet}\label{sec:stochastic-calculus}
In this section, we shall apply the stochastic calculus on the plane developed in \cite{CW75} to derive Green's formula
for the multi-dimensional Brownian sheet, which is a key ingredient for studying SPDEs for the eigenvalues
in Section \ref{sec:SPDE}.
\subsection{Some preliminaries on stochastic calculus on the plane}
In this subsection, we recall from Cairoli and
Walsh \cite{CW75} some preliminaries for stochastic
calculus on the plane.
Define the partial order ``$\prec$'' on $\mathbb{R}^2$ as follows. For any $(s_1, t_1), \, (s_2, t_2) \in \mathbb{R}^2$,
\[(s_1,t_1)\prec (s_2, t_2), \text{ iff } s_1\le s_2, \, t_1\le t_2,\]
and write
\[(s_1,t_1)\prec \prec(s_2, t_2), \text{ iff } s_1< s_2, \, t_1< t_2.\]
Let $(\Omega, \mathcal G, \mathbb{P})$ be a probability space and let the filtration $\mathcal{F}=\{\mathcal{F}_z, z\in \mathbb{R}_+^2\}$ be a family of
sub-$\sigma$-field of $\mathcal G$ satisfying
\begin{enumerate}
\item $\mathcal{F}_z\subset \mathcal{F}_{z'}$ if $z\prec z'$;
\item $\mathcal{F}_0$ contains all null sets of $\mathcal G$;
\item for each $z, \mathcal{F}_z=\bigcap\limits_{z\prec\prec z'}\mathcal{F}_{z'}$;
\item for each $z, \mathcal{F}_{z}^1$ and $\mathcal{F}_{z}^2$ are conditionally independent given $\mathcal{F}_z$.
\end{enumerate}
Here, for $z=(s,t)\in \mathbb{R}_+^2$,
\begin{align*}
\mathcal{F}_{z}^1=\mathcal{F}_{s\infty}:= \underset{v}\vee \mathcal{F}_{sv};~~~
\mathcal{F}_{z}^2=\mathcal{F}_{\infty t}:= \underset{u}{\vee}\mathcal{F}_{ut}.
\end{align*}
In particular, the augmented filtration generated by a finite family of independent Brownian sheets satisfies the
above conditions.
Let $Y=\{Y_z, z\in \mathbb{R}_+^2\}$ be a process such that for each $z$ the random variable $Y_z$ is integrable.
We recall the definitions of martingale, strong martingale, weak martingale, and increasing process relative to
$\mathcal{F}$ in \cite{CW75}.
\begin{definition}
$Y$ is a \emph{martingale} if
\begin{enumerate}
\item $Y$ is adapted;
\item $\mathbb{E}}\def\E{\mathbb{E}[Y_{z'}|\mathcal{F}_z]=Y_z,$ for each $z\prec z'$.
\end{enumerate}
\end{definition}
Suppose $z=(s,t)$ and $z'=(s',t')$ such that $z\prec\prec z'$. We denote by $(z, z']$ the rectangle
$(s,s']\times(t,t']$. The increment of $Y$ over the rectangle $(z, z']$ is
\[Y((z,z'])=Y_{s't'}-Y_{st'}-Y_{s't}+Y_{st}.\]
\begin{definition}~
\begin{itemize}
\item[(a)] $Y$ is a \emph{weak martingale} if
\begin{enumerate}
\item $Y$ is adapted;
\item $\mathbb{E}}\def\E{\mathbb{E}[Y((z,z'])|\mathcal{F}_z]=0$ for each $z\prec\prec z'.$
\end{enumerate}
\end{itemize}
\begin{itemize}
\item[(b)] $Y$ is an \emph{$i$-martingale} ($i=1,2$) if
\begin{enumerate}
\item $Y$ is $\mathcal{F}_z^i$-adapted;
\item $\mathbb{E}}\def\E{\mathbb{E}[Y((z,z'])|\mathcal{F}_z^i]=0$ for each $z\prec\prec z'.$
\end{enumerate}
\end{itemize}
\begin{itemize}
\item[(c)] $Y$ is a \emph{strong martingale} if
\begin{enumerate}
\item $Y$ is adapted;
\item $Y$ vanishes on the axes;
\item $\mathbb{E}}\def\E{\mathbb{E}[Y((z,z'])|\mathcal{F}_z^1\vee \mathcal{F}_z^2]=0$ for each $z\prec\prec z'.$
\end{enumerate}
\end{itemize}
\end{definition}
\begin{definition}
$Y$ is an \emph{increasing process} if
\begin{enumerate}
\item $Y$ is right-continuous and adapted;
\item $Y_z=0$ on the axes;
\item $Y(A)\ge0$ for each rectangle $A\subset \mathbb{R}_+^2$.
\end{enumerate}
\end{definition}
Let $M=\{M_z, z\in\mathbb{R}_+^2\}$ be a martingale relative to $\mathcal{F}$. Then $M$ is both a 1-martingale and 2-martingale,
i.e., $\{M_{s0}, \mathcal{F}_{s0}^1, s\in\mathbb{R}_+\}$ and $\{M_{0t}, \mathcal{F}_{0t}^2, t\in\mathbb{R}_+\}$ are martingales. The converse is also true.
Now we assume that $M$ is a square integrable martingale. By \cite[Theorem 1.5]{CW75}, there exists an increasing
process $\langle M\rangle$ such that $M^2-\langle M\rangle$ is a weak martingale. For each fixed $t$, let
$\{[M]^1_{st}, s\in\mathbb{R}_+ \}$ be the unique increasing process which is predictable relative to $\{\mathcal{F}_{st}, s\in \mathbb{R}_+\}$
such that $\{M^2_{st}-[M]^1_{st}, s\in \mathbb{R}_+\}$ is a martingale. Similarly, one can define $[M]^2$. As pointed by
\cite[p.121]{CW75}, for a strong martingale $M$, either $[M]^1$ or $[M]^2$ can serve as the process $\langle M\rangle$.
Furthermore, by \cite[Theorem 1.9]{CW75}, if either $\mathcal{F}$ is generated by the Brownian sheet or $M$ has finite fourth
moment, then $[M]^1=[M]^2$, and hence we can choose $\langle M \rangle=[M]^1=[M]^2$. As a consequence,
for any fixed $t$, $\{M^2_{st}-\langle M \rangle_{st}, s\in \mathbb{R}_+\}$ is a martingale, and similarly, for any fixed $s$,
$\{M^2_{st}-\langle M \rangle_{st}, t\in \mathbb{R}_+\}$ is a martingale. As in \cite{CW75}, we shall use $d_s \langle M \rangle_{st}$
($d_t \langle M \rangle_{st}$, resp.) to denote the differential of $\langle M\rangle $ with respect to $s$ ($t$, resp.).
For two square integrable martingales $M$ and $N$, we denote by $\langle M, N\rangle$ any process which is the
difference of two increasing processes such that $MN-\langle M, N\rangle$ is a weak martingale.
One can choose, for instance,
\begin{equation}\label{e:quad-cov}
\langle M, N\rangle =\frac12\big(\langle M+N \rangle- \langle M\rangle -\langle N\rangle \big).
\end{equation}
Define $[M,N]^i=\frac12\left([M+N]^i-[M]^i-[N]^i\right)$ for $i=1,2.$ Then either $[M,N]^1$ or $[M,N]^2$ can serve as the process
$\langle M,N \rangle.$ Two martingales $M$ and $N$ are said to be \emph{orthogonal} if $MN$ is a weak martingale,
and we write $M\perp N$.
For $p\ge 1$, let $\mathfrak M^p$ denote the set of right-continuous martingales $M=\{M_z, z\in \mathbb{R}_+^2\}$ such that
$M=0$ on the axes and $\mathbb{E}}\def\E{\mathbb{E}[|M_z|^p]<\infty$ for all $z\in \mathbb{R}_+^2$. Let $\mathfrak M^p_c$ (resp. $\mathfrak M_s^p$) be the set of
continuous (resp. strong) martingales in $\mathfrak M^p$. Similarly, let $\mathfrak M^p(z_0)$ (resp. $\mathfrak M^p_c(z_0), \mathfrak M^p_s(z_0)$)
be the set of right-continuous (resp. continuous, strong) martingales $M=\{M_z, z\prec z_0\}$ such that $M_z=0$
on the axes and $\mathbb{E}}\def\E{\mathbb{E}[|M_z|^p] <\infty$ for all $z\prec z_0$.
Below we recall some results which will be used in our proofs.
\begin{theorem}{\cite[Theorem 1.2]{CW75}} \label{Thm-BDG 2dim}
Let $\{M_z:z \in \mathbb{R}_+^2\}$ be a right-continuous martingale. Then for $p > 1$,
\begin{align*}
\mathbb{E}}\def\E{\mathbb{E} \left[ \sup_z |M_z|^p \right]
\le \left( \dfrac{p}{p-1} \right)^{2p} \sup_z \mathbb{E}}\def\E{\mathbb{E} \left[ |M_z|^p \right].
\end{align*}
\end{theorem}
For any $z\in \mathbb{R}_+^2$, we denote the rectangle $(0, z]$ by $R_z$. We also fix $z_0\in \mathbb{R}_+^2$.
\begin{theorem}{\cite[Proposition 1.6]{CW75}} \label{Thm-prop1.6}
Let $M,N\in \mathfrak M^2(z_0)$. Then
\begin{enumerate}
\item $\mathbb{E}}\def\E{\mathbb{E}[(MN)(D)|\mathcal{F}_z]=\E[M(D) N(D)|\mathcal{F}_z]$ for each rectangle $D=(z, z']\subset R_{z_0}$;
\item $M\perp N$ iff ~$\E[M(D)N(D)|\mathcal{F}_z]=0$ for each rectangle $D=(z, z']\subset R_{z_0}$.
\end{enumerate}
\end{theorem}
\begin{theorem}{\cite[Proposition 1.8]{CW75}} \label{Thm-prop1.8}
If $M\in \mathfrak M_s^2(z_0)$, then $[M]^i$ is the unique $\mathcal{F}_z^i$-predictable increasing process such that
for $i=1,2$,
\begin{align*}
\mathbb{E}}\def\E{\mathbb{E} \left[ M(D)^2|\mathcal{F}_z^i \right]
= \mathbb{E}}\def\E{\mathbb{E} \left[ (M^2)(D)|\mathcal{F}_z^i \right]
= \mathbb{E}}\def\E{\mathbb{E} \left[ [M]^i(D)|\mathcal{F}_z^i \right]
\end{align*}
for each rectangle $D = (z,z'] \subseteq R_{z_0}$. Consequently, for $M,N\in \mathfrak M_s^2(z_0)$, noting
that $MN=\frac12\left((M+N)^2-M^2-N^2\right)$, we have for $i=1,2$,
\begin{align*}
\mathbb{E}}\def\E{\mathbb{E} \left[ M(D)N(D)|\mathcal{F}_z^i \right]
= \mathbb{E}}\def\E{\mathbb{E} \left[ (MN)(D)|\mathcal{F}_z^i \right]
= \mathbb{E}}\def\E{\mathbb{E} \left[ [M,N]^i(D)|\mathcal{F}_z^i \right]
\end{align*}
\end{theorem}
\begin{theorem}{\cite[Theorem 1.9]{CW75}} \label{Thm-prop1.9}
Let $M\in \mathfrak M_s^2$. Assuming either the filtration $\mathcal{F}$ is generated by the Brownian sheet or $M$ is
continuous with finite fourth moment, we have $[M]^1 = [M]^2$.
\end{theorem}
\subsection{On $\psi\cdot MN$ and $J_{MN}$ }\label{sec:J-MN}
Let us recall from \cite[Section 6]{CW75} the notion $J_M$ of a continuous martingale $M\in \mathfrak M_s^4$
on $\mathbb{R}_+^2$.
Recall the notation $R_{st} = (0,s] \times (0,t]$. By \cite[Eq.~(6.3)]{CW75},
\begin{align*}
J_M(s_0,t_0)
=& \int_0^{s_0} M(s,t_0) M(ds,t_0) - \int_{R_{s_0t_0}} M(s,t) dM(s,t)\\
=& \int_0^{t_0} M(s_0,t) M(s_0,dt) - \int_{R_{s_0t_0}} M(s,t) dM(s,t)\\
=& \frac12 M^2(s_0, t_0)-\frac12\langle M\rangle_{s_0, t_0}-\int_{R_{s_0t_0}} M(s,t) dM(s,t).
\end{align*}
Heuristically, one has $dJ_M(s,t)= M(s,dt) M(ds,t)$ (see \cite{CW75}). Similarly, for two $\mathcal{F}$-adapted martingales
$M$ and $N$, we introduce the following generalization $J_{MN}$ which induces the measure $M(s, dt) N(ds,t)$ on $\mathbb{R}_+^2$,
\begin{align} \label{def-J}
J_{MN}(s_0,t_0)
= \int_0^{s_0} M(s,t_0) N(ds,t_0) - \int_{R_{s_0t_0}} M(s,t) dN(s,t),
\end{align}
assuming that the right-hand side is well-defined. Clearly we have $J_M=J_{MM}$. Analogous to $J_M$ in \cite[Theorem 6.1]{CW75},
$J_{MN}$ will play a key role in the multi-dimensional Green's formula in the forthcoming Theorems \ref{Thm-Green formula} and
\ref{Coro-Green formula}.
Similar to \cite {CW75}, we shall represent $J_{MN}$ by a new type of stochastic integral denoted by $\psi\cdot MN$ which will be
defined in the sequel. Firstly, we need to introduce another order relation ``$\curlywedge$'' in $\mathbb{R}_+^2$ which is complementary
to ``$\prec$'' and plays an essential role in the definition of $\psi\cdot MN$. For $z=(s,t)$ and $z'=(s', t')$, we say $z \curlywedge z'$
if $s\le s'$ and $t\ge t'$, and $z{\curlywedge\atop \curlywedge} z'$ if $s<s'$ and $t>t'$. In the $st$-plane where the $s$-axis is
horizontal and the $t$-axis is vertical, $z\curlywedge z'$ means that $z$ is on the upper left of $z'$ in the plane. As a comparison,
$z\prec z'$ means that $z$ is on the lower left of $z'$.
\begin{proposition} \label{Prop 2.1}
Suppose $M,N\in \mathfrak M_s^2(z_0)$. Let $A=(z_A, z_A']$ and $B=(z_B, z_B']$ be two rectangles such that $A\curlywedge B$,
i.e., $z_1\curlywedge z_2$ for all $z_1\in A$ and $z_2\in B$.
Define the process $X= \{X_z, ~z \in \mathbb{R}_+^2\}$ by
\begin{align*}
X_z = \xi M(A\cap R_z) N(B \cap R_z), ~z \in \mathbb{R}_+^2,
\end{align*}
where $\xi$ is bounded and $\mathcal{F}_{z_A\vee z_B}$-measurable. Then $X$ belongs to $\mathfrak M^2(z_0)$, it is continuous if $M$ is, and
\begin{align}\label{e:quad-var}
\langle X \rangle_z
=\xi^2 \iint_{R_z \times R_z} \mathbf 1_A(z_1) \mathbf 1_B(z_2) d[M]^2_{z_1} d[N]^1_{z_2} .
\end{align}
\end{proposition}
\begin{proof}
We will follow the proof of \cite[Proposition 2.4]{CW75}.
For $D = (z,z']$ with $z = (s,t) \prec \prec z' = (s',t')$, the increment of $X$ over $D$ is
\begin{equation}\label{e:increment-X}
X(D) = M({\tilde A})N({\tilde B}),
\end{equation}
where $\tilde A = A \cap (R_{s't'}\backslash R_{s't})$ and ${\tilde B} = B \cap (R_{s't'}\backslash R_{st'})$.
Suppose $z_{\tilde A}$ is the lower-left corner of $\tilde A$. Then both $\xi$ and $N({\tilde B})$ are $\mathcal{F}_{z_{\tilde A}}^2$-measurable,
and hence
\begin{align*}
\mathbb{E}}\def\E{\mathbb{E} \left[ X(D)|\mathcal{F}_z^2 \right]
= \mathbb{E}}\def\E{\mathbb{E} \left[ \left.\mathbb{E}}\def\E{\mathbb{E}\left[\xi M({\tilde A})N({\tilde B})|\mathcal{F}_{z_{\tilde A}}^2\right]\right|\mathcal{F}_z^2 \right]= \mathbb{E}}\def\E{\mathbb{E} \left[\left.
\xi N({\tilde B})\mathbb{E}}\def\E{\mathbb{E}\left[M({\tilde A})|\mathcal{F}_{z_{\tilde A}}^2\right]\right|\mathcal{F}_z^2 \right]= 0.
\end{align*}
Similarly, one can show $\mathbb{E}}\def\E{\mathbb{E} \left[ X(D)|\mathcal{F}_z^1 \right] = 0$. Hence, $X$ is a martingale.
Let $z_{\tilde B}$ be the lower left-hand corner of $\tilde B$, and denote $z_0=z_{\tilde A}\vee z_{\tilde B}$. Then
$z_A\vee z_B\prec z_0$, and hence $\xi$ is $\mathcal{F}_{z_0}$-measurable. Thus, by Theorem \ref{Thm-prop1.6},
\[\mathbb{E}}\def\E{\mathbb{E}[X^2(D)|\mathcal{F}_{z}]=\mathbb{E}}\def\E{\mathbb{E}\left[\left.\xi^2\mathbb{E}}\def\E{\mathbb{E}\left[ M(\tilde A)^2N(\tilde B)^2|\mathcal{F}_{z_0}\right] \right|\mathcal{F}_{z}\right]. \]
Now we have
\begin{align*}
\mathbb{E}}\def\E{\mathbb{E} \left[ M({\tilde A})^2 N({\tilde B})^2|\mathcal{F}_{z_0} \right] =& \mathbb{E}}\def\E{\mathbb{E} \left[ M({\tilde A})^2|\mathcal{F}_{z_0} \right] \mathbb{E}}\def\E{\mathbb{E} \left[ N({\tilde B})^2|\mathcal{F}_{z_0} \right] \\
=& \mathbb{E}}\def\E{\mathbb{E} \left[\left. \mathbb{E}}\def\E{\mathbb{E} \left[ M({\tilde A})^2 | \mathcal{F}_{z_0}^2 \right] \right|\mathcal{F}_{z_0} \right] \mathbb{E}}\def\E{\mathbb{E} \left[ \mathbb{E}}\def\E{\mathbb{E}\left. \left[ N({\tilde B})^2|\mathcal{F}_{z_0}^1 \right] \right|\mathcal{F}_{z_0} \right] \\
=& \mathbb{E}}\def\E{\mathbb{E} \left[ [M]^2({\tilde A}) |\mathcal{F}_{z_0} \right] \mathbb{E}}\def\E{\mathbb{E} \left[ [N]^1({\tilde B}) |\mathcal{F}_{z_0} \right] \\
=& \mathbb{E}}\def\E{\mathbb{E} \left[ [M]^2({\tilde A})[N]^1({\tilde B}) |\mathcal{F}_{z_0} \right],
\end{align*}
where the first and the last equalities follow from the assumption that $\mathcal{F}_{z_0}^1$ and $\mathcal{F}_{z_0}^2$ are conditionally
independent given $\mathcal{F}_{z_0}$, and the third equality follows from Theorem \ref{Thm-prop1.8}. Thus
$$\mathbb{E}}\def\E{\mathbb{E}[X^2(D) - \xi^2 [M]^2({\tilde A})[N]^1({\tilde B})|\mathcal{F}_z]=0$$
and hence $X_z^2-\langle X\rangle_z$ is a weak martingale where $\langle X\rangle_z$ is given by \eqref{e:quad-var}.
The proof is concluded.
\end{proof}
With Proposition \ref{Prop 2.1} in mind, we define a new type of stochastic integral denoted by $\psi\cdot MN$,
following the approach in \cite{CW75}.
Fix an integer $n$ and $z_0=(s_0, t_0)\in\mathbb{R}_+^2$. Divide $R_{z_0}$ into rectangles $\rectangle_{i,j} = (z_{i,j},z_{i+1,j+1}]$,
where $z_{i,j} = (is_0/2^n,jt_0/2^n)$ for $i, j=0,1, \dots, 2^n-1$. We first define $\psi \cdot MN$ for an indicator function $\psi.$
If $i,j,k,l$ are positive integers with $1\le i < k \le 2^n$ and $1\le l < j \le 2^n$, i,e. $\rectangle_{i,j}\curlywedge \rectangle_{k,l} $,
define the so-called indicator function
\begin{align}\label{eq:psi}
\psi_{ijkl}(z_1,z_2) = \xi \mathbf 1_{\rectangle_{i,j}}(z_1) \mathbf 1_{\rectangle_{k,l}}(z_2),
\end{align}
where $\xi$ is bounded and $\mathcal{F}_{z_{k,j}}$-measurable,
and define
\begin{align*}
(\psi_{ijkl} \cdot MN)_z
= \xi M(\rectangle_{i,j} \cap R_z) N(\rectangle_{k,l} \cap R_z), \ \ \ \ z \in R_{z_0}.
\end{align*}
Then by Proposition \ref{Prop 2.1}, $\psi_{ijkl} \cdot MN$ is a well-defined square integrable martingale with quadratic variation
\begin{align*}
\langle \psi_{ijkl} \cdot MN \rangle_z
= \iint_{R_z \times R_z} \psi_{ijkl}^2(z_1,z_2) d[ M]^2_{z_1} d[ N]^1_{z_2},
\end{align*}
and thus we have the following isometry
\begin{equation}\label{e:iso}
\E[|\psi_{ijkl}\cdot MN|^2] =\E\left[\iint_{R_z \times R_z} \psi_{ijkl}^2(z_1,z_2) d[ M]^2_{z_1} d[ N]^1_{z_2}\right].
\end{equation}
We shall define $\psi\cdot MN$ for a more general class of integrands $\psi$ following the standard approximation procedure. For this purpose, one needs the isometry \eqref{e:iso} to hold for finite sum of indicator functions, and it suffices to prove the following equality
\begin{align}\label{e:qua-var'}
\langle \psi_{ijkl} \cdot MN, \psi_{mpqr} \cdot MN \rangle_z
= \iint_{R_z \times R_z} \psi_{ijkl}(z_1,z_2) \psi_{mpqr}(z_1,z_2) d[M]^2_{z_1} d[ N]^1_{z_2}.
\end{align}
Here, $\psi_{mpqr}(z_1, z_2)=\eta \mathbf 1_{\rectangle_{m,p}}(z_1) \mathbf 1_{\rectangle_{q,r}}(z_2)$ with $m<q\le 2^n, r<p\le 2^n$ and
$\eta$ being a bounded $\mathcal{F}_{z_{q,p}}$-measurable random variable. To prove \eqref{e:qua-var'}, we consider the following more general situation.
Suppose $M,N,M',N' \in \mathcal \mathfrak M_s^2(z_0)$, and let $(A,B)$ and $(A',B')$ be two pairs of rectangles satisfying the conditions
in Proposition 2.1, i.e., $A\curlywedge B$ and $A'\curlywedge B'$. Furthermore, we assume $A, A', B, B'$ are from the set
$\{\rectangle_{i,j}, i,j=0, 1, \dots, 2^n-1\}$. Thus, any two of the rectangles $A, A', B, B'$ are either coincide or disjoint. Denote
$z_0=(z_A\vee z_B)\vee (z_{A'}\vee z_{B'})$. We claim that the following equality holds
\begin{align} \label{e:increment}
&\mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A)M'(A')N(B)N'(B') \right| \mathcal{F}_{z_0} \right]
+ \mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A')M'(A)N(B)N'(B') \right| \mathcal{F}_{z_0} \right] \nonumber \\
&\qquad + \mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A)M'(A')N(B')N'(B) \right| \mathcal{F}_{z_0} \right]
+ \mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A')M'(A)N(B')N'(B) \right| \mathcal{F}_{z_0} \right] \nonumber \\
&= \mathbb{E}}\def\E{\mathbb{E} \left[ \left. \Big( M(A)M'(A') + M(A')M'(A) \Big) \Big( N(B)N'(B') + N(B')N'(B) \Big) \right| \mathcal{F}_{z_0} \right] \nonumber \\
&= 4 \mathbb{E}}\def\E{\mathbb{E} \Big[ \left. [M,M']^2(A \cap A') [N,N']^1(B \cap B') \right| \mathcal{F}_{z_0} \Big].
\end{align}
\begin{proof}[\it Proof of (\ref{e:increment})]
The first equality is straightforward. In the following, we shall prove the second equality.
Recall that the four rectangles $A, A', B, B'$ are either disjoint or coincide; furthermore, $A\curlywedge B$ and $A'\curlywedge B'$,
i.e., $A$ (resp. $A'$) is on the upper left side of $B$ (resp. $B'$). We prove the second inequality in \eqref{e:increment}
by separating the relative locations of $A, A', B',B'$ into four cases. In the following, we denote the lower left corner of a rectangle
$E$ by $z_E$.
{\bf Case 1.} If $A$ is on the top of $A'$ and $A\cap A'=\emptyset$, noting that $A$ (resp. $A'$) is to the upper left of $B$ (resp. $B'$),
we have that $M'(A'), N(B), N'(B')$ are all $\mathcal{F}_{z_A}^2$-measurable. Since $M$ is a $2$-martingale, we have, noting that
$\mathcal{F}_{z_0}\subset \mathcal{F}_{z_A}^2$
\begin{align*}
\mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A)M'(A')N(B)N'(B') \right| \mathcal{F}_{z_0} \right]
&= \mathbb{E}}\def\E{\mathbb{E} \Big[ \left. \mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A)M'(A')N(B)N'(B') \right| \mathcal{F}_{z_A}^2 \right] \right| \mathcal{F}_{z_0} \Big] \\
&= \mathbb{E}}\def\E{\mathbb{E} \Big[ \left. M'(A')N(B)N'(B') \mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A) \right| \mathcal{F}_{z_A}^2 \right] \right| \mathcal{F}_{z_0} \Big] \\
&= 0.
\end{align*}
Similarly, for the other terms on the left-hand side of \eqref{e:increment} we also have
\begin{align*}
&\mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A')M'(A)N(B)N'(B') \right| \mathcal{F}_{z_0} \right] = 0, \\
&\mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A)M'(A')N(B')N'(B) \right| \mathcal{F}_{z_0} \right] = 0, \\
&\mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A')M'(A)N(B')N'(B) \right| \mathcal{F}_{z_0} \right] = 0.
\end{align*}
Summing over all the above equalities, we get \eqref{e:increment}.
{\bf Case 2.} If $A'$ is on the top of $A$ and $A\cap A'=\emptyset$, the proof is the same by considering the $\sigma$-field
$\mathcal{F}_{z_{A'}}^2$. If $B$ is to the right (resp. left) of $B'$ with $B\cap B'=\emptyset$, then the proof is also the same by considering
the $\sigma$-field $\mathcal{F}_{z_B}^1$ (resp. $\mathcal{F}_{z_{B'}}^1$).
{\bf Case 3.} Now we only have one situation left: $A$ and $A'$ are at the same horizontal level, which is on the top of $B$ and
$B'$, and $B$ and $B'$ are at the same vertical level, which is to the right of $A$ and $A'$. We denote $z_0 := z_A \vee z_B
=z_{A'}\vee z_{B'}$. Note that $M(A), M'(A'), M(A'), M'(A)$ are $\mathcal{F}_{z_0}^1$ measurable and $N(B), N'(B'), N(B'), N'(B)$
are $\mathcal{F}_{z_0}^2$ measurable. We have
\begin{align} \label{e:increment'}
&\quad \mathbb{E}}\def\E{\mathbb{E} \left[ \left. \Big( M(A)M'(A') + M(A')M'(A) \Big) \Big( N(B)N'(B') + N(B')N'(B) \Big) \right| \mathcal{F}_{z_0} \right] \nonumber \\
&= \mathbb{E}}\def\E{\mathbb{E} \Big[ \left. \mathbb{E}}\def\E{\mathbb{E} \left[ M(A)M'(A') + M(A')M'(A) \right| \mathcal{F}_{z_0} \right] \mathbb{E}}\def\E{\mathbb{E} \left[ \left. N(B)N'(B') + N(B')N'(B) \right|
\mathcal{F}_{z_0} \right] \Big| \mathcal{F}_{z_0} \Big],
\end{align}
where the equality follows from the conditional independence of $\mathcal{F}_{z_0}^1$ and $\mathcal{F}_{z_0}^2$ given $\mathcal{F}_{z_0}$.
To compute
\begin{align*}
\mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A)M'(A') + M(A')M'(A) \right| \mathcal{F}_{z_0} \right],
\end{align*}
we split it into the following three cases.
(a) If $A = A'$, noting that $\mathcal{F}_{z_0}^2 = \mathcal{F}_{z_A}^2$, by Theorem \ref{Thm-prop1.8},
\begin{align} \label{eq-conditional A=A'}
\mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A) M'(A) \right| \mathcal{F}_{z_0}^2 \right]
= \mathbb{E}}\def\E{\mathbb{E} \left[ \left. (MM')(A) \right| \mathcal{F}_{z_0}^2 \right]
= \mathbb{E}}\def\E{\mathbb{E} \left[ \left. [M,M']^2(A) \right| \mathcal{F}_{z_0}^2 \right].
\end{align}
(b) If $A$ and $A'$ are two adjacent disjoint rectangles on the same horizontal level, then $A \cup A'$ is also a rectangle.
Without loss of generality, we may assume that $A$ is to the left of $A'$, then $z_A$ is also the lower left corner of
$A \cup A'$. Thus, by Case (a), we have
\begin{align*}
&\mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A) M'(A) \right| \mathcal{F}_{z_0}^2 \right]
= \mathbb{E}}\def\E{\mathbb{E} \left[ \left. [M,M']^2(A) \right| \mathcal{F}_{z_0}^2 \right], \\
&\mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A') M'(A') \right| \mathcal{F}_{z_0}^2 \right]
= \mathbb{E}}\def\E{\mathbb{E} \left[ \left. [M,M']^2(A') \right| \mathcal{F}_{z_0}^2 \right], \\
&\mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A \cup A') M'(A \cup A') \right| \mathcal{F}_{z_0}^2 \right]
= \mathbb{E}}\def\E{\mathbb{E} \left[ \left. [M,M']^2(A \cup A') \right| \mathcal{F}_{z_0}^2 \right].
\end{align*}
Noting that $M(A \cup A') = M(A) + M(A')$, $M'(A \cup A') = M'(A) + M'(A')$ and $[M,M']^i(A \cup A') = [M,M']^i(A)
+ [M,M']^i(A')$, we subtract the first two equations from the third one and obtain
\begin{align} \label{eq-eq-conditional A not= A'}
\mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A)M'(A') + M(A')M'(A) \right| \mathcal{F}_{z_0}^2 \right] = 0.
\end{align}
(c) If $A$ and $A'$ are two non-adjacent rectangles on the same horizontal level, we denote by $A''$ the rectangle
between $A$ and $A'$. Note that $A''$ is the union of small rectangles in the set $\{\rectangle_{i, j}, i, j=1, \dots, 2^n\}$.
By Case (b), we have
\begin{align*}
&\mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A)M'(A'') + M(A'')M'(A) \right| \mathcal{F}_{z_0}^2 \right] = 0, \\
&\mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A)M'(A'' \cup A') + M(A'' \cup A')M'(A) \right| \mathcal{F}_{z_0}^2 \right] = 0.
\end{align*}
Noting that $M'(A'' \cup A') = M'(A'') + M'(A')$, one can subtract the first equality from the second one to obtain
\eqref{eq-eq-conditional A not= A'}.
Therefore, summarizing the three cases (a-c), we can write
\begin{align*}
\mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A)M'(A') + M(A')M'(A) \right| \mathcal{F}_{z_0}^2 \right]
= 2 \mathbb{E}}\def\E{\mathbb{E} \left[ \left. [M,M']^2(A \cap A') \right| \mathcal{F}_{z_0}^2 \right].
\end{align*}
Hence, by taking conditional expectation with respect to the $\sigma$-field $\mathcal{F}_{z_0}$, we have
\begin{align} \label{eq-quadratic on M A}
\mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A)M'(A') + M(A')M'(A) \right| \mathcal{F}_{z_0} \right]
= 2 \mathbb{E}}\def\E{\mathbb{E} \left[ \left. [M,M']^2(A \cap A') \right| \mathcal{F}_{z_0} \right].
\end{align}
In the same spirit, we can also prove
\begin{align} \label{eq-quadratic on N B}
\mathbb{E}}\def\E{\mathbb{E} \left[ \left. N(B)N'(B') + N(B')N'(B) \right| \mathcal{F}_{z_0} \right]
= 2 \mathbb{E}}\def\E{\mathbb{E} \left[ \left. [N,N']^1(B \cap B') \right| \mathcal{F}_{z_0} \right].
\end{align}
Finally, substituting \eqref{eq-quadratic on M A} and \eqref{eq-quadratic on N B} into \eqref{e:increment'}, we have
\begin{align*}
& \mathbb{E}}\def\E{\mathbb{E} \left[ \left. \left( M(A)M'(A') + M(A')M'(A) \right) \left( N(B)N'(B') + N(B')N'(B) \right) \right| \mathcal{F}_{z_0} \right] \\
&= 4 \mathbb{E}}\def\E{\mathbb{E} \left[ \left. \mathbb{E}}\def\E{\mathbb{E} \left[ \left. [M,M']^2(A \cap A') \right| \mathcal{F}_{z_0} \right] \mathbb{E}}\def\E{\mathbb{E} \left[ \left. [N,N']^1(B \cap B') \right|
\mathcal{F}_{z_0} \right] \right| \mathcal{F}_{z_0} \right] \\
&= 4 \mathbb{E}}\def\E{\mathbb{E}
\left[ \left. [M,M']^2(A \cap A') [N,N']^1(B \cap B') \right| \mathcal{F}_{z_0} \right],
\end{align*}
where the conditional independence of $\mathcal{F}_{z_0}^1$ and $\mathcal{F}_{z_0}^2$ given $\mathcal{F}_{z_0}$ is used again in the last equality.
This proves \eqref{e:increment}.
\end{proof}
By choosing $M'=M$ and $N'=N$, eq. \eqref{e:increment} degenerates to
\begin{align} \label{e:increment1}
\mathbb{E}}\def\E{\mathbb{E} \left[ \left. M(A)M(A')N(B)N(B') \right| \mathcal{F}_{z_0} \right]
=& \mathbb{E}}\def\E{\mathbb{E} \Big[ \left. [M,M]^2(A \cap A') [N,N]^1(B \cap B') \right| \mathcal{F}_{z_0} \Big].
\end{align}
Now, as in Proposition \ref{Prop 2.1}, we can define
\begin{equation}\label{e:xx'}
X_z = \xi M(A\cap R_z) N(B \cap R_z)~ \text{ and }~ X'_z = \xi' M'(A'\cap R_z) N'(B' \cap R_z)
\end{equation}
for some bounded variables $\xi\in\mathcal F_{z_A\vee z_B}$ and $\xi'\in\mathcal F_{z_{A'}\vee z_{B'}}$. Denote
$z_0:=(z_A\vee z_B)\vee (z_{A'}\vee z_{B'})$ and we assume $z_0\prec z=(s,t)$, since otherwise at least one of
$X_z$ and $X_z'$ is zero. Let $z'=(s',t')$ be such that $z\prec\prec z'$ and let $D:=(z, z']$.
Assuming $M=M'$ and $N=N'$ in \eqref{e:xx'}, following the approach used in the proof of Proposition \ref{Prop 2.1},
we can show by \eqref{e:increment},
\[\E[(XX')(D)|\mathcal{F}_z] =
\xi\xi' \mathbb{E}}\def\E{\mathbb{E} \Big[ \left. [M,M]^2(\tilde A \cap \tilde A') [N,N]^1(\tilde B \cap \tilde B') \right| \mathcal{F}_{z} \Big], \]
where $\tilde A = A \cap (R_{s't'}\backslash R_{s't})$, ${\tilde B} = B \cap (R_{s't'}\backslash R_{st'})$, and
$\tilde A' = A' \cap (R_{s't'}\backslash R_{s't})$ and ${\tilde B'} = B' \cap (R_{s't'}\backslash R_{st'})$. This leads to
\begin{align}\label{e:quad-var1}
\langle X, X' \rangle_z
=\xi\xi' \iint_{R_z \times R_z} \mathbf 1_{A\cap A'}(z_1) \mathbf 1_{B\cap B'}(z_2) d[M, M]^2_{z_1} d[N, N]^1_{z_2},
\end{align}
and hence \eqref{e:qua-var'} is verified.
Now we are ready to define $\psi\cdot MN$ for a more general integrand $\psi$. We say $\psi$ is a simple function if it is a finite sum of $\psi_{ijkl}$
given in \eqref{eq:psi}. Let $\mathcal D$ be the $\sigma$-filed on $\mathbb{R}_+^2\times \mathbb{R}_+^2\times \Omega$
generated by all the simple functions. We call $\mathcal D$ the field of predictable sets. Let $\mathcal{L}_{MN}^2(z_0)$
be the class of all predictable processes such that
\begin{align}
\mathbb{E}}\def\E{\mathbb{E} \left[ \iint_{R_{z_0} \times R_{z_0}} \psi^2(z_1,z_2) d[ M]^2_{z_1} d[ N]^1_{z_2}\right] < \infty.
\end{align}
Then $\mathcal{L}_{MN}^2(z_0)$ is a Hilbert space with the inner product
\begin{align}\label{e:isometry}
(\psi, \phi)
= \mathbb{E}}\def\E{\mathbb{E} \left[ \iint_{R_{z_0} \times R_{z_0}} \psi(z_1,z_2) \phi(z_1,z_2) d[ M]^2_{z_1} d[ N]^1_{z_2} \right],
\end{align}
and the simple functions form a dense subset. By \eqref{e:qua-var'} and \eqref{e:isometry}, the mapping
$\psi\mapsto \psi\cdot MN$ defines an isometry between the set of simple functions and $ \mathfrak M^2(z_0)$. Then, by a
standard approximation argument, one can extend the definition of $
\psi\cdot MN$ for each process $\psi\in \mathcal{L}^2_{MN}(z_0)$. Furthermore, \eqref{e:qua-var'} also yields for $z\prec z_0$,
\begin{equation}\label{e:qua-var}
\langle \psi\cdot MN, \phi\cdot MN \rangle_{z}
= \iint_{R_{z} \times R_{z}} \psi(z_1,z_2) \phi(z_1,z_2) d\langle M\rangle_{z_1} d\langle N\rangle_{z_2} , ~~\forall \psi, \phi\in \mathcal{L}^2_{MN}(z_0).
\end{equation}
{\bf Throughout the rest of this section, we only consider continuous strong martingales with finite fourth moments, unless otherwise stated.}
Then based on Theorem \ref{Thm-prop1.9}, we have
\begin{equation}\label{e:quad1=quad2}
[M]^1=[M]^2=\langle M\rangle; ~~[N]^1=[N]^2=\langle N\rangle.
\end{equation}
To end this subsection, we shall follow the approach used in \cite[Section 6]{CW75} to show that $J_{MN}$ defined in \eqref{def-J}
can be represented by $\psi\cdot MN$ with $\psi(z_1, z_2) =\mathbf 1_{[z_1{\curlywedge\atop\curlywedge} z_2]}$.
Recall the notations $z_{i,j} = (is_0/2^n,jt_0/2^n)$ and $\rectangle_{i,j} = (z_{i,j},z_{i+1,j+1}]$. We also denote $\epsilon_{i,j}
= (z_{i0}, z_{i+1,j}]$ and $\delta_{i,j} = (z_{0,j}, z_{i,j+1}]$. Denote
\begin{align}
J_{MN}^n(z)
&:= \sum_{i,j=0}^{2^n-1} M( \delta_{i,j} \cap R_z ) N( \epsilon_{i,j} \cap R_z )\notag \\
&= \sum_{i,j=0}^{2^n-1} \left( \sum_{k=0}^{i-1} M( (z_{k,j},z_{k+1,j+1}] \cap R_z ) \right)
\left( \sum_{l=0}^{j-1} N( (z_{i,l},z_{i+1,l+1}] \cap R_z ) \right)\notag \\
&= \sum_{k<i} \sum_{l<j} \left(\psi_{kjil} \cdot MN\right)_z,\label{e:J^n}
\end{align}
where $\psi_{kjil}$ is given in \eqref{eq:psi}.
Thus, letting $n \rightarrow \infty$, we have
\begin{align*}
J_{MN}^n(z) \rightarrow (\psi \cdot MN)_z,
\end{align*}
where
\begin{align*}
\psi(z_1,z_2) =\mathbf 1_{[z_1{\curlywedge\atop\curlywedge} z_2]}= \begin{cases}
1, &\text{ if } z_1{\curlywedge\atop\curlywedge} z_2, \\
0, & \mathrm{otherwise}.
\end{cases}
\end{align*}
Define $M^n = \sum\limits_{i,j=0}^{2^n-1} \mathbf 1_{\rectangle_{i,j}}(z) M_{z_{i,j}}$. Then $M^n$ is a sequence of simple
functions that approximate $M$ and
\begin{align}
\int_{R_{z_0}} M^n dN &= \sum_{i,j=0}^{2^n-1} M_{z_{i,j}} N(\rectangle_{i,j}) \nonumber \\
&= \sum_{i,j=0}^{2^n-1} M_{z_{i,j}} \left( N(\epsilon_{i,j+1}) - N(\epsilon_{i,j}) \right) \nonumber \\
&= \sum_{i,j=0}^{2^n-1} \left( M_{z_{i,j+1}} N(\epsilon_{i,j+1}) - M_{z_{i,j}} N(\epsilon_{i,j}) \right) +
\sum_{i,j=0}^{2^n-1} \left( M_{z_{i,j}} - M_{z_{i,j+1}} \right) N(\epsilon_{i,j+1}) \nonumber \\
&= \sum_{i=0}^{2^n-1} M_{z_{i,2^n}} N(\epsilon_{i,2^n}) - \sum_{i,j=0}^{2^n-1} M(\delta_{i,j}) \left( N(\epsilon_{i,j})
+ N(\rectangle_{i,j}) \right).\label{e:int-MdN}
\end{align}
If we define $\widetilde{M}_{s,t_0}^n = M_{is_0/2^n,t_0}$ for $s \in (is_0/2^n, (i+1)s_0/2^n]$, and $\delta^n(z) = M(\delta_{i,j})$
if $z \in \rectangle_{i,j}$. Let $H_{z_0}$ be the line segment with endpoints $(0,t_0)$ and $z_0 = (s_0,t_0)$, then
\begin{align}\label{e:0}
\int_{R_{z_0}} M^n dN
= \int_{H_{z_0}} \widetilde{M}_{s,t_0}^n(s)N(ds,t)
- J_{MN}^n(z_0) - \int_{R_{z_0}} \delta^n dN.
\end{align}
By the Cauchy-Schwarz inequality,
\begin{align}
\mathbb{E}}\def\E{\mathbb{E} \left[ \bigg( \int_{R_{z_0}} \delta^n dN \bigg)^2 \right]
&= \mathbb{E}}\def\E{\mathbb{E} \left[ \int_{R_{z_0}} \left( \delta^n(s,t) \right)^2 d \langle N \rangle_{st} \right] \notag\\
&\le \mathbb{E}}\def\E{\mathbb{E} \left[ \int_{R_{z_0}} \sup_{i,j} M(\delta_{i,j})^2 d \langle N \rangle_{st} \right] \notag\\
&\le \mathbb{E}}\def\E{\mathbb{E} \left[ \sup_{i,j} M(\delta_{i,j})^2 \langle N \rangle_{z_0} \right] \notag\\
&\le \left( \mathbb{E}}\def\E{\mathbb{E} \left[ \sup_{i,j} M(\delta_{i,j})^4 \right] \mathbb{E}}\def\E{\mathbb{E} \left[ \langle N \rangle_{z_0}^2 \right] \right)^{1/2}\notag \\
&\to 0, ~ n\to \infty,\label{e:1}
\end{align}
where the last step holds due to the continuity of $M$ and the dominated convergence theorem, noting that
$\mathbb{E}}\def\E{\mathbb{E}[\sup_{i,j,n} M(\delta_{i,j})^4]$ is dominated by $\mathbb{E}}\def\E{\mathbb{E}[\sup_{z\prec z_0} |M_z|^4]$, which is dominated by
$(4/3)^8\mathbb{E}}\def\E{\mathbb{E}[ |M_{z_0}|^4]$ due to Theorem \ref{Thm-BDG 2dim} and the existence of the fourth moment of the $M$.
Furthermore, Theorem \ref{Thm-BDG 2dim} yields
\begin{align*}
\mathbb{E}}\def\E{\mathbb{E} \left[ \sup_n\sup_{z\prec z_0} (M_z^n-M_z)^4 \right]
\le 8\mathbb{E}}\def\E{\mathbb{E} \left[ \sup_n\sup_{z\prec z_0} |M_z^n|^4 + \sup_{z\prec z_0} |M_z|^4 \right]
\le 16 \mathbb{E}}\def\E{\mathbb{E} \left[ \sup_{z\prec z_0} |M_z|^4 \right]
< \infty.
\end{align*}
By the Cauchy-Schwarz inequality, the dominated convergence theorem and the continuity of $M$, we have
\begin{align}
\mathbb{E}}\def\E{\mathbb{E} \left[ \bigg( \int_{R_{z_0}} (M^n-M) dN \bigg)^2 \right]
&= \mathbb{E}}\def\E{\mathbb{E} \left[ \int_{R_{z_0}} (M^n-M)^2 d \langle N \rangle \right] \notag\\
&\le \mathbb{E}}\def\E{\mathbb{E} \left[ \sup_{n,z} (M_z^n-M_z)^2 \langle N \rangle_{z_0} \right] \notag\\
&\le \left( \mathbb{E}}\def\E{\mathbb{E} \left[ \sup_{n,z} (M_z^n-M_z)^4 \right] \mathbb{E}}\def\E{\mathbb{E} \left[ \langle N \rangle_{z_0}^2 \right] \right)^{1/2} \notag\\
&\rightarrow 0, ~n \rightarrow \infty.\label{e:2}
\end{align}
Similarly, we can show the following $L^2$-convergence,
\begin{align}\label{e:3}
\int_{H_{z_0}} \widetilde{M}_{s,t_0}^n N(ds,t)
\rightarrow \int_{H_{z_0}} M(s,t) N(ds,t),~ n \rightarrow \infty.
\end{align}
Recalling that $\lim_{n\to\infty}J_{MN}^n(z) =(\psi \cdot MN)_z$ with $\psi(z_1,z_2) =\mathbf 1_{[z_1{\curlywedge\atop\curlywedge} z_2]}$,
by \eqref{e:0}, \eqref{e:1}, \eqref{e:2} and \eqref{e:3}, we have
\begin{align*}
(\psi \cdot MN)_{z_0} = \int_{H_{z_0}} M(s,t) N(ds,t) - \int_{R_{z_0}} M dN,
\end{align*}
and hence by the definition \eqref{def-J} of $J_{MN}$, we have
\begin{equation}\label{e:J}
J_{MN}(z_0)=(\psi \cdot MN)_{z_0}.
\end{equation}
Therefore, we can calculate $\langle J_{MN}\rangle$ by \eqref{e:qua-var},
\begin{align*}
\langle J_{MN}\rangle_z= \iint_{R_{z} \times R_{z}} \mathbf 1_{[z_1{\curlywedge \atop \curlywedge} z_2]} d\langle M\rangle_{z_1}
d\langle N\rangle_{z_2}=\int_{R_z} d_t\langle M\rangle_{st} d_s\langle N \rangle_{st},
\end{align*}
and hence
\begin{equation}\label{e:quad-var-J}
d\langle J_{MN}\rangle_{st} = d_t\langle M\rangle_{st} d_s\langle N \rangle_{st}.
\end{equation}
Furthermore, the following equality holds,
\begin{align}\label{def-J'}
J_{MN}(s_0,t_0)
=& \int_0^{s_0} M(s,t_0) N(ds,t_0) - \int_{R_{s_0t_0}} M(s,t) dN(s,t)\notag\\
=& \int_0^{t_0} N(s_0,t) M(s_0,dt) - \int_{R_{s_0t_0}} N(s,t) dM(s,t).
\end{align}
This can be deduced by rewriting \eqref{e:int-MdN} as follows
\begin{align*}
\int_{R_{z_0}} M^n dN
&= \sum_{i,j=0}^{2^n-1} M_{z_{i,j}} N(\rectangle_{i,j}) \nonumber \\
&= \sum_{i,j=0}^{2^n-1} M_{z_{i,j}} \left( N(\delta_{i+1,j}) - N(\delta_{i,j}) \right) \nonumber \\
&= \sum_{i,j=0}^{2^n-1} \left( M_{z_{i+1,j}} N(\delta_{i+1,j}) - M_{z_{i,j}} N(\delta_{i,j}) \right) +
\sum_{i,j=0}^{2^n-1} \left( M_{z_{i,j}} - M_{z_{i+1,j}} \right) N(\delta_{i+1,j}) \nonumber \\
&= \sum_{j=0}^{2^n-1} M_{z_{2^n, j}} N(\delta_{2^n,j}) - \sum_{i,j=0}^{2^n-1} M(\epsilon_{i,j}) \left( N(\delta_{i,j}) + N(\rectangle_{i,j}) \right).
\end{align*}
By letting $n$ go to infinity, we get for $\psi(z_1, z_2)=\mathbf 1_{[z_1{\curlywedge \atop \curlywedge} z_2]}$,
\[(\psi\cdot NM)_{z_0}= \int_0^{t_0} M(s_0, t) N(s_0, dt)-\int_{R_{z_0}} MdN.\]
This together with \eqref{e:J} implies \eqref{def-J'}.
\subsection{Multi-dimensional Green's formula for martingales on the plane}
Now we are ready to prove Theorem \ref{Thm-Green formula}, the multi-dimensional Green's formula on the plane.
Let $\{M^{(i)}(s,t), (s,t)\in\mathbb{R}_+^2\}_{1 \le i \le d}$ be a family of independent continuous strong martingales on $\mathbb{R}_+^2$
with finite fourth moment. We assume that the increasing process $\langle M^{(i)} \rangle$ is deterministic for every$1\le i \le d$.
Let $F_j = F_j(s,t), 1 \le j \le d$ be a sequence of predictable processes of the form,
\begin{align} \label{eq-stochastic partial derivative}
F_j(s,t) = F_j(s,0) + \sum_{i=1}^d \int_0^t f_{j,i}(s,r) M^{(i)}(s,dr) + \int_0^t f_{j,0} (s,r) dr,
\end{align}
where $f_{j,i}, 1\le j \le d, 0\le i\le d$ are $\mathcal{F}$-predictable processes.
\begin{theorem} \label{Thm-Green formula}
Fix $s_0, t_0 > 0$. Suppose that $\{F_j(s,t)\}_{1 \le j \le d}$ are predictable processes given by \eqref{eq-stochastic partial derivative}.
Assume
\begin{align}\label{e:c1}
\mathbb{E}}\def\E{\mathbb{E} \left[ \int_0^{s_0} \int_0^{t_0} f_{j,i}(s,t)^2 d_t\big\langle M^{(i)} \big\rangle_{s_0 t} d_s\big\langle M^{(j)} \big\rangle_{st_0} \right]
< \infty, \qquad \forall 1 \le i, j \le d,
\end{align}
and
\begin{align}\label{e:c2}
\mathbb{E}}\def\E{\mathbb{E} \left[ \int_0^{t_0} \int_0^{s_0} f_{j,0}(s,t)^2 d_s \big\langle M^{(i)}\big\rangle_{st_0} dt \right] < \infty, \qquad \forall 1 \le j \le d.
\end{align}
Then for any rectangle $A \subseteq R_{s_0 t_0}$, we have
\begin{align} \label{eq-Thm1}
\sum_{j=1}^d \int_{\partial A} F_j(s,t) M^{(j)}(ds,t) &= \sum_{j=1}^d \int_A F_j(s,t) dM^{(j)}(s,t)
+ \sum_{i,j=1}^d \int_A f_{j,i}(s,t) dJ_{M^{(i)},M^{(j)}}(s,t) \nonumber \\
& \qquad + \sum_{j=1}^d \int_A f_{j,0}(s,t) M^{(j)} (ds,t) dt.
\end{align}
\end{theorem}
\begin{proof}
We will follow the argument in the proof of \cite[Theorem 6.1]{CW75}. Let $A = [s_1, s_2] \times [t_1, t_2]\subset [0,s_0]\times[0,t_0]$.
Without loss of generality, we may assume that $F_j = 0$ on the line segment with endpoints $(s_1,t_1)$ and $(s_2,t_1)$.
Indeed, noting $F_j(s,t)= F_j(s, t_1) + (F_j(s,t)-F_j(s, t_1)),$ it follows from
\begin{align*}
\int_A F_j(s,t_1) dM^{(j)}(s,t)
= \int_{s_1}^{s_2} F_j(s,t_1) \left( M^{(j)}(ds,t_2) - M^{(j)}(ds,t_1) \right)
= \int_{\partial A} F_j(s,t_1) M^{(j)}(ds,t),
\end{align*}
that \eqref{eq-Thm1} holds for $F_j(s,t)$ if and only if
it holds for $F_j(s,t) - F_j(s,t_1)$.
Next, we consider the case that each stochastic partial derivative $f_{j,i}(s,t)\equiv f_{j,i}\in \mathcal{F}_{s_1, t_1}$ is a constant function
for $1 \le j \le d, 0 \le i \le d$. Then by \eqref{eq-stochastic partial derivative}, we have
\begin{align} \label{eq-2.18}
F_j(s,t) = \sum_{i=1}^d f_{j,i} \left( M^{(i)}(s,t) - M^{(i)}(s,t_1) \right) + f_{j,0}(t-t_1), (s,t) \in A, 1 \le j \le d.
\end{align}
On one hand, noting that $J_{MN}(A)=J_{MN}(s_2,t_2)-J_{MN}(s_1,t_2)-J_{MN}(s_2,t_1)+J_{MN}(s_1,t_1)$, it follows from \eqref{def-J} that
\begin{align} \label{eq-2.19}
& \int_A f_{j,i}(s,t) dJ_{M^{(i)},M^{(j)}}(s,t) \nonumber \\
&= \int_{\partial A} f_{j,i} M^{(i)}(s,t) M^{(j)}(ds,t) - \int_A f_{j,i} M^{(i)}(s,t) dM^{(j)}(s,t) \nonumber \\
&=\int_{\partial A} f_{j,i} \left( M^{(i)}(s,t) - M^{(i)}(s,t_1) \right) M^{(j)}(ds,t) - \int_A f_{j,i} \left( M^{(i)}(s,t) - M^{(i)}(s,t_1) \right) dM^{(j)}(s,t).
\end{align}
Here $\int_{\partial A}$ is a line integral on $\partial A$ with clockwise as its positive direction.
On the other hand, It\^{o}'s formula yields
\begin{align} \label{eq-2.20}
& \int_A f_{j,0} (t-t_1) dM^{(j)}(s,t) \nonumber \\
&= f_{j,0} \int_{t_1}^{t_2} (t-t_1) \left( M^{(j)}(s_2,dt) - M^{(j)}(s_1,dt) \right) \nonumber \\
&= f_{j,0} (t_2-t_1) \left( M^{(j)}(s_2,t_2) - M^{(j)}(s_1,t_2) \right) - f_{j,0} \int_{t_1}^{t_2} \left( M^{(j)}(s_2,t) - M^{(j)}(s_1,t) \right) dt \nonumber \\
&= \int_{\partial A} f_{j,0} (t-t_1) M^{(j)}(ds,t) - \int_{t_1}^{t_2} \left( \int_{s_1}^{s_2} f_{j,0} M^{(j)}(ds,t) \right) dt.
\end{align}
By \eqref{eq-2.18}, \eqref{eq-2.19} and \eqref{eq-2.20}, we get \eqref{eq-Thm1}. Thus, we have proved the theorem
for the case that all stochastic partial derivatives are constant functions. Note that for $A=\cup_{i=1}^k A_i$ where
$A_i$ are disjoint rectangles, one has $\int_{\partial A}=\sum_{i=1}^k \int_{\partial A_i} $ and $\int_A=\sum_{i=1}^k \int_{A_i}$.
Therefore, \eqref{eq-Thm1} also holds for the case that all stochastic partial derivatives are simple functions.
For the general case, recall that the martingales $\{M^{(i)}\}_{1 \le i \le d}$ are independent and the
increasing processes $\{\left\langle M^{(i)} \right\rangle\}_{1 \le i \le d}$ are deterministic. By \eqref{e:c1} and \eqref{e:c2}, for $0 \le i \le d$, $1 \le j \le d$,
we can find sequences $\{f_{j,i}^{(n)}\}_{n \in \mathbb{N}}$ of bounded simple functions such that as $n\to \infty$,
\begin{align} \label{eq-condition on stochastic derivative 1}
\int_0^{s_0} \int_0^{t_0} \mathbb{E}}\def\E{\mathbb{E} \left[ \left( f_{j,i}^{(n)}(s,t) - f_{j,i}(s,t) \right)^2 \right] d_t\big\langle M^{(i)} \big\rangle_{s_0t}d_s
\big\langle M^{(j)} \big\rangle_{st_0} \rightarrow 0, ~1 \le i,j \le d,
\end{align}
and
\begin{align} \label{eq-condition on stochastic derivative 2}
\int_0^{s_0} \int_0^{t_0} \mathbb{E}}\def\E{\mathbb{E} \left[ \left( f_{j,0}^{(n)}(s,t) - f_{j,0}(s,t) \right)^2 \right] dt d_s \big\langle M^{(j)}\big\rangle_{st_0}
\rightarrow 0, \qquad 1 \le j \le d.
\end{align}
Define
\begin{align*}
F_j^{(n)}(s,t) = \sum_{i=1}^d \int_0^t f_{j,i}^{(n)}(s,r) M^{(i)}(s,dr) + \int_0^t f_{j,0}^{(n)} (s,r) dr, ~1 \le j \le d.
\end{align*}
Then \eqref{eq-Thm1} holds for the family $\big\{ F_j^{(n)} \big\}_{1 \le j \le d}$, and it remains to take the limit as $n \rightarrow \infty$.
We deal with the left-hand side of \eqref{eq-Thm1} first. It follows from \eqref{eq-condition on stochastic derivative 1} that,
as $n \rightarrow \infty$, for $1 \le i, j \le d$
\begin{align} \label{eq-2.23}
& \mathbb{E}}\def\E{\mathbb{E} \left[ \left( \int_{\partial A} \int_0^t f_{j,i}(s,r) M^{(i)}(s,dr) M^{(j)}(ds,t) - \int_{\partial A} \int_0^t f_{j,i}^{(n)}(s,r) M^{(i)}(s,dr) M^{(j)}(ds,t) \right)^2 \right] \nonumber \\
&\le 2\sum_{k=1,2} \mathbb{E}}\def\E{\mathbb{E} \left[ \left( \int_{s_1}^{s_2} \int_0^{t_k} \left( f_{j,i}(s,r) - f_{j,i}^{(n)}(s,r) \right) M^{(i)}(s,dr) M^{(j)}(ds,t_k) \right)^2 \right] \nonumber \\
&= 2\sum_{k=1,2} \int_{s_1}^{s_2} \int_0^{t_k} \mathbb{E}}\def\E{\mathbb{E} \left[ \left( f_{j,i}(s,r) - f_{j,i}^{(n)}(s,r) \right)^2 \right] d_r\big \langle M^{(i)} \big\rangle_{sr} d_s \big\langle M^{(j)} \big\rangle_{st_k} \nonumber \\
&\le 2\sum_{k=1,2} \int_{s_1}^{s_2} \int_0^{t_k} \mathbb{E}}\def\E{\mathbb{E} \left[ \left( f_{j,i}(s,r) - f_{j,i}^{(n)}(s,r) \right)^2 \right] d_r\big\langle M^{(i)} \big\rangle_{s_0r} d_s \big\langle M^{(j)} \big\rangle_{st_0} \nonumber \\
&\rightarrow 0.
\end{align}
Similarly, by \eqref{eq-condition on stochastic derivative 2}, we have as $n\to\infty$, for $1\le j\le d$,
\begin{align} \label{eq-2.24}
&\mathbb{E}}\def\E{\mathbb{E} \left[ \left( \int_{\partial A} \int_0^t f_{j,0}(s,r) dr M^{(j)}(ds,t) - \int_{\partial A} \int_0^t f_{j,0}^{(n)}(s,r) dr M^{(j)}(ds,t) \right)^2 \right] \nonumber \\
&\le 2\sum_{k=1,2} \int_{s_1}^{s_2} \mathbb{E}}\def\E{\mathbb{E} \left[ \left( \int_0^{t_k} \left( f_{j,0}(s,r) - f_{j,0}^{(n)}(s,r) \right) dr \right)^2 \right] d_s\big\langle M^{(j)} \big\rangle_{st_0} \nonumber \\
&\le 4 \int_{s_1}^{s_2} t_0 \int_0^{t_0} \mathbb{E}}\def\E{\mathbb{E} \left[ \left( f_{j,0}(s,r) - f_{j,0}^{(n)}(s,r) \right)^2 \right] dr d_s\big\langle M^{(j)} \big\rangle_{st_0} \nonumber \\
&\rightarrow 0.
\end{align}
Hence, combing \eqref{eq-2.23} with \eqref{eq-2.24}, we get
\begin{align*}
\lim_{n \rightarrow \infty} \sum_{j=1}^d \int_{\partial A} F_j^{(n)}(s,t) M^{(j)}(ds,t) = \sum_{j=1}^d \int_{\partial A} F_j(s,t) M^{(j)}(ds,t)
\end{align*}
in $L^2(\Omega)$.
Next, we deal with the first term on the right-hand side of \eqref{eq-Thm1}. By \eqref{eq-condition on stochastic derivative 1},
\begin{align} \label{eq-2.25}
& \mathbb{E}}\def\E{\mathbb{E} \left[ \left( \int_A \int_0^t f_{j,i}(s,r) M^{(i)}(s,dr) dM^{(j)}(s,t) - \int_A \int_0^t f_{j,i}^{(n)}(s,r) M^{(i)}(s,dr) dM^{(j)}(s,t) \right)^2 \right] \nonumber \\
&= \int_A \mathbb{E}}\def\E{\mathbb{E} \left[ \left( \int_0^t \left( f_{j,i}(s,r) - f_{j,i}^{(n)}(s,r) \right) M^{(i)}(s,dr) \right)^2 \right] d\big\langle M^{(j)} \big\rangle_{st} \nonumber \\
&= \int_A \int_0^t \mathbb{E}}\def\E{\mathbb{E} \left[ \left( f_{j,i}(s,r) - f_{j,i}^{(n)}(s,r) \right)^2 \right] d_r\big\langle M^{(i)}\big\rangle_{sr} d\big\langle M^{(j)} \big\rangle_{st} \nonumber \\
&\le \int_A \int_0^{t_0} \mathbb{E}}\def\E{\mathbb{E} \left[ \left( f_{j,i}(s,r) - f_{j,i}^{(n)}(s,r) \right)^2 \right] d_r\big\langle M^{(i)}\big\rangle_{sr} d\big \langle M^{(j)} \big\rangle_{st} \nonumber \\
&\le \int_{0}^{s_0} \int_0^{t_0} \mathbb{E}}\def\E{\mathbb{E} \left[ \left( f_{j,i}(s,r) - f_{j,i}^{(n)}(s,r) \right)^2 \right] d_r\big \langle M^{(i)}\big\rangle_{s_0r} d_s\big\langle M^{(j)} \big\rangle_{st_0} \nonumber \\
&\rightarrow 0, \quad n \rightarrow \infty,\ \ \ \forall 1 \le i, j \le d.
\end{align}
Similarly, by \eqref{eq-condition on stochastic derivative 2},
\begin{align} \label{eq-2.26}
&\mathbb{E}}\def\E{\mathbb{E} \left[ \left( \int_A \int_0^t f_{j,0}(s,r) dr dM^{(j)}(s,t) - \int_A \int_0^t f_{j,0}^{(n)}(s,r) dr dM^{(j)}(s,t) \right)^2 \right] \nonumber \\
&\le \int_A t_0 \int_0^{t_0} \mathbb{E}}\def\E{\mathbb{E} \left[ \left( f_{j,0}(s,r) - f_{j,0}^{(n)}(s,r) \right)^2 \right] dr d\big \langle M^{(j)} \big\rangle_{st} \nonumber \\
&\le t_0 \int_{0}^{s_0} \int_0^{t_0} \mathbb{E}}\def\E{\mathbb{E} \left[ \left( f_{j,0}(s,r) - f_{j,0}^{(n)}(s,r) \right)^2 \right] dr d_s \big\langle M^{(j)} \big\rangle_{st_0}\nonumber \\
&\rightarrow 0, \quad n \rightarrow \infty,\ \forall 1 \le j \le d.
\end{align}
Hence, \eqref{eq-2.25} and \eqref{eq-2.26} imply
\begin{align*}
\lim_{n \rightarrow \infty} \sum_{j=1}^d \int_A F_j^{(n)}(s,t) M^{(j)}(ds,t) \rightarrow \sum_{j=1}^d \int_A F_j(s,t) M^{(j)}(ds,t)
\end{align*}
in $L^2(\Omega)$.
Next we deal with the limit of the second term on the right hand side of \eqref{eq-Thm1}. By \eqref{e:quad-var-J} and
\eqref{eq-condition on stochastic derivative 1}, we have
\begin{align} \label{eq-2.27}
&\mathbb{E}}\def\E{\mathbb{E} \left[ \left( \int_A f_{j,i}(s,t) dJ_{M^{(i)},M^{(j)}}(s,t) - \int_A f_{j,i}^{(n)}(s,t) dJ_{M^{(i)},M^{(j)}}(s,t) \right)^2 \right] \nonumber \\
&\le \mathbb{E}}\def\E{\mathbb{E} \left[ \int_A \left| f_{j,i}(s,t) - f_{j,i}^{(n)}(s,t) \right|^2 d_t \big\langle M^{(i)} \big\rangle_{s_0t} d_s \big\langle M^{(j)}\big\rangle_{st_0} \right] \nonumber \\
&\rightarrow 0, \quad n \rightarrow \infty, \ ~\forall 1 \le i, j \le d.
\end{align}
Hence, we have
\begin{align*}
\sum_{i,j=1}^d \int_A f_{j,i}^{(n)}(s,t) dJ_{M^{(i)},M^{(j)}}(s,t) \rightarrow \sum_{i,j=1}^d \int_A f_{j,i}(s,t) dJ_{M^{(i)},M^{(j)}}(s,t)
\end{align*}
in $L^2(\Omega)$.
Lastly, we deal with the limit of the third term on the right hand side of \eqref{eq-Thm1}. By the Cauchy-Schwarz inequality
and \eqref{eq-condition on stochastic derivative 2},
\begin{align} \label{eq-2.29}
& \mathbb{E}}\def\E{\mathbb{E} \left[ \left( \int_A f_{j,0}(s,t) M^{(j)} (ds,t) dt - \int_A f_{j,0}^{(n)}(s,t) M^{(j)} (ds,t) dt \right)^2 \right] \nonumber \\
&= \mathbb{E}}\def\E{\mathbb{E} \left[ \left( \int_{t_1}^{t_2} \int_{s_1}^{s_2} \left( f_{j,0}(s,t) - f_{j,0}^{(n)}(s,t) \right) M^{(j)} (ds,t) dt \right)^2 \right] \nonumber \\
&\le \left( t_2-t_1 \right) \int_{T}^{t_2} \mathbb{E}}\def\E{\mathbb{E} \left[ \left( \int_{S}^{s_2} \left( f_{j,0}(s,t) - f_{j,0}^{(n)}(s,t) \right) M^{(j)} (ds,t) \right)^2 \right] dt \nonumber \\
&= \left( t_2-t_1 \right) \int_{T}^{t_2} \int_{S}^{s_2} \mathbb{E}}\def\E{\mathbb{E} \left[ \left( f_{j,0}(s,t) - f_{j,0}^{(n)}(s,t) \right)^2 \right] \big\langle M^{(j)} (ds,t) \big\rangle dt \nonumber \\
&\rightarrow 0, \quad n \rightarrow \infty,\ \forall 1 \le j \le d.
\end{align}
Thus, we have the following convergence in $L^2(\Omega)$,
\begin{align*}
\sum_{j=1}^d \int_A f_{j,0}^{(n)}(s,t) M^{(j)} (ds,t) dt
\rightarrow \sum_{j=1}^d \int_A f_{j,0}(s,t) M^{(j)} (ds,t) dt.
\end{align*}
The proof is concluded.
\end{proof}
Similarly, for predictable processes of the form
\begin{align}\label{e:G}
F_j(s,t) = F_j(0,t) + \sum_{i=1}^d \int_0^s f_{j,i}(r,t) M^{(i)}(dr,t) + \int_0^s f_{j,0} (r,t) dr, \ \ 1 \le j \le d,
\end{align}
where $f_{j,i}, \,1\le j\le d, \, 0 \le i \le d$ are $\mathcal{F}$-predictable processes,
we have the following Green's formula.
\begin{theorem} \label{Coro-Green formula} Fix $s_0, t_0 > 0$. Suppose that $\{F_j(s,t)\}_{1\le j\le d}$ are predictable processes
given by \eqref{e:G}. Assume
\begin{align*}
\mathbb{E}}\def\E{\mathbb{E} \left[ \int_0^{s_0} \int_0^{t_0} f_{j,i}(s,t)^2 d_s\big\langle M^{(i)} \big\rangle_{st_0}d_t \big\langle M^{(j)}\big\rangle_{st_0} \right]
< \infty,\quad \forall 1 \le i, j \le d,
\end{align*}
and
\begin{align*}
\mathbb{E}}\def\E{\mathbb{E} \left[ \int_0^{s_0} \int_0^{t_0} f_{j,0}(s,t)^2 d_t\big\langle M^{(i)} \big\rangle_{s_0 t} ds \right] < \infty, \quad \forall 1 \le j \le d.
\end{align*}
Then for any rectangle $A \subseteq R_{s_0 t_0}$, we have
\begin{align*}
\sum_{j=1}^d \int_{\partial A} F_j(s,t) M^{(j)}(s,dt) &= \sum_{j=1}^d \int_A F_j(s,t) dM^{(j)}(s,t) +
\sum_{i,j=1}^d \int_A f_{j,i}(s,t) dJ_{M^{(j)},M^{(i)}}(s,t)\nonumber \\
&\qquad \qquad + \sum_{j=1}^d \int_A f_{j,0}(s,t) M^{(j)} (s,dt) ds.
\end{align*}
\end{theorem}
\begin{proof}
Noting that by the second equality of \eqref{def-J'}, we have that for the left-hand side of \eqref{eq-2.19},
\begin{align*}
\int_A f_{j,i}(s,t) dJ_{M^{(j)},M^{(i)}}(s,t
&= \int_{\partial A} f_{j,i} M^{(i)}(s,t) M^{(j)}(s,dt) - \int_A f_{j,i} M^{(i)}(s,t) dM^{(j)}(s,t).
\end{align*}
The rest of the proof is the same as that of Theorem \ref{Thm-Green formula} and thus is omitted.
\end{proof}
\subsection{Quadratic covariations of $J_{MN}$ and $J_{M'N'}$ }
Let $M,N, M', N'$ be continuous martingales belonging to $\mathfrak M_s^4(z_0).$ In this subsection, for the completion of the theory, we shall derive the quadratic covariation for $J_{MN} = \psi \cdot MN$ and $J_{M'N'} = \psi \cdot M'N'$
with $\psi (z_1, z_2) = \mathbf 1_{[z_1 {\curlywedge\atop \curlywedge} z_2]}$ which are defined in Section \ref{sec:J-MN}.
More specifically, we aim to show
\begin{equation}\label{e:quad-var-J'}
d\langle J_{MN}, J_{M',N'} \rangle_{st} = d_t\langle M, M'\rangle_{st} d_s\langle N, N' \rangle_{st}.
\end{equation}
Recall that $J_{MN}$ can be approximated by $J^n_{MN}$ as in \eqref{e:J^n}, and that one can approximate the
function $\psi (z_1, z_2) = \mathbf 1_{[z_1 {\curlywedge\atop \curlywedge} z_2]}$ by
\begin{align*}
\psi = \lim_{n\to \infty} \sum_{i,j, k,l\in \mathbf I_n} \psi_{ijkl},
\end{align*}
where $\psi_{ijkl}(z_1, z_2) = \mathbf 1_{\rectangle_{i,j}}(z_1) \mathbf 1_{\rectangle_{k,l}}(z_2)$ and $\mathbf I_n$ is a subset of
$\{(i,j, k, l), i, j, k,l \in 1, \dots, 2^n\}$ which consists of $(i,j,k,l)$ satisfying $0\le i<k\le 2^n-1$ and $0\le l<j\le 2^n-1$. Denote by $\mathbf J_n$ the subset of $\mathbf I_n \times \mathbf{I}_n$ such that for $((i,j,k,l),(i',j',k',l')) \in \mathbf J_n$, the four rectangles $A=\rectangle_{i,j}$, $B=\rectangle_{k,l}$, $A'=\rectangle_{i',j'}$, $B'=\rectangle_{k',l'}$ are of the same position as in Case 3 in the proof of \eqref{e:increment} in Section \ref{sec:J-MN}. That is, $A$ and $A'$ are at the same horizontal level and are at the upper left of $B$ and $B'$, while $B$ and $B'$ are at the same vertical level. Now the quadratic covariation can be computed as follows,
\begin{align} \label{eq-quadratic JMN and JM'N'}
\left\langle J_{MN}, J_{M'N'} \right\rangle_{z_0}
&= \lim_{n\to \infty} \sum_{(i,j,k,l) \in \mathbf I_n} \sum_{(i',j',k',l') \in \mathbf I_n} \left\langle \psi_{ijkl} \cdot MN, \psi_{i'j'k'l'} \cdot M'N' \right\rangle_{z_0} \nonumber \\
&= \lim_{n\to \infty} \Bigg( \sum_{((i,j,k,l), (i',j',k',l')) \in \mathbf J_n} \left\langle \psi_{ijkl} \cdot MN, \psi_{i'j'k'l'} \cdot M'N' \right\rangle_{z_0} \notag\\
&\qquad \qquad\qquad + \sum_{((i,j,k,l), (i',j',k',l')) \notin \mathbf J_n} \left\langle \psi_{ijkl} \cdot MN, \psi_{i'j'k'l'} \cdot M'N' \right\rangle_{z_0} \Bigg).
\end{align}
For the first term on the right-hand side of \eqref{eq-quadratic JMN and JM'N'}, observing that the indices
$((i,j,k',l'), (i',j',k,l)),$ $ ((i',j',k,l), (i,j,k',l'))$, and $((i',j',k',l'), (i,j,k,l))$ all belong to $\mathbf J_n$ as long as $((i,j,k,l), (i',j',k',l'))$
$ \in\mathbf J_n$. Thus, we have
\begin{align} \label{eq-quadratic symmetric}
&\sum_{((i,j,k,l), (i',j',k',l')) \in \mathbf J_n} \left\langle \psi_{ijkl} \cdot MN, \psi_{i'j'k'l'} \cdot M'N' \right\rangle_{z_0}\nonumber \\
&= \dfrac{1}{4} \sum_{((i,j,k,l), (i',j',k',l')) \in \mathbf J_n} \Big( \left\langle \psi_{ijkl} \cdot MN, \psi_{i'j'k'l'} \cdot M'N' \right\rangle_{z_0}
+ \left\langle \psi_{ijk'l'} \cdot MN, \psi_{i'j'kl} \cdot M'N' \right\rangle_{z_0} \nonumber \\
&\qquad \qquad \qquad \qquad+ \left\langle \psi_{i'j'kl} \cdot MN, \psi_{ijk'l'} \cdot M'N' \right\rangle_{z_0}
+ \left\langle \psi_{i'j'k'l'} \cdot MN, \psi_{ijkl} \cdot M'N' \right\rangle_{z_0} \Big) \nonumber \\
&= \sum_{((i,j,k,l), (i',j',k',l')) \in\mathbf J_n} \langle M,M'\rangle(\rectangle_{i,j} \cap \rectangle_{i',j'})
\langle N,N'\rangle (\rectangle_{k,l} \cap \rectangle_{k',l'})\nonumber\\
&= \sum_{(i,j,k,l)\in \mathbf I_n} \langle M,M'\rangle (\rectangle_{i,j}) \langle N,N'\rangle (\rectangle_{k,l}),
\end{align}
where the second equality follows from \eqref{e:increment}.
For the second term in \eqref{eq-quadratic JMN and JM'N'}, noting that when $((i,j,k,l), (i',j',k',l')) \notin \mathbf J_n$, the four
rectangles $A=\rectangle_{i,j}$, $B=\rectangle_{k,l}$, $A'=\rectangle_{i',j'}$ and $B'=\rectangle_{k',l'}$ are of the same position
as in Case 1 or Case 2 in the proof of \eqref{e:increment} in Section \ref{sec:J-MN}. Thus, we have
\begin{align} \label{eq-quadratic nonsymmetric}
\sum_{((i,j,k,l), (i',j',k',l')) \notin \mathbf J_n} \left\langle \psi_{ijkl} \cdot MN, \psi_{i'j'k'l'} \cdot M'N' \right\rangle_{z_0}= 0.
\end{align}
Therefore, substituting \eqref{eq-quadratic symmetric} and \eqref{eq-quadratic nonsymmetric} into \eqref{eq-quadratic JMN and JM'N'},
one has
\begin{align*}
\left\langle J_{MN}, J_{M'N'} \right\rangle_{z_0}
&= \lim_{n\to \infty} \sum_{(i,j,k,l)\in \mathbf I_n}\langle M,M'\rangle (\rectangle_{ij}) \langle N,N'\rangle(\rectangle_{kl}) \\
&= \int_{R_{z_0}}\int_{R_{z_0}} \mathbf 1_{[z_1{\curlywedge \atop \curlywedge }z_2]} d\langle M,M'\rangle_{z_1}d\langle N,N'\rangle_{z_2},
\end{align*}
and this implies \eqref{e:quad-var-J'}.
\begin{remark} One can easily check that the computation is still valid if the function $\psi(z_1, z_2)$ is the limit of
$\psi^{(n)}(z_1,z_2)$ in $\mathcal L_{MN}^2(z_0) $ and in $\mathcal L_{M'N'}^2(z_0)$ satisfying
\begin{align} \label{eq-condition on psi}
\psi^{(n)}(z_1,z_2) \psi^{(n)}(z_1',z_2') = \psi^{(n)}(z_1,z_2') \psi^{(n)}(z_1',z_2),
\end{align}
for all $z_1=(s_1, t_1), z_2=(s_2, t_2), z_1'=(s_1',t_1'), z_2'=(s_2', t_2')$ satisfying $\max \{s_1, s_2\} \le s_1' = s_2'$ and
$\max \{t_1', t_2'\} \le t_1 = t_2$. Clearly, $\psi^{(n)}(z_1,z_2) = h_1(z_1)h_2(z_2)$ satisfies \eqref{eq-condition on psi}.
Moreover, by fixing $(z_1', z_2')$, one can check that all the functions satisfying \eqref{eq-condition on psi} must have the
form $\psi^{(n)}(z_1,z_2) = h_1(z_1)h_2(z_2)$. In this situation, we have
\begin{equation}\label{e:qua-var2}
\langle \psi\cdot MN, \psi\cdot M'N' \rangle_{z}
= \iint_{R_{z} \times R_{z}} |\psi(z_1,z_2)|^2 d\langle M, M'\rangle_{z_1} d\langle N,N'\rangle_{z_2} .
\end{equation}
\end{remark}
\section{SPDEs for the eigenvalue processes}\label{sec:SPDE}
In this section, we will derive a system of SPDEs satisfied by the eigenvalue processes of the Brownian sheet
matrix $\boldsymbol X$ defined in \eqref{e:X}. We assume that the family $b(s,t)$ of independent Brownian sheets have deterministic initial values such that the eigenvalues of the symmetric matrix $\boldsymbol X(0, 0)$ are distinct.
Recall that the standard $1$-dimensional Brownian sheet $\{B(s,t), (s,t)\in\mathbb{R}_+^2\}$ is a centered Gaussian random
field with covariance function
\begin{align*}
\mathbb{E}}\def\E{\mathbb{E} \left[ B(s_1,t_1) B(s_2,t_2) \right] = (s_1 \wedge s_2) (t_1 \wedge t_2).
\end{align*}
It follows directly from L\'evy's characterization of Brownian motion that for any fixed $t_1,\ t_2 > 0$,
\begin{align*}
\dfrac{1}{\sqrt{t_1}} B(t_1, \cdot), \
\dfrac{1}{\sqrt{t_2}} B(\cdot, t_2)
\end{align*}
are standard $1$-dimensional Brownian motions.
Consider the Brownian sheet matrix defined in \eqref{e:X}. As in Appendix \ref{sec:appendix}, for $1\le i \le d$,
let $\lambda_i (s,t) := \Phi_i(b (s,t)) =\tilde \Phi_i(\boldsymbol X(s,t))$ be the $i$-th biggest eigenvalue of $\boldsymbol X(s,t)$, where the
function $\tilde \Phi_i: \mathbf S_d\to \mathbb{R}$ maps a $d\times d$ symmetric matrix $A\in\mathbf S_d$ to its
$i$-th biggest eigenvalue $\tilde \Phi_i(A)$.
Let $S, T > 0$ be constants. By applying It\^{o}'s formula to $\lambda_i(S, \cdot)$, we have
\begin{align} \label{eq-3.1}
&\lambda_i (S,T) - \lambda_i(0,0)= \lambda_i (S,T) - \lambda_i(S,0)
= \Phi_i(b (S, T)) - \Phi_i(b (S, 0)) \nonumber \\
&= \sum_{k \le h} \int_0^{T} \dfrac{\partial \Phi_i}{\partial b_{kh}}(b(S,t)) b_{kh}(S,dt)
+ \dfrac{1}{2} \sum_{k \le h} \int_0^{T} \dfrac{\partial^2 \Phi_i}{\partial b_{kh}^2}(b(S,t)) \langle b_{kh}(S,dt) \rangle \nonumber \\
&= \sum_{k \le h} \int_0^{T} \dfrac{\partial \Phi_i}{\partial b_{kh}}(b(S,t)) b_{kh}(S,dt)
+ \dfrac{S}{2} \sum_{k \le h} \int_0^{T} \dfrac{\partial^2 \Phi_i}{\partial b_{kh}^2}(b(S,t)) dt.
\end{align}
By \eqref{eq-3.1} and \eqref{eq-2.2}, we have
\begin{align} \label{eq-3.2}
\lambda_i (S,T) - \lambda_i(0,0)
= \sum_{k \le h} \int_0^{T} \dfrac{\partial \Phi_i}{\partial b_{kh}}(b(S,t)) b_{kh}(S,dt) +
\sum_{j:j \neq i}\int_0^{T} \dfrac{S}{\lambda_i (S,t) - \lambda_j (S,t)} dt.
\end{align}
We shall express the right-hand side of \eqref{eq-3.2} as a sum of double integrals on $[0,S]\times[0,T]$.
We first deal with the second term.
For $i \neq j$, as in Appendix \ref{sec:appendix} we denote for any $x \in \mathbb{R}^{d(d+1)/2}$,
\begin{equation}\label{e:psi}
\Psi_{ij}(x)=\frac{1}{\Phi_i(x)-\Phi_j(x)}.
\end{equation}
By It\^{o}'s formula, we have
\begin{align} \label{eq-3.4}
& \dfrac{S}{\lambda_i (S,t) - \lambda_j (S,t)} = S \Psi_{ij}(b(S,t))\nonumber \\
&= \int_0^{S} \dfrac{1}{\lambda_i(s,t) - \lambda_j(s,t)} ds
+ \sum_{k \le h} \int_0^{S} s \dfrac{\partial \Psi_{ij}}{\partial b_{kh}}(b(s,t)) b_{kh}(ds,t) \nonumber \\
&\qquad\quad + \dfrac{1}{2} \sum_{k \le h} \int_0^{S} s \dfrac{\partial^2 \Psi_{ij}}{\partial b_{kh}^2}(b(s,t))
\langle b_{kh}(ds,t) \rangle\nonumber \\
&= \int_0^{S} \dfrac{1}{\lambda_i(s,t) - \lambda_j(s,t)} ds
+ \sum_{k \le h} \int_0^{S} s \dfrac{\partial \Psi_{ij}}{\partial b_{kh}}(b(s,t)) b_{kh}(ds,t) \nonumber \\
&\qquad \quad + \dfrac{1}{2} \sum_{k \le h} \int_0^{S} st \dfrac{\partial^2\Psi_{ij}}{\partial b_{kh}^2}(b(s,t)) ds.
\end{align}
Substituting \eqref{eq-2.11} into \eqref{eq-3.4}, we have
\begin{align} \label{eq-3.5}
&\dfrac{S}{\lambda_i (S,t) - \lambda_j (S,t)} \nonumber \\
&= \int_0^{S} \dfrac{1}{\lambda_i(s,t) - \lambda_j(s,t)} ds + \sum_{k \le h} \int_0^{S} s
\dfrac{\partial \Psi_{ij}}{\partial b_{kh}}(b(s,t)) b_{kh}(ds,t) \nonumber \\
&\qquad + \int_0^{S} \dfrac{2st}{(\lambda_i (s,t) - \lambda_j (s,t))^3} ds \nonumber \\
&\qquad + \int_0^{S} \dfrac{1}{(\lambda_i (s,t) - \lambda_j (s,t))} \sum_{l:l \neq i,j} \dfrac{st}{\left( \lambda_i (s,t) - \lambda_l (s,t) \right)
\left( \lambda_j (s,t) - \lambda_l (s,t) \right)} ds.
\end{align}
Lastly, we substitute \eqref{eq-3.5} to \eqref{eq-3.2} to obtain
\begin{align} \label{eq-3.6}
& \lambda_i (S,T) - \lambda_i(0,0) \nonumber \\
&= \sum_{k \le h} \int_0^{T} \dfrac{\partial \Phi_i}{\partial b_{kh}}(b(S,t)) b_{kh}(S,dt)
+ \sum_{j:j \neq i}\int_0^{T} \int_0^{S} \dfrac{1}{\lambda_i(s,t) - \lambda_j(s,t)} ds dt \nonumber \\
&\qquad + \sum_{j:j \neq i} \sum_{k \le h} \int_0^{T} \int_0^{S} s \dfrac{\partial \Psi_{ij}}{\partial b_{kh}}(b(s,t)) b_{kh}(ds,t) dt \nonumber \\
&\qquad + \sum_{j:j \neq i} \int_0^{T} \int_0^{S} \dfrac{2st}{(\lambda_i (s,t) - \lambda_j (s,t))^3} ds dt \nonumber \\
&\qquad + \sum_{j\neq l: j \neq i, l\neq i} \int_0^{T} \int_0^{S} \dfrac{1}{(\lambda_i (s,t) - \lambda_j (s,t))}
\dfrac{st}{\left( \lambda_i (s,t) - \lambda_l (s,t) \right) \left( \lambda_j (s,t) - \lambda_l (s,t) \right)} dsdt.
\end{align}
The last term on the right-hand side of \eqref{eq-3.6} vanishes, noting that it sums over all $j\neq l$ for $j,l \neq i$ and that
$ \frac{1}{(\lambda_i - \lambda_j)} \frac{st}{\left( \lambda_i - \lambda_l \right) \left( \lambda_j - \lambda_l \right)}$ changes
its sign by interchanging the indices $j$ and $l$. Therefore, we have
\begin{align} \label{eq:pde}
& \lambda_i (S,T) - \lambda_i(0,0) \nonumber \\
&= \sum_{k \le h} \int_0^{T} \dfrac{\partial \Phi_i}{\partial b_{kh}}(b(S,t)) b_{kh}(S,dt)
+ \sum_{j:j \neq i} \int_0^{T} \int_0^{S} \dfrac{1}{\lambda_i(s,t) - \lambda_j(s,t)} ds dt \nonumber \\
&\qquad + \sum_{j:j \neq i} \sum_{k \le h}\int_0^{T} \int_0^{S} s \dfrac{\partial \Psi_{ij}}{\partial b_{kh}}(b(s,t)) b_{kh}(ds,t) dt \nonumber \\
&\qquad + \sum_{j:j \neq i} \int_0^{T} \int_0^{S} \dfrac{2st}{(\lambda_i (s,t) - \lambda_j (s,t))^3} dsdt.
\end{align}
Now, we apply the multi-dimensional Green's formula (Theorem \ref{Coro-Green formula}) to the first term on the right-hand
side of \eqref{eq:pde}. By \cite[Theorem 2.1]{Jaramillo2018} (see also \cite[Theorem 1.1]{SXY20}), it has positive probability
for the eigenvalues $\{\lambda_i(s,t), 1\le i\le d\}$ of the Brownian sheet matrix $\boldsymbol X$ to collide. To avoid the singularity at the
collisions, we shall restrict $(s,t)$ in a region where all eigenvalues keep a distance from each other.
Define the region $D_\epsilon$ for $\epsilon>0$ by
\begin{align*}
D_{\epsilon} = \left\{ (x_1, \ldots, x_d) \in \mathbb{R}^d: x_i - x_{i+1} > \epsilon, 1 \le i \le d-1 \right\}.
\end{align*}
Let $\chi_{\epsilon} \in C_b^{\infty}(\mathbb{R}^d)$ such that $\chi_{\epsilon}(x) = 1$ for $x \in D_{\epsilon}$ and
$\chi_{\epsilon}(x) = 0$ for $x \in \mathbb{R}^d \setminus D_{\frac\epsilon2}$. For simplicity, we denote
$\Phi = (\Phi_1, \ldots, \Phi_d)$. By It\^{o}'s formula, we have
\begin{align} \label{eq-3.69}
&\left( \dfrac{\partial \Phi_i}{\partial b_{kh}} \chi_{\epsilon} (\Phi) \right) (b(s,t)) \nonumber \\
&= \left( \dfrac{\partial \Phi_i}{\partial b_{kh}} \chi_{\epsilon} (\Phi) \right) (b(0,t)) \nonumber \\
&\qquad + \sum_{k' \le h'} \int_0^{s} \left( \dfrac{\partial^2 \Phi_i}{\partial b_{kh} \partial b_{k'h'}} \chi_{\epsilon} (\Phi)
+ \dfrac{\partial \Phi_i}{\partial b_{kh}} \sum_{l=1}^d \dfrac{\partial \chi_{\epsilon}}{\partial x_l} (\Phi)
\dfrac{\partial \Phi_l}{\partial b_{k'h'}} \right)(b(r,t)) b_{k'h'}(dr,t) \nonumber \\
&\qquad + \dfrac{t}{2} \sum_{k' \le h'} \int_0^s \left( \dfrac{\partial^3 \Phi_i}{\partial b_{kh} \partial b_{k'h'}^2} \chi_{\epsilon} (\Phi)
+ 2 \dfrac{\partial^2 \Phi_i}{\partial b_{kh} \partial b_{k'h'}} \sum_{l=1}^d \dfrac{\partial \chi_{\epsilon}}{\partial x_l} (\Phi)
\dfrac{\partial \Phi_l}{\partial b_{k'h'}} \right. \nonumber \\
&\qquad \qquad \left. + \dfrac{\partial \Phi_i}{\partial b_{kh}} \sum_{l,l'=1}^d \dfrac{\partial^2 \chi_{\epsilon}}{\partial x_l \partial x_{l'}} (\Phi)
\dfrac{\partial \Phi_l}{\partial b_{k'h'}} \dfrac{\partial \Phi_{l'}}{\partial b_{k'h'}} + \dfrac{\partial \Phi_i}{\partial b_{kh}}
\sum_{l=1}^d \dfrac{\partial \chi_{\epsilon}}{\partial x_l} (\Phi) \dfrac{\partial^2 \Phi_l}{\partial b_{k'h'}^2} \right) (b(r,t)) dr.
\end{align}
Note that the function $\chi_{\epsilon}$ and all its partial derivatives vanish when $x \in \mathbb{R}^d \setminus D_{\frac\epsilon2}$,
by Lemma \ref{Lemma-eigen derivative}, all the integrand functions in \eqref{eq-3.69} are bounded. Hence, we can apply
Theorem \ref{Coro-Green formula} to obtain
\begin{align}
& \sum_{k \le h} \int_0^{T} \left( \dfrac{\partial \Phi_i}{\partial b_{kh}} \chi_{\epsilon} (\Phi) \right) (b(S,t)) b_{kh}(S,dt) \nonumber \\
&= \sum_{k \le h} \int_{\partial R_{ST}} \left( \dfrac{\partial \Phi_i}{\partial b_{kh}} \chi_{\epsilon} (\Phi) \right)(b(s,t)) b_{kh}(s,dt) \nonumber \\
&= \sum_{k \le h} \iint_{R_{ST}} \left( \dfrac{\partial \Phi_i}{\partial b_{kh}} \chi_{\epsilon} (\Phi) \right)(b(s,t)) db_{kh}(s,t) \nonumber \\
&\qquad + \sum_{k \le h} \sum_{k' \le h'} \iint_{R_{ST}} \left( \dfrac{\partial^2 \Phi_i}{\partial b_{kh} \partial b_{k'h'}} \chi_{\epsilon} (\Phi)
+ \dfrac{\partial \Phi_i}{\partial b_{kh}} \sum_{l=1}^d \dfrac{\partial \chi_{\epsilon}}{\partial x_l} (\Phi) \dfrac{\partial \Phi_l}
{\partial b_{k'h'}} \right) (b(s,t)) dJ_{ b_{kh} b_{k'h'}}(s,t) \nonumber \\
&\qquad + \sum_{k \le h} \iint_{R_{ST}} \dfrac{t}{2} \sum_{k' \le h'} \Bigg( \dfrac{\partial^3 \Phi_i}{\partial b_{kh} \partial b_{k'h'}^2} \chi_{\epsilon} (\Phi) + 2 \dfrac{\partial^2 \Phi_i}{\partial b_{kh} \partial b_{k'h'}} \sum_{l=1}^d \dfrac{\partial \chi_{\epsilon}}{\partial x_l}
(\Phi) \dfrac{\partial \Phi_l}{\partial b_{k'h'}} \nonumber \\
&\qquad + \dfrac{\partial \Phi_i}{\partial b_{kh}} \sum_{l,l'=1}^d \dfrac{\partial^2 \chi_{\epsilon}}{\partial x_l \partial x_{l'}} (\Phi)
\dfrac{\partial \Phi_l}{\partial b_{k'h'}} \dfrac{\partial \Phi_{l'}}{\partial b_{k'h'}} + \dfrac{\partial \Phi_i}{\partial b_{kh}}
\sum_{l=1}^d \dfrac{\partial \chi_{\epsilon}}{\partial x_l} (\Phi) \dfrac{\partial^2 \Phi_l}{\partial b_{k'h'}^2} \Bigg) (b(s,t)) b_{kh}(s,dt) ds. \label{e:3.9}
\end{align}
Denote
\begin{align*}
T_{\epsilon} = \left\{ (s,t): (\Phi_1(b (s, t)), \ldots, \Phi_d(b (s,t))) \notin D_{\epsilon} \right\}.
\end{align*}
We shall construct a sequence of adapted random time pairs $\{(\sigma_{\frac1n}, \tau_{\frac1n})\}_{n\ge1}$ such that
$(\sigma_{\frac1n}, \tau_{\frac1n})\prec (\sigma_{\frac1{n+1}}, \tau_{\frac1{n+1}})$. First, we choose a pair of random
times $(\sigma_1, \tau_1)$ as follows. For each fixed $\omega\in \Omega$, if $T_{1}(\omega) = \emptyset$, then we
choose $\sigma_1(\omega) = \tau_1(\omega) = \infty$; if $T_1(\omega) \not= \emptyset$, then by Zorn's lemma, there
exists a minimal element $(s_1, t_1)$ in $T_{1}(\omega)$, and we set $(\sigma_1(\omega), \tau_1(\omega)) = (s_1, t_1).$
By the meaning of minimal element, we have $[(s,t)\prec\prec (\sigma_1, \tau_1)] = [(\Phi_1(b (s, t)), \ldots, \Phi_d(b (s,t)))
\in D_{1} ] \in \mathcal{F}_{st}$. Next, for an arbitrary fixed $\omega\in \Omega$, let $(\sigma_{\frac12}, \tau_{\frac12})$ be a minimal
element of the set
\[\left\{ (\sigma_1(\omega),\tau_1(\omega))\prec (s,t): (\Phi_1(b (s, t)), \ldots, \Phi_d(b (s,t))) \notin D_{\frac12} \right\},\]
and $ (\sigma_{\frac12}, \tau_{\frac12}) = (\infty, \infty)$ if the set is empty.
Clearly $(\sigma_{1}, \tau_1)\prec (\sigma_{\frac12},\tau_{\frac12})$,
\[[(\sigma_1, \tau_1)\prec (s,t)\prec\prec(\sigma_{\frac12}, \tau_{\frac12})]=[(\Phi_1(b (s, t)), \ldots, \Phi_d(b (s,t)))
\in D_{\frac12}\backslash D_{1} ] \in \mathcal{F}_{st},\]
and hence $[ (s,t)\prec\prec(\sigma_{\frac12}, \tau_{\frac12})]\in \mathcal{F}_{st}$. The rest of random time pairs
$(\sigma_{\frac1n},\tau_{\frac1n})$ can be constructed in the same way. Define
\begin{equation}\label{e:sigma-tau}
(\sigma,\tau)=\sup_{n\ge 1} (\sigma_{\frac1n}, \tau_{\frac1n}).
\end{equation}
Thus, $[ (s,t)\prec\prec(\sigma, \tau)]=\cup_{n\ge1}[ (s,t)\prec\prec(\sigma_{\frac1n},\tau_{\frac1n})]\in \mathcal{F}_{st}$.
For each $n\ge 1$, on the set $[\omega\in\Omega: (S,T)\prec\prec (\sigma_{\frac1n}(\omega), \tau_{\frac1n}(\omega))]$,
we have that for $(s,t)\prec (S,T)$, $\Phi(b(s,t)) = (\Phi_1(b(s,t), \ldots, \Phi_d(b(s,t))$ belongs to $D_{\frac1n}$ and all
the partial derivatives of the function $\chi_{\frac1n}$ vanish.
Thus, by \eqref{e:3.9}, we have for $(S,T)\prec\prec (\sigma, \tau)$,
\begin{align} \label{eq-3.11}
&\sum_{k \le h} \int_0^{T} \dfrac{\partial \Phi_i}{\partial b_{kh}}(b(S,t)) b_{kh}(S,dt)
= \sum_{k \le h} \int_{\partial R_{ST}} \dfrac{\partial \Phi_i}{\partial b_{kh}}(b(s,t)) b_{kh}(s,dt) \nonumber \\
&= \sum_{k \le h} \iint_{R_{ST}} \dfrac{\partial \Phi_i}{\partial b_{kh}}(b(s,t)) db_{kh}(s,t)+ \sum_{k \le h}
\sum_{k' \le h'} \iint_{R_{ST}} \dfrac{\partial^2 \Phi_i}{\partial b_{kh} \partial b_{k'h'}}(b(s,t)) dJ_{ b_{kh} b_{k'h'}}(s,t) \nonumber \\
&\quad + \sum_{k \le h} \iint_{R_{ST}} \dfrac{t}{2} \sum_{k' \le h'} \dfrac{\partial^3 \Phi_i}{\partial b_{kh}
\partial b_{k'h'}^2}(b(s,t)) b_{kh}(s,dt) ds.
\end{align}
Therefore, substitute \eqref{eq-3.11} to \eqref{eq:pde}, we have for $(S, T)\prec\prec (\sigma, \tau)$ and $1\le i\le d$,
\begin{align} \label{eq-spde}
&\lambda_i (S,T) - \lambda_i(0,0)\notag\\
& = \sum_{k \le h} \iint_{R_{ST}} \dfrac{\partial \Phi_i}{\partial b_{kh}}(b(s,t)) db_{kh}(s,t)+ \sum_{k \le h} \sum_{k' \le h'}
\iint_{R_{ST}} \dfrac{\partial^2 \Phi_i}{\partial b_{kh} \partial b_{k'h'}}(b(s,t)) dJ_{ b_{kh} b_{k'h'}}(s,t) \nonumber \\
&\qquad+ \sum_{k \le h} \sum_{k' \le h'}\iint_{R_{ST}} \dfrac{t}{2} \dfrac{\partial^3 \Phi_i}{\partial b_{kh}
\partial b_{k'h'}^2}(b(s,t)) b_{kh}(s,dt) ds + \sum_{j:j \neq i} \int_0^{T} \int_0^{S} \dfrac{1}{\lambda_i(s,t) - \lambda_j(s,t)} ds dt \nonumber \\
&\qquad + \sum_{j:j \neq i} \sum_{k \le h}\int_0^{T} \int_0^{S} \dfrac{\partial \Psi_{ij}}{\partial b_{kh}}(b(s,t)) s b_{kh}(ds,t) dt
+ \sum_{j:j \neq i} \int_0^{T}\int_0^{S} \dfrac{2st}{(\lambda_i (s,t) - \lambda_j (s,t))^3} ds dt.
\end{align}
Noting that by \eqref{e:psi},
\begin{align*}
\dfrac{\partial \Psi_{ij}}{\partial b_{kh}} (b(s,t))
&= \dfrac{-1}{(\lambda_i(s,t) - \lambda_j(s,t))^2} \left( \dfrac{\partial \Phi_i}{\partial b_{kh}} - \dfrac{\partial \Phi_j}{\partial b_{kh}} \right),
\end{align*}
we have, by \eqref{eq-2.4'},
\begin{align*}
\sum_{j:j\neq i} \dfrac{\partial \Psi_{ij}}{\partial b_{kh}} (b(s,t))
= \dfrac{1}{2} \sum_{k' \le h'} \dfrac{\partial^3 \Phi_i}{\partial b_{kh} \partial b_{k'h'}^2}.
\end{align*}
Therefore, \eqref{eq-spde} can be written in a symmetric form: for $(S, T)\prec\prec (\sigma, \tau)$ and $1\le i\le d$,
\begin{align} \label{eq-spde'}
&\lambda_i (S,T) - \lambda_i(0,0)\notag\\
=& \sum_{k \le h} \iint_{R_{ST}} \dfrac{\partial \Phi_i}{\partial b_{kh}}(b(s,t)) db_{kh}(s,t) + \sum_{k \le h} \sum_{k' \le h'}
\iint_{R_{ST}} \dfrac{\partial^2 \Phi_i}{\partial b_{kh} \partial b_{k'h'}}(b(s,t)) dJ_{ b_{kh} b_{k'h'}}(s,t) \nonumber \\
&\quad+ \frac12 \sum_{k \le h} \sum_{k' \le h'} \iint_{R_{ST}} \dfrac{\partial^3 \Phi_i}{\partial b_{kh} \partial b_{k'h'}^2}(b(s,t))
\Big( tb_{kh}(s,dt) ds+ sb_{kh}(ds,t) dt \Big)\nonumber \\
&\quad + \sum_{j:j \neq i}\int_0^{T} \int_0^{S} \left(\dfrac{1}{\lambda_i(s,t) - \lambda_j(s,t)} + \dfrac{2st}
{(\lambda_i (s,t) - \lambda_j (s,t))^3} \right)ds dt.
\end{align}
Recalling that we have assumed the initial eigenvalues are distinct, by the continuity of eigenvalue functions, we have $(0,0)\prec\prec (\sigma, \tau)$ a.s.
Thus,
for almost all $\omega\in \Omega$, we have the following formal partial differential equations
near the initial point $(0,0)$:
for $1\le i\le d$,
\begin{align} \label{eq-spde''}
d\lambda_i (s,t) \notag
=& \sum_{k \le h} \dfrac{\partial \Phi_i}{\partial b_{kh}}(b(s,t)) db_{kh}(s,t) + \sum_{k \le h} \sum_{k' \le h'}\dfrac{\partial^2 \Phi_i}
{\partial b_{kh} \partial b_{k'h'}}(b(s,t)) dJ_{ b_{kh} b_{k'h'}}(s,t) \nonumber \\
&\quad+ \frac12 \sum_{k \le h} \sum_{k' \le h'} \dfrac{\partial^3 \Phi_i}{\partial b_{kh} \partial b_{k'h'}^2}(b(s,t))
\Big( tb_{kh}(s,dt) ds+ sb_{kh}(ds,t) dt \Big)\nonumber \\
&\quad + \sum_{j:j \neq i}\left( \dfrac{1}{\lambda_i(s,t) - \lambda_j(s,t)}+ \dfrac{2st}{(\lambda_i (s,t) - \lambda_j (s,t))^3}\right) ds dt.
\end{align}
\section{High-dimensional limit of the empirical spectral distributions}\label{sec:limit}
In this section, we study the high-dimensional limit of empirical spectral measure of the rescaled Brownian sheet matrices. In Section \ref{sec:tightness}, we first obtain the tightness of the empirical spectral measures (Theorem \ref{Thm-tightness}), and then show the convergence by Wigner's theorem (Theorem \ref{Thm-limit measure}). In Section \ref{sec:PDE}, we derive a PDE for the Stieltjes transform of the limiting measure and also a McKean-Vlasov equation for the limiting measure.
\subsection{Tightness and high-dimensional limit}
\label{sec:tightness}
For every integer $d \ge 1$, let $\boldsymbol X^d(s,t)$ be a $d\times d$ matrix given by \eqref{e:X}, and $\{\lambda_i^d(s,t):1 \le i \le d\}$
be the set of eigenvalues of $\boldsymbol X^d(s,t)$. Define the empirical spectral measure of $\boldsymbol X^d(s,t)/\sqrt d$
\begin{align}\label{e:em}
L_d(s,t)(dx)
= \dfrac{1}{d} \sum_{i=1}^d \delta_{\lambda_i^d(s,t)/\sqrt{d}} (dx).
\end{align}
For a measurable function $g:\mathbb{R}\to \mathbb{R}$, we write
\begin{align} \label{eq-emprical}
\langle g, L_d(s,t) \rangle
:=\int_{\mathbb{R}} g(x) L_d(s,t) (dx)
= \dfrac{1}{d} \sum_{i=1}^d g \left( \dfrac{\lambda_i^d(s,t)}{\sqrt{d}} \right).
\end{align}
Let $\mathcal{P}(\mathbb{R})$ be the set of probability measures
on $\mathbb{R}$ equipped with its weak topology and $C_0(\mathbb{R})$ be the set of continuous functions on $\mathbb{R}$ that vanish at infinity.
Throughout this subsection, let $S$ and $T$ be two fixed positive numbers, and recall the notation
$R_{ST}=[0,S]\times[0,T].$
The following tightness criterion for probability-measure-valued stochastic processes is a straightforward generalization of
\cite[Proposition B.3]{SYY20} (see also \cite[Section 3]{Rogers1993} where this criterion was applied implicitly).
\begin{lemma} \label{Lemma-tightness}
Let $\{\mu_d(s,t), (s,t) \in R_{ST}\}_{d \in \mathbb{N}}\subset C(R_{ST}, \mathcal{P}(\mathbb{R}))$ be a sequence of probability-measure-valued
random fields. Assume the following conditions are satisfied:
\begin{enumerate}
\item[(A)] there exists a non-negative function $\varphi(x)$ satisfying $\lim\limits_{|x| \to +\infty} \varphi(x) = +\infty$ and
\begin{align*}
\sup_{d \in \mathbb{N}} \mathbb{E}}\def\E{\mathbb{E} \left[ \left| \left\langle \varphi, \mu_d(s,t) \right\rangle \right|^{\gamma} \right] < \infty,
~ \forall (s,t) \in R_{ST},
\end{align*}
for some $\gamma > 0$;
\item[(B)] there exists a countable dense subset $\{f_i(x), x\in\mathbb{R}\}_{i \in \mathbb{N}}$ of $C_0(\mathbb{R})$, such that for some positive constants $a_1>1$ and $a_2>1$,
\begin{align*}
&\mathbb{E}}\def\E{\mathbb{E} \left[ \left| \left\langle f_i, \mu_d(s_2,t_2) \right\rangle - \left\langle f_i, \mu_d(s_1,t_1) \right\rangle \right|^{a_1} \right]
\le C_{f_i,S,T} |(s_2,t_2) - (s_1,t_1)|^{a_2}
\end{align*}
for all $(s_1,t_1), (s_2,t_2) \in R_{ST}, d \in \mathbb{N}$ and $i\in N$, where $C_{f_i,S,T}$ is a constant depending only on $S$, $T$ and $f_i$.
\end{enumerate}
Then the set $\{\mu_d(s,t), (s,t) \in R_{ST}\}_{d \in \mathbb{N}}$ of $C(R_{ST}, \mathcal{P}(\mathbb{R}))$-valued random elements is tight, i.e., it induces a tight family of probability measures on $C(R_{ST}, \mathcal{P}(\mathbb{R}))$.
\end{lemma}
The Kolmogorov continuity theorem for random fields (see e.g. \cite[Theorem 2.5.1 in Chapter 5]{K02}) implies that, on every compact interval, the
Brownian sheet is $\beta$-H\"{o}lder continuous for $\beta\in(0, \frac12)$. The following lemma is a direct consequence of Fernique's theorem (\cite{fernique}).
\begin{lemma} \label{Lemma-Holder of Bs}
For any $\beta\in(0,\frac12)$, there exists a positive constant $\delta=\delta(\beta, S, T)$ depending only on $(\beta, S, T)$ such that
\begin{align*}
\mathbb{E}}\def\E{\mathbb{E} \left[ \exp \left(\delta \left\| B \right\|_{\beta;R_{ST}}^2 \right) \right] < \infty,
\end{align*}
where
\begin{align} \label{eq-def-Holder of Bs}
\left\|B \right\|_{\beta;R_{ST}}
= \sup_{(s_1,t_1), (s_2,t_2) \in R_{ST}} \dfrac{\left| B(s_2,t_2) - B(s_1,t_1) \right|}{|(s_2,t_2)-(s_1,t_1)|^{\beta}}
\end{align}
is the $\beta$-H\"{o}lder norm of $B$ on the rectangle $R_{ST}$.
\end{lemma}
Now we are ready to derive the following result on the tightness of the sequence $\{L_d(s,t), (s,t)\in R_{ST}\}_{d\in\mathbb{N}}$.
\begin{theorem} \label{Thm-tightness}
Assume that there exists a nonnegative function $\varphi(x) \in C^1(\mathbb{R})$ with bounded derivative, such that
\begin{align} \label{eq-initial condition}
\lim\limits_{|x| \to +\infty} \varphi(x) = +\infty ~ \mbox{ and } ~ \sup_{d \in \mathbb{N}} \left\langle \varphi, L_d(0,0) \right\rangle < \infty.
\end{align}
Then the sequence $\{L_d(s,t), (s,t) \in R_{ST}\}_{d \in \mathbb{N}}$ is tight on $C(R_{ST}, \mathcal{P}(\mathbb{R}))$.
\end{theorem}
\begin{proof}
Let $f$ be an arbitrary continuously differentiable function with bounded derivative. By the mean value theorem and
the Hoffman-Wielandt inequality (see e.g. \cite[Lemma 2.1.19]{anderson2010}), we have for $(s_2,t_2), (s_1,t_1) \in R_{ST}$,
\begin{align} \label{eq-pathwise holder}
&\left| \left\langle f, L_d(s_2,t_2) \right\rangle - \left\langle f, L_d(s_1,t_1) \right\rangle \right|^2
=\left| \dfrac{1}{d} \sum_{i=1}^d \left( f \left( \dfrac{\lambda_i^d(s_2,t_2)}{\sqrt{d}} \right) - f \left( \dfrac{\lambda_i^d(s_1,t_1)}{\sqrt{d}} \right) \right) \right|^2 \nonumber \\
&\le \dfrac{1}{d} \sum_{i=1}^d \left| f \left( \dfrac{\lambda_i^d(s_2,t_2)}{\sqrt{d}} \right) - f \left( \dfrac{\lambda_i^d(s_1,t_1)}{\sqrt{d}} \right) \right|^2
\le \dfrac{\|f'\|_{L^{\infty}}^2}{d^2} \sum_{i=1}^d \left| \lambda_i^d(s_2,t_2) - \lambda_i^d(s_1,t_1) \right|^2 \nonumber \\
&\le \dfrac{\|f'\|_{L^{\infty}}^2}{d^2} \sum_{i,j=1}^d \left| X_{ij}^d(s_2,t_2) - X_{ij}^d(s_1,t_1) \right|^2
= \dfrac{2\|f'\|_{L^{\infty}}^2}{d^2} \sum_{i \le j}\left| b_{ij}(s_2,t_2) - b_{ij}(s_1,t_1) \right|^2.
\end{align}
Noting that $\{b_{ij}(s,t)\}_{1 \le i \le j \le d}$ are standard Brownian sheets, by \eqref{eq-pathwise holder} and the Minkowski inequality,
we have for some $\beta\in (0,\frac12)$,
\begin{align} \label{eq-holder in expectation}
& \mathbb{E}}\def\E{\mathbb{E} \left[ \left| \left\langle f, L_d(s_2,t_2) \right\rangle - \left\langle f, L_d(s_1,t_1) \right\rangle \right|^4 \right] \nonumber \\
\le& \dfrac{4\|f'\|_{L^{\infty}}^4}{d^4} \mathbb{E}}\def\E{\mathbb{E} \left[ \left( \sum_{i\le j} \left| b_{ij}(s_2,t_2) - b_{ij}(s_1,t_1) \right|^2 \right)^2 \right] \nonumber \\
\le& \dfrac{4\|f'\|_{L^{\infty}}^4}{d^4} \left( \sum_{i\le j} \left( \mathbb{E}}\def\E{\mathbb{E} \left[ \left| b_{ij}(s_2,t_2) - b_{ij}(s_1,t_1) \right|^4 \right] \right)^{1/2} \right)^2 \nonumber \\
=& \dfrac{4\|f'\|_{L^{\infty}}^4}{d^4} \left( \dfrac{d(d+1)}{2} \left( \mathbb{E}}\def\E{\mathbb{E} \left[ \left| b_{11}(s_2,t_2) - b_{11}(s_1,t_1) \right|^4 \right] \right)^{1/2} \right)^2 \nonumber \\
=& \dfrac{(d+1)^2 \|f'\|_{L^{\infty}}^4}{d^2} \mathbb{E}}\def\E{\mathbb{E} \left[ \left| b_{11}(s_2,t_2) - b_{11}(s_1,t_1) \right|^4 \right] \nonumber \\
\le& 4 \|f'\|_{L^{\infty}}^4\mathbb{E}}\def\E{\mathbb{E} \left[ \left\|b_{11} \right\|_{\beta;R_{ST}}^4 |(s_2,t_2)-(s_1,t_1)|^{4\beta} \right] \nonumber \\
=& C(\beta, f', S,T) |(s_2,t_2)-(s_1,t_1)|^{4\beta},
\end{align}
where $C(\beta, f', S,T)$ is a finite positive constant by Lemma \ref{Lemma-Holder of Bs}.
As a consequence, Condition (A) in Lemma \ref{Lemma-tightness} is satisfied with $\gamma=4$. Moreover, if we choose $\beta \in (\frac14,\frac12)$, then assumption \eqref{eq-initial condition} and \eqref{eq-holder in expectation} together yield Condition (B) in Lemma \ref{Lemma-tightness} with $a_1=4$, $a_2=4\beta$ and $\{f_i\}_{i \in \mathbb{N}}$ being a sequence of functions in $C^1(\mathbb{R})$ with bounded derivative that is dense in $C_0(\mathbb{R})$. Then the proof is concluded by Lemma \ref{Lemma-tightness}.
\end{proof}
\begin{remark}
In the above proof, the independence of the Brownian sheets $b_{ij}$ ($i \le j$) actually is not used.
\end{remark}
Denote by $\mu_{sc}(dx)$ the semicircle distribution, i.e. $\mu_{sc}(dx) = p_{sc}(x) dx$, where the density function is given by
\begin{align*}
p_{sc}(x) = \dfrac{\sqrt{4-x^2}}{2\pi} 1_{[-2,2]}(x).
\end{align*}
Let $\{\tilde \mu(s,t), (s,t) \in R_{ST}\}$ be an element in $C(R_{ST}, \mathcal{P}(\mathbb{R}))$ such that $\tilde \mu(s,t)$ is a probability measure with density function $\tilde p_{s,t}(x) = \frac{1}{\sqrt{st}} p_{sc}(x/\sqrt{st})$. That is, $\tilde \mu(s,t)$ is a rescaled semicircle distribution. Here, we use the convention that $\tilde \mu(s,t)(dx) = \delta_0(dx)$ if $st=0$.
\begin{theorem} \label{Thm-limit measure}
Assume the same condition as in Theorem \ref{Thm-tightness}. Also assume that $\{\boldsymbol X^d(0,0), d\in \mathbb{N}\}$ are symmetric deterministic matrices such that
\begin{align*}
D := \sup_{d \in \mathbb{N}} \left\| \frac{1}{\sqrt{d}} \boldsymbol X^d(0,0) \right\|
< \infty,
\end{align*}
where $\|\cdot\|$ is the operator norm (the operator norm of a symmetric matrix is its largest eigenvalue), and that $L_d(0,0)$ converges weakly to some probability measure $\mu(0,0)$ as $d$ goes to infinity.
Then, as $d\to \infty$, $\{L_d(s,t), (s,t) \in R_{ST}\}$ converges in probability to $\{\mu(s,t), (s,t) \in R_{ST}\}$ in $C(R_{ST}, \mathcal{P}(\mathbb{R}))$ which is given by
\begin{align}\label{e:limit-measure}
\mu(s,t) = \tilde{\mu}(s,t) \boxplus \mu(0,0),
\end{align}
where $\boxplus$ is the free additive convolution of two probability measures (\cite[Definition 5.3.20]{anderson2010}).
\end{theorem}
\begin{proof}
For any fixed $(s,t) \in R_{ST}$ with $st > 0$, we have
\begin{align*}
\dfrac{1}{\sqrt{d}} \boldsymbol X^d(s,t)
= \dfrac{1}{\sqrt{d}} \left( \boldsymbol X^d(s,t) - \boldsymbol X^d(0,0) \right) + \dfrac{1}{\sqrt{d}} \boldsymbol X^d(0,0).
\end{align*}
By the self-similarity property of the Brownian sheet, one can see that $\frac{1}{\sqrt{std}} \left( \boldsymbol X^d(s,t) - \boldsymbol X^d(0,0) \right)$ is a $d \times d$
Wigner matrix (see e.g. \cite[Section 2.1]{anderson2010} for the definition). By Wigner's semicircle law (see e.g. \cite[Theorem 2.1.1]{anderson2010}),
the empirical spectral measure of $\frac{1}{\sqrt{std}} \left( \boldsymbol X^d(s,t) - \boldsymbol X^d(0,0) \right)$ converges in probability to the semicircle distribution
$\mu_{sc}$ in $\mathcal P(\mathbb{R})$ as $d \to \infty$. Thus, the empirical spectral measure of $\frac{1}{\sqrt{d}} \left( \boldsymbol X^d(s,t) - \boldsymbol X^d(0,0) \right)$
converges in probability to the measure $\tilde \mu(s,t)$ in $\mathcal P(\mathbb{R})$ as $d \to \infty$. Note that the empirical spectral measure of the
matrix $\frac{1}{\sqrt{d}} \boldsymbol X^d(0,0)$ is $L_d(0,0)$, which converges to $\mu(0,0)$ in $\mathcal P(\mathbb{R})$. Therefore, by \cite[Theorem 5.4.5]{anderson2010},
for every $(s,t) \in R_{ST}$ with $st > 0$, the empirical spectral measure of the matrix $\frac{1}{\sqrt{d}} \boldsymbol X^d(s,t)$ converges in probability
to the measure $\mu(s,t)$ given by \eqref{e:limit-measure} in $\mathcal P(\mathbb{R})$ as $d$ goes to infinity. Moreover, when $st=0$, $\frac{1}{\sqrt{d}} \boldsymbol X^d(s,t) = \frac{1}{\sqrt{d}} \boldsymbol X^d(0,0)$, and the empirical spectral measures converge to $\mu(0,0)$ in $\mathcal P(\mathbb{R})$.
By Theorem \ref{Thm-tightness}, the sequence $\{L_d(s,t), (s,t) \in R_{ST}\}_{d \in \mathbb{N}}$ is tight. Let $\{\nu(s,t), (s,t) \in R_{ST}\}$
be the weak limit of an arbitrary convergent subsequence of $\{L_d(s,t), (s,t) \in R_{ST}\}_{d \in \mathbb{N}}$. Noting that for every fixed
$(s,t) \in R_{ST}$, $L_d(s,t)$ is the empirical spectral measure of the matrix $\frac{1}{\sqrt{d}} \boldsymbol X^d(s,t)$ and it converges
in probability to the deterministic measure $\mu(s,t)$, we can conclude that $\nu(s,t)=\mu(s,t)$ for $(s,t) \in R_{ST}$.
Therefore, the limit of any convergent subsequence of $\{L_d(s,t), (s,t) \in R_{ST}\}_{d \in \mathbb{N}}$ is the deterministic measure
$\{\mu(s,t), (s,t) \in R_{ST}\}$ given by \eqref{e:limit-measure}. The proof is concluded.
\end{proof}
\subsection{PDEs for the limit measure} \label{sec:PDE}
It is known (see e.g. \cite{anderson2010}) that the high-dimensional limit $\hat \mu_t(dx)$ of the empirical measures of Dyson Brownian
motion \eqref{Eq:Dyson} satisfies the following McKean-Vlasov equation,
\begin{align} \label{eq-Dyson limit equation}
\dfrac{\partial}{\partial t} \int_{\mathbb{R}} f(x) \hat \mu_t(dx)
= \dfrac{1}{2} \iint_{\mathbb{R}^2} \dfrac{f'(x) - f'(y)}{x-y} \hat \mu_t(dx) \hat \mu_t(dy), ~~\mbox{for } f \in C_b^2(\mathbb{R}).
\end{align}
The Stieltjes transform
\[\hat G_t(z)=\int_{\mathbb{R}} \frac1{z-x} \hat \mu_t(dx), \mbox{ for } z\in \mathbb C\backslash \mathbb{R}\]
of $\hat \mu_t(dx)$ solves the following complex version of inviscid Burgers' equation
\begin{align*}
\partial_t \hat G_t(z) + \hat G_t(z) \partial_z \hat G_t(z) = 0.
\end{align*}
In this subsection, we will derive parallel PDEs for the limit $\mu(s,t)$ (see Theorem \ref{Thm-limit measure}) of the empirical spectral
measures of the rescaled Brownian sheet matrices. We remark that the equations are obtained by the properties the semicircle
distribution and may have other equivalent forms.
Assume $\mu(0,0)(dx) = \delta_0(dx)$, then the limiting measure $\mu(s,t) (dx) = \tilde{\mu}(s,t) (dx)$, recalling that
\[\tilde{\mu}(s,t) (dx) = \tilde p_{s,t}(x)dx=\frac{1}{\sqrt{st}} p_{sc}(x/\sqrt{st})dx\] is a rescaled semicircle distribution. Thus,
for a test function $f$, we have
\begin{align} \label{eq-PDE}
\dfrac{\partial^2}{\partial s \partial t} \left\langle f, \mu(s,t) \right\rangle
=& \dfrac{\partial^2}{\partial s \partial t} \int_{\mathbb{R}} \dfrac{1}{\sqrt{st}} f(x) p_{sc}(x/\sqrt{st}) dx \nonumber \\
=& \dfrac{\partial^2}{\partial s \partial t} \int_{\mathbb{R}} f(\sqrt{st} x) p_{sc}(x) dx \nonumber \\
=& \dfrac{\partial}{\partial s} \int_{\mathbb{R}} \dfrac{\sqrt{s} x}{2 \sqrt{t}} f'(\sqrt{st} x) p_{sc}(x) dx \nonumber \\
=& \int_{\mathbb{R}} \dfrac{x^2}{4} f''(\sqrt{st} x) p_{sc}(x) dx + \int_{\mathbb{R}} \dfrac{x}{4 \sqrt{st}} f'(\sqrt{st} x) p_{sc}(x) dx \nonumber \\
=& \int_{\mathbb{R}} \dfrac{x^2}{4 (st)^{3/2}} f''(x) p_{sc}(x/\sqrt{st}) dx + \int_{\mathbb{R}} \dfrac{x}{4 (st)^{3/2}} f'(x) p_{sc}(x/\sqrt{st}) dx \nonumber \\
=& \dfrac{1}{4st} \left\langle x^2 f''(x), \mu(s,t) \right\rangle + \dfrac{1}{4st} \left\langle x f'(x), \mu(s,t) \right\rangle.
\end{align}
Noting that the density of the measure $\hat \mu_t(dx)$ is $\hat p_t(x) = \frac{1}{\sqrt{t}} p_{sc}(x/\sqrt{t})$, the left-hand side of
\eqref{eq-Dyson limit equation} can be written as
\begin{align} \label{eq-limit left}
\dfrac{\partial}{\partial t} \int_{\mathbb{R}} f(x) \hat \mu_t(dx)
=& \dfrac{\partial}{\partial t} \int_{\mathbb{R}} \dfrac{1}{\sqrt{t}} f(x) p_{sc}(x/\sqrt{t}) dx \nonumber \\
=& \dfrac{\partial}{\partial t} \int_{\mathbb{R}} f(\sqrt{t} x) p_{sc}(x) dx \nonumber \\
=& \int_{\mathbb{R}} \dfrac{x}{2\sqrt{t}} f'(\sqrt{t} x) p_{sc}(x) dx.
\end{align}
Similarly, the right-hand side of \eqref{eq-Dyson limit equation} can be written as
\begin{align} \label{eq-limit right}
\dfrac{1}{2} \iint_{\mathbb{R}^2} \dfrac{f'(x) - f'(y)}{x-y} \hat \mu_t(dx) \hat \mu_t(dy)
= \dfrac{1}{2} \iint_{\mathbb{R}^2} \dfrac{f'(\sqrt{t} x) - f'(\sqrt{t} y)}{\sqrt{t} (x-y)} p_{sc}(x) p_{sc}(y) dxdy.
\end{align}
Substituting \eqref{eq-limit left} and \eqref{eq-limit right} into \eqref{eq-Dyson limit equation}, we get
\begin{align} \label{eq-semicircle identity}
\int_{\mathbb{R}} x f'(\sqrt{t} x) p_{sc}(x) dx
= \iint_{\mathbb{R}^2} \dfrac{f'(\sqrt{t} x) - f'(\sqrt{t} y)}{x-y} p_{sc}(x) p_{sc}(y) dxdy, ~ \forall t > 0.
\end{align}
Taking derivative with respect to $t$ for both sides, we have
\begin{align} \label{eq-semicircle identity derivative}
\int_{\mathbb{R}} x^2 f''(\sqrt{t} x) p_{sc}(x) dx
= \iint_{\mathbb{R}^2} \dfrac{xf''(\sqrt{t} x) - yf''(\sqrt{t} y)}{x-y} p_{sc}(x) p_{sc}(y) dxdy, ~ \forall t > 0.
\end{align}
Now, combining \eqref{eq-PDE}, \eqref{eq-semicircle identity} and \eqref{eq-semicircle identity derivative}, we have
\begin{align}
\dfrac{\partial^2}{\partial s \partial t} \left\langle f, \mu(s,t) \right\rangle
=& \int_{\mathbb{R}} \dfrac{x^2}{4} f''(\sqrt{st} x) p_{sc}(x) dx
+ \int_{\mathbb{R}} \dfrac{x}{4 \sqrt{st}} f'(\sqrt{st} x) p_{sc}(x) dx \nonumber \\
=& \dfrac{1}{4} \iint_{\mathbb{R}^2} \dfrac{xf''(\sqrt{st} x) - yf''(\sqrt{st} y)}{x-y} p_{sc}(x) p_{sc}(y) dxdy \nonumber \\
&+ \dfrac{1}{4\sqrt{st}} \iint_{\mathbb{R}^2} \dfrac{f'(\sqrt{st} x) - f'(\sqrt{st} y)}{x-y} p_{sc}(x) p_{sc}(y) dxdy \nonumber \\
=& \dfrac{1}{4} \iint_{\mathbb{R}^2} \dfrac{xf''(x) - yf''(y)}{x-y} \cdot \dfrac{1}{st} p_{sc}(x/\sqrt{st}) p_{sc}(y/\sqrt{st}) dxdy \nonumber \\
&+ \dfrac{1}{4} \iint_{\mathbb{R}^2} \dfrac{f'(x) - f'(y)}{x-y} \cdot \dfrac{1}{st} p_{sc}(x/\sqrt{st}) p_{sc}(y/\sqrt{st}) dxdy \nonumber \\
=& \dfrac{1}{4} \iint_{\mathbb{R}^2} \dfrac{xf''(x) - yf''(y)}{x-y} \mu(s,t)(dx) \mu(s,t)(dy) \nonumber \\
&+ \dfrac{1}{4} \iint_{\mathbb{R}^2} \dfrac{f'(x) - f'(y)}{x-y} \mu(s,t)(dx) \mu(s,t)(dy).
\end{align}
Therefore, we get the following McKean-Vlasov equation for $\mu(s,t)(dx)$:
\begin{equation}\label{e:MV}
\dfrac{\partial^2}{\partial s \partial t} \left\langle f, \mu(s,t) \right\rangle=
\frac14 \iint_{\mathbb{R}^2} \frac{\left(xf'(x)\right)'-\left(yf'(y)\right)'}{x-y} (\mu(s,t))^{\otimes 2}(dx,dy).
\end{equation}
Now we consider the Stieltjes transform of $\mu(s,t)(dx)$:
\begin{align*}
G_{s,t}(z) = \left\langle \dfrac{1}{z-x}, \mu(s,t) \right\rangle, \mbox{ for } z \in \mathbb{C}\backslash \mathbb{R}.
\end{align*}
Note that the Stieltjes transform $G(z)$ of the semicircle distribution $p_{sc}(x) dx$ can be written as
\begin{align} \label{eq-08}
G(z) = \left\langle (z-x)^{-1}, \mu_{sc} \right\rangle
=& \int_{\mathbb{R}} \dfrac{1}{z-x} p_{sc}(x) dx \nonumber \\
=& \int_{\mathbb{R}} \dfrac{1}{z-x/\sqrt{st}} p_{sc}(x/\sqrt{st}) \dfrac{dx}{\sqrt{st}} \nonumber \\
=& \int_{\mathbb{R}} \dfrac{\sqrt{st}}{\sqrt{st}z-x} \tilde p_{s,t}(x) dx \nonumber \\
=& \sqrt{st} G_{s,t}(\sqrt{st} z).
\end{align}
By \cite[(2.4.6)]{anderson2010} (see also \cite[(2.103)]{Tao2012}), $G(z)$ solves
\begin{align} \label{eq-09}
G(z)^2 - z G(z) + 1 = 0.
\end{align}
Substituting \eqref{eq-08} into \eqref{eq-09}, we have
\begin{align*}
st \left( G_{s,t}(\sqrt{st} z) \right)^2 - z \sqrt{st} G_{s,t}(\sqrt{st} z) + 1 = 0,
\end{align*}
which can be rewritten as
\begin{align} \label{eq-derivative 0 order}
st \left( G_{s,t}(z) \right)^2 - z G_{s,t}(z) + 1 = 0.
\end{align}
Taking the derivative with respect to $z$ in \eqref{eq-derivative 0 order}, we get
\begin{align} \label{eq-derivative 1 order}
2st G_{s,t}(z) \partial_z G_{s,t}(z) - z \partial_z G_{s,t}(z) - G_{s,t}(z) = 0.
\end{align}
Take the derivative with respect to $z$ in \eqref{eq-derivative 1 order}, we have
\begin{align} \label{eq-derivative 2 order}
2st \left( G_{s,t}(z) \partial_z^2 G_{s,t}(z) + \left( \partial_z G_{s,t}(z) \right)^2 \right) - z \partial_z^2 G_{s,t}(z) - 2\partial_z G_{s,t}(z) = 0.
\end{align}
Now, by choosing $f(x) = (z-x)^{-1}$ in \eqref{eq-PDE}, we have
\begin{align} \label{eq-Stieltjes transform-PDE}
\dfrac{\partial^2}{\partial s \partial t} G_{s,t}(z)
=& \dfrac{1}{2st} \left\langle \dfrac{x^2}{(z-x)^3}, \mu(s,t) \right\rangle
+ \dfrac{1}{4st} \left\langle \dfrac{x}{(z-x)^2}, \mu(s,t) \right\rangle \nonumber \\
=& \dfrac{1}{2st} \left\langle \dfrac{(z-x)^2 - 2z(z-x) + z^2}{(z-x)^3}, \mu(s,t) \right\rangle
+ \dfrac{1}{4st} \left\langle \dfrac{(x-z) + z}{(z-x)^2}, \mu(s,t) \right\rangle \nonumber \\
=& \dfrac{1}{4st} \left\langle \dfrac{1}{z-x}, \mu(s,t) \right\rangle
- \dfrac{3z}{4st} \left\langle \dfrac{1}{(z-x)^2}, \mu(s,t) \right\rangle
+ \dfrac{z^2}{2st} \left\langle \dfrac{1}{(z-x)^3}, \mu(s,t) \right\rangle \nonumber \\
=& \dfrac{1}{4st} G_{s,t}(z) + \dfrac{3z}{4st} \partial_z G_{s,t}(z) + \dfrac{z^2}{4st} \partial_z^2 G_{s,t}(z) \nonumber \\
=& \dfrac{1}{4st} \big( G_{s,t}(z) + z \partial_z G_{s,t}(z) \big)
+ \dfrac{z}{4st} \big( 2 \partial_z G_{s,t}(z) + z \partial_z^2 G_{s,t}(z) \big) \nonumber \\
=& \dfrac{1}{2} G_{s,t}(z) \partial_z G_{s,t}(z) + \dfrac{z}{2} \left( G_{s,t}(z) \partial_z^2 G_{s,t}(z) + \left( \partial_z G_{s,t}(z) \right)^2 \right),
\end{align}
where the last equality follows from \eqref{eq-derivative 1 order} and \eqref{eq-derivative 2 order}. Therefore, we have the following
generalized Burgers' equation for $G_{s,t}(z)$
\begin{equation}\label{e:Burgers'}
\dfrac{\partial^2}{\partial s \partial t} G_{s,t}(z)=\dfrac{1}{2} G_{s,t}(z) \partial_z G_{s,t}(z) + \dfrac{z}{2}
\left( G_{s,t}(z) \partial_z^2 G_{s,t}(z) + \left( \partial_z G_{s,t}(z) \right)^2 \right).
\end{equation}
|
\section{Introduction}\label{intro}
Buildings in the United States consume about 40\% of total primary energy resulting in a fifth of total greenhouse gas emissions. Energy efficiency in buildings is a climate change imperative and a significant business opportunity. Electric utilities benefit significantly from energy efficiency through deferral of transmission and distribution system infrastructure upgrades. With increasing penetration of renewables, the conventional power generation profile has changed drastically; it now has to deal with increased variability. Buildings can provide flexibility from the demand side to adapt to variability of generation. To take full advantage of greenhouse gas emission reductions as well as grid flexibility arising from building operation and energy conservation measures, one needs to perform simulations of numerous scenarios. Although, for some limited purposes, such as monitoring and verification (M\&V), fault detection and simple modifications to building operation, one can use empirical/statistical models, for more comprehensive cases properly reconciled physics-based detailed simulations are essential.
Buildings rarely perform as simulated, even after accounting for design vs as-built differences. Hence the need for reconciliation using monitored data. As noted by Reddy \cite{reddy2006}, “Historically, the calibration process has been an art form that inevitably relies on user knowledge, past experience, statistical expertise, engineering judgment, and an abundance of trial and error”. This approach hardly informs design or simulation. The building performance gap – the serious discrepancy between modeled and actual energy performance – has often caused frustration \cite{perfgap} \cite{teleg}. In the context of building energy, the term calibration is pervasively used to denote any method of reconciling simulations with performance. In this paper we identify and distinguish between tuning, calibration and parameter estimation. Most methods fall within the tuning category.
ASHRAE initiated a research project (RP1051) in 2004 intended to identify the best tools, techniques, approaches, procedures from the existing body of research and to develop a coherent and systematic calibration methodology involving a Monte Carlo approach that includes both “parameter estimation” and determination of the uncertainty in the calibrated simulation. The research was restricted to reconciling detailed building energy simulation programs against utility bills \cite{reddyetal}. Subsequently, several researchers have made computational advances, still relying on a variant of Monte Carlo, resulting in the development of AutoBEM (Automatic Building Energy Model) \cite{autotune} to automate the calibration process. The Autotune methodology uses multiparameter optimization techniques, in combination with data mining-informed artificial intelligence agents, to automatically modify software inputs so that simulation output matches measured data. The approach involves identifying about 150 of the most important parameters and using machine learning algorithms to “learn” successful versus unsuccessful paths to optimization. However, the approach relies on brute-force with little attempt made to analyze the data in a manner that attaches error estimates and informs building design. Further, the problem is intrinsically over-parametrized and this adds a large degree of indeterminacy at a fundamental level. There is little expectation that the reconciled model has captured the right numerical values of the various input parameters.
A solution, in principle, to this problem of a “scientific” gap between simulations and performance data is to perform a true parameter estimation using standard techniques, such as the Levenberg-Marquardt method; see, for example, Chapter 15 of the book “Modeling of Data” \cite{numrec}. This approach also gives an estimate of the errors in the parameters. A brute force approach of estimating the input parameters to the simulation is not feasible, because there are far too many of them. A new reformulation that transforms the problem into one with a manageably few parameters is given in this paper.
Such a view was taken by Sun and Reddy \cite{sun_reddy} who proposed a building calibration methodology, along the lines of parameter estimation, that “involves several distinct concepts, namely, sensitivity analysis (to identify a subset of strong influential variables), identifiability analysis (to determine how many parameters of this subset can be tuned mathematically and identifies which ones are the best candidates), numerical optimization (to determine the numerical values of this best subset of parameters), and uncertainty analysis (to deduce the range of variation of these parameters)”. Some aspects of this study are adopted in the new methodology for HVAC calibration, but there are important differences as described later.
A true parameter estimation method for reconciling an in-house simulation with actual performance, called PSTAR (Primary and Secondary Terms Analysis and Renormalization) was introduced by Subbarao. \cite{pstar}. In conjunction, an experimental procedure called Short-Term Energy Monitoring (STEM) was developed that provides short-term test specifications and analysis to elicit physically relevant parameters \cite{stemres} \cite{stem}. PSTAR provides a method for identifying and computing primary and secondary terms in energy balances, and then introduces certain parameters that are estimated from monitored data to enforce energy balance in actual buildings. PSTAR analysis of heat flows provides a conceptual framework for the method described here.
At this point, it is necessary to have unambiguous terminology. We will assign precise meaning to the terms: tuning, calibration and parameter estimation subsequently. The following Section lays the ground work for that discussion.
\section{Macro Heat Flows: A Simple Case}\label{macrohf}
Working with macro heat flows contributing to energy balance is central to our methodology of reconciling simulations with measured data. A typical simulation, e.g., EnergyPlus, determines micro heat flows reaching, for example, the indoor air node from each of the other nodes (as well as any direct input). It keeps track of a large number of nodes, and thereby a large number of micro heat flows, but does not distinguish between how much of the heat flow is due to the major external drivers such as indoor temperature, outdoor temperature, solar radiation etc. By contrast, we will consider macro heat flows reaching, for example the indoor air node, from all nodes that can be attributed to a specific driving function. Crucially, we develop a method to obtain these heat flows from specialized EnergyPlus simulations.
Let us start with a simple case. Consider a box with walls (floor included) with nominal conductance $U_{wall,nom}$, area $A_{wall}$, ceiling with nominal conductance $U_{ceiling,nom}$, area $A_{ceiling}$ and windows with nominal conductance $U_{window,nom}$, area $A_{window}$. Assume the heat capacity of all elements to be zero. The box is electrically heated to maintain an inside temperature of $T_{in}(n)$ at time step n, while the ambient temperature is $T_{amb}(n)$. A “simulation” or an expression for the combined energy heat flows gives the heating energy use as
\begin{multline} \label{simple_ebe}
Q_{sim,nom}(n)= U_{wall,nom} A_{wall} (T_{in}(n)-T_{amb}(n))+ U_{ceiling,nom} A_{ceiling} (T_{in}(n)-T_{amb}(n))+\\
U_{window,nom} A_{window}(T_{in}(n)-T_{amb}(n))
\end{multline}
where the subscript “nom” refers to nominal or first-guess or assumed values.
Compare this with measured heat $Q_{mea}(n)$ (which from physical considerations, under the above-stated assumption of zero heat capacity, will be proportional to $T_{in}(n)-T_{amb}(n))$. It is most likely that the measured heating energy will differ from the simulated heating energy because the assumed input values of $U_{wall,nom}$, $U_{ceiling,nom}$, and $U_{window,nom}$ are estimates and will differ from the actual values.
We can tune by tweaking individual values to $U_{wall,tune}$, $U_{ceiling,tune}$, and $U_{window,tune}$ such that the resulting $Q_{sim,tune}(n)$ is as close as possible (in the least squares error sense) to $Q_{mea}(n)$ . This can be done in an infinite number of ways, and serves to illustrates the concept of over-parametrization as stated earlier. An attempt to perform estimation of the micro parameters $U_{wall,tune}$, $U_{ceiling,tune}$, and $U_{window,tune}$ from the regression equation
\begin{multline}
Q_{mea}(n) \approx U_{wall,tune} A_{wall}(T_{in}(n)-T_{amb}(n))+ U_{ceiling,tune} A_{ceiling}(T_{in}(n)-T_{amb}(n))+\\U_{window,tune} A_{window}(T_{in}(n)-T_{amb}(n))
\end{multline}
is mathematically impossible. Only the combined term
\begin{equation}
BLC \triangleq U_{wall,tune} A_{wall}+ U_{ceiling,tune} A_{ceiling}+U_{window,tune} A_{window}
\label{eqn:BLCdef}
\end{equation}
is well-determined. This quantity is called Building Load Coefficient (BLC). In other words, the regression equation should be modified to
\begin{equation}
Q_{mea}(n) \approx p_{BLC} Q_{sim,nom}(n)
\end{equation}
where $p_{BLC}$ is the single parameter to be estimated. The deviation of the estimate of $p_{BLC}$ from 1 is a measure of the over or under estimation of BLC. This clearly shows the advantage of working with macro heat flows, in this simple case, namely $Q_{sim,nom}(n)$
Consider now the more realistic case when the box elements have varying amount of thermal mass. The energy balance equation is more complex, and will need to consider transient behavior, i.e., past values of temperatures etc. It is no longer clear what specific combination is well-determined. We can surmise that it will involve effects such as admittances rather than U-values. Once again, we seek a formulation that handles these effects implicitly while only exposing a small number of “knobs” to turn such that a true parameter estimation can be done. What these knobs are is considered at length later. Again, we find it is advantageous to work with macro heat flows.
Let us now add an HVAC system to the box. In cases where an overall equipment/system efficiency can realistically capture or account for actual behavior, it is added as an additional knob. For example, the rated COP of a chiller or the full-load efficiency of a boiler. Even for more complex systems, the building envelope parameters estimated above enable us to use the box as a dynamical calorimeter (i.e., an “instrument” that measures the amount of heat introduced or extracted by the HVAC system knowing the driving forces: inside and outside temperatures, solar radiation etc.) to determine heating and cooling provided in each time step. This would then allow us to estimate the HVAC characteristics.
The PSTAR \cite{pstar} approach uses building description to compute zonal admittances and resulting heat flows by an in-house simulation, and then renormalizes (i.e., introduces and estimates parameters such as $p_{BLC}$) the heat flows to achieve the most accurate energy balance. The patent \cite{patent} describes a method to compute heat flows using the simulation program with specialized forcing functions. The method described here represents major improvement over the previous method.
\section{Tuning, Calibration and Parameter Estimation}\label{sec:whatiscalib}
Some discussion on terminology is warranted considering the fact that calibration is loosely used as an umbrella term for any method to reconcile simulation with performance data. The terms “tuning”, “calibration”, and “parameter estimation” need to be defined more precisely. Tuning is essentially a process involving varying the large number of input parameters either subjectively or based on some expected probability. A set of “best performing inputs” are retained, and an iterative process is followed until the (set of) simulation inputs is deemed satisfactory. Such subjective approaches have intrinsic limitations due to the large number of inputs that can be selected in a numerous different permutations to yield equivalent outputs (as illustrated earlier in this paper). Tuning is currently common in both research and professional studies.
Calibration is a well-established term in metrology; it essentially means modifying the raw output of a measuring device with a previously-determined correlation to obtain an improved measurement. The metrology-type calibration process is shown in Figure \ref{fig1}. The figure depicts the process that uses the previously determined calibration function say by comparing the instrument raw readings against those from a more accurate reference standard. During training, the known output is used to determine the calibration function. We can think of a simulator such as EnergyPlus as the measuring device, and, its energy use and indoor temperature time series as raw outputs. In this case, there cannot be a simple one-to-one relationship between the raw outputs and calibrated outputs. A generalization is depicted in the Figure \ref{fig2}. EnergyPlus raw output is used as a starting point to empirically obtain improved predictions. For example, using weather and indoor temperature time series as well as energy use time series from EnergyPlus as inputs, we can train a neural net using measured energy use; this can be subsequently used for improved predictions. (This approach must be contrasted with the common method of training a neural net with weather and indoor temperature time series as inputs wherein no simulation is involved). The network parameters have no direct physical interpretation. Further, implementing thermostatic constraints is a problem. We will not pursue this metrology-like calibration approach. What we propose and formulate in detail is an enhanced parameter estimation approach we call EPE (Figure \ref{fig3}). The term "calibration" is so ubiquitous for any reconciliation method, we will use that term loosely.
\begin{figure}[htbp]
\caption{Schematic to illustrate how a calibration function is used to modify the measured output of a simple device e.g., thermometer. The calibration function is determined initially from known outputs through obvious modifications of the flow chart}
\centering
\includegraphics[width=0.8\textwidth]{1}
\label{fig1}
\end{figure}
\begin{figure}[htbp]
\caption{Schematics of a generalization of the “calibration” process to simulations. Initially, the neural net is trained from measured data through obvious modifications of the flow chart}
\centering
\includegraphics[width=0.8\textwidth]{2}
\label{fig2}
\end{figure}
\begin{figure}[htbp]
\caption{Schematics of how to deploy the Enhanced Parameter Estimation Method. Initially, the parameters are estimated and neural net trained from measured data as described in the text. Note that no modification of shell input parameters of the EnergyPlus model is done. Deviations of simulation inputs from actual are accounted for by corrective heat flows.}
\centering
\includegraphics[width=0.8\textwidth]{3}
\label{fig3}
\end{figure}
\section{Enhanced Parameter Estimation (EPE)}\label{sec:epe}
This section describes the mathematical formulation of the method, the multiple specialized EnergyPlus simulations needed, parameters and their estimations, and using machine learning approach to the residuals. Results of applying this methodology to a synthetic building and to a real building are described below.
\subsection{Formulation of the Method}
A simulator such as EnergyPlus performs an energy balance at a number of nodes, in particular at each surface node and each zone air node. Measured performance data on buildings does not permit examining each of these nodes. Let us examine the energy balance at each of the zone air nodes. Consider a linearized version of energy flows in a one zone building; nonlinearities are addressed later through a neural net model for the residuals. The heat flow at time t due to current and past indoor and outdoor temperatures $T_{in}(t')$ and $T_{out}(t')$ can be written as (with the convention that a positive value implies heat gain by the air node) a convolution integral (see \cite{doe2engman}):
\begin{equation}
\int_{-\infty}^t dt' \, \left[ -V(t-t')T_{in}(t')+W(t-t')T_{out}(t') \right]
\label{eqn:conv}
\end{equation}
This can be rewritten as a sum of two heat flows $Q_{in}(t)$ and $Q_{BLC}(t)$
\begin{equation}
Q_{in}(t) =-\int_{-\infty}^t dt' \, V(t-t')T_{in}(t') +LT_{in}(t)
\label{eqn:QinDef}
\end{equation}
\begin{equation}
Q_{BLC}(t) = -L \left[ T_{in}(t) - T_{out}^{eff}(t) \right]
\label{eqn:QBLCgendef}
\end{equation}
\begin{equation}
T_{out}^{eff}(t) \triangleq \int_{-\infty}^t dt' \, \frac{W(t-t')}{L} T_{out}(t')
\label{eqn:Touteffdef}
\end{equation}
The quantities $V(t-t')$ and $W(t-t')$ are the appropriate transfer functions. They can be computed from a building description. We do not need to explicitly compute them. If all exterior walls are extremely light, $W(t-t')=L\delta(t-t')$ and the effective outside temperature is equal to the actual instantaneous outside temperature. As the walls become more massive, the effective outside temperature is a weighted average of current and past temperatures, the weights depending on the wall characteristics. For a lightweight wall, the weights decrease rapidly unlike for a heavyweight wall. (It is possible to use only the current outside temperature to define this term but this results in undesirable for heavyweight walls). Sky temperature effects can easily be included by redefining the effective outside temperature:
\begin{equation}
T_{out}^{eff}(t) \triangleq \int_{-\infty}^t dt' \, \left[ \frac{W(t-t')}{L} T_{out}(t') + \frac{W_{sky}(t-t')}{L} T_{sky}(t') \right]
\label{eqn:Teffdefinclsky}
\end{equation}
Note that a continuous time formulation is used. It is straightforward to switch between continuous and discrete formulation as convenient.
We will now add $Q_{sun}(t)$ to account for solar gains, and $Q_{LEP}(t)$ to account for gains from light, equipment and people that appear as heat gain at the indoor-air node. Note in particular that the term $Q_{LEP}(t)$ is the heat gain by the air node after accounting for the delayed release of the gains and any loss through the walls without appearing as gains at the air node. Additional energy flow terms such as infiltration, ventilation and ground flow will be included when necessary. For the time being, for pedagogical simplicity, we have set them to zero.
The heating/cooling load at the zone air node can be written as:
\begin{equation}
Q_{load}(t) \triangleq Q_{BLC}(t)+Q_{in}(t)+Q_{sun}(t)+Q_{LEP}(t)
\label{eqn:load}
\end{equation}
Let us denote by $Q_{HC}(t)$ the heating/cooling delivered to the air node by the HVAC system. Finally, the energy balance equation for the air node is:
\begin{equation}
Q_{load}(t) + Q_{HC}(t) = 0
\label{eqn:ebe}
\end{equation}
The building load $Q_{HC}(t)$ is met by energy $E(t)$ (electricity and/or gas collectively) that is a function of $Q_{HC}(t)$, weather variables such as outdoor temperature and humidity collectively denoted by Weather(t), and a set of system parameters such as COP, part load efficiency etc. collectively denoted by \{$\alpha$\}:
\begin{equation}
E(t) = Function(Q_{HC}(t), Weather(t), \{\alpha\})
\label{eqn:ener}
\end{equation}
The energy use $E(t)$ is an output of the simulation. Our main problem is how best to reconcile simulated values of $E(t)$ with the measured values.
The basic methodology is best explained through a one zone building. Variations can be addressed as needed.
\newline
\newline
We start with:
\begin{itemize}
\item
An inexpensively created EnergyPlus simulation model, called audit model, that is the one to be reconciled with performance data
\item{The available performance data consists of time series of}
\begin{itemize}
\item{Indoor temperature $T_{in,measured}(t)$}
\item{Weather (outdoor temperature, solar radiation,...) collectively denoted by $Weather(t)$}
\item{Heat input from lights, equipment and people $LEP(t)$}
\item{Energy Use (electricity, gas) collectively denoted by $E(t)$}
\item{Additionally, short-term data for the actual heat input $Q_{HC}^{mea}(t)$ (may be negative) possibly through flow $\Delta$T type measurements}
\end{itemize}
\end{itemize}
The analysis consists of two stages. In Stage 1, the building shell related parameters are estimated; HVAC system is replaced by an ideal system that simply delivers the required heating/cooling. In Stage 2, the HVAC related parameters are estimated; the building shell is replaced by a box with only process loads.
\newline
\newline
Stage 1: Building Shell parameters
\begin{itemize}
\item
Perform an EnergyPlus simulation of the audit building, for the period for which $Q_{measured}(t)$ is available using $Weather(t)$, $LEP(t)$ with $T_{in,measured}(t)$ as the set-points. Denote the resulting load by $Q_{1}(t)$. If the audit building shell accurately represents the real building shell (and the simulator is accurate), $Q_{1}(t)$ would be equal to $Q_{measured}(t)$. We do not expect this to be the case and now need to reconcile simulated and actual performances.
\item
Perform a second EnergyPlus run, this time with the solar radiation set to 0, and obtain the required input $Q_{2}(t)$. The heat flow
\begin{equation}
Q_{sun}(t) \triangleq Q_{2}(t) - Q_{1}(t)
\label{eqn:Qsundef}
\end{equation}
is identified as the heat input to the indoor air node due to solar radiation after accounting for any delayed release.
\item
Perform a third EnergyPlus run, this time with the indoor temperature fixed at a constant value $T_{in,fixed1}$ and the solar radiation set to 0, to obtain the required input $Q_{3}(t)$.
\item
Perform a fourth EnergyPlus run, this time with the indoor temperature fixed at a constant value $T_{in,fixed2}$ and the solar radiation set to 0, to obtain the required input $Q_{4}(t)$.
\item
Perform a fifth EnergyPlus run, this time with the indoor temperature fixed at a constant value $T_{in,fixed1}$ and the solar radiation as well as Light, Equipment and People heat input set to 0, to obtain the required input $Q_{5}(t)$.
\item
Obtain the following heat flows:
\begin{equation}
Q_{in}(t) \triangleq -Q_{2}(t)+Q_{3}(t) \frac{T_{in,measured}(t)-T_{in,fixed2}}{T_{in,fixed1}-T_{in,fixed2}} +\\Q_{4}(t) \frac{T_{in,fixed1}-T_{in,measured}(t)}{T_{in,fixed1}-T_{in,fixed2}}
\label{eqn:Qindef}
\end{equation}
\begin{equation}
Q_{BLC}(t) \triangleq -Q_{2}(t) + Q_{3}(t)-Q_{5}(t) - Q_{in}(t)
\label{eqn:QBLCdef}
\end{equation}
\begin{equation}
Q_{LEP}(t) \triangleq -Q_{3}(t) + Q_{5}(t)
\label{eqn:QLEPdef}
\end{equation}
Eqs \ref{eqn:Qsundef} and \ref{eqn:QLEPdef} make intuitive sense. Note that $Q_{LEP}(t)$ is the effect of ${LEP}(t)$ after accounting for any time delays and losses. The term $Q_{in}(t)$ is the heat flow to the air node based on the history of indoor temperatures. $Q_{BLC}(t)$ is the heat flow to the air node based on the current indoor temperature and an effective outdoor temperature that is based on the history of outdoor temperatures (including contributions from the history of sky temperature depression)
\item Introduce parameters to be estimated and estimate them
\newline
By construction, the following equation is satisfied:
\begin{equation}
Q_{BLC}(t)+Q_{in}(t)+Q_{sun}(t)+Q_{LEP}(t)+Q_{1}(t)=0
\label{eqn:QIdentdef}
\end{equation}
We now bring in $Q_{HC}^{mea}(t)$, and require that Eq.\ref{eqn:QIdentdef} be satisfied with $Q_{HC}^{mea}(t)$, not $Q_{1}(t)$. We modify the rest of the heat flows by introducing parameters and estimate them to satisfy the equation in the least squares error sense. The simplest modification is:
\begin{equation}
p_{BLC}Q_{BLC}(t)+p_{in}Q_{in}(t)+p_{sun}Q_{sun}(t))+p_{LEP}Q_{LEP}(t)+Q_{HC}^{mea}(t)\approx 0
\label{eqn:Qlinmoddef}
\end{equation}
This results in a simple linear regression problem. These parameters change only the scale of a heat flow but not the shape. To change the shape we introduce additional parameters. A heat flow, e.g., $Q_{sun}(t)$ is modified through a transfer function as follows:
\begin{equation}
Q_{sun}^{mod}=p_{sun}Q_{sun}(t))+Q_{sun}^{TF}(t)
\label{eqn:Qsunmoddef}
\end{equation}
where
\begin{equation}
Q_{sun}^{TF}(t)=\alpha_{sun} Q_{sun}^{TF}(t-1)+\beta_{sun} (Q_{sun}(t)-Q_{sun}(t-1))
\label{eqn:Qsuntfdef}
\end{equation}
The new energy balance equation is
\begin{multline}
p_{BLC}Q_{BLC}(t) + p_{in}Q_{in}(t) + p_{sun}Q_{sun}(t) + p_{LEP}Q_{LEP}(t) + \\Q_{TF,sun}(t) + Q_{HC}^{mea}(t) \approx 0
\label{eqn:regressimproved}
\end{multline}
This introduces 2 additional parameters for the solar heat flow $\alpha_{sun}$, $\beta_{sun}$. Similar parameters can be introduced to the other heat flows. We have to be judicious in introducing additional parameters through an error analysis to ensure that the data supports their estimation. We now have a nonlinear regression problem. This can be solved using Levenberg-Marquardt method.
Finally, it is obvious that the simulated building would behave close to the real building if we add the following process load to the simulated building:
\begin{multline}
Q_{correction}(t) \triangleq (p_{BLC}-1)Q_{BLC}(t)+(p_{in}-1)Q_{in}(t)+(p_{sun}-1)Q_{sun}(t))+\\
(p_{LEP}-1)Q_{LEP}(t)+Q_{sun}^{TF}(t)
\label{eqn:QCorrection}
\end{multline}
Similar modifications are needed to accommodate transfer functions for any of the heat flows. One can create any number of corrective heat flows (for example, one due to sky temperature depression), but judicious, parsimonious choice based on experimentation is necessary. The EnergyPlus simulation we started with is modified by adding $Q_{correction}(t)$ as a process heat. The resulting output is EPE's calibrated output. EPE accomplishes shell calibration without any modification of the simulation inputs (as noted in the Introduction section, data does not support any systematic modification of the vastly undetermined problem); the deviation of the simulated building from the actual is corrected by introducing corrective heat flows.
\item
Examine the parameters in light of their physical significance
\newline
The deviation of, for example, $p_{BLC}$ from $1$ is a measure of the extent to which the initial estimate of BLC is over or under the actual estimate. The simulation input file can be reexamined in light of such deviations, but there is no systematic way to assign the correction to individual walls, windows etc. (Large deviations may point to possible bugs in inputs; such erroneous inputs should be obviously fixed.) Similarly all the other parameters should be examined.
\item
Develop a neural net for the residuals
\newline
With the various heat flows as inputs and residuals as output, a neural net is trained to improve the energy balance
\end{itemize}
Stage 2: HVAC System
\begin{itemize}
\item
Replace the building shell by a process load whose time series is given by:
\begin{equation}
Q_{reconciled}(t)\triangleq-Q_{BLC}^{mod}(t)-Q_{in}^{mod}(t)-Q_{sun}^{mod}(t)-Q_{LEP}^{mod}(t)+Q_{residual}(t)
\label{eqn:Qrecon}
\end{equation}
Specialized simulations are done to obtain the heat flows which are then modified to obtain the process load. (we can, of course, use $Q_{measured}(t)$ directly for the periods when it is available)
\item
Perform an EnergyPlus simulation with the audit HVAC system and the process load to determine the simulated energy use $E_{sim}(t)$.
\newline
If the audit description of the HVAC system is accurate, then $E_{sim}(t)$ will be close to $E_{measured}(t)$. This is generally not expected.
\item
Identify the parameters to be estimated and estimate them
\newline
In some ways this is similar to the shell problem but in many ways, there are important differences. First of all, each time step is essentially independent of the previous time steps (unless one is considering minute-by-minute control sequence, or storage elements as part of the HVAC system). There are as for the shell a large number of parameters and not all can be determined with whole building data.
\end{itemize}
It is important to note that we use hourly (or subhourly) data for temperatures, weather, and energy. This is a consistent match with a simulator that works with such inputs and outputs. Such data is available from energy management systems, smart meters and if necessary from nearby weather station. There is a disconnect in calibrating hourly simulation with monthly data (highly overparametrized and so a large number of diverse input variable sets will yield very good fits to utility bills). Unfortunately, this is quite common. In \cite{autotune}, the authors calibrate an EnergyPlus simulation model using hourly data, but without indoor zone temperatures. The thermal mass behavior in response to indoor temperature variations is important in terms of the relative magnitude of internal heat flows and neglecting this cannot yield proper calibrated models.
\newline We emphasize that EPE accomplishes shell calibration without any modification of the simulation inputs (as noted in the Introduction section, data does not support any systematic modification of the vastly undetermined problem); the deviation of the simulated building from the actual is corrected by introducing corrective heat flow (Eq. \ref{eqn:QCorrection}). One can come up with ways to modify the EnergyPlus input so that the resulting $p_{BLC}$ and other similar parameters are (close to) 1. For example, we can multiply all exterior wall layer conductivities by $p_{BLC}$. Any such revision can only be justified with additional component-level data.
\subsection {Synthetic Building Example:}
In this section, we will demonstrate the multi-stage methodology for a synthetic building i.e., a software model of a building from which we generate simulated data and use it as the “real” data. The “real” building is then taken to be the one simulated in EnergyPlus to give hourly time series of indoor temperatures and delivered heating and cooling energy in step 1 and electrical consumption by HVAC system in step 2. In this section, the terms “measured” and “real” refer to this building.
For this study, the "real" building is the 53,600 sq. ft. building with normal construction which corresponds to the medium office building prototype selected by USDOE for work related to ASHARE 90.1 standard development. The audit building has been intentionally modified so as to have different insulation, mass and solar characteristics; more specifically it has more insulation, more mass (added concrete wall layer) and smaller solar heat gain coefficient. Two locations were selected for analyzing such a building: Phoenix, AZ and Philadelphia, PA. A more detailed study was done for the Phoenix location.
Note that evaluation with a synthetic building allows us to vary the inputs at will and evaluate the accuracy of the Enhanced Parameter Estimation process with more certainty.
Step 1: Obtain the measured (hourly) values of indoor temperature $T_{in}(t)$ and delivered heating/cooling $Q_{HC}^{mea}(t)$ for the "real" building
Step 2: Run the five EnergyPlus simulations of the audit building as specified and determine the four heat flows $Q_{BLC}(t)$,$Q_{in}(t)$,$Q_{sun}(t)$,$Q_{LEP}(t)$.
\begin{figure}[htbp]
\caption{The heat flow $Q_{mea}(t)$ is measured and the rest computed (synthetic building). }
\centering
\includegraphics[width=0.8\textwidth]{4}
\label{fig4}
\end{figure}
These and the “measured” heating/cooling energy delivered are shown in Figure \ref{fig4}. The period of analysis covered two months: May-June. For clarity, only a representative 2-week period is shown in the figure. If the audit building were identical to the real building (and if EnergyPlus is a fully accurate simulator) the five heat flows would add up to 0, or equivalently the negative of the sum of the four computed flows would equal the delivered heating/cooling energy amounts.
\begin{figure}[htbp]
\caption{The delivered heating/cooling energy flows (equal to the negative sum of the four heat flows) before and after the best fit, compared to measured values}
\centering
\includegraphics[width=0.8\textwidth]{5}
\label{fig5}
\end{figure} This hourly sequence of negative values of the sum is labeled “Before” in Figure \ref{fig5}.
Step 3: Perform a non-linear least squares best fit. The negative sum after the best fit is also shown in Figure \ref{fig5}. Note the improvement. The “After” plot is much closer to the “mea” and captures peaks and valleys better. The parameters from the best fit are shown in Table \ref{tab:parphx}. The errors quoted are purely statistical. (To avoid collinearity, $p_{LEP}$ was set to 1 and not estimated in these runs. Such correlations can often be avoided by choosing suitable data windows. This issue needs further study).
\begin{table}[htbp]
\centering
\caption{The Estimated parameters (May-June data)}
\begin{tabular}{|p{4.5em}|p{7.315em}|p{7.315em}|}
\midrule
$p_{BLC}$ & 1.48$\pm$0.02 \\
\midrule
$p_{in}$ & 0.63$\pm$0.05 \\
\midrule
$p_{sun}$ & 1.29$\pm$0.01 \\
\midrule
$p_{LEP}$ & 1 \\
\midrule
$p_{in,phase}$ & 0.43$\pm$0.04 \\
\midrule
$p_{TF,in}$ & 0.94$\pm$0.01 \\
\midrule
$p_{sun,phase}$ & 1.35$\pm$0.02 \\
\midrule
$p_{TF,sun}$ & 0.988$\pm$0.002 \\
\bottomrule
\end{tabular}%
\label{tab:parphx}%
\end{table}%
Step 4: Identify a neural net model for the residuals of the best fit least squares model. To avoid confusion, let us call this PE\_Residuals (for Parameter Estimation Residuals). With the usual caveats, we proceed to improve the estimate by using neural nets. The PE\_residuals show some residual pattern not captured by the parameter estimation process. With the six heat flows $Q_{BLC}(t)$, $Q_{in}(t)$, $Q_{sun}(t)$, $Q_{LEP}(t)$, $Q_{TF,in}(t)$, $Q_{TF,sun}(t)$ as inputs and PE\_Residuals as output with 9 hidden nodes, a multilayer perceptron network was trained on the PE\_Residuals from the May-June data. The fitted PE\_Residuals are now added to the delivered heating/cooling from the parameter estimation step to improve the estimate. The improvement is significant as can be seen from a scatter plot, Figure \ref{fig6}, of the three heat flows vs measured flow:
\begin{figure}[htbp]
\caption{Scatter plot showing the extent to which the Enhanced Parameter Fit process improves the agreement with measured data (May-June data)}
\centering
\includegraphics[width=0.8\textwidth]{6}
\label{fig6}
\end{figure}
\begin{table}[htbp]
\centering
\caption{Goodness-of-fit statistics of the EPE process for the synthetic building for the period May-June}
\begin{tabular}{|p{4.5em}|p{7.315em}|p{7.315em}|p{7.315em}|}
\toprule
\multicolumn{1}{|r|}{} & Audit Building & After Parameter Estimation & After Parameter Estimation and Residual Neural Net (Enhanced Parameter Estimation) \\
\midrule
Mean Bias Error & 5.13 MJ & -0.02 MJ & -0.05 MJ \\
\midrule
Root Mean Square Error & 10.33 MJ & 3.23 MJ & 2.47 MJ \\
\bottomrule
\end{tabular}%
\label{tab:err}%
\end{table}%
The mean bias error and RMS error are shown in Table \ref{tab:err}. Note that after the parameter estimation step, the mean bias error is essentially zero after calibration while the RMSE value has been reduced three-fold. The RMSE is further reduced (from 3.23 MJ to 2.47 MJ) after the neural net fitting procedure is applied.
One of the thorny problems in applying neural networks to time series is handling past values of driving functions. In our case, this is a non-issue since the inputs are heat flows that already incorporate the history of driving functions. Thus, each row is treated independently. The time series problem has been reduced to an independent case problem!
\subsection{Physical Significance of the Parameters}
The parameters have physical significance which helps us better understand the performance of the building.
The parameter $p_{BLC}$ was estimated to be 1.48. This implies that the building load coefficient of the actual building is roughly 48\% higher than that of the audit building. We cannot go beyond this to determine which components (walls or windows etc.) are causing this discrepancy; reasonable conjectures would be circumstance specific. Since the “real” building is synthetic, we know the actual numerical values of the parameters . The differences between the actual and the audit building were such that the load coefficient of the actual building is roughly 40\% higher; this is in good agreement with our parameter estimation. The value of the load coefficient is somewhat dependent on the treatment of convection and radiation at the interior and exterior surfaces and this must be resolved if the differences are to be reconciled further.
Similarly, $p_{sun}$ = 1.29 implies that the actual building has roughly 29\% higher solar gains. The solar heat gain coefficient of the windows of the actual building was about 47\% higher; we need to consider also opaque gains. Similarly $p_{in}$ = 0.63 means that the actual building has less effective thermal mass; this is consistent with the differences between audit and actual building inputs to EnergyPlus. The other parameters modify the delay characteristics of solar gains and of heat flows due to indoor fluctuations beyond what is in the audit description.
\subsection{HVAC System}
So far, we have used May-June data with hourly delivered heating/cooling data available to us. This enabled us to estimate envelope parameters as well as train a neural net to address the envelope characteristics. We now will use the EPE methodology to predict loads on the mechanical systems, and thereby perform HVAC reconciliation.
If the delivered heating/cooling data is available, for example, through a BTU meter, we can simply use that data to analyze the HVAC system; in particular, we could estimate certain HVAC parameters. We can visualize a use-case situation where the needed instrumentation is installed temporarily and then removed to reduce operating costs, while being able to keep monitoring the HVAC performance continuously. The proposed procedure involving enhanced parameter estimation would permit using the building envelope as a calorimeter, essentially as a substitute for flow-$\Delta$T type instrumentation. We will discuss this aspect next.
Suppose we have July-August data with indoor temperatures and HVAC electricity use. We can use the indoor temperature data in EnergyPlus to compute the four heat flows as described in the previous section: $Q_{BLC}(t)$, $Q_{in}(t)$, $Q_{sun}(t)$, $Q_{LEP}(t)$ for the period July-August. From the previously estimated $p_{TF,in}$ and $p_{TF,sun}$ parameters, we can determine the two additional flows $Q_{TF,in}(t)$, $Q_{TF,sun}(t)$. We can now apply Eq. \ref{eqn:regressimproved} and include the hourly residual time series data from the neural net to get the best estimate of delivered cooling:
\begin{multline}
Q_{HVAC}^{estimate}(t)=p_{BLC}Q_{BLC}(t) + p_{in}Q_{in}(t) + p_{sun}Q_{sun}(t) + p_{LEP}Q_{LEP}(t) + \\ p_{TF,sun2}Q_{TF,sun}(p_{TF,sun1},Q_{sun}(t)) + p_{TF,in2}Q_{TF,in}(p_{TF,in1},Q_{in}(t)) +NN_{Res}
\label{eqn:delivered}
\end{multline}
The term $NN_{Res}$ represents the residual time series estimated from the neural net.
We can now run an EnergyPlus simulation with the estimated loads as process load inputs to an isolated box; the building envelope is replaced by this isolated box for this simulation. Energy Plus takes the load time series and outputs electricity use time series using the user-specified input for the HVAC parameters - COP under rated conditions, and part load, and off-rated performance. We can now perform a non-linear least squares estimation of the HVAC parameters. To do this using software packages requires repeated calls to EnergyPlus. It is important to recognize that the data permits estimating a very small number of HVAC parameters, and these should be chosen judiciously.
To illustrate the general approach, we assume that the only parameter to be estimated is the COP under rated conditions and leave part-load and off-rated corrections at their default values. We can run EnergyPlus for a range of values of COP. We compute the RMS error as a function of rated COP and identify the value at the minimum error; this is a manual process for accomplishing the nonlinear estimation. This process of searching for a COP values which results in minimum RMSE is shown in Figure \ref{fig7}.
\begin{figure}[htbp]
\caption{Search process for the best estimate for cooling system COP that minimizes RMSE}
\centering
\includegraphics[width=0.8\textwidth]{7}
\label{fig7}
\end{figure}
\newline
The rated COP that minimizes the error turned out to be 3.625. Because the real building is in fact synthetic, we know the real COP value was 3.5. The difference is due, in addition to uncertainties of the methodology, mainly to an artifact of EnergyPlus. For the building we can deduce the electricity time series from EnergyPlus in two ways: One is via the normal simulation. The second is to first get the cooling loads (with the right indoor temperatures specified) and introduce them to another simulation with the shell replaced by an isolated box. The two are expected to be essentially identical; however, we found a 2\% increase for the isolated box. This suggests that for some unidentified reason (to be investigated), EnergyPlus increases the load by 2\% while simulating the box. To match the normal electricity with a 2\% higher load use requires about a 2\% increase in COP.
\subsection{How repeatable are the parameter estimates}
To evaluate whether the parameter estimates are impacted by climate, we have repeated the analyses for the same months (May-June) in two different climates: Philadelphia and Phoenix. The results are shown in Table \ref{tab:table3}. The agreement is generally satisfactory considering we do not expect them to be identical. As is to be expected from physical considerations, differences in sky temperature, ground reflectivity between the two cities affect the parameters.
\begin{table}[htbp]
\centering
\caption{Estimated Parameters in two different climates along with uncertainty values}
\begin{tabular}{|p{4.065em}|p{7.315em}|p{7.315em}|}
\multicolumn{1}{r}{} & \multicolumn{1}{l}{Philadelphia} & \multicolumn{1}{l}{Phoenix} \\
\midrule
$p_{BLC}$ & 1.57$\pm$0.03 & 1.48$\pm$0.02 \\
\midrule
$p_{in}$ & 0.69$\pm$0.02 & 0.63$\pm$0.05 \\
\midrule
$p_{sun}$ & 1.42$\pm$0.01 & 1.29$\pm$0.01 \\
\midrule
$p_{LEP}$ & \multicolumn{1}{r|}{1} & \multicolumn{1}{r|}{1} \\
\midrule
$p_{in,phase}$ & 0.45$\pm$0.02 & 0.43$\pm$0.04 \\
\midrule
$p_{TF,in}$ & 0.70$\pm$0.02 & 0.94$\pm$0.01 \\
\midrule
$p_{sun,phase}$ & 1.37$\pm$0.03 & 1.35$\pm$0.02 \\
\midrule
$p_{TF,sun}$ & 0.954$\pm$0.002 & 0.988$\pm$0.002 \\
\bottomrule
\end{tabular}%
\label{tab:table3}%
\end{table}%
\subsection{Actual Building In Philadelphia}
For this study, we have identified a building, located in the Philadelphia Navy Yard, for which an EnergyPlus simulation input model was also developed from a previous study carried out by Ke Xu. \cite{KeXu}. Building 101 was initially built in 1911 as a marine barrack with the gross building floor area about 75,000 ft$^2$. This was a brick building with double-pane windows and underwent a major HVAC system renovation in 1999. The building had 3 above grade floors and one Basement floor which primarily served as mechanical and storage rooms. The HVAC system in the building comprised of three variable-air-volume (VAV) air handling units (AHUs) with total design airflow of 53,240 CFM. The terminal units have hydraulic reheat coils to meet the heat / reheat loads which are served by a 1,632 MBH gas-fired boiler. The AHUs are served by DX cooling coils. The boiler supplies 180$^o$F hot water to AHU heating coils, Terminal Box reheat coils, and hot water radiators. The attic spaces are heated by hot water unit heaters. Further details can be found in \cite{KeXu}.
For our study, we had access to the 2016 end-use energy data for each minute, which included lighting electric, AHU fans electric, DX Coils electric, pumps electric, total building electric and miscellaneous equipment electric energy as residual. Gas consumption data by the boilers were also available for each minute. We carried out the analysis at an hourly level. The weather file for this location, for the year 2016, was obtained from Whitebox Technologies \cite{whitebox}.
Real buildings, unlike synthetic buildings, present unique challenges; the EPE methodology should be adapted to meet the challenges. Such inevitable complexities in dealing with real buildings should not detract from the methodology itself.
\begin{figure}[htbp]
\caption{Measured data for the actual building. Only a representative winter period is shown}
\centering
\includegraphics[width=0.8\textwidth]{8}
\label{fig8}
\end{figure}
\begin{figure}[htbp]
\caption{Measured data for the actual building. Only a representative summer period is shown}
\centering
\includegraphics[width=0.8\textwidth]{9}
\label{fig9}
\end{figure}
The building has 7-zones. The measured data is shown, for representative winter and summer periods, in Figure \ref{fig8} and Figure \ref{fig9}. For ease of comparing magnitudes of various heat flows, which is a crucial step, a common scale is used in the graphs. Although some data channels are available for each zone or for each air handler, for the purposes of this graph the sum is presented. LEP represents the energy due to lights, equipment and people. Delayed release of this energy, loss of some energy incident radiatively on exterior walls etc. are considered later during analysis. Similarly, measured gas input to the boiler and electrical energy input to the HVAC systems are shown. Efficiency considerations are also done later during the analysis.
From the 7-zone input file, the four audit heat flows $Q_{BLC}(t)$, $Q_{in}(t)$, $Q_{sun}(t)$, $Q_{LEP}(t)$ are computed for each zone as explained before. In these computations, each zone is scheduled to maintain its measured temperature at each hour. The audit heat flows are shown in Figure \ref{fig10} Computed Heat Flows (shown for 2 weeks in winter)and Figure \ref{fig11} for the summer and winter periods respectively. The varying magnitudes of the individual heat flows both diurnally as well as seasonally are noteworthy.
\begin{figure}[htbp]
\caption{Computed Heat Flows (shown for 2 weeks in winter)}
\centering
\includegraphics[width=0.8\textwidth]{10}
\label{fig10}
\end{figure}
\begin{figure}[htbp]
\caption{Some computed heat flows (shown for two summer weeks)}
\centering
\includegraphics[width=0.8\textwidth]{11}
\label{fig11}
\end{figure}
Additionally, ventilation is estimated from each of the fan flows and associated return, outdoor and mixed air temperatures. Infiltration is modeled following \cite{Infil}
The next step of parameter estimation requires great care.
\begin{itemize}
\item
One can attempt to perform energy balance for each zone and estimate parameters at individual zone level. Given the uncertainties in delivered heating/cooling for each zone and interzonal heat transfer and airflow, this was deemed too problematic. We concluded that one should first perform the energy balance by summing over zones, even though certain flows are calculated at zone level.
\item
If we have a known (i.e. measured) heat flow (not just gas input or electrical input to a HVAC system), whose magnitude was large enough for at least for some portion of the data period, the parameter estimation would be more robust. Otherwise, only combinations of parameters can be estimated. A simple example serves to illustrate this: In the Princeton Scorekeeping method (PRISM) analysis [18] for gas heated homes, only the building load coefficient divided by the furnace efficiency can be deduced by fitting the utility bills with variable base degree day data. For our building, delivered heating/cooling energy use was not available. Inspection of the graphs (at least during the plotted period) shows that no suitable period with a large enough known heat flow, even if we were to treat $Q_{LEP}$ as known. (Note: $Q_{LEP}$ is the heat released to the air node after accounting for delayed release, loss through exterior walls etc. It is not equal to $LEP$ input; even the total $Q_{LEP}$ is less than $LEP$ by an amount that depends on the details of radiative couplings as can be seen from the data in the graphs. So $Q_{LEP}$, the actual heat input to the air node, is close to the known electricity consumption by lights and equipment plus heat attributed to people; that is as far as this can be estimated. One can, as an approximation, treat $Q_{LEP}$ as a known heat flow. This is tantamount to accepting that the simulation correctly handles delays and losses of $LEP$. We can then filter on periods when it has large numerical values.)
\item
As proof of concept, we filtered for periods when $Q_{BLC}$ is high and $Q_{LEP}$,$Q_{sun}$,$Q_{in}$ are low, and found 6 contiguous hours from midnight to 6 am on February 15 to be suitable. The heat flows for that period are shown in Figure \ref{fig12}. The energy balance equation for each hour in this period is
\begin{multline}
p_{BLC}Q_{BLC}(t)+p_{BoilerEff}Q_{gas}(t)+\\
Q_{LEP}(t)+Q_{sun}(t)+Q_{in}(t)+Q_{ventilation}(t)+Q_{infiltration}(t)\approx0
\end{multline}
\begin{figure}[htbp]
\caption{Heat flows, gas, and electricity use for a 6-hour period}
\centering
\includegraphics[width=0.8\textwidth]{12}
\label{fig12}
\end{figure}
Because of the relatively low magnitude of $Q_{LEP}$, $Q_{sun}$, $Q_{in}$ in the early morning hours, they are secondary terms and we are justified in not associating parameters with them. We can thereby determine a relationship between $p_{BLC}$, $p_{BoilerEff}$. Knowing one parameter would allow the other to be determined. As we vary $p_{BLC}$, the corresponding value of $p_{BoilerEff}$ can be determined by least squares fit. The resulting relationship is shown in Figure \ref{fig13}
\begin{figure}[htbp]
\caption{Boiler Efficiency as a function of $p_{BLC}$}
\centering
\includegraphics[width=0.8\textwidth]{13}
\label{fig13}
\end{figure}
\end{itemize}
To summarize, the dominant heat flow is $p_{BLC}Q_{BLC}$ for the time period selected. The terms $Q_{LEP}$, $Q_{sun}$, $Q_{in}$ are small and contributed to improve the accuracy of $p_{BLC}$ estimate. These heat flows constitute a net heat loss that is compensated by the heat introduced by the boiler with the measured gas input. For any given boiler efficiency (with default part-load and off-rated curves) we can determine the heat introduced by the boiler. This relationship is shown in Figure \ref{fig13}. If we assume the boiler efficiency to be in the 80\% to 90\% range, $p_{BLC}$, will be in the 1.2 to 1.3 range, implying that the envelope is 20\% to 30\% more lossy than implied by the EnergyPlus input file.
The subsequent steps, involving introduction of the usual additional transfer function parameters and training a neural net for the residuals, required data that was not available to us.
\section{Summary and Future Work}
The Enhanced parameter Estimation (EPE) methodology described in this paper is significantly different from conventional approaches to calibration. It allows one to systematically and objectively reconcile building energy simulations with actual performance data. In summary, EPE consists of the following steps:
\begin{itemize}
\item Various sources of zone level heat flows are defined and computed using a set of specially formulated EnergyPlus simulation runs
\item Least Squares fit to the model of the energy balance is accomplished by introducing and estimating parameters with physical significance. This aspect has been demonstrated in both synthetic and real buildings
\item A two-stage calibration method is developed: (i) for analyzing the building shell, the HVAC system is replaced by an ideal system; and (ii) for analyzing the HVAC system, the building shell is replaced by a box with only process loads.
\item Improved model prediction accuracy is achieved by training a neural net to the time series of residuals remaining after the least squares parameter estimation.
\item Inputs to the neural net are computed heat flows that automatically incorporate the history of temperatures, solar radiation, internal gains, etc., thus simplifying a time series problem into one in which each hour is independent as far as the neural net is concerned.
\end{itemize}
Future work is planned to address (a) incorporation of the parameters into EnergyPlus simulation with thermostatic constraints, (b) refinements to multizone buildings, (c) automatic determination of data windows in which different parameters are best elicited through classification methods, (d) advanced HVAC parameter estimation (e) refining the methodology with data from additional actual buildings.
\section{Acknowledgments}
The authors gratefully acknowledge an SBIR grant from the Department of Energy, with Dr. Amir Roth as Program Manager. Dr James Freihaut and Mr. Scott Wagner of Pennsylvania State University kindly provided building data and a simulation model for the actual building used to illustrate/evaluate the EPE methodology.
\printbibliography
\end{document}
|
\section{Introduction}
Convolutional Neural Networks (CNN) have become a widely used architecture since they have proven to be an accurate universal function estimator in many domains, such as Image Classification and Natural Language Processing \cite{CNN-lecun}. As a particular type of Neural Network (NN), a CNN consists of a collection of parameters, which represent weights and biases, that are tuned depending on the problem that is faced. These parameters, put together after long training sessions, have proven to be able to solve complex problems even outperforming humans \cite{GO-deepmind}. The tuning is performed automatically and the values that the parameters take are outside of the human comprehension, turning the models into black boxes.
The use of CNNs in custom applications demands high computational power, since there is no physical limitation to the complexity of the designed models. Simple augmentation of the size of the collection of trainable parameters has proven to work as long as new data keeps being fed to the model \cite{hestness2017deep}. Furthermore, following the exponential evolution of the availability of computational power, the tendency of solving complex tasks with complex models is increasing \cite{anthes2017artificial}. Language models are good examples of this, e.g. Mann et al. with the description of a NN with 175 billion parameters \cite{gpt3}.
According to this tendency, the most common deployment platforms used historically for DL-based software are servers that enable cloud computing. For the purpose of making use of a trained NN, a data pipeline is developed manually so that it links the pre-processing steps, computation of results, post-processing and eventually the display of the predictions. When all these steps are implemented in a server/cloud, the requirement of an internet connection for using the model is a must. However, with the growth of the mobile applications market, approaches on the integration of NNs in lightweight devices are becoming popular given their ability to provide great services to the user. Within this setting, DL-based software consists of software-reliant systems that \textbf{include data and components} that implement algorithms mimicking learning and problem solving \cite{arpteg2018software}.
In this work, we aim to perform an exploratory and descriptive analysis on: {\it (i)} what are the current challenges regarding the deployment of DL-based software as mobile applications; and {\it (ii)} how can complexity be controlled while keeping a good level of accuracy. The subject is analysed by means of a practical study in which we perform all the required cycles from data acquisition, DL modelling, classification and application, and operation in real context. In this way, we study the relation between the DL-based software development and the accuracy when operating with it in the production settings \cite{9238323}. Ensuring the generalization of the DL-based software in the operation environment is key and can be especially concerning in safety-critical applications \cite{martinez-fernandez}.
Our work is performed over the German Traffic Sign Recognition Benchmark dataset \cite{Benchmarking-tsr}, created in order to standardize the traffic sign recognition literature. In the DL literature, this faced task is an instance of the Image Classification problem \cite{doi:10.1080/01431160600746456}. Each image in the data is from one and only one class so the models will be trained using classification accuracy as the success criteria. CNNs are the standard approach to Image Classification problems for their ability to fit in image processing. The CNN architecture has already shown that can achieve very good results in the literature but also for this specific dataset \cite{li2014medical} \cite{Benchmarking-tsr}.
This work has four main contributions:
\begin{itemize}
\item Show experiences of the end-to-end DL lifecycle from a practical study on traffic sign recognition.
\item Discuss the challenges encountered in the deployment of CNNs.
\item Discuss the trade-offs between accuracy and complexity in environments with limited computational power (e.g., mobile applications).
\item An open science package of the developed DL-based software freely available on GitHub, following the "Cookiecutter Data Science" project structure \footnote{https://drivendata.github.io/cookiecutter-data-science/} to foster correctness and reproducibility.
\end{itemize}
The document is structured as follows. In Section 2 we describe Related Work on both the challenges of integrating DL models in mobile applications and the focuses on optimizing the performance trade-off. In Section 3 we describe the Study Goal and Research Questions. In Section 4 we describe the study design and the end-to-end DL software lifecycle. In Section 5 we show the results of practical study that allow to answer the Research Questions. In Section 6 we discuss what is strictly analysed in our study and what specific parts of the whole topic are addressed.
To end with, in Section 7 we draw conclusions from the research performed and motivate future work.
\section{Related Work}
In this section, we respectively describe the challenges of deploying CNNs in DL-based software, and related work on the optimization of the performance trade-off in systems with limited computation power.
\subsection{Challenges in deployment}\label{sec:sota-challenges}
Regarding the recent interests in the deployment of DL-based software, Chen et al. state that mobile applications come with the strongest popularity trend over other platforms like servers/clouds or browsers for integrating DL models \cite{deployment-DL}. They also show that this increasing trend in the deployment of DL models in mobile applications comes with an inverse proportional relation with the knowledge that the community has about the subject. This way, they demonstrate that the deployment of DL-based software becomes the most challenging part in the life cycle of DL, and that this effect is even more critical when the target are mobile devices.
When analysing the challenges of deployment of DL-based software, several studies focus on the quality attributes that are more relevant to this type of software. The importance of each quality attribute varies depending on the application but it is always the case that DL-based software implies taking care of measures that might not be considered when building traditional software systems. Remarkably, DL-based software effectiveness strongly depends on the structure of the data.
Indeed, Pons and Ozkaya clearly state that, compared to software systems that do not integrate DL-based components, the deployment of DL-based software increases the risk in many quality attributes \cite{quality-AI}. They identify Data centricity as the cause that affects the robustness, security, privacy and sustainability of these systems. Additionally, they analyse the methodology for architecting in Artificial Intelligence (AI) engineering. With the focus put on the software-data dependency, Ozkaya makes the difference between the development of software systems and DL-based software in the processes of building and sustaining \cite{engineering-AI}. Also, Lwakatare et al. provide a list of challenges and solutions regarding the life cycle of DL-based software within industry settings. The challenges and solutions are synthesized into four quality attributes: adaptability (e.g. unstable data dependencies and quality problems), scalability (e.g. balancing efficiency and effectiveness in DL workflow), safety (e.g. explainability of DL models) and privacy (e.g. difficult data exploration in private datasets) \cite{large-scale-AI}.
The identified challenges in the above related work that are objects of study in our work are:
\begin{itemize}
\item (C1) \textbf{Frameworks in early stage} of its development to support DL-based mobile applications.
\item (C2) \textbf{Software--Data dependency}.
\item (C3) \textbf{Explainability} of the models.
\item (C4) \textbf{Sustainability} of the models when deployed.
\end{itemize}
\subsection{In the quest for optimizing the performance}
In the pursuit of optimizing the performance of DL-based software, there exist several key points that affect the overall system implementation. Important bottlenecks that involve limitations to the efficiency are found in both the modelling stage and in the usage of different frameworks to deploy the models. On the one hand, recent contributions \cite{mobile-nets} show an optimized CNN architecture for enabling its usage in devices with less energy capacity. This allows to reduce the model's complexity (e.g. storage weight, computing power) while keeping the desired accuracy, hence optimizing the performance. In this case, the increase in the performance is due to the increase in the efficiency.
On the other hand, the availability of different frameworks supporting the implementation of DL-based software in all its stages has a strong influence on the accuracy of the integrated DL models, as shown in \cite{empirical-frameworks}.
Compared to the aforementioned studies, in our work we focus on analysing the challenges when deploying a specific type of NN, namely CNN, into mobile devices. Furthermore, we do so by means of a practical study. We relate the challenges found with the analysis of the performance, defined as the trade-off between two quality attributes: accuracy and complexity. We provide a comparative analysis between different configurations for each of the models in order to gather evidence of the implications of the accuracy and complexity. Moreover, the sustainability of the models once deployed is also studied to enhance the capabilities of the applications during the DL-based software lifecycle. Our work is developed under the PyTorch framework and Android operating system.
\section{Research Questions}
In this Section, we define the Research Questions (RQs), following the GQM guidelines \cite{GQM}.
\begin{itemize}
\item RQ1 - What are the challenges of the creation and integration of complex CNNs in DL-based mobile applications?
\item RQ2 - What criteria are needed to reason about the trade-off between accuracy and complexity of DL models in mobile applications?
\end{itemize}
The motivation of RQ1 within this project is to study the \textbf{challenges of the creation, training and integration of a CNN in a mobile application}, while identifying, exploring and documenting as many challenges as possible. The obtained conclusions aim to be generic for anyone whose desire is deploying complex models in quotidian software applications.
Furthermore, RQ2 motivates a \textbf{comparative analysis between different solutions to the problem of optimizing the performance trade-off}, since balancing the accuracy of the models and their complexity can be studied from different viewpoints (e.g. design of optimized operations or limitation of architecture's complexity).
\section{Study design}\label{study-design}
In this section we describe the proposed pipeline for achieving a functional implementation of the traffic sign recognition mobile application.
There are four main cycles in the DL-based software lifecycle \cite{9238323}. First, the \textit{Data Management} which consists of the collection and processing of raw data. Second, the \textit{DL Modelling} which consists of adjusting different models to obtain the best-performing possible solution. Third, the \textit{Development} of the environment (mobile application in our work) and the integration of the DL model in it. Fourth, the \textit{DL-based System Operation} with the application that integrates the DL model. This last phase enables sustainability of the model and experiments the performance trade-off within the production settings.
A global overview of the study design and DL-based software lifecycle is shown in Figure \ref{global_overview}. The goals required over the system are shown in the following list:
\begin{enumerate}
\item Use of the device integrated camera
\item Real-time response for a call to the model
\item High accuracy in traffic sign recognition
\item Data augmentation by the community
\item Integration of updated models
\end{enumerate}
We implement these goals through iterations along the aforementioned DL-based software lifecycle phases. We also provide a solution for the \textit{Send Data} operation in Figure \ref{global_overview}. This allows the possibility that the user collects the generated data at run-time during the DL-based software system operation, which is then manually annotated and used for re-training the models on the performant computer. Then, this updated models can be integrated again for inference in the mobile application.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{static/GlobalProject.png}
\caption{Software development lifecycle in our DL-based software}
\label{global_overview}
\end{figure*}
\subsection{Platforms and technologies used}
For joining the DL modelling and the development of DL-based software, there exist recent DL frameworks, all of them under continuous development. These frameworks offer up to two different approaches to model transferring operations. On the one hand, \emph{model conversion}. CoreML for iOS software and TensorFlow-Light for Android arguably are the most famous examples that offer this capability. On the other hand, \emph{model export} using the Open Neural Network Exchange (.ONNX) file format, which was presented in 2017 by Microsoft and Facebook. This file format allows to store serialized versions of trained networks in compressed files.
Since the .ONNX file format is widely used in the current state of the art, it is defined as a key part part of the proposed project pipeline for the practical study.
With all this, the focus is put on finding two different platforms that allow development-sided and operation-sided implementation of both the DL-component and the application that integrates it. The development-sided platform has to allow training and writing of a CNN into a .ONNX file. The operation-sided one has to be capable of reading and using the model and also of supporting the use of the camera in a mobile device. In this work, the development-sided platform is set to be PyTorch and the operation-sided platform is set to be Unity. We make use of the Unity Barracuda libraries for connecting the two platforms. Also, as observed in Figure \ref{global_overview} the models built on PyTorch are trained on the Kaggle GPU, which is available freely for 40 hours a week.
Moreover, it is also seen in Figure \ref{global_overview} that only two of the five models that are trained are later deployed to the application. For an appropriate analysis of the performance of these two models when deployed, we built two applications that only vary in the integrated models.
\subsection{The Dataset}
The GTSRB traffic sign data consists of 144,769 labelled images of 2,416 traffic sign instances of 70 classes belonging to a portion of the real traffic signs that can be found in the roads of Germany. However, after following several criteria for guaranteeing the quality of data, explained in \cite{Benchmarking-tsr}, the final dataset consists of a collection of 51,840 images from 43 classes fulfilling: {\it (i)} images are of sizes from 15×15 to 222×193 pixels, stored in PPM format; {\it (ii)} it contains the definitions of the region of interest with a separating margin of the size equal to the 10\% of the pixels; {\it (iii)} it records the temporal order of the creation of the images; {\it (iv)} it provides post-processed features like HOG descriptors, Haar-like features and Color Histograms. No usage of the post-processed features is made for the purpose of the deployment of a full-stack convolutional-based architecture for solving the image classification problem.
\subsubsection{The Pre-processing}
We applied several pre-processing steps that can be considered critical in the context of image processing model training: {\it (i)} the images are resized to a fixed and squared size; {\it (ii)} a center crop is applied to the images; {\it (iii)} The images are encoded as tensors of the form \textit{[B, C, H, W]}, where \textit{B} is the batch size, \textit{C} is the number of channels, \textit{H} is height and \textit{W} is width. The values of these tensors are normalized into the 0-1 range; {\it (iv)} the tensors are standardized in the \textit{H}x\textit{W} channels with the pre-computed means and standard deviations of the 3 RGB channels of the images in the dataset.
\subsubsection{The Augmentation}
For the purpose of supporting a comparative analysis of the performance of different models, we implemented two variants for applying the technique of data augmentation to the pre-processed dataset: {\it (i)} rotation in 90º degrees plus the vertical and horizontal flips to the images; {\it (ii)} addition of manually tagged images by the users of the application during the system operation phase.
\subsection{The DL/CNN models}
The models that are applied for solving the traffic sign recognition task are the standard CNN, the MobileNet v2 (MBv2) \cite{mobile-nets} and the ResNet34 (RN34) \cite{resnet}. For the first one, we test and evaluate different configurations in order to choose an optimal one that balances accuracy and complexity. For the last two, we test and compare two different ways of training the model: pre-training with feature extraction and fine-tuning, and full-training from scratch. The CNN works with standard convolutional blocks that consist of convolutions followed by batch normalization, an activation function, a max-pooling layer and a dropout layer. It can learn representations of images from scratch which are then passed to a classifier that is built on top of the convolutional blocks. The MBv2 is similar but works with separable depthwise convolutions. It is a form of factorized convolutions which factorize a standard convolution into a depthwise one and a 1x1 one called a pointwise convolution. The depthwise convolution applies a single filter to each input channel and the pointwise convolution then applies a 1x1 convolution to combine the outputs of the depthwise convolution \cite{mobile-nets}. Finally, the RN34 is a more complex CNN that has been successfully applied in the image classification domain and has become a benchmark model for related tasks in the DL literature \cite{resnet}. Both the MBv2 and the RN34 used in the practical study are downloaded from PyTorch and are pre-trained on ImageNet data \cite{deng2009imagenet}.
The results on the CNN models are shown in Table \ref{tab:cnn-results}, where we compare the number of Convolutional Layers, the sizes of the two implemented Fully Connected layers (FC1 and FC2), the total number of parameters, and also the accuracy achieved in the test set with 10 epochs of training. Furthermore, we write down if the augmented dataset is used, the fixed image sizes chosen and the model weight in MegaBytes (MB). Note that the input size of FC1 is equal to the number of output features of the convolutional blocks. Also note that the output size of FC1 is the input size of FC2 and the output size of FC2 is 43, which is the number of unique classes of traffic signs in the data. The resulting optimal configuration from Table \ref{tab:cnn-results} is promising because of the little complexity, the small storage weight and the great performance shown.
\begin{table*}[t]
\begin{center}
\caption{\label{tab:cnn-results}Performance analysis of the different CNN configurations.}
\scalebox{1}{%
\begin{tabular}{|c|c|r|c|c|c|r|r|r|}
\hline
\textbf{Data Aug.} & \textbf{Image Size} & \textbf{Batch Size} & \textbf{Conv. Layers} & \textbf{FC1 input} & \textbf{FC2 input}
& \textbf{Num. Parameters} & \textbf{Test accuracy} & \textbf{Model Weight} \\
\hline
NO & 256x256 & 256 & 5 & 8x8x256 & 4x4x64 & 17.21M & 88.02\% & 65.68MB \\
NO & 512x512 & 64 & 6 & 8x8x512 & 4x4x128 & 67.13M & 95.00\% & 259.70MB \\
YES & 256x256 & 64 & 5 & 8x8x256 & 4x4x64 & 17.21M & 89.31\% & 65.68MB \\
YES & 512x512 & 64 & 6 & 8x8x512 & 4x4x128 & 67.13M & \textbf{95.52\%} & 259.70MB \\
NO & 256x256 & 64 & 6 & 4x4x512 & 2x2x128 & \textbf{5.10M} & \textbf{95.00\%} & \textbf{22.11MB} \\
\hline
\end{tabular}}
\end{center}
\end{table*}
As it can be seen in Table \ref{tab:cnn-results}, no increase in performance appears when the augmented set of data is used. This might be because very fast training is happening, thanks to the non-saturating ReLU function in combination of the convolutional operations plus the use of minibatches. More importantly, it can be seen that the highest tier classification accuracy can be achieved with the lightest of the tested configurations.
Regarding the MBv2, the experiments do not change its architecture but only its training methodology. Three techniques are distinguished: Feature Extraction (FE), which uses a pre-trained version of the model and just tunes the built classifier on top of it, training its weights and biases. Fine-Tuning (FT), which also uses the pre-trained version of the model but tunes the weights of all its layers. And Training from Scratch (TfS), which loads the architecture with a random initialization and fully trains it. The results on the training stage with the MBv2 for 10 epochs are presented in Table \ref{tab:mobilenets-test}.
\begin{table}[h!]
\centering
\caption{Performance analysis of the Mobile Net v2}
\label{tab:mobilenets-test}
\scalebox{1}{%
\begin{tabular}{|c|c|r|}
\hline
\textbf{Data Aug.} & \textbf{Training method} &
\textbf{Test accuracy} \\
\hline
NO & FE & 4.00\% \\
NO & FT & 4.00\% \\
NO & TfS & 4.00\% \\
YES & FE & 68.00\% \\
YES & FT & \textbf{95.20\%} \\
YES & TfS & 92.50\% \\
\hline
\end{tabular}
}
\end{table}
An important drawback of the MBv2 is spotted in Table \ref{tab:mobilenets-test}, where it can be seen that without the augmented set of data the network is not capable of learning on the traffic sign recognition task. Anyway, when the augmented set of data is used it can be seen that the same performance as the best CNN is achieved with a much smaller number of parameters, showing that this architecture performs very efficiently.
Finally, the results on the benchmark RN34 is shown in Table \ref{tab:resnet-test}. Training experiments follow the same logic as with the MBv2 but without trying the augmented set of data, since it is rapidly seen that it is not needed for this architecture.
\begin{table}[h!]
\centering
\caption{Performance analysis of the RN34}\label{tab:resnet-test}
\scalebox{1}{%
\begin{tabular}{|c|c|r|}
\hline
\textbf{Data Aug.} & \textbf{Training method} & \textbf{Test accuracy} \\
\hline
NO & FE & 20.00\% \\
NO & FT & \textbf{97.80\%} \\
NO & TfS & 94.10\% \\
\hline
\end{tabular}
}
\end{table}
As it is seen in Table \ref{tab:resnet-test} the training method of FT gets to the highest seen accuracy in the test split of the training set during the practical study. The output model of this training stage is the one deployed in the application.
\subsection{The DL-based software}
The architecture of the DL-based software system operation is shown in Figure \ref{app_DL}, following design patterns from \cite{washizaki2020machine}. We build two applications that have the same architecture but only differ in the DL model they integrate. The application that loads the SmallCNN has a total weight of 68.84MB and the one that loads the RN34 of 131MB. The baseline weight of the application bundle generated by Unity3D is of 46.73MB without the models.
The two built applications implement the goals 1,2 and 3 defined in Section \ref{study-design}. However, for ensuring the sustainability of the models when deployed, and for satisfying the applications goals 4 and 5, an external server/computer is needed for performing the re-training of the model with the new tagged data added by the community. This implementation still lets the user obtain recognition predictions without external calls. The user has this option because the application client will only make use of the server/computer for downloading new updates on the model and for manually submit new tagged data in a personally controlled way, whenever it is desired. We consider allowing the user to annotate the data that it generates at run-time an efficient solution to the sustainability of the model when deployed. This could be done either by verifying correct model recognition predictions or by correcting wrong ones. However, the architecture that we implement in this practical study makes use of the mobile device storage to save the taken pictures, which are then manually annotated and sent to the local machine, where they are fed to the model again in the Kaggle GPU. This is not a scalable implementation but releasable for research purposes.
\subsection{Threats to validity}
Regarding construct validit
, we build two DL models following a small CNN and ResNet34 (RN34) \cite{resnet}, in order to mitigate mono-operation bias. As for conclusion validit
, due to the exploratory nature of the study we do not execute statistical tests. Regarding internal validit
, the DL models are executed in the same context (e.g., computational power, mobile technologies), mitigating that the efficiency results are caused accidentally rather than by the DL models themselves. Finally, regarding external validit
, our results are tied to the context of traffic sign recognition in mobile applications. Furthermore, the analysis of the performance in the operation-sided environment is carried out in a mobile device with the Qualcomm Adreno 618 GPU running over Android, which determines the extent to which the results obtained that regard the complexity can be generalized.
\begin{figure*}[h]
\centering
\scalebox{0.65}{%
\includegraphics[width=\linewidth]{static/sysops.png}
}
\caption{DL-based Software System Operation Architecture}
\label{app_DL}
\end{figure*}
\section{Results}
In this section we present and discuss the results of our practical study for each RQ.
\subsection{RQ1: Challenges of the creation and integration of complex CNNs in DL-based mobile applications}
For answering the RQ1, we analyze both the identified challenges in the state-of-the-art (see Section \ref{sec:sota-challenges}) and our results. This analysis is shown in Table \ref{results-challenges}, where: \textbf{v} indicates that the challenge has been verified to be real and concerning; \textbf{±} indicates that the challenge has been identified but not faced in the practical study; \textbf{new} defines newly spotted challenges up to our knowledge emerging from our study.
\begin{table}[h]
\begin{center}
\caption{\label{results-challenges}Diagnoses of identified challenges}
\scalebox{1.15}{
\begin{tabular}{|c|c|}
\hline
\textbf{Challenges} & \textbf{Diagnostic}\\
\hline
Frameworks in early stage (C1) & v \\
Software-Data dependency (C2) & v \\
Explainability (C3) & ± \\
Sustainability (C4) & v \\
Software dependencies (C5) & new \\
Model performance (C6) & new \\
\hline
\end{tabular}
}
\end{center}
\end{table}
We verify C1 because we find few alternatives to develop DL-based mobile applications. Furthermore, we encounter difficulties in the support of DL architectures in these frameworks. In our work we experience limitations in the ONNX file format and in the Unity Barracuda package, so we motivate further development of these. From C1 we uncover C5, stating that \textbf{we see a too marked separating line between the development-sided frameworks and the operation-sided ones}. C5 reveals a constraint we found in DL-based software defined as follows. \textbf{No matter what framework one wishes to use, it will always need to wrap at least two different technologies}: one for the data management and model training outside the device, and the other for the model operation inside the user’s device. This implies challenges regarding the mutual support between the two technologies. In the following paragraph, the faced challenges regarding C5 are listed.
First, \textbf{the pre-processing steps that are applied to both the initial dataset in the development phase and the ones in the operation phase (which are applied in real time to the images taken by the user) have to be the same despite being implemented in different frameworks}. DL-based software developers shall ensure that the inputs on both platforms are the same to the network, so it gives the same predictions for the same images.
Secondly, \textbf{the operation-sided framework must support model importing and must provide functional libraries that enable the appropriate usage of the model in its exported format}. In Unity Barracuda, a challenging limitation found is the inability to support a lot of the basic operations in the treatment of DL models (a reason for which we verify C1). For example, let's consider the \textit{view()} function from PyTorch. For the hand-made CNNs, this operation has been replaced by \textit{Flatten()} which does exactly the same, but for the pre-trained MBv2 this method is used and hence the architecture cannot be deployed directly to Unity. In conclusion, due to the limitations in the used operation-sided framework, a model architecture that is built for ensuring efficiency on mobile devices cannot be deployed to the mobile application.
Regarding C2, \textbf{we have experienced high volatility in the applications' effectiveness when testing them with real-world data outside of the training set}. We verify C2 because we experience that the DL-based mobile applications effectiveness strongly depends on the data that is fed to it. This fact makes sustainability (C4) \textbf{become a critical quality attribute for reducing the data dependency and ensuring high performance of the models when deployed}. In the practical study, we implement a solution for the challenge of ensuring sustainability in DL-based software that consists of real world context data augmentation. By doing this, we smooth the differences between the development and operation phases of the DL-based software lifecycle by adapting the models to the real world environment. Hence, we also smooth the dependency between the applications' effectiveness and the data that is fed to it, which makes the DL-based mobile applications more sustainable and makes them generalize better. See the performance of the two deployed models in the two iterations of the operation phase in Table \ref{tab:operation-performance} where it is shown how the data augmentation technique applied in Iteration 2 reduces the data-software dependency (C2) and hence increases the sustainability (C4) of the applications.
\begin{table}[ht]
\centering
\caption{Performance of the deployed models when receiving operation-sided data}
\label{tab:operation-performance}
\scalebox{1}{%
\begin{tabular}{|c|c|r|r|}
\hline
\textbf{Model} & \textbf{Iteration} &
\textbf{Accuracy} & \textbf{Storage Weight} \\
\hline
Small CNN & 1 & 45.07\% & 22.11MB \\
RN34 & 1 & 64.78\% & 83.24MB \\
Small CNN & 2 & \textbf{92.95\%} & 22.11MB \\
RN34 & 2 & \textbf{98.59\%} & 83.24MB \\
\hline
\end{tabular}
}
\end{table}
Regarding C3, our approach for providing explainability in the built applications is based on providing standard confidence indicators of the models' predictions. These indicators consist of the output probabilities of class membership which have been scaled to become percentages of confidence for a given input. \textbf{These simple indicators provide a level of explainability of the models} that is acceptable for the applications in the practical study, meaning that \textbf{they allow some level of interpretation to the user}. We do not consider ensuring this quality attribute a critical challenge because it might not generalize to lots of applications that integrate a DL-based component. However, \textbf{we recognize the importance of the presence of any measure that provides the user with information about why the model has given a result}.
Our practical study uncovered C6. Compared to the usage of a server, the model exporting operation and its integration into the mobile application \textbf{has the drawback of the inclusion of the model's storage weight in the application, hence, in the mobile storage}. For this reason the \textbf{optimization of the performance trade-off is a must when deploying DL-based software to mobile devices}. We have shown in the practical study that not limiting the complexity of the model's architecture, as in the case of the RN34, yields a cost in performance in the mobile application in terms of real time response and storage weight. For the deployment of DL-based software to mobile devices, ensuring efficiency is not an option, it is a must, and facing it becomes a critical challenge.
\subsection{RQ2: Criteria needed to reason about the trade-off between accuracy and complexity of DL models in mobile applications}
Most of the criteria applied to optimize the performance trade-off are approaches based on the reduction of the models' complexity. In an environment where the developer has few alternatives in the choice of both development-sided and operation-sided frameworks, the deployment of these systems becomes a bottleneck of the DL pipeline. This way, \textbf{the evolution of the performance of DL-based software is strictly linked to the evolution of the current model conversion and exporting technologies}, a field that is in its very early stages, where few file formats can be used and with several limitations. For this reason, the deployment of DL-based software to the server cloud and the remote usage through internet connections is still a more flexible choice in terms of DL-based services that can be provided to a mobile user. However, this is not fault-tolerant, due to the dependency on hosting and internet connections.
\textbf{A fact that makes the deployment of DL-based software to mobile devices not always suitable is that it collides with the benefits of transfer learning}. The major capabilities of transfer learning include great results with little task-specific data, the ease in fine-tuning pre-built architectures and omitting the model design stage. The common available models for transferring learning are of massive weight, and the revolution of transfer learning has created a motivation to use these largest-scaled models for solving any task before designing task-specific architectures \cite{transfer-learn}. This has never been a problem until now, when one wants the massive model to be deployed in an environment with limited computational power.
We show in Table \ref{tab:operation-performance} that the most famous architecture amongst all the experimented ones, which is the RN34 \cite{resnet}, is the one carrying the biggest drawbacks in terms of storage weight. In this way, the simple CNN architecture with the appropriate configuration outperforms the massive model in terms of performance because of the ability to provide almost the same accuracy when solving the traffic sign recognition task and because of its reduction in the complexity of the architecture.
Finally, the MBv2 looks very promising in this field for its low complexity and high accuracy. Although it has not been deployed for the proposed practical study due to framework incompatibilities, it is concluded that the modifications in its architecture provide the best capabilities for the purpose of deploying DL-based software into environments with limited computational power, in this case the mobile devices.
This way, in order to lean towards a specific architecture for DL models that are going to be deployed in mobile applications we look for the ones that keep low complexity (e.g. fast real-time response, limited number of parameters, reasonable storage weight) and accomplish high accuracy and generalization.
\section{Discussion}
In the following we review our findings and discuss their implications. First, we have verified that the identified challenges that regard the software-data dependency and sustainability of the models deployed in the software are of major concern. To solve them, we have proposed the implementation of a data augmentation technique that adapts the models to the real world environment, hence providing sustainability. We have shown significant increases in the accuracy when sustainability is ensured in the two systems by means of the data augmentation technique (see Table \ref{tab:operation-performance}). This step is represented by the "Send Data" operation in Figure \ref{global_overview}. We have also verified the need of the evolution of frameworks that support the development of DL-based software, which is another challenge that has been identified in related work. Additionally, we have uncovered a new challenge related to the mutual dependencies of these frameworks, which enriches the identification of the framework capabilities challenge. Furthermore, we have provided identification and description of the criteria that can be applied to approach efficiency in the deployment of DL models in mobile applications. Also, we have related the reasoning of the accomplishment of efficiency to a newly identified challenge.
Taking into account the differences between the performance of the two DL-based mobile applications according to our criteria, we now discuss a possible functional and efficient implementation. This implementation could consist of the integration of the two deployed models in the same DL-based mobile application. The default model could be set to the most efficient one, and in case the results given by this were not satisfactory the user would have the option to switch to the most robust and least efficient model.
\section{Conclusions}
In this work we have studied the DL-based software lifecycle in a practical manner. We have provided an analysis of the challenges found when developing this type of software. Specifically, we have put the focus on the deployment of DL-based software in mobile applications. Concretely, we have tested the performance of different CNN architectures under the PyTorch and Android frameworks. With all this, we have experienced, solved and documented many of the previously identified challenges in the related work and also have highlighted new ones. Furthermore, we have analysed the roles of the accuracy and the complexity in the DL-based software performance, relating this quality attribute to a newly identified challenge of the deployment of DL-based software.
This study motivates several key points of future work. First, the search for optimized model architectures that allow more efficient solutions to the deployment of DL-based components in mobile applications. These can both implement more efficient operations in its architecture or have limitations in its complexity when designed. The study also encourages the community to develop alternative solutions to the conversion and export of models together with software technologies that support mobile application development with DL-based components. The latter would reduce the severity of the challenges that are met while implementing this type of software.
Moreover, the study motivates the design of more sophisticated confidence indicators that describe in a more precise way the uncertainty in the models predictions. The uncertainty of a model prediction given the input conditions is a measure that is hard to determine but useful to influence the decisions of autonomous DL-based systems. These can be key in many applications related to safety-critical environments. Also, sophisticated confidence indicators increase the quality of the systems since these enhance the explainability to the user. Last, we expect long evolution in the design of empowered GPUs for mobile devices which can relax the limitations of complexity in the architectures for achieving efficiency within the mobile device settings.
\section{Data Availability}
In this section we provide a demo video of the developed and evolved DL-based software, available at \url{https://www.youtube.com/watch?v=yFsp6kxO5kI}. Additionally, we provide an open source code repository of the whole project in \url{https://github.com/yuyecreus/CNN-in-mobile-device}.
\section*{Acknowledgment}
We thank Lisa J{\"o}ckel for having reviewed our work and for being supportive and useful to enhance it.
The research presented in this paper has been developed in the context of the CBI course at the GCED@FIB.
{\small
\bibliographystyle{IEEE/IEEEtran}
|
\section{Introduction}
The dynamical Casimir effect \cite{moore1970quantum} is the general model encompassing the gravitational analog model of scalar particle creation by a single perfectly reflecting moving mirror \cite{DeWitt:1975ys,Davies:1976hi,Davies:1977yv}. The usual approach to the analog is to assume a prescribed trajectory that fulfills given physical requirements and compute the resulting radiative measures \cite{Good:2016oey,Walker_1982,walker1985particle,Good:2020fjz,Akal:2020twv,Good:2020byh}. A key theoretical success of the dynamical Casimir effect has been the demonstration that accelerated point mirrors disturb the quantum vacuum via a non-zero Bogoliubov transformation and renormalized stress tensor, resulting in principal outputs: particle production, energy flux and entanglement (see e.g. \cite{Bianchi:2014qua, Bianchi:2014vea, Good:2017ddq, Romualdo:2019eur, Good:2019tnf, Cong:2018vqx, Lee:2019adw}). To reconcile the usual divergent stress tensor, point-splitting regularization is used to construct meaningful finite results consistent with particle production (e.g. \cite{Good:2020uff,Stargen:2016euf}). In this prescription, it is found that the particle production and energy flux are a result of the mirror’s acceleration and jerk, respectively \cite{Fabbri,Birrell:1982ix}. The entropy associated with the moving mirror has motivated investigations into thermodynamic puzzles (see e.g. \cite{Davies:1982cn,Helfer:2000fg}) and quantum information issues (see e.g. \cite{Chen:2017lum,Good:2018aer,Giulio}). Efforts are underway to directly\footnote{Superconducting quantum interference device can act as moving mirrors whereas the dynamical Casimir effect can be measured in the case of a Bose-Einstein condensates, see the review \cite{Dodonov:2020eto}.} measure moving mirror radiation \cite{Chen:2020sir,Chen:2015bcg}.
Recently, perfectly reflecting mirror solutions in $(1+1)$-dimensions have been found that demonstrate unexpected resemblances to strong gravitational systems in $(3+1)$-dimensions. Particularly, $(1+1)D$ mirrors could emulate the radiation provided by accelerating boundaries in $(3+1)D$ in terms of particle production and radiated energy. A typical example of the emulated $(3+1)D$ radiation is given by objects undergoing a gravitational collapse into a black hole, leading to Hawking radiation. Currently, analogy between mirrors and well-known spacetimes, e.g.\ Schwarzschild \cite{Good:2016oey}, Reissner-Nordstr\"om \cite{good2020particle}, Kerr \cite{Good:2020fjz}, and de Sitter/AdS have been found \cite{Good:2020byh}.
Even if the apparent issue related to different dimensions seems to occur, this analogy is predictive and shows the goodness of mirrors in $(1+1)D$ with particular trajectories in describing such physical cases.
Non-thermal or quasi-thermal perfectly reflecting solutions closely characterize other well-known curved spacetime end-states, including extremal black holes (asymptotic uniformly accelerated mirrors \cite{Liberati:2000sq,good2020extreme,Good:2020fjz,Rothman:2000mm,Foo:2020bmv}), black hole remnants (asymptotic constant-velocity mirrors \cite{Good:2016atu,Good:2018ell,Good:2018zmx,Myrzakul:2018bhy,Good:2015nja,Good:2016yht}) and complete black hole evaporation (asymptotic zero-velocity mirrors \cite{Walker_1982, Good:2019tnf,GoodMPLA,Good:2017kjr,B,Good:2017ddq,Good:2018aer}). However, it is worth saying that the reduction from a $(3+1)$-dimensional spacetime to a $(1+1)$-dimensional spacetime does not yield a conformally invariant action. For example, starting from a massless $(3+1)$-dimensional theory in the curved background of spherically symmetric Schwarzschild geometry, transversal (angular) momenta will effectively induce a non-zero mass term in the reduced $(1+1)$-dimensional theory. Hence, the $(1+1)D$ spacetime we are going to deal with does not aim to represent the $(3+1)D$ system\footnote{In the mirror framework, one direct and precise connection that holds in both dimensional contexts, $(3+1)D$ and $(1+1)D$, is the \emph{Lorentz invariant power} as demonstrated in Ref. \cite{Zhakenuly:2021pfm}}.
Remarkably, generalizing Bogolyubov coefficients in $(3+1)$-dimensional theory has so far been intractable. Consequently, attempts toward particle production analysis in $(3+1)D$ spacetime is beyond the scope of this paper.
Despite impressive progress over the last half-century \cite{Dodonov:2020eto}, the moving mirror model is still evolving. The extension to realistic conditions for partially transmitting mirrors has had success in generalizing the specialized case of perfectly reflecting mirrors which often posses infrared divergences \cite{Frolov:1999bi,Nicolaevici:1999ga,Nicolaevici_2001,Nicolaevici:2011zza,Fosco:2017jjf,Nicolaevici:2010zz}. Semi-transparent mirrors can also be used to simulate a null-shell collapse to form a black hole and provide new insights in determining the physics of particle production \cite{Calogeracos,Nicolaevici:2009zz}.
Considering semitransparent mirrors, we provide a more realistic case for the dynamical Casimir effect since, in real mirrors, perfect reflection is only an approximation valid for a small range of frequencies. For the black hole-mirror analogy, perfect reflection models the regularity condition at the center of the collapsing ball $r=0$, where $r$ is the radial coordinate in $(3+1)D$ spacetime\footnote{{We remind that the analogy occurs between $(1+1)D$ mirrors and $(3+1)D$ collapsing balls or shells}}. This condition says that the field vanishes at $r=0$ because no field can exist behind $r<0$, as the coordinate itself is defined only for $r\geq 0$. The semitransparency of the mirror stresses out this condition: this may seem unphysical. However, in important and interesting contexts where it becomes impossible to impose regularity, say e.g. 4-dimensional Schwarzschild spacetime with Eddington-Finkelstein coordinates such that $r=0$ is a spacelike curvature singularity (see e.g. Eq. (5.137) of Ref. \cite{Fabbri}), perfect reflection might indeed need to be relaxed. For this reason, there is good physical motivation to study semi-transparent moving mirrors with respect to black hole radiation.
Another intriguing issue is the interplay between mirrors and quantum information theory. A mirror can be seen as a fundamental tool to model a quantum communication channel.
Since the relation between input and output modes through a mirror is linear, it actually realizes a bosonic Gaussian channel. In this perspective, the mirror is however always considered at rest. Only recently quantum channels arising from the reflection of a one-mode bosonic input upon a perfectly reflecting moving mirror have been characterized \cite{Giulio}.
It seems then quite natural to investigate the information transmission capabilities of quantum channels arising in the broader context of semitransparent moving mirrors. This would allow us to also explore the information capabilities across the null shell of a collapsing black hole.
In this paper we investigate the particle production from a semitransparent moving mirror, obtaining analytical expressions of Bogoliubov coefficients. To do this, we consider a very short acceleration period compared with the wavelength of the produced particles, namely we focus on impulsive accelerated semitransparent mirrors. Consequently, we obtain a finite spectrum of the radiated particles. Then, we investigate the transmission of a signal, carried by the field, through a semitransparent moving mirror.
The above mentioned spectrum permits to understand if the mirror motion can improve the quality of the signal transmission, or if it only creates additional noise, compared with the static case.
Actually, we shall realize the field coming from left-past spacetime region, passing through the semitransparent moving mirror and ending up to right-future spacetime region as undergoing the action of a Gaussian quantum channel,
obtaining an average transmission coefficient, $\tau$, and an average number of noisy particles created, $\overline{n}$. For a mirror with a short acceleration period we find $\tau<1$ and $\overline{n}=0$, yielding a beam splitter bosonic channel. Therefore, an exact expression of the classical and quantum capacity is provided. The most interesting property arising from this line of investigation is that, for each frequency of the input signal, $\tau$ is maximized when the final speed of the mirror is equal to a critical speed, which is different from the speed of light.
The paper is organized in the following: in Sec.~\ref{Sec2} we provide the general expressions for Bogoliubov coefficients for semitransparent moving mirrors, assuming the trajectory of the mirror starting from time-like past and ending at time-like future in proper null coordinates. in Sec.~\ref{Sec3} we focus on trajectories which have a finite acceleration period leading to analytic expressions when this period is very small. In Sec.~\ref{Sec4} we show that the transmission of a signal through a semitransparent moving mirror corresponds to the transmission through a bosonic Gaussian quantum channel, following the same procedure used in Ref. \cite{Giulio}. Finally, in Sec. \ref{Sec5} we provide an exact expression for the classical and quantum capacity of the quantum channel created by an impulsive accelerated mirror. Throughout we use natural units, namely $\hbar=c=1$.
\color{black}
\section{Bogoliubov coefficients for moving mirrors with proper null coordinates}\label{Sec2}
In this section we propose a general, thought implicit, expression for the Bogoliubov coefficients relating input (in) and output (out) modes in the presence of a semitransparent moving mirror. These coefficients give information about the spectrum of particles produced by the mirror and
will be at the heart of the communication properties of the mirror.
We work on a $(1+1)D$ spacetime, which can be compactly portrayed through Penrose diagrams, as in Fig.~\ref{penrose}.
As stressed in the introduction, the $(3+1)$-dimensional case is more realistic, even if very harsh to study analytically. Nevertheless, the $(1+1)D$ results provide a suitable matching with theoretical expectations in $(3+1)D$ spacetimes, e.g. recovering the Hawking radiation and/or obtaining the dynamical Casimir effect when the mirror is very large. Consequently, the next results are thought to hold a relevant guidance for $(3+1)D$ spacetimes.
There, $i^-$ and $i^+$ represent time-like past and future infinities, respectively. The null surfaces $\mathcal{J}_{R/L}^\pm$ are instead the boundaries of the Penrose diagram. Since only massless scalar particles will be considered as input and output, the input mode should necessarily come from a past null-like surface $\mathcal{J}_{R/L}^-$, whereas the output mode should end up at a future null-like surface $\mathcal{J}_{R/L}^{\pm}$. To this end, we simply introduce the null coordinates $u=t-x$ and $v=t+x$. The trajectory of a mirror is usually expressed via null coordinates through the function $p(u)\coloneqq v_{mirror}$ and its inverse $f(v)\coloneqq u_{mirror}$, e.g. \cite{Davies:1976hi}. To guarantee the mirror does not evolve faster than light both $p(u)$ and $f(v)$ are increasing monotonic functions.
As anticipated, we only consider a massless scalar field $\Phi$ (since it describes the vast majority of radiation fields) without self-interaction. The Lagrangian density for this field interacting with a static mirror at the position $x_m$ is described by \cite{Calogeracos}:
\begin{equation}
\mathcal{L}=\frac{1}{2}\partial_\mu\Phi\partial^\mu\Phi+\eta\Phi\delta(x-x_m).
\end{equation}
From this, one can obtain the following reflection and transmission amplitudes:
\begin{equation}
r(\omega)=-\frac{i\eta}{\omega+i\eta}\,,\hspace{1 cm} s(\omega)=\frac{\omega}{\omega+i\eta},
\end{equation}
where $\omega$ is the frequency of the reflected/transmitted mode.
The field $\Phi$ can be expanded as:
\begin{equation}
\Phi=\sum_{J=L,R}\int_0^\infty\left(\phi_\omega^Ja_\omega^J+\phi_\omega^{J*}a_\omega^{J\dagger}\right)d\omega,
\end{equation}
where $\phi^R_\omega$ (resp. $\phi^L_\omega$) is the input mode with frequency $\omega$ incoming from the right, i.e. $\mathcal{J}_R^-$ (resp. left, i.e. $\mathcal{J}_L^-$) and $a_\omega^R$ (resp. $a_\omega^L$) is the corresponding annihilation operator. Considering the boundary condition given by the mirror at the position $x=x_m$, the modes $\phi_\omega^R$ and $\phi_\omega^L$ can be written, respectively, as:
\begin{align}\label{semitstaticR}
\phi_{\omega}^{R}(u,v)&=\frac{1}{\sqrt{4\pi|\omega|}}\left(s(\omega)e^{-i\omega v}\theta(u-v+2x_m)\right.\notag\\
&\left.+\left(e^{-i\omega v}+r(\omega)e^{-i\omega u}\right)\theta(v-u-2x_m)\right),\\
\phi_{\omega}^L(u,v)&=\frac{1}{\sqrt{4\pi|\omega|}}\left(s(\omega)e^{-i\omega u}\theta(v-u-2x_m)\right.
\notag\\&\left.+\left(e^{-i\omega u}+r(\omega)e^{-i\omega v}\right)\theta(u-v+2x_m)\right).
\label{semitstaticL}
\end{align}
\begin{figure}
\centering
\includegraphics[scale=0.5]{Figure1.pdf}
\caption{Penrose diagram showing the trajectories of high acceleration mirrors for different values of the parameter $\nu$, defined in Sec.~\ref{Sec3}. The trajectory is like the one described by Eq.~\eqref{pacc} with $u_0$ very small. In particular, $\nu=1.5$ for the green dashed line, $\nu=2$ for the green line, $\nu=4$ for the black dashed line and $\nu=100$ for the black line. We can imagine the infinitesimal acceleration period to be in a neighborhood of $t=0$.}
\label{penrose}
\end{figure}
The expressions for the input modes \eqref{semitstaticR} and \eqref{semitstaticL} are valid when the mirror is static. When the mirror moves along a trajectory $x_m(t)$ also the amplitudes $r(\omega)$ and $s(\omega)$ change in time (see Appendix A of \cite{Nicolaevici:2009zz}). Hence, there is a great mathematical complication, since the boundary condition between the left side and right side of the mirror becomes time-dependent. To overcome this problem, we resort to the strategy used in Ref.~\cite{C}. Namely, we put ourselves in the mirror frame, using proper coordinates. The proper distance from the mirror is indicated by $\rho$ and the proper time by $\tau$. From them, the proper null coordinates are defined as $\overline{u}\coloneqq\tau-\rho$ and $\overline{v}=\tau+\rho$. The proper null coordinates $\overline{u}$ and $\overline{v}$ could also be written in terms of the external null coordinates $u=t-x$ and $v=t+x$ through $\overline{u}(u)$ and $\overline{v}(v)$ which depend on the trajectory of the mirror (specified by $p(u)$ and $f(v)$). This dependence comes from the metric conservation (see Eq.~(20) of \cite{Obadia:2001hj}):
\begin{equation}
ds^2=d\overline{u}^2=d\overline{v}^2=\partial_up(u)du^2=\partial_vf(v)dv^2,
\end{equation}
from which one obtains:
\begin{equation}
\frac{d\overline{u}(u)}{du}=\sqrt{\partial_up(u)},\hspace{1 cm}\frac{d\overline{v}(v)}{dv}=\sqrt{\partial_vf(v)}.
\end{equation}
Finally, we define the input modes in proper coordinates as $g_\omega^R$ and $g_\omega^L$.
In the mirror frame, the mirror is obviously static. Hence, the modes $g_\omega^R$ and $g_\omega^L$ could be written analogously to the modes $\phi_\omega^R$ and $\phi_\omega^L$ in the static case, i.e. like Eqs.~\eqref{semitstaticR} and \eqref{semitstaticL}:
\begin{align}\label{propermodeR}
&g_{\omega}^{R}(\overline{u},\overline{v})=\frac{1}{\sqrt{4\pi|\omega|}}\notag\\
&\times\left(s(\omega)e^{-i\omega\overline{v}}\theta(\overline{u}-\overline{v})+\left(e^{-i\omega\overline{v}}+r(\omega)e^{-i\omega\overline{u}}\right)\theta(\overline{v}-\overline{u})\right),\\
&g_{\omega}^{L}(\overline{u},\overline{v})=\frac{1}{\sqrt{4\pi|\omega|}}\notag\\
&\times\left(s(\omega)e^{-i\omega\overline{u}}\theta(\overline{v}-\overline{u})+\left(e^{-i\omega\overline{u}}+r(\omega)e^{-i\omega\overline{v}}\right)\theta(\overline{u}-\overline{v})\right).
\label{propermodeL}
\end{align}
From now on, we consider only time-like trajectories for the mirror, i.e. we consider trajectories starting at $i^-$ and ending up at $i^+$, referring to the Penrose diagram in Fig.~\ref{penrose}. In that case, both the modes in external coordinates $\{\phi_\omega^R\}_v$ and the ones in proper coordinates $\{g_\omega^R\}_{\overline{v}}$ form a complete set of input modes incoming from $\mathcal{J}_{R}^{-}$. Analogously, both the sets $\{\phi_\omega^L\}_u$ and $\{g_\omega^L\}_{\overline{u}}$ are complete sets of modes incoming from $\mathcal{J}_L^{-}$. As a consequence, the modes in external coordinates $\phi_\omega^J$ are related to the ones in proper coordinates through the following Bogoliubov transformation \cite{C}:
\begin{equation}\label{inmodes}
\phi^{J}_{\omega}=\int_{-\infty}^{+\infty}\chi(\omega')\left(\phi_{\omega}^{J},g_{\omega'}^{J}\right)g_{\omega'}^{J}d\omega',
\end{equation}
where $\chi$ is the sign function.
Using Eq.~\eqref{inmodes} one may obtain a general expression for the input modes in external coordinates $\phi_\omega^J$. Let us turn our attention to the scalar product $\left(\phi_{\omega}^{R},g_{\omega'}^{R}\right)$. To single out a convenient integration surface for the integration, we select $\mathcal{J}_R^-$, since here $\phi_{\omega}^{R}=\frac{1}{\sqrt{4\pi|\omega|}}e^{-i\omega v}$ and $g_\omega^R=\frac{1}{\sqrt{4\pi|\omega|}}e^{-i\omega \overline{v}(v)}$. So, the scalar product becomes
\begin{equation}\label{scprR}
\left(\phi_\omega^R,g_{\omega'}^R\right)=-i\int_{-\infty}^{+\infty}\left((\partial_vg_{\omega'}^{R*})\phi_{\omega}^{R}-g_{\omega'}^{R*}\partial_v\phi_{\omega}^R\right)dv.
\end{equation}
Using the fact that these modes vanish for $v\rightarrow\pm\infty$ we can integrate Eq.~\eqref{scprR} by parts, simplifying the scalar product to
\begin{align}
\left(\phi_{\omega}^{R},g_{\omega'}^{R}\right)&=2i\int_{-\infty}^{+\infty}g_{\omega'}^{R}\partial_v\phi_{\omega}^Rdv\notag\\
&=\frac{\omega}{2\pi\sqrt{|\omega||\omega'|}}\int_{-\infty}^{+\infty}e^{-i\left(\omega v-\omega'\overline{v}(v)\right)}dv.
\end{align}
The same thing is done for the scalar product $\left(g_{\omega'}^{L},\phi_{\omega}^{L}\right)$ integrating on the surface $\mathcal{J}_L^-$ and obtaining
\begin{align}
\left(\phi_{\omega}^{L},g_{\omega'}^{L}\right)&=2i\int_{-\infty}^{+\infty}g_{\omega'}^{L*}\partial_v\phi_{\omega}^Ldv\notag\\
&=\frac{\omega}{2\pi\sqrt{|\omega||\omega'|}}\int_{-\infty}^{+\infty}e^{-i\left(\omega u-\omega'\overline{u}(u)\right)}du.
\end{align}
From now on, we also consider a detector positioned on the right of the mirror. Hence, the terms of $g_\omega^J$ (from Eqs.~\eqref{propermodeR} and \eqref{propermodeL}) proportional to $\theta(\overline{u}-\overline{v})$ are neglected. For the mode coming from the right of the mirror $\phi_\omega^R$, applying some contour integration in the variable $\omega'$ and using the fact that $\overline{u}(u)$ and $\overline{v}(v)$ are increasing monotonic functions (to guarantee that the mirror speed is not faster than the speed of light), we obtain
\begin{equation}
\phi^{R}_{\omega}=\frac{1}{\sqrt{4\pi\omega}}e^{-i\omega v}+\phi^{refl}_{\omega,R}(u),
\end{equation}
where $\phi^{refl}_{\omega}(u)$ is the reflected wave of the input mode:
\begin{align}
\phi^{refl}_{\omega}(u)&=\sqrt{\frac{\omega}{4\pi}}i\int_{-\infty}^{+\infty}\left(\frac{1}{2}\chi(\overline{u}(u)-\overline{v}(v'))\right.\notag\\
&\left.-\theta(\overline{u}(u)-\overline{v}(v'))e^{-\eta(\overline{u}(u)-\overline{v}(v')}\right)e^{-i\omega v'}dv'.
\end{align}
With a similar calculation, we can obtain $\phi_{\omega}^{L}$ as
\begin{align}
\phi_{\omega}^{L}&=\sqrt{\frac{\omega}{4\pi}}(-i)\int_{-\infty}^{+\infty}\theta(u-u')e^{-\eta\left(\overline{u}(u)-\overline{u}(u')\right)}e^{-i\omega u'}du'\nonumber\\
&=\sqrt{\frac{\omega}{4\pi}}(-i)\int_{-\infty}^{u}e^{-\eta\left(\overline{u}(u)-\overline{u}(u')\right)}e^{-i\omega u'}du'\,.
\end{align}
In the presence of an accelerating boundary, we expect a difference between the input and the output spacetime structure \cite{Hawking:1974sw, Birrell:1982ix,CarlitzWilley}. The output mode (outgoing to the right, i.e. $\mathcal{J}_R^+$) can be written in terms of the input ones through the following Bogoliubov transformation:
\begin{equation}
\phi_\omega^{out}=\sum_{J=R,L}\int_0^\infty\left(\alpha_{\omega\omega'}^{RJ}\phi_{\omega'}^J+\beta_{\omega\omega'}^{RJ}\phi_{\omega'}^{J*}\right)d\omega',
\end{equation}
where $\alpha_{\omega\omega'}^{RR}=\left(\phi_{\omega}^{out},\phi_{\omega'}^{R}\right)$, $\alpha_{\omega\omega'}^{RL}=\left(\phi_{\omega}^{out},\phi_{\omega'}^{L}\right)$, $\beta_{\omega\omega'}^{RR}=\left(\phi_{\omega}^{out*},\phi_{\omega'}^{R}\right)^*$ and $\beta_{\omega\omega'}^{RL}=\left(\phi_{\omega}^{out*},\phi_{\omega'}^{L}\right)^*$ are the Bogoliubov coefficients we want to compute. We are particularly interested in the $\beta$ Bogoliubov coefficients, since they give us information about the production of the particles by the mirror. Since we consider mirrors with a trajectory ending at $i^+$, $\mathcal{J}^+_R$ is a surface we can use as an integration surface for the scalar product of Bogoliubov coefficients. In fact on this surface we have
\begin{subequations}
\begin{align}
\phi_{\omega}^{out*}&\longrightarrow\frac{1}{\sqrt{4\pi\omega}}e^{i\omega u},\\
\phi_{\omega'}^{R}&\longrightarrow\phi^{refl}_{\omega',R},
\end{align}
\end{subequations}
Applying the derivative of the scalar product to the input modes $\phi_\omega^{J}$ and changing the variable from $u$ to $\overline{u}$ (we define $u(\overline{u})$ as the inverse of $\overline{u}(u)$) we obtain the following general expression for the Bogoliubov coefficients:
\pagebreak
\begin{widetext}
\begin{align}\label{betacoeffR}
&\beta_{\omega\omega'}^{RR}=-\frac{\eta}{2\pi}\sqrt{\frac{\omega'}{\omega}}\int_{-\infty}^{+\infty}\int_{-\infty}^{+\infty}\theta(\overline{u}-\overline{v}(v'))e^{\eta\overline{v}(v')-i\omega'v'}e^{-\eta\overline{u}-i\omega u(\overline{u})}dv'd\overline{u},\\
&\beta_{\omega\omega'}^{RL}=-\frac{\eta}{2\pi}\sqrt{\frac{\omega'}{\omega}}\int_{-\infty}^{+\infty}\int_{-\infty}^{+\infty}\theta(\overline{u}-\overline{u}(u'))e^{\eta\overline{u}(u')-i\omega'u'}e^{-\eta\overline{u}-i\omega u(\overline{u})}du'd\overline{u},\label{bogcoeffL}
\end{align}
\end{widetext}
and
\begin{widetext}
\begin{align}\label{alphaR}
&\alpha_{\omega\omega'}^{RR}=-\frac{\eta}{2\pi}\sqrt{\frac{\omega'}{\omega}}\int_{-\infty}^{+\infty}\int_{-\infty}^{+\infty}\theta(\overline{u}-\overline{v}(v'))e^{\eta\overline{v}(v')+i\omega'v'}e^{-\eta\overline{u}-i\omega u(\overline{u})}dv'd\overline{u},\\
&\label{alphaL}\alpha_{\omega\omega'}^{RL}=\sqrt{\frac{\omega'}{\omega}}\delta(\omega-\omega')-\frac{\eta}{2\pi}\sqrt{\frac{\omega'}{\omega}}\int_{-\infty}^{+\infty}\int_{-\infty}^{+\infty}\theta(\overline{u}-\overline{u}(u'))e^{\eta\overline{u}(u')+i\omega'u'}e^{-\eta\overline{u}-i\omega u(\overline{u})}du'd\overline{u}.
\end{align}
\end{widetext}
Considering Eqs. \eqref{betacoeffR} and \eqref{bogcoeffL}, we have that $N_\omega^R=\int_{0}^{+\infty}\left|\beta_{\omega\omega'}^{RR}\right|^2d\omega'$ is the spectrum of the particles produced due the reflection of the modes at the right of the mirror and $N_\omega^L=\int_{0}^{+\infty}\left|\beta_{\omega\omega'}^{RL}\right|^2d\omega'$ is the spectrum of particles produced due the transmission of the modes at the left of the mirror. The total spectrum of produced particles is therefore given by
\begin{equation}
N_\omega=N_\omega^R+N_\omega^L.
\end{equation}
\section{Approximated trajectories with impulsive acceleration} \label{Sec3}
Once we have found a general expression for the Bogoliubov coefficients, as reported in Eqs. \eqref{betacoeffR}, \eqref{bogcoeffL}, \eqref{alphaR} and \eqref{alphaL}, we aim at finding trajectories for which the Bogoliubov coefficients can be explicitly and analytically computed.
In this Section we show that this task can be accomplished for trajectories corresponding to impulsive acceleration.
First, looking at Eqs. \eqref{betacoeffR} and \eqref{bogcoeffL}, we notice that finding explicit Bogoliubov coefficients turns out to be non-analytical and quite hard for some physical aspects that we summarize below. \begin{itemize}[noitemsep]
\item[{\bf 1)}] The trajectories in null comoving coordinates are arguments of exponentials, and the only functions which can be easily analytically computed in an exponential are the linear ones, which corresponds to a non-accelerating mirror, giving trivially the Bogoliubov coefficients $\beta=0$.
\item[{\bf 2)}] In order to find a non-trivial trajectory we have to use integral functions (such as the Euler gamma), but even if we solve one integral with it, we also need to solve the other. This will be an integral of an exponential multiplied by an integral function, which is almost always not computable analytically.
\item[{\bf 3)}] In particular for Eq.~\eqref{bogcoeffL} we need to find a function $\overline{u}(u)$ such that the exponential of it and the exponential of its inverse $u(\overline{u})$ is integrable analytically. Even if such a function exists, it is needed also that Eq.~\eqref{betacoeffR} is analytically computable.
\item[{\bf 4)}] For some trajectories (such as the Carlitz-Willey's and the uniform acceleration trajectory \cite{CarlitzWilley}) we can pass easily from the expressions of the trajectories in null coordinates to $\overline{u}(u)$ and $\overline{v}(v)$. However, for the other non-trivial trajectories which provides an exact solution for the Bogoliubov coefficients in the perfectly reflecting case (Walker-Davies \cite{Walker_1982}, Arctx \cite{B}, Dlogex \cite{Good_2015BirthCry}, Proex \cite{Good:2016yht}, etc.) it is impossible even to find an analytic expression for the trajectories in null coordinates $\overline{u}(u)$ and $\overline{v}(v)$. Reversely, if we succeed finding a function $\overline{u}(u)$, such that we can find an analytic function for the Bogoliubov coefficients, we need to have the respective functions $p(u)$ and $f(v)$ describing the trajectory of a mirror. To this aim, the quantities $p(u)$ and $f(v)$ should be real, without asymptotes, without singularities in their domain, and monotonic in their domain.
\end{itemize}
We thus consider a class of trajectories that will allow us to simplify the treatment.
It consists of trajectories fulfilling the conditions below:
\begin{itemize}[noitemsep]
\item the mirror is static at $x=0$ for $t<0$,
\item it begins to accelerate toward its left along a certain trajectory $z(t)$ at $t=0$ until it reaches the time $t_0$, arriving to the point $x_0$,
\item after a time interval $t_0$, the mirror continues travelling with the same velocity $V$ reached after the acceleration period. Such $V$ must be smaller than $1$, otherwise the proper acceleration of the mirror would become infinite.
\end{itemize}
Such a trajectory can be described though the simplest choice, represented by the following function
\begin{equation}\label{pacc}
p_{tot}(u)=\begin{cases}
u, & \text{if}\;u\le 0,\\
p(u), & \text{if}\;0<u\le u_0,\\
p(u_0)+\nu^{-1}(u-u_0), & \text{if}\;u>u_0.
\end{cases}
\end{equation}
For simplicity, hereafter we refer to the complete trajectories with the subscript ``tot" and with the usual $p(u)$, $f(v)$, $\overline{u}(u)$ and $\overline{v}(v)$ we refer to the trajectory of the mirror only on its period of acceleration. In Eq.~\eqref{pacc} $u_0=t_0-x_0$ quantifies the width of the acceleration period, alongside with $\nu^{-1}:=\partial_u p(u)|_{u_0}$, which ensures the continuity of the derivative on $u_0$\footnote{The continuity of $p_{tot}(u)$ and of its first derivative should be imposed at $u=0$ as well. Hence, the trajectory $p(u)$ should be chosen to satisfy this condition.}. The quantity $\nu$ is related to the final speed of the mirror $V$ through $\nu=\frac{1+V}{1-V}$ and any deviations from the linearity in terms of $\sim u$ would imply that, at asymptotic regimes, the velocity is no longer a constant.
Analogously, we can write the inverse of $p_{tot}(u)$, i.e., $f_{tot}(v)$ by
\begin{equation}\label{facc}
f_{tot}(v)=\begin{cases}
v, & \textrm{if}\;v\le 0,\\
f(v), & \textrm{if}\;0<v\le v_0,\\
f(v_0)+\nu(v-v_0), & \textrm{if}\;v>v_0,
\end{cases}
\end{equation}
where $v_0=p(u_0)$. It is worth noticing that all the quantities derived from $u_0$ ,i.e., $v_0$, $\overline{u}_0=\overline{u}(u_0)$ and $\overline{v}_0=\overline{v}(v_0)$ satisfy the equations describing the trajectories of the mirror ,i.e., $v=p(u)$, $u=f(v)$, $\overline{u}=\overline{v}(v)$ and $\overline{v}=\overline{u}(u)$. As consequence we have: $v_0=p(u_0)$, $u_0=f(v_0)$ and $\overline{u}(u_0)=\overline{v}(v_0)=\overline{u}_0=\overline{v}_0$.\\
\indent In proper coordinates Eq.~\eqref{pacc} and Eq.~\eqref{facc} become:
\begin{equation}\label{trajgenu}
\overline{u}_{tot}(u)=\begin{cases}
u, & \text{if}\;u\le 0,\\
\overline{u}(u), & \text{if}\;0<u\le u_0,\\
\overline{u}(u_0)+\nu^{-1/2}(u-u_0), & \text{if}\;u>u_0,
\end{cases}
\end{equation}
\begin{equation}\label{trajgenv}
\overline{v}_{tot}(v)=\begin{cases}
v, & \text{if}\;v\le 0,\\
\overline{v}(v), & \text{if}\;0<v\le v_0,\\
\overline{v}(v_0)+\sqrt{\nu}(v-v_0), & \text{if}\;v>v_0.
\end{cases}
\end{equation}
We now focus on the $\beta$ Bogoliubov coefficients (in order to obtain the $\alpha$ it is sufficient to switch $\omega'\rightarrow-\omega'$ everywhere except in the external factor $-\frac{\eta}{2\pi}\sqrt{\frac{\omega'}{\omega}}$, and add the term $\sqrt{\frac{\omega'}{\omega}}\delta(\omega-\omega')$ in $\alpha_{\omega\omega'}^{RL}$). For this class of trajectories we obtain, separating the integrals, through the Eq.~\eqref{betacoeffR} and Eq.~\eqref{bogcoeffL} the following
\begin{widetext}
\begin{equation}\label{accrbc}
\begin{split}
\beta_{\omega\omega'}^{RR}=-\frac{\eta}{2\pi}\sqrt{\frac{\omega'}{\omega}}\left[\frac{e^{-i\omega u_0-\eta\overline{u}_0}}{\left(\omega\sqrt{\nu}-i\eta\right)(\omega'+i\eta)}-\frac{1}{(\omega'+i\eta)(\omega+\omega'+i\epsilon)}-\frac{e^{-i\omega u_0-i\omega'v_0}}{\left(\omega\nu+\omega'-i\nu\epsilon\right)\left(\omega\sqrt{\nu}-i\eta\right)}\right.\\\left.+\frac{1}{\eta-i\omega'}\int_{0}^{\overline{u}_0}e^{-\eta\overline{u}-i\omega u(\overline{u})}d\overline{u}+\frac{e^{-i\omega u_0-\eta\overline{u}_0}}{\eta+i\omega\sqrt{\nu}}\int_{0}^{v_0}e^{\eta\overline{v}(v')-i\omega'v'}dv'\right.\\\left.+\int_{0}^{\overline{u}_0}e^{-\eta\overline{u}-i\omega u(\overline{u})}\left(\int_{0}^{v(\overline{u})}e^{\eta\overline{v}(v')-i\omega'v'}dv'\right)d\overline{u}\right];
\end{split}
\end{equation}
\end{widetext}
\begin{widetext}
\begin{equation}\label{acclbg}
\begin{split}
\beta_{\omega\omega'}^{RL}=-\frac{\eta}{2\pi}\sqrt{\frac{\omega'}{\omega}}\left[\frac{e^{-i\omega u_0-\eta\overline{u}_0}}{\left(\omega\sqrt{\nu}-i\eta\right)(\omega'+i\eta)}-\frac{1}{(\omega'+i\eta)(\omega+\omega'+i\epsilon)}-\frac{e^{-i(\omega+\omega')u_0}}{\left(\omega+\omega'-i\epsilon\right)\left(\omega\sqrt{\nu}-i\eta\right)}\right.\\
\left.+\frac{1}{\eta-i\omega'}\int_{0}^{\overline{u}_0}e^{-\eta\overline{u}-i\omega u(\overline{u})}d\overline{u}+\frac{e^{-i\omega u_0-\eta\overline{u}_0}}{\eta+i\sqrt{\nu}\omega}\int_{0}^{u_0}e^{\eta\overline{u}(u')-i\omega'u'}du'\right.\\
\left.+\int_{0}^{\overline{u}_0}e^{-\eta\overline{u}-i\omega u(\overline{u})}\left(\int_{0}^{u(\overline{u})}e^{\eta\overline{u}(u')-i\omega'u'}du'\right)d\overline{u}\right],
\end{split}
\end{equation}
\end{widetext}
where $\epsilon$ is an exponential cutoff in $u$ and $v$, needed in order to make some integrals convergent. For the $\beta$ Bogoliubov coefficients, we can set $\epsilon=0$ without problems. However, for the $\alpha$ Bogoliubov coefficients we cannot neglect $\epsilon$, otherwise a divergence for them occurs for $\omega=\omega'$\footnote{For Eq.~\eqref{alphaL}, considering $\epsilon\ne0$, the Dirac delta can be seen as $\frac{1}{\pi}\frac{\epsilon}{\epsilon^2+(\omega-\omega')^2}$.}.
For computing particle production, we only need the $\beta$ Bogoliubov coefficients. So, limiting to $\epsilon=0$ and looking at Eqs. \eqref{accrbc} and \eqref{acclbg}, we soon notice that we need a strategy to neglect those integrals that are non-analytical.
Since the arguments of all the integrals have no singularities in their integration range, and since for $u_0\rightarrow0$ we have also $v_0\rightarrow0$ and $\overline{u}(u)\rightarrow0$, the first option is to consider $u_0\rightarrow0$, taking the acceleration period so much short to neglect its contribution. Further, with the recipe $u_0\sim0$, we stress that, before this period the mirror was at rest, whereas after it the mirror shows a finite velocity.
Since we are minimizing the period in which the mirror accelerates, we could maximize the acceleration. In this respect, we refer to these mirrors as ``impulsive accelerated mirrors", in which particular physical consequences are expected. In particular, to clarify why we need to maximize the acceleration, let us first consider the well-consolidate Carlitz-Willey trajectory \cite{CarlitzWilley}. Here, we have $\overline{u}(u)=\frac{2}{k}\left(1-e^{-ku/2}\right)$ and $\overline{v}(v)=\frac{2}{k}\left(1-\sqrt{1-kv}\right)$, where $k$ is intimately related to the mirror acceleration.
Even though we cannot fix the acceleration, since it is not constant for the Carlitz-Willey trajectory, we can fix the parameter $k$ to be arbitrarily large enough. Thus, the approximation for impulsive accelerated mirrors consists in setting $u_0\ll1$ and $k\gg1$, leading to the single main assumption $ku_0=\text{const}$. The value of this constant is related to the final speed of the mirror itself. Indeed, the parameter $\nu$ in Eqs.~\eqref{accrbc} and Eq.~\eqref{acclbg} is simply given by $\nu=e^{ku_0}$ for the Carlitz-Willey trajectory. The trajectories of such mirrors are portrayed in Fig. \ref{penrose}.
It is worth noticing that the above described approximation is valid for all those trajectories provided that we can associate to the acceleration of the mirror a constant parameter $k$.
\indent Finally we can apply the approximation to Eqs.~\eqref{accrbc} and ~\eqref{acclbg}, using $k=\ln(\nu)/u_0$ and expanding in series for $u_0\rightarrow0$. For short acceleration periods, only the lowest expansion orders are clearly needful.
To the zeroth order the integrals on Eq.~\eqref{accrbc} and Eq.~\eqref{acclbg} can be completely neglected. In this case, the Bogoliubov coefficients read
\begin{align}
\label{0termR}
\beta_{\omega\omega'}^{RR}&=-\frac{\eta}{2\pi}\sqrt{\omega'\omega}\frac{\sqrt{\nu}\left(\omega'-\omega\sqrt{\nu}\right)\left(\omega\sqrt{\nu}-1\right)+i\eta\left(\nu-1\right)}{\left(\omega\sqrt{\nu}-i\eta\right)\left(\omega'+i\eta\right)\left(\omega\nu+\omega'\right)\left(\omega+\omega'\right)},\\
\beta_{\omega\omega'}^{RL}&=-\frac{\eta}{2\pi}\sqrt{\omega'\omega}\frac{1-\sqrt{\nu}}{\left(\omega\sqrt{\nu}-i\eta\right)\left(\omega'+i\eta\right)\left(\omega+\omega'\right)}.
\label{0termL}
\end{align}
Here, we are neglecting the acceleration period and Eqs.~\eqref{0termR} and~\eqref{0termL} are valid for each trajectory for which we associate a constant parameter $k$ to the acceleration. Both the Bogoliubov coefficients Eqs.~\eqref{0termR} and~\eqref{0termL} have the factor $\sqrt{\omega'\omega}$ at the beginning.
So, their modulus squares show the factor $\omega'\omega$, that is not cancelled by any term in the denominator. This ensures that in these modulus squares no infrared divergences occur\footnote{This is valid for $\beta_{\omega\omega'}^{RR}$, if $\omega\ne0$. If not, we could have possible infrared divergences, leading to unphysical particle production. } for $\omega'\rightarrow0$. Furthermore, for $\omega'\rightarrow+\infty$ Eqs.~\eqref{0termR} and~\eqref{0termL} are asymptotic to $\omega'^{-3/2}$. This means that their modulus squares do not provide any ultraviolet divergence. As a consequence the number of particles produced with this approximation is finite and different from zero for finite values of $\omega$. Finally, checking the case in which the mirror lies at rest, namely $\nu=1$, we immediately notice that the Bogoliubov coefficients are zero, as expected.\\
\indent To the first order expansion, we restrict the trajectories to the ones with non singular $\partial_u\overline{u}(u)$ and $\partial_u\overline{v}(v')$ throughout the range $0\le u\le u_0$. Hence, first order Bogoliubov coefficients are
\begin{widetext}
\begin{equation}\label{1orderR}
\beta_{\omega\omega'}^{RR}\sim\nonumber\\
-\frac{\eta}{2\pi}\sqrt{\omega'\omega}\left\{\frac{\sqrt{\nu}\left(\omega'-\omega\sqrt{\nu}\right)\left(\sqrt{\nu}-1\right)+i\eta\left(\nu-1\right)}{\left(\omega\sqrt{\nu}-i\eta\right)\left(\omega'+i\eta\right)\left(\omega\nu+\omega'\right)\left(\omega+\omega'\right)}+\frac{(\omega+\omega')\left[\eta\left(V_0\nu-1\right)+i\omega\nu\left(U_0\sqrt{\nu}-1\right)-i\omega'\sqrt{\nu}\left(V_0\sqrt{\nu}-U_0\right)\right]}{\left(\omega\sqrt{\nu}-i\eta\right)\left(\omega'+i\eta\right)\left(\omega\nu+\omega'\right)\left(\omega+\omega'\right)}u_0\right\},
\end{equation}
\end{widetext}
\begin{equation}\label{1orderL}
\beta_{\omega\omega'}^{RL}\sim-\frac{\eta}{2\pi}\sqrt{\omega'\omega}\frac{1-\sqrt{\nu}+i(\omega+\omega')\left(U_0\sqrt{\nu}-1\right)u_0}{\left(\omega\sqrt{\nu}-i\eta\right)\left(\omega'+i\eta\right)\left(\omega+\omega'\right)},
\end{equation}
where $U_0\coloneqq\partial_{u_0}\overline{u}_0|_{u_0=0}$ and $V_0\coloneqq\partial_{u_0}v_0|_{u_0=0}$. For the Carlitz Willey trajectory \cite{carlitz1987reflections}, we get
\begin{equation}
U_0=\frac{2}{\ln\nu}\left(1-\frac{1}{\sqrt{\nu}}\right),
\end{equation}
\begin{equation}
V_0=\frac{1}{\ln\nu}\left(1-\frac{1}{\nu}\right).
\end{equation}
In terms of $\nu$, the first order Bogoliubov coefficients for the Carlitz-Willey trajectory become
\begin{widetext}
\begin{equation}\label{betaCWR}
\begin{split}
\beta_{\omega\omega'}^{RR}\sim-\frac{\eta}{2\pi}\sqrt{\omega'\omega}\left\{\frac{\sqrt{\nu}\left(\omega'-\omega\sqrt{\nu}\right)\left(\sqrt{\nu}-1\right)+i\eta\left(\nu-1\right)}{\left(\omega\sqrt{\nu}-i\eta\right)\left(\omega'+i\eta\right)\left(\omega\nu+\omega'\right)\left(\omega+\omega'\right)}\right.\\
\left.+\frac{(\omega+\omega')\left[\eta\left(\frac{1}{\ln(\nu)}(\nu-1)-1\right)+i\omega\nu\left(\frac{2}{\ln(\nu)}\left(\sqrt{\nu}-1\right)-1\right)-i\omega'\frac{\sqrt{\nu}}{\ln(\nu)}\left(\frac{1}{\sqrt{\nu}}+\sqrt{\nu}-2\right)\right]}{\left(\omega\sqrt{\nu}-i\eta\right)\left(\omega'+i\eta\right)\left(\omega\nu+\omega'\right)\left(\omega+\omega'\right)}u_0\right\},
\end{split}
\end{equation}
\end{widetext}
\noindent and
\begin{align}\label{betaCWL}
&\beta_{\omega\omega'}^{RL}\sim\nonumber\\
&-\frac{\eta}{2\pi}\sqrt{\omega'\omega}\frac{1-\sqrt{\nu}+i(\omega+\omega')\left(\frac{2}{\ln(\nu)}\left(\sqrt{\nu}-1\right)-1\right)u_0}{\left(\omega\sqrt{\nu}-i\eta\right)\left(\omega'+i\eta\right)\left(\omega+\omega'\right)}.
\end{align}
Since only the first order of $u_0$ is considered, in the computation of the modulus square of the Bogoliubov coefficients, the terms proportional to $u_0^2$ are neglected.
\indent Summing the modulus square of Eq.~\eqref{betaCWL} and Eq.~\eqref{betaCWR}, we obtain
\begin{widetext}
\begin{equation}\label{partprodintegrand}
\begin{split}
\left|\beta_{\omega\omega'}^{RR}\right|^2+\left|\beta_{\omega\omega'}^{RL}\right|^2=\frac{\eta^2}{4\pi^2}\omega\omega'\left[\frac{(\sqrt{\nu}-1)^2(\omega\nu+\omega')^2+\nu(\omega\sqrt{\nu}-\omega')^2(\sqrt{\nu}-1)^2+\eta^2(\nu-1)^2}{(\omega^2\nu+\eta^2)(\omega'^2+\eta^2)(\omega\nu+\omega')^2(\omega+\omega')^2}\right.\\+\left.
\frac{2(\omega+\omega')\left(\eta\sqrt{\nu}(\omega'-\omega\sqrt{\nu})(\sqrt{\nu}-1)\left(\frac{1}{\ln(\nu)}(\nu-1)-1\right)+\eta(\nu-1)\left(\frac{\omega'}{\ln\nu}(\sqrt{\nu}-1)^2\right)-\omega\nu\left(\frac{2}{\ln\nu}(\sqrt{\nu}-1)-1\right)\right)}{(\omega^2\nu+\eta^2)(\omega'^2+\eta^2)(\omega\nu+\omega')^2(\omega+\omega')^2}u_0
\right].
\end{split}
\end{equation}
\end{widetext}
Thus, by integrating over $d\omega'$ we get the total number of particles with frequency $\omega$ created by the mirror. Even though the corresponding number could be analytically computed, its expression turns out to be extremely complicated. Thus, we omitted it explicitly.\\
Studying Eq.~\eqref{partprodintegrand}, one can prove that the zero order term increases as $\nu$ increases. In particular, it converges to an asymptotic value $\mathcal{A}$ as
$\propto\sqrt{\nu}^{\,-1}$. Instead, the first order term, as $\nu$ increases, goes to zero faster, namely as $\frac{1}{\sqrt{\nu}\ln{\nu}}$. Starting from this fact, one can prove that, for $u_0$ enough small, an upper bound for the particle production is provided when $\nu\to\infty$. This upper bound is provided by the following analytical expression.
\begin{equation}
N_\omega\rightarrow\frac{\eta^2}{2\pi^2\omega}\left[\frac{\omega^2-\eta^2}{(\omega^2+\eta^2)^2}\ln\left(\frac{\omega}{\eta}\right)+\frac{\pi\eta\omega-\omega^2-\eta^2}{(\omega^2+\eta^2)^2}\right].
\end{equation}
The spectrum of the particles production for the trajectories shown in Fig.~\ref{penrose} is depicted in Fig.~\ref{tot}. In the Figs.~\ref{comp1},~\ref{comp2},~\ref{comp3} some comparisons among the contributions of the right and left part of the mirror are shown: it can be seen from them that the two contributions are comparable when $\eta\ll u_0^{-1}$. We also stress an infrared divergence for the spectrum, in agreement with the modified Carlitz-Willey trajectory as found in Refs. \cite{B} and \cite{C}. Moreover, we can observe that the contribution of the vacuum modes in the right of the mirror dominates over the one in the left for low frequencies. For high frequencies the contribution of the vacuum modes in the left is slightly higher than the one in the right, becoming the same for $\omega\rightarrow\infty$ only.
\color{black}
\begin{figure*}
\centering
\subfigure[Spectrum of the particles produced by the mirror with $\eta=1$ for different values of $A$, obtained integrating in $d\omega'$ Eq. \eqref{partprodintegrand}. It is $u_0=0.0001$.\label{tot}]{\includegraphics[height=0.3\hsize,clip]{TotPartProdalpha1u00001.pdf}}
\subfigure[Comparison of the spectra of the particles produced at the right hand side of the mirror. $N_\omega^R$ (resp. $N_\omega^L$) is the contribution from the vacuum modes at the right (resp. left) hand side of the mirror, obtained integrating in $d\omega'$ the modulus squared of Eq.~\eqref{betaCWR} (resp. Eq.~\eqref{betaCWL}). The complete spectrum is the sum $N^L_\omega + N^R_\omega$. Here $\eta=1$, $u_0=0.0001$ and $A=0.1$.\label{comp1}]{\includegraphics[height=0.3\hsize,clip]{Figure2bres.pdf}}
\subfigure[The same of \ref{comp1} but with $\eta=0.1$ and $A=0.5$.\label{comp2}]{\includegraphics[height=0.3\hsize,clip]{Figure2cres.pdf}}
\subfigure[The same of \ref{comp1} but with $\eta=0.1$ and $A=0.1$.\label{comp3}]{\includegraphics[height=0.3\hsize,clip]{Figure2dres.pdf}}
\caption{Plots showing the behaviour of the particle production as function of the frequency for a mirror with a trajectory like the ones shown in Fig.~\ref{penrose}.}
\label{figura}
\end{figure*}
\section{Quantum channel arising from a moving mirror}\label{Sec4}
We now revisit the model of semitransparent moving mirrors through an information communication perspective. In particular, the aim is to realize the field coming from left-past spacetime region, passing through the semitransparent moving mirror and ending up to right-future spacetime region as undergoing the action of a quantum channel. Then study the capacities in transmitting classical and quantum information of such a channel.
We consider as input a mode from $\mathcal{J}_L^-$ and as output the same mode transmitted by the mirror and outgoing toward $\mathcal{J}_R^+$. The bosonic modes in these two zones are related to each other through the Bogoliubov transformations, which are linear. Hence, taking the mode say of frequency $\omega$ as an input mode, while the other modes as environment modes (initially in vacuum, and after the process traced out), we end up with a Gaussian quantum channel.
To formalize such a mapping, following \cite{Giulio}, it is sufficient to know that a Gaussian quantum channel maps a bosonic Gaussian state into another bosonic Gaussian state. Considering only one mode $\omega$, the canonical variables in phase space are given by the vector $\mathbf{z}_\omega=(q_\omega,p_\omega)\in\mathbb{R}^2$ and a Gaussian bosonic state is represented by a characteristic function
\begin{equation}\label{carbeforechannel}
\chi^{in}({\bf z}_\omega)=\exp\left(-\frac{1}{4}{\bf z}_\omega^T\cdot\sigma_\omega\cdot{\bf z}_\omega+i{\bf d}_\omega^T\cdot{\bf z}_\omega\right),
\end{equation}
where ${\bf d}_\omega=\left(\langle Q_\omega\rangle,\langle P_\omega\rangle\right)$ and $\sigma_j$ is the so called covariance matrix, defined as\footnote{$Q_\omega=\frac{1}{\sqrt{2}}\left(a_\omega^\dagger+a_\omega\right)$ and $P_\omega=\frac{1}{i\sqrt{2}}\left(a_\omega-a_\omega^\dagger\right)$ are the canonical quadrature operators for the mode $\omega$.}
\begin{equation}
\sigma_\omega=\left(\begin{matrix}
\langle Q_\omega^2\rangle & \frac{1}{2}\langle P_\omega Q_\omega+Q_\omega P_\omega\rangle\\\frac{1}{2}\langle P_\omega Q_\omega+Q_\omega P_\omega\rangle& \langle P_\omega^2\rangle
\end{matrix}\right).\label{comatrix}
\end{equation}
A one mode Gaussian quantum channel maps the one mode characteristic function, Eq.~\eqref{carbeforechannel}, into
\begin{equation}\label{carafterchannel}
\begin{split}
\chi^{out}({\bf z}_\omega)=\exp\big(-\frac{1}{4}\mathbf{z}_\omega^T\cdot\left(\mathbb{T}\sigma_\omega\mathbb{T}^T+\mathbb{N}\right)\cdot\mathbf{z}_\omega\\
+i\left(\mathbf{d}_\omega^T\cdot\mathbb{T}^T+\mathbf{v}_\omega^T\right)\cdot\mathbf{z}_\omega\big),
\end{split}
\end{equation}
where a Gaussian quantum channel is so characterized by the triad $\left(\mathbb{T}, \mathbb{N}, \mathbf{v}\right)$ with $\mathbb{T}$ and $\mathbb{N}$ as two $2\times2$ matrices respectively related to the attenuation/amplification of the mode, and to the noise which affects the input signal. In particular, the attenuation/amplification is given by $\tau=\det\mathbb{T}$. Since the information transmission capabilities of the channel can be characterized in terms of entropic quantities, that do not depend on the vector $\mathbf{v}$, we can investigate the evolution of the covariance matrix, $\sigma_j$. Thus, from Eqs.~\eqref{carbeforechannel} and~\eqref{carafterchannel}, we write
\begin{equation}\label{covmap}
\sigma_\omega^{in}\longmapsto\sigma_\omega^{out}=\mathbb{T}\sigma_\omega^{in}\mathbb{T}^T+\mathbb{N}.
\end{equation}
Our focus is on the covariance matrix, Eq.~(\ref{comatrix}), of the input mode and its output as a result of Eq.~(\ref{covmap}). In line with our aim at the beginning of this section, all modes at $\mathcal{J}_L^-$ are considered uncorrelated and in the vacuum, but the single frequency mode of interest is $\omega$. This can be formalized with the following values
\begin{equation}\label{1imp}
\langle a_{\omega'}^La_{\omega''}^L\rangle=M\delta(\omega-\omega')\delta(\omega'-\omega''),
\end{equation}
\begin{equation}\label{2imp}
\langle a_{\omega'}^{L\dagger}a_{\omega''}^L\rangle=N\delta(\omega-\omega')\delta(\omega'-\omega''),
\end{equation}
and the expectation values of all the other combination of bosonic operators (including the ones relative to the right side of the mirror) equal to zero (except for the Hermitian of Eq.~\eqref{1imp} and the commutation of Eq.~\eqref{2imp}). Moreover, following Eqs.~\eqref{1imp} and \eqref{2imp}, $N$ is the mean number of particles in the input state, and $M$ specifies the correlation between $Q_\omega$ and $P_\omega$. \\
\indent Using Eqs.~(\ref{1imp}) and (\ref{2imp}), we calculate $\sigma_\omega^{in}$ and $\sigma_\omega^{out}$ taking the input and output bosonic operator for the mode $\omega$ (related to each other by a Bogoliubov transformation). We calculate the quadrature operators for both of them using Eq.~\eqref{1imp} and Eq.~\eqref{2imp}. It turns out that the input and output covariance matrices are related by a relation equal to Eq.~\eqref{covmap} from which we can obtain the entries of the $\mathbb{T}=\left(\begin{matrix}T_1&T_2\\T_3&T_4
\end{matrix}\right)$ and $\mathbb{N}=\left(\begin{matrix}N_1&N_2\\N_3&N_4
\end{matrix}\right)$ in terms of Bogoliubov coefficients, i.e.,
\begin{align}
&T_1=\epsilon\pi\Re\left(\alpha_{\omega\omega}^{RL}-\beta_{\omega\omega}^{RL}\right)\,,\\
&T_2=\epsilon\pi\Im\left(\alpha_{\omega\omega}^{RL}+\beta_{\omega\omega}^{RL}\right)\,,\\
&T_3=-\epsilon\pi\Im\left(\alpha_{\omega\omega}^{RL}-\beta_{\omega\omega}^{RL}\right)\,,\\
&T_4=\epsilon\pi\Re\left(\alpha_{\omega\omega}^{RL}+\beta_{\omega\omega}^{RL}\right)\,,
\end{align}
\begin{align}\label{Ns}
N_1=&-\frac{\epsilon\pi}{2}\left|\alpha_{\omega\omega}^{RL*}-\beta_{\omega\omega}^{RL}\right|^2+\\
&+\frac{1}{2}\int_{0}^{\infty}\left(\left|\alpha_{\omega\omega'}^{RL*}-\beta_{\omega\omega'}^{RL}\right|^2+\left|\alpha_{\omega\omega'}^{RR*}-\beta_{\omega\omega'}^{RR}\right|^2\right)d\omega'\,,\nonumber\\
\label{Ns2}N_2=&N_3=-\epsilon\pi\Im\left(\alpha_{\omega\omega}^{RL}\beta_{\omega\omega}^{RL}\right)+\\
&+\int_{0}^{\infty}\Im(\alpha_{\omega\omega'}^{RL}\beta_{\omega\omega'}^{RL}+\alpha_{\omega\omega'}^{RR}\beta_{\omega\omega'}^{RR})d\omega'\,,\nonumber\end{align}
\begin{align}\label{Ns3}
N_4=&-\frac{\epsilon\pi}{2}\left|\alpha_{\omega\omega}^{RL*}+\beta_{\omega\omega}^{RL}\right|^2+\\
&+\frac{1}{2}\int_{0}^{\infty}\left(\left|\alpha_{\omega\omega'}^{RL*}+\beta_{\omega\omega'}^{RL}\right|^2+\left|\alpha_{\omega\omega'}^{RR*}+\beta_{\omega\omega'}^{RR}\right|^2\right)d\omega',\nonumber
\end{align}
where $\epsilon$ is the cutoff\footnote{We ignored it for $\beta$ coefficients, although it is needful for $\alpha$ coefficients, otherwise as $\omega=\omega'$ possible divergences arise. } of Eqs.~\eqref{accrbc} and~\eqref{acclbg}. Here $\Re$ and $\Im$ denote the real and imaginary parts, respectively. \\
\indent For the average attenuation/amplification in time $\tau\coloneqq\det\mathbb{T}$, we have the following general expression
\begin{equation}\label{taugeneric}
\tau=\epsilon^2\pi^2\left(|\alpha_{\omega\omega}^{RL}|^2-|\beta_{\omega\omega}^{RL}|^2\right).
\end{equation}
Applying the Bogoliubov coefficients of an impulsive accelerated mirror to first order in a $u_0$ expansion around $u_0=0$ and using the Carlitz-Willey's acceleration we get
\begin{equation}\label{tau}
\tau=\frac{4\Omega^4\nu+\Omega^2\left(1+\sqrt{\nu}\right)^2}{4\left(\Omega^2\nu+1\right)\left(\Omega^2+1\right)},
\end{equation}
where $\Omega\coloneqq\omega/\eta$. It is easy to see that $\tau = \frac{\omega^2}{\omega
^2+\eta
^2}$, as expected, for $\nu=1$. We are reminded that as $\Omega\to \infty$ one has $\tau =1$ (perfect transparency) and as $\Omega \to 0$, one has $\tau\to 0$ (perfect reflection), even when $\nu\neq 1$.
Transmission through a semitransparent mirror should result in three main effects:
\begin{itemize}
\item[{\bf 1.}] \emph{a loss of the input signal, since part of that is reflected};
\item[{\bf 2.}] \emph{an interference of the input signal with the other modes of the initial vacuum environment};
\item[{\bf 3.}] \emph{a particle production contribution which eventually amplifies the input signal}.
\end{itemize}
One can prove that $\tau<1$ for all frequencies and $\nu$, providing that no signal amplification occurs, as evidence of the first effect. Nevertheless, for $\nu>1$, there is a reduction of such loss. The second and third effect also show themselves in this way (nevertheless, the effect of the particle production mostly arises as noise, since it occurs even without an input signal).
Another relevant aspect of Eq. \eqref{tau} is that to the first order in $u_0$ around $u_0=0$ does not contribute to $\tau$. In general, one can prove that also the other orders of $u_0$ does not give a contribution to $\tau$ in the limit $\epsilon\rightarrow0$. In fact, calculating $\tau$ from the general expression Eq.~\eqref{acclbg} (taking the corresponding $\alpha$ coefficient), for finite acceleration periods result in convergent integrals. Under these conditions, in the limit $\epsilon\rightarrow0$, only the divergent part of $\alpha_{\omega\omega}^{RL}$ gives a contribution:
\begin{equation}
\lim_{\epsilon\rightarrow\infty}\alpha_{\omega\omega}^{RL}=\frac{1}{\pi\epsilon}-\frac{\eta}{2\pi\epsilon}\left[\frac{\left(1+\sqrt{\nu}\right)\omega-2i\eta}{i\left(\omega\sqrt{\nu}-i\eta\right)\left(\omega-i\eta\right)}\right].
\end{equation}
After a calculation, one can prove that this will lead to the same $\tau$ of Eq.~\eqref{tau}. In conclusion, we have shown that, for the calculation of $\tau$ we are justified in removing the restriction over a small acceleration period. It is then realized that this $\tau$
is valid for any mirror trajectory with a finite acceleration period, at least in the limit $\epsilon\rightarrow0$.\\
Studying $\tau$ from Eq. \eqref{tau}, for $\nu\rightarrow\infty$, we have an asymptotic behaviour of $\tau$ equal to
\begin{equation}\label{asymptotictau}
\tau=\frac{\Omega^2+\frac{1}{4}}{\Omega^2+1}.
\end{equation}
However, differently from the particle production $N_\omega$ (see Sec.~\ref{Sec3}), $\tau$ does not increase in a monotonic way by increasing the final speed of the mirror toward its left (as considered in Fig. \ref{penrose}). In fact, for each value of $\Omega$ we have a finite value of $\nu$ which maximizes $\tau$. In other words, it exists a critical final $V_{\textrm{crit}}(\Omega)$ of a mirror (accelerating toward the left) with finite acceleration period, for which $\tau$ reaches a maximum and slightly decreases for $V>V_{\textrm{crit}}$, asymptotically reaching $\tau$ described by Eq. \eqref{asymptotictau}. The critical value of $\nu$, say $\nu_{\textrm{crit}}$, in function of $\Omega$, can be written as
\begin{equation}\label{critA}
\nu_{\textrm{crit}}(\Omega)=\frac{9}{2}+\frac{4}{\Omega^2} +\frac{1+(3\Omega^2+1)\sqrt{(9\Omega^2+1)(\Omega^2+1)}}{2\Omega^4}.
\end{equation}
For $\Omega\ll1$ we have $\nu_{\textrm{crit}}\rightarrow\infty$ which corresponds to $V_{\textrm{crit}}=1$. For $\Omega\gg1$ the critical speed reaches the asymptotic value $V_{\textrm{crit}}(\Omega\to\infty)=0.8$. The decreasing of $\tau$ after $V_{\textrm{crit}}$ is sharper for low frequencies, although $V_{\textrm{crit}}$ is closer to the speed of light in this range. A plot for $\tau$ is portrayed in Fig. \ref{tau figure}.
\begin{figure}
\centering
\includegraphics[scale=0.6]{tau.pdf}
\caption{Behaviour of $\tau$ vs $\Omega$ from Eq.~\eqref{tau}. It was considered: a static mirror ($\nu=1$), a mirror accelerating toward the left with a final speed comparable to $V_{\textrm{crit}}\sim 0.8$, i.e. the critical speed at high frequencies ($\nu=10$) and a mirror accelerating toward the left with a final speed really close to the speed of light ($\nu=1000$).}
\label{tau figure}
\end{figure}
One expects a contribution to the noise created by the mirror due to the particle production and characterized by $N_\omega$ with frequency mode $\omega$. Since $\tau<1$ for each $\omega$, $\eta$ and $A$, we anticipate a lossy and noisy quantum channel. By the classification of one-mode Gaussian channels made in Ref. \cite{Holevo}, we perform two unitary Gaussian transformations, one before and one after the quantum channel (respectively named, pre-processing and post-processing), in order to reduce the matrix $\mathbb{N}$, of a lossy and noisy quantum channel, to its canonical form
\begin{equation}\label{Ncan}
\mathbb{N}_c=(1-\tau)\left(\frac{1}{2}+\overline{n}\right)\mathbb{I},
\end{equation}
where $\overline{n}$ is the number of noisy particles created by the quantum channel. The term ${1\over2}$ is naturally related to vacuum energy induced by mode $\omega$. In the continuous limit we expect that it takes the value $\frac{1}{2\pi\epsilon}$.
Moreover, instead of finding the average number of noisy particles arriving to the detector $\overline{n}$, in the continuous limit we expect to have a spectrum of particles expressed as $\overline{n}_\omega$. The former, once integrated in a range of frequencies, provides a dimensionless number. Hence, in the continuous case, Eq.~\eqref{Ncan} becomes
\begin{equation}\label{contNcan}
\det\mathbb{N}=\left(1-\tau\right)\left(\frac{1}{2\pi\epsilon}+\overline{n}_\omega\right).
\end{equation}
In the continuous limit only an infinitesimal range of frequencies would be detected. We can therefore write $\overline{n}=\epsilon\pi\overline{n}_\omega$.\\
\indent Since the determinant of $\mathbb{N}$ does not change when reducing it to its canonical form, we can study it from Eqs.~\eqref{Ns}, \eqref{Ns2} and \eqref{Ns3}, leading to
\begin{equation}\label{Ncont}
\det\mathbb{N}=\det\mathbb{N}_c=\left(\frac{1-\tau}{2\pi\epsilon}+B\right)^2-C^2,
\end{equation}
where
\begin{equation}
B\coloneqq-\epsilon\pi\left|\beta_{\omega\omega}^{RL}\right|^2+N_\omega,
\end{equation}
and
\begin{equation}
C\coloneqq\left|\epsilon\pi\alpha_{\omega\omega}^{RL}\beta_{\omega\omega}^{RL}-\sum_{S=L,R}\int_{0}^{\infty}\alpha_{\omega\omega'}^{RS}\beta_{\omega\omega'}^{RS}d\omega'\right|.
\end{equation}
Comparing Eq.~\eqref{contNcan} with Eq.~\eqref{Ncont}, we get
\begin{equation}\label{spectrumnoisy}
\overline{n}_\omega=\frac{1}{2\pi\epsilon}\left[-1+\sqrt{1+4\pi\epsilon\frac{B}{1-\tau}+4\pi^2\epsilon^2\frac{B^2-C^2}{(1-\tau)^2}}\right].
\end{equation}
Thus, the average number of noisy particles arriving to the detector is
\begin{equation}\label{contlimitnoise}
\overline{n}=\frac{1}{2}\left[-1+\sqrt{1+4\frac{\tilde{B}}{1-\tau}+4\frac{\tilde{B}^2-\tilde{C}^2}{(1-\tau)^2}}\right],
\end{equation}
where $\tilde{B}\coloneqq\epsilon\pi B$ and $\tilde{C}=\epsilon\pi C$.\\
\indent If $B$ and $C$ are not divergent, for $\epsilon\rightarrow0$ we have $B=N_\omega$ and expanding the square root in the last term of Eq.~\eqref{spectrumnoisy}, we get the spectrum
\begin{equation}\label{overn}
\overline{n}_\omega=\frac{N_\omega}{1-\tau}.
\end{equation}
Consequently, in this case, $\overline{n}=0$, that corresponds to our stand-alone approximation, i.e., to an impulsive accelerated mirror. Indeed, we already demonstrated $N_\omega$ is convergent as well as $B$, with $\tilde{B}=0$. Analogously, one can easily prove the convergence of $C$ as well, leading to $\tilde{C}=0$. This behaviour naturally suggests that the impulsive semitransparent mirror acts like a beam splitter.
A different expectation occurs when $B$ and $C$ are divergent, e.g. for the perfectly reflecting Carlitz-Willey mirror. In this case a rigorous approach to get $\tilde{B}$ and $\tilde{C}$ requires the use of wave packets, where the frequency range is supposed to vanish. In fact, applying this approach to the Carlitz-Willey trajectory furnishes a finite $\tilde{B}$, see e.g. \cite{B} for further details.
\section{Quantum channel capacities}\label{Sec5}
In this section, we evaluate classical and quantum capacities of the quantum channel described in the previous section. In so doing, we quantify the capability of an impulsive accelerated mirror to transmit both classical and quantum information.
For bosonic Gaussian channels the regularization of the capacities is a hard task and this problem is not fully solved, neither for classical nor quantum capacities. Fortunately, the channel we obtained in Sec. \ref{Sec4} becomes a beam splitter in the continuous limit. For this kind of channels the additivity is proved both for classical \cite{LossyCapacity} and quantum capacity \cite{Calculationoneshot}.
We start by studying the classical capacity. Let us take the classical information we want to transmit with continuous random variable $X$ and probability distribution $p_x$. The encoding procedure is identified by a map which associates to each value $x$ of the random variable a state $\rho_x$. Let $\Phi$ be the quantum channel of communication. The maximum that we can extract about $X$ at the channel output is given by \emph{Holevo information} \cite{Holevoinf1,Holevoinf2}:
\begin{equation}
\chi(\rho,\Phi)=S\left(\Phi(\rho)\right)-\int p_xS(\Phi(\rho_x))dx.
\end{equation}
where $S$ is the von Neumann entropy and $\rho\coloneqq\int p_x\rho_xdx$.\\
For one-mode Gaussian (OMG) channels it is possible to express the Holevo information in terms of covariance matrices if we restrict the possible encodings to Gaussian ones, see e.g. \cite{OMGclassicalc}. Namely, we have to restrict the possible inputs of the OMG channel $(\mathbb{T},\mathbb{N},\mathbf{v})$ to be bosonic Gaussian states with covariance matrix $\sigma$ and $\mathbf{d}=(x,0)$. Moreover we assume $p_x$ to be a Gaussian probability distribution with mean equal to zero and covariance matrix $\sigma'$. In the reference \cite{ConjectureGaussian} it is proved that, if the channel is a beam splitter, such encodings maximize the Holevo information, becoming \cite{HolevoCM}
\begin{equation}\label{Holevoinformation}
S\left(\mathbb{T}(\sigma+\sigma')\mathbb{T}^T+\mathbb{N}\right)-S\left(\mathbb{T}\sigma\mathbb{T}^T+\mathbb{N}\right),
\end{equation}
where the Von Neumann entropy $S$, referring to a covariance matrix $\sigma$, can be written by $S(\sigma)=h(d)$, with $d\coloneqq\sqrt{\det(\sigma)}$ and
\begin{equation*}
S(\sigma)=\left(d+\frac{1}{2}\right)\log\left(d+\frac{1}{2}\right)-\left(d-\frac{1}{2}\right)\log\left(d-\frac{1}{2}\right).
\end{equation*}
At this point, the classical capacity $C$ is given by the maximum of Eq.~\eqref{Holevoinformation} over the inputs $\sigma$ and $\sigma'$. However, since the bosonic Gaussian states are in an infinite-dimensional Hilbert space, as input we can take a state with an infinite particle amount. Obviously, this maximizes the Holevo information and it leads to an infinite classical capacity. This case is unrealistic, since we need an infinite amount of energy for the encoding process. In order to remove this possibility, we have to impose a restriction on the maximum energy $E$ which can be used for the encoding, by
\begin{equation}
\frac{1}{2}\omega\text{Tr}(\sigma+\sigma')\le E.
\end{equation}
With this prescription, the following classical capacity for a noiseless, lossy channel (beam splitter), has been obtained by \cite{LossyCapacity}
\begin{equation}\label{cgused}
\begin{split}
C&=\frac{\tau E}{\omega}\log\left(\frac{\tau E+\omega}{\tau E}\right)+\log\left(\frac{\tau E+\omega}{\omega}\right).
\end{split}
\end{equation}
Plots of $C$ are shown is Fig. \ref{ClassicalCapacity}. Since the channel is asymptotically without loss for $\omega\rightarrow\infty$, one can expect a constant capacity in this limit. However, for great frequencies the encoding necessitates more energy. For this reason, if the energy is constrained, the number of photons we can use for the encoding decreases linearly. In fact, studying the asymptotic behaviour of $\tau$ for $\omega\rightarrow\infty$, it turns out that $C$ goes to zero very slowly, $\sim\frac{\ln\omega}{\omega}$. Instead, for $\omega\rightarrow0$, $C$ goes to zero linearly. Moreover, from Eq.~\eqref{cgused} we can see that the capacity increases with $\tau$.
As a consequence, we expect the capacity to be maximized for $\nu=\nu_{\textrm{crit}}$. Further, it is interesting to notice, from Fig. \ref{ClassicalCapacity}, how for different values of $\nu$ the maximum of the lower bound of the classical capacity occurs for different values of $\omega$. This ``maximum capacity frequency" seems to decrease when we increase $\nu$. However, since the peak becomes higher increasing the final speed of the mirror, we can conclude that the capability of an impulsive accelerated mirror to transmit classical information always increases with the mirror final speed.\\
\begin{figure}
\centering
\includegraphics[scale=0.6]{CGE5.pdf}
\caption{Classical capacity of an impulsive accelerated mirror constraining the encoding energy to $E=5$ according to Eq.~\eqref{cgused} for various values of $\nu$. $\tau$ is provided by the Eq.~\eqref{tau} with $\eta=1$. }
\label{ClassicalCapacity}
\end{figure}
Next, we proceed to evaluate the quantum capacity, obtained maximizing the coherent information over the input (for a beam-splitter channel, the quantum capacity is additive \cite{HolevoCM}). Br\'adler \cite{OMGquantumc} proved that the one-shot quantum capacity of a lossy and noisy OMG channel is maximized: either when the number of photons $N$ used for the encoding is null, or when this number $N$ is infinite (so, we have an infinite amount of energy for the encoding). In the first case the maximized coherent information is zero and no quantum information can be transmitted reliably. However, unlike the classical capacity framework, in case of infinite amount of energy of the encoding we have a finite value for the coherent information. This means that there is no need to impose a constraint for the energy of the encoding in order to have a finite value for the one-shot quantum capacity. Nevertheless, the infinite energy of the encoding is unrealistic. For this reason, the quantum capacity that we intend to study might be considered as upper bound of the ``real one" with a finite encoding energy. However, the quantum capacity, obtained with a finite $E$, is basically the same of the one obtained with $E\rightarrow\infty$ in the region $\omega\ll E$. Consequently, even if we compute a quantum capacity for $E\rightarrow\infty$, it is realistic to choose $E$ large enough within the range of employed frequencies.
For the coherent information $J_c$ of a OMG lossy channel, as $E\rightarrow\infty$ we have \cite{Holevo,Calculationoneshot}
\begin{equation}
J_c(E\rightarrow\infty)=\log\frac{\tau}{1-\tau}.
\end{equation}
Using $\tau$ from Eq. \eqref{tau} we get
\begin{equation}\label{coherentmirror}
J_c(E\rightarrow\infty)=\log\left[\Omega^2\frac{\frac{(1+\sqrt{\nu})^2}{4}+\Omega^2\nu}{1+\Omega^2\left(\frac{3}{4}-2\sqrt{\nu}+\frac{3}{4}\nu\right)}\right],
\end{equation}
whose behaviour is shown in Fig. \ref{QuantumCapacity}.
\begin{figure}
\centering
\includegraphics[scale=0.6]{OSQC.pdf}
\caption{Coherent information for the impulsive accelerated mirror with an input having an infinite number of photons, according to Eq.~\eqref{coherentmirror}.}
\label{QuantumCapacity}
\end{figure}
The quantum capacity of an impulsive accelerated mirror is given by
\begin{equation}\label{osqc}
Q(\omega)=\max\{0,J_c(E\rightarrow\infty,\omega)\}.
\end{equation}
Analyzing Eq.~\eqref{coherentmirror}, we have that $Q$ diverges logarithmically as $\omega\rightarrow\infty$, in agreement with the fact that the channel becomes without loss in this limit. Even in this case, for each frequency, the maximum of the quantum capacity is obtained for $\nu=\nu_{\textrm{crit}}$. The quantum capacity is non-zero only if $\tau>\frac{1}{2}$. This is in agreement with the no-cloning theorem, for which it is impossible to transmit quantum information reliably if we have a loss $1-\tau$ larger than $\frac{1}{2}$, otherwise the input state can be cloned, (see e.g. \cite{Erasure}). Further, from Fig. \ref{osqc} we observe that in the range of frequencies in which $Q>0$ the curve with $\nu=10$ is larger than the one with $\nu=1000$. This suggests that, unlike the classical capacity, in order to maximize the quantum capacity with an impulsive accelerated mirror we are forced to take a final speed of the mirror comparable with the critical speed for high frequencies, i.e., $V\sim0.8$.
The minimum frequency required for a non-null quantum capacity depends upon $\nu$ through the relation
\begin{equation}
\Omega^2=\frac{(1-\sqrt{\nu})^2}{2\nu}+\sqrt{\frac{(1-\sqrt{\nu})^4}{4\nu^2}+\frac{1}{\nu}}.
\end{equation}
For both $\nu\to1$ and $\nu\to\infty$ the frequency converges to $\Omega=1$. By construction, we thus expect a minimum in the range $1<\nu<\infty$.
\section{Final remarks}\label{Sec6}
In this work, we studied partially reflecting accelerating mirrors finding general expressions for the Bogoliubov coefficients. This work has been motivated by analog models, called accelerated boundary correspondences, that describe the correspondence between the particle production from a null-shell of a collapsing black hole and the particle production from a perfectly reflecting accelerating mirror.
The natural extension to semi-transparency indicated potential signatures of new effects. Along this line, we have studied the trajectories in which the mirror satisfies a few physical conditions: in the past, it lies at rest and in the future it shows a finite acceleration period, ending with a constant sub-light speed. We introduced the concept of impulsive accelerated mirrors and we computed the Bogoliubov coefficients by considering a very short acceleration period. Consequently, we evaluated particle production from the so-obtained Bogoliubov coefficients, providing explicit analytical expressions dependent on frequency and on the final speed of the mirror.
The particles considered are non-interactive scalar particles, hence with coupling constant $\lambda\rightarrow0$. It was proved in Refs.\cite{Interaction1,Interaction2} that the interaction (even if infinitesimal) gives a non-negligible contribute to the particle production at the time $t\sim\lambda^{-1}$. Here we considered an error on the frequency $\Delta(\omega)=\epsilon\pi$. As a consequence, the particles produced are the ones in the time interval $\left(-(\epsilon)^{-1},+(\epsilon)^{-1}\right)$. Since we have taken the continuous limit $\epsilon\to0$, in order to neglect the contribute of the interaction, it is sufficient to have $\lambda\rightarrow0$ faster than $\epsilon$. Moreover, considering mirrors with a finite acceleration period, we expect the particle production (and its consequent effects) to occur during the acceleration period. Hence, even if we consider $\lambda$ finite but very small, we do not care about what happens at times $t>\lambda^{-1}$, since the acceleration is over at such times.
Next, we have recognized the mirror as a Gaussian quantum channel acting between the spacetime regions of left-past and right-future. The evolution of an input signal crossing the mirror could then be studied using the previously obtained Bogoliubov coefficients. For these quantum channels, we investigated the properties of transmission of an input signal, the noise created by the mirror over the channel and we finally evaluated both the classical and quantum capacities. Since we were searching for analytic solutions for the Bogoliubov coefficients, the continuous limit for the frequencies was considered. As a consequence, all the properties of the mirror as a Gaussian channel (i.e., $\tau$, $\overline{n}$ and the capacities) are an average in time from $-\infty$ to $+\infty$.
In addition, we speculated about the physical consequences of our framework. In particular, the simplicity and flexibility of the moving mirror model, coupled to its unique collection of radiative properties, demonstrate that with use of appropriate trajectories the moving mirror idealization of evaporating black hole radiation and information transfer are remarkably suitable. In harmony with black hole complementarity \cite{Susskind:1993if}, observers on both sides of the mirror cannot make simultaneous physical measurements, much the same way that one cannot both simultaneously measure, to within the uncertainty principle, the position and momentum of a particle in quantum mechanics. With non-horizon perfect reflection, the information stays on one side of the mirror, carried by the radiation providing full knowledge of the initial quantum state; this necessarily requires that the radiation is never precisely thermal but quasi-thermal \cite{Good:2019tnf}. That is, the particles are not distributed in an exact Planck distribution but carry small imprinting evidence of collapse. With non-horizon semi-transparency (e.g. Eq.~(\ref{pacc})), the right (left) observer collects the information from both the right (left)-movers which reflect (transmit) through the mirror, giving complete information about the initial state. For the sake of completeness, one can also consider the complementary communication scheme, namely the reflection case of an input signal incoming from right-past and outgoing to right-future. This prescription is likely less interesting than the transmission case. Indeed, the former works better in modeling black holes and information theory and holds a more appropriate physical meaning.
From the results of this work, we speculate that this picture in general, and further use of this class of trajectories in particular, will contribute towards the resolution of information transfer in the black hole evaporation process.
Concluding, to get relevant information about the time in which such properties occur, future works will generalize our treatment considering wave packets, whose wave packet width, $\Delta\omega$, satisfies $\Delta\omega\sim u_0^{-1}$ and so we will investigate physical properties during this small acceleration period, i.e., $u_0$.
The results lead smoothly to further investigations with respect to astrophysical applications to compact objects. For example, possible scenarios of high-energy astrophysical explosions could be object of future works modeled by means of our approach.
\acknowledgments
Funding from state-targeted program ``Center of Excellence for Fundamental and Applied Physics" (BR05236454) by the Ministry of Education and Science of the Republic of Kazakhstan is acknowledged, as well as the FY2021-SGP-1-STMM Faculty Development Competitive Research Grant No. 021220FD3951 at Nazarbayev University. OL is also thankful to the Ministry of Education and Science of the Republic of Kazakhstan, Grant: IRN AP08052311 for financial support. SM is thankful to the funding from European Union's Horizon 2020 research and innovation program under grant agreement no. 862644 (FET-Open Project: QUARTET).
|
\section{Introduction}
Recent advances in deep learning have shown state-of-the-art (SOTA) results on the challenging single-view depth estimation (SVDE) and stereo disparity estimation (SDE) tasks. In particular, self-supervised methods for SVDE have reached performance levels similar or even superior to the fully-supervised networks \cite{packing3d, semguide, falnet}. However, the previous SOTA self-supervised methods are unable to predict accurate pixel-level depth estimates, which are often observed along the object's depth boundaries. Predicting pixel-level accurate 3D geometries is essential for robotic grasping, augmented reality, navigation, and 3D object detection.
In this paper, we present a pixel-level accurate depth estimation network (PLADE-Net) with neural positional encoding and a distilled matting Laplacian loss, both of which allow for consistent depth estimates in homogeneous areas and sharp depth predictions along the object boundaries. Our PLADE-Net outperforms the most recent self-supervised SOTA methods \cite{packing3d,semguide,falnet,unos,reversing} (both mono and stereo) by large margins, achieving an unprecedented accuracy on the challenging KITTI dataset while keeping a low number of parameters. This paper's contributions are:
\begin{enumerate}[leftmargin=*,noitemsep,topsep=0.5pt]
\vspace*{1mm}
\item We propose to exploit and distill the closed-form solution of the matting Laplacian \cite{matting} for self-supervision, leading to a novel loss function that allows for pixel-level accuracy in self-supervised single- and stereo-view DE.
\vspace*{1mm}
\item We show that neural positional encoding (NPE) can be usefully incorporated into CNNs for depth estimation, as it allows the network to reason about camera distortions, scene orientation, and non-local relationships.
\vspace*{1mm}
\item We present PLADE-Net, a novel network architecture that incorporates NPE. Our PLADE-Net incorporates multi-scale inputs and a single-scale output, opposite to single-scale inputs and multi-scale outputs in previous works \cite{monodepth1,monodepth2,packing3d, depth_hints}. Relative to previous works, our PLADE-Net doubles the number of filter channels in the early feature extraction layers, and halves the number of filter channels in its bottleneck. These seemingly trivial design choices, already make our PLADE-Net, even without our newly proposed loss functions, to outperform the previous SOTA methods.
\vspace*{1mm}
\item The PLADE-Net is the first work that shows unprecedented accuracy levels for SVDE, exceeding 95\% in terms of $\delta^1$ metric on the challenging KITTI\cite{kitti2012} dataset.
\end{enumerate}
\vspace*{1mm}
Figure \ref{fig:op_img} compares the depth estimate performances of the most recent SOTA methods \cite{depth_hints, packing3d, falnet} with respect to our PLADE-Net. As shown in the detailed view of the estimated depth regions (dotted boxes numbered from \raisebox{.5pt}{\textcircled{\raisebox{-.9pt} {1}}}\ to \raisebox{.5pt}{\textcircled{\raisebox{-.9pt} {4}}}), our PLADE-Net produces very precisely estimated depths along the object boundaries. Simultaneously, the SOTA methods fail by yielding inaccurate object depths that partially leak into the background.
Our paper is organized as follows: In Section 2, we review the most relevant self-supervised methods for our work; Section 3 presents our PLADE-Net with neural positional encoding and a distilled matting Laplacian loss with in-depth explanations; In Section 4, we provide extensive ablation studies and experiments that support the effectiveness of our contributions; Finally, we conclude our work in Section 5.
\section{Related Works}
Learning-based self-supervised single view depth estimation (SVDE) is a relatively new problem and has rapidly advanced since it was first proposed in the work of Garg \etal \cite{garg}. Self-supervised SVDE is usually achieved by exploiting the 3D information embedded in datasets that contain multiple captures from the same scene. Previous methods have successfully learned SVDE from stereo pairs \cite{garg, monodepth1, refinedistill, superdepth, net3, depth_hints, infuse_classic, falnet} and video \cite{sfmlearner, monodepth2, packing3d, semguide, depthwild}, by training their CNNs for the backward or forward synthesis of the training image samples, given the target view as input. Similarly, other works \cite{monodepth1, epc, bridging, unos, reversing} have addressed the less ill-posed problem of stereo depth estimation (SDE), where the left and right views are available during training and testing. As our proposed PLADE-Net learns from stereo images, we only review the methods that learn from stereo in this section for the sake of simplicity.
\textbf{Learning SVDE from stereo.} Among the top-performing SVDE methods that learn from stereo we find the works of \cite{infuse_classic, depth_hints, falnet}. The contemporary works of Tosi \etal \cite{infuse_classic} and Watson \etal \cite{depth_hints} proposed to guide the training of their SVDE networks with distilled stereo disparity estimates obtained from the classical approach of semi-global matching (SGM) \cite{sgm0, sgm1}. While Watson \etal \cite{depth_hints} used the SGM disparity as a proxy label when the resulting photometric loss is lower than the CNN-estimated depth, Tosi \etal \cite{infuse_classic} distilled the SGM proxy label via left-right (LR) consistency checks. Inspired by \cite{infuse_classic, depth_hints}, we distill the matting Laplacian with both photometric and LR-consistency checks in this work.
The recent work of Gonzalez and Kim \cite{falnet}, proposed to \enquote{forget about the LiDAR}, by learning high-quality depths with a multi-view occlusion module and exponentially quantified disparity volumes. Additionally, they proposed a two-stage training strategy to learn view synthesis and refine their network, called FAL-net\cite{falnet}, for SVDE. While their method obtains the SOTA metrics on the KITTI \cite{kitti2012} dataset, their approach is still far from generating pixel-level accurate depths, as shown in Figure \ref{fig:op_img}-\raisebox{.5pt}{\textcircled{\raisebox{-.9pt} {3}}}. Their second stage loss functions cannot enforce sharp object depth boundaries, as they are limited by the computed occlusions' quality, leading to sub-optimal estimates.
\textbf{Learning SDE.} Interestingly, the less ill-posed problem of learning stereo disparity estimation (SDE) in a self-supervised manner has been studied less extensively than the single-view case. The most prominent works include those of Wang \etal \cite{unos} and Aleotti \etal \cite{reversing}. Wang \etal \cite{unos} proposed to exploit spatiotemporal information by learning SDE from stereo videos. Their \enquote{UnOS} learns optical flow, stereo disparity, and camera pose by spatially and temporally projecting the target views into the spatiotemporal reference images and measuring the reconstruction errors to provide means of self-supervision. The work of Aleotti \etal proposes to distill the disparity estimates from a monocular disparity competition network to provide additional proxy labels for the SDE task. Aleotti \etal achieve the SOTA by training existing networks \cite{monodepth2, piramid_sm} with their monocular proxy labels, which remove the well-known stereo artifacts caused by occlusions \cite{monodepth1, sgm0, sgm1, reversing}.
\section{Method}
We propose a novel Pixel-Level Accurate Depth Estimation network, called PLADE-Net, with neural positional encoding and a distilled matting Laplacian loss. Architecture-wise, neural positional encoding is incorporated into our PLADE-Net to learn location-specific image features. Training-wise, our PLADE-Net learns single-view depth from stereo pairs in a two-stage training strategy following the previous work \cite{falnet}. In the first stage of training, our PLADE-Net is trained for simple stereoscopic view synthesis with a combination of $l_1$, perceptual \cite{perceptual}, and smoothness losses. In the second stage, our network is fine-tuned with an occlusions-free reconstruction loss with the multi-view occlusion module and other secondary smoothness and mirror losses defined in \cite{falnet}. More importantly, a distilled matting Laplacian loss is newly proposed, allowing the learning of highly accurate pixel-level depth estimates.
In the following subsections, we introduce an image (and inverse depth) formation model (which follows the one defined in \cite{falnet}) and describe the intuition behind our main contributions for \textit{neural positional encoding} and \textit{a distilled matting Laplacian loss}. We then describe the details of our network architecture and training loss functions in Subsections 3.4 and 3.5, respectively.
\subsection{Stereoscopic Image Formation Model}
We build our PLADE-Net based on the work of Gonzalez and Kim \cite{falnet}, as their method showed SOTA results for learning single-view depth from stereo images. Therefore, we adopt their image formation model in which the convolutional neural network (CNN) outputs a disparity probability logit volume ${\tens{D}}^L_L$ for a given left input view ${\tens{I}}_L$. ${\tens{D}}^L_L$ can be either progressively projected to the right-view and soft-maxed channel-wise to form the right-from-left disparity probability volume ${\tens{D}}^{PR}_L$ or simply soft-maxed to generate the left disparity probability volume ${\tens{D}}^{PL}_L$. ${\tens{D}}^{PR}_L$ can be used for stereoscopic view synthesis by
\begin{equation} \label{eq:synth_right}
{\tens{I}}'_R = \textstyle \sum_{n=0}^{N} g\left({\tens{I}}_L, d_n\right) \odot {\tens{D}}^{PR}_{L_n},
\end{equation}
where $\odot$ indicates the Hadamard product, $g(\cdot)$ denotes a shifting of the input image to the left by $d_n$ pixels, and $N$ is the number of planes in the probability volume ${\tens{D}}^{PR}_L$. On the other hand, ${\tens{D}}^{PL}_L$ can be used to extract the disparity map ${\tens{D}}'_L$, which is learned as a by-product from the view synthesis task, as defined by
\begin{equation} \label{eq:disp}
{\tens{D}}'_L = \textstyle \sum_{n=0}^{N} d_n {\tens{D}}^{PL}_{L_n}
\end{equation}
We also adopt the exponential disparity quantization in \cite{falnet}, as it is well-posed for the depth estimation task. Exponential quantization takes into account the inverse relationship between disparity and depth by distributing far- and close-by quantization levels more evenly and is given by
\begin{equation} \label{eq:exp_disc}
d_n = d_{max}e^{\ln{d_{max}/d_{min}}(n/N-1)},
\end{equation}
where $d_{max}$ and $d_{min}$ are the minimum and maximum disparity hyper-parameters. For a fair comparison with \cite{falnet}, we set $d_{max}=300$ and $d_{min}=2$ for all our experiments.
\subsection{Neural Positional Encoding}
It is well-known that convolutional neural networks on their own are very well capable of encoding positional information \cite{how_much_pos}. However, since the local CNN filters are shared across spatial locations, a network, trained with randomly cropped patches from the original image data, will struggle to learn location-specific features, such as lens or projection distortions, ground versus sky regions, and potentially non-local relationships. In particular, lens and projection distortions make the objects near image borders appear to be more stretched than those in the image centers, the degree of which often depends on the camera focal length. This is a potential source of confusion for the CNNs, as two objects in the same distance to a camera will be projected differently on the camera plane, depending on their relative position to the resulting image, as illustrated in Figure \ref{fig:rot_effect}.
The relative object size is an essential cue for depth estimation. It can affect the estimation accuracy if a network does not have a means of understanding the locations of the training patches in their original images. To provide the network with a mechanism to account for the likelihood of objects being stretched when they are located close to the image borders, we propose neural positional encoding (NPE) for depth estimation. We realize NPE into our PLADE-Net as the concatenation of deep positional features at each encoder stage. A deep positional feature map ${\tens{F}}_{npe}$ is obtained by processing the pixel location ${\bm{p}}=(x,y)$ information of each patch with two fully-connected layers with exponential linear unit (ELU) activations, which is given by:
\begin{equation} \label{eq:npe}
{\tens{F}}_{npe}({\bm{p}}) = elu({\bm{w}}_2 \cdot elu({\bm{w}}_1 \cdot {\bm{p}} + b_1) + b_2),
\end{equation}
where $w_{1,2}$ and $b_{1,2}$ are the learnable weights and biases of our neural positional encoding layers. The operation in Eq. \ref{eq:npe} can be trivially realized with $1\times1$ convolutions in available deep learning libraries. Note that in contrast with \cite{camconvs}, we do not concatenate $x$- and $y$-coordinates of ${\bm{p}}$, but do concatenate our deep positional feature maps into the downstream convolutional layers in our PLADE-Net. It should be noted again that, in our neural positional encoding, ${\bm{p}}=(x,y)$ are the pixel locations of the patches relative to their original images before cropping.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{figures/rot_effect.jpg}
\vspace*{-3mm}
\caption{An illustration of projection distortions in image borders. Objects closer to the image borders appear stretched and closer, like the three in (a) in comparison with the three in (b).}
\label{fig:rot_effect}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figures/matting_distill.jpg}
\vspace*{-7mm}
\caption{Matted disparity distillation process.}
\label{fig:matting_dist}
\vspace*{-4mm}
\end{figure}
\subsection{Distilling the Matting Laplacian}
The closed-form solution to the matting Laplacian \cite{matting} is a useful tool in classical low-level computer vision. It can sharply segment an input image based on pixel intensity and proximity in the input image, a roughly estimated or user-defined segmentation map, and a confidence map. The matting Laplacian has been used to refine depth estimates, generating structurally sharp but incorrectly labeled matted depth maps. We exploit the strong features of the matting Laplacian to learn highly accurate pixel-level depth in a self-supervised fashion, remedying its weak points by distilling the matted depth maps with photometric and left-right consistency \cite{monodepth1} checks.
Our matting Laplacian distillation process is depicted in Figure \ref{fig:matting_dist}. Given an input image ${\tens{I}}_L$ and its initial depth estimate ${\tens{D}}'_{L0}$ by our PLADE-Net, we generate a matted disparity ${\tens{D}}'_{LM0}$ following \cite{matting}. As can be observed in Figure \ref{fig:matting_dist}-(c), ${\tens{D}}'_{LM0}$ is very sharp, but many pixels are wrongly labeled, such as the tree at the left-hand side of ${\tens{I}}_L$, the biker in the background, and the overall road depths. To remedy this, we apply a $5 \times 5$ local window mean scaling to obtain the locally scaled matted disparity map ${\tens{D}}'_{LM}$, as shown in Figure \ref{fig:matting_dist}-(d). Then, we estimate a distillation mask ${\tens{M}}_L$ via photometric and left-right consistency checks as given by:
\begin{equation} \label{eq:dist_md}
\begin{split}
{\tens{M}}&_L = \big[\big|{\tens{I}}_L - g({\tens{I}}_R,{\tens{D}}'_{LM})\big| < \big|{\tens{I}}_L - g({\tens{I}}_R,{\tens{D}}'_L)\big|\big] \odot \\
&\big[\big|{\tens{D}}'_{LM} - g({\tens{D}}'_{RM},{\tens{D}}'_{LM})\big| < \big|{\tens{D}}'_L - g({\tens{D}}'_R,{\tens{D}}'_L)\big|\big],
\end{split}
\end{equation}
where ${\tens{D}}'_{R0}$ and ${\tens{D}}'_{RM}$ are the initial disparity estimate and the locally scaled matted disparity map for the corresponding right-view input image ${\tens{I}}_R$, respectively. $g(\cdot)$ works as a backward-warping operation. In Eq. \ref{eq:dist_md}, if the inequalities in both brackets are satisfied at a pixel location, the resulting mask value at that location is 1, otherwise 0. The distillation map for the right view ${\tens{M}}_R$ is obtained by swapping the L and R sub-scripts in Eq. \ref{eq:dist_md}. Eq. \ref{eq:dist_md} selects as a source for self-supervisions the pixel depths in ${\tens{D}}'_{LM}$ that (i) generate better backward warped images and (ii) are more consistent with their corresponding right view pixel depths. As can be noted in the detailed view of the hand of the biker in Figure \ref{fig:matting_dist}-(e), the matted disparity becomes dramatically sharper than the initial disparity estimate but has incorrect values. The mean-scaled matted disparity is both sharper and correctly scaled. As expected, the distillation mask is active on the biker's hand edges, which guides our network to generate pixel-level accurate depth estimates.
\subsection{Network Architecture}
Our proposed PLADE-Net adopts the simple auto-encoder backbone from \cite{falnet}, with residual blocks in the encoder side and nearest-upscale-based up-convolutions followed by skip-connections in the decoder side. However, we considerably change the learned feature maps' distribution by doubling the extracted features in the shallow convolutional layers (from 32 to 64) and halving the number of feature maps in the bottleneck (from 512 to 256). Our PLADE-Net is depicted in Figure \ref{fig:network} and incorporates our proposed NPE by concatenating (denoted by \raisebox{.5pt}{\textcircled{\raisebox{-.5pt} {c}}}) deep positional features at each encoder stage's input.
Contrary to the previous works \cite{monodepth1, monodepth2, depth_hints, packing3d, falnet} that incorporate a single-scale input and multi-scale outputs, our PLADE-Net adopts multi-scale inputs and a single scale output. In our PLADE-Net, low-level features are extracted from a bilinearly downscaled version of the input image ${\tens{I}}_L$ and concatenated into the second encoder stage's input, as depicted to the left-hand side of Figure \ref{fig:network}. Our PLADE-Net outputs a single-scale disparity logit volume ${\tens{D}}^L_L$, which can be employed for novel view synthesis and single-view depth estimation, as shown to the right-hand side of Figure \ref{fig:network}, according to Eqs. \ref{eq:synth_right} and \ref{eq:disp}, respectively.
Our PLADE-Net delivers higher performance with an equal or lower number of parameters than the previous works, with 15M versus 17M of the previous SOTA \cite{falnet}. It is worth noting that our PLADE-Net achieves the SOTA performance without the need for any advanced layer such as attention, batch/group normalization, sub-pixel convolution, or spatial pyramid pooling. Detailed architecture layer information can be found in the supplementary materials.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figures/network.jpg}
\vspace*{-5mm}
\caption{High-level overview of our proposed PLADE-Net.}
\label{fig:network}
\vspace*{-4mm}
\end{figure}
\subsection{Loss Functions}
Following the training strategy in \cite{falnet}, we train our PLADE-Net in two stages. In the first stage, we focus on learning stereoscopic view synthesis, which can be understood as training the top output branch of our PLADE-Net in Figure \ref{fig:network}, which generates a synthetic right view. In the second stage, we train our PLADE-Net with an occlusions-free reconstruction loss. Still, more importantly, we incorporate additional loss functions that affect the lower output branch in Figure \ref{fig:network}, which generates a near pixel-accurate disparity estimate.
The total loss function ($l_{s1}$) in the first stage of training is a combination of $l_1$, perceptual \cite{perceptual} ($l_p$), and disparity smoothness ($l_{ds}$) losses, as given by
\begin{equation} \label{eq:l_s1}
l_{s1} = l_1 + \alpha_pl_p + \alpha_{ds}l_{ds},
\end{equation}
where $\alpha_p$ and $\alpha_{ds}$ are empirically set to 0.01 and 0.0004, respectively, to balance their contributions. The total loss $l_{s2}$ for the second training stage adds our novel distilled matting Laplacian loss ($l_{dm}$), a deep corr-$l_1$ loss ($l_{dc}$), and the mirror loss ($l_{m}$) in \cite{falnet} to $l_{s1}$, and is defined by:
\begin{equation} \label{eq:l_s2}
l_{s2} = l_{s1} + l_{m} + \alpha_{dm}l_{dm} + \alpha_{dc}l_{dc},
\end{equation}
where $\alpha_{dm}=0.25$ and $\alpha_{dc}=0.01$ are empirically set to weight the contributions of the distilled matting Laplacian and the deep corr-$l_1$ losses, respectively.
In the first training stage, $l_{s1}$ is computed for the left view only. In the second stage, $l_{s2}$ is computed for both left and right views, giving the actual total loss of $l=(l^L_{s2}+l^R_{s2})/2$. For the sake of completeness, we describe $l_{s1}$ and $l_{m}$ in detail in our supplemental materials.
\vspace*{1mm}
\textbf{Deep Corr-$l_1$ Loss.} Inspired by \cite{structurepre}, we explored training our PLADE-Net with a deep corr-$l_1$ loss, to encourage the generation of depth estimates with structural details similar to the ones in the single-view input. However, we observed marginal performance improvements and depth artifacts, as further shown in Section 4. Nevertheless, we observed an affinity between the deep corr-$l_1$ loss and our proposed distilled matting Laplacian loss. Our deep corr-$l_1$ loss ($l_{dc}$) penalizes the deep-auto-correlation difference between the input image and the predicted depth map. Deep-auto-correlation is obtained by measuring the auto-correlation of the deep features of ${\tens{I}}_L$ or ${\tens{D}}'_L$, extracted by the third maxpool layer of a pre-trained VGG19 \cite{vgg} for the image classification task, denoted by $\phi(\cdot)$. $l_{dc}$ is then given by:
\begin{equation} \label{eq:dcl_loss}
l_{dc} = ||\text{acorr}(\phi({\tens{D}}'_L), k)-\text{acorr}(\phi({\tens{I}}_L), k)||_1,
\end{equation}
where $\text{acorr}(\cdot, k)$ is the auto-correlation operator on a $k$$\times$$k$ window, empirically set to $k=3$ in all our experiments.
\vspace*{1mm}
\textbf{Distilled Matting Laplacian Loss.} We previously detailed our matting Laplacian distillation process in Subsection 3.3. Given the locally scaled matted left disparity map ${\tens{D}}'_{LM}$ and distillation mask ${\tens{M}}_L$, our distilled matting Laplacian loss $l_{dm}$ is simply given by:
\begin{equation} \label{eq:dm_loss}
l_{dm} = (1 / \text{max}({\tens{D}}'_L))||{\tens{M}}_L \odot({\tens{D}}'_L - {\tens{D}}'_{LM})||_1,
\end{equation}
where $\text{max}({\tens{D}}'_L)$ is the maximum disparity value in the scene and normalizes the loss between 0 and 1. By incorporating ${\tens{M}}_L$ into Eq. (\ref{eq:dm_loss}), we can keep the highly detailed matted depths while filtering out the incorrectly labeled pixels commonly present in image matting.
\begin{table}[t]
\small
\centering
\setlength{\tabcolsep}{2.5pt}
\input{tables/table_ablation}
\vspace*{-2mm}
\caption{Ablation studies of our PLADE-Net on KITTI\cite{kitti2012}. Metrics are \colorbox{c_lowbest}{the lower the better} and \colorbox{c_highbest}{the higher the better}.}
\label{tab:ablation}
\vspace*{-2mm}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figures/loss_ablation.jpg}
\vspace*{-8mm}
\caption{Ablation studies on our distilled matting Laplacian loss.}
\label{fig:loss_ablation}
\vspace*{-3mm}
\end{figure}
\section{Experiments and Results}
\subsection{Datasets}
\textbf{KITTI\cite{kitti2012}.} To compare with a wider spectrum of recent works, we utilize the Eigen train split \cite{eigen} (K), which is a subset of the KITTI \cite{kitti2012} training set, consisting of 22,600 left-right training image pairs captured from a moving car. Following the standard practice, we test our method on the KITTI Eigen test split in its original \cite{eigen} and improved \cite{kitti_official} versions, which contain 697 and 652 images with projected LiDAR ground truths (GT), respectively. The improved Eigen test split contains denser GTs by selectively accumulating LiDAR points from 5 consecutive frames. Performance is measured with the metrics defined in \cite{eigen} (up to 80m). Additionally, in our experiments, we propose a \enquote{naive} stereo input extension of our PLADE-Net, which is trained with a split obtained from the intersection of the KITTI Eigen train set \cite{eigen} and the KITTI Split \cite{monodepth1}. The resulting Stereo-Split excludes scenes from the KITTI Eigen test split \cite{eigen} and the KITTI2015 training set \cite{kitti2015}. The KITTI2015 \cite{kitti2015} training set consists of 200 image pairs with CAD-refined LiDAR GT and is the default benchmark to evaluate self-supervised stereo networks.
\textbf{CityScapes\cite{cityscapes}.} In most of our ablation studies, we concurrently train the PLADE-Net variations with the CityScapes \cite{cityscapes} dataset (following the multi-dataset training procedure in \cite{deep3dpan, falnet}) to ensure that they do not under-perform due to the lack of enough data. The CityScapes \cite{cityscapes} dataset consists of 24,500 stereo pairs without depth GTs, and similar to KITTI \cite{kitti2012}, it is captured from a driving perspective. We follow the car hood and border artifacts removal procedures from \cite{monodepth1, deep3dpan, falnet}.
\textbf{Make3D\cite{make_3d}.} To test the generalization power of our PLADE-Net, we evaluate it on the Make3D \cite{make_3d} Test134 dataset, which is made of 134 high-resolution RGB outdoor images with low-resolution depth GTs. We followed the evaluation procedure defined in \cite{monodepth1}, with the C1 metrics (up to 70m) defined in \cite{liudisc}.
\subsection{Implementation Details}
Following the training procedure in \cite{falnet}, we trained our PLADE-Net for 50 epochs in the first training stage and 10 epochs in the second stage by the Adam\cite{adam} optimizer with an initial learning rate of 0.0001 and 0.00005 for the first and second training stages, respectively. The learning rate was reduced to half at epochs [30, 40, 50] in the first training stage and epochs [5, 10] in the second stage. Data augmentations on-the-fly were incorporated into our network training. For a fair comparison with previous works, we adopted random resizing from 0.75 to 1.5, followed by $192\times640$ random cropping, random left-right flipping, random gamma, random brightness, and random individual color brightness. Inference was run at full image resolution.
\subsubsection{Computing the Matting Laplacian}
Computing the closed-form solution to the matting Laplacian \cite{matting} is expensive, taking up to 30 (60) seconds in matting a KITTI \cite{kitti2012} (CityScapes \cite{cityscapes}) depth-RGB sample. For this reason, we first generated a matted disparity complementary dataset instead of matting on-the-fly. We used our PLADE-Net with one stage of training to build such a matted disparity dataset. We procured to apply the corresponding spatial data augmentations (resizes, crops, and flips) to the matted disparity samples during training. Our distillation process re-scales the matted disparity values, thus not requiring to apply scaling factors during data sampling.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figures/res_make3d.jpg}
\vspace*{-9mm}
\caption{Qualitative comparisons on the Make3D dataset \cite{make_3d}.}
\label{fig:make_3d}
\vspace*{-2mm}
\end{figure}
\begin{table}
\small
\centering
\setlength{\tabcolsep}{5.5pt}
\input{tables/table_make3d}
\vspace*{-2mm}
\caption{Results on Make3D \cite{make_3d}. All self-supervised methods benefit from median scaling. M3D: Training on the Make3D\cite{make_3d}.}
\label{tab:make3d}
\vspace*{-2mm}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{figures/stereo_k2015.jpg}
\vspace*{-7mm}
\caption{Stereo results on the KITTI2015\cite{kitti2015} dataset.}
\label{fig:kitti2015}
\vspace*{-2mm}
\end{figure}
\begin{table}[t]
\small
\centering
\setlength{\tabcolsep}{1.7pt}
\input{tables/table_kitti2015}
\vspace*{-2mm}
\caption{Comparison of existing self-supervised SDE methods on the KITTI2015 \cite{kitti2015} training set. SV: Training from stereo videos. \textbf{Best} and \underline{second-best} metrics. Results capped to 80m.}
\label{tab:kitti_2015}
\vspace*{-3mm}
\end{table}
\begin{figure*}
\centering
\includegraphics[width=0.97\textwidth]{figures/ablation_pe.jpg}
\vspace*{-4mm}
\caption{Qualitative ablation studies on our proposed neural positional encoding (NPE) in our PLADE-Net.}
\label{fig:ablation_npe}
\vspace*{-4mm}
\end{figure*}
\subsection{Ablation Studies}
Table \ref{tab:ablation} shows our ablation studies on the improved KITTI Eigen test split \cite{kitti_official}. We first ablate the effects of our NPE in the PLADE-Net for the first stage of training. As it can be noted, the positional encoding (PE) that simply concatenates the pixel location $(x, y)$ values directly into the encoder stage can even yield slight performance improvements in most metrics in comparison with the cases without it (denoted as \enquote{w/o PE}). However, our PLADE-Net shows substantial performance improvements in all metrics by incorporating our \textit{neural} positional encoding (NPE).
The ablation studies on the effects of our NPE in the second stage of training are shown in the second section of Table \ref{tab:ablation}. Interestingly, our PLADE-Net w/o PE gets stuck in bad local minima, with marginal performance improvements in the second training stage. In contrast, our PLADE-Net with simple PE outperforms all previous SOTA methods in terms of $\delta^1$ accuracy. On the other hand, our PLADE-Net with NPE exhibits the best performance by considerable margins. Our PLADE-Net with NPE trained only with the KITTI Eigen train split \cite{eigen} shows the efficacy of robust learning regardless of the training data size. The effects of our NPE in both training stages (TS1 and TS2) are depicted in Figure \ref{fig:ablation_npe}. Our PLADE-Net without NPE struggles to estimate depths for objects close to the image borders, yielding depth artifacts regardless of the training stage.
The third section of Table \ref{tab:ablation} and Figure \ref{fig:loss_ablation} show respectively the quantitative and qualitative ablation studies for our proposed loss functions. As can be noted, our distilled matting Laplacian loss is the main contributor to achieving such high-performance. Our PLADE-Net trained with distilled matting Laplacian loss only ($a_{dc}=0$, $a_{dm}=0.25$) achieves an $\delta^1$ accuracy of 94.8\% while our PLADE-net with deep corr-$l_1$ loss only ($a_{dc}=0.01$, $a_{dm}=0$) obtains the lower accuracy of 94.5\%. In addition, training with only deep corr-$l_1$ loss induces depth artifacts seen as bright spots in Figure \ref{fig:loss_ablation}-(d). Our PLADE-Net without our proposed loss functions ($a_{dc}=0$, $a_{dm}=0$) shows the lowest performance with the most blur depth estimates as shown in Figure \ref{fig:loss_ablation}-(c). Interestingly, our PLADE-Net with ($a_{dc}$=$0$, $a_{dm}$=$0$) still shows considerably better metrics than the previous SOTA FAL-net \cite{falnet}, demonstrating the effectiveness of our NPE and design choices of multi-scale inputs and more learned features on the shallow feature extractors, suggesting that SVDE benefits from richer low-level features.
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{figures/res_kitti.jpg}
\vspace*{-9mm}
\caption{Qualitative comparisons on the KITTI Eigen test split \cite{eigen}. Our PLADE-Net consistently estimates much more detailed depths.}
\label{fig:res_eigen}
\end{figure*}
\subsection{Resutls}
\textbf{Results on KITTI.} Table \ref{tab:kitti_eigen} and Figure \ref{fig:res_eigen} present quantitative and qualitative comparisons among the previous methods and our PLADE-Net on the KITTI Eigen test split \cite{eigen}. Our PLADE-Net clearly outperforms all previous self-supervised methods in most metrics on the original Eigen test split \cite{eigen}, and in all the metrics on the improved test split \cite{kitti_official}. Our PLADE-Net shows sharper and pixel-level accurate depth estimates in complex and cluttered image regions, as shown in every zoom-box of Figure \ref{fig:res_eigen}. Quantitatively, our PLADE-Net without any post-processing (PP), even trained only on the KITTI (K) dataset, outperforms the previous methods that were trained on KITTI + CityScapes (K+CS). Following the PP step in \cite{falnet}, our method achieves even higher accuracies and lower error metrics.
\begin{table*}[t]
\small
\centering
\input{tables/table_kitti}
\vspace*{-2mm}
\caption{Evaluations on the KITTI Eigen test split \cite{eigen}. Models are trained on the KITTI Eigen\cite{eigen} train-split (K) and CityScapes\cite{cityscapes} (CS). CS$\rightarrow$K indicates CS pre-training. K+CS indicates concurrent K and CS training. DoF and D denote depth-of-field and depth supervision. S, S$_{\text{SGM}}$, V, V+Se indicate stereo, stereo+SGM, video, and video + semantics self-supervision. V methods benefit from median-scaling. \textbf{Best} and \underline{second-best} metrics. Methods that use post-processing (PP) are checked \redcheck. Results capped to 80m.}
\label{tab:kitti_eigen}
\vspace*{-7mm}
\end{table*}
\textbf{Results on KITTI (stereo)}
To further evaluate the effectiveness of our PLADE-Net, we define an stereo input variant, the PLADE-NetS, which is evaluated and compared with the SOTA methods on the KITTI2015 dataset. Our stereo variant is a clone of the PLADE-net, with the difference that the PLADE-NetS \enquote{naively} incorporates the right-view image information in a second encoder, whose bottleneck features are concatenated to the left view bottleneck features. In our PLADE-NetS we do not incorporate any advanced stereo matching layers such as 1D-Correlation or 3D convolutions, and still, our network manages to outperform the most recent self-supervised SOTA methods \cite{reversing, unos} in most metrics by a considerable margin, as indicated in Table \ref{tab:kitti_2015}. Figure \ref{fig:kitti2015} shows that our network with stereo inputs keeps generating very sharp and pixel-level accurate depth estimates with clear object boundaries.
\textbf{Results on Make3D.} Table \ref{tab:make3d} compares our PLADE-Net against the SOTA self-supervised methods on Make3D \cite{make_3d}. Our approach generalizes the best among the self-supervised methods under comparison and is very close to the fully-supervised method of Laina \etal \cite{lainadeeper}. It is clear in Figure \ref{fig:make_3d} that our PLADE-Net generates sharper depth estimates on the previously unseen Make3D dataset \cite{make_3d} in comparison with the recent FAL-net \cite{falnet} SOTA.
\section{Conclusions}
We showed that our PLADE-Net with neural positional encoding (NPE) could generalize better than the conventional CNN approaches. NPE allows our PLADE-Net to learn location-specific features, which aid in predicting consistent disparities in all image regions. Furthermore, our proposed distilled matting Laplacian loss provides strong self-supervision signals to learn sharp and pixel-level accurate depth estimation. Our PLADE-Net outperforms all previous self-, semi-, and fully-supervised methods on the challenging KITTI dataset with unprecedented accuracy levels and exhibits superior generalization capacities on the Make3D and CityScapes datasets.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}\label{IMR}
The $H^2$-regularity of variational solutions to a two-dimensional transmission problem with geometric constraint is investigated, in particular when part of the interface becomes part of the outer boundary of the domain due to the geometric constraint, a situation in which the domain includes some non-Lipschitz subdomains with cusp points.
To set up the geometric framework, let $D:=(-L,L)$ be a finite interval of $\mathbb{R}$, $L>0$, and let $H>0$ and $d>0$ be two positive parameters. Given a function $u\in C(\bar{D},[-H,\infty))$ with $u(\pm L)=0$, we define the subdomain $\Omega(u)$ of $D\times (-H,\infty)$ by
$$
\Omega(u) := \left\{ (x,z)\in D\times \mathbb{R} \,:\, -H< z < u(x)+d \right\} = \Omega_1(u)\cup \Omega_2(u)\cup \Sigma(u)\,,
$$
where
$$
\Omega_1(u) := \left\{ (x, z)\in D\times \mathbb{R} \,:\, -H< z< {u}(x)\right\}
$$
and
$$
\Omega_2(u):= \left\{ (x,z)\in D\times \mathbb{R}\,:\, u(x)< z < u(x)+d\right\}
$$
are separated by the interface
$$
\Sigma(u) := \left\{ (x,z)\in D\times \mathbb{R}\,:\, z= u(x)>-H \right\}\,.
$$
\begin{figure}
\begin{tikzpicture}[scale=0.9]
\draw[black, line width = 1.5pt, dashed] (-7,0)--(7,0);
\draw[black, line width = 2pt] (-7,0)--(-7,-2.5);
\draw[black, line width = 2pt] (7,-2.5)--(7,0);
\draw[black, line width = 2pt] (-7,-2.5)--(7,-2.5);
\draw[blue, line width = 2pt] plot[domain=-7:7] (\x,{-0.75-0.75*cos((pi*\x/7) r)});
\draw[blue, line width = 2pt] plot[domain=-7:7] (\x,{-1.25-0.75*cos((pi*\x/7) r)});
\draw[blue, line width = 1pt, arrows=->] (2,0)--(2,-1.65);
\node at (2.2,-0.6) {${\color{blue} v}$};
\node at (-5,-1.5) {${\color{blue} \Omega_1(v)}$};
\node at (-3,0.5) {${\color{blue} \Omega_2(v)}$};
\draw (-3.6,0.5) edge[->,bend right,line width = 1pt] (-4.7,-0.7);
\node at (0,-3.25) {$D$};
\node at (5.75,-1.75) {{\color{blue} $\Sigma(v)$}};
\draw (5.25,-1.75) edge[->,bend left, line width = 1pt] (3.5,-1.3);
\node at (-7.8,1) {$z$};
\draw[black, line width = 1pt, arrows = ->] (-7.5,-3)--(-7.5,1);
\node at (-8,-2.5) {$-H$};
\draw[black, line width = 1pt] (-7.6,-2.5)--(-7.4,-2.5);
\node at (-7.8,-0.5) {$0$};
\draw[black, line width = 1pt] (-7.6,-0.5)--(-7.4,-0.5);
\node at (-7.8,0) {$d$};
\draw[black, line width = 1pt] (-7.6,0)--(-7.4,0);
\node at (-7,-3.25) {$-L$};
\node at (7,-3.25) {$L$};
\draw[black, line width = 1pt, arrows = <->] (-7,-2.75)--(7,-2.75);
\end{tikzpicture}
\caption{Geometry of $\Omega(v)$ for a state $v\in S$ with empty coincidence set.}\label{Fig1}
\begin{tikzpicture}[scale=0.9]
\draw[black, line width = 1.5pt, dashed] (-7,0)--(7,0);
\draw[black, line width = 2pt] (-7,0)--(-7,-2.5);
\draw[black, line width = 2pt] (7,-2.5)--(7,0);
\draw[black, line width = 2pt] (-7,-2.5)--(7,-2.5);
\draw[red, line width = 2pt] plot[domain=-7:7] (\x,{-1-cos((pi*\x/7) r)});
\draw[red, line width = 2pt] plot[domain=-7:7] (\x,{-1.5-cos((pi*\x/7) r)});
\draw[red, line width = 1pt, arrows=->] (2,0)--(2,-2.05);
\node at (2.3,-0.6) {${\color{red} w}$};
\node at (-5,-1.5) {${\color{red} \Omega_1(w)}$};
\node at (-3,0.5) {${\color{red} \Omega_2(w)}$};
\draw (-3.65,0.5) edge[->,bend right,line width = 1pt] (-4.7,-0.75);
\node at (0,-3.25) {$D$};
\node at (5.75,-1.75) {{\color{red} $\Sigma(w)$}};
\draw (5.25,-1.75) edge[->,bend left, line width = 1pt] (3.5,-1.6);
\node at (-7.8,1) {$z$};
\draw[black, line width = 1pt, arrows = ->] (-7.5,-3)--(-7.5,1);
\node at (-8,-2.5) {$-H$};
\draw[black, line width = 1pt] (-7.6,-2.5)--(-7.4,-2.5);
\node at (-7.8,-0.5) {$0$};
\draw[black, line width = 1pt] (-7.6,-0.5)--(-7.4,-0.5);
\node at (-7.8,0) {$d$};
\draw[black, line width = 1pt] (-7.6,0)--(-7.4,0);
\node at (-7,-3.25) {$-L$};
\node at (7,-3.25) {$L$};
\draw[black, line width = 1pt, arrows = <->] (-7,-2.75)--(7,-2.75);
\node at (-1.5,-3.25) {${\color{red} \mathcal{C}(w)}$};
\draw (-1,-3.25) edge[->,bend right, line width = 1pt] (0,-2.55);
\draw[black, line width = 2pt] (-7,-2.5)--(7,-2.5);
\end{tikzpicture}
\caption{Geometry of $\Omega(w)$ for a state $w\in \mathcal{S}$ with non-empty coincidence set.}\label{Fig2}
\end{figure}
Owing to the (geometric) constraint $u\ge -H$, the lower boundary of $\Omega_2(u)$, given by the graph of the function $u$, cannot go beyond the lower boundary $D\times \{-H\}$ of $\Omega_1(u)$ but may coincide partly with it, along the so-called coincidence set
\begin{equation}
\mathcal{C}(u) := \{x\in D\,:\, u(x)=-H\}\,, \label{CS}
\end{equation}
see Figures~\ref{Fig1} and~\ref{Fig2}. Clearly, the geometry of $\Omega(u)$, as well as the regularity of its boundary, heavily depends on whether $\min_D\{u\}>-H$ or $\min_D\{u\}=-H$. Indeed, if $\min_D\{u\}>-H$ (i.e. the graph of $u$ is strictly separated from $D\times \{-H\}$ as in Figure~\ref{Fig1}), then the coincidence set $\mathcal{C}(u)$
is empty and $\Omega_1(u)$ is connected. In contrast, if $\min_D\{u\}=-H$ so that the graph of $u$ intersects $D\times \{-H\}$, then $\mathcal{C}(u)\ne\emptyset$ and $\Omega_1(u)$ is disconnected with at least two (and possibly infinitely many) connected components, see Figures~\ref{Fig2} and~\ref{Fig3}.
For such a geometry, we study the regularity of variational solutions to the transmission problem
\begin{subequations}\label{psi}
\begin{align}
\mathrm{div}(\sigma\nabla\psi_u) &=0 \quad\text{in }\ \Omega(u)\,,\label{a1a}\\
\llbracket \psi_u \rrbracket =\llbracket \sigma\nabla \psi_u \rrbracket \cdot \mathbf{n}_{ \Sigma(u)} &=0 \quad\text{on }\ \Sigma(u)\,,\label{a1b}\\
\psi_u&=h_u\quad\text{on }\ \partial\Omega(u)\,,\label{a1c}
\end{align}
\end{subequations}
where
\begin{equation*}
\sigma := \sigma_1 \mathbf{1}_{\Omega_1(u)} + \sigma_2 \mathbf{1}_{\Omega_2(u)}
\end{equation*}
for some positive constants $\sigma_1\ne \sigma_2$, and $\mathbf{n}_{ \Sigma(u)}$ denotes the unit normal vector field to $\Sigma(u)$ (pointing into ${\Omega}_2(u)$) given by
$$
\mathbf{n}_{ \Sigma(u)}:=\frac{(-\partial_x u, 1)}{\sqrt{1+(\partial_x u)^2}}\,.
$$
In \eqref{a1c}, $h_u$ is a suitable function reflecting the boundary behavior of $\psi_u$, see Section~\ref{sec.abv} for details. In addition, $\llbracket \cdot \rrbracket$ denotes the (possible) jump across the interface $\Sigma(u)$; that is,
\begin{equation*}
\llbracket f \rrbracket(x,u(x)) := f|_{\Omega_1(u)}(x,u(x)) - f|_{\Omega_2(u)}(x,u(x))\,, \qquad x\in D\,,
\end{equation*}
whenever meaningful for a function $f:\Omega(u)\to\mathbb{R}$.
\medskip
Let us already mention that there are several features of the specific geometry of $\Omega(u)$ which may hinder the $H^2$-regularity of the solution $\psi_u$ to \eqref{psi}. Indeed, on the one hand, the interface $\Sigma(u)$ always intersects with the boundary $\partial\Omega(u)$ of $\Omega(u)$ and it follows from \cite{Lem77} that this sole property prevents the $H^2$-regularity of $\psi_u$, unless $\sigma$ and the angles between $\Sigma(u)$ and $\partial\Omega(u)$ at the intersection points satisfy some additional conditions. On the other hand, $\Omega(u)$ and $\Omega_2(u)$ are at best Lipschitz domains, while $\Omega_1(u)$ may consist of non-Lipschitz domains with cusp points.
\medskip
The particular geometry $\Omega(u)= \Omega_1(u)\cup \Omega_2(u)\cup \Sigma(u)$, in which the boundary value problem~\eqref{psi} is set, is encountered in the investigation of an idealized electrostatically actuated microelectromechanical system (MEMS) as described in detail in~\cite{LW18}. Such a device consists of an elastic plate of thickness~$d$ which is fixed at its boundary $\{\pm L\} \times (0,d)$ and suspended above a rigid conducting ground plate located at $z=-H$. The elastic plate is made up of a dielectric material and deformed by a Coulomb force induced by holding the ground plate and the top of the elastic plate at different electrostatic potentials. In this context, $u$ represents the vertical deflection of the bottom of the elastic plate, so that the elastic plate is given by $\Omega_2(u)$, while $\Omega_1(u)$ denotes the free space between the elastic plate and the ground plate. An important feature of the model is that the elastic plate cannot penetrate the ground plate, resulting on the geometric constraint $u\ge -H$. Still, a contact between the elastic plate and the ground plate --~corresponding to a non-empty coincidence set $\mathcal{C}(u)$~-- is explicitly allowed. The dielectric properties of $\Omega_1(u)$ and $\Omega_2(u)$ are characterized by positive constants $\sigma_1$ and $\sigma_2$, respectively. The electrostatic potential $\psi_u$ is then supposed to satisfy \eqref{psi} and is completely determined by the deflection $u$. The state of the MEMS device is thus described by the deflection $u$, and equilibrium configurations of the device are obtained as critical points of the total energy which is the sum of the mechanical and electrostatic energies, the former being a functional of $u$ while the latter is the Dirichlet integral of $\psi_u$. Owing to the nonlocal dependence of $\psi_u$ on $u$, minimizing the total energy and deriving the associated Euler-Lagrange equation demand quite precise information on the regularity of the electrostatic potential $\psi_u$ for an arbitrary, but {\it fixed} function $u$ and its continuous dependence thereon. This first step of provisioning the required information is the main purpose of the present research, and we refer to the forthcoming paper~\cite{LW2?} where the minimizing problem leading to the determination of~$u$ is analyzed.
\medskip
Since the regularity of the variational solution $\psi_u$ to \eqref{psi} is intimately connected with the regularity of the boundaries of $\Omega(u)$, $\Omega_1(u)$, and $\Omega_2(u)$, let us first mention that $\Omega(u)$ and $\Omega_2(u)$ are always Lipschitz domains and that the measures of the angles at their vertices do not exceed $\pi$, a feature which complies with the $H^2$-regularity of $\psi_u$ away from the interface $\Sigma(u)$ \cite{Gr85}. This property is shared by $\Omega_1(u)$ when the coincidence set $\mathcal{C}(u)$ is empty, see Figure~\ref{Fig1}, so that it is expected that $\psi|_{\Omega_i(u)}$ belongs to $H^2(\Omega_i(u))$, $i=1,2$, in that case. However, when $\mathcal{C}(u)$ is non-empty, the open set $\Omega_1(u)$ is no longer connected and the boundary of its connected components is no longer Lipschitz, but features cusp points. Moreover, there is an interplay between the transmission conditions~\eqref{a1b} and the boundary condition~\eqref{a1c} when $\mathcal{C}(u)\ne\emptyset$. Whether $\psi|_{\Omega_i(u)}$ still belongs to $H^2(\Omega_i(u))$, $i=1,2$, in this situation is thus an interesting question, that we answer positively in our first result. For the precise statement, we introduce the functional setting we shall work with in the sequel. Specifically, we set
$$
\bar{\mathcal{S}} := \{v\in H^2(D) \cap H_0^1(D)\,:\, v\ge -H \text{ in } D \;\text{ and }\; \pm \llbracket \sigma \rrbracket \partial_x v(\pm L) \le 0 \}\,,
$$
and
$$
\mathcal{S}:=\{v\in H^2(D) \cap H_0^1(D)\,:\, v> -H \text{ in } D \;\text{ and }\; \pm \llbracket \sigma \rrbracket \partial_x v(\pm L) \le 0 \}\,.
$$
Clearly, the coincidence set $\mathcal{C}(u)$ is empty if and only if $u\in \mathcal{S}$. In addition, the situation already alluded to, where $\mathcal{C}(u)$ is non-empty and $\Omega_1(u)$ is a disconnected open set in $\R^2$ with a non-Lipschitz boundary, corresponds to functions $u\in\bar{\mathcal{S}}\setminus \mathcal{S}$. Also, we include the constraint $\pm \llbracket \sigma \rrbracket \partial_x u(\pm L) \le 0$ in the definition of $\mathcal{S}$ and $\bar{\mathcal{S}}$ to guarantee that the way $\Sigma(u)$ and $\partial\Omega(u)$ intersect does not prevent the $H^2$-regularity of $\psi_u$ in smooth situations (i.e. $u\in \mathcal{S}\cap W_\infty^2(D)$), see \cite{Lem77}.
\begin{theorem}\label{Thm1}
Suppose \eqref{bobbybrown} below.
\begin{itemize}
\item[(a)] For each $u\in \bar{\mathcal{S}}$, there is a unique variational solution $\psi_u \in h_{u}+H_{0}^1(\Omega(u))$ to \eqref{psi}. Moreover, $\psi_{u,1}:= \psi_{u}|_{\Omega_1(u)} \in H^2(\Omega_1(u))$ and $\psi_{u,2} := \psi_{u}|_{\Omega_2(u)} \in H^2(\Omega_2(u))$, and $\psi_{u}$ is a strong solution to the transmission problem~\eqref{psi}.
\item[(b)] Given $\kappa>0$, there is $c(\kappa)>0$ such that, for every $u\in\bar{\mathcal{S}}$ satisfying $\|u\|_{H^2(D)}\le \kappa$,
\begin{equation*}
\|\psi_u\|_{H^1(\Omega(u))} + \|\psi_{u,1}\|_{H^2(\Omega_1(u))} + \|\psi_{u,2}\|_{H^2(\Omega_2(u))} \le c(\kappa)\,.
\end{equation*}
\end{itemize}
\end{theorem}
It is worth emphasizing that, for $i\in\{1,2\}$, the restriction of $\psi_u$ to $\Omega_i(u)$ belongs to $H^2(\Omega_i(u))$ for all $u\in\bar{\mathcal{S}}$. In particular, there is no regularity breakdown when the coincidence set $\mathcal{C}(u)$ is non-empty. A similar observation is made in \cite{ARMA20} for a different geometric setting when one of the two subsets does not depend on the function $u$.
\begin{remark}\label{rem.clamped}
When the upper part $\Omega_2(v)$ is clamped at its lateral boundaries in the sense that
\begin{equation*}
u\in H_0^2(D) := \{v\in H^2(D)\cap H_0^1(D)\,:\, \partial_x v(\pm L) = 0\}\,,
\end{equation*}
Theorem~\ref{Thm1} applies whatever the values of $\sigma_1$ and $\sigma_2$.
\end{remark}
Theorem~\ref{Thm1} is an immediate consequence of Proposition~\ref{ACDC} below. Its proof begins with quantitative $H^2$-estimates on $\psi_u$ depending only on $\|u\|_{H^2(D)}$ for sufficiently smooth functions in $\mathcal{S}$, the $H^2$-regularity of $\psi_u$ being guaranteed by \cite{Lem77} in that case. Since the class of functions for which these estimates are valid is dense in $\bar{\mathcal{S}}$, we complete the proof with a compactness argument, the main difficulty to be faced being the dependence of $\Omega(u)$ on $u$.
\noindent More precisely, we begin with a variational approach to \eqref{psi} and first show in Section~\ref{sec.vs} by classical arguments that, given $u\in \bar{\mathcal{S}}$, the variational solution $\psi_{u}$ to \eqref{psi} corresponds to the minimizer on $h_{u}+H_{0}^1(\Omega(u))$ of the associated Dirichlet energy
\begin{equation*}
\mathcal{J}(u)[\theta]:= \frac{1}{2}\int_{\Omega(u)} \sigma |\nabla \theta|^2\,\rd (x,z)\,, \qquad \theta \in h_u + H_0^1(\Omega(u))\,.
\end{equation*}
Thanks to this characterization, we use $\Gamma$-convergence tools to show the $H^1$-stability of $\psi_u$ with respect to $u$ in Section~\ref{sec.h1s}. Section~\ref{sec.h2r} is devoted to the study of the $H^2$-regularity of $\psi_u$ which we first establish in Section~\ref{sec.h2r1} for smooth functions $u\in \mathcal{S}\cap W_\infty^2(D)$ (thus having an empty coincidence set), relying on the analysis performed in \cite{Lem77}. It is worth mentioning that the constraint involving $\llbracket \sigma\rrbracket$ in the definition of $\mathcal{S}$ comes into play here. For $u\in \mathcal{S}\cap W_\infty^2(D)$, we next derive quantitative $H^2$-estimates on $\psi_u$ which only depend on $\|u\|_{H^2(D)}$ as stated in Theorem~\ref{Thm1}~(b), see Section~\ref{sec.h2e}. The building block is an identity in the spirit of \cite[Lemma~4.3.1.2]{Gr85} allowing us to interchange derivatives with respect to $x$ and $z$ in some integrals involving second-order derivatives, its proof being provided in Appendix~\ref{sec.id}. We then combine these estimates with the already proved $H^1$-stability of variational solutions to \eqref{psi} and use a compactness argument to extend the $H^2$-regularity of $\psi_u$ to arbitrary functions $u\in \bar{\mathcal{S}}$ in Section~\ref{sec.h2e1}. In this step, special care is required to cope with the variation of the functional spaces with $u$. In fact, as a side product of the proof of Theorem~\ref{Thm1}, we obtain qualitative information on the continuous dependence of $\psi_u$ with respect to $u$, which we collect in the next result.
\begin{theorem}\label{Thm2}
Suppose \eqref{bobbybrown} below.
Let $\kappa>0$, $u\in \bar{\mathcal{S}}$, and consider a sequence $(u_n)_{n\ge 1}$ in $\bar{\mathcal{S}}$ such that
\begin{equation}
\|u_n\|_{H^2(D)}\le \kappa\,, \quad n\ge 1\,, \qquad \lim_{n\to\infty} \|u_n-u\|_{H^1(D)} = 0\,. \label{y0}
\end{equation}
Setting $M := d + \max\left\{ \|u\|_{L_\infty(D)} \,,\, \sup_{n\ge 1}\{\|u_n\|_{L_\infty(D)}\} \right\}$,
\begin{subequations}\label{y}
\begin{equation}
\lim_{n\to \infty} \big\| (\psi_{u_n} - h_{u_n}) -(\psi_{u} - h_{u}) \big\|_{H^1(\Omega_M)} = 0\,. \label{y1}
\end{equation}
In addition, if $i\in\{1,2\}$ and $U_i$ is an open subset of $\Omega_i(u)$ such that $\bar{U}_i$ is a compact subset of $\Omega_i(u)$, then
\begin{equation}
\psi_{u_n,i}\rightharpoonup \psi_{u,i} \quad\text{in}\quad H^2(U_i)\,. \label{y2}
\end{equation}
Also, for any $p\in [1,\infty)$,
\begin{equation}
\begin{split}
\lim_{n\to\infty} \big\|\nabla\psi_{u_n,2}(\cdot,u_n) - \nabla\psi_{u,2}(\cdot,u) \big\|_{L_p(D,\R^2)} & = 0 \,, \\
\lim_{n\to\infty} \big\| \nabla\psi_{u_n,2}(\cdot,u_n+d) - \nabla\psi_{u,2}(\cdot,u+d) \big\|_{L_p(D,\R^2)} & = 0 \,.
\end{split} \label{y3}
\end{equation}
\end{subequations}
\end{theorem}
Clearly, the quantity $M$ introduced in Theorem~\ref{Thm2} is finite due to \eqref{y0} and the continuous embedding of $H^1(D)$ in $C(\bar{D})$.
\medskip
\paragraph{\textbf{Notation}}
Given $v\in \bar{\mathcal{S}}$, $f\in L_2(\Omega(v))$, and $i\in\{1,2\}$, we denote the restriction of $f$ to $\Omega_i(v)$ by $f_i$; that is, $f_i := f|_{\Omega_i(v)}$.
Throughout the paper, $c$ and $(c_k)_{k\ge 1}$ denote positive constants depending only on $L$, $H$, $d$, $V$, $\sigma_1$, and $\sigma_2$. The dependence upon additional parameters will be indicated explicitly.
\section{The Boundary Values}\label{sec.abv}
We state the precise assumptions on the function $h_v$ occurring in \eqref{a1c}. Roughly speaking, we assume that it is the trace on $\partial\Omega(v)$ of a function $h_v\in H^1(\Omega(u))$ which is such that $h|_{\Omega_i(v)}$ belongs to $H^2(\Omega_i(v))$ for $i=1,2$ and satisfies the transmission conditions \eqref{a1b}, as well as suitable boundedness and continuity properties with respect to $u$.
Specifically, for every $v\in \bar{\mathcal{S}}$, let
$$
h_v: D\times (-H,\infty)\rightarrow \R
$$
be such that
\begin{subequations}\label{bobbybrown}
\begin{equation}\label{200}
h_v\in H^1(\Omega(v))\,,\qquad h_{v,i}:=h_v\vert_{\Omega_i(v)}\in H^2\big(\Omega_i(v)\big)\,,\quad i=1,2\,,
\end{equation}
and suppose that $h_v$ satisfies the transmission conditions
\begin{equation}\label{201}
\llbracket h_v\rrbracket=\llbracket \sigma\nabla h_v\rrbracket\cdot {\bf n}_{\Sigma(v)}=0 \ \text{ on }\ \Sigma(v)\,.
\end{equation}
For $\kappa>0$ given, there is $c(\kappa)>0$ such that, for all $v\in \bar{\mathcal{S}}$ satisfying $\|v\|_{H^2(D)}\le \kappa$,
\begin{equation}\label{202}
\| h_{v,i}\|_{H^2(\Omega_i(v))}\le c(\kappa)\,, \quad i=1,2\,.
\end{equation}
Moreover, given $v\in\bar{\mathcal{S}}$ and a sequence $(v_n)_{n\ge 1}$ in $\bar{\mathcal{S}}$ satisfying
\begin{equation*}
\lim_{n\to \infty} \|v_n - v\|_{H^1(D)} = 0\,,
\end{equation*}
we assume that
\begin{equation}\label{204}
\lim_{n\rightarrow\infty} \| h_{v_n} - h_{v}\|_{H^1(D\times (-H,M))}=0
\end{equation}
and
\begin{equation}\label{205}
\lim_{n\rightarrow \infty} \|h_{v_n}(\cdot, v_n+d) - h_{v}(\cdot, v+d)\|_{C(\bar{D})} = 0\,,
\end{equation}
where
\begin{equation*}
M := d + \max\left\{ \|v\|_{L_\infty(D)} \,,\, \sup_{n\ge 1}\{\|v_n\|_{L_\infty(D)}\} \right\}<\infty\,.
\end{equation*}
\end{subequations}
Observe that the convergence of $(v_n)_{n\ge 1}$, the continuous embedding of $H^1(D)$ in $C(\bar{D})$, and \eqref{204} imply that
\begin{equation}\label{203}
\lim_{n\rightarrow\infty} \int_{\Omega(v_n)}\sigma\vert\nabla h_{v_n}\vert^2\,\rd (x,z)= \int_{\Omega(v)}\sigma\vert\nabla h_{v}\vert^2\,\rd (x,z)\,.
\end{equation}
From now on, we impose the conditions~\eqref{bobbybrown} throughout.
\medskip
We finish this short section by providing an example of $h_v$ satisfying the imposed conditions~\eqref{bobbybrown}.
\begin{example}\label{ex1}
Let $\zeta\in C^2(\R)$ be such that $\zeta|_{(-\infty,1]}\equiv 0$ and $\zeta|_{[1+d,\infty)} \equiv V$ for some $V>0$. Given $v\in\bar{\mathcal{S}}$, put
\begin{equation}\label{exx1}
h_v(x,z) :=\zeta(z- v(x)+1)\,, \qquad -H \le z \,, \quad x\in \bar{D}\,.
\end{equation}
Then \eqref{200}-\eqref{205} are satisfied. In addition,
$$
h_v(x,-H)=0\,, \quad h_v(x,v(x)+d)=V\,,\qquad x\in D\,.
$$
In the context of a MEMS device alluded to in the introduction, these additional properties mean that the ground plate and the top of the elastic plate are kept at constant potential. For instance, $\zeta(r):=V\min\{1,(r-1)^2/d^2\}$ for $r>1$ and $\zeta\equiv 0$ on $(-\infty,1]$ will do.
\end{example}
\section{Variational Solution to \eqref{psi}}\label{sec.vs}
In this section we investigate the properties of the variational solution $\psi_v$ to~\eqref{psi} for~$v\in \bar{\mathcal{S}}$ and, in particular, its $H^1$-stability.
\subsection{A Variational Approach to \eqref{psi}}\label{sec.min}
Given $v\in \bar{\mathcal{S}}$ we introduce the set of admissible potentials
$$
\mathcal{A}(v):=h_{v}+H_{0}^1(\Omega(v))\,,
$$
on which we define the functional
\begin{equation}\label{sos}
\mathcal{J}(v)[\theta]:=\frac{1}{2}\int_{ \Omega(v)} \sigma \vert\nabla \theta\vert^2\,\rd (x,z)\,,\qquad \theta\in \mathcal{A}(v)\,.
\end{equation}
The variational solution $\psi_v$ to the transmission problem \eqref{psi} is then the minimizer of the functional $\mathcal{J}(v)$ on the set $\mathcal{A}(v)$:
\begin{lemma}\label{L1}
For each $v\in \bar{\mathcal{S}}$ there is a unique minimizer $\psi_v \in \mathcal{A}(v)$ of $\mathcal{J}(v)$ on $\mathcal{A}(v)$; that is,
\begin{equation}
\mathcal{J}(v)[\psi_v] = \min_{\theta\in \mathcal{A}(v)} \mathcal{J}(v)[\theta]\,. \label{z0}
\end{equation}
In addition,
\begin{equation}
\int_{\Omega(v)}\sigma\vert\nabla\psi_v\vert^2\,\rd(x,z)\le \int_{\Omega(v)}\sigma\vert\nabla h_v\vert^2\,\rd(x,z)\,. \label{z1}
\end{equation}
\end{lemma}
\begin{proof}
Let $v\in \bar{\mathcal{S}}$ and recall that $h_v\in H^1(\Omega(v))$ according to \eqref{200}. Thus, the existence of a minimizer $\psi_v$ of $\mathcal{J}(v)$ on $\mathcal{A}(v)$ readily follows from the direct method of calculus of variations due to the lower semicontinuity and coercivity of $\mathcal{J}(v)$ on $\mathcal{A}(v)$, the latter being ensured by the assumption $\sigma\ge \min\{\sigma_1,\sigma_2\}>0$ and Poincar\'e's inequality. The uniqueness of $\psi_v$ is guaranteed by the strict convexity of $\mathcal{J}(v)$. Next, since obviously $h_v\in \mathcal{A}(v)$, the inequality \eqref{z1} is an immediate consequence of the minimizing property \eqref{z0} of $\psi_v$.
\end{proof}
For further use, we report the following version of Poincar\'e's inequality for functions in $H^1_0(\Omega(v))$ with a constant depending mildly on $v\in\bar{\mathcal{S}}$.
\begin{lemma}\label{LP}
Let $v\in\bar{\mathcal{S}}$ and $\theta\in H_0^1(\Omega(v))$. Then
\begin{equation*}
\|\theta\|_{L_2(\Omega(v))} \le 2 \|H+d+v\|_{L_\infty(D)} \|\partial_z \theta\|_{L_2(\Omega(v))}\,.
\end{equation*}
\end{lemma}
\begin{proof}
For $x\in D$ and $z\in (-H,v(x)+d)$,
\begin{equation*}
\theta(x,z)^2 = 2 \int_{-H}^z \theta(x,y) \partial_z\theta(x,y)\ \mathrm{d}y\,.
\end{equation*}
Hence, after integration with respect to $(x,z)$ over $\Omega(v)$,
\begin{align*}
\|\theta\|_{L_2(\Omega(v))}^2 & = \int_{\Omega(v)} \theta(x,z)^2\ \mathrm{d}(x,z) \\
& \le 2 \|H+d+v\|_{L_\infty(D)} \int_{\Omega(v)} |\theta(x,y)| |\partial_z \theta(x,y)|\ \mathrm{d}(x,z) \\
& \le 2 \|H+d+v\|_{L_\infty(D)} \|\theta\|_{L_2(\Omega(v))} \|\partial_z \theta\|_{L_2(\Omega(v))}\,,
\end{align*}
from which we deduce the stated inequality.
\end{proof}
\subsection{$H^1$-Stability of $\psi_v$}\label{sec.h1s}
The purpose of this section is to study the continuity properties of the solution $\psi_v$ to \eqref{z0} with respect to $v$. More precisely, we aim at establishing the following result.
\begin{proposition}\label{C3}
Consider $v\in \bar{\mathcal{S}}$ and a sequence $(v_n)_{n\ge 1}$ in $\bar{\mathcal{S}}$ such that
\begin{equation}\label{o1}
v_n\rightarrow v \ \text{ in }\ H_0^1(D)\,,
\end{equation}
and set
\begin{equation}
M := d + \max\left\{ \|v\|_{L_\infty(D)} \,,\, \sup_{n\ge 1}\{\|v_n\|_{L_\infty(D)}\} \right\}\,, \label{z10}
\end{equation}
which is finite by \eqref{o1} and the continuous embedding of $H^1(D)$ in $C(\bar{D})$. Then
$$
\lim_{n\to\infty} \left\| (\psi_{v_n}-h_{v_n}) - (\psi_v - h_v) \right\|_{H_0^1(D\times (-H,M))} = 0
$$
and
$$
\lim_{n\to\infty} \mathcal{J}(v_n)[\psi_{v_n}] = \mathcal{J}(v)[\psi_v]\,.
$$
\end{proposition}
To prove Proposition~\ref{C3}, we make use of a $\Gamma$-convergence approach and argue as in \cite[Section~3.2]{ARMA20} with minor changes. For the sake of completeness we provide a complete proof in Appendix~\ref{sec.pp3.3}.
\section{$H^2$-Regularity}\label{sec.h2r}
In the previous section we introduced the variational solution $\psi_v\in H^1(\Omega(v)$ to \eqref{psi} for arbitrary $v\in\bar{\mathcal{S}}$ and noticed its continuous dependence in $H^1(\Omega(v)$ with respect to $v$. We now aim at improving the $H^1$-regularity of $\psi_v|_{\Omega_i(v)}$ to $H^2(\Omega_i(v))$ for $i=1,2$. To this end we first consider the case of smooth functions $v\in \mathcal{S} \cap W_\infty^2(D)$ with empty coincidence sets and provide in Section~\ref{sec.h2r1} and Section~\ref{sec.h2e} the corresponding $H^2$-estimates that depend only on the norm of $v$ in $H^2(D)$ (but {\it not} on its $W_\infty^2(D)$-norm). In Section~\ref{sec.h2e1} we extend these estimates to the general case $v\in\bar{\mathcal{S}}$ by means of a compactness argument.
\subsection{$H^2$-Regularity for $v\in \mathcal{S}\cap W_\infty^2(D)$}\label{sec.h2r1}
Assuming that $v$ is smoother with an empty coincidence set, see Figure~\ref{Fig1}, the existence of a strong solution $\psi_v$ to \eqref{psi} is a consequence of the analysis performed in \cite{Lem77}.
\begin{proposition}\label{prz1}
If $v\in \mathcal{S} \cap W_\infty^2(D)$, then the variational solution $\psi_v$ to \eqref{z0} satisfies
\begin{equation*}
\psi_{v,i} := \psi_v|_{\Omega_i(v)} \in H^2(\Omega_i(v))\,, \quad i=1,2\,,
\end{equation*}
and the transmission problem
\begin{subequations}\label{psiS}
\begin{align}
\mathrm{div}(\sigma\nabla\psi_v)&=0 \quad\ \text{in }\ \Omega(v)\,,\label{a1aS}\\
\llbracket \psi_v \rrbracket =\llbracket \sigma\nabla \psi_v \rrbracket \cdot \mathbf{n}_{ \Sigma(v)} &=0\quad\ \text{on }\ \Sigma(v)\,,\label{a1bS}\\
\psi_v&=h_v\quad\text{on }\ \partial\Omega(v)\,.\label{a1cS}
\end{align}
\end{subequations}
Moreover, $\partial_x \psi_v + \partial_x v \partial_z \psi_v$ and $-\sigma\partial_x v \partial_x \psi_v +\sigma \partial_z\psi_v$ both belong to $H^1(\Omega(v))$.
\end{proposition}
Besides \cite{Lem77}, the proof of Proposition~\ref{prz1} requires the following auxiliary result.
\begin{lemma}\label{L0}
Let $v\in\bar{\mathcal{S}}$ and consider $\phi\in L_2(\Omega(v))$ such that
\begin{equation*}
\phi_i := \phi|_{\Omega_i(v)} \in H^1(\Omega_i(v))\,, \quad i=1,2\,,
\end{equation*}
and $\llbracket \phi\rrbracket=0$ on $\Sigma(v)$. Then $\phi\in H^1(\Omega(v))$ and
\begin{equation}
\|\phi\|_{H^1(\Omega(v))} \le \|\phi_1\|_{H^1(\Omega_1(v))} + \|\phi_2\|_{H^1(\Omega_2(v))} \,. \label{z2}
\end{equation}
\end{lemma}
\begin{proof}
We set $e_x=(1,0)$ and $e_z=(0,1)$. Given $\theta\in C_c^\infty\big(\Omega(v)\big)$ and $j\in \{x,z\}$ we note that
\begin{equation*}
\begin{split}
\int_{\Omega(v)}\phi \partial_j\theta \,\rd (x,z) & = \int_{\Omega(v)} \mathrm{div}(\phi\theta e_j)\, \rd (x,z) - \sum_{i=1}^2 \int_{\Omega_i(v)} \theta \partial_j \phi_i\,\rd (x,z)\\
&=\int_{\Sigma(v)} \llbracket \phi\rrbracket\,\theta e_j\cdot \mathbf{n}_{ \Sigma(v)}\,\rd \sigma_{\Sigma(v)} - \sum_{i=1}^2 \int_{\Omega_i(v)} \theta \partial_j \phi_i\,\rd (x,z)\,,
\end{split}
\end{equation*}
due to Gau\ss' theorem. Thus, since $\llbracket \phi\rrbracket=0$ on $\Sigma(v)$,
$$
\left\vert \int_{\Omega(v)}\phi \partial_j\theta\,\rd (x,z)\right\vert\le\big(\Vert\phi_1\Vert_{H^1(\Omega_1(v))}+\Vert\phi_2\Vert_{H^1(\Omega_2(v))}\big)\,\Vert\theta\Vert_{L_2(\Omega(v))}\,,
$$
for $j=x,z$ and $\theta\in C_c^\infty\big(\Omega(v)\big)$.
Consequently, $\phi\in H^1(\Omega(v))$.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{prz1}]
We check that the transmission problem \eqref{psiS} fits into the framework of \cite{Lem77}. Since $v\in \mathcal{S} \cap W_\infty^2(D)$ and $v(\pm L) = 0$, the boundaries of $\Omega_1(v)$ and $\Omega_2(v)$ are $W_\infty^2$-smooth curvilinear polygons and the interface $\Sigma(v)$ meets the boundary $\partial\Omega(v)$ of $\Omega(v)$ at the vertices $A_\pm := (\pm L,0)$. Moreover, at the vertex $A_\pm$, the measures $\omega_{\pm,1}$ and $\omega_{\pm,2}$ of the angles between $-e_z$ and $(1,\mp \partial_x v(\pm L))$ and between $(1,\mp \partial_x v(\pm L))$ and $e_z$, respectively, satisfy $\omega_{\pm,1}+\omega_{\pm,2} = \pi$, as well as
\begin{equation*}
\begin{split}
\omega_{\pm,2} & \ge \frac{\pi}{2} \;\text{ if }\; \llbracket \sigma \rrbracket < 0\,, \\
\omega_{\pm,2} & \le \frac{\pi}{2} \;\text{ if }\; \llbracket \sigma \rrbracket > 0\,,
\end{split}
\end{equation*}
by definition of $\mathcal{S}$. According to the analysis performed in \cite{Lem77}, these conditions guarantee that the variational solution $\psi_v$ to \eqref{z0} provided by Lemma~\ref{L1} satisfies $\psi_{v,i} = \psi_v|_{\Omega_i(v)} \in H^2(\Omega_i(v))$ for $i=1,2$ and solves the transmission problem \eqref{psi} in a strong sense.
Next, owing to the just established $H^2$-regularity of $\psi_{v,1}$ and $\psi_{v,2}$, we may differentiate with respect to $x$ the transmission condition $\llbracket \psi_v \rrbracket(x,v(x))=0$, $x\in D$, and find that
\begin{equation*}
\llbracket \partial_x \psi_v + \partial_x v \partial_z \psi_v \rrbracket = 0 \quad\ \text{ on }\ \Sigma(v)\,.
\end{equation*}
The stated $H^1$-regularity of $\partial_x \psi_v + \partial_x v \partial_z \psi_v$ then follows from Lemma~\ref{L0} and the boundedness of $\partial_x v$ and $\partial_x^2 v$. In the same vein, due to \eqref{a1b}, the regularity of $v$, and the identity
\begin{equation*}
\frac{\llbracket - \sigma \partial_x v \partial_x \psi_v + \sigma \partial_z \psi_v \rrbracket}{\sqrt{1+(\partial_x v)^2}} = \llbracket \sigma\nabla \psi_v \rrbracket \cdot \mathbf{n}_{ \Sigma(v)} = 0\,,
\end{equation*}
the claimed $H^1$-regularity of $- \sigma \partial_x v \partial_x \psi_v + \sigma \partial_z \psi_v$ is again a consequence of Lemma~\ref{L0} and the boundedness of $\partial_x v$ and $\partial_x^2 v$.
\end{proof}
\subsection{$H^2$-Estimates on $\psi_v$ for $v\in \mathcal{S} \cap W_\infty^2(D)$}\label{sec.h2e}
The $H^2$-regularity of $\psi_v$ being guaranteed by Proposition~\ref{prz1} for $v\in \mathcal{S}\cap W_\infty^2(D)$, the next step is to show that this property extends to any $v\in\bar{\mathcal{S}}$. To this end, we shall now derive quantitative $H^2$-estimates on $\psi_v$, paying special attention to their dependence upon the regularity of $v$. As in \cite{ARMA20}, it turns out to be more convenient to study a non-homogeneous transmission problem with homogeneous Dirichlet boundary conditions instead of \eqref{psiS}. Specifically, for $v\in \mathcal{S}\cap W_\infty^2(D)$, we define
\begin{equation}
\chi=\chi_v:=\psi_v-h_v\in H_{0}^1(\Omega(v))\,, \label{chi0}
\end{equation}
where $\psi_v\in H^1(\Omega(v))$ is the unique solution to \eqref{psiS} provided by Proposition~\ref{prz1}. Since $\psi_{v,i}=\psi_v|_{\Omega_i(v)}$ belongs to $H^2(\Omega_i(v))$ for $i=1, 2$, we readily infer from \eqref{200} and \eqref{chi0} that
\begin{equation}\label{chi}
\chi_{i} := \chi_{v}|_{\Omega_i(v)}\in H^2(\Omega_i(v))\,, \quad i=1,2\,.
\end{equation}
We omit in the following the dependence of $\chi$ on $v$ for ease of notation.
According to \eqref{200}, \eqref{201}, and Proposition~\ref{prz1}, $\chi$ solves the transmission problem
\begin{subequations}\label{a2}
\begin{align}
\mathrm{div}(\sigma\nabla\chi) & = - \mathrm{div}(\sigma\nabla h_v) \quad\text{in }\ \Omega(v)\,,\label{a2a}\\
\llbracket \chi \rrbracket = \llbracket \sigma\nabla \chi\rrbracket\cdot {\bf n}_{\Sigma(v)} &=0\quad\text{on }\ \Sigma(v)\,,\label{a2b}\\
\chi&=0\quad\text{on }\ \partial\Omega(v)\,, \label{a2c}
\end{align}
\end{subequations}
and it follows from \eqref{200} that it is equivalent to derive $H^2$-estimates on $(\psi_{v,1},\psi_{v,2})$ or $(\chi_1,\chi_2)$.
For that purpose, we transform \eqref{a2} to a transmission problem on the rectangle $\mathcal{R}:=D\times (0,1+d)$. More precisely, we introduce the transformation
\begin{equation}\label{t1}
T_1(x,z):=\left(x,\frac{z+H}{v(x)+H}\right)\,,\qquad (x,z)\in \Omega_1(v)\,,
\end{equation}
mapping $\Omega_1(v)$ onto the rectangle \mbox{$\mathcal{R}_1:=D\times (0,1)$}, and the transformation
\begin{equation}\label{t2}
T_2(x,z):=\left(x,z-v(x)+1\right)\,,\qquad (x,z)\in \Omega_2(v)\,,
\end{equation}
mapping $\Omega_2(v)$ onto the rectangle \mbox{$\mathcal{R}_2:=D\times (1,1+d)$}. The interface separating $\mathcal{R}_1$ and $\mathcal{R}_2$ is
$$
\Sigma_0:=D\times \{1\}\,,
$$
so that
$$
\mathcal{R}=D\times (0,1+d)=\mathcal{R}_1\cup \mathcal{R}_2\cup \Sigma_0\,.
$$
It is worth pointing out here that $T_1$ is well-defined due to $v\in \mathcal{S}$. Let $(x,\eta)$ denote the new variables in $\mathcal{R}$; that is, $(x,\eta)=T_1(x,z)$ for $(x,z)\in \mathcal{R}_1$ and $(x,\eta)=T_2(x,z)$ for $(x,z)\in \mathcal{R}_2$.
Then, \eqref{chi} implies
\begin{equation}\label{C0}
\Phi:=\Phi_1 \mathbf{1}_{\mathcal{R}_1} + \Phi_2 \mathbf{1}_{\mathcal{R}_2}\in H_0^1(\mathcal{R})\,,\qquad
\Phi_i:= \chi_i\circ (T_i)^{-1}\in H^2(\mathcal{R}_i)\,,\quad i=1,2\,.
\end{equation}
For further use, we also introduce
$$
\hat\sigma(x,\eta):=\left\{\begin{array}{cl} \dfrac{\sigma_1}{v(x)+H}\,, & (x,\eta)\in \mathcal{R}_1\,,\\
\hphantom{x}\vspace{-3.5mm}\\
\sigma_2 \,, & (x,\eta)\in \mathcal{R}_2\,,
\end{array}\right.
$$
and derive the following fundamental identity for $\Phi$, which provides a connection between some integrals involving products of second-order derivatives of $\Phi$ and is in the spirit of \cite[Lemma~4.3.1.2]{Gr85}, \cite[Lemma~3.4]{ARMA20}, and \cite[Lemme~II.2.2]{Lem77}.
\begin{lemma}\label{L2a}
Given $v\in \mathcal{S} \cap W_\infty^2(D)$, the function $\Phi$ defined in \eqref{C0} satisfies
\begin{align*}
\sum_{i=1}^2 \int_{\mathcal{R}_i} \hat{\sigma} \partial_x^2\Phi_i\, \partial_\eta^2\Phi_i\, \rd (x,\eta) & = \sum_{i=1}^2 \int_{\mathcal{R}_i} \hat{\sigma} |\partial_{x}\partial_{\eta}\Phi_i|^2\,\rd (x,\eta) \\
& \qquad - \sigma_1 \int_{\mathcal{R}_1} \frac{\partial_x v}{(v+H)^2} \partial_\eta\Phi_1 \partial_{x}\partial_{\eta}\Phi_1\, \rd (x,\eta)\\
& \qquad +\frac{1}{2}\int_D\frac{\partial_x^2 v \big((\partial_x v)^2-1\big)}{(1+(\partial_x v)^2)^2} \, \left\llbracket\sigma(\partial_x\Phi )^2\right\rrbracket (x,1)\,\rd x\,.
\end{align*}
\end{lemma}
\begin{proof}
We adapt the proof of \cite[Lemma~3.4]{ARMA20} and \cite[Lemme~II.2.2]{Lem77}. Note that \eqref{a2b}, \eqref{t1}, \eqref{t2}, and \eqref{C0} imply $\llbracket \Phi\rrbracket=0$ on $\Sigma_0$, so that
\begin{equation}\label{i}
\llbracket \partial_x\Phi\rrbracket=0\quad\text{ on }\ \Sigma_0\,.
\end{equation}
Consequently, since $(\partial_x\Phi_1,\partial_x\Phi_2)$ lies in $H^1(\mathcal{R}_1)\times H^1(\mathcal{R}_2)$ by \eqref{C0}, we may argue as in the proof of Lemma~\ref{L0} and deduce from \eqref{i} that
$$
F:=\partial_x\Phi \in H^1(\mathcal{R})\,.
$$
Moreover, by \eqref{C0},
\begin{equation}
F(x,0) = F(x,1+d) = 0\,, \qquad x\in D\,. \label{z4}
\end{equation}
Similarly, setting
$$
G:=-\sigma\frac{\partial_x v}{1+(\partial_x v)^2}\partial_x\Phi+\hat\sigma\partial_\eta\Phi
$$
we derive from \eqref{C0} that $G_i := G|_{\mathcal{R}_i}\in H^1(\mathcal{R}_i)$ for $i=1,2$, while \eqref{a2b}, \eqref{t1}, \eqref{t2}, and \eqref{C0} imply that, for $x\in D$,
\begin{align*}
G_1(x,1) & = \frac{\sigma_1}{\sqrt{1+(\partial_x v(x))^2}} \left[ - \partial_x v(x) \partial_x \chi_1(x,v(x)) + \partial_z \chi_1(x,v(x)) \right] \\
& = \frac{\sigma_2}{\sqrt{1+(\partial_x v(x))^2}} \left[ - \partial_x v(x) \partial_x \chi_2(x,v(x)) + \partial_z \chi_2(x,v(x)) \right] = G_2(x,1)\,;
\end{align*}
that is, $\llbracket G\rrbracket=0$ on~$\Sigma_0$, and we argue as in the proof of Lemma~\ref{L0} to conclude that
$$
G \in H^1(\mathcal{R})\,.
$$
In addition, by \eqref{C0},
\begin{align*}
G(\pm L, \eta) & = - \sigma (\pm L,\eta) \left( \frac{\partial_x v}{1 + (\partial_x v)^2} \right)(\pm L) \partial_x \Phi(\pm L,\eta) + \hat{\sigma}(\pm L,\eta) \partial_\eta \Phi(\pm L,\eta) \\
& = - \sigma (\pm L,\eta) \left( \frac{\partial_x v}{1 + (\partial_x v)^2} \right)(\pm L) \partial_x \Phi(\pm L,\eta)
\end{align*}
for $\eta\in (0,1+d)$. Hence,
\begin{equation}
G(\pm L, \eta) + \sigma (\pm L,\eta) \left( \frac{\partial_x v}{1 + (\partial_x v)^2} \right)(\pm L) F(\pm L,\eta)= 0\,, \qquad \eta\in (0,1+d)\,. \label{z5}
\end{equation}
Owing to \eqref{z4}, \eqref{z5}, and the $H^1$-regularity of $F$ and $G$, we are in a position to apply Lemma~\ref{ID} (see Appendix~\ref{sec.id}) with
\begin{equation*}
(V,W)=(F,G) \;\;\text{ and }\;\; \tau^{\pm} = \sigma \left( \frac{\partial_x v}{1+(\partial_x v)^2} \right)(\pm L)\,,
\end{equation*}
to obtain the identity
\begin{equation}\label{DI}
\int_\mathcal{R}\partial_x F\partial_\eta G\,\rd (x,\eta)=\int_\mathcal{R}\partial_{\eta}F\partial_x G\,\rd (x,\eta)\,.
\end{equation}
Using the definitions of $F$ and $G$, the identity \eqref{DI} reads
\begin{equation*}
\begin{split}
\sum_{i=1}^2\int_{\mathcal{R}_i} \partial_x^2\Phi_i &\left(-\sigma\frac{\partial_x v}{1+(\partial_x v)^2} \partial_{x}\partial_{\eta}\Phi_i + \hat{\sigma} \partial_\eta^2 \Phi_i \right) \,\rd (x,\eta)\\
& = \sum_{i=1}^2\int_{\mathcal{R}_i}\partial_{x}\partial_{\eta}\Phi_i \left( -\sigma\frac{\partial_x v}{1+(\partial_x v)^2} \partial_{x}^2\Phi_i - \sigma \frac{\partial_x^2 v [1 -(\partial_x v)^2]}{[1+(\partial_x v)^2]^2} \partial_x\Phi_i \right)\, \rd (x,\eta) \\
& \qquad + \sum_{i=1}^2\int_{\mathcal{R}_i}\partial_{x}\partial_{\eta}\Phi_i \Big( \partial_x\hat\sigma \partial_\eta\Phi_i +\hat\sigma\partial_{x}\partial_{\eta}\Phi_i \Big) \,\rd (x,\eta)\,.
\end{split}
\end{equation*}
Noticing that the first terms on both sides of the above identity are the same and that
$$
\partial_x\Phi_i\partial_{x}\partial_{\eta}\Phi_i=\frac{1}{2} \partial_\eta\big((\partial_x\Phi_i)^2\big)
$$
implies that
\begin{equation*}
\begin{split}
\sum_{i=1}^2\int_{\mathcal{R}_i} \sigma & \frac{\partial_x^2 v \big[(\partial_x v)^2)-1\big]}{[1+(\partial_x v)^2]^2} \partial_x\Phi_i\partial_{x}\partial_{\eta}\Phi_i \,\rd (x,\eta)\\
& =\frac{1}{2}\int_D\frac{\partial_x^2 v \big[(\partial_x v)^2-1\big]}{[1+(\partial_x v)^2]^2} \, \left\llbracket\sigma(\partial_x\Phi )^2\right\rrbracket (x,1)\,\rd x\,,
\end{split}
\end{equation*}
the assertion follows, recalling that $\partial_x \hat{\sigma}=0$ in $\mathcal{R}_2$.
\end{proof}
\begin{remark}\label{rem.gr}
If $\partial_x v(\pm L)=0$, then \eqref{z5} reduces to $G(\pm L,\eta) = 0$ for $\eta\in (0,1+d)$ and the crucial identity~\eqref{DI} used in the proof of Lemma~\ref{L2a} directly follows from \cite[Lemma~4.3.1.2]{Gr85}. For the general case $v\in \mathcal{S}$, we require the extension given in Lemma~\ref{ID}.
\end{remark}
We now translate the outcome of Lemma~\ref{L2a} in terms of the solution $\chi$ to \eqref{a2}.
\begin{lemma}\label{L2}
Let $v\in \mathcal{S} \cap W_\infty^2(D)$. The solution $\chi=\psi_v-h_v$ to \eqref{a2} satisfies
\begin{equation*}
\begin{split}
\sum_{i=1}^2 \int_{\Omega_i(v)}\sigma\,\partial_x^2\chi_i\,\partial_z^2\chi_i\,\rd (x,z) & = \sum_{i=1}^2 \int_{\Omega_i(v)}\sigma |\partial_{x}\partial_{z}\chi_i|^2\,\rd (x,z) \\
& \qquad - \frac{\sigma_2}{2} \int_D \partial_x^2 v(x)\, \big( \partial_z\chi_2(x,v(x)+d) \big)^2\, \rd x\\
& \qquad -\frac{1}{2}\int_D\frac{\partial_x^2 v(x)}{1+(\partial_x v(x))^2}\,\left\llbracket \sigma\vert\nabla\chi\vert^2\right\rrbracket \big(x,v(x)\big)\,\rd x\,.
\end{split}
\end{equation*}
\end{lemma}
\begin{proof}
Let us first recall the regularity of $\Phi$ stated in \eqref{C0} which validates the subsequent computations. Using the transformations $T_1$ and $T_2$ introduced in \eqref{t1} and \eqref{t2}, respectively, we obtain
\begin{align*}
\sum_{i=1}^2 \int_{\Omega_i(v)}\sigma\,&\partial_x^2\chi_i\,\partial_z^2\chi_i\,\rd (x,z)\\
=& \int_{\mathcal{R}_1}\frac{\sigma_1}{v+H} \bigg[\partial_x^2\Phi_1 + \eta \Big( 2\Big(\frac{\partial_x v}{v+H}\Big)^2 - \frac{\partial_x^2 v}{v+H}\Big) \, \partial_\eta\Phi_1 - 2 \eta \frac{\partial_x v}{v+H}\partial_{x}\partial_{\eta}\Phi_1\\
&\hspace{4cm} +\eta^2\Big(\frac{\partial_x v}{v+H}\Big)^2\partial_\eta^2\Phi_1\bigg] \,\partial_\eta^2\Phi_1\,\rd (x,\eta)\\
&+\int_{\mathcal{R}_2}\sigma_2\Big[\partial_x^2\Phi_2-2\partial_xv\partial_{x}\partial_{\eta}\Phi_2-\partial_x^2 v\partial_\eta\Phi_2 +(\partial_x v)^2 \partial_\eta^2\Phi_2 \Big] \, \partial_\eta^2\Phi_2\,\rd (x,\eta) \\
=& \sum_{i=1}^2 \int_{\mathcal{R}_i} \hat{\sigma} \partial_x^2\Phi_i \,\partial_\eta^2\Phi_i\,\rd (x,\eta)\\
& +\int_{\mathcal{R}_1}\frac{\sigma_1}{v+H} \bigg[ \eta \Big( 2 \Big(\frac{\partial_x v}{v+H}\Big)^2 - \frac{\partial_x^2 v}{v+H} \Big) \, \partial_\eta\Phi_1 - 2\eta\frac{\partial_x v}{v+H} \partial_{x}\partial_{\eta}\Phi_1\\
&\hspace{4cm} +\eta^2\Big(\frac{\partial_x v}{v+H}\Big)^2\partial_\eta^2\Phi_1\bigg] \,\partial_\eta^2\Phi_1\,\rd (x,\eta)\\
&+\int_{\mathcal{R}_2}\sigma_2\Big[-2\partial_xv\partial_{x}\partial_{\eta}\Phi_2-\partial_x^2 v\partial_\eta\Phi_2 +(\partial_x v)^2 \partial_\eta^2\Phi_2 \Big] \, \partial_\eta^2\Phi_2\,\rd (x,\eta)\,.
\end{align*}
We use Lemma~\ref{L2a} to express the first integral on the right-hand side and get
\begin{align}
& \sum_{i=1}^2 \int_{\Omega_i(v)}\sigma\, \partial_x^2\chi_i\,\partial_z^2\chi_i\,\rd (x,z) \nonumber \\
& \qquad = \int_{\mathcal{R}_1} \hat{\sigma} |\partial_{x}\partial_{\eta}\Phi_1|^2 \,\rd (x,\eta) + \int_{\mathcal{R}_2} \hat{\sigma} |\partial_{x}\partial_{\eta}\Phi_2|^2 \,\rd (x,\eta) \nonumber \\
& \qquad\qquad +\int_{\mathcal{R}_1} \frac{\sigma_1}{v+H} \bigg[ -\frac{\partial_x v}{v+H} \partial_{\eta}\Phi_1\partial_{x}\partial_{\eta}\Phi_1 - 2 \eta \frac{\partial_x v}{v+H} \partial_{x}\partial_{\eta}\Phi_1 \partial_\eta^2\Phi_1 \nonumber \\
&\hspace{4cm} + \eta^2 \Big(\frac{\partial_x v}{v+H}\Big)^2 \big\vert\partial_\eta^2\Phi_1\big\vert^2 + 2\eta \Big(\frac{\partial_x v}{v+H}\Big)^2 \,\partial_\eta\Phi_1 \partial_\eta^2\Phi_1 \nonumber \\
& \hspace{8cm} - \eta \frac{\partial_x^2 v}{v+H}\,\partial_\eta\Phi_1 \partial_\eta^2\Phi_1 \bigg] \, \,\rd (x,\eta) \nonumber \\
& \qquad\qquad +\int_{\mathcal{R}_2}\sigma_2\Big[-2\partial_x v \partial_{x}\partial_{\eta}\Phi_2 \partial_\eta^2\Phi_2-\partial_x^2 v\partial_\eta\Phi_2\partial_\eta^2\Phi_2 +(\partial_x v)^2\big\vert\partial_\eta^2\Phi_2\big\vert^2\Big]\,\rd (x,\eta) \nonumber \\
& \qquad\qquad +\frac{1}{2}\int_D\frac{\partial_x^2 v \big((\partial_x v)^2-1\big)}{(1+(\partial_x v)^2)^2} \, \left\llbracket\sigma(\partial_x\Phi )^2\right\rrbracket (x,1)\,\rd x \label{E1}\,.
\end{align}
We then compute separately the integrals over $\mathcal{R}_i$, $i=1,2$, and begin with the contribution of $\mathcal{R}_1$. We complete the square to get
\begin{align*}
I_1& := \int_{\mathcal{R}_1} \frac{\sigma_1}{v+H} \bigg[ \vert\partial_{x}\partial_{\eta}\Phi_1\vert^2 -\frac{\partial_x v}{v+H} \partial_{\eta}\Phi_1\partial_{x}\partial_{\eta}\Phi_1 - 2 \eta \frac{\partial_x v}{v+H} \partial_{x}\partial_{\eta}\Phi_1 \partial_\eta^2\Phi_1 \\
&\hspace{3cm} +\eta^2 \Big(\frac{\partial_x v}{v+H}\Big)^2 \big\vert\partial_\eta^2\Phi_1\big\vert^2 + 2\eta \Big(\frac{\partial_x v}{v+H}\Big)^2 \,\partial_\eta\Phi_1 \partial_\eta^2\Phi_1 \\
& \hspace{8cm} - \eta \frac{\partial_x^2 v}{v+H}\,\partial_\eta\Phi_1 \partial_\eta^2\Phi_1 \bigg] \, \,\rd (x,\eta) \\
& = \int_{\mathcal{R}_1} \frac{\sigma_1}{v+H}\bigg[ \big| \partial_{x}\partial_{\eta}\Phi_1 \big|^2 + \Big(\frac{\partial_x v}{v+H}\Big)^2 \big| \partial_\eta\Phi_1 \big|^2 + \eta^2\Big(\frac{\partial_x v}{v+H}\Big)^2 \big| \partial_\eta^2\Phi_1\big|^2 \\
& \hspace{4cm} -2\eta\frac{\partial_x v}{v+H} \partial_{x}\partial_{\eta}\Phi_1 \partial_\eta^2\Phi_1 + 2\eta\Big(\frac{\partial_x v}{v+H}\Big)^2\,\partial_\eta\Phi_1 \partial_\eta^2\Phi_1 \\
& \hspace{8cm} - 2 \frac{\partial_x v}{v+H} \partial_{\eta}\Phi_1 \partial_{x}\partial_{\eta}\Phi_1 \bigg]\ \rd (x,\eta)\\
& \qquad + \int_{\mathcal{R}_1} \frac{\sigma_1}{v+H} \bigg[ - \Big(\frac{\partial_x v}{v+H}\Big)^2 \big| \partial_\eta\Phi_1 \big|^2 + \frac{\partial_x v}{v+H} \partial_{\eta}\Phi_1 \partial_{x}\partial_{\eta}\Phi_1 \\
& \qquad \hspace{6cm} - \eta \frac{\partial_x^2 v}{v+H} \partial_\eta\Phi_1 \partial_\eta^2 \Phi_1\bigg]\ \rd (x,\eta) \\
& = \int_{\mathcal{R}_1} \sigma_1 (v+H) \bigg[ \frac{\partial_{x}\partial_{\eta}\Phi_1}{v+H} -\frac{\partial_x v}{(v+H)^2} \partial_{\eta}\Phi_1-\eta \frac{\partial_x v}{(v+H)^2}\, \partial_\eta^2\Phi_1\bigg]^2\,\rd (x,\eta)\\
&\qquad +\int_{\mathcal{R}_1} \sigma_1 \partial_x v\,\bigg[\frac{1}{(v+H)^2}\partial_{\eta}\Phi_1\partial_{x}\partial_{\eta}\Phi_1 -\frac{\partial_x v}{(v+H)^3}\big(\partial_\eta\Phi_1\big)^2\bigg] \,\rd (x,\eta)\\
&\qquad - \int_{\mathcal{R}_1} \sigma_1 \frac{\partial_x^2 v}{(v+H)^2}\,\eta\,\partial_\eta\Phi_1 \partial_\eta^2\Phi_1 \,\rd (x,\eta)\,.
\end{align*}
Thanks to the identities
\begin{equation*}
\frac{1}{(v+H)^2}\partial_\eta\Phi_1\partial_{x}\partial_{\eta}\Phi_1 - \frac{\partial_x v}{(v+H)^3} \big(\partial_\eta\Phi_1\big)^2 = \frac{1}{2} \partial_x \left( \left( \frac{\partial_\eta\Phi_1}{v+H} \right)^2\right)\,,
\end{equation*}
\begin{equation*}
\partial_\eta\Phi_1\partial_\eta^2\Phi_1=\frac{1}{2} \partial_\eta \big(\partial_\eta\Phi_1\big)^2\,
\end{equation*}
and the property $\partial_\eta\Phi_1(\pm L,\eta)= 0$ for $\eta\in (0,1)$ stemming from \eqref{C0}, we may perform integration by parts in the last two integrals on the right-hand side of the previous identity and obtain
\begin{align*}
I_1 & =\int_{\mathcal{R}_1} \sigma_1 (v+H)\bigg[\frac{\partial_{x}\partial_{\eta}\Phi_1 }{v+H} -\frac{\partial_x v}{(v+H)^2} \partial_{\eta}\Phi_1-\eta \frac{\partial_x v}{(v+H)^2}\, \partial_\eta^2\Phi_1\bigg]^2\,\rd (x,\eta)\\
&\qquad -\frac{\sigma_1}{2}\int_{D}\frac{\partial_x^2 v}{(v+H)^2}\,\big(\partial_\eta\Phi_1(x,1)\big)^2\,\rd x\,.
\end{align*}
Transforming the above identity back to $\Omega_1(v)$ yields
\begin{equation}\label{R1}
I_1=\int_{\Omega_1(v)}\sigma_1\big\vert\partial_{x}\partial_{z}\chi_1\big\vert^2\,\rd (x,z) - \frac{\sigma_1}{2} \int_{D} \partial_x^2 v(x)\,\big( \partial_z\chi_1 (x,v(x))\big)^2\,\rd x\,.
\end{equation}
Next, arguing in a similar way,
\begin{align*}
I_2 := &\, \sigma_2 \int_{\mathcal{R}_2}\Big[ \big\vert \partial_x\partial_\eta \Phi_2\big|^2 -2\partial_x v \partial_{x}\partial_{\eta}\Phi_2 \partial_\eta^2\Phi_2-\partial_x^2 v\partial_\eta\Phi_2\partial_\eta^2\Phi_2 +(\partial_x v)^2\big\vert\partial_\eta^2\Phi_2\big\vert^2\Big]\,\rd (x,\eta) \\
= &\, \sigma_2 \int_{\mathcal{R}_2} \Big[\big| \partial_{x}\partial_{\eta}\Phi_2 \big|^2 - 2 \partial_x v\partial_{x}\partial_{\eta}\Phi_2 \partial_\eta^2\Phi_2+(\partial_x v)^2\big\vert\partial_\eta^2\Phi_2\big\vert^2\Big]\,\rd (x,\eta) \\
& \qquad - \frac{\sigma_2}{2}\int_{\mathcal{R}_2}\partial_x^2 v \partial_\eta \left( \partial_\eta \Phi_2 \right)^2 \,\rd (x,\eta) \\
= &\, \sigma_2 \int_{\mathcal{R}_2} \Big[\partial_{x}\partial_{\eta}\Phi_2 - \partial_x v \partial_\eta^2\Phi_2\Big]^2\,\rd (x,\eta) - \frac{\sigma_2}{2}\int_{D} \partial_x^2 v(x)\,\big( \partial_\eta\Phi_2 (x,1+d)\big)^2\,\rd x\\
& \qquad +\frac{\sigma_2 }{2}\int_{D} \partial_x^2 v(x)\,\big( \partial_\eta\Phi_2 (x,1)\big)^2\,\rd x\,.
\end{align*}
Transforming this formula back to $\Omega_2(v)$ yields
\begin{equation}\label{R2}
\begin{split}
I_2 &= \sigma_2 \int_{\Omega_2(v)} \big|\partial_{x}\partial_{z}\chi_2\big|^2\,\rd (x,z)
- \frac{\sigma_2}{2} \int_{D} \partial_x^2 v(x)\,\big( \partial_z\chi_2 (x,v(x)+d)\big)^2\,\rd x\\
&\qquad + \frac{\sigma_2}{2}\int_{D} \partial_x^2 v(x)\,\big( \partial_z\chi_2 (x,v(x))\big)^2\,\rd x \,.
\end{split}
\end{equation}
Finally,
\begin{align*}
& \int_D\frac{\partial_x^2 v \big[(\partial_x v)^2-1\big]}{[1+(\partial_x v)^2]^2} \, \left\llbracket\sigma(\partial_x\Phi )^2\right\rrbracket (x,1)\,\rd x \\
& \qquad = \int_D\frac{\partial_x^2 v \big[(\partial_x v)^2-1\big]}{[1+(\partial_x v)^2]^2} \, \left\llbracket \sigma(\partial_x\chi + \partial_x v \partial_z \chi )^2\right\rrbracket (x,1)\,\rd x\,,
\end{align*}
and we deduce from \eqref{E1}, \eqref{R1}, \eqref{R2}, and the above identity that
\begin{equation}\label{E2}
\begin{split}
\sum_{i=1}^2 \int_{\Omega_i(v)}\sigma\,&\partial_x^2\chi_i\,\partial_z^2\chi_i\,\rd (x,z)\\
=& \sum_{i=1}^2 \int_{\Omega_i(v)}\sigma\,\big|\partial_{x}\partial_{z}\chi_i\big|^2\,\rd (x,z)
-\frac{\sigma_2}{2}\int_{D} \partial_x^2 v(x)\,\big( \partial_z\chi_2 (x,v(x)+d)\big)^2\,\rd x\\
&-\frac{1}{2}\int_{D} \partial_x^2 v(x)\,\left\llbracket\sigma \big(\partial_z\chi_2\big)^2\right\rrbracket (x,v(x))\,\rd x\\
&+\frac{1}{2}\int_D\frac{\partial_x^2 v \big[(\partial_x v)^2-1\big]}{[1+(\partial_x v)^2]^2} \, \left\llbracket\sigma\big(\partial_x\chi+\partial_x v\partial_z\chi\big)^2\right\rrbracket (x,v(x))\,\rd x\,.
\end{split}
\end{equation}
It remains to simplify the last two integrals on the right-hand side of \eqref{E2}. To this end, we first recall that the regularity of $\chi$ allows us to differentiate with respect to $x$ the transmission condition $\llbracket \chi\rrbracket =0$ on $\Sigma(v)$ to deduce that
\begin{equation}
\llbracket \partial_x \chi + \partial_x v \partial_z \chi \rrbracket = 0 \;\text{ on }\; \Sigma(v)\,, \label{z6}
\end{equation}
while the second transmission condition in \eqref{a2b} reads
\begin{equation}
\llbracket \sigma \big( \partial_x v \partial_x \chi - \partial_z \chi \big) \rrbracket = 0 \;\text{ on }\; \Sigma(v)\,. \label{z7}
\end{equation}
In particular, \eqref{z6} and \eqref{z7} imply that, on $\Sigma(v)$,
\begin{align*}
\llbracket \sigma \big( \partial_x v \partial_x \chi - \partial_z \chi \big) \big( \partial_x \chi + \partial_x v \partial_z \chi \big) \rrbracket & = \big( \partial_x \chi_1 + \partial_x v \partial_z \chi_1 \big) \llbracket \sigma \big( \partial_x v \partial_x \chi - \partial_z \chi \big) \rrbracket \\
& \qquad + \sigma_2 \big( \partial_x v \partial_x \chi_2 - \partial_z \chi_2 \big) \llbracket \big( \partial_x \chi + \partial_x v \partial_z \chi \big) \rrbracket \\
& = 0\,.
\end{align*}
Therefore,
\begin{align*}
J & := \big[ (\partial_x v)^2-1 \big] \Big\llbracket \sigma \big(\partial_x\chi+\partial_x v\partial_z\chi\big)^2 \Big\rrbracket - [1+(\partial_x v)^2]^2 \Big\llbracket \sigma \big( \partial_z \chi \big)^2 \Big\rrbracket \\
& = \big[ (\partial_x v)^2-1 \big] \Big\llbracket \sigma \big(\partial_x\chi+\partial_x v\partial_z\chi\big)^2 \Big\rrbracket - [1+(\partial_x v)^2]^2 \Big\llbracket \sigma \big( \partial_z \chi \big)^2 \Big\rrbracket \\
& \qquad - 2\partial_x v \Big\llbracket \sigma \big( \partial_x v \partial_x \chi - \partial_z \chi \big) \big( \partial_x \chi + \partial_x v \partial_z \chi \big) \Big\rrbracket \\
& = \Big\llbracket \sigma \big[ (\partial_x v)^2 - 1 - 2(\partial_x v)^2 \big] \big( \partial_x\chi \big)^2 \Big\rrbracket \\
& \qquad + \Big\llbracket \sigma \big[ 2 \partial_x v \big( (\partial_x v)^2-1 \big) - 2 (\partial_x v)^3 + 2 \partial_x v \big] \partial_x \chi \partial_z \chi \Big\rrbracket \\
& \qquad + \Big\llbracket \sigma \big[ (\partial_x v)^2 \big( (\partial_x v)^2 - 1\big) + 2 (\partial_x v)^2 - [1+(\partial_x v)^2]^2 \big] \big( \partial_z \chi \big)^2 \Big\rrbracket \\
& = - \big[ 1 + (\partial_x v)^2 \big] \Big\llbracket \sigma \big( \partial_x \chi \big)^2 + \sigma \big( \partial_z \chi \big)^2 \Big\rrbracket \\
& = - \big[ 1 + (\partial_x v)^2 \big] \Big\llbracket \sigma |\nabla\chi|^2 \Big\rrbracket\,.
\end{align*}
Hence,
\begin{equation}
\frac{(\partial_x v)^2-1}{[1+(\partial_x v)^2]^2} \Big\llbracket \sigma \big(\partial_x\chi+\partial_x v\partial_z\chi\big)^2 \Big\rrbracket - \Big\llbracket \sigma \big( \partial_z \chi \big)^2 \Big\rrbracket = - \frac{1}{1+(\partial_x v)^2} \left\llbracket \sigma |\nabla\chi|^2 \right\rrbracket\,. \label{17}
\end{equation}
Consequently, \eqref{E2} and \eqref{17} entail
\begin{equation*}
\begin{split}
\sum_{i=1}^2 \int_{\Omega_i(v)}\sigma\,\partial_x^2\chi_i\,\partial_z^2\chi_i\,\rd (x,z) & = \sum_{i=1}^2 \int_{\Omega_i(v)}\sigma |\partial_{x}\partial_{z}\chi_i|^2\,\rd (x,z) \\
& \quad - \frac{1}{2}\int_D \sigma_2\partial_x^2 v(x)\,\big(\partial_z\chi_2(x,v(x)+d)\big)^2\,\rd x\\
& \quad -\frac{1}{2}\int_D\frac{\partial_x^2 v(x)}{1+(\partial_x v(x))^2}\,\left\llbracket \sigma |\nabla\chi|^2\right\rrbracket \big(x,v(x)\big)\,\rd x\,,
\end{split}
\end{equation*}
as claimed.
\end{proof}
In order to estimate the boundary and the transmission terms in Lemma~\ref{L2}, we first report the following trace estimates.
\begin{lemma}\label{L3}
Given $\kappa>0$ and $\alpha\in (0,1/2]$, there is $c(\alpha,\kappa)>0$ such that, for any $v\in\bar{\mathcal{S}}$ satisfying $\|v\|_{H^2(D)}\le \kappa$ and $\theta\in H^1(\Omega_2(v))$,
\begin{equation*}
\|\theta_2(\cdot,v)\|_{H^\alpha(D)}+ \|\theta_2(\cdot,v+d)\|_{H^\alpha(D)}\le c(\alpha,\kappa)\, \|\theta_2\|_{L_2(\Omega_2(v))}^{(1-2\alpha)/2}\, \|\theta_2\|_{H^1(\Omega_2(v))}^{(2\alpha+1)/2}\,.
\end{equation*}
\end{lemma}
\begin{proof}
Let $\theta\in H^1(\Omega_2(v))$. Using the transformation $T_2$ defined in \eqref{t2} which maps $\Omega_2(v)$ onto the rectangle $\mathcal{R}_2 =D\times (1,1+d)$, we note that $\phi:=\theta\circ T_2^{-1}$ belongs to $H^1(\mathcal{R}_2)$ with
\begin{equation}
\|\phi\|_{L_2(\mathcal{R}_2)} = \|\theta\|_{L_2(\Omega_2(v))} \label{n2}
\end{equation}
and
\begin{equation*}
\|\nabla\phi \|_{L_2(\mathcal{R}_2)}^2 = \|\partial_x \theta + \partial_x v \partial_z \theta \|_{L_2(\Omega_2(v))}^2 + \|\partial_z \theta \|_{L_2(\Omega_2(v))}^2\,,
\end{equation*}
so that the continuous embedding of $H^2(D)$ in $W_\infty^1(D)$ and the assumed bound on $v$ readily imply that
\begin{equation}
\|\phi\|_{H^1(\mathcal{R}_2)} \le c(\kappa) \|\theta\|_{H^1(\Omega_2(v))} \,. \label{n1}
\end{equation}
By complex interpolation,
$$
[L_2(\mathcal{R}_2 ),H^1(\mathcal{R}_2 )]_{\alpha+1/2}\doteq H^{\alpha+1/2} (\mathcal{R}_2 )\,,
$$
from which we deduce that
$$
\|\phi\|_{H^{\alpha+1/2}(\mathcal{R}_2 )}\le c(\alpha) \|\phi\|_{L_2(\mathcal{R}_2 )}^{(1-2\alpha)/2} \|\phi\|_{H^1(\mathcal{R}_2 )}^{(2\alpha+1)/2}\,.
$$
Since $\alpha>0$, the trace maps $H^{\alpha+1/2}(\mathcal{R}_2 )$ continuously on $H^{\alpha}(D\times \{1\})$, and we thus infer from \eqref{n2} and \eqref{n1} that
\begin{equation*}
\begin{split}
\|\theta(\cdot,v)\|_{H^{\alpha}(D)} &=\|\phi(\cdot, 1) \|_{H^{\alpha}(D)}\le c(\alpha) \|\phi\|_{H^{\alpha+1/2}(\mathcal{R}_2)} \\
& \le c(\alpha)\|\phi\|_{L_2(\mathcal{R}_2 )}^{(1-2\alpha)/2} \|\phi\|_{H^1(\mathcal{R}_2 )}^{(2\alpha+1)/2}\\
&\le c(\alpha,\kappa)\|\theta\|_{L_2(\Omega_2(v))}^{(1-2\alpha)/2}\|\theta\|_{H^1(\Omega_2(v))}^{(2\alpha+1)/2}\,.
\end{split}
\end{equation*}
The estimate for $\|\theta(\cdot,v+d)\|_{H^{\alpha}(D)} $ is proved in a similar way.
\end{proof}
Based on Lemma~\ref{L3} we are in a position to estimate the boundary and transmission terms in the identity provided by Lemma~\ref{L2}.
\begin{lemma}\label{L4}
Let $\zeta\in (3/4,1)$ and $\kappa>0$. There is $c(\zeta,\kappa)>0$ such that, if $v\in \mathcal{S} \cap W_\infty^2(D)$ satisfies $\|v\|_{H^2(D)}\le \kappa$, then the solution $\chi=\chi_v$ to \eqref{a2} satisfies
\begin{equation}\label{a}
\begin{split}
& \left| \frac{\sigma_2}{2} \int_D\partial_x^2 v(x) \big(\partial_z\chi_2(x,v(x)+d) \big)^2 \,\rd x \right| \\
& \hspace{3cm} \le c(\zeta,\kappa) \, \|\partial_z\chi_2\|_{L_2(\Omega_2(v))}^{2(1-\zeta)} \, \|\partial_z\chi_2\|_{H^1(\Omega_2(v))}^{2\zeta}
\end{split}
\end{equation}
and
\begin{equation}\label{b}
\begin{split}
& \left| \frac{1}{2} \int_D \frac{\partial_x^2 v(x)}{1+(\partial_x v(x))^2}\, \left\llbracket \sigma\vert\nabla\chi\vert^2 \right\rrbracket \big(x,v(x)\big)\,\rd x\right| \\
& \hspace{3cm} \le c(\zeta,\kappa)\, \|\nabla\chi_2\|_{L_2(\Omega_2(v))}^{2(1-\zeta)}\, \|\nabla\chi_2\|_{H^1(\Omega_2(v))}^{2\zeta}\,.
\end{split}
\end{equation}
\end{lemma}
\begin{proof}
To prove \eqref{a}, let us first note that $H^{\zeta-1/2}(D)$ embeds continuously into $L_4(D)$. We use the Cauchy-Schwarz inequality and Lemma~\ref{L3} with $\alpha=\zeta-1/2$ and deduce
\begin{equation*}
\begin{split}
\left| \frac{\sigma_2}{2}\int_D \partial_x^2 v(x) \big(\partial_z\chi_2(x,v(x)+d)\big)^2\, \rd x\right|& \le \frac{\sigma_2}{2} \|\partial_x^2 v\|_{L_2(D)}\, \|\partial_z\chi_2(\cdot,v+d)\|_{L_4(D)}^2\\
& \le c(\kappa)\, \|\partial_z\chi_2(\cdot,v+d)\|_{H^{\zeta-1/2}(D)}^2\\
& \le c(\zeta,\kappa)\, \|\partial_z\chi_2\|_{L_2(\Omega_2(v))}^{2(1-\zeta)}\, \|\partial_z\chi_2\|_{H^1(\Omega_2(v))}^{2\zeta}\,.
\end{split}
\end{equation*}
As for \eqref{b} we obtain analogously
\begin{align}
&\left| \frac{\sigma_2}{2}\int_D \frac{\partial_x^2 v(x)}{1+(\partial_x v(x))^2}\, \left[ \big(\partial_x\chi_2(x,v(x))\big)^2 + \big(\partial_z\chi_2(x,v(x))\big)^2 \right] \,\rd x\right|\nonumber \\
& \qquad\qquad \le \frac{\sigma_2}{2} \|\partial_x^2 v\|_{L_2(D)} \|\nabla \chi_2(\cdot,v)\|_{L_4(D)}^2 \nonumber \\
& \qquad\qquad \le c(\zeta,\kappa)\, \|\nabla\chi_2\|_{L_2(\Omega_2(v))}^{2(1-\zeta)}\, \|\nabla\chi_2\|_{H^1(\Omega_2(v))}^{2\zeta} \label{pp}
\end{align}
and
\begin{equation}
\begin{split}
&\left| \frac{\sigma_1}{2}\int_D \frac{\partial_x^2 v(x)}{1+(\partial_x v(x))^2}\, \left[ \big(\partial_x\chi_1(x,v(x))\big)^2 + \big(\partial_z\chi_1(x,v(x))\big)^2 \right] \,\rd x\right| \\
& \qquad\qquad \le \frac{\sigma_1}{2} \|\partial_x^2 v\|_{L_2(D)} \|\nabla \chi_1(\cdot,v)\|_{L_4(D)}^2 \,. \label{z8}
\end{split}
\end{equation}
At this point, we use \eqref{z6} and \eqref{z7} to show that
\begin{align*}
\partial_x \chi_1 & = \frac{\sigma_1+\sigma_2(\partial_x v)^2}{\sigma_1\big(1+(\partial_x v)^2\big)} \partial_x \chi_2 +\frac{\llbracket\sigma\rrbracket\partial_x v}{\sigma_1\big(1+(\partial_x v)^2\big)}\partial_z\chi_2 \quad\text{on }\ \Sigma(v)\,, \\
\partial_z \chi_1 & = \frac{\llbracket\sigma\rrbracket\partial_x v}{\sigma_1\big(1+(\partial_x v)^2\big)}\partial_x\chi_2 + \frac{\sigma_1+\sigma_2(\partial_x v)^2}{\sigma_1\big(1+(\partial_x v)^2\big)} \partial_z \chi_2 \quad\text{on }\ \Sigma(v)\,.
\end{align*}
Consequently,
\begin{align*}
|\partial_x \chi_1| & \le \frac{\max\{\sigma_1,\sigma_2\}}{\sigma_1} \left( |\partial_x \chi_2| + |\partial_z \chi_2| \right) \quad\text{on }\ \Sigma(v)\,, \\
|\partial_z \chi_1| & \le \frac{\max\{\sigma_1,\sigma_2\}}{\sigma_1} \left( |\partial_x \chi_2| + |\partial_z \chi_2| \right) \quad\text{on }\ \Sigma(v)\,,
\end{align*}
so that
\begin{equation*}
\|\nabla \chi_1(\cdot,v)\|_{L_4(D)} \le c \|\nabla \chi_2(\cdot,v)\|_{L_4(D)}\,.
\end{equation*}
Owing to \eqref{z8} and the above inequality, we may then argue as in the proof of \eqref{pp} to conclude that
\begin{align*}
&\left| \frac{\sigma_1}{2}\int_D \frac{\partial_x^2 v(x)}{1+(\partial_x v(x))^2}\, \left[ \big(\partial_x\chi_1(x,v(x))\big)^2 + \big(\partial_z\chi_1(x,v(x))\big)^2 \right] \,\rd x\right| \\
& \qquad\qquad \le c(\zeta,\kappa)\, \|\nabla\chi_2\|_{L_2(\Omega_2(v))}^{2(1-\zeta)}\, \|\nabla\chi_2\|_{H^1(\Omega_2(v))}^{2\zeta}\,,
\end{align*}
as claimed in \eqref{b}.
\end{proof}
We now gather the previous findings to deduce the following crucial $H^2$-estimate on the solution $\psi_v$ to \eqref{psiS} for $v\in \mathcal{S}\cap W_\infty^2(D)$, which only depends on the $H^2(D)$-norm of $v$ (but not on its $W_\infty^2(D)$-norm).
\begin{proposition}\label{P2}
Let $\kappa>0$ and $v\in\mathcal{S}\cap W_\infty^2(D)$ be such that $\|v\|_{H^2(D)}\le \kappa$. There is a constant $c_0(\kappa)>0$ such that the solution $\psi_v$ to \eqref{psiS} satisfies
\begin{subequations}\label{est}
\begin{equation}\label{est1}
\|\chi\|_{H^1(\Omega(v))} + \|\chi_1\|_{H^2(\Omega_1(v))} + \|\chi_2\|_{H^2(\Omega_2(v))}\le c_0(\kappa)\,,
\end{equation}
and
\begin{equation}\label{est2}
\|\psi_v\|_{H^1(\Omega(v))}+\|\psi_{v,1}\|_{H^2(\Omega_1(v))}+\|\psi_{v,2}\|_{H^2(\Omega_2(v))} \le c_0(\kappa)\,,
\end{equation}
\end{subequations}
recalling that $\chi=\psi_v-h_v$ and $\chi_i = \chi|_{\Omega_i(v)}$, $i=1,2$.
\end{proposition}
\begin{proof}
Let $v\in \mathcal{S} \cap W_\infty^2(D)$ with $\|v\|_{H^2(D)}\le \kappa$. Since $\sigma$ is constant on $\Omega_1(v)$ and on $\Omega_2(v)$, it readily follows from \eqref{a2a} that
$$
\sum_{i=1}^2 \int_{\Omega_i(v)}\sigma |\Delta\chi_i|^2\, \rd (x,z)= \sum_{i=1}^2 \int_{\Omega_i(v)}\sigma | \Delta h_{v,i}|^2\, \rd (x,z)\,.
$$
Since
\begin{equation*}
|\Delta\chi_i|^2 = |\partial_x^2\chi_i|^2 + |\partial_z^2\chi_i|^2 + 2 \partial_x^2\chi_i \partial_z^2 \chi_i\,, \qquad i=1,2\,,
\end{equation*}
we infer from Lemma~\ref{L2} and the above two formulas that
\begin{align*}
& \sum_{i=1}^2 \int_{\Omega_i(v)} \sigma \big\{ |\partial_x^2\chi_i|^2+ 2|\partial_{x}\partial_{z}\chi_i|^2+|\partial_z^2\chi_i|^2\big\}\,\rd (x,z)\\
& \qquad = \sum_{i=1}^2 \int_{\Omega_i(v)}\sigma |\Delta h_{v,i}|^2\,\rd (x,z) + 2 \sum_{i=1}^2 \int_{\Omega_i(v)} \sigma \left( |\partial_x\partial_z \chi_i|^2 - \partial_x^2 \chi_i \partial_z^2\chi_i \right)\, \rd (x,z) \\
& \qquad \le \sum_{i=1}^2 \int_{\Omega_i(v)}\sigma |\Delta h_{v,i}|^2\,\rd (x,z) + \sigma_2 \int_D \partial_x^2 v(x) \,\big(\partial_z\chi_2(x,v(x)+d)\big)^2\,\rd x\\
&\qquad\qquad + \int_D \frac{\partial_x^2 v(x)}{1+(\partial_x v(x))^2}\, \left\llbracket \sigma |\nabla\chi|^2 \right\rrbracket \big(x,v(x)\big)\, \rd x\,.
\end{align*}
Using Lemma~\ref{L4} with $\zeta=7/8$, along with the identity
\begin{equation*}
\begin{split}
& \sum_{i=1}^2 \int_{\Omega_i(v)}\sigma\big\{|\partial_x^2\chi_i|^2+ 2|\partial_{x}\partial_{z}\chi_i|^2+|\partial_z^2\chi_i|^2\big\}\,\rd (x,z) \\
& \hspace{2cm} = \sigma_1 \|\nabla\chi_1\|_{H^1(\Omega_1(v))}^2 + \sigma_2 \|\nabla\chi_2\|_{H^1(\Omega_2(v))}^2\,,
\end{split}
\end{equation*}
we further obtain
\begin{align*}
& \sigma_1 \|\nabla\chi_1\|_{H^1(\Omega_1(v))}^2 + \sigma_2 \|\nabla\chi_2\|_{H^1(\Omega_2(v))}^2\\
& \qquad\qquad \le \sum_{i=1}^2 \int_{\Omega_i(v)}\sigma |\Delta h_{v,i}|^2\,\rd (x,z) + c(\kappa)\, \|\nabla\chi_2\|_{L_2(\Omega_2(v))}^{1/4}\, \|\nabla\chi_2\|_{H^1(\Omega_2(v))}^{7/4}\,.
\end{align*}
Hence, thanks to Young's inequality,
\begin{align*}
& \sigma_1 \|\nabla\chi_1\|_{H^1(\Omega_1(v))}^2 + \sigma_2 \|\nabla\chi_2\|_{H^1(\Omega_2(v))}^2\\
& \qquad\qquad\le \sum_{i=1}^2 \int_{\Omega_i(v)}\sigma |\Delta h_{v,i}|^2\,\rd (x,z) + \frac{\sigma_2}{2}\, \|\nabla\chi_2\|_{H^1(\Omega_2(v))}^2 + c(\kappa) \|\nabla\chi_2\|_{L_2(\Omega_2(v))}^2\,.
\end{align*}
Recalling that
\begin{align*}
\|\nabla \chi_2\|_{L_2(\Omega_2(v))}^2 & \le \frac{1}{\sigma_2} \int_{\Omega(v)} \sigma |\nabla\chi|^2\, \rd(x,z) \le \frac{1}{\sigma_2} \int_{\Omega(v)} \sigma |\nabla h_v|^2\, \rd (x,z) \\
& \le \frac{\max\{\sigma_1,\sigma_2\}}{\sigma_2} \|\nabla h_v\|_{L_2(\Omega(v))}^2
\end{align*}
by \eqref{a2} and that $\min\{\sigma_1,\sigma_2\}>0$, we conclude that
\begin{equation}
\begin{split}
& \|\nabla\chi_1\|_{H^1(\Omega_1(v))}^2 + \|\nabla\chi_2\|_{H^1(\Omega_2(v))}^2 \\
& \hspace{2cm} \le c(\kappa) \left( \|\Delta h_{v,1}\|_{L_2(\Omega_1(v))}^2 + \|\Delta h_{v,2}\|_{L_2(\Omega_2(v))}^2 + \|\nabla h_v\|_{L_2(\Omega(v))}^2 \right)\,.
\end{split} \label{25}
\end{equation}
Owing to the continuous embedding of $H^2(D)$ in $C(\bar{D})$, combining \eqref{25} and Lemma~\ref{LP} leads us to the estimate
\begin{equation*}
\begin{split}
& \|\chi\|_{H^1(\Omega(v))} + \|\chi_{1}\|_{H^2(\Omega_1(v))} + \|\chi_{2}\|_{H^2(\Omega_2(v))} \\
& \hspace{2cm} \le c(\kappa) \big(\|\nabla h_v \|_{L_2(\Omega(v))} + \|\Delta h_{v,1}\|_{L_2(\Omega_1(v))}^2 + \|\Delta h_{v,2}\|_{L_2(\Omega_2(v))}^2 \big)\,.
\end{split}
\end{equation*}
The bound~\eqref{est1} then readily follows from the assumptions~\eqref{200} and~\eqref{202}. Finally, \eqref{est1}, together with \eqref{200} and \eqref{202}, yields~\eqref{est2}.
\end{proof}
\subsection{$H^2$-Regularity and $H^2$-Estimates on $\psi_v$ for $v\in \bar{\mathcal{S}}$}\label{sec.h2e1}
Finally, we extend Proposition~\ref{prz1} and Proposition~\ref{P2} by showing the $H^2$-regularity of $\psi_v$ and the corresponding $H^2$-estimates for an arbitrary $v\in\bar{\mathcal{S}}$; that is, we drop the additional $W_\infty^2$-regularity of $v$ assumed in the previous sections and also allow for a non-empty coincidence set.
\begin{figure}
\begin{tikzpicture}[scale=0.9]
\draw[black, line width = 1.5pt, dashed] (-7,0)--(7,0);
\draw[black, line width = 2pt] (-7,0)--(-7,-2.5);
\draw[black, line width = 2pt] (7,-2.5)--(7,0);
\draw[black, line width = 2pt] (-7,-2.5)--(7,-2.5);
\draw[red, line width = 2pt] plot[domain=-7:-2.5] (\x,{-1-cos((pi*(11.5+\x)/4.5) r)});
\draw[red, line width = 2pt] plot[domain=-7:-2.5] (\x,{-1.5-cos((pi*(11.5+\x)/4.5) r)});
\draw[red, line width = 2pt] (-2.5,-2)--(-2,-2);
\draw[red, line width = 2pt] plot[domain=-2:3] (\x,{-2.0-0.5*cos((pi*(4+2*\x)/5) r)});
\draw[red, line width = 2pt] plot[domain=-2:3] (\x,{-1.5-0.5*cos((pi*(4+2*\x)/5) r)});
\draw[red, line width = 2pt] plot[domain=3:7] (\x,{-1-cos((pi*(11-\x)/4) r)});
\draw[red, line width = 2pt] plot[domain=3:7] (\x,{-1.5-cos((pi*(11-\x)/4) r)});
\draw[red, line width = 1pt, arrows=->] (2,0)--(2,-2.1);
\node at (2.3,-0.6) {${\color{red} w}$};
\node at (-5.5,-1.8) {${\color{red} \Omega_1(w)}$};
\node at (-3,0.5) {${\color{red} \Omega_2(w)}$};
\draw (-3.65,0.5) edge[->,bend right,line width = 1pt] (-5,-1.15);
\node at (0,-3.5) {$D$};
\node at (6,-1.75) {{\color{red} $\Sigma(w)$}};
\draw (5.5,-1.75) edge[->,bend left, line width = 1pt] (4.7,-1.8);
\node at (-7.8,1) {$z$};
\draw[black, line width = 1pt, arrows = ->] (-7.5,-3)--(-7.5,1);
\node at (-8,-2.5) {$-H$};
\draw[black, line width = 1pt] (-7.6,-2.5)--(-7.4,-2.5);
\node at (-7.8,-0.5) {$0$};
\draw[black, line width = 1pt] (-7.6,-0.5)--(-7.4,-0.5);
\node at (-7.8,0) {$d$};
\draw[black, line width = 1pt] (-7.6,0)--(-7.4,0);
\node at (-7,-3.25) {$-L$};
\node at (7,-3.25) {$L$};
\draw[black, line width = 1pt, arrows = <->] (-7,-2.75)--(7,-2.75);
\node at (-4,-3.25) {${\color{red} \mathcal{C}(w)}$};
\draw (-3.5,-3.25) edge[->,bend right, line width = 1pt] (3,-2.55);
\draw (-3.5,-3.25) edge[->,bend right,line width = 1pt] (-2.25,-2.55);
\draw[black, line width = 2pt] (-7,-2.5)--(7,-2.5);
\end{tikzpicture}
\caption{Geometry of $\Omega(w)$ for a state $w\in \mathcal{S}$ with non-empty and disconnected coincidence set.}\label{Fig3}
\end{figure}
\begin{proposition}\label{ACDC}
Let $\kappa>0$ and $v\in\bar{\mathcal{S}}$ be such that $\|v\|_{H^2(D)} \le \kappa$.
\begin{itemize}
\item [\textbf{(a)}] The unique minimizer $\psi_v\in \mathcal{A}(v)$ of $\mathcal{J}(v)$ on $\mathcal{A}(v)$ provided by Lemma~\ref{L1} satisfies
\begin{equation*}
\psi_{v,i} = \psi_v|_{\Omega_i(v)} \in H^2(\Omega_i(v))\,, \qquad i=1,2\,,
\end{equation*}
and is a strong solution to the transmission problem \eqref{psiS}. Moreover, there is $c_1(\kappa)>0$ such that
\begin{equation}\label{king3}
\|\psi_v\|_{H^1(\Omega(v))} + \|\psi_{v,1}\|_{H^2(\Omega_1(v))} + \|\psi_{v,2}\|_{H^2(\Omega_2(v))}\le c_1(\kappa)\,.
\end{equation}
\item[\textbf{(b)}] Consider a sequence $(v_n)_{n\ge 1}$ in $\bar{\mathcal{S}}$ satisfying
\begin{equation}\label{vh2}
\|v_n\|_{H^2(D)}\le \kappa\,, \quad n\ge 1\,, \;\;\text{ and }\;\; \lim_{n\to\infty} \|v_n-v\|_{H^1(D)} = 0 \,.
\end{equation}
If $i\in\{1,2\}$ and $U_i$ is an open subset of $\Omega_i(v)$ such that $\bar U_i$ is a compact subset of $\Omega_i(v)$, then
\begin{equation*}
\psi_{v_n,i}\rightharpoonup \psi_{v,i} \quad\text{in}\quad H^2(U_i)\,,
\end{equation*}
recalling that $\psi_{v_n,i} = \psi_{v_n}|_{\Omega_i(v_n)}$.
\end{itemize}
\end{proposition}
The proof involves three steps: we first establish Proposition~\ref{ACDC}~\textbf{(b)} under the additional assumption
\begin{equation*}
\sup_{n\ge 1}\left\{ \|\psi_{v_n,1}\|_{H^2(\Omega_1(v_n))} + \|\psi_{v_n,2}\|_{H^2(\Omega_2(v_n))} \right\} < \infty\,.
\end{equation*}
Building upon this result, we take advantage of the density of $\mathcal{S}\cap W_\infty^2(D)$ in $\bar{\mathcal{S}}$ and of the estimates derived in Proposition~\ref{P2} to verify Proposition~\ref{ACDC}~\textbf{(a)} by a compactness argument. Combining the previous steps leads us finally to a complete proof of Proposition~\ref{ACDC}~\textbf{(b)}. We thus start with the proof of Proposition~\ref{ACDC}~\textbf{(b)} when the solutions $(\psi_{v_n})_{n\ge 1}$ to \eqref{psiS} associated with the sequence $(v_n)_{n\ge 1}$ satisfies the above additional bound. We state this result as a separate lemma for definiteness.
\begin{lemma}\label{lez1}
Let $\kappa>0$ and $v\in\bar{\mathcal{S}}$ be such that $\|v\|_{H^2(D)} \le \kappa$ and consider a sequence $(v_n)_{n\ge 1}$ in $\bar{\mathcal{S}}$ satisfying \eqref{vh2}. Assume further that, for each $n\ge 1$, $(\psi_{v_n,1},\psi_{v_n,2})$ belongs to $H^2(\Omega_1(v_n))\times H^2(\Omega_2(v_n))$ and that there is $\mu>0$ such that
\begin{equation}
\|\psi_{v_n,1}\|_{H^2(\Omega_1(v_n))} + \|\psi_{v_n,2}\|_{H^2(\Omega_2(v_n))} \le \mu\,, \qquad n\ge 1\,. \label{uh2}
\end{equation}
Then $\psi_{v,i} \in H^2(\Omega_i(v))$, $i=1,2$. In addition, if $i\in\{1,2\}$ and $U_i$ is an open subset of $\Omega_i(v)$ such that $\bar U_i$ is a compact subset of $\Omega_i(v)$, then
\begin{equation*}
\psi_{v_n,i}\rightharpoonup \psi_{v,i} \quad\text{in}\quad H^2(U_i)
\end{equation*}
and
\begin{equation}
\|\psi_{v,1}\|_{H^2(\Omega_1(v))} + \|\psi_{v,2}\|_{H^2(\Omega_2(v))} \le \mu\,. \label{z101}
\end{equation}
\end{lemma}
The proof of Lemma~\ref{lez1} is very close to that of \cite[Proposition~3.13 \& Corollary~3.14]{ARMA20}. For the sake of completeness we provide a detailed proof in Appendix~\ref{pr.acdc}.
\begin{proof}[Proof of Proposition~\ref{ACDC}~\textbf{(a)}]
Let $v\in \bar{\mathcal{S}}$ be such that $\|v\|_{H^2(D)}\le \kappa$. We may choose a sequence $(v_n)_{n\ge 1}$ in $\mathcal{S}\cap W_\infty^2(D)$ satisfying
\begin{equation}
v_n\rightarrow v \ \text{ in }\ H^2(D)\,,\qquad \sup_{n\ge 1}\,\|v_n\|_{H^2(D)}\le 2\kappa\,. \label{z1103}
\end{equation}
Owing to \eqref{z1103} and the regularity property $v_n\in \mathcal{S}\cap W_\infty^2(D)$, $n\ge 1$, Proposition~\ref{C3} guarantees that $(\psi_{v_n,1},\psi_{v_n,2})$ belongs to $H^2(\Omega_1(v_n))\times H^2(\Omega_2(v_n))$ and $(\psi_{v_n})_{n\ge 1}$ satisfies \eqref{uh2} with $\mu=c_0(2\kappa)$. We then infer from Lemma~\ref{lez1} that $(\psi_{v,1},\psi_{v,2})$ belongs to $H^2(\Omega_1(v))\times H^2(\Omega_2(v))$ and satisfies
\begin{equation*}
\|\psi_{v,1}\|_{H^2(\Omega_1(v))} + \|\psi_{v,2}\|_{H^2(\Omega_2(v))} \le c_0(2\kappa)\,.
\end{equation*}
Combining the above bound with \eqref{204} and Lemma~\ref{lez1} gives \eqref{king3}. It remains to check that $\psi_v$ is a strong solution to \eqref{psiS} which can be done as in \cite[Corollary~3.14]{ARMA20}, see Appendix~\ref{pr.acdc} for details.
\end{proof}
\begin{proof}[Proof of Proposition~\ref{ACDC}~\textbf{(b)}]
Proposition~\ref{ACDC}~\textbf{(b)} is now a straightforward consequence of Proposition~\ref{ACDC}~\textbf{(a)} and Lemma~\ref{lez1}.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{Thm1}]
The proof of Theorem~\ref{Thm1} readily follows from Proposition~\ref{ACDC}~{\bf (a)}.
\end{proof}
We supplement the $H^2$-weak continuity of $\psi_v$ with respect to $v$ reported in Proposition~\ref{ACDC} with the continuity of the traces of $\nabla\psi_{v,2}$ on the upper and lower boundaries of $\Omega_2(v)$.
\begin{proposition}\label{U2}
Let $\kappa>0$ and $v\in\bar{\mathcal{S}}$ be such that $\|v\|_{H^2}(D)\le \kappa$ and consider a sequence $(v_n)_{n\ge 1}$ in~$\bar{\mathcal{S}}$ satisfying \eqref{vh2}. Then, for $p\in [1,\infty)$,
\begin{align}
\nabla\psi_{v_n,2}(\cdot,v_n) & \rightarrow \nabla\psi_{v,2}(\cdot,v) \quad \text{ in }\quad L_p(D,\R^2)\,, \label{x}\\
\nabla\psi_{v_n,2}(\cdot,v_n+d) & \rightarrow \nabla\psi_{v,2}(\cdot,v+d) \quad \text{ in }\quad L_p(D,\R^2)\,, \label{x1}
\end{align}
and
\begin{equation}\label{xx}
\|\nabla\psi_{v,2}(\cdot,v) \|_{L_p(D,\R^2)} + \|\nabla\psi_{v,2}(\cdot,v+d) \|_{L_p(D,\R^2)} \le c(p,\kappa)\,.
\end{equation}
\end{proposition}
\begin{proof}
Recall first from \eqref{king3} that
\begin{equation}\label{king3a}
\|\psi_{v_n,2}\|_{H^2(\Omega_2(v_n))} \le c_1(\kappa)\,,\qquad n\ge 1\,.
\end{equation}
As in the proof of Lemma~\ref{L3} we map $\Omega_2(v)$ onto the rectangle $\mathcal{R}_2 =D\times (1,1+d)$ and define, for $(x,\eta)\in \mathcal{R}_2$ and $n\ge 1$,
\begin{equation*}
\phi_n(x,\eta) := \psi_{v_n,2}(x,\eta+v_n(x)-1)\,,\qquad \phi(x,\eta) := \psi_{v,2}(x,\eta+v(x)-1)\,.
\end{equation*}
Let $q\in (1,2)$. Since
\begin{align*}
\nabla\phi_n(x,\eta) & = \Big( \partial_x \psi_{v_n} + \partial_x v_n \partial_z \psi_{v_n} \,,\, \partial_z \psi_{v_n} \Big)(x,\eta+v_n(x)-1)\,, \\
\partial_x^2 \phi_n(x,\eta) & = \Big( \partial_x^2 \psi_{v_n} + 2\partial_x v_n \partial_x \partial_z \psi_{v_n} + (\partial_x v_n)^2 \partial_z^2 \psi_{v_n} + \partial_x^2 v_n \partial_z \psi_{v_n} \Big)(x,\eta+v_n(x)-1)\,, \\
\partial_x\partial_\eta\phi_n(x,\eta) & = \Big( \partial_x\partial_z \psi_{v_n} + \partial_x v_n \partial_z^2 \psi_{v_n}\Big)(x,\eta+v_n(x)-1)\,, \\
\partial_\eta^2 \phi_n(x,\eta) & = \partial_z^2\psi_{v_n}(x,\eta+v_n(x)-1)\,,
\end{align*}
it follows from \eqref{vh2}, \eqref{king3a}, the continuous embedding of $H^2(D)$ in $C^1(\bar{D})$, and that of $H^1(\mathcal{R}_2)$ in $L^{2q/(2-q)}(\mathcal{R}_2)$ that
\begin{equation}\label{q1}
\phi_n\in W_q^2(\mathcal{R}_2)\;\;\text{ with }\;\; \| \phi_n\|_{W_q^2(\mathcal{R}_2)}\le c(q,\kappa)\,,\qquad n\ge 1\,.
\end{equation}
Now, given $p\in [1,\infty)$, we choose $q\in (1,\min\{2,p\})$ satisfying $1<2/q<1 + 1/p$ and $s\in (2/q-1/p,1)$. Since
$$
\phi_n \rightharpoonup \phi \quad \text{ in }\quad W_q^2(\mathcal{R}_2)
$$
by \eqref{204}, \eqref{q1}, and Proposition~\ref{ACDC}, the continuity of the trace as a mapping from $W_q^1(\mathcal{R}_2)$ to $W_q^{1-1/q}(D\times\{1\})$ and the compactness of the embedding of $W_q^{1-1/q}(D)$ in $L_p(D)$ imply that
\begin{equation}
\nabla\phi_n (\cdot,1) \rightarrow \nabla\phi (\cdot,1) \quad \text{ in }\quad W_q^{s-1/q}(D)\label{z107}
\end{equation}
and
\begin{equation}
\|\nabla\phi(\cdot,1)\|_{L_p(D)} \le c(p,\kappa)\,. \label{z108}
\end{equation}
That is,
$$
\partial_z\psi_{v_n,2}(\cdot,v_n)= \partial_\eta\phi_n (\cdot,1) \rightarrow \partial_\eta\phi (\cdot,1)=\partial_z\psi_{v,2}(\cdot,v) \quad \text{ in }\quad L_p(D)
$$
and, recalling \eqref{vh2} and the continuous embedding of $H^2(D)$ in $C^1(\bar{D})$,
\begin{equation*}
\begin{split}
\partial_x\psi_{v_n,2}(\cdot,v_n)= \partial_x \phi_n (\cdot,1)&-\partial_x v_n\partial_\eta\phi_n (\cdot,1)\\
& \rightarrow \partial_x \phi (\cdot,1)-\partial_x v\partial_\eta\phi (\cdot,1)
=\partial_x\psi_{v,2}(\cdot,v) \quad \text{ in }\quad L_p(D)\,.
\end{split}
\end{equation*}
Furthermore, \eqref{z107} and \eqref{z108}, along with the bound $\|v\|_{H^2(D)}\le \kappa$ and the continuous embedding of $H^2(D)$ in $C^1(\bar{D})$, entail that
$$
\|\nabla\psi_{v,2}(\cdot,v)\|_{L_p(D)} \le c(p,\kappa)\,,
$$
which proves \eqref{x} and the first bound in \eqref{xx}. Clearly, \eqref{x1} and the second bound in \eqref{xx} are shown in the same way.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{Thm2}]
The proof of Theorem~\ref{Thm2} is now a consequence of Proposition~\ref{C3} for \eqref{y1}, Proposition~\ref{ACDC}~{\bf (b)} for \eqref{y2}, and Proposition~\ref{U2} for \eqref{y3}.
\end{proof}
|
\section{#1} \setcounter{equation}{0}}
\renewcommand{\dagger}{\#}
\newcommand{\proofc}{{\sc Proof} \ }
\newcommand{\be}{\begin{equation} \label}
\newcommand{\ee}{\end{equation}}
\newcommand{\bea}{\begin{eqnarray}\label}
\newcommand{\eea}{\end{eqnarray}}
\newcommand{\bas}{\begin{eqnarray*}}
\newcommand{\eas}{\end{eqnarray*}}
\newcommand{\bit}{\begin{itemize}}
\newcommand{\eit}{\end{itemize}}
\newcommand{\nn}{\nonumber}
\newcommand{\R}{\mathbb{R}}
\newcommand{\N}{\mathbb{N}}
\newcommand{\pO}{\partial\Omega}
\newcommand{\dN}{\partial_\nu}
\newcommand{\eps}{\varepsilon}
\newcommand{\dist}{{\rm dist} \, }
\newcommand{\supp}{{\rm supp} \, }
\newcommand{\subsubset}{\subset\subset}
\newcommand{\wto}{\rightharpoonup}
\newcommand{\wsto}{\stackrel{\star}{\rightharpoonup}}
\newcommand{\hra}{\hookrightarrow}
\newcommand{\io}{\int_\Omega}
\newcommand{\na}{\nabla}
\newcommand{\Del}{\Delta}
\newcommand{\lam}{\lambda}
\newcommand{\pa}{\partial}
\newcommand{\bom}{\overline{\Omega}}
\newcommand{\Om}{\Omega}
\newcommand{\mint}{- \hspace*{-3.3mm} \int}
\newcommand{\Mint}{- \hspace*{-4mm} \int}
\newcommand{\mult}{\otimes}
\newcommand{\ov}{\overline}
\newcommand{\wh}{\widehat}
\newcommand{\wt}{\widetilde}
\newcommand{\vs}{\vspace*}
\newcommand{\hs}{\hspace*}
\newcommand{\cd}{\cdot}
\newcommand{\proj}{{\mathcal{P}}}
\newcommand{\vp}{\varphi}
\newcommand{\cb}{\color{blue}}
\newcommand{\cred}{\color{red}}
\newcommand{\ab}{\\[3mm]}
\newcommand{\abs}{\\[5pt]}
\newcommand{\Abs}{\\[5mm]}
\newcommand{\Aabs}{\\[7mm]}
\newcommand{\adb}{\allowdisplaybreaks}
\newcommand{\tm}{T_{max}}
\newcommand{\tme}{T_{max,\eps}}
\newcommand{\ueps}{u_\eps}
\newcommand{\veps}{v_\eps}
\newcommand{\Feps}{F_\eps}
\newcommand{\beps}{b_\eps}
\newcommand{\heps}{h_\eps}
\newcommand{\yeps}{y_\eps}
\newcommand{\te}{\tau_\eps}
\newcommand{\vst}{v_\star}
\newcommand{\neps}{n_\eps}
\newcommand{\ceps}{c_\eps}
\newcommand{\Peps}{P_\eps}
\newcommand{\zeps}{z_\eps}
\newcommand{\Kf}{K_F}
\newcommand{\kf}{k_F}
\newcommand{\red}[1]{{\color{red}#1}}
\usepackage{newunicodechar}
\newunicodechar{ε}{\varepsilon}
\newunicodechar{α}{\alpha}
\newunicodechar{β}{\beta}
\newunicodechar{Ω}{\Omega}
\newunicodechar{Δ}{\Delta}
\newunicodechar{∇}{\nabla}
\newunicodechar{∂}{\partial}
\newunicodechar{γ}{\gamma}
\newunicodechar{Φ}{\Phi}
\newunicodechar{ℝ}{\R}
\newunicodechar{ℕ}{\mathbb{N}}
\newunicodechar{ν}{\nu}
\newunicodechar{∞}{\infty}
\newunicodechar{τ}{\tau}
\newcommand{\Ombar}{\overline{Ω}}
\newcommand{\norm}[2][]{\|#2\|_{#1}}
\newcommand{\matr}[1]{\begin{pmatrix}#1\end{pmatrix}}
\allowdisplaybreaks
\begin{document}
\enlargethispage{10mm}
\title{Radial solutions to a chemotaxis-consumption model\\
involving prescribed signal concentrations on the boundary}
\author{
Johannes Lankeit\footnote{<EMAIL>}\\
{\small Leibniz Universität Hannover, Institut für Angewandte Mathematik}\\
{\small Welfengarten~1, 30167 Hannover, Germany}
\and
Michael Winkler\footnote{<EMAIL>}\\
{\small Institut f\"ur Mathematik, Universit\"at Paderborn, }\\
{\small Warburger Str.~100, 33098 Paderborn, Germany}
}
\date{}
\maketitle
\begin{abstract}
\noindent
The chemotaxis system
\begin{align*}
u_t &= \Delta u - \nabla \cdot (u\nabla v), \\
v_t &= \Delta v - uv,
\end{align*}
is considered under the boundary conditions $\frac{\partial u}{\partial\nu}- u\frac{\partial v}{\partial\nu}=0$ and $v=v_\star$ on $\partial\Omega$,
where $\Omega\subset\mathbb{R}^n$ is a ball and $v_\star$ is a given positive constant.\abs
In the setting of radially symmetric and suitably regular initial data, a result on global existence of bounded classical
solutions is derived in the case $n=2$, while global weak solutions are constructed when $n\in \{3,4,5\}$.
This is achieved by analyzing an energy-type inequality reminiscent of global structures previously observed
in related homogeneous Neumann problems. Ill-signed boundary integrals newly appearing therein are controlled
by means of spatially localized smoothing arguments revealing higher order regularity features outside the spatial origin.\abs
Additionally, unique classical solvability in the corresponding stationary problem is asserted, even in nonradial frameworks.\abs
\noindent {\bf Key words:} chemotaxis; consumption; global existence; stationary states; inhomogeneous boundary conditions\\
{\bf MSC (2020):} 35K55 (primary); 35J61, 35Q92, 92C17 (secondary)
\end{abstract}
\newpage
\mysection{Introduction}\label{intro}
Chemotaxis systems, if posed in bounded domains, are usually studied with homogeneous Neumann boundary conditions. Especially where the chemotactic agents partially direct their motion toward higher concentrations of a signal which they consume instead of produce, however, other boundary conditions may become relevant.\abs
In this article, we consider the chemotaxis consumption system
\be{firstsystem}
\left\{ \begin{array}{ll}
u_t &= \Delta u - \na \cdot (u\na v), \\
v_t &= \Delta v - uv,
\end{array} \right.
\ee
posing Dirichlet boundary conditions for the signal concentration $v$ and no-flux conditions for the bacterial population density $u$. \abs
Arising from a line of investigations concerned with pattern formation in colonies of {\em B. subtilis} in a fluid environment \cite{tuval}, such chemotaxis systems with signal consumption, additionally coupled to a Stokes- or Navier-Stokes fluid have been studied extensively over the last decade (for pointers to the literature see e.g. Section 4.1 of the survey \cite{BBTW} or the introduction of \cite{cao_lankeit}; for the model in the context of coral spawing, see e.g. \cite{zheng_corals}). While most works prescribed no-flux, homogeneous Neumann and homogeneous Dirichlet conditions for bacterial population density, signal concentration and fluid velocity,
it turned out that these
failed to adequately capture the colourful dynamics
observed in the form of
the patterns previously alluded to.\abs
In particular, a common result for the long-term behaviour was convergence to a constant state (i.e. a state without any patterns), as e.g. in \cite{taowin_evsmooth_stabil_consumption}, \cite{zhang_li}, \cite{win_arma}, \cite{win_TRAN}, in
\cite{lankeit_m3as}, \cite{lankeit_wang} and \cite{win_food_supported}
in a system additionally including population growth, or also \cite{fan_jin} in a related system with nonlinear diffusion.\abs
Not least because of this,
it has been suggested to use different, more realistic, inhomogeneous boundary conditions for the chemical signal
(see \cite{braukhoff} and \cite{braukhoff_lankeit}, but also \cite{tuval}), namely either
Robin type boundary conditions where the rate of oxygen influx is controlled by the local oxygen concentration at the boundary or nonzero Dirichlet conditions directly prescribing the latter.
(It has been confirmed \cite[Prop.~5.1]{braukhoff_lankeit} that the latter kind of conditions arises as a limit case of the former.)\abs
In general, altering boundary conditions can have a profound impact on the solution behaviour in chemotaxis systems, see the appearance
of a second critical mass in the Keller--Segel type system with Diricihlet conditions for $v$ studied in \cite{flw} if compared to the
same system with Neumann conditions; in both cases homogeneous.
While -- in these particular settings related to \eqref{firstsystem} -- more realistic from a modelling perspective, the change of boundary conditions to inhomogeneous conditions brings about additional mathematical challenges. \abs
In particular, a large part of the analysis of \eqref{firstsystem} and its relatives relies on certain energy-like structures such
as that expressed in the inequality
\begin{equation}\label{energy}
\frac{d}{dt} \left(\io u\ln u + 2 \io |\nabla \sqrt{v}|^2 \right) + \io \frac{|\nabla u|^2}u + \io v|D^2\ln v|^2 + \frac12 \io \frac{u|\nabla v|^2}v \le 0,
\end{equation}
as documented for the Neumann problem for (\ref{firstsystem}) in \cite[(3.1)]{taowin_evsmooth_stabil_consumption}.
Replacing homogeneous Neumann by inhomogeneous Dirichlet boundary conditions for $v$, in the derivation of \eqref{energy} additional boundary integrals arise, destroying the (quasi-)Lyapunov structure of \eqref{energy} or relatives thereof on which existence results in, e.g., \cite{lankeit_m3as}, \cite{taowin_locallybounded}, \cite{win_weak_ctNS}, \cite{black_stokes_limit}, \cite{zhang_li}, \cite{taowin_evsmooth_stabil_consumption}, \cite{win_CPDE12} or also \cite{jiang_han} essentially relied. \abs
Accordingly, only few results for chemotaxis-consumption models with boundary conditions different from homogeneous Neumann conditions
are available: Those concerning the related system with slightly different chemotaxis and energy consumption studied in \cite{knosalla_global,knosalla_nadzieja_stationary} with inhomogeneous Neumann and Dirichlet conditions are restricted to spatially one-dimensional domains.
For Robin-type conditions of the form introduced in \cite{braukhoff}, also in higher dimensions, the stationary problem of \eqref{firstsystem} has been shown to be uniquely solvable (for any prescribed total mass $\io u$ of the first component, see \cite{braukhoff_lankeit}), and (under a moderate smallness condition) features as the limit of a parabolic-elliptic simplification of \eqref{firstsystem} (cf.~\cite{FLM}). Also in fluid-coupled systems solutions have been found in the presence of logistic source terms
(\cite{braukhoff}), or superlinear diffusion (\cite{wu_xiang,tian_xiang}), or without both (\cite{braukhoff_tang}).\abs
In the Dirichlet setting this article is concerned with
we mention \cite{lee_wang_yang} where an asymptotic analysis of the vanishing diffusivity limit for $v$ in the stationary system seems to confirm the potential of \eqref{firstsystem}
to capture pattern dynamics.
In the time-dependent problem (including fluid flow), solutions in $ℝ^2\times[0,1]$ were constructed in \cite{peng_xiang} if the signal consumption was strong, at least quadratic with respect to $u$, and in $\Omega\subset ℝ^2$ in \cite{wang_win_xiang10}, in both cases under a smallness condition on the initial data.
Without smallness conditions, solutions to \eqref{firstsystem} coupled to a fluid flow governed by the Stokes equations were constructed in \cite{wang_win_xiang6} ($\Omega\subset ℝ^N$ with linear diffusion for $N=2$ and porous medium type diffusion in higher dimensions) and in \cite{wang_win_xiang_CPDE} ($\Omega\subset ℝ^3$). Nevertheless, the solution concepts pursued in these works are rather weak
and do not yield comparable regularity as \cite{win_CPDE12} and \cite{tao_bdoxygenconsumption} for solutions to the system with homogeneous Neumann conditions.\abs
As to the above-mentioned difficulties concerning \eqref{energy}, different strategies have been employed:
Exploiting the Robin condition in their systems, the works \cite{braukhoff} and subsequently \cite{wu_xiang} and \cite{tian_xiang} rely on a Lions-Magenes type transformation converting $v$ to a function with homogeneus Neumann boundary conditions.
The energy functional is enhanced by additional ``boundary energy'' terms in \cite{braukhoff_tang}.
In \cite{peng_xiang} and \cite{FLM}, a trace theorem is used to control the boundary integrals by integrals over the domain involving higher derivatives, which are available either due to the simpler elliptic form of the second equation (in \cite{FLM}) or due to a smallness condition (\cite{peng_xiang}, also in the result on long-term limit in \cite{FLM}).
In \cite{wang_win_xiang_CPDE}, a localized modification of the energy functional was investigated, the localization being detrimental to the regularity information near the boundary. Leaving \eqref{energy} behind, the approaches in
\cite{wang_win_xiang6} and \cite{wang_win_xiang10} used different energy functionals (or small-data energy functionals), giving rise to less potent a priori estimates, as reflected in the generalized sense of solvability obtained.\abs
{\bf Regularity control on the boundary for radial solutions. Main results.} \quad
In this article, we plan to use radial symmetry as a mean to unravel difficulties related to possible effects
that the change from Neumann to Dirichlet boundary conditions for the signal may have on boundary regularity of solutions.
Specifically, in a ball $\Omega=B_R(0)\subset\R^n$ with $R>0$ and $n\ge 2$, and with a given positive constant $\vst$,
we shall consider the initial-boundary value problem
\be{0}
\left\{ \begin{array}{ll}
u_t = \Delta u - \na \cdot (u\na v),
\qquad & x\in \Omega, \ t>0, \\[1mm]
v_t = \Delta v - uv,
\qquad & x\in \Omega, \ t>0, \\[1mm]
\frac{\partial u}{\partial\nu}-u \frac{\partial v}{\partial\nu}=0, \quad v=\vst,
\qquad & x\in \pO, \ t>0, \\[1mm]
u(x,0)=u^{(0)}(x), \quad v(x,0)=v^{(0)}(x),
\qquad & x\in\Omega,
\end{array} \right.
\ee
assuming that
\be{init}
\left\{ \begin{array}{l}
u^{(0)}\in W^{1,\infty}(\Om) \mbox{ is nonnegative in $\Om$ and radially symmetric with $u^{(0)}\not\equiv 0$,\quad and} \\[1mm]
v^{(0)}\in W^{1,\infty}(\Om) \mbox{ is positive in $\bom$ and radially symmetric with $v^{(0)}=\vst$ on $\pO$,}
\end{array} \right.
\ee
where, as throughout the sequel, radial symmetry of a function $\varphi$ on $\Om$ is to be understood as referring to the
spatial origin.\abs
To make appropriate use of these symmetry assumptions,
at a first stage of our analysis
we shall rely on the essentially one-dimensional framework thereby generated in order to step by step turn the
basic properties of mass conservation in the first component, and uniform $L^\infty$ boundedness in the second,
into knowledge on higher order regularity features locally outside the spatial origin (see Section \ref{sect3} and especially
Corollary \ref{cor10}).
This will particularly enable us to appropriately control boundary integrals which due to the presence of possibly nonzero
normal derivatives arise in a spatially global energy analysis related to that in (\ref{energy}) (Section \ref{sect4}).\abs
In the spatially planar case, this will be found to entail a priori bounds actually in $L^\infty \times W^{1,\infty}$,
and to thus imply the following statement on global classical solvability and boundedness in (\ref{0}):
\begin{theo}\label{theo16}
Let $R>0$ and $\Omega=B_R(0)\subset\R^2$, and suppose that $\vst\ge 0$, and that $u^{(0)}$ and $v^{(0)}$ satisfy (\ref{init}).
Then there exist unique functions
\bas
\left\{ \begin{array}{l}
u\in C^0(\bom\times [0,\infty)) \cap C^{2,1}(\bom\times (0,\infty)) \qquad \mbox{and} \\[1mm]
v\in \bigcap_{q>2} C^0([0,\infty);W^{1,q}(\Om)) \cap C^{2,1}(\bom\times (0,\infty))
\end{array} \right.
\eas
which are such that $u>0$ and $v>0$ in $\bom\times [0,\infty)$, that $(u(\cdot,t),v(\cdot,t))$ is radially symmetric for all $t>0$,
and that $(u,v)$ solves (\ref{0}) in the classical sense in $\Om\times (0,\infty)$.
Moreover, there exists $C>0$ such that
\be{16.1}
\|u(\cdot,t)\|_{L^\infty(\Om)} + \|v(\cdot,t)\|_{W^{1,\infty}(\Om)} \le C
\qquad \mbox{for all } t>0.
\ee
\end{theo}
But also in some higher-dimensional situations, the regularity information gained from our energy analysis can be used to
establish a result on global solvability, albeit in a slightly weaker framework:
\begin{theo}\label{theo25}
Let $n\in \{3,4,5\}$, $R>0$ and $\Omega=B_R(0)\subset\R^n$, let $\vst\ge 0$, and assume (\ref{init}).
Then one can find nonnegative functions
\be{25.01}
\left\{ \begin{array}{l}
u\in L^\infty((0,\infty);L^1(\Om)) \cap L^\frac{n+2}{n}_{loc}(\bom\times [0,\infty))
\cap L^\frac{n+2}{n+1}_{loc}([0,\infty);W^{1,\frac{n+2}{n+1}}(\Om)) \qquad \mbox{and} \\[1mm]
v\in L^\infty(\Om\times (0,\infty)) \quad \mbox{with} \quad
v-\vst \in L^\infty((0,\infty);W_0^{1,2}(\Om)) \cap L^4_{loc}([0,\infty);W^{1,4}(\Om))
\end{array} \right.
\ee
such that $u(\cdot,t)$ and $v(\cdot,t)$ are radially symmetric for a.e.~$t>0$, and that $(u,v)$ forms a global weak solution of
(\ref{0}) in the sense of Definition \ref{dw} below.
Furthermore, there exists $C>0$ such that
\be{25.1}
\io u(\cdot,t)\ln u(\cdot,t) \le C
\qquad \mbox{for almost all } t>0
\ee
and
\be{25.2}
\io |\na v(\cdot,t)|^2 \le C
\qquad \mbox{for almost all } t>0
\ee
as well as
\be{25.3}
\int_t^{t+1} \io \Big\{ u^\frac{n+2}{n} + |\na u|^\frac{n+2}{n+1} + |\na v|^4 \Big\} \le C
\qquad \mbox{for all } t>0.
\ee
\end{theo}
We have to leave open here the question how far information on the large time behaviour of the above solutions
that goes beyond the boundedness features in (\ref{16.1}) and in (\ref{25.1})-(\ref{25.3}) can be derived,
especially in the presence of large initial data.
After all, a steady state analysis guided by the approach developed in \cite{braukhoff_lankeit} provides the following
result which may be viewed as an indication for nontrivial dynamics involving structured states in (\ref{0}):
\begin{theo}\label{th:stationary}
Let $Ω\subset ℝ^n$ be a bounded domain with smooth boundary, and suppose that
$\vst \in \bigcup_{\beta\in (0,1)} C^{2+\beta}(\bom)$ is nonnegative.
Then for every $m\ge 0$, the stationary problem \eqref{stationary} has a unique solution $(u,v)\in (C^2(\Ombar))^2$
satisfying $\io u=m$.
If $Ω=B_R(0)$ and $\vst$ is constant, then this solution is radially symmetric, and both $u$ and $v$ are convex.
\end{theo}
\mysection{Local solvability, approximation and basic properties}
In order to simultaneously address, throughout large parts of our analysis,
the case $n=2$ in which classical solvability is strived for, and the case $n\in \{3,4,5\}$
in which we intend to construct a solution via approximation, for $\eps\in [0,1)$ let us consider the variants of (\ref{0})
given by
\be{0eps}
\left\{ \begin{array}{ll}
u_{\eps t} = \Delta\ueps - \na \cdot \big( \ueps\Feps'(\ueps)\na\veps\big),
\qquad & x\in\Om, \ t>0, \\[1mm]
v_{\eps t} = \Delta \veps - \Feps(\ueps)\veps,
\qquad & x\in\Om, \ t>0, \\[1mm]
\frac{\partial\ueps}{\partial\nu} - \ueps\Feps'(\ueps)\frac{\partial\veps}{\partial\nu}=0, \quad \veps=\vst,
\qquad & x\in\pO, \ t>0, \\[1mm]
\ueps(x,0)=u^{(0)}(x), \quad \veps(x,0)=v^{(0)}(x),
\qquad & x\in\Om,
\end{array} \right.
\ee
where
\be{F0}
\Feps(\xi):=\frac{\xi}{1+\eps\xi},
\qquad \xi\ge 0, \ \eps\in [0,1),
\ee
satisfies
\be{F}
0 \le \Feps(\xi)\le \xi
\quad \mbox{and} \quad
0 \le \Feps'(\xi)=\frac{1}{(1+\eps\xi)^2} \le 1
\qquad \mbox{for all $\xi\ge 0$ and } \eps\in [0,1);
\ee
indeed, these choices ensure that (\ref{0eps}) coincides with (\ref{0}) when $\eps=0$.\abs
Local solvability and a handy extensibility criterion can be obtained by resorting to standard literature:
\begin{lem}\label{lem1}
Let $\eps\in [0,1)$. Then there exist $\tme\in (0,\infty]$ and uniquely determined functions
\bas
\left\{ \begin{array}{l}
\ueps\in C^0(\bom\times [0,\tme)) \cap C^{2,1}(\bom\times (0,\tme)) \qquad \mbox{and} \\[1mm]
\veps\in \bigcap_{q>n} C^0([0,\tme);W^{1,q}(\Om)) \cap C^{2,1}(\bom\times (0,\tme))
\end{array} \right.
\eas
such that $\ueps>0$ and $\veps>0$ in $\bom\times [0,\tme)$, that $\ueps(\cdot,t)$ and $\veps(\cdot,t)$ are radially symmetric
for all $t\in (0,\tme)$,
that $(\ueps,\veps)$ solves (\ref{0eps}) classically in $\Om\times (0,\tme)$, and that
\be{ext}
\mbox{either $\tme=\infty$, \quad or \quad}
\limsup_{t\nearrow\tme} \Big\{ \|\ueps(\cdot,t)\|_{W^{1,q}(\Om)} + \|\veps(\cdot,t)\|_{W^{1,q}(\Om)} \Big\}
= \infty
\quad
\mbox{for all $q>n$}.
\ee
\end{lem}
\begin{proof}
This results from \cite[Thm.~14.4 and Thm.~14.6]{amann_nonhomogeneous} when, for $U=\matr{u\\v-\vst}$, applied to
the evolution problem given by $U_t=\nabla\cdot (A(U)\nabla U)+f(U)$,
$\left(\matr{1&0\\0&0} A(U)ν + \matr{0&0\\0&1} U\right)\vert_{∂Ω} =0$, $U(0)=\matr{u^{(0)}\\v^{(0)}-\vst}$,
with $A(U)=\matr{1&-U_1F'(U_1)\\0&1}$, $f(U)=\matr{0\\-F(U_1)(U_2+\vst)}$.
\end{proof}
These solutions clearly preserve mass in their first component and are bounded in the second.
\begin{lem}\label{lem2}
Let $\eps\in [0,1)$. Then
\be{mass}
\io \ueps(\cdot,t)=\io u^{(0)}
\qquad \mbox{for all } t\in (0,\tme)
\ee
and
\be{vinfty}
\|\veps(\cdot,t)\|_{L^\infty(\Om)} \le \|v^{(0)}\|_{L^\infty(\Om)}
\qquad \mbox{for all } t\in (0,\tme).
\ee
\end{lem}
\begin{proof}
While (\ref{mass}) can directly be seen upon integrating the first equation in (\ref{0eps}), the inequality in (\ref{vinfty})
can be verified by means of the comparison principle applied to the second equation in (\ref{0eps}), because
$\ov{v}(x,t):=\|v^{(0)}\|_{L^\infty(\Om)}$, $(x,t)\in\bom\times [0,\infty)$, satisfies
$\ov{v}_t- \Delta\ov{v}+\Feps(\ueps)\ov{v}=\Feps(\ueps)\ov{v}\ge 0$ in $\Om\times (0,\tme)$ for all $\eps\in (0,1)$ as well as
$\ov{v}|_{t=0}\ge v^{(0)}$ and $\ov{v}|_{\pO}\ge \vst$, the latter due to the fact that (\ref{init}) necessarily requires that
$\vst \le \|v^{(0)}\|_{L^\infty(\Om)}$.
\end{proof}
Also for the gradient of the second solution component some first a priori estimates are available.
\begin{lem}\label{lem4}
There exists $C>0$ such that
\be{4.1}
\|\na\veps(\cdot,t)\|_{L^1(\Om)} \le C
\qquad \mbox{for all $t\in (0,\tme)$ and $\eps\in [0,1)$.}
\ee
\end{lem}
\begin{proof}
According to well-known smoothing estimates for the Dirichlet heat semigroup $(e^{t\Delta})_{t\ge 0}$ on $\Om$
(\cite{eidelman_ivasishen}, \cite[Section 48.2]{quittner_souplet}), there exist $\lam>0$, $c_1>0$ and $c_2>0$ such that for all $t>0$,
\bas
\|\na e^{t\Delta}\varphi\|_{L^1(\Om)} \le c_1 \|\varphi\|_{W^{1,\infty}(\Om)}
\qquad \mbox{for all $\varphi \in W^{1,\infty}(\Om)$ such that $\varphi=\vst$ on $\pO$,}
\eas
and
\bas
\|\na e^{t\Delta}\varphi\|_{L^1(\Om)} \le c_2 \cdot (1+t^{-\frac{1}{2}}) e^{-\lam t} \|\varphi\|_{L^1(\Om)}
\qquad \mbox{for all $\varphi \in C^0(\bom)$ such that $\varphi=\vst$ on $\pO$.}
\eas
Since (\ref{F}) together with (\ref{mass}) and (\ref{vinfty}) ensures that for all $\eps\in [0,1)$ we have
$0\le \Feps(\ueps)\le \ueps$ and hence
\bas
\|\Feps(\ueps)\veps\|_{L^1(\Om)}
\le \|\Feps(\ueps)\|_{L^1(\Om)} \|\veps\|_{L^\infty(\Om)}
\le \|\ueps\|_{L^1(\Om)} \|\veps\|_{L^\infty(\Om)}
\le c_3:=\|u^{(0)}\|_{L^1(\Om)} \|v^{(0)}\|_{L^\infty(\Om)}
\eas
for all $t\in (0,\tme)$, in view of (\ref{0eps}) this implies that for any such $\eps$,
\bas
\|\na\veps(\cdot,t)\|_{L^1(\Om)}
&=& \big\| \na \big(\veps(\cdot,t)-\vst\big)\big\|_{L^1(\Om)} \\
&=& \bigg\| \na e^{t\Delta} (v^{(0)}-\vst)
- \int_0^t \na e^{(t-s)\Delta} \Big\{ \Feps\big(\ueps(\cdot,s)\big)\veps(\cdot,s) \Big\} ds \bigg\|_{L^1(\Om)} \\
&\le& c_1 \|v^{(0)}-\vst\|_{W^{1,\infty}(\Om)}
+ c_2 c_3 \int_0^t \Big(1+(t-s)^{-\frac{1}{2}}\Big) e^{-\lam (t-s)} ds \\
&\le& c_1 \|v^{(0)}-\vst\|_{W^{1,\infty}(\Om)}
+ c_2 c_3 \int_0^\infty (1+\sigma^{-\frac{1}{2}}) e^{-\lam\sigma} d\sigma
\qquad \mbox{for all } t\in (0,\tme),
\eas
which establishes (\ref{4.1}).
\end{proof}
\mysection{Local estimates outside the origin}\label{sect3}
In line with common abuse of notation, we occasionally write $\ueps(r,t)$ and $\veps(r,t)$,
instead of $\ueps(x,t)$ and $\veps(x,t)$, for $r=|x|\in[0,R]$, $t\in[0,\tme)$ and $\eps\in [0,1)$,
and in order to prepare our derivation of local estimates outside the origin, we
observe that whenever $\chi\in C_0^\infty((0,R])$, for each $\eps\in [0,1)$ we have
\be{chiv}
\Big( \chi \cdot (\veps-\vst)\Big)_t
= \Big( \chi\cdot (\veps-\vst)\Big)_{rr}
+ \beps(r,t),
\qquad r\in (0,R), \ t\in (0,\tme),
\ee
where
\bea{b}
& & \hs{-20mm}
\beps(r,t) \equiv \beps^{(\chi)}(r,t)
:= \Big( \frac{n-1}{r}\chi(r)-2\chi_r(r)\Big) v_{\eps r}(r,t)
- \chi_{rr}(r) \cdot \big(\veps(r,t)-\vst\big) \nn\\
& & \hs{40mm}
- \chi(r)\cdot \Feps\big(\ueps(r,t)\big) \veps(r,t),
\qquad r\in (0,R), \ t\in (0,\tme).
\eea
As the above basic estimates imply $L^1$ bounds for $\beps$, a straightforward argument based on parabolic smoothing
in the one-dimensional heat equation (\ref{chiv}) yields the following information on regularity of the taxis gradient
outside the origin.
\begin{lem}\label{lem5}
Let $q\in (1,\infty)$ and $\delta\in (0,R)$. Then there exists $C(q,\delta)>0$ with the property that
\be{5.1}
\|v_{\eps r}(\cdot,t)\|_{L^q((\delta,R))} \le C(q,\delta)
\qquad \mbox{for all $t\in (0,\tme)$ and $\eps\in [0,1)$.}
\ee
\end{lem}
\begin{proof}
Given $\delta\in (0,R)$, we fix $\chi\in C^\infty([0,R])$ such that $0\le\chi\le 1$, that $\chi\equiv 0$ in $[0,\frac{\delta}{2}]$,
and that $\chi\equiv 1$ in $[\delta,R]$, and let $(e^{-tA})_{t\ge 0}$ denote the one-dimensional heat semigroup generated
by the operator $A:=-(\cdot)_{rr}$ under homogeneous Dirichlet boundary conditions on $(\frac{\delta}{2},R)$.
Then known regularization features of the latter (\cite{eidelman_ivasishen}, \cite{quittner_souplet}) ensure that if we fix
$q\in (1,\infty)$, then we can find $\lam=\lam(q,\delta)>0, c_1=c_1(q,\delta)>0$ and $c_2=c_2(q,\delta)>0$ such that whenever $t>0$,
\be{5.2}
\|\partial_r e^{-tA}\varphi\|_{L^q((\frac{\delta}{2},R))}
\le c_1\|\varphi\|_{W^{1,\infty}((\frac{\delta}{2},R))}
\qquad \mbox{for all $\varphi\in W^{1,\infty}((\frac{\delta}{2},R))$ such that $\varphi(\frac{\delta}{2})=\varphi(R)=0$}
\ee
and
\be{5.3}
\|\partial_r e^{-tA}\varphi\|_{L^q((\frac{\delta}{2},R))}
\le c_2 \cdot \big(1+t^{-1+\frac{1}{q}}\big) e^{-\lam t} \|\varphi\|_{L^1((\frac{\delta}{2},R))}
\qquad \mbox{for all $\varphi\in C^0([\frac{\delta}{2},R])$ with $\varphi(\frac{\delta}{2})=\varphi(R)=0$}.
\ee
Apart from that, a combination of Lemma \ref{lem4} with (\ref{mass}) and (\ref{vinfty}) shows that since
$\supp \chi \subset [\frac{\delta}{2},R]$, we can pick $c_3=c_3(\delta)>0$ in such a way that for any $\eps\in [0,1)$, the
function $\beps=\beps^{(\chi)}$ defined in (\ref{b}) satisfies
\bas
\|\beps(\cdot,t)\|_{L^1((\frac{\delta}{2},R))} \le c_3
\qquad \mbox{for all } t\in (0,\tme).
\eas
On the basis of (\ref{chiv}) and the fact that $\chi \cdot (\veps-\vst)=0$ on $\{\frac{\delta}{2},R\} \times (0,\tme)$,
we can therefore utilize (\ref{5.2}) and (\ref{5.3}) to estimate
\bea{5.4}
\Big\| \partial_r \Big\{ \chi \cdot \big(\veps(\cdot,t)-\vst\big) \Big\} \Big\|_{L^q((\frac{\delta}{2},R))}
&=& \bigg\| \partial_r e^{-tA} \Big\{ \chi \cdot (v^{(0)}-\vst)\Big\}
+ \int_0^t \partial_r e^{-tA} \beps(\cdot,s) ds \bigg\|_{L^q((\frac{\delta}{2},R))} \nn\\
&\le& c_1 \big\| \chi \cdot (v^{(0)}-\vst)\big\|_{W^{1,\infty}((\frac{\delta}{2},R))} \nn\\
& & + c_2 \int_0^t \Big(1+(t-s)^{-1+\frac{1}{q}}\Big) e^{-\lam (t-s)} \|\beps(\cdot,s)\|_{L^1((\frac{\delta}{2},R))} ds \nn\\
&\le& c_1 \big\| \chi \cdot (v^{(0)}-\vst)\big\|_{W^{1,\infty}((\frac{\delta}{2},R))}
+ c_2 c_3 c_4
\eea
for all $t\in (0,\tme)$ and $\eps\in [0,1)$,
with $c_4:=\int_0^\infty(1+\sigma^{-1+\frac{1}{q}}) e^{-\lam\sigma} d\sigma$ being finite since we are assuming that $q<\infty$.
As $\chi\equiv 1$ in $[\delta,R]$, (\ref{5.1}) directly results from (\ref{5.4}).
\end{proof}
This has consequences for bounds on certain powers of $\ueps$ and their derivative outside a neighbourhood of the spatial origin.
\begin{lem}\label{lem6}
For any choice of $p\in (0,1)$ and $\delta\in (0,R)$ one can find $C(p,\delta)>0$ such that whenever $\eps\in [0,1)$,
\be{6.1}
\int_t^{t+\te} \int_\delta^R \Big| \big(\ueps^\frac{p}{2}\big)_r \Big|^2 \le C(p,\delta)
\qquad \mbox{for all $t\in [0,\tme-\te)$,}
\ee
and that
\be{6.2}
\int_t^{t+\te} \int_\delta^R \ueps^{p+2} \le C(p,\delta)
\qquad \mbox{for all $t\in [0,\tme-\te)$,}
\ee
where $\te:=\min\{1,\frac{1}{2}\tme\}$.
\end{lem}
\begin{proof}
We fix $\zeta\in C^\infty(\bom)$ such that $0\le\zeta\le 1$, and that $\zeta\equiv 0$ in $\ov{B}_\frac{\delta}{2}(0)$
as well as $\zeta\equiv 1$ in $\bom\setminus \ov{B}_\delta(0)$.
Relying on the positivity of $\ueps$ in $\bom\times (0,\tme)$ for $\eps\in [0,1)$, from (\ref{0eps}) we then obtain that due to
Young's inequality,
\bea{6.3}
\frac{1}{p} \frac{d}{dt} \io \zeta^2 \ueps^p
&=& \io \zeta^2 \ueps^{p-1} \na \cdot \Big\{ \na\ueps-\ueps\Feps'(\ueps)\na\veps\Big\} \nn\\
&=& (1-p) \io \zeta^2 \ueps^{p-2} |\na\ueps|^2
-(1-p) \io \zeta^2 \ueps^{p-1} \Feps'(\ueps) \na\ueps\cdot\na\veps \nn\\
& & - 2\io \zeta \ueps^{p-1} \na\ueps\cdot\na\zeta
+ 2 \io \zeta\ueps^p \Feps'(\ueps) \na\veps\cdot\na\zeta \nn\\
&\ge& \frac{1-p}{2} \io \zeta^2 \ueps^{p-2} |\na\ueps|^2
- (2-p) \io \zeta^2 \ueps^p |\na\veps|^2 \nn\\
& & - \frac{5-p}{1-p} \io |\na\zeta|^2 \ueps^p
\qquad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1),
\eea
because $0\le\Feps'\le 1$ for all $\eps\in [0,1)$.
Here by the H\"older inequality,
\bas
\io \zeta^2 \ueps^p |\na\veps|^2
\le \bigg\{ \io \ueps \bigg\}^p \cdot \bigg\{ \io \zeta^\frac{2}{1-p} |\na\veps|^\frac{2}{1-p}\bigg\}^{1-p}
\qquad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1)
\eas
and
\bas
\io |\na\zeta|^2 \ueps^p
\le \|\na\zeta\|_{L^\frac{2}{1-p}(\Om)}^2 \cdot \bigg\{ \io \ueps \bigg\}^p
\qquad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1),
\eas
so that since $\supp \zeta\subset \bom\setminus B_\frac{\delta}{2}(0)$ we may apply Lemma \ref{lem5} to $q:=\frac{2}{1-p}$ to see
that thanks to (\ref{mass}), with some $c_1=c_1(p,\delta)>0$ we have
\bas
(2-p)\io \zeta^2 \ueps^p |\na\veps|^2
+ \frac{5-p}{1-p} \io |\na\zeta|^2 \ueps^p
\le c_1
\qquad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1).
\eas
Therefore, an integration in (\ref{6.3}) shows that again due to the H\"older inequality and (\ref{mass}),
\bas
\frac{1-p}{2} \int_t^{t+\te} \int_{\Om\setminus B_\delta(0)} \ueps^{p-2} |\na\ueps|^2
&\le& \frac{1-p}{2} \int_t^{t+\te} \io \zeta^2 \ueps^{p-2} |\na\ueps|^2 \\
&\le& \frac{1}{p} \io \zeta^2 \ueps^p(\cdot,t-\te)
- \frac{1}{p} \io \zeta^2 \ueps^p(\cdot,t)
+ c_1\te \\
&\le& \frac{1}{p} \cdot \bigg\{ \io u^{(0)} \bigg\}^p
+ c_1
\quad \mbox{for all $t\in [0,\tme-\te)$ and } \eps\in [0,1),
\eas
because $\te\le 1$.
This implies (\ref{6.1}), whereupon (\ref{6.2}) readily results from (\ref{6.1}) according to the fact that the Gagliardo-Nirenberg
inequality provides $c_2=c_2(p,\delta)>0$ fulfilling
\bas
\int_\delta^R \ueps^{p+2}
&=& \|\ueps^\frac{p}{2}\|_{L^\frac{2(p+2)}{p}((\delta,R))}^\frac{2(p+2)}{p} \\
&\le& c_2\Big\| \big(\ueps^\frac{p}{2}\big)_r \Big\|_{L^2((\delta,R))}^2
\big\|\ueps^\frac{p}{2}\big\|_{L^\frac{2}{p}((\delta,R))}^\frac{4}{p}
+ c_2 \big\|\ueps^\frac{p}{2}\big\|_{L^\frac{2}{p}((\delta,R))}^\frac{2(p+2)}{p}
\quad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1),
\eas
and because for all $t\in (0,\tme)$ and $\eps\in [0,1)$,
\bas
\big\|\ueps^\frac{p}{2}\big\|_{L^\frac{2}{p}((\delta,R))}^\frac{2}{p}
= \int_\delta^R \ueps(r,t) dr
\le \delta^{1-n} \int_0^R r^{n-1} \ueps(r,t) dr
= \delta^{1-n} \int_0^R r^{n-1} u^{(0)}(r) dr
\eas
by (\ref{mass}).
\end{proof}
In contrast to settings with homogeneous boundary conditions, in the present situation it will become necessary to deal with non-vanishing boundary terms. While this section will culminate in corresponding estimates, a key to these becomes visible in the following corollary already.
\begin{cor}\label{cor7}
There exists $C>0$ such that
\be{7.1}
\int_t^{t+\te} \ueps(R,s) ds \le C
\qquad \mbox{for all $t\in [0,\tme-\te)$,}
\ee
where again $\te=\min\{1,\frac{1}{2}\tme\}$ for $\eps\in [0,1)$.
\end{cor}
\begin{proof}
By means of the Gagliardo-Nirenberg inequality, we can pick $c_1>0$ with the property that
\bas
\|\varphi\|_{L^\infty((\frac{R}{2},R))}^6
\le c_1\|\varphi_r\|_{L^2((\frac{R}{2},R))}^2 \|\varphi\|_{L^4((\frac{R}{2},R))}^4
+ c_1\|\varphi\|_{L^4((\frac{R}{2},R))}^6
\qquad \mbox{for all $\varphi\in W^{1,2}((\frac{R}{2},R))$,}
\eas
whence
\bas
\int_t^{t+\te} \ueps^\frac{3}{2}(R,s) ds
&\le& \int_t^{t+\te} \big\| \ueps^\frac{1}{4}(\cdot,s)\big\|_{L^\infty((\frac{R}{2},R))}^6 ds \\
&\le& c_1 \int_t^{t+\te} \Big\| \big(\ueps^\frac{1}{4}\big)_r(\cdot,s) \Big\|_{L^2((\frac{R}{2},R))}^2
\big\|\ueps^\frac{1}{4}(\cdot,s)\big\|_{L^4((\frac{R}{2},R))}^4 ds \\
& & + c_1\int_t^{t+\te} \big\|\ueps^\frac{1}{4}(\cdot,s)\big\|_{L^4((\frac{R}{2},R))}^6 ds
\qquad \mbox{for all $t\in [0,\tme-\te)$ and } \eps\in [0,1).
\eas
Combining (\ref{mass}) with an application of Lemma \ref{lem6} to $p:=\frac{1}{2}$ thus shows that with some $c_2>0$ we have
\bas
\int_t^{t+\te} \ueps^\frac{3}{2}(R,s) ds \le c_2
\qquad \mbox{for all $t\in [0,\tme-\te)$ and } \eps\in [0,1),
\eas
from which (\ref{7.1}) follows upon employing the H\"older inequality.
\end{proof}
The following elementary observation, a proof of which can be found in \cite[Lemma 3.4]{win_JFA}, will be referred to
in Lemma~\ref{lem8}, Lemma \ref{lem11} and Lemma \ref{lem14}.
\begin{lem}\label{lem77}
Let $T\in (0,\infty]$ and $\tau\in (0,T)$, and let $h\in L^1_{loc}((0,T))$ be nonnegative and such that
\bas
\int_t^{t+\tau} h(s) ds \le b
\qquad \mbox{for all } t\in (0,T-\tau)
\eas
with some $b>0$. Then
\bas
\int_0^t e^{-\lam(t-s)} h(s) ds \le \frac{b\tau}{1-e^{-\lam\tau}}
\qquad \mbox{for all $t\in (0,T)$ and any } \lam>0.
\eas
\end{lem}
Whereas the previous estimates for $\ueps$ were concerned with temporally integrated quantities, the following lemma provides a temporally uniform bound.
\begin{lem}\label{lem8}
Let $p\in (1,3)$ and $\delta\in (0,R)$. Then there exists $C(p,\delta)>0$ such that
\be{8.1}
\int_\delta^R \ueps^p(r,t) dr
\le C(p,\delta)
\qquad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1).
\ee
\end{lem}
\begin{proof}
We again take a function $\zeta\in C^\infty(\bom)$ fulfilling $0\le\zeta\le 1$ and $\zeta|_{\ov{B}_\frac{\delta}{2}(0)}\equiv 0$
as well as $\zeta|_{\bom\setminus B_\delta(0)}\equiv 1$, and once more rely on (\ref{0eps}) to see by means of Young's inequality
and (\ref{F}) that
\bea{8.2}
\frac{d}{dt} \io \zeta^2 \ueps^p
+ \io \zeta^2 \ueps^p
&=& - p(p-1) \io \zeta^2 \ueps^{p-2} |\na\ueps|^2
+ p(p-1) \io \zeta^2 \ueps^{p-1} \Feps'(\ueps) \na\ueps\cdot\na\veps \nn\\
& & -2p \io \zeta \ueps^{p-2} \na\ueps\cdot\na\zeta
+ 2p \io \zeta\ueps^p \Feps'(\ueps) \na\veps\cdot\na\zeta \nn\\
& & + \io \zeta^2 \ueps^p \nn\\
&\le& \frac{p(p+1)}{2} \io \zeta^2 \ueps^p |\na\veps|^2
+ \frac{p(p+1)}{p-1} \io |\na\zeta|^2 \ueps^p \nn\\
& & + \io \zeta^2 \ueps^p
\qquad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1).
\eea
Here, taking any $p_0=p_0(p)>p$ such that $p_0<3$, we may again draw on Young's inequality to estimate
\bas
\io \zeta^2 \ueps^p |\na\veps|^2
\le \int_{\Omega\setminus B_\frac{\delta}{2}(0)} \ueps^{p_0}
+ \int_{\Omega\setminus B_\frac{\delta}{2}(0)} |\na\veps|^\frac{2p_0}{p_0-p}
\qquad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1)
\eas
and
\bas
\io |\na\zeta|^2 \ueps^p
\le \int_{\Omega\setminus B_\frac{\delta}{2}(0)} \ueps^{p_0}
+ \io |\na\zeta|^\frac{2p_0}{p_0-p}
\qquad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1)
\eas
as well as
\bas
\io \zeta^2 \ueps^p
\le \int_{\Omega\setminus B_\frac{\delta}{2}(0)} \ueps^{p_0}
+ |\Omega|
\qquad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1),
\eas
so that invoking Lemma \ref{lem5} we find $c_1=c_1(p,\delta)>0$ fulfilling
\bas
& & \hs{-30mm}
\frac{p(p+1)}{2} \io \zeta^2 \ueps^p |\na\veps|^2
+ \frac{p(p+1)}{p-1} \io |\na\zeta|^2 \ueps^p
+ \io \zeta^2 \ueps^p \\
&\le& \heps(t):=c_1\int_{\Omega\setminus B_\frac{\delta}{2}(0)} \ueps^{p_0}
+ c_1
\qquad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1).
\eas
From (\ref{8.2}) we therefore obtain that
\bas
\frac{d}{dt} \io \zeta^2 \ueps^p + \io \zeta^2 \ueps^p
\le \heps(t)
\qquad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1),
\eas
so that since $c_2=c_2(p,\delta):=\sup_{\eps\in [0,1)} \sup_{t\in (0,\tme-\te)} \int_t^{t+\te} \heps(s) ds$ with $\te=\min\{1,\frac{1}{2}\tme\}$ is finite according to Lemma \ref{lem6} and the fact that $p_0<3$, by using an ODE comparison argument along with Lemma~\ref{lem77}
we infer that
\bas
\io \zeta^2 \ueps^p
&\le& e^{-t} \cdot \io \zeta^2 (u^{(0)})^p
+ \int_0^t e^{-(t-s)} \heps(s) ds \\
&\le& \io (u^{(0)})^p
+ \frac{c_2\te}{1-e^{-\te}}
\qquad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1),
\eas
and hence conclude as intended, because $\frac{\tau}{1-e^{-\tau}} \le \frac{1}{1-e^{-1}}$ for all $\tau\in (0,1]$,
and because $\zeta\equiv 1$ in $\Om\setminus B_\delta(0)$.
\end{proof}
With these bounds at hand, we can even estimate the derivative of the second component uniformly, again
outside a neighbourhood of the origin.
\begin{lem}\label{lem9}
For each $\delta\in (0,R)$ there exists $C(\delta)>0$ satisfying
\be{9.1}
|v_{\eps r}(r,t)| \le C(\delta)
\qquad \mbox{for all $r\in [\delta,R]$, $t\in (0,\tme)$ and } \eps\in [0,1).
\ee
\end{lem}
\begin{proof}
We once more take $\chi\in C^\infty([0,R])$ such that $0\le\chi\le 1$ and $\chi|_{[0,\frac{\delta}{2}]}\equiv 0$
as well as $\chi|_{[\delta,R]} \equiv 1$, and then infer from Lemma \ref{lem8} in conjunction with (\ref{F}), (\ref{vinfty}) and
Lemma \ref{lem5} that there exists $c_1=c_1(\delta)>0$ such that with
$(\beps)_{\eps\in [0,1)} \equiv (\beps^{(\chi)})_{\eps \in [0,1)}$
as defined in (\ref{b}) we have
\bas
\|\beps(\cdot,t)\|_{L^2((\frac{\delta}{2},R))} \le c_1
\qquad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1).
\eas
As for the Dirichlet heat semigroup $(e^{-tA})_{t\ge 0}$ on $(\frac{\delta}{2},R))$ it is
known (\cite{eidelman_ivasishen}, \cite{quittner_souplet}) that there exist $\lam=\lam(\delta)>0$,
$c_2=c_2(\delta)>0$ and $c_3=c_3(\delta)>0$ with the property that for all $t>0$,
\bas
\|\partial_r e^{-tA} \varphi\|_{L^\infty((\frac{\delta}{2},R))}
\le c_2\|\varphi\|_{W^{1,\infty}((\frac{\delta}{2},R))}
\qquad \mbox{for all $\varphi\in W^{1,\infty}((\frac{\delta}{2},R))$ such that
$\varphi(\frac{\delta}{2})=\varphi(R)=0$}
\eas
as well as
\bas
\|\partial_r e^{-tA} \varphi\|_{L^\infty((\frac{\delta}{2},R))}
\le c_3\cdot (1+t^{-\frac{3}{4}}) e^{-\lam t} \|\varphi\|_{L^2((\frac{\delta}{2},R))}
\quad \mbox{for all $\varphi\in C^0([\frac{\delta}{2},R])$ with
$\varphi(\frac{\delta}{2})=\varphi(R)=0$,}
\eas
from (\ref{chiv}) we obtain that
\bas
\|v_{\eps r}(\cdot,t)\|_{L^\infty((\delta,R))}
&\le& \Big\|\partial_r \Big\{ \chi \cdot \big(\veps(\cdot,t)-\vst\big) \Big\} \Big\|_{L^\infty((\frac{\delta}{2},R))} \\
&\le& c_2 \big\| \chi\cdot (v^{(0)}-\vst)\big\|_{W^{1,\infty}((\frac{\delta}{2},R))}
+ c_1 c_3 \int_0^\infty (1+\sigma^{-\frac{3}{4}}) e^{-\lam\sigma} d\sigma
\eas
for all $t\in (0,\tme)$ and any $\eps\in [0,1)$.
\end{proof}
In conclusion, this allows for appropriately controlling all the boundary integrals that will turn out to appear
in the course of our subsequent energy analysis:
\begin{cor}\label{cor10}
There exists $C>0$ such that
\be{10.1}
\int_t^{t+\te} \int_{\pO} \frac{1}{\veps} \frac{\partial |\na\veps|^2}{\partial\nu} \le C
\qquad \mbox{for all $t\in [0,\tme-\te)$}
\ee
and
\be{10.2}
\int_t^{t+\te} \int_{\pO} |\na\veps|^2 \frac{\partial |\na\veps|^2}{\partial\nu} \le C
\qquad \mbox{for all $t\in [0,\tme-\te)$}
\ee
as well as
\be{10.3}
\bigg| \int_{\pO} \frac{|\na\veps|^2}{\veps^2} \frac{\partial\veps}{\partial\nu} \bigg| \le C
\qquad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1),
\ee
where, as before, $\te=\min\{1,\frac{1}{2}\tme\}$ for $\eps\in [0,1)$.
\end{cor}
\begin{proof}
Once more explicitly relying on radial symmetry, we may use that according to Lemma \ref{lem1} the second equation in (\ref{0eps})
holds up to $\pO$ throughout $(0,\tme)$, which namely ensures that on $\pO$ we have the one-sided inequality
\bas
\frac{1}{\veps} \cdot \frac{\partial |\na\veps|^2}{\partial\nu}
&=& \frac{2}{\veps} v_{\eps r} v_{\eps rr} \\
&=& \frac{2}{\veps} v_{\eps r} \cdot \Big\{ v_{\eps rr} + \frac{n-1}{R} v_{\eps r} \Big\}
- \frac{2(n-1)}{R\veps} v_{\eps r}^2 \\
&=& \frac{2}{\veps} \Feps(\ueps) \veps v_{\eps r}
- \frac{2(n-1)}{R\veps} v_{\eps r}^2 \\
&\le& 2\Feps(\ueps) v_{\eps r}
\qquad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1).
\eas
Again thanks to (\ref{F}), this implies that
\bas
\frac{1}{2} \int_t^{t+\te} \int_{\pO} \frac{1}{\veps} \frac{\partial |\na\veps|^2}{\partial\nu}
&\le& |\pO| \cdot \int_t^{t+\te} \ueps(R,s) \cdot |v_{\eps r}(R,s)| ds \\
&\le& |\pO| \cdot \|v_{\eps r}\|_{L^\infty((\frac{R}{2},R)) \times (0,\tme))}
\cdot \int_t^{t+\te} \ueps(R,s) ds \\[2mm]
& & \hs{20mm}
\qquad \mbox{for all $t\in [0,\tme-\te)$ and } \eps\in [0,1),
\eas
and that, similarly,
\bas
\int_t^{t+\te} \int_{\pO} |\na\veps|^2 \frac{\partial |\na\veps|^2}{\partial\nu}
&\le& 2 |\pO| \vst \cdot \|v_{\eps r}\|_{L^\infty((\frac{R}{2},R)) \times (0,\tme))}^3
\cdot \int_t^{t+\te} \ueps(R,s) ds \\[2mm]
& & \hs{20mm}
\qquad \mbox{for all $t\in [0,\tme-\te)$ and } \eps\in [0,1),
\eas
so that since furthermore
\bas
\bigg| \int_{\pO} \frac{|\na\veps|^2}{\veps^2} \frac{\partial\veps}{\partial\nu} \bigg|
\le \frac{|\pO|}{2\vst^2} \|v_{\eps r}\|_{L^\infty((\frac{R}{2},R)) \times (0,\tme))}^3
\qquad \mbox{for all $t\in (0,\tme)$ and } \eps\in [0,1),
\eas
the claim results from Lemma \ref{lem9} when combined with Corollary \ref{cor7}.
\end{proof}
\mysection{Energy analysis}\label{sect4}
Our approach toward deriving a suitable relative of (\ref{energy}) is now launched by the following observation.
\begin{lem}\label{lem3}
Let $\eps\in [0,1)$. Then
\bea{3.1}
& & \hs{-16mm}
\frac{d}{dt} \bigg\{ \io \ueps\ln\ueps + \frac{1}{2} \io \frac{|\na\veps|^2}{\veps} \bigg\}
+ \io \frac{|\na\ueps|^2}{\ueps}
+ \io \veps |D^2\ln\veps|^2 \nn\\
&=&
-\frac{1}{2} \io \frac{\Feps(\ueps)}{\veps} |\na\veps|^2
+ \frac{1}{2} \int_{\pO} \frac{1}{\veps} \frac{\partial |\na\veps|^2}{\partial\nu}
- \frac{1}{2} \int_{\pO} \frac{|\na\veps|^2}{\veps^2} \frac{\partial\veps}{\partial\nu}
\qquad \mbox{for all $t\in (0,\tme)$.}
\eea
\end{lem}
\begin{proof}
According to the no-flux boundary condition accompanying the first equation in (\ref{0eps}),
\be{3.2}
\frac{d}{dt} \io \ueps\ln\ueps
+ \io \frac{|\na\ueps|^2}{\ueps}
= \io \Feps'(\ueps) \na\ueps\cdot\na\veps
\qquad \mbox{for all } t\in (0,\tme),
\ee
while on the basis of the second equation in (\ref{0eps}) we first compute
\bea{3.3}
\frac{1}{2} \frac{d}{dt} \io \frac{|\na\veps|^2}{\veps}
&=& \io \frac{1}{\veps} \na\veps\cdot\na \Big\{ \Del\veps - \Feps(\ueps)\veps\Big\}
- \frac{1}{2} \io \frac{|\na\veps|^2}{\veps^2} \cdot \Big\{ \Del\veps-\Feps(\ueps)\veps\Big\} \nn\\
&=& \frac{1}{2} \io \frac{1}{\veps} \Del|\na\veps|^2
- \io \frac{1}{\veps} |D^2\veps|^2
- \frac{1}{2} \io \frac{|\na\veps|^2}{\veps^2} \Del\veps \nn\\
& & - \frac{1}{2} \io \frac{\Feps(\ueps)}{\veps} |\na\veps|^2
- \io \Feps'(\ueps) \na\ueps\cdot\na\veps
\qquad \mbox{for all } t\in (0,\tme),
\eea
because $\na\veps\cdot\na\Del\veps=\frac{1}{2} \Del |\na\veps|^2 - |D^2\veps|^2$.
Here two integrations by parts show that
\bas
\frac{1}{2} \io \frac{1}{\veps} \Del |\na\veps|^2
= \frac{1}{2} \io \frac{1}{\veps^2} \na\veps\cdot\na|\na\veps|^2
+ \frac{1}{2} \int_{\pO} \frac{1}{\veps} \frac{\pa |\na\veps|^2}{\pa\nu}
\qquad \mbox{for all } t\in (0,\tme),
\eas
and that
\bas
- \frac{1}{2} \io \frac{|\na\veps|^2}{\veps^2} \Del\veps
= \frac{1}{2} \io \frac{1}{\veps^2} \na\veps\cdot\na |\na\veps|^2
- \io \frac{|\na\veps|^4}{\veps^3}
- \frac{1}{2} \int_{\pO} \frac{|\na\veps|^2}{\veps^2} \frac{\pa\veps}{\pa\nu}
\quad \mbox{for all } t\in (0,\tme),
\eas
so that since $\na |\na\veps|^2 = 2D^2\veps\cdot\na\veps$, we obtain that
\bas
& & \hs{-20mm}
\frac{1}{2} \io \frac{1}{\veps} \Del |\na\veps|^2
- \io \frac{1}{\veps} |D^2\veps|^2
- \frac{1}{2} \io \frac{|\na\veps|^2}{\veps^2} \Del\veps
- \frac{1}{2} \int_{\pO} \frac{1}{\veps} \frac{\pa |\na\veps|^2}{\pa\nu}
+ \frac{1}{2} \int_{\pO} \frac{|\na\veps|^2}{\veps^2} \frac{\pa\veps}{\pa\nu} \nn\\
&=& - \io \frac{1}{\veps} |D^2\veps|^2
+ 2\io \frac{1}{\veps^2} \na\veps\cd (D^2\veps\cd\na\veps)
- \io \frac{|\na\veps|^4}{\veps^3} \\
&=& - \sum_{i,j=1}^n \io \veps \cd \bigg| \frac{\pa_{x_i x_j} \veps}{\veps}
- \frac{\pa_{x_i} \veps \pa_{x_j}\veps}{\veps^2} \bigg|^2 \\
&=& - \sum_{i,j=1}^n \io \veps \cd \Big| \pa_{x_i x_j} \ln \veps \Big|^2 \\
&=& - \io \veps |D^2 \ln\veps|^2
\qquad \mbox{for all } t\in (0,\tme).
\eas
Therefore, (\ref{3.3}) is equivalent to (\ref{3.1}).
\end{proof}
In order to make use of the last term on the left of \eqref{3.1}, we will employ the following variant of
a functional inequality which for functions with vanishing normal derivative on $\pO$ has been
documented in \cite[Lemma~3.3]{win_CPDE12}.
\begin{lem}\label{lem21}
Let $\vp\in C^2(\bom)$ be such that $\vp>0$ in $\bom$. Then
\be{21.1}
\io \frac{|\na\vp|^4}{\vp^3}
\le (2+\sqrt{n})^2 \io \vp |D^2\ln \vp|^2
+ 2 \int_{\pO} \frac{|\na\vp|^2}{\vp^2} \frac{\pa\vp}{\pa\nu}.
\ee
\end{lem}
\begin{proof}
We integrate by parts to see that
\bas
\io \frac{|\na\vp|^4}{\vp^3}
&=& \io |\na\ln \vp|^2 \na\ln \vp\cd \na\vp \\
&=& - \io \vp\na\ln\vp\cd \na |\na\ln\vp|^2
- \io \vp |\na\ln \vp|^2 \Del \ln \vp
+ \int_{\pO} \vp |\na\ln \vp|^2 \frac{\pa\ln\vp}{\pa\nu} \\
&=& -2 \io \frac{1}{\vp} \na\vp \cdot (D^2\ln\vp \cdot\na\vp)
- \io \frac{|\na\vp|^2}{\vp} \Del\ln\vp
+ \int_{\pO} \frac{|\na\vp|^2}{\vp^2} \frac{\pa\vp}{\pa\nu}.
\eas
As
\bas
-2 \io \frac{1}{\vp} \na\vp \cdot (D^2\ln\vp \cdot\na\vp)
- \io \frac{|\na\vp|^2}{\vp} \Del\ln\vp
&\le& (2+\sqrt{n}) \io \frac{|\na\vp|^2}{\vp} |D^2\ln\vp| \\
&\le& \frac{1}{2} \io \frac{|\na\vp|^4}{\vp^3}
+ \frac{(2+\sqrt{n})^2}{2} \io \vp |D^2\ln\vp|^2
\eas
by Young's inequality, this implies (\ref{21.1}).
\end{proof}
Now an exploitation of the latter in the context of (\ref{3.1}) shows that the boundary regularity features obtained
in Corollary \ref{cor10} imply the following spatially global estimates.
\begin{lem}\label{lem11}
There exists $C>0$ such that for each $\eps\in [0,1)$, writing $\te=\min\{1,\frac{1}{2}\tme\}$, we have
\be{11.1}
\io \ueps(\cdot,t) \ln \ueps(\cdot,t) \le C
\qquad \mbox{for all $t\in (0,\tme)$}
\ee
and
\be{11.01}
\io |\na\veps(\cdot,t)|^2 \le C
\qquad \mbox{for all $t\in (0,\tme)$}
\ee
as well as
\be{11.2}
\int_t^{t+\te} \io \frac{|\na\ueps|^2}{\ueps} \le C
\qquad \mbox{for all $t\in [0,\tme-\te)$}
\ee
and
\be{11.4}
\int_t^{t+\te} \io |\na\veps|^4 \le C
\qquad \mbox{for all $t\in [0,\tme-\te)$.}
\ee
\end{lem}
\begin{proof}
We first employ the Gagliardo-Nirenberg inequality to pick $c_1>0$ such that
\be{11.5}
\|\varphi\|_{L^\frac{2(n+2)}{n}(\Om)}^\frac{2(n+2)}{n}
\le c_1\|\na\varphi\|_{L^2(\Om)}^2 \|\varphi\|_{L^2(\Om)}^\frac{4}{n}
+ c_1 \|\varphi\|_{L^2(\Om)}^\frac{2(n+2)}{n}
\qquad \mbox{for all } \varphi\in W^{1,2}(\Om),
\ee
and use that $\frac{\xi\ln\xi}{\xi^\frac{n+2}{n}} \to 0$ as $\xi\to +\infty$ in choosing $c_2>0$ such that abbreviating
$c_3:=\io u^{(0)}$ we have
\be{11.6}
\xi\ln\xi \le \frac{2}{c_1 c_3^\frac{2}{n}} \xi^\frac{n+2}{n} + c_2
\qquad \mbox{for all } \xi>0.
\ee
Then writing $c_4:=\frac{1}{(2+\sqrt{n})^2}$ and $c_5:=\frac{|\Om|}{8c_4} \|v^{(0)}\|_{L^\infty(\Om)}$,
by means of (\ref{11.6}), Young's inequality, (\ref{11.5}), (\ref{mass}) and (\ref{vinfty}) we see that for each $\eps\in [0,1)$,
\bas
\yeps(t):=\io \ueps(\cdot,t)\ln\ueps(\cdot,t) + \frac{1}{2} \io \frac{|\na\veps(\cdot,t)|^2}{\veps(\cdot,t)},
\qquad t\in [0,\tme),
\eas
has the property that
\bas
\yeps(t)
&\le& \frac{2}{c_1 c_3^\frac{2}{n}} \io \ueps^\frac{n+2}{n} + c_2 |\Om|
+ \frac{c_4}{2} \io \frac{|\na\veps|^4}{\veps^3}
+ \frac{1}{8c_4} \io \veps \\
&=& \frac{2}{c_1 c_3^\frac{2}{n}} \|\sqrt{\ueps}\|_{L^\frac{2(n+2)}{n}(\Om)}^\frac{2(n+2)}{n} + c_2|\Om|
+ \frac{c_4}{2} \io \frac{|\na\veps|^4}{\veps^3}
+ \frac{1}{8c_4} \io \veps \\
&\le& \frac{2}{c_3^\frac{2}{n}} \|\na\sqrt{\ueps}\|_{L^2(\Om)}^2 \|\sqrt{\ueps}\|_{L^2(\Om)}^\frac{4}{n}
+ \frac{2}{c_3^\frac{2}{n}} \|\sqrt{\ueps}\|_{L^2(\Om)}^\frac{2(n+2)}{n}
+ c_2|\Om|
+ \frac{c_4}{2} \io \frac{|\na\veps|^4}{\veps^3}
+ c_5 \\
&=& \frac{1}{2} \io \frac{|\na\ueps|^2}{\ueps}
+ 2c_3 + c_2|\Om|
+ \frac{c_4}{2} \io \frac{|\na\veps|^4}{\veps^3}
+ c_5
\qquad \mbox{for all } t\in (0,\tme),
\eas
so that since Lemma \ref{lem21} warrants that
\bas
c_4 \io \frac{|\na\veps|^4}{\veps^3}
\le \io \veps |D^2\ln\veps|^2
+ 2c_4 \int_{\pO} \frac{|\na\veps|^2}{\veps^2} \frac{\pa\veps}{\pa\nu}
\qquad \mbox{for all } t\in (0,\tme),
\eas
it follows that
\bea{11.7}
\hs{-6mm}
\yeps(t)
+ \frac{1}{2} \io \frac{|\na\ueps|^2}{\ueps}
+ \frac{c_4}{2} \io \frac{|\na\veps|^4}{\veps^3}
&\le& \io \frac{|\na\ueps|^2}{\ueps}
+ \io \veps |D^2\ln\veps|^2 \nn\\
& & + 2c_4 \int_{\pO} \frac{|\na\veps|^2}{\veps^2} \frac{\pa\veps}{\pa\nu}
+ c_6
\qquad \mbox{for all } t\in (0,\tme),
\eea
with $c_6:=2c_3 + c_2|\Om|+c_5$.
Accordingly, from (\ref{3.1}) we infer upon dropping a favorably signed summand therein that
\bea{11.8}
& & \hs{-17mm}
\yeps'(t)+\yeps(t)
+\frac{1}{2} \io \frac{|\na\ueps|^2}{\ueps}
+ \frac{c_4}{2} \io \frac{|\na\veps|^4}{\veps^3} \nn\\
& & \hs{-5mm}
\le \heps(t)
:= \frac{1}{2} \int_{\pO} \frac{1}{\veps} \frac{\pa |\na\veps|^2}{\pa\nu}
+ \Big(2c_4-\frac{1}{2}\Big) \int_{\pO} \frac{|\na\veps|^2}{\veps^2} \frac{\pa\veps}{\pa\nu}
+ c_6
\qquad \mbox{for all } t\in (0,\tme).
\eea
As Corollary \ref{cor10} provides $c_7>0$ such that
\be{11.9}
\int_t^{t+\te} \heps(s) ds \le c_7
\qquad \mbox{for all $t\in [0,\tme-\te)$ and } \eps\in [0,1),
\ee
through Lemma \ref{lem77} this firstly ensures that for all $t\in (0,\tme)$ and $\eps\in [0,1)$,
\bea{11.10}
\yeps(t)
&\le& \yeps(0) e^{-t} + \int_0^t e^{-(t-s)} \heps(s) ds \nn\\
&\le& |\yeps(0)| + \frac{c_7\te}{1-e^{-\te}} \nn\\
&\le& c_8:= \io u^{(0)}|\ln u^{(0)}| + \frac{1}{2} \io \frac{|\na v^{(0)}|^2}{v^{(0)}}
+ \frac{c_7}{1-e^{-1}},
\eea
again because $\frac{\te}{1-e^{-\te}} \le \frac{1}{1-e^{-1}}$ for all $\eps\in [0,1)$.
Going back to (\ref{11.8}), from this we thereupon infer that
\bea{11.11}
& & \hs{-30mm}
\frac{1}{2} \int_t^{t+\te} \io \frac{|\na\ueps|^2}{\ueps}
+ \frac{c_4}{2} \int_t^{t+\te} \io \frac{|\na\veps|^4}{\veps^3} \nn\\
&\le& \yeps(t) - \yeps(t+\te)
- \int_t^{t+\te} \yeps(s) ds
+ \int_t^{t+\te} \heps(s) ds \nn\\
&\le& c_8 + \frac{2|\Om|}{e} + c_7
\qquad \mbox{for all $t\in [0,\tme-\te)$ and } \eps\in [0,1),
\eea
since evidently
\be{11.12}
\yeps(t) \ge \io \ueps\ln \ueps \ge - \frac{|\Om|}{e}
\qquad \mbox{for all $t\in [0,\tme)$ and } \eps\in [0,1).
\ee
Once more relying on (\ref{vinfty}), from (\ref{11.11}) we obtain both (\ref{11.2}) and (\ref{11.4}), whereas (\ref{11.1})
and (\ref{11.01}) similarly result from (\ref{11.10}) due to the second inequality in (\ref{11.12}).
\end{proof}
\mysection{The two-dimensional case. Proof of Theorem \ref{theo16}}
In this section we concentrate on the two-dimensional setting of Theorem~\ref{theo16}.
Since the solutions there will already turn out to be bounded and classical, it is not necessary to resort to an approximation by means of \eqref{0eps} for $ε>0$. Throughout this section, we will therefore directly address the solutions $(u,v):=(u_0,v_0)$ of \eqref{0eps}
obtained for $\eps=0$.\abs
Based on the information provided by Lemma \ref{lem11}, we can combine the outcomes of two further testing procedures applied
to (\ref{0eps}) in quite a standard manner, and thereby achieve the following key toward higher order bounds:
\begin{lem}\label{lem14}
Let $n=2$. Then there exists $C>0$ such that the solution $(u,v)\equiv (u_0,v_0)$ of (\ref{0eps}), as corresponding
to the choice $\eps=0$, satisfies
\be{14.1}
\io |\na v(\cdot,t)|^4 \le C
\qquad \mbox{for all } t\in (0,\tm),
\ee
where $\tm:=T_{max,0}$ is as accordingly provided by Lemma \ref{lem1}.
\end{lem}
\begin{proof}
On the basis of (\ref{0eps}) when restricted to $\eps=0$, by means of Young's inequality we see that
\be{14.2}
\frac{d}{dt} \io u^2 + \io |\na u|^2
= - \io |\na u|^2
+ 2\io u\na u\cdot\na v
\le \io u^2 |\na v|^2
\qquad \mbox{for all } t\in (0,\tm),
\ee
and that for all $t\in (0,\tm)$,
\bea{14.3}
\frac{1}{4} \frac{d}{dt} \io |\na v|^4
&=& \io |\na v|^2 \na v\cdot\na \big\{ \Del v - uv \big\} \nn\\
&=& \frac{1}{2} \io |\na v|^2 \Del |\na v|^2
- \io |\na v|^2 |D^2 v|^2
- \io u |\na v|^4
- \io v |\na v|^2 \na u\cdot\na v \nn\\
&=& - \frac{1}{2} \io \Big|\na |\na v|^2 \Big|^2
+ \frac{1}{2} \int_{\pO} |\na v|^2 \frac{\pa |\na v|^2}{\pa\nu} \nn\\
& & - \io |\na v|^2 |D^2 v|^2
- \io u |\na v|^4
- \io v |\na v|^2 \na u\cdot\na v \nn\\
&\le& - \frac{1}{2} \io \Big| \na |\na v|^2 \Big|^2
+ \frac{1}{2} \int_{\pO} |\na v|^2 \frac{\pa |\na v|^2}{\pa\nu}
- \io v |\na v|^2 \na u\cdot\na v \nn\\
&\le& - \frac{1}{2} \io \Big| \na |\na v|^2 \Big|^2
+ \frac{1}{2} \int_{\pO} |\na v|^2 \frac{\pa |\na v|^2}{\pa\nu}
+ \|v_0\|_{L^\infty(\Om)} \io |\na u| \cd |\na v|^3
\eea
because of (\ref{vinfty}).
To proceed from this, we employ the Gagliardo-Nirenberg inequality to find $c_1>0$ such that
\bea{14.4}
\io |\na v|^6
&=& \Big\| |\na v|^2 \Big\|_{L^3(\Om)}^3 \nn\\
&\le& c_1 \Big\| \na |\na v|^2 \Big\|_{L^2(\Om)}^2
\Big\| |\na v|^2 \Big\|_{L^1(\Om)}
+ c_1 \Big\| |\na v|^2 \Big\|_{L^1(\Om)}^3 \nn\\
&\le& c_1 c_2 \Big\| \na |\na v|^2 \Big\|_{L^2(\Om)}^2
+ c_1 c_2^3
\qquad \mbox{for all } t\in (0,\tm),
\eea
with finiteness of $c_2:=\sup_{t\in (0,\tm)} \io |\na v(\cdot,t)|^2$ being asserted by Lemma \ref{lem11}.
We then fix $a>0$ suitably small such that
\be{14.5}
8a^2 \|v^{(0)}\|_{L^\infty(\Om)}^2
\le \frac{a}{c_1 c_2},
\ee
and take $\eta>0$ small enough fulfilling
\be{14.6}
\frac{1}{2\eta} \ge 1 + \Big( \frac{2c_1 c_2}{a}\Big)^\frac{1}{2},
\ee
whereupon an application of a well-known variant of the Gagliardo-Nirenberg inequality (\cite{biler_hebisch_nadzieja})
shows that since Lemma \ref{lem11} warrants boundedness of $(u(\cdot,t)\ln u(\cdot,t))_{t\in (0,\tm)}$ in $L^1(\Om)$,
there exists $c_3>0$ such that
\be{14.7}
\io u^3 \le \eta \io |\na u|^2 + c_3
\qquad \mbox{for all } t\in (0,\tm).
\ee
We now let
\bas
y(t):=\io u^2(\cdot,t) + a \io |\na v(\cdot,t)|^4,
\qquad t\in [0,\tm),
\eas
and combine (\ref{14.2}) with (\ref{14.3}) to obtain that since
\bas
\io |\na u|^2 \ge \frac{1}{2} \io |\na u|^2 + \frac{1}{2\eta} \io u^3 - \frac{c_3}{2\eta}
\qquad \mbox{for all } t\in (0,\tm)
\eas
and
\bas
2 \io \Big|\na |\na v|^2 \Big|^2
\ge \frac{2}{c_1c_2} \io |\na v|^6 - 2c_2^2
\qquad \mbox{for all } t\in (0,\tm)
\eas
by (\ref{14.7}) and (\ref{14.4}), we have
\bea{14.8}
& & \hs{-20mm}
y'(t) + y(t)
+ \frac{1}{2} \io |\na u|^2
+ \frac{1}{2\eta} \io u^3
+ \frac{2a}{c_1c_2} \io |\na v|^6 \nn\\
&\le& \io u^2 + a\io |\na v|^4 \nn\\
& & + \frac{c_3}{2\eta} + 2ac_2^2 \nn\\
& & + \io u^2 |\na v|^2 \nn\\
& & + 4a \|v^{(0)}\|_{L^\infty(\Om)} \io |\na u| \cdot |\na v|^3 \nn\\
& & + 2a \int_{\pO} |\na v|^2 \frac{\pa |\na v|^2}{\pa\nu}
\qquad \mbox{for all } t\in (0,\tm).
\eea
Here due to Young's inequality,
\bas
& & \hs{-20mm}
\io u^2
+ a \io |\na v|^4
+ \io u^2 |\na v|^2
+ 4a \|v^{(0)}\|_{L^\infty(\Om)} \io |\na u|\cd |\na v|^3 \nn\\
&\le& \io u^2 + a\io |\na v|^4 + \io u^2 |\na v|^2
+ \frac{1}{2} \io |\na u|^2
+ 8a^2 \|v^{(0)}\|_{L^\infty(\Om)}^2 \io |\na v|^6 \\
&=& \io u^2 \\
& & + \io \Big\{ \frac{a}{2c_1 c_2} |\na v|^6 \Big\}^\frac{2}{3} \cdot (2c_1 c_2)^\frac{2}{3} a^\frac{1}{3} \\
& & + \io \Big\{ \frac{a}{2c_1 c_2} |\na v|^6 \Big\}^\frac{1}{3} \cdot \Big(\frac{2c_1c_2}{a}\Big)^\frac{1}{3} u^2 \\
& &
+ \frac{1}{2} \io |\na u|^2
+ 8a^2 \|v^{(0)}\|_{L^\infty(\Om)}^2 \io |\na v|^6 \\
&\le& \io u^3 + |\Om| \\
& &
+ \frac{a}{2c_1c_2} \io |\na v|^6
+ (2c_1c_2)^2 a|\Om| \\
& & + \frac{a}{2c_1c_2} \io |\na v|^6
+ \Big(\frac{2c_1 c_2}{a}\Big)^\frac{1}{2} \io u^3 \\
& & + \frac{1}{2} \io |\na u|^2
+ 8a^2 \|v^{(0)}\|_{L^\infty(\Om)}^2 \io |\na v|^6 \\
&=& \Big\{ 1 + \Big(\frac{2c_1c_2}{a}\Big)^\frac{1}{2}\Big\} \cdot \io u^3
+ \Big\{ \frac{a}{c_1c_2} + 8a^2 \|v^{(0)}\|_{L^\infty(\Om)}^2 \Big\} \cdot \io |\na v|^6 \\
& & + |\Om| + (2c_1c_2)^2 a|\Om| \\
& & + \frac{1}{2} \io |\na u|^2
\qquad \mbox{for all } t\in (0,\tm),
\eas
whence drawing on (\ref{14.6}) and (\ref{14.5}) we infer from (\ref{14.8}) that
\bas
y'(t) + y(t)
\le h(t):=\frac{c_3}{2\eta} + 2ac_2^2 + |\Om| + (2c_1c_2)^2 a|\Om|
+ 2a\int_{\pO} |\na v|^2 \frac{\pa |\na v|^2}{\pa\nu}
\qquad \mbox{for all } t\in (0,\tm).
\eas
Since Corollary \ref{cor10} ensures that $\sup_{t\in (0,\tm-\tau_0)} \int_t^{t+\tau_0} h(s)ds$, with $τ_0=\min\{1,\frac12\tm\}$, is finite, by means of
Lemma \ref{lem77} this entails that $y$ is bounded in $(0,\tm)$, which in particular implies (\ref{14.1})
with some suitably large $C>0$.
\end{proof}
Indeed, this implies boundedness in the respective first solution components.
\begin{lem}\label{lem15}
Let $n=2$. Then there exists $C>0$ such that with $(u,v)\equiv (u_0,v_0)$ and $\tm=T_{max,0}$ taken from Lemma \ref{lem1}
we have
\be{15.1}
\|u(\cdot,t)\|_{L^\infty(\Om)} \le C
\qquad \mbox{for all } t\in (0,\tm).
\ee
\end{lem}
\begin{proof}
We write the first equation in (\ref{0eps}) for $\eps=0$ in the form
$u_t = \Delta u + \na\cdot (b(x,t) u)$, $(x,t)\in\Om\times (0,\tm)$, and note that according to Lemma \ref{lem14},
$b:=-\na v$ belongs to $L^\infty((0,\tm);L^q(\Om))$ with $q:=4$ exceeding the considered spatial dimension.
Since $(\na u+b(x,t)u)\cdot\nu=0$ on $\pO\times (0,\tm)$, we may therefore refer to a boundedness statement derived by means
of a straightforward Moser-type iteration (\cite{ding_win}) to directly obtain (\ref{15.1}).
\end{proof}
For the proof of Theorem~\ref{theo16}, we are merely lacking a
transfer of the boundedness properties we have just obtained
to the spaces that actually occur in the extensibility criterion \eqref{ext}:
\begin{proof}[Proof of Theorem \ref{theo16}]
Based on the outcome of Lemma \ref{lem15}, we may again utilize known smoothing properties of the Dirichlet heat semigroup
on $\Omega$, and additionally employ a standard result on gradient H\"older
regularity in scalar parabolic equations (\cite{lieberman}),
to find $c_1>0, c_2>0$ and $\theta_1\in (0,1)$ such that
\be{16.2}
\|v(\cdot,t)\|_{W^{1,\infty}(\Om)} \le c_1
\qquad \mbox{for all } t\in (0,\tm)
\ee
and
\bas
\|\na v\|_{C^{\theta_1,\frac{\theta_1}{2}}(\bom\times [t,t+\tau_0])} \le c_2
\qquad \mbox{for all $t\in (\frac{\tau_0}{4},\tm-\tau_0)$,}
\eas
where again $\tau_0=\min\{1,\frac{1}{2}\tm\}$.
According to (\ref{16.2}), we may thereafter rely on the latter token once again to infer from Lemma \ref{lem15} and the first
sub-problem in (\ref{0eps}) that with some $c_3>0$ and $\theta_2\in (0,1)$ we have
\bas
\|u\|_{C^{1+\theta_2,\theta_2}(\bom\times [t,t+\tau_0])} \le c_3
\qquad \mbox{for all $t\in (\frac{\tau_0}{2},\tm-\tau_0)$,}
\eas
which combined with (\ref{16.2}) and (\ref{ext}) shows that Lemma \ref{lem1} indeed asserts that $\tm=\infty$, whereupon
(\ref{16.1}) becomes a consequence of Lemma \ref{lem15} and (\ref{16.2}).
\end{proof}
\mysection{The case \texorpdfstring{$n\ge 3$}{n > 2}. Proof of Theorem \ref{theo25}}
The solution concept to be pursued in higher-dimensional cases appears to be quite natural.
\begin{defi}\label{dw}
Let
\be{w1}
\left\{ \begin{array}{l}
u\in L^1_{loc}([0,\infty);W^{1,1}(\Om)) \qquad \mbox{and} \\[1mm]
v\in L^1_{loc}([0,\infty);W^{1,1}(\Om))
\end{array} \right.
\ee
be nonnegative and such that $v(\cdot,t)-\vst\in W_0^{1,1}(\Om)$ for a.e.~$t>0$, and that
\be{w2}
u\na v\in L^1_{loc}(\bom\times [0,\infty);\R^n)
\qquad \mbox{and} \qquad
uv\in L^1_{loc}(\bom\times [0,\infty)).
\ee
Then $(u,v)$ will be called a {\em global weak solution of (\ref{0})} if
\be{wu}
-\int_0^\infty \io u\vp_t - \io u^{(0)}\vp(\cdot,0)
= - \int_0^\infty \io \na u\cdot\na\vp
+ \int_0^\infty \io u\na v\cdot\na\vp
\ee
for all $\vp\in C_0^\infty(\bom\times [0,\infty))$, and if
\be{wv}
-\int_0^\infty \io v\vp_t - \io v^{(0)}\vp(\cdot,0)
= - \int_0^\infty \io \na v\cdot\na\vp
- \int_0^\infty uv\vp
\ee
for all $\vp\in C_0^\infty(\Om\times [0,\infty))$.
\end{defi}
In order to construct such solutions, we now utilize solutions of the approximate versions of \eqref{0eps}, that is, those
corresponding to positive values of $\eps$.
As can easily be seen, the strength of the accordingly regularizing features $F_\eps$ is sufficient to ensure
that each of these solutions is global in time:
\begin{lem}\label{lem13}
Let $n\ge 3$ and $\eps\in (0,1)$. Then $\tme=\infty$.
\end{lem}
\begin{proof}
Supposing for contradiction that $\tme$ be finite for some $\eps\in (0,1)$, we note that since $0\le\Feps\le\frac{1}{\eps}$
and hence $\Feps(\ueps)\veps$ belongs to $L^\infty(\Om\times (0,\tme))$ by (\ref{vinfty}), the standard result on
parabolic $C^{1+\theta,\theta}$ regularity from \cite{lieberman} would become applicable so as to ensure that
\be{13.1}
\veps\in C^{1+\theta_1,\theta_1}(\bom\times \mbox{$[\frac{1}{4}\tme,\tme]$})
\ee
for some $\theta_1\in (0,1)$.
As $0 \le \xi\Feps'(\xi) \le \frac{1}{\eps}$ for all $\xi\ge 0$, this would especially assert boundedness of
$\ueps\Feps'(\ueps)\na\veps$ in $\Om\times (\frac{1}{4}\tme,\tme)$ and hence, firstly, warrant the inclusion
$\ueps\in L^\infty(\Om\times (\frac{1}{4}\tme,\tme))$ through the Moser iteration result from \cite{ding_win}.
Thereafter, another application of standard parabolic regularity theory (\cite{lieberman}) would entail that, in fact,
$\ueps\in C^{1+\theta_2,\theta_2}(\bom\times [\frac{1}{2}\tme,\tme])$ for some $\theta_2\in (0,1)$, which together with
(\ref{13.1}) would clearly contradict (\ref{ext}).
\end{proof}
In passing to the limit $\eps\searrow0$, we will use the following consequences of Lemma \ref{lem11} on further
spatio-temporal bounds.
\begin{lem}\label{lem22}
Let $n\ge 3$. Then there exists $C>0$ such that
\be{22.1}
\int_t^{t+1} \io \ueps^\frac{n+2}{n} \le C
\qquad \mbox{for all $t\ge 0$ and } \eps\in (0,1)
\ee
and
\be{22.2}
\int_t^{t+1} \io |\na\ueps|^\frac{n+2}{n+1} \le C
\qquad \mbox{for all $t\ge 0$ and } \eps\in (0,1).
\ee
\end{lem}
\begin{proof}
From the Gagliardo-Nirenberg inequality and (\ref{mass}) we obtain $c_1>0$ and $c_2>0$ such that
\bas
\io \ueps^\frac{n+2}{n}
= \|\sqrt{\ueps}\|_{L^\frac{2(n+2)}{n}(\Om)}^{\frac{2(n+2)}n}
&\le& c_1 \|\na\sqrt{\ueps}\|_{L^2(\Om)}^2 \|\sqrt{\ueps}\|_{L^2(\Om)}^\frac{4}{n}
+ c_1 \|\sqrt{\ueps}\|_{L^2(\Om)}^\frac{2(n+2)}{n} \\
&\le& c_2 \io \frac{|\na\ueps|^2}{\ueps} + c_2
\qquad \mbox{for all $t> 0$ and } \eps\in (0,1),
\eas
and an application of Young's inequality shows that
\bas
\io |\na\ueps|^\frac{n+2}{n+1}
&=& \io \Big\{ \frac{|\na\ueps|^2}{\ueps} \Big\}^\frac{n+2}{2(n+1)} \cdot \ueps^\frac{n+2}{2(n+1)} \\
&\le& \io \frac{|\na\ueps|^2}{\ueps} + \io \ueps^\frac{n+2}{n}
\qquad \mbox{for all $t> 0$ and } \eps\in (0,1).
\eas
Therefore, both (\ref{22.1}) and (\ref{22.2}) result from (\ref{11.2}).
\end{proof}
In preparation of an Aubin--Lions type argument, we additionally require some weak estimates for time derivatives.
\begin{lem}\label{lem23}
Let $n\in \{3,4,5\}$. Then for all $T>0$ one can find $C(T)>0$ such that
\be{23.1}
\int_0^T \|u_{\eps t}(\cdot,t)\|_{(W^{1,1+\frac{5n+2}{6-n}}(\Om))^\star}^{1+\frac{6-n}{5n+2}} dt \le C(T)
\qquad \mbox{for all } \eps\in (0,1)
\ee
and
\be{23.2}
\int_0^T \|v_{\eps t}(\cdot,t)\|_{(W_0^{1,1+\frac{(n+1)(n-2)}{3n+2}}(\Om))^\star}^\frac{n+2}{n} \le C(T)
\qquad \mbox{for all } \eps\in (0,1).
\ee
\end{lem}
\begin{proof}
We note that since $2\le n<6$, we have $1+\frac{5n+2}{6-n}=\frac{4(n+2)}{6-n}\ge n+2$, so that we can fix $c_1>0$ with the property
that $\|\na\psi\|_{L^{n+2}(\Om)} + \|\na\psi\|_{L^\frac{4(n+2)}{6-n}(\Om)} \le c_1$ for all $\psi\in C^1(\bom)$ such that
$\|\psi\|_{W^{1,1+\frac{5n+2}{6-n}}(\Om)} \le 1$.
Given any such $\psi$, using (\ref{0eps}) along with the H\"older inequality, we thus obtain that since $|\Feps'|\le 1$ for
all $\eps\in (0,1)$ by (\ref{F}),
\bas
\bigg| \io u_{\eps t} \psi \bigg|
&=& \bigg| - \io \na\ueps\cdot\na\psi
+ \io \ueps\Feps'(\ueps) \na\veps\cdot\na\psi \bigg| \\
&\le& \|\na\ueps\|_{L^\frac{n+2}{n+1}(\Om)} \|\na\psi\|_{L^{n+2}(\Om)}
+ \|\ueps\|_{L^\frac{n+2}{n}(\Om)} \|\na\veps\|_{L^4(\Om)} \|\na\psi\|_{L^\frac{4(n+2)}{6-n}(\Om)} \\
&\le& c_1 \|\na\ueps\|_{L^\frac{n+2}{n+1}(\Om)}
+ c_1 \|\ueps\|_{L^\frac{n+2}{n}(\Om)} \|\na\veps\|_{L^4(\Om)}
\qquad \mbox{for all $t>0$ and } \eps\in (0,1).
\eas
For all $t>0$ and $\eps\in (0,1)$, we therefore have
\bea{23.3}
\|u_{\eps t}\|_{(W^{1,1+\frac{5n+2}{6-n}}(\Om))^\star}^{1+\frac{6-n}{5n+2}}
&\le& (2c_1)^\frac{4n+8}{5n+2} \cdot \bigg\{ \|\na\ueps\|_{L^\frac{n+2}{n+1}(\Om)}^\frac{4n+8}{5n+2}
+ \|\ueps\|_{L^\frac{n+2}{n}(\Om)}^\frac{4n+8}{5n+2} \|\na\veps\|_{L^4(\Om)}^\frac{4n+8}{5n+2} \bigg\} \nn\\
&\le& (2c_1)^\frac{4n+8}{5n+2} \cdot \bigg\{ \|\na\ueps\|_{L^\frac{n+2}{n+1}(\Om)}^\frac{n+2}{n+1}
+ 1
+ \|\ueps\|_{L^\frac{n+2}{n}(\Om)}^\frac{n+2}{n}
+ \|\na\veps\|_{L^4(\Om)}^4 \bigg\}
\eea
by Young's inequality, because $1+\frac{6-n}{5n+2}=\frac{4n+8}{5n+2}\le \frac{n+2}{n+1}$, again due to the fact that $n\ge 2$.
Integrating (\ref{23.3}) in time shows that (\ref{23.1}) is implied by Lemma \ref{lem11} and Lemma \ref{lem22}.\abs
Likewise, to derive (\ref{23.2}) we observe that since
the inequality $n\ge 3$ warrants that $1+\frac{(n+1)(n-2)}{3n+2}=\frac{n(n+2)}{3n+2} \ge \frac{4}{3}$, and since
$W^{1,\frac{n(n+2)}{3n+2}}(\Om) \hra L^\frac{n+2}{2}(\Om)$ due to the fact that $1-\frac{3n+2}{n+2}=-\frac{2n}{n+2}$,
with some $c_2>0$ we have
$\|\na\psi\|_{L^\frac{4}{3}(\Om)} + \|\psi\|_{L^\frac{n+2}{2}(\Om)} \le c_1$ for all
$\psi\in {\cal B} :=\Big\{ \wh{\psi} \in C_0^\infty(\Om) \ \Big| \ \|\wh{\psi}\|_{W^{1,1+\frac{(n+1)(n-2)}{3n+2}}(\Om)} \le 1\Big\}$.
Therefore, the second equation in (\ref{0eps}) shows that due to the H\"older inequality, (\ref{F}) and Young's inequality,
\bas
\|v_{\eps t}\|_{W_0^{1,1+\frac{(n+1)(n-2)}{3n+2}}(\Om))^\star}^\frac{n+2}{n}
&=& \sup_{\psi\in {\cal B}}
\bigg| \io v_{\eps t} \psi \bigg|^\frac{n+2}{n} \\
&=& \sup_{\psi\in {\cal B}}
\bigg| - \io \na\veps\cdot\na\psi - \io \Feps(\ueps)\veps \psi \bigg|^\frac{n+2}{n} \\
&\le& \sup_{\psi\in {\cal B}}
\bigg\{ \|\na\veps\|_{L^4(\Om)} \|\na\psi\|_{L^\frac{4}{3}(\Om)}
+ \|\ueps\|_{L^\frac{n+2}{n}(\Om)} \|\veps\|_{L^\infty(\Om)} \|\psi\|_{L^\frac{n+2}{2}(\Om)} \bigg\}^\frac{n+2}{n} \\
&\le& c_2^\frac{n+2}{n} \cdot
\bigg\{ \|\na\veps\|_{L^4(\Om)}
+ \|\ueps\|_{L^\frac{n+2}{n}(\Om)} \|\veps\|_{L^\infty(\Om)} \bigg\}^\frac{n+2}{n} \\
&\le& (2c_2)^\frac{n+2}{n} \cdot
\bigg\{ \|\na\veps\|_{L^4(\Om)}^\frac{n+2}{n}
+ \|\ueps\|_{L^\frac{n+2}{n}(\Om)}^\frac{n+2}{n} \|\veps\|_{L^\infty(\Om)}^\frac{n+2}{n} \bigg\} \\
&\le& (2c_2)^\frac{n+2}{n} \cdot
\bigg\{ \|\na\veps\|_{L^4(\Om)}^4 + 1
+ \|\ueps\|_{L^\frac{n+2}{n}(\Om)}^\frac{n+2}{n} \|\veps\|_{L^\infty(\Om)}^\frac{n+2}{n} \bigg\}
\eas
for all $t>0$ and $\eps\in (0,1)$,
as clearly $\frac{n+2}{n} \le 4$. In view of Lemma \ref{lem11}, Lemma \ref{lem22} and (\ref{vinfty}), the inequality in (\ref{23.2})
thus results upon an integration.
\end{proof}
Our limit passage has thereby been prepared:
\begin{lem}\label{lem24}
Let $n\in \{3,4,5\}$. Then there exist $(\eps_j)_{j\in\N}\subset (0,1)$, fulfilling $\eps_j\searrow 0$ as $j\to\infty$,
as well as nonnegative functions
$u$ and $v$ on $\Om\times (0,\infty)$ which satisfy (\ref{25.01}), for which $(u(\cdot,t),v(\cdot,t))$ is radially
symmetric for a.e.~$t>0$, for which as $\eps=\eps_j\searrow 0$ we have
\begin{align}
\ueps&\to u
&&\mbox{in } \bigcap_{p\in [1,\frac{n+2}{n})} L^p_{loc}(\bom\times [0,\infty)) \mbox{ and a.e.~in } \Om\times (0,\infty),
\label{24.2} \\
\ueps&\wto u
&&\mbox{in } L^{\frac{n+2}n}_{loc}(\bom\times [0,\infty)),
\label{u-weak-1plus2overn} \\
\na\ueps&\wto\na u
&&\mbox{in } L^\frac{n+2}{n+1}_{loc}(\bom\times [0,\infty)),
\label{24.3} \\
\veps&\to v
&&\mbox{in } \bigcap_{p\in [1,\infty)} L^p_{loc}(\bom\times [0,\infty)) \mbox{ and a.e.~in } \Om\times (0,\infty)
\qquad \mbox{and}
\label{24.4} \\
\na\veps&\wto\na v
&&\mbox{in } L^4_{loc}(\bom\times [0,\infty)),
\label{24.5} \\
\na\veps&\wsto\na v
&&\mbox{in } L^\infty_{loc}((0,\infty);L^2(\Om)),
\label{24.5b}
\end{align}
and such that $(u,v)$ is a global weak solution of (\ref{0}) in the sense of Definition \ref{dw}.
\end{lem}
\begin{proof}
Given $T>0$, from Lemma \ref{lem22} and Lemma \ref{lem23} we know that
\bas
(\ueps)_{\eps\in (0,1)} \mbox{ is bounded in } L^\frac{n+2}{n+1}\big( (0,T); W^{1,\frac{n+2}{n+1}}(\Om)\big)
\eas
and that
\bas
(u_{\eps t})_{\eps\in (0,1)} \mbox{ is bounded in }
L^{1+\frac{6-n}{5n+2}} \Big( (0,T); \big(W^{1,1+\frac{5n+2}{6-n}}(\Om)\big)^\star \Big),
\eas
while according to Lemma \ref{lem11}, (\ref{vinfty}) and Lemma \ref{lem23},
\bas
(\veps-\vst)_{\eps\in (0,1)} \mbox{ is bounded in } L^4\big( (0,T); W^{1,4}(\Om)\big)
\eas
and
\bas
\big(\partial_t (\veps-\vst)\big)_{\eps\in (0,1)} \mbox{ is bounded in }
L^\frac{n+2}{n} \Big( (0,T); \big(W_0^{1,1+\frac{(n+1)(n-2)}{3n+2}}(\Om)\big)^\star \Big),
\eas
because clearly $\partial_t (\veps-\vst) = v_{\eps t}$.
Therefore, two applications of an Aubin-Lions lemma (\cite{simon}) provide $(\eps_j)_{j\in\N} \subset (0,1)$
as well as nonnegative radially symmetric functions $u\in L^\frac{n+2}{n+1}_{loc}([0,\infty);W^{1,\frac{n+2}{n+1}}(\Om))$
and $(v-\vst)\in L^4_{loc}([0,\infty);W_0^{1,4}(\Om))$ such that $ \eps_j\searrow 0$ as $j\to\infty$, that
(\ref{24.3}) and (\ref{24.5}) hold, and that
$(\ueps,\veps)\to (u,v)$ a.e.~in $\Omega\times (0,\infty)$ as
$\eps=\eps_j\searrow 0$.
Since furthermore $(\ueps)_{\eps\in (0,1)}$ is bounded in $L^\infty((0,\infty);L^1(\Om))$ and in
$L^\frac{n+2}{n}(\Om\times (0,T))$ for all $T>0$ by (\ref{mass}) and Lemma \ref{lem21}, and since
$(\veps)_{\eps\in (0,1)}$ is bounded in $L^\infty(\Om\times (0,\infty))$ and in $L^\infty((0,\infty);W^{1,2}(\Om))$
according to (\ref{vinfty}) and Lemma \ref{lem11}, it is clear that actually \eqref{u-weak-1plus2overn} and \eqref{24.5b} hold and $u$ and $v$ have all the regularity features
in (\ref{25.01}), and that the Vitali convergence theorem along with (\ref{F0}) and (\ref{F})
ensures that
\bas
\ueps\to u,
\quad \Feps(\ueps)\to u
\quad \mbox{and} \quad
\ueps\Feps'(\ueps) \to u
\qquad \mbox{in } \bigcap_{p\in [1,\frac{n+2}{n})} L^p_{loc}(\bom\times [0,\infty))
\eas
as well as
\bas
\veps\to v
\qquad \mbox{in } \bigcap_{p\in [1,\infty)} L^p_{loc}(\bom\times [0,\infty))
\eas
as $\eps=\eps_j\searrow 0$.
Besides especially completing the verification of (\ref{24.2}) and (\ref{24.4}), and hence especially also of all the
regularity requirements in Definition \ref{dw}, together with
(\ref{24.5})
these two latter properties guarantee that
\be{24.6}
\ueps\Feps'(\ueps)\na\veps \wto u\na v
\qquad \mbox{in } L^1_{loc}(\bom\times [0,\infty))
\ee
and
\be{24.7}
\Feps(\ueps)\veps \to uv
\qquad \mbox{in } L^1_{loc}(\bom\times [0,\infty))
\ee
as $\eps=\eps_j\searrow 0$, because
\bas
\frac{1}{\lim_{p\nearrow \frac{n+2}{n}} p} + \frac{1}{4} = \frac{5n+2}{4(n+2)} = 1-\frac{6-n}{4(n+2)} <1
\eas
and
\bas
\frac{1}{\lim_{p\nearrow \frac{n+2}{n}} p} + \frac{1}{\lim_{p\to \infty} p} = \frac{n}{n+2}<1.
\eas
Since for each $\eps\in (0,1)$ we have
\bas
-\int_0^\infty \io \ueps\vp_t - \io u^{(0)} \vp(\cdot,0)
= - \int_0^\infty \io \na\ueps\cdot\na\vp
+ \int_0^\infty \io \ueps\Feps'(\ueps)\na\veps\cdot\na\vp
\eas
for all $\vp\in C_0^\infty(\bom\times [0,\infty))$ and
\bas
- \int_0^\infty \io \veps \vp_t - \io v^{(0)} \vp(\cdot,0)
= - \int_0^\infty \io \na\veps\cdot\na\vp
- \int_0^\infty \io \Feps(\ueps)\veps\vp
\eas
for all $\vp\in C_0^\infty(\Om\times [0,\infty))$, taking $\eps=\eps_j\searrow 0$ we therefore readily obtain (\ref{wu})
from (\ref{24.2}), (\ref{24.3}) and (\ref{24.6}), whereas (\ref{wv}) results from (\ref{24.4}), (\ref{24.5}) and (\ref{24.7}).
Consequently, it follows that $(u,v)$ indeed forms a global weak solution of (\ref{0}) in the sense of Definition \ref{dw}.
\end{proof}
This essentially establishes our main result on global weak solvability in (\ref{0}) already.
\begin{proof}[Proof of Theorem \ref{theo25}
The statement on existence of a global weak solution fulfilling (\ref{25.01}) directly results from Lemma \ref{lem24},
whereas the boundedness properties in (\ref{25.1}), (\ref{25.2}) and (\ref{25.3}) can readily be obtained upon combining (\ref{11.1}),
(\ref{11.01}), (\ref{22.1}), (\ref{22.2}) and (\ref{11.4}) with (\ref{24.2}), \eqref{24.5b}, \eqref{u-weak-1plus2overn}, (\ref{24.3}) and (\ref{24.5}).
\end{proof}
\section{Stationary states}
We finally consider the stationary problem associated with (\ref{0}), that is, the boundary value problem
\begin{align}\label{stationary}
0&=Δu-∇\cdot(u∇v)&&\text{ in } Ω\nn\\
0&=Δv-uv&&\text{ in } Ω\\
\frac{∂u}{∂ν}&=u\frac{∂v}{∂ν}, \quad v=\vst&&\text{ on }∂Ω,\nn
\end{align}
and our arguments in this regard will be closely
related to those in \cite{braukhoff_lankeit}, where the second equation was instead supplemented by Robin-type boundary conditions.
Dropping the requirement of radial symmetry here, we will assume that
$Ω\subset ℝ^n$ is a bounded bounded with smooth boundary, and that with some $\beta\in (0,1)$, $\vst$ belongs to
$C^{2+β}(\Ombar)$ and is nonnegative on $\pO$.\abs
A first essential observation is that we can eliminate $u$ from the stationary system and subsequently deal with a single equation only. Especially regarding the question of uniqueness, the appearance of a constant parameter $\alpha$ in said equation could turn out to be unfortunate. However, we will later show that $\alpha$ is in one-to-one correspondence with $\io u$. (An alternative would be to compute $α=\frac{m}{\io e^v}$ and work with the nonlocal equation for $v$, see \cite{lee_wang_yang}, where this approach was used for the special case of constant boundary value.)
\begin{lem}\label{lem:stationary-u}
Let $v\in C^2(\Ombar)$. If $u\in C^2(\Ombar)$ satisfies
\begin{equation}\label{u-eq}
\begin{cases}
0=Δu-∇\cdot(u∇v)&\text{in } Ω\\
\frac{∂u}{∂ν} = u \frac{∂v}{∂ν}&\text{ on }∂Ω
\end{cases}
\end{equation}
then there is $α\inℝ$ such that
\begin{equation}\label{u-from-v}
u=αe^v.
\end{equation}
If, on the other hand, \eqref{u-from-v} holds for some $v\in C^2(\Ombar)$ and $α\inℝ$, then \eqref{u-eq}. Furthermore, the signs of $\io u$ and $α$ coincide.
\end{lem}
\begin{proof}
While the second part of the statement directly follows from the chain rule and the last part is obvious after integration of \eqref{u-from-v}, the first is identical to \cite[Lemma~4.1]{braukhoff_lankeit}.
\end{proof}
We now take care of solvability and some a priori estimates for solutions of the second equation of \eqref{stationary} if we insert \eqref{u-from-v}, firstly in a related linear problem.
\begin{lem}\label{lem:v-linear-regularity}
Let $α\ge 0$, $\vst\in C^{2+β}(\Ombar)$, $\vst\ge 0$ and $v\in C^{β}(\Ombar)$. Then
\begin{equation}\label{v-linear}
\begin{cases}
Δ\tilde{v} = α\tilde{v} e^{v} &\text{in } Ω\\
\tilde{v}=\vst&\text{on } ∂Ω
\end{cases}
\end{equation}
has a unique solution $\tilde{v}\in C^{2+β}(\Ombar)$. This solution satisfies
\begin{equation}\label{v-linear-uniformbound}
0\le \tilde{v} \le γ:= \max_{∂Ω} \vst.
\end{equation}
Moreover, for every $α$ and $\vst$ as above, there is $C>0$ such that for every $v\in C^{β}(\Ombar)$ with $0\le v\le γ$ the corresponding solution $\tilde{v}$ satisfies
\begin{equation}\label{v-holderbound}
\norm[C^{β}(\Ombar)]{v} \le C.
\end{equation}
\end{lem}
\begin{proof}
Unique solvability results from \cite[Thm.~6.14]{GT}. If $α=0$, then \eqref{v-linear-uniformbound} follows from the classical maximum principle (\cite[Thm.~3.1]{GT}) for the harmonic function $\tilde{v}$ and the assumptions on $\vst$. For $α>0$, we note that if $\tilde{v}$ is minimal at some $x_0\in Ω$, then $0\le Δ\tilde{v}(x_0)=α\tilde{v}(x_0)e^{v(x_0)}$, due to the positivity of $αe^{v(x_0)}$ entails that $\tilde{v}\ge \min\tilde{v}=\tilde{v}(x_0)\ge 0$. If $\tilde{v}$, however, is minimal at some $x_0\in ∂Ω$, then, again, $\tilde{v}\ge \tilde{v}(x_0)=\vst(x_0)\ge 0$. With nonnegativity of $\tilde{v}$ thus ensured and hence $Δ\tilde{v}\ge 0$ in $Ω$, the second part of \eqref{v-linear-uniformbound} follows from the maximum principle (\cite[Cor.~3.2]{GT}). The Hölder bound in \eqref{v-holderbound} thereby can be concluded from the boundedness of the right-hand side in \eqref{v-linear} and elliptic regularity \cite[Thm.~8.29]{GT}.
\end{proof}
With this, solving the second equation of \eqref{stationary} with \eqref{u-from-v} is possible:
\begin{lem}\label{lem:stationary-v}
For every $α\ge 0$, the boundary value problem
\begin{equation}\label{stationary-v}
\begin{cases}
Δv = αve^v& \text{ in } Ω\\
v=\vst&\text{ on } ∂Ω
\end{cases}
\end{equation}
has a solution $v\in C^2(\Ombar)$, and $v$ satisfies \eqref{v-linear-uniformbound} and \eqref{v-holderbound}.
\end{lem}
\begin{proof}
With $c_1>0$ taken from \eqref{v-holderbound}, we introduce
\[
X=\{w\in C^{β}(\Ombar)\mid 0\le w\le γ, \norm[C^{β}(\Ombar)]{w}\le c_1\}\subset C^{β}(\Ombar)
\]
and for $v\in X$ let $Φ(v)=\tilde{v}$ denote the solution $\tilde{v}\in C^{2+β}(\Ombar)$ of \eqref{v-linear}. According to elliptic regularity theory (\cite[Thm.~6.6]{GT}) in conjunction with \eqref{v-holderbound}, $Φ(X)$ is bounded in $C^{2+β}(\Ombar)$, thus relatively compact in $C^{β}(\Ombar)$ and, again by Lemma~\ref{lem:v-linear-regularity}, $Φ(X)\subseteq X$. As $Φ\colon X\to X$ moreover is continuous,
Schauder's theorem asserts the existence of a fixed point $v=Φ(v)$.
\end{proof}
\begin{lem}\label{lem:v-monotone-wrt-alpha}
Let $v_1$ and $v_2$ be two solutions of \eqref{stationary-v} with $α_1\ge 0$ and $α_2\ge 0$, respectively.
If $α_1\ge α_2$, then $v_1\le v_2$.
\end{lem}
\begin{proof}
Both $v_1$ and $v_2$ -- being solutions to \eqref{v-linear} with $v=v_1$ or $v=v_2$, respectively -- are nonnegative. We let $Ω_1:=\{x\inΩ\mid v_1(x)>v_2(x)\}$ and $\bar{v}=v_1-v_2$. As $x\mapsto xe^x$ is monotone for $x\in[0,\infty)$,
\begin{align*}
Δ\bar{v} &= α_1v_1e^{v_1}-α_2v_2e^{v_2}\ge α_2(v_1e^{v_1}-v_2e^{v_2}) \ge 0 \qquad &&\text{in }Ω_1\\
\bar{v}&=0\qquad&&\text{on }∂Ω_1.
\end{align*}
By the maximum principle therefore $\max_{\Ombar}\bar{v} = \max_{∂Ω}\bar{v} = 0$ and thus $Ω_1=\emptyset$, so that $v_1\le v_2$ in $Ω\setminus Ω_1 = Ω$.
\end{proof}
As a particular consequence of Lemma~\ref{lem:v-monotone-wrt-alpha}, for every $α\ge 0$ the solution to \eqref{stationary-v} is unique. From now on, we will denote it by $v_{α}$.
That, according to Lemma~\ref{lem:v-monotone-wrt-alpha}, $v_{α}$ is decreasing with respect to $α$ is of little help with regard to the monotonicity of $αe^{v_{α}}$ (or rather $α\io e^{v_{α}}$). For further information we study the derivative of $v_{α}$ w.r.t. $α$.
\begin{lem}\label{lem:vprime}
For every $α_1> 0$, the function
\begin{equation}\label{vprime}
v'_{α_1} = \frac{d}{dα} v_{α} \Big|_{α=α_1} = \lim_{α_2\to α_1} \frac{v_{α_2}-v_{α_1}}{α_2-α_1}
\end{equation}
exists (with the limit taken in $C^2(\Ombar)$) and satisfies
\begin{equation}\label{vprime-bvp}
\begin{cases}
Δv'_{α_1} = v_{α_1} + (α_1 e^{v_{α_1}}+α_1v_{α_1}e^{v_{α_1}})v'_{α_1}&\text{ in }Ω\\
v'_{α_1}=0&\text{on } ∂Ω.
\end{cases}
\end{equation}
\end{lem}
\begin{proof}
For $α_1\ge 0$, $α_2\ge 0$, we let $w_{α_2,α_1}=\frac{v_{α_2}-v_{α_1}}{α_2-α_1}$ and note that $w=w_{α_2,α_1}$ solves
\begin{equation}\label{w}
\begin{cases}
Δw = f_{1,α_2}+f_{2,α_2,α_1} w &\text{ in } Ω\\
w=0&\text{ on } ∂Ω
\end{cases}
\end{equation}
with $f_{1,α_2}=v_{α_2}e^{v_{α_2}}$, $f_{2,α_2,α_1}=α_1e^{v_{α_1}}+α_1v_{α_1}e^{v_{α_1}}F(v_{α_2}-v_{α_1})$, $F(z)=\frac{e^z-1}z$, $z\neq 0$, $F(0)=1$.
From the Hölder bounds on $f_{1,α_2}$ and $f_{2,α_2,α_1}$ resulting from Lemma~\ref{lem:stationary-v} and \eqref{v-holderbound}, together with elliptic regularity theory (in the shape of \cite[Thm.~6.6]{GT}), we conclude that for every $A>0$ there is $C>0$ such that
\[
\norm[C^{2+β}(\Ombar)]{w_{α_2,α_1}}\le C \qquad \text{for all } α_1,α_2\in[0,A].
\]
If $(α_n)_{n\inℕ}\subset [0,∞)$ is a sequence with limit $α\in[0,∞)$, by Arzelà-Ascoli's theorem for every subsequence $(α_{n_k})_{k\inℕ}$ there are $w\in C^2(\Ombar)$ and $(α_{n_{k_l}})_{l\inℕ}$ such that $w_{α_{n_{k_l}}}\to w$ in $C^2(\Ombar)$ as $l\to\infty$ and, by \eqref{w}, $w$ satisfies
\begin{equation}\label{w-limit}
\begin{cases}
Δw= f_{1,α}+f_{2,α,α} w&\text{in } Ω\\
w=0&\text{on } ∂Ω.
\end{cases}
\end{equation}
The solution to \eqref{w-limit} is unique (since $-\io |∇(w_1-w_2)|^2 =\io f_{2,α,α} (w_1-w_2)^2 \ge 0$ whenever $w_1$ and $w_2$ solve \eqref{w-limit}), hence actually $w_{α_n,α}\to w=v'$ in $C^2(\Ombar)$ as $n\to \infty$ and both existence of the limit in \eqref{vprime} and \eqref{vprime-bvp} follow.
\end{proof}
\begin{lem}\label{lem:vprime-estimates}
For every $α>0$,
\[
0 \ge v_{α}'>-\frac1{α}\qquad \text{in }Ω.
\]
\end{lem}
\begin{proof}
We abbreviate $v'=v'_{α}$ and $v=v_{α}$.
From Lemma~\ref{lem:v-monotone-wrt-alpha}, we obtain that $0\ge v'$. We let $x_0\in \Ombar$ be such that $v'(x_0)=\min_{\Ombar}v'$. Then $v'\equiv 0$ (which would finish the proof) or $x_0\inΩ$ and due to \eqref{vprime-bvp}
\[
0\le Δv' = ve^v + (αe^v+αve^v)v' \qquad \text{at } x_0,
\]
so that by positivity of $e^{v(x_0)}$
\[
0\le v+α(1+v)v'\qquad\text{at } x_0,
\]
which yields
\[
v'(x_0)\ge -\frac1{α} \cdot \frac{v(x_0)}{1+v(x_0)}>-\frac1{α}.\qedhere
\]
\end{proof}
Consequences of Lemma~\ref{lem:vprime-estimates} on the desired relation between $α$ and $m=\io u$ are as follows:
\begin{lem}\label{lem:m}
The map
\begin{equation}\label{def:m}
m:\begin{cases}
[0,∞)\to [0,∞)\\α\mapsto \io αe^{v_α}
\end{cases}
\end{equation}
is bijective.
\end{lem}
\begin{proof}
Computing the derivative of $m$ (which uses Lemma~\ref{lem:vprime}), like in \cite[L.3.15]{braukhoff_lankeit}, we obtain
\[
m'(α)=\io e^{v_{α}}+\io αe^{v_{α}}v'_{α} = \io e^{v_{α}}(1+αv'_{α}),
\]
so that $m'(α)>0$ by Lemma~\ref{lem:vprime-estimates}, ensuring injectivity. Since $m(0)=0$ and $m(α)=\io αe^{v_{α}}\ge α|Ω|\to ∞$ as $α\to∞$, surjectivity is obvious.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{th:stationary}]
Combining Lemma~\ref{lem:stationary-u} with \eqref{def:m} shows that $(u,v)\in (C^2(\Ombar))^2$ solves \eqref{stationary} with
$\io u=m_0$ if and only if $m_0=m(α)$, $u=αe^v$ and $v=v_{α}$ solves \eqref{stationary-v}.
Bijectivity of $m$ (Lemma~\ref{lem:m}), existence and uniqueness of $v_{α}$ (Lemma~\ref{lem:stationary-v}
and Lemma~\ref{lem:v-monotone-wrt-alpha}) therefore imply the first part of Theorem~\ref{th:stationary}.\abs
In the case when $\Om$ is a ball and $\vst$ is constant, radial symmetry follows from the above uniqueness statement.
Since $u=α\exp(v)$ by Lemma~\ref{lem:stationary-u} and $\exp$ is monotone and convex, to complete the proof
it is sufficient to show convexity of $v$, that is of
the solution to \eqref{stationary-v}. But when written in radial coordinates, \eqref{stationary-v} turns into
\[
(r^{n-1}v_r)_r = αr^{n-1} ve^v,\qquad r\in(0,R),\qquad v(R)=\vst,
\]
with $v_r(0)=0$ due to radial symmetry and differentiability of $v$. Hence,
\begin{equation}\label{vr}
v_r(r)=r^{1-n} \int_0^r αs^{n-1} v(s)e^{v(s)}ds = αr \int_0^1 t^{n-1}v(rt)e^{v(rt)}dt.
\end{equation}
Nonnegativity of $v$ (cf. \eqref{v-linear-uniformbound}) shows that hence $v_r\ge 0$; thus the rightmost expression in \eqref{vr} is
clearly increasing with respect to $r$, which shows monotonicity of $v_r$ and therefore convexity of $v$.
\end{proof}
{\small
\section*{Acknowledgement}
The second author acknowledges support of the \textit{Deutsche Forschungsgemeinschaft} in the contex of the project \textit{Emergence of structures and advantages in cross-diffusion systems} (Project No. 411007140, GZ: WI 3707/5-1).
}
{
\small
\setlength{\parskip}{0pt}
\setlength{\itemsep}{0pt plus 0.3ex}
\def\cprime{$'$}
|
\section*{Introduction}
Non-ambiguous trees (NATs for short) were introduced in a previous paper \cite{AvaBouBouSil14}.
We propose in the present article a sequel to this work.
Tree-like tableaux \cite{ABN} are certain fillings of Ferrers diagram, in simple bijection
with permutations or alternative tableaux \cite{postnikov,viennot}. They are the subject of an intense
research activity in combinatorics, mainly because they appear as the key tools
in the combinatorial interpretation of the well-studied model of statistical mechanics
called PASEP: they naturally encode the states of the PASEP, together with
the transition probabilities through simple statistics \cite{CorteelWilliams}.
Among tree-like tableaux, NATs were defined as rectangular-shaped objects in \cite{AvaBouBouSil14}.
In this way, they are in bijection with permutations $\sigma=\sigma_1\,\sigma_2\,\dots\,\sigma_n$
such that the excedances ($\sigma_i>i$) are placed at the beginning of the word $\sigma$.
Such permutations were studied by Ehrenborg and Steingrimsson \cite{ES},
who obtained an explicit enumeration formula. Thanks to NATs, a bijective proof
of this formula was described in \cite{AvaBouBouSil14}.
In the present work, we define NATs as labelled binary trees (see Definition
\ref{def_nat}, which is equivalent to the original definition). This new
presentation allows us to obtain many new results about these objects. The plan
of the article is the following.
In Section \ref{nat}, we (re-)define NATs as
binary trees whose right and left children are respectively labelled with two
sets of labels. We show how the generating series for these objects satisfies
differential equations (Prop. \ref{diff_eq_nat}), whose solution is quite
simple and explicit (Prop. \ref{gen_ser_nat}). A combinatorial interpretation
of this expression involves the (new) notion of hooks in binary trees, linked
to the notion of leaves in ordered trees. Moreover this expression yields a new
formula for the number of NATs as a positive sum (see Proposition~\ref{thm_p}),
where Ehrenborg-Steingrimsson's formula is alternating. It should be noted that
Prop.~\ref{gen_ser_nat} and Proposition~\ref{thm_p} (in the case $\alpha=\beta=1)$
were already proven by Clark and Ehrenborg \cite{CE}. To conclude with Section
\ref{nat}, we obtain $q$-analogues of our formula, which are similar to those
obtained for binary trees in \cite{BjornerWachs,HivNovThib08} (see Theorem
\ref{thm-q-hook}, the relevant statistics are either the number of inversions
or the inverse major index).\\
Section \ref{gnat} presents a generalisation of
NATs in higher dimension. For any $k\le d$, we consider NATs of dimension
$(d,k)$, embedded in $\mathbb{Z}^d$, and with edges of dimension $k$
\footnote{A definition in terms of labelled trees is given in Subsection
\ref{def_natdk}.}.
The original case corresponds to dimension $(2,1)$. Our main result on this
question is a differential equation satisfied by the generating series of these
new objects.
Finally, we study the (new) notion of hooks on binary trees in
Section \ref{hook}. We prove (through the use of generating series, and
bijectively) that the number of hooks is distributed on binary trees as another
statistics: the childleaf statistic, defined as the number of vertices who have
at least one leaf as a child.
\section{Non-ambiguous trees} \label{nat}
\subsection{Definitions} \label{definitions}
We recall that a \emph{binary tree} is a rooted tree whose vertices may have
no child, one left child, one right child or both of them.
The size of a binary tree is its number of vertices.
Usually, it is considered that there is a unique binary tree with
no vertex, it is called the \emph{empty binary tree}. In this article, we
consider that there are two binary trees of size 0: the \emph{left empty binary
tree} and the \emph{right empty binary tree}, they are respectively denoted by
$\emptyset_L$ and $\emptyset_R$. Having no child in the left direction (resp.
right direction) is the same as having the left (resp. right) empty subtree in
this direction. We denote by $\BT$ the set of binary trees.
Given a binary tree $B$, we denote by $\LV(B)$ and $\RV(B)$ the set of left
children (also called left vertices) and the set of right children (also called
right vertices). By convention, $\LV(\emptyset_L)=\RV(\emptyset_R)=-1$ and
$\LV(\emptyset_R)=\RV(\emptyset_L)=0$. We shall extend this notation to NATs.
Let $U$ and $V$ be two vertices of a binary tree $B$. If $V$ is a vertex of the
subtree of $B$ whose root is $U$, then $V$ is a \emph{descendant} of $U$ and $U$
an \emph{ancestor} of $V$.
We now define the notion of non-ambiguous trees:
\begin{defi} \label{def_nat}
A \emph{non-ambiguous tree} (NAT) $T$ is a labelling of a binary tree $B$
such that:
\begin{itemize}
\item the left (resp. right) children are labelled from $1$ to
$|\LV(B)|$ (resp. $|\RV(B)|$), such that different left (resp.
right) vertices have different labels. In other words, each left
(right) label appears exactly once.
\item if $U$ and $V$ are two left (resp. right) children in the tree,
such that $U$ is an ancestor of $V$, then the label of $U$ in $T$ is
strictly greater than the label of $V$.
\end{itemize}
The underlying binary tree of a non-ambiguous tree is called its \emph{shape}.
By convention, there is a unique NAT whose shape is $\emptyset_L$ (resp.
$\emptyset_R$) which is also denoted $\emptyset_L$ (resp. $\emptyset_R$).
We denote by $\NAT(B)$ the set of NATs of shape $B$.
\end{defi}
It is sometimes useful to label the root as well. In
this case, it is considered as both a left and right child so that it carries
a pairs of labels, namely $(|\LV(T)|+1, |\RV(T)|+1)$. On pictures, to ease the
reading, we color the labels of left and right vertices in red and blue
respectively.
In \cite{AvaBouBouSil14}, NATs were defined in a slightly different, more geometrical way.
We recall it here, and show that the two definitions are equivalent.
Formally, a (geometric) NAT of size $n$ is a set $A$ of $n$ points $(x,y)\in\mathbb{N}\times\mathbb{N}$ such that:
\begin{enumerate
\item \label{condition_1_ana} $(0,0)\in A$; we call this point the \emph{root} of $A$;
\item \label{condition_2_ana} given a non-root point $p=(x,y)\in A$, there exists one point $q=(x',y')\in A$ such that $y'<y$ and $x'=x$, or one point $r=(x',y')\in A$ such that $x'<x$, $y'=y$, but not both (which means that the pattern $\begin{array}{c}\includegraphics[scale=1.0]{pattern_1}\end{array}$
is avoided);
\item \label{condition_3_ana} there is no empty line between two given points: if there exists a point $p=(x,y)\in A$, then for every $x'<x$ (resp. $y'<y$) there exists $q=(x'',y'')\in A$ such that $x''=x'$ (resp. $y''=y'$).
\end{enumerate}
\begin{lem}
This geometric definition is equivalent to Definition
\ref{def_nat}.
\end{lem}
\begin{proof}
Let us consider a NAT $T$ presented as a labelled tree
(Definition \ref{def_nat}).
Each vertex of $T$ may be given a pair of coordinates $
(x,y)$ as follows.
For a left (resp. right) child, its $x$ (resp. $y$)
coordinate is its label in $T$.
and its $y$ (resp. $x$) coordinate is the one of its
closest ancestor which is a right child, or the root.
We thus get a geometrical object, which can be drawn as
shown in Figure \ref{fig_exple_nat} (top, right).
With the minor change of coordinates $(\tilde{x},
\tilde{y})=(|\LV(T)|+1-x, |\RV(T)|+1-y)$, we get a set of
points which
satisfies the geometrical definition.
The only thing that needs to be checked is the avoidance
of the patter
.
Let us proceed by reductio ad absurdum. We suppose that
there are three vertices $p=(\tilde{x},\tilde{y})$,
$q=(\tilde{x'},\tilde{y})$ and $r=(\tilde{x},\tilde{y'})$,
with $\tilde{x'}<\tilde{x}$ and $\tilde{y'}<\tilde{y}$. We
can suppose moreover that $p$ is a right child and that
its parent is then $r$. Thus, $q$ is not an ancestor of $p
$. Let us consider the closest ancestor $c$ of $q$ which
is a right vertex. Then the label of $c$ in $T$ is the
same as the one of $p$, but it is not $p$ as $\tilde{x'}<
\tilde{x}$. It is then absurd that $p$ and $c$ have the
same label in $T$.
Conversely, to go from the geometric version of a NAT to
Definition \ref{def_nat},
we just have to forget the redundant coordinate.
\end{proof}
The top part of Figure \ref{fig_exple_nat} shows an
example of a NAT,
and illustrates the correspondence between the geometrical
presentation
of \cite{AvaBouBouSil14} and Definition \ref{def_nat}.
The dimension $\wi_L(T)\times \wi_R(T)$ of the rectangle containing the
geometrical presentation of $T$, is called the geometric size of $T$ and
satisfies
\[
(\wi_L(T),\wi_R(T)) = (|\LV(T)|+1, |\RV(T)|+1).
\]
By convention, the non-ambiguous trees $\emptyset_L$ and $\emptyset_R$ satisfy
respectively $(\wi_L(\emptyset_L),\wi_R(\emptyset_L)) = (0,1)$ and $(\wi_L(\emptyset_R),\wi_R(\emptyset_R)) = (1,0)$.
\tikzset{every node/.style={inner sep=1pt,scale=0.8}}
\begin{figure}
\begin{center}
\[T=\begin{tikzpicture}[baseline=(current bounding box.center),scale=0.5]
\node (r) at (0,8) {(\textcolor{red}{11},\textcolor{blue}{12})};
\node (11b) at (1,7) {\textcolor{blue}{11}};
\node (10b) at (-1,6) {\textcolor{blue}{10}};
\node (9b) at (0,5) {\textcolor{blue}{9}};
\node (8b) at (-1,4) {\textcolor{blue}{8}};
\node (7b) at (2,6) {\textcolor{blue}{7}};
\node (6b) at (0,3) {\textcolor{blue}{6}};
\node (5b) at (2,4){\textcolor{blue}{5}};
\node (4b) at (3,5) {\textcolor{blue}{4}};
\node (3b) at (4,4){\textcolor{blue}{3}};
\node (2b) at (3,1){\textcolor{blue}{2}};
\node (1b) at (5,3){\textcolor{blue}{1}};
\node (10r) at (-2,7){\textcolor{red}{10}};
\node (9r) at (1,5){\textcolor{red}{9}};
\node (8r) at (1,3){\textcolor{red}{8}};
\node (7r) at (3,3){\textcolor{red}{7}};
\node (6r) at (-2,5){\textcolor{red}{6}};
\node (5r) at (2,2){\textcolor{red}{5}};
\node (4r) at (-2,3){\textcolor{red}{4}};
\node (3r) at (-1,2){\textcolor{red}{3}};
\node (2r) at (-3,6){\textcolor{red}{2}};
\node (1r) at (-3,4){\textcolor{red}{1}};
\draw (2r)--(10r)--(r)--(11b)--(7b)--(4b)--(3b)--(1b);
\draw (7b)--(9r);
\draw (9r)--(5b);
\draw (5b)--(8r);
\draw (3b)--(7r);
\draw (7r)--(5r);
\draw (5r)--(2b);
\draw (10r)--(10b);
\draw (9b)--(10b)--(6r)--(1r);
\draw (6b)--(8b)--(4r);
\draw (3r)--(6b);
\draw (6r)--(8b);
\end{tikzpicture}
\hspace*{1cm}
\begin{tikzpicture}[baseline=(current bounding box.center),
scale=0.35,
every node/.style={inner sep=0pt,scale=0.7}]
\draw (0,0.75) node{\textcolor{blue}{12}};
\draw (1,0.75) node{\textcolor{blue}{11}};
\draw (2,0.75) node{\textcolor{blue}{10}};
\draw (3,0.75) node{\textcolor{blue}{9}};
\draw (4,0.75) node{\textcolor{blue}{8}};
\draw (5,0.75) node{\textcolor{blue}{7}};
\draw (6,0.75) node{\textcolor{blue}{6}};
\draw (7,0.75) node{\textcolor{blue}{5}};
\draw (8,0.75) node{\textcolor{blue}{4}};
\draw (9,0.75) node{\textcolor{blue}{3}};
\draw (10,0.75) node{\textcolor{blue}{2}};
\draw (11,0.75) node{\textcolor{blue}{1}};
\draw (-0.75,0) node{\textcolor{red}{11}};
\draw (-0.75,-1) node{\textcolor{red}{10}};
\draw (-0.75,-2) node{\textcolor{red}{9}};
\draw (-0.75,-3) node{\textcolor{red}{8}};
\draw (-0.75,-4) node{\textcolor{red}{7}};
\draw (-0.75,-5) node{\textcolor{red}{6}};
\draw (-0.75,-6) node{\textcolor{red}{5}};
\draw (-0.75,-7) node{\textcolor{red}{4}};
\draw (-0.75,-8) node{\textcolor{red}{3}};
\draw (-0.75,-9) node{\textcolor{red}{2}};
\draw (-0.75,-10) node{\textcolor{red}{1}};
\draw[gray!50] (0,0.5)--(0,-10.5);
\draw[gray!50] (1,0.5)--(1,-10.5);
\draw[gray!50] (2,0.5)--(2,-10.5);
\draw[gray!50] (3,0.5)--(3,-10.5);
\draw[gray!50] (4,0.5)--(4,-10.5);
\draw[gray!50] (5,0.5)--(5,-10.5);
\draw[gray!50] (6,0.5)--(6,-10.5);
\draw[gray!50] (7,0.5)--(7,-10.5);
\draw[gray!50] (8,0.5)--(8,-10.5);
\draw[gray!50] (9,0.5)--(9,-10.5);
\draw[gray!50] (10,0.5)--(10,-10.5);
\draw[gray!50] (11,0.5)--(11,-10.5);
\draw[gray!50] (-0.5,0)--(11.5,0);
\draw[gray!50] (-0.5,-1)--(11.5,-1);
\draw[gray!50] (-0.5,-2)--(11.5,-2);
\draw[gray!50] (-0.5,-3)--(11.5,-3);
\draw[gray!50] (-0.5,-4)--(11.5,-4);
\draw[gray!50] (-0.5,-5)--(11.5,-5);
\draw[gray!50] (-0.5,-6)--(11.5,-6);
\draw[gray!50] (-0.5,-7)--(11.5,-7);
\draw[gray!50] (-0.5,-8)--(11.5,-8);
\draw[gray!50] (-0.5,-9)--(11.5,-9);
\draw[gray!50] (-0.5,-10)--(11.5,-10);
\node (r) at (0,0) {$\bullet$};
\node (11b) at (12-11,0) {$\bullet$};
\node (10b) at (12-10,10-11) {$\bullet$};
\node (9b) at (12-9,10-11) {$\bullet$};
\node (8b) at (12-8,6-11) {$\bullet$};
\node (7b) at (12-7,0) {$\bullet$};
\node (6b) at (12-6,6-11) {$\bullet$};
\node (5b) at (12-5,9-11){$\bullet$};
\node (4b) at (12-4,0) {$\bullet$};
\node (3b) at (12-3,0){$\bullet$};
\node (2b) at (12-2,5-11){$\bullet$};
\node (1b) at (12-1,0){$\bullet$};
\node (10r) at (0,10-11){$\bullet$};
\node (9r) at (12-7,9-11){$\bullet$};
\node (8r) at (12-5,8-11){$\bullet$};
\node (7r) at (12-3,7-11){$\bullet$};
\node (6r) at (12-10,6-11){$\bullet$};
\node (5r) at (12-3,5-11){$\bullet$};
\node (4r) at (12-8,4-11){$\bullet$};
\node (3r) at (12-6,3-11){$\bullet$};
\node (2r) at (0,2-11){$\bullet$};
\node (1r) at (12-10,1-11){$\bullet$};
\draw (2r)--(10r)--(r)--(11b)--(7b)--(4b)--(3b)--(1b);
\draw (7b)--(9r);
\draw (9r)--(5b);
\draw (5b)--(8r);
\draw (3b)--(7r);
\draw (7r)--(5r);
\draw (5r)--(2b);
\draw (10r)--(10b);
\draw (9b)--(10b)--(6r)--(1r);
\draw (6b)--(8b)--(4r);
\draw (3r)--(6b);
\draw (6r)--(8b);
\end{tikzpicture}\]
\[T_L=
\begin{tikzpicture}[baseline=(current bounding box.center),scale=0.5]
\node (10b) at (-1,6) {\textcolor{blue}{4}};
\node (9b) at (0,5) {\textcolor{blue}{3}};
\node (8b) at (-1,4) {\textcolor{blue}{2}};
\node (6b) at (0,3) {\textcolor{blue}{1}};
\node (10r) at (-2,7){(\textcolor{red}{6},\textcolor{blue}{5})};
\node (6r) at (-2,5){\textcolor{red}{5}};
\node (4r) at (-2,3){\textcolor{red}{4}};
\node (3r) at (-1,2){\textcolor{red}{3}};
\node (2r) at (-3,6){\textcolor{red}{2}};
\node (1r) at (-3,4){\textcolor{red}{1}};
\draw (2r)--(10r);
\draw (10r)--(10b);
\draw (9b)--(10b)--(6r)--(1r);
\draw (6b)--(8b)--(4r);
\draw (3r)--(6b);
\draw (6r)--(8b);
\end{tikzpicture}
\hspace{0.5cm}
T_R=
\begin{tikzpicture}[baseline=(current bounding box.center),scale=0.5]
\node (11b) at (1,7) {(\textcolor{red}{5},\textcolor{blue}{7})};
\node (7b) at (2,6) {\textcolor{blue}{6}};
\node (5b) at (2,4){\textcolor{blue}{5}};
\node (4b) at (3,5) {\textcolor{blue}{4}};
\node (3b) at (4,4){\textcolor{blue}{3}};
\node (2b) at (3,1){\textcolor{blue}{2}};
\node (1b) at (5,3){\textcolor{blue}{1}};
\node (9r) at (1,5){\textcolor{red}{4}};
\node (8r) at (1,3){\textcolor{red}{3}};
\node (7r) at (3,3){\textcolor{red}{2}};
\node (5r) at (2,2){\textcolor{red}{1}};
\draw (11b)--(7b)--(4b)--(3b)--(1b);
\draw (7b)--(9r);
\draw (9r)--(5b);
\draw (5b)--(8r);
\draw (3b)--(7r);
\draw (7r)--(5r);
\draw (5r)--(2b);
\end{tikzpicture}\]
\caption{A non-ambiguous tree, its geometrical presentation,
and its left and right subtrees}\label{fig_exple_nat}
\end{center}
\end{figure}
Figure \ref{fig_exple_nat_B} gives an example of a class $NAT(B)$,
in the case where the binary tree is $B=\begin{tikzpicture}[baseline=(current bounding box.center),scale=0.2]
\node (r) at (0,2) {$\sbullet$};
\node (1b) at (2,1) {$\sbullet$};
\node (2b) at (-0.7,0) {$\sbullet$};
\node (1r) at (-2,1){$\sbullet$};
\node (2r) at (0.7,0){$\sbullet$};
\draw (2r)--(1b)--(r)--(1r)--(2b);
\end{tikzpicture}
$.
\tikzset{every node/.style={inner sep=1pt,scale=0.8}}
\begin{figure}
\begin{center}
\[
\begin{tikzpicture}[baseline=(current bounding box.center),scale=0.5]
\node (r) at (0,2) {(\textcolor{red}{2},\textcolor{blue}{2})};
\node (1b) at (2,1) {\textcolor{blue}{1}};
\node (2b) at (-1,0) {\textcolor{blue}{2}};
\node (1r) at (-2,1){\textcolor{red}{1}};
\node (2r) at (1,0){\textcolor{red}{2}};
\draw (2r)--(1b)--(r)--(1r)--(2b);
\end{tikzpicture}
\hspace*{1cm}
\begin{tikzpicture}[baseline=(current bounding box.center),scale=0.5]
\node (r) at (0,2) {(\textcolor{red}{2},\textcolor{blue}{2})};
\node (2b) at (2,1) {\textcolor{blue}{2}};
\node (1b) at (-1,0) {\textcolor{blue}{1}};
\node (1r) at (-2,1){\textcolor{red}{1}};
\node (2r) at (1,0){\textcolor{red}{2}};
\draw (2r)--(2b)--(r)--(1r)--(1b);
\end{tikzpicture}
\hspace*{1cm}
\begin{tikzpicture}[baseline=(current bounding box.center),scale=0.5]
\node (r) at (0,2) {(\textcolor{red}{2},\textcolor{blue}{2})};
\node (1b) at (2,1) {\textcolor{blue}{1}};
\node (2b) at (-1,0) {\textcolor{blue}{2}};
\node (2r) at (-2,1){\textcolor{red}{2}};
\node (1r) at (1,0){\textcolor{red}{1}};
\draw (1r)--(1b)--(r)--(2r)--(2b);
\end{tikzpicture}
\hspace*{1cm}
\begin{tikzpicture}[baseline=(current bounding box.center),scale=0.5]
\node (r) at (0,2) {(\textcolor{red}{2},\textcolor{blue}{2})};
\node (2b) at (2,1) {\textcolor{blue}{2}};
\node (1b) at (-1,0) {\textcolor{blue}{1}};
\node (2r) at (-2,1){\textcolor{red}{2}};
\node (1r) at (1,0){\textcolor{red}{1}};
\draw (1r)--(2b)--(r)--(2r)--(1b);
\end{tikzpicture}
\]
\caption{An example of $NAT(B)$}\label{fig_exple_nat_B}
\end{center}
\end{figure}
\subsection{Differential equations on non-ambiguous trees}
\label{differential}
The goal of this section is to get (new) formulas for the number of NATs
with prescribed shape.
The crucial argument is the following remark: let $T$ be a NAT
whose shape is a non-empty binary tree $B = \NodeBT{L}{R}$.
Restricting the labellings of the left and right children of $T$
to $L$ and $R$ gives non-decreasing labelling of their respective left and
right children. Note that the root of $L$ (resp. $R$) is a left
(resp. right) child in $T$. By renumbering the labels so that they are
consecutive numbers starting from $1$, we get two non-ambiguous labellings for
$L$ and $R$, that is two non-ambiguous trees $T_L$ and
$T_R$. See Figure~\ref{fig_exple_nat} for an example.
\begin{rque}
The geometric size of $T$ satisfies:
\begin{align}
\wi_L(T) &= \wi_L(T_L)+\wi_L(T_R) \\
\wi_R(T) &= \wi_R(T_L)+\wi_R(T_R)
\end{align}
\end{rque}
Conversely, knowing the labelling of $L$ and $R$, to recover the labelling
of $T$, one has to choose which labels among $\{ 1, \ldots, \LV(T) \}$ will be used for
$L$ (including its root) and the same for right labels.
As a consequence, the number of NAT with shape $B$ is given by:
\begin{equation}\label{Equation:BNat-binom}
\left|\NATB[\NodeBT{L}{R}]\right| =
\binom{|\LV(T)|}{|\LV(R)|}
\binom{|\RV(T)|}{|\RV(L)|}\,
|\NATB[L]|\,
|\NATB[R]|.
\end{equation}
Our first step is to recover the hook-length formula for the number of
NATs of fixed shape (\cite{AvaBouBouSil14}). We
use the method from~\cite{HivNovThib08}, namely, applying recursively a
bilinear integro-differential operator called here a \emph{pumping function}
along a binary tree.
First of all, we consider the $\operatorname{\mathbb{Q}}$-vector space $\operatorname{\mathbb{Q}}\NAT$ of formal sums of
non-ambiguous trees and identifies $\NATB$ with the formal sum of its elements.
We consider also the $\operatorname{\mathbb{Q}}$-vector spaces $\operatorname{\mathbb{Q}}\NAT_L$ and $\operatorname{\mathbb{Q}}\NAT_R$ generated
respectively by $\NAT\setminus\{\emptyset_R\}$ and
$\NAT\setminus\{\emptyset_L\}$. Let $\BNAT$ be the linear map
\[
\BNAT:\operatorname{\mathbb{Q}}\NAT_L\times\operatorname{\mathbb{Q}}\NAT_R\mapsto\operatorname{\mathbb{Q}}\NAT
\]
sending a pair of non-ambiguous trees $(T_1,T_2)$ to the formal sum of NATs $T$
such that $T_L=T_1$ and $T_R=T_2$. The main remark is that $\NATB$ can be
computed by a simple recursion using $\BNAT$.
\begin{lem}\label{lem:recur_B}
The formal sum $\NATB[B]$ of non-ambiguous trees of shape $B$ satisfies
the following recursion: if $B=\emptyset_L$ or $B=\emptyset_R$ then
$\NATB[B] = B$, else
\[
\NATB[\NodeBT{L}{R}] = \BNAT\left(\NATB[L], \NATB[R]\right).
\]
\end{lem}
To count non-ambiguous trees, and as suggested by the binomial coefficients
in~\eqref{Equation:BNat-binom}, we shall use \emph{doubly exponential
generating functions} in two variables $x$ and $y$, where $x$ and $y$ count the
geometric size: the weight of a NAT $T$ is $\Phi(T):=
\frac{x^{\wi_L(T)}}{\wi_L(T)!} \frac{y^{\wi_R(T)}}{\wi_R(T)!}\,.$ We extend
$\Phi(T)$ by linearity to a map $\operatorname{\mathbb{Q}}\NAT\mapsto\operatorname{\mathbb{Q}}[[x, y]]$. Consequently,
$\Phi(\NATB[B])$ is the generating series of the non-ambiguous trees of shape
$B$. Thanks to~\eqref{Equation:BNat-binom} the image in $\operatorname{\mathbb{Q}}[[x, y]]$ of the
bilinear map $\BNAT$ under the map $\Phi$ is a simple differential operator:
\begin{defi}\label{def:pumping}
The \emph{pumping function} $\Bxy$ is the bilinear map
$\operatorname{\mathbb{Q}}[[x,y]]\times\operatorname{\mathbb{Q}}[[x,y]]\mapsto\operatorname{\mathbb{Q}}[[x,y]]$ defined by
\begin{equation}
\Bxy(f, g) = \int_0^x\int_0^y
\partial_yf(u,v)\cdot
\partial_xg(u,v)
\diff u\diff v.
\end{equation}
We further define recursively, for any binary tree $B$
an element $\Bx(B)\in\operatorname{\mathbb{Q}}[[x,y]]$ by
\begin{equation}
\Bx(\emptyset_L) = y,
\quad
\Bx(\emptyset_R) = x
\quad\text{and}\qquad
\Bx\left({\NodeBT{L}{R}}\right) = \Bxy\left(\Bx(L), \Bx(R)\right)\,.
\end{equation}
\end{defi}
A simple computation gives us that
for $f=\frac{x^{a_1}}{a_1!}\frac{y^{b_1}}{b_1!}$ and $g=\frac{x^{a_2}}{a_2!} \frac{y^{b_2}}{b_2!}$
one has
$$
\Bxy(f,g)=\binom{a_1+a_2-1}{a_2-1} \binom{b_1+b_2-1}{b_1-1} \frac{x^{a_1+a_2}}{(a_1+a_2)!} \frac{y^{b_1+b_2}}{(b_1+b_2)!}.
$$
Whence for $(T_1, T_2)$ a pair of NATs in $\operatorname{\mathbb{Q}}\NAT_L\times\operatorname{\mathbb{Q}}\NAT_R$, we get:
\begin{align*}
\Bxy(\Phi(T_1),\Phi(T_2)) =& \binom{w_L(T_1)+w_L(T_2)-1}{w_L(T_2)-1} \binom{w_R(T_1)+w_R(T_2)-1}{w_R(T_1)-1}\\
&\times \frac{x^{w_L(T_1)+w_L(T_2)}}{(w_L(T_1)+w_L(T_2))!} \frac{y^{w_R(T_1)+w_R(T_2)}}{(w_R(T_1)+w_R(T_2))!}
\end{align*}
Thus~\eqref{Equation:BNat-binom} implies the following lemma.
\begin{lem}\label{lem:com_phi}
For $(T_1, T_2)$ a pair of NATs in $\operatorname{\mathbb{Q}}\NAT_L\times\operatorname{\mathbb{Q}}\NAT_R$, one has
$$\Phi(\BNAT(T_1, T_2)) = \Bxy(\Phi(T_1),\Phi(T_2)).$$
\end{lem}
We derive from these two lemmas the following proposition.
\begin{prop}
\label{prop:commutation_phi}
For any non-empty binary tree $B$,
$$\Phi(\NATB[B])=\Bx(B).$$
\end{prop}
\begin{proof}
We apply Lemmas \ref{lem:recur_B} and \ref{lem:com_phi} to write (with $L$ and $R$ the left and right sub-tree of $B$):
\begin{align*}
\Phi(\NATB[B])
&=
\Phi(\BNAT(\NAT(L),\NAT(R)))
\\
& =
\Phi(\BNAT(
\sum_{T_1 \in \NAT(L)} T_1,
\sum_{T_2 \in \NAT(R)} T_2
))
\\
&=
\sum_{T_1 \in \NAT(L)}
\
\sum_{T_2 \in \NAT(R)}
\Phi(\BNAT(
T_1, T_2
))
\\
&=
\sum_{T_1 \in \NAT(L)}
\
\sum_{T_2 \in \NAT(R)}
\Bxy(
\Phi(T_1), \Phi(T_2)
)
\\
&=
\Bxy\left(
\sum_{T_1 \in \NAT(L)} \Phi(T_1),
\sum_{T_2 \in \NAT(R)} \Phi(T_2)
\right)
=
\Bxy( \Bx(L), \Bx(R) ).
\end{align*}
\end{proof}
We are now able to recover the hook-length formula
of~\cite{AvaBouBouSil14} for non-ambiguous trees of a given shape.
\begin{prop} \label{prop_hook_nat}
Let $\B$ be a binary tree. For each non-root left (resp. right) vertex $U$, we
denote by $\EL(U)$ (resp. $\ER(U)$) the number of left (resp. right) vertices of
the subtree with root $U$ (itself included in the count). Then
\begin{equation}\label{eq:HLT}
|\NAT(\B)|
=
\frac{ |\LV(\B)|! \cdot |\RV(\B)|!}{
\displaystyle
\prod_{U: \text{left child}}\EL(U) \cdot
\prod_{U: \text{right child}}\ER(U)
}\,.
\end{equation}
\end{prop}
Before proving this Proposition, let us illustrate it on an example.
\begin{exple} \label{exple_hook}
Let $B = \scalebox{0.5}
{ \newcommand{\node (a) {(\color{red}$4$,\color{blue}$5$)}{\node[draw,fill,circle] (a) {$$};}
\newcommand{\node (b) {\color{red}$3$}{\node[draw,fill,circle] (b) {$$};}
\newcommand{\node (c) {\color{red}$1$}{\node[draw,fill,circle] (c) {$$};}
\newcommand{\node (d) {\color{blue}$2$}{\node[draw,fill,circle] (d) {$$};}
\newcommand{\node (e) {\color{blue}$4$}{\node[draw,fill,circle] (e) {$$};}
\newcommand{\node (f) {\color{red}$2$}{\node[draw,fill,circle] (f) {$$};}
\newcommand{\node (g) {\color{blue}$3$}{\node[draw,fill,circle] (g) {$$};}
\newcommand{\node (h) {\color{blue}$1$}{\node[draw,fill,circle] (h) {$$};}
\begin{tikzpicture}[scale=0.5,baseline=(current bounding box.center),
every node/.style={inner sep=2pt}]
\matrix[column sep=.15cm, row sep=.15cm,ampersand replacement=\&]{
\& \& \& \node (a) {(\color{red}$4$,\color{blue}$5$) \& \& \& \& \& \\
\& \node (b) {\color{red}$3$ \& \& \& \& \& \& \node (e) {\color{blue}$4$ \& \\
\node (c) {\color{red}$1$ \& \& \node (d) {\color{blue}$2$ \& \& \& \node (f) {\color{red}$2$ \& \& \& \node (h) {\color{blue}$1$ \\
\& \& \& \& \& \& \node (g) {\color{blue}$3$ \& \& \\
};
\path[thick] (b) edge (c) edge (d)
(f) edge (g)
(e) edge (f)
(a) edge (b) edge (e)
(e) edge (h);
\end{tikzpicture}}$. The hook formula is given by:
\begin{equation}
|\NAT(\B)|
=
\frac{ 3! 4!}{
\left( 1 \cdot 2 \cdot 1\right) \cdot
\left(1 \cdot 1 \cdot 3 \cdot 1 \right)
}=24\,.
\end{equation}
\end{exple}
\begin{proof}
Proposition \ref{prop:commutation_phi} may be rewritten as:
\begin{equation}\label{eq:com_phi}
\Bx(B) = |\NAT(\B)| \frac{x^{\wi_L(B)}y^{\wi_R(B)}}{\wi_L(B)! \wi_R(B)!}.
\end{equation}
So, we get by a simple computation :
\begin{align*}
\Bx(B)
&=
\Bx\left({\NodeBT{L}{R}}\right)
=
\Bxy\left(\Bx(L), \Bx(R)\right)
\\
&=
\Bxy\left(
|\NAT(L)|
\frac{
x^{\wi_L(L)}y^{\wi_R(L)}
}{
\wi_L(L)! \wi_R(L)!
}
,
|\NAT(R)|
\frac{
x^{\wi_L(R)}y^{\wi_R(R)}
}{
\wi_L(R)! \wi_R(R)!
}
\right)
\\
&=
\frac{
|\NAT(L)| \cdot |\NAT(R)|
}{
\wi_L(L)! \wi_R(L)! \wi_L(R)! \wi_R(R)!
}
\Bxy\left(
x^{\wi_L(L)}y^{\wi_R(L)},
x^{\wi_L(R)}y^{\wi_R(R)}
\right)
\\
&=
\frac{
|\NAT(L)| \cdot |\NAT(R)| \wi_R(L) \wi_L(R)
}{
\wi_L(L)! \wi_R(L)! \wi_L(R)! \wi_R(R)!
}
\frac{
x^{\wi_L(L)+\wi_L(R)}y^{\wi_R(L)+\wi_R(R)}
}{
(\wi_L(L)+\wi_L(R)) \cdot (\wi_R(L)+\wi_R(R))
}
\\
&=
\frac{
\wi_R(L) \wi_L(R)
}{
\wi_L(B) \wi_R(B)
}
|\NAT(L)| \frac{x^{\wi_L(L)}y^{\wi_R(L)}}{\wi_L(L)! \wi_R(L)!}
\cdot
|\NAT(R)| \frac{x^{\wi_L(R)}y^{\wi_R(R)}}{\wi_L(R)! \wi_R(R)!}
\end{align*}
which, together with \eqref{eq:com_phi} gives:
\begin{align*}
|\NAT(\B)| \frac{x^{\wi_L(B)}y^{\wi_R(B)}}{\wi_L(B)! \wi_R(B)!}
&=
\frac{
\wi_R(L) \wi_L(R)
}{
\wi_L(B) \wi_R(B)
}
\cdot
|\NAT(L)| \frac{x^{\wi_L(L)}y^{\wi_R(L)}}{\wi_L(L)! \wi_R(L)!}
\\
& \cdot
|\NAT(R)| \frac{x^{\wi_L(R)}y^{\wi_R(R)}}{\wi_L(R)! \wi_R(R)!}\,.
\end{align*}
We obtain
\begin{align*}
\frac{
|\NAT(\B)|
}{
(\wi_L(B)-1)! (\wi_R(B)-1)!
}
&=
\wi_R(L) \wi_L(R)
\frac{
|\NAT(L)| \cdot |\NAT(R)|
}{
\wi_L(L)! \wi_R(L)! \wi_L(R)! \wi_R(R)!
}
\\
&=
\frac{
1
}{
\wi_L(L) \wi_R(R)
}
\frac{
|\NAT(L)|
}{
(\wi_L(L)-1)! (\wi_R(L)-1)!
}
\\
&\cdot
\frac{
|\NAT(R)|
}{
(\wi_L(R)-1)! (\wi_R(R)-1)!
}\,.
\end{align*}
We deduce that
\begin{align*}
\frac{
|\NAT(\B)|
}{
\LV(B)! \RV(B)!
}
&=
\frac{
1
}{
\wi_L(L) \wi_R(R)
}
\cdot
\frac{
|\NAT(L)|
}{
\LV(L)! \RV(L)!
}
\cdot
\frac{
|\NAT(R)|
}{
\LV(R)! \RV(R)!
}\,.
\end{align*}
The coefficient $\frac{1}{\wi_L(L) \wi_R(R)}$ recursively gives the denominator of the hook-legth formula, and we get \eqref{eq:HLT}.
\end{proof}
\bigskip
Let $\LGFN$ be the \emph{exponential generating function of non-ambiguous trees}
with weight $\Phi$:
\begin{equation}
\LGFN(x,y):= \sum_{\N \in \NAT} \Phi(\N) =
\sum_{\N \in \NAT} \frac{x^{\wi_L(T)}}{\wi_L(T)!} \frac{y^{\wi_R(T)}}{\wi_R(T)!}\,.
\end{equation}
and $\GFN$ its derivative $\GFN=\partial_x\partial_y\LGFN$. Naturally,
they are linked by the relation
\[
\LGFN(x,y) = y + x + \int_0^x \int_0^y \GFN(u,v)\diff u\diff v.
\]
They are both solutions of a fixed point differential equation.
\begin{prop}
\label{prop_equ_diff_nat}
The generating function $\GFN$ and $\LGFN$ can be computed by the following
fixed point differential equations:
\begin{equation}
\label{equ_gfn}
\LGFN = y + x + \int_x \int_y \partial_x \LGFN \cdot \partial_y \LGFN
\hspace{.5cm}
\text{and}
\hspace{.5cm}
\GFN =
\left( 1 + \int_{x} \GFN \right)
\cdot
\left( 1 + \int_{y} \GFN \right)
\end{equation}
\end{prop}
\begin{proof}
The first equation is a consequence of the definition of the bilinear
map $\Bxy$:
\begin{align*}
\LGFN &= \sum_{B \in \BT} \Bx(B)\\
&= y + x + \sum_{(L, R) \in \BT_L\times\BT_R} \Bx\left({\NodeBT{L}{R}}\right)\\
&= y + x + \sum_{(L, R) \in \BT_L\times\BT_R} \Bxy(\Bx(L), \Bx(R))\\
&= y + x + \Bxy(\LGFN-x, \LGFN-y)\\
&= y + x + \Bxy(\LGFN, \LGFN),
\end{align*}
with $\BT_L=\BT\setminus\{\emptyset_R\}$ and
$\BT_R=\BT\setminus\{\emptyset_L\}$. To prove the second equation, remark that
the first equation implies the identity
\[
\partial_x \partial_y \LGFN = \partial_x \LGFN. \partial_y \LGFN
\]
and moreover that we have
\[
\partial_x \LGFN = 1 + \int_y \GFN
\hspace{.5cm}
\text{and}
\hspace{.5cm}
\partial_y \LGFN = 1 + \int_x \GFN.\
\]
\end{proof}
From these identities, a closed formula can be computed for $\GFN$ and $\LGFN$.
The expression of $\GFN$ was already proven in \cite{CE} using permutations.
\begin{prop}\label{prop:gf_expressions}
The doubly exponential generating functions for non-ambiguous trees are given by
$$
\LGFN = x+y-\log( 1 - (e^x-1)(e^y-1) )\,,
\quad\text{and}\quad
\GFN
=
\frac{
e^{x+y}
}{
\left(
1 - (e^x - 1)(e^y - 1)
\right)^2
}.
$$
\end{prop}
\begin{proof}
We know that $\LGFN$ is a solution of
\begin{equation}
\label{equ_lgfn}
\begin{cases}
\partial_x \partial_y f = \partial_x f \times \partial_y f,\\
f(x,y)=f(y,x)
\end{cases}
\end{equation}
This system of equations satisfies the two following properties:
\begin{itemize}
\item if $s$ is a solution of Equation~\ref{equ_lgfn} then for each power
series $\varphi$ with constant term equal to zero,
$f(\varphi(x),\varphi(y))$ is also a solution;
\item if we fix the initial condition $f(x,0)$, there exists a unique
formal power series solution to Equation~\ref{equ_lgfn}.
\end{itemize}
Let $f$ be a particular solution. Let us consider the notation
$f_x:=\partial_xf$ and $f_y:=\partial_yf$, then
\[
\partial_xf_y\cdot\partial_yf_x=f_x^2f_y^2.
\]
We suppose that $\partial_xf_y=f_y^2$ and $\partial_yf_x=f_x^2$, hence
\[
f_y=\frac{-1}{x+c_1(y)} \text{ and } f_x=\frac{-1}{y+c_2(x)}.
\]
Since $f_x^2=\partial_y f_x=f_xf_y$, we get $f_y=f_x$, which implies
\[
x+c_1(y)=y+c_2(x).
\]
As a consequence, $c_1(z)=c_2(z)=z+c$ with $c$ a real number.
Finally
\[
f(x,y)=-\ln(x+y+c).
\]
Conversely, $-\ln(x+y+c)$ satisfies Equation~\eqref{equ_lgfn}. It remains to
find a real number $c$ and formal power series $\varphi$ such that
$\LGFN(x,0)=-\ln(\varphi(x)+c)$. Since $\LGFN(x,0)=x$, we get
$\varphi(x)=e^{-x}-c$. Moreover, the condition $\varphi(0)=0$ implies $c=1$.
As a consequence
\[
\LGFN(x,y)=-\ln(e^{-x}+e^{-y}-1),
\]
which can be rewritten as
\[
\LGFN(x,y)=y+x-\ln(1-(e^x-1)(e^y-1)).
\]
Differentiating with respect to $x$ and $y$, we find the expression of $\GFN$.
\end{proof}
In the context of the PASEP \cite{CorteelWilliams}, it is natural to consider statistics $\LO(\B)$ and $\RO(\B)$.
Let us briefly recall that the partially asymmetric exclusion process (PASEP) is a model for a system
of interacting particles hopping left and right on a one-dimensional lattice of $n$ sites.
In the general case, the probability of a given state of the model depends on five parameters $(q,\alpha,\beta,\gamma,\delta)$
which give the probability of transitions (a particle moving to the right or the left, or going in or out the model, when possible).
Tree-like tableaux \cite{ABN} have been proven to give a combinatorial interpretation for the steady state of the PASEP when $\gamma=\delta=0$,
the weight in $\alpha$ (resp. $\beta$) corresponding to the statistic $\LO(\B)$ (resp. $\RO(\B)$) defined as follows.
\begin{defi}
The \emph{leftmost branch} of a binary tree $\B$ is the set of vertices
$\{s_0,\ldots,s_k\}$ such that $s_0$ is the root of $\B$, $s_k$ is a leaf and
$s_{i+1}$ is the left child of $s_i$, for each $i<k$. Similarly, we define the \emph{rightmost
branch} of a binary tree.
We denote by $\LO(\B)$ and $\RO(\B)$ the number of non-root
vertices respectively in the leftmost and rightmost branches.
\end{defi}
We extend these
definitions to non-ambiguous trees. For example, in
Figure~\ref{fig_exple_nat}, we have $\LO(T)=2$ and $\RO(T)=5$.
Let us define the following $(\alpha, \beta)$-generating function for
non-ambiguous trees:
\[
\GFN(x,y;\alpha, \beta)
=
\sum_{\N \in \NAT}
\frac{
x^{|\LV(\N)|}
\cdot
y^{|\RV(\N)|}
\cdot
\alpha^{\LO(\N)}
\cdot
\beta^{\RO(\N)}
}{
|\LV(\N)|! \cdot |\RV(\N)|!
}.
\]
It satisfies an $(\alpha,\beta)$-analogue for the identity of
Proposition~\ref{prop_equ_diff_nat}.
\begin{prop}\label{diff_eq_nat}
A differential equation for $\GFN(x,y;\alpha, \beta)$ is
$$
\GFN(x,y;\alpha, \beta)=
\left(
1 + \alpha \int_x \GFN(u,y;\alpha, 1)\diff u
\right)
\cdot
\left(
1 + \beta \int_y \GFN(x,v;1, \beta)\diff v
\right).
$$
\end{prop}
\begin{proof}
We just need to define a new pumping function:
\[
\Bxy^{(\alpha, \beta)}(f,g)
=
\alpha\beta \Bxy(\left.f\right|_{\beta=1},\left.g\right|_{\alpha=1})
\]
and deduce the expected differential equation.
\end{proof}
The solution of the new differential equation is given by
Proposition~\ref{pro_equ_diff_alpha_beta}, a bijective proof is given in
Section~\ref{bijectionslabelledtree}.
\begin{prop}\label{gen_ser_nat}
The $(\alpha,\beta)$-exponential generating function for non-ambiguous trees is equal to
\label{pro_equ_diff_alpha_beta}
\[
\GFN(x,y;\alpha, \beta)
=
\frac{
e^{\alpha x + \beta y}
}{
\left(
1 - (e^x -1)(e^y - 1)
\right)^{\alpha+\beta}
}.
\]
\end{prop}
If we develop this expression we obtain an $(\alpha,\beta)$-analogue of the
enumeration of non-ambiguous trees of fixed geometric size.
Let us recall that the $q$-analogue of the Stirling
numbers of the second kind $\operatorname S_{2,q}(n,k)$
is the number of ways to partition a set of size $n$ in $k$ parts, where the power of $q$ counts the number of elements different from $n$ in the subset containing $n$. For instance, $S_{2,q}(3,2) = 1+2q$, $S_{2,q}(4,2) = 1+3q+3q^2$ and $S_{2,q}(4,3) =3+3q$.
\begin{prop}\label{pro_ana_enum_alpha_beta} \label{thm_p}
Let $i$ and $j$ be two positive integers. The $(\alpha,\beta)$-analogue of
the number of $NATs$ of geometric size $i\times j$ is
\[
\left[\frac{x^{i-1}\,y^{j-1}}{(i-1)!\,(j-1)!}\right]
\GFN(x,y;\alpha,\beta)
=
\sum_{p\geqslant 1}(p-1)!\,(\alpha+\beta)_{p-1}
\operatorname S_{2,\alpha}(i,p)\operatorname S_{2,\beta}(j,p).
\]
with $q^{(n)}:=q(q+1)\cdots(q+n-1)$
the rising factorial and $\operatorname S_{2,q}(n,k)$ the
$q$-analogue of the Stirling
numbers of the second kind.
\end{prop}
\subsection{Combinatorial interpretation with the zigzag bijection of Burstein}
\label{bijectionslabelledtree}
The purpose of this subsection is to explain combinatorially Propositions~\ref{gen_ser_nat} and \ref{pro_ana_enum_alpha_beta}.
To do so, we use the ``zigzag'' bijection introduced and studied in \cite{SteiWill}. This bijection, that we will denote by $\varphi$, was further studied by Burstein \cite{Bur07}.
First, let us introduce the statistic that corresponds to the integer $p$ in the
enumeration formula of Proposition~\ref{pro_ana_enum_alpha_beta}.
\begin{defi}\label{def_hook}
Let $B$ be a binary tree and $v$ one of its node. The \emph{hook} of a
vertex $v$ is the union of $\{v\}$, its leftmost branch and its rightmost
branch. We say that $v$ is the root of its hook. There is a unique way to
partition the vertices into hooks. The number of hooks in such a partition is
the \emph{hook number of the tree} and it is denoted by $hook(T)$. We extend this
definition to non-ambiguous trees.
\end{defi}
\begin{rque}
We can obtain recursively the unique partition of a binary tree into hooks by
extracting the root's hook and iterating the process on each tree of the
remaining forest.
\end{rque}
\begin{exple} In Figure \ref{fig:fig_nat}, the hook whose root is \textcolor{blue}{10} is highlighted in purple. The partition of all the vertices into hooks can be seen by
keeping only bold edges. The roots of the hooks are $\{(\textcolor{red}{11},\textcolor{blue}{12}), \textcolor{blue}{10}, \textcolor{red}{9}, \textcolor{blue}{8}, \textcolor{red}{8}, \textcolor{red}{7}, \textcolor{red}{3}, \textcolor{blue}{2}\}$, and so the hook number of the tree is $8$.
\begin{figure}
\begin{center}
\begin{tikzpicture}[scale=0.45]
\node (r) at (0,8) {(\textcolor{red}{11},\textcolor{blue}{12})};
\node (11b) at (1,7) {\textcolor{blue}{11}};
\node[fill=purple!50] (10b) at (-1,6) {\textcolor{blue}{10}};
\node[fill=purple!50] (9b) at (0,5) {\textcolor{blue}{9}};
\node (8b) at (-1,4) {\textcolor{blue}{8}};
\node (7b) at (2,6) {\textcolor{blue}{7}};
\node (6b) at (0,3) {\textcolor{blue}{6}};
\node (5b) at (2,4){\textcolor{blue}{5}};
\node (4b) at (3,5) {\textcolor{blue}{4}};
\node (3b) at (4,4){\textcolor{blue}{3}};
\node (2b) at (3,1){\textcolor{blue}{2}};
\node (1b) at (5,3){\textcolor{blue}{1}};
\node (10r) at (-2,7){\textcolor{red}{10}};
\node (9r) at (1,5){\textcolor{red}{9}};
\node (8r) at (1,3){\textcolor{red}{8}};
\node (7r) at (3,3){\textcolor{red}{7}};
\node[fill=purple!50] (6r) at (-2,5){\textcolor{red}{6}};
\node (5r) at (2,2){\textcolor{red}{5}};
\node (4r) at (-2,3){\textcolor{red}{4}};
\node (3r) at (-1,2){\textcolor{red}{3}};
\node (2r) at (-3,6){\textcolor{red}{2}};
\node[fill=purple!50] (1r) at (-3,4){\textcolor{red}{1}};
\draw[line width=3pt] (2r)--(10r)--(r)--(11b)--(7b)--(4b)--(3b)--(1b);
\draw (7b)--(9r);
\draw[line width=3pt] (9r)--(5b);
\draw (5b)--(8r);
\draw (3b)--(7r);
\draw[line width=3pt] (7r)--(5r);
\draw (5r)--(2b);
\draw (10r)--(10b);
\draw[line width=3pt, purple!50] (9b)--(10b)--(6r)--(1r);
\draw[line width=3pt] (6b)--(8b)--(4r);
\draw (3r)--(6b);
\draw (6r)--(8b);
\end{tikzpicture}
\caption{Hooks (bold edges) on a non-ambiguous tree}\label{fig:fig_nat}
\end{center}
\end{figure}
\end{exple}
The correspondence between $p$ and this new statistic is proven hereafter.
Let us now define the bijection $\varphi$ between non-ambiguous trees and permutations. From now on, we juggle between the geometric representation and the labelled binary
tree representation of non-ambiguous trees. Let $T$
be a non-ambiguous tree. We remove the first column of $T$, and denote by $T'$ the result of this deletion. We number, starting with
1, the south-east border, starting from the westmost edge
(Figure~\ref{fig:ana_zigzag}). Let $\sigma$ be the permutation $\varphi(T)$ and
$i$ the positive integer corresponding to a border edge. The image $\sigma(i)$
is defined as follows. Let $e$ be the border edge numbered by $i$ and suppose
that $e$ is vertical. If $e$ has no point to its left in the same row in $T'$, then
$\sigma(i)=i$. Else, starting from the leftmost point of the row in $T'$, we go down to
the closest point in the same column, then right to the closest point in the
same row and so on by doing a "zigzag" which alternates going down and right, until we reach a border edge $e'$. The image $\sigma(i)$
corresponds to the integer associated to $e'$. If $e$ is horizontal, we start
with the topmost point of the same column and then we "zigzag", starting from
with right direction and going alternatively down and right, to find $\sigma(i)$. For example, if $T$ if the
non-ambiguous tree of Figure~\ref{fig:ana_zigzag}, then $\sigma(23)=13$ and
$\sigma(3)=7$, and more generally
\[
\sigma=(13\;1\;6\;20\;12\;5\;22\;10\;2\;23)\;(21)\;(18\;3\;7\;17\;15\;4\;19)\;(14\;9\;8\;16)\;(11)
\]
where we use the cyclic notation for permutations.
\begin{rque}\cite[Section 2]{Bur07}\label{rem:BurCyc}
It is a simple consequence of the construction that every cycle of $\varphi(T)$ corresponds either to an empty line or to a binary tree in $T'$.
\end{rque}
\begin{figure}[h]
\begin{center}
\includegraphics[scale=.6]{ana_zigzag}
\caption{Example of the bijection of Burstein. }
\label{fig:ana_zigzag}
\end{center}
\end{figure}
Let us recall that an excedance in a permutation $\sigma$ is an entry $i$ such that $\sigma(i)>i$. In the cyclic notation, it corresponds to an ascent (an integer followed -cyclically- by a greater one).
\begin{prop}\cite[Theorem 14]{SteiWill}
Let $\w_L$ and $\w_R$ be two positive integers. The map $\varphi$ is a
bijection between non-ambiguous trees of geometric size $\w_L\times\w_R$ and
permutations of size $\w_L+\w_R-1$ such that all their excedances are at
positions $1,\cdots,\w_R-1$.
\end{prop}
If we do the same construction without deleting the first column of $T$ (in this case the corresponding horizontal border edge is numbered with $0$), we get another map, which we denote $\psi$. As observed in Remark~\ref{rem:BurCyc}, $\psi(T)$ is a cyclic permutation.
\begin{figure}[h]
\begin{center}
\includegraphics[scale=.6]{ana_zigzag_cycle}
\caption{Example of the bijection $\psi$.}
\label{fig:ana_zigzag_cycle}
\end{center}
\end{figure}
The link between $\varphi$ and $\psi$ is given by the following lemma.
\begin{lem}\label{lem:ana_perm_cycles_cycle_bleus_rouges}
Let $T$ be a non-ambiguous tree. Let $c_1\cdots c_k$ be the decomposition in
cycles of $\varphi(T)$ such that the biggest element of $c_i$
is larger than the biggest element of $c_{i+1}$.
We denote by $m_i$ the representative of $c_i$ ending by its biggest element.
Then, a representative of $\psi(T)$ is the word $0m_1\cdots m_k$.
\end{lem}
This is illustrated on the running example of Figure \ref{fig:ana_zigzag_cycle}. On this non-ambiguous tree, denoted by $T$, $\psi(T)$ is the cycle
\[
\psi(T)=\left(
0\;
13\;1\;6\;20\;12\;5\;22\;10\;2\;23\;
21\;
18\;3\;7\;17\;15\;4\;19\;
14\;9\;8\;16\;
11
\right).
\]
Recall that $\varphi(T)$ is given by
\[
\varphi(T)=(13\;1\;6\;20\;12\;5\;22\;10\;2\;23)\;(21)\;(18\;3\;7\;17\;15\;4\;19)\;(14\;9\;8\;16)\;(11)
\]
In particular, $\psi(19)=14=\varphi(16)$.
\begin{proof}
Let us denote $\sigma=\varphi(T)$ and $c = \psi(T)$.
It is clear that only the images of entries that have a point in the leftmost column of $T$ (the one deleted in $T'$) are modified.
Let us denote by $i_1 < \cdots < i_k$ these entries.
The construction implies that $c(i_j ) = \sigma(i_{j-1})$ for $1<j\le k$, $c(i_1) = 0$ and $c(0) = \sigma(i_k)$. Whence the result.
\end{proof}
This lemma has two consequences.
The first one is that, as $\varphi$ is a bijection, $\psi$ is also one.
The second consequence is that the excedances of $\psi(T)$ are precisely those of $\varphi(T)$, together with $0$.
\begin{cor}\label{cor:ana_bij_cycle}
Let $\w_L$ and $\w_R$ be two positive integers. The map $\psi$ is a
bijection between non-ambiguous trees of geometric size $\w_L\times\w_R$ and
cycles of $\llbracket 0,\w_L+\w_R\rrbracket$ such that all their excedances are at
positions $0,\cdots,\w_R-1$.
\end{cor}
In order to ease future explanations, we number independently rows and columns.
We replace the integers $\llbracket 0,\w_R-1\rrbracket$ with $\llbracket
\b{1},\b{\w_R}\rrbracket$ by using the map $i\mapsto \b{(\w_R-i)}$, and the
integers $\llbracket \w_R,\w_R+\w_L-1\rrbracket$ with $\llbracket
\r{1},\r{\w_L}\rrbracket$ by using the map $i\mapsto \r{(i-\w_R+1)}$, as shown
in Figure~\ref{fig:ana_zigzag_cycle}. If we denote $T$ the non-ambiguous tree of
this figure, then $\psi(T)$ is equal to the cycle
\[
\left(
\b{9}\;\r{5}\;\b{8}\;\b{3}\;\r{12}\;\r{4}\;\b{4}\;\r{14}\;
\r{2}\;\b{7}\;\r{15}\;\r{13}\;\r{10}\;\b{6}\;\b{2}\;\r{9}\;
\r{7}\;\b{5}\;\r{11}\;\r{6}\;\r{1}\;\b{1}\;\r{8}\;\r{3}
\right).
\]
\begin{defi}\label{def:ana_cycle_bi}
Let $i$ and $j$ be two positive integers. We define a \emph{2-coloured block
decreasing cycle of size $i\times j$} as a cycle of the set
$\llbracket \r{1},\r{i} \rrbracket \cup \llbracket \b{1},\b{j} \rrbracket$
such that, if the image of an element $a$ is an element $b$ of the same
colour then $a>b$.
\end{defi}
Using this definition, the map $\psi$ is a bijection between non-ambiguous trees
of geometric size $i\times j$ and 2-coloured block decreasing cycles of size
$i\times j$. Moreover, the number of blue blocks of the 2-coloured block
decreasing cycles has a simple interpretation over non-ambiguous trees.
\begin{lem}\label{lem:ana_enumeration_nb_bloc_fixe}
Let $T$ be a non-ambiguous tree, then the hook number of $T$ is equal to the
number of blue block in $\psi(T)$.
\end{lem}
For example, if $T$ is the non-ambiguous tree of
Figure~\ref{fig:ana_zigzag_cycle}, its hook number is 7, which is also the number of
blue blocks in $\psi(T)$.
\begin{proof}
To prove this lemma, we show that, on the geometrical interpretation of a non-ambiguous tree, the vertical border edges on which ends a zigzag path starting from an horizontal border edge are exactly the vertical border edges of a line containing the right branch of a hook. For instance, on Figure~\ref{fig:ana_zigzag_cycle}, these vertical border edges are these numbered by $\textcolor{red}{15}$, $\textcolor{red}{14}$, $\textcolor{red}{12}$, $\textcolor{red}{11}$, $\textcolor{red}{9}$, $\textcolor{red}{8}$ and $\textcolor{red}{5}$.
A zigzag starting from an horizontal border edge first reach a node on the right branch of a hook (as it has no vertices above). Then every vertical step reaches either the root of a hook or an horizontal border edge, while every horizontal step reaches either a non-root node on the right branch of a hook or a vertical border edge. The only reached nodes are then on the right branch of a hook or its root. If such a zigzag path ends on a vertical border edge, the last step is horizontal from the root of a hook or a node on its right branch: the corresponding line contains the right branch of a node.
Similarly, a zigzag starting from a vertical border edge first reach a node on the left branch of a hook (as it has no vertices on its left). Then every horizontal step reaches either the root of a hook or a vertical border edge, while every vertical step reaches either a non-root point on the left branch of a hook or an horizontal border edge. The only reached nodes are then on the left branch of a hook or its root. The root of a hook in this case can only be reached after an horizontal step. If such a zigzag path ends on a vertical border edge, the last step is then horizontal from a non-root node on the left branch of a hook: the corresponding line does not contain the right branch of a node, which proves the result.
\end{proof}
From this lemma, we deduce the following proposition.
\begin{prop}
Let $i, j$ and $p$ be positive integers. The number of non-ambiguous
trees of geometric size $i\times j$ and hook number $p$ is
\[
(p-1)!\,p!\,\S(i,p)\,\S(j,p).
\]
Moreover, the doubly exponential generating series of non-empty
non-ambiguous trees, with weight on a NAT $T$ given by
\[
z^{hook(T)}
\frac{
x^{\w_L(T)}\,y^{\w_R(T)}
}{
\w_L(T)!\,\w_R(T)!
},
\]
is
\[
-\ln\left(1-z(e^x-1)(e^y-1)\right).
\]
\end{prop}
\begin{proof}
According to Corollary~\ref{cor:ana_bij_cycle} and
Lemma~\ref{lem:ana_enumeration_nb_bloc_fixe}, non-ambiguous trees with geometric size $i \times j$ with hook number $p$ is in bijection with 2-coloured
blocks decreasing cycles of size $i\times j$ with $p$ blue blocks, which are counted by the enumeration formula.
The generating series is the one corresponding to a cycle whose elements are pairs formed by a non-empty blue set and a non-empty red set. Recalling that the generating series of cycles is $-\ln(1-u)$ and the one of pairs of sets is $z(e^x-1)(e^y-1)$, we get the result by composition.
\end{proof}
We extend naturally this proposition with the parameters $\alpha$ and $\beta$.
\begin{thm}\label{th:ana_zigzag_alpha_beta}
Let $i, j$ and $p$ be positive integers. The number of non-ambiguous
trees of geometric size $i\times j$ and hook number $p$ is
\[
(p-1)!\,(\alpha+\beta)_{p-1}\,\operatorname S_{ 2,\alpha }(i,p)\,\operatorname S_{ 2,\beta }(j,p).
\]
Moreover, the doubly exponential generating series of non-empty
non-ambiguous trees, with weight on a NAT $T$ given by
\[
\alpha^{\LO(T)}
\beta^{\RO(T)}
z^{hook(T)}
\frac{
x^{\w_L(T)-1}\,y^{\w_R(T)-1}
}{
(\w_L(T)-1)!\,(\w_R(T)-1)!
},
\]
is
\[
\frac{
ze^{\alpha x+\beta y}
}{
\left(1-z(e^x-1)(e^y-1)\right)^{\alpha+\beta}
}.
\]
\end{thm}
\begin{proof}
We first prove the expression $\frac{
ze^{\alpha x+\beta y}
}{
\left(1-z(e^x-1)(e^y-1)\right)^{\alpha+\beta}
}$. Let us first remark that this can be rewritten as $z \times e^{\alpha x+\beta y} \times e^{-\alpha\ln\left(1-z(e^x-1)(e^y-1)\right)} \times e^{-\beta\ln\left(1-z(e^x-1)(e^y-1)\right)}$.
This expression is the translation in terms of generating function of the following decomposition. A non-ambiguous tree can be decomposed into three pieces:
\begin{itemize}
\item the set of child-free (i.e. with no child outside of the hook) vertices in the hook $h$ of the root (in yellow on figure \ref{fig:fig_nat2})
\item the set of non-ambiguous trees attached to the left branch of $h$ (in green on figure \ref{fig:fig_nat2})
\item and the set of non-ambiguous trees attached to the right branch of
$h$ (in orange on figure \ref{fig:fig_nat2}).
\end{itemize}
\begin{figure}
\begin{center}
\begin{tikzpicture}[scale=0.45]
\node (r) at (0,0) {(\textcolor{red}{15},\textcolor{blue}{9})};
\node[fill=orange!50,below right =1cm of r] (8b) {\textcolor{blue}{8}};
\node[fill=orange!50,below right =1cm of 8b] (7b) {\textcolor{blue}{7}};
\node[fill=orange!50,below left =0.5cm of 7b] (14r) {\textcolor{red}{14}};
\node[fill=orange!50,below right =0.5cm of 14r] (4b) {\textcolor{blue}{4}};
\node[fill=orange!50,below left =0.5cm of 4b] (12r) {\textcolor{red}{12}};
\node[fill=orange!50,below right =0.5cm of 12r] (3b) {\textcolor{blue}{3}};
\node[fill=yellow!50,below left =0.5cm of r] (13r) {\textcolor{red}{13}};
\node[fill=green!50,below left =1cm of 13r] (11r) {\textcolor{red}{11}};
\node[fill=green!50,below left =1cm of 11r] (8r) {\textcolor{red}{8}};
\node[fill=green!50,below right =0.5cm of 8r] (1b) {\textcolor{blue}{1}};
\node[fill=green!50, below right =0.5cm of 11r] (6b) {\textcolor{blue}{6}};
\node[fill=green!50,below right =0.5cm of 6b] (5b) {\textcolor{blue}{5}};
\node[fill=green!50,below left =0.5cm of 5b] (9r) {\textcolor{red}{9}};
\node[fill=green!50,below right =0.5cm of 9r] (2b) {\textcolor{blue}{2}};
\node[fill=green!50,below left =0.5cm of 6b] (10r) {\textcolor{red}{10}};
\node[fill=green!50,below left =0.5cm of 9r] (7r) {\textcolor{red}{7}};
\node[fill=green!50,below left =0.5cm of 1b] (6r) {\textcolor{red}{6}};
\node[fill=orange!50,below left =0.5cm of 8b] (5r){\textcolor{red}{5}};
\node[fill=orange!50,below left =0.5cm of 12r] (4r){\textcolor{red}{4}};
\node[fill=yellow!50,below left =0.5cm of 8r] (3r){\textcolor{red}{3}};
\node[fill=orange!50,below left =0.5cm of 14r] (2r) {\textcolor{red}{2}};
\node[fill=green!50,below left =0.5cm of 6r] (1r) {\textcolor{red}{1}};
\draw[line width=3pt] (3r)--(8r)--(11r)--(13r)--(r)--(8b)--(7b);
\draw (8b)--(5r);
\draw (7b)--(14r);
\draw[line width=3pt] (2r)--(14r)--(4b);
\draw (4b)--(12r);
\draw[line width=3pt] (4r)--(12r)--(3b);
\draw (11r)--(6b);
\draw[line width=3pt] (10r)--(6b)--(5b);
\draw (5b)--(9r);
\draw[line width=3pt] (7r)--(9r)--(2b);
\draw (8r)--(1b);
\draw[line width=3pt] (1b)--(6r)--(1r);
\end{tikzpicture}
\includegraphics[scale=0.6]{ana_zigzag2.pdf}
\caption{Illustration of the proof of Theorem \ref{th:ana_zigzag_alpha_beta} with both tree and geometrical representation of the same NAT}\label{fig:fig_nat2}
\end{center}
\end{figure}
There is one more hook in the initial tree than in the 3-tuple of sets
described above as the hook of the root does not appear in this decomposition:
this first remark justifies the first term $z$ in the product.
The yellow set is constituted by a set of red labels and a set of blue labels,
all red (resp. blue) elements contributing to $\LO(T)$ (resp. $\RO(T)$). The
generating series associated to this set is then $e^{\alpha x}e^{\beta y}$.
The green set is the set of sub-non-ambiguous trees attached to a
vertex of the hook of the root. Each subtree gives a contribution
$\alpha$ as its root belongs to the hook of the root of the initial
tree. Following Lemma \ref{lem:ana_enumeration_nb_bloc_fixe}, the
associated generating series is then
$e^{-\alpha\ln\left(1-z(e^x-1)(e^y-1)\right)}$.
In the same way, the generating series associated with the orange set is $e^{-\beta\ln\left(1-z(e^x-1)(e^y-1)\right)}$.
For the enumeration formula, let us consider a non-ambiguous tree $T$. We
use the same idea of decomposition. We delete the topmost row and the
leftmost column, before using the zigzags paths. Then, for example, one of the zigzag on the tree of Figure~\ref{fig:ana_zigzag_cycle} is illustrated on Figure~\ref{fig:fig_nat2}. For this tree, we get
\[
\r{(}\r{3}\r{)}\,
\r{(}\r{13}\r{)}\,
\r{(}\b{6}\;\b{2}\;\r{9}\;\r{7}\;\b{5}\;\r{11}\;\r{10}\r{)}\,
\r{(}\b{1}\;\r{8}\;\r{6}\;\r{1}\r{)}\,
\b{(}\b{8}\;\r{5}\b{)}\,
\b{(}\b{7}\;\b{3}\;\r{12}\;\r{4}\;\b{4}\;\r{14}\;\r{2}\b{)}.
\]
Red (resp. blue) parentheses means that the corresponding sub-non-ambiguous
tree is attached to the leftmost (resp. rightmost) branch. We
illustrate on the right of Figure~\ref{fig:fig_nat2} the cycle
$ \r{(}\b{6}\;\b{2}\;\r{9}\;\r{7}\;\b{5}\;\r{11}\;\r{10}\r{)}$.
This way, we obtain a partition of red (resp. blue)
labels in $p=hook(T)$ non-empty sets. The number of non-root points in the
leftmost column (resp. topmost row), with no right (resp. left) child, is
equal to the number of elements minus 1 in the subset containing the biggest
element. This explains the $\alpha$-analogue (resp. $\beta$-analogue) of
$\operatorname S_2$. Let us order and number the $p-1$ other blue subsets
with respect to their biggest element and pair each blue subset with the
red block to its right (in the same cycle). Keeping the same example,
we obtain
\[
\begin{array}{ccccccc}
& 6 & 5 & 4 & 3 & 2 & 1\\
\{\b{9}\} &\{\b{8}\} &\{\b{7}\;\b{3}\} &\{\b{6}\;\b{2}\}
&\{\b{5}\} &\{\b{4}\} &\{\b{1}\}\\
\{\r{15}\;\r{13}\;\r{3}\} &\{\r{5}\} &\{\r{12}\;\r{4}\}
&\{\r{9}\;\r{7}\} &\{\r{11}\;\r{10}\} &\{\r{14}\;\r{2}\}
&\{\r{8}\;\r{6}\;\r{1}\}\\
\end{array}.
\]
In a general setting, there are $(p-1)!$ pairing possibilities. Let us now
replace each pair with its corresponding number. We get
\[
\b{(}6\b{)}\;
\b{(}5\;2\b{)}\;
\r{(}4\;3\r{)}\;
\r{(}1\r{)}.
\]
In the end, in addition to the two partitions, we have a permutation of size
$p-1$ decomposed in cycles and whose cycles are coloured in red or in blue.
Each red (resp. blue) cycle counts for an $\alpha$ (resp. $\beta$), hence,
the generating polynomial of such permutations is $(\alpha+\beta)_{p-1}$.
We finally get the desired formula.
\end{proof}
As stated in the introduction, Proposition~\ref{gen_ser_nat} and
Theorem~\ref{thm_p} (in the case $\alpha=\beta=1$) were already proven by Clark
and Ehrenborg \cite{CE}. In the proof they gave, the statistic $p$ is
interpreted on permutations as follows.
\begin{defi}\label{def:ana_CE}
Let $i,j$ and $n$ be positive integers such that $n=i+j-1$. Let $\p$ be
a permutation of size $n$ such that all its excedances are at position
$\llbracket 1,j-1 \rrbracket$. The \emph{CE-statistic}
\footnote{
The "+1" does not appear in the definition of Clark and Ehrenborg. We
introduced it because there is a shift between the hook statistic and the
CE-statistic.
}
of $\p$ is the
positive integer
\[
\operatorname{CE}(\p)=|\{u\in\llbracket 1,j-1 \rrbracket,\, \p(u)>i\}|+1.
\]
\end{defi}
According to Lemma~\ref{lem:ana_enumeration_nb_bloc_fixe}, the hook
statistic corresponds to the number of blue blocks in 2-coloured block
decreasing cycles. The bijection between 2-coloured block decreasing cycles
and permutations with all their excedances at the beginning that we will
study is $\Theta=\varphi\circ\psi^{-1}$. The following lemma describes the
difference between the number of blue blocks and the CE-statistic.
\begin{lem}\label{lem:ana_blocs_bleus_CE}
Let $i$ and $j$ be positive integers and $c$ a 2-coloured block
decreasing cycle of size $i\times j$. Then, the number of blue blocks of
$c$ is equal to the CE-statistic of $\Theta(c)$
\begin{itemize}
\item minus 1, if $\b{j}$ has a blue element to its right and
\item plus 1, if $\r{1}$ has a blue integer, different from $\b{j}$
to its left.
\end{itemize}
\end{lem}
\begin{proof}
Let $\p$ be a permutation of size $n=i+j-1$ such that all its excedances
are at position $\llbracket 1,j-1 \rrbracket$. If we keep the
interpretation with red and blue integers, the CE-statistic can be interpreted as
\[
|\{
\b{u}\in\llbracket \b{1},\b{j-1} \rrbracket,\,
\p(\b{u})\in\llbracket \r{2},\r{i} \rrbracket
\}|+1.
\]
Hence, the CE-statistic is the number of blue blocks of $\p$ without
$\r{1}$ to their right plus 1. Hence, we should study how the number of
blue blocks behave with respect to $\Theta$. Using
Lemma~\ref{lem:ana_perm_cycles_cycle_bleus_rouges} we obtain the
conditions of Lemma~\ref{lem:ana_blocs_bleus_CE}.
\end{proof}
Following the previous lemma, $\Theta$ is not sufficient to prove the
equidistribution between the hook statistic and the CE-statistic. We need
one last involution. Let $m$ be the representative of a 2-coloured block
decreasing sequence $c$ of size $i\times j$ such that $\b{j}$ is on the left of
$m$. The word $m$ can be factorised as $m=\b{j}b_1\cdots b_k\r{1}m'$ where
the $b_i$ are maximal blocks of same colours. Let $\omega$ be the involution
such that if $k$ is even then a representative of $\omega(c)$ is
$m=\b{j}b_2b_1\cdots b_kb_{k-1}\r{1}m'$, and if $k$ is odd then
$\omega(c)=c$.
\begin{prop}\label{prop:ana_blocs_bleus_CE_equi}
Let $c$ be a 2-coloured block decreasing sequence. The number of blue
blocks of $c$ is equal to the CE-statistic of $\Theta(\omega(c))$.
\end{prop}
\begin{cor}
The hook statistic on non-ambiguous trees and the CE-statistic on
permutations with all their excedances at the beginning are
equidistributed.
\end{cor}
\subsection{q-analogues of the hook formula}\label{q_analogs}
As for binary trees, there exist $q$-analogues of the hook formula for NATs
of a given shape, associated to either the number of inversions
or the major index. There are two ingredients: first we need to associate a pair of
permutations to a non-ambiguous tree, and second we need to give a $q$-analogue
of the bilinear map $\Bxy$. It turns out that it is possible to use two
different $q$'s namely $q_R$ and $q_L$ for the derivative and integral in $x$
and $y$.
The first step, in order to formulate a $q$-hook formula, is to associate to any
non-empty non-ambiguous tree $T$ a pair of permutations
\begin{equation*}
\sigma(T)=(\sigma_L(T),
\sigma_R(T))\in\SG_{\LV(T)} \times \SG_{\RV(T)}.
\end{equation*}
\begin{defi}
Let $T$ be a non-ambiguous tree. Then $\sigma_L(T)$ is obtained by
performing a left postfix reading of the left labels: precisely we
recursively read trees $\NodeBT{L}{R}$ by reading the left labels of $L$,
then the left labels of $R$ and finally the label of the root if it is a left
child. The permutation $\sigma_R(T)$ is defined similarly by reading the
right labels in the right subtree, then in the left subtree and finally reading
the root.
\end{defi}
If we consider the example of Figure~\ref{fig_exple_nat} the two associated permutations are $\sigma_L(T) = (2,$ $1,$ $4,$ $3,$ $6,$ $10,$ $8,$ $9,$ $5,$ $7)$ and $\sigma_R(T) = (1,$ $2, 3,$ $4, 5,$ $7, 11,$ $9, 6,$ $8, 10)$.
Recall that the \emph{number of inversions} of a permutation $\sigma\in\SG_n$
is the number of $i<j\leqslant n$ such that $\sigma(i)>\sigma(j)$ (denoted by Inv from now on). A \emph{descent} of
$\sigma$ is a $i<n$ such that $\sigma(i)>\sigma(i+1)$ and the \emph{inverse
major index} of $\sigma$ is the sum of the descents of $\sigma^{-1}$ (denoted by iMaj from now on). Finally
for a repetition free word $w$ of length~$l$ we denote by $\std(w)$ the
permutations in $\SG_l$ obtained by renumbering $w$ keeping the order of the
letters. For example $\std(36482)=24351$. We define as usual the $q$-integer
$[n]_q:= \frac{1-q^n}{1-q}$, and the $q$-factorial $[n]_q!:= \prod_{i=1}^{n}
[i]_q$.
For a non-ambiguous tree $\N$ and a statistic $S\in\{\inv,\imaj\}$, we define the weight of a NAT $T$ as
\begin{equation}
\omega_S(T):= q_L^{S(\sigma_L(T))}q_R^{S(\sigma_R(T))}.
\end{equation}
The following theorem is a $q$-analogue of Proposition \ref{prop_hook_nat}.
\begin{thm}\label{thm-q-hook}
For any non-empty binary tree $B$,
\begin{equation}
\sum_{T\in\NAT(\B)} \omega_{S}(T) =
\frac{ |\LV(\B)|_{q_L}! \cdot |\RV(\B)|_{q_R}!}{
\displaystyle\prod_{U: \text{left child}}[\EL(U)]_{q_L} \cdot
\prod_{U: \text{right child}}[\ER(U)]_{q_R}
}\,.
\end{equation}
\end{thm}
Going back to the non-ambiguous tree of Figure~\ref{fig_exple_nat}, the
inversions numbers are $\inv(\sigma_L(T)) = 11$ and $\inv(\sigma_R(T)) = 7$, so
that $w_{\inv}(T) = q_L^{11}q_R^{7}$. For the inverse major index, we get the permutations $\sigma_L(T)^{-1} = (2, 1, 4, 3, 9, 5, 10, 7, 8, 6)$ and $
\sigma_R(T)^{-1} = (1, 2, 3, 4, 5, 9, 6, 10, 8, 11, 7)$.
Consequently, $\imaj(\sigma_L(T)) = 1 + 3 + 5 + 7 + 9 = 25$
and $\imaj(\sigma_R(T)) = 6 + 8 + 10 = 24$ so that
$w_{\imaj}(T) = q_L^{25}q_R^{24}$.
The argument of the proof follows the same path as for the hook formula, using
pumping functions. Recall that the $q$-derivative and $q$-integral are
defined as $\partial_{x, q} x^n:= [n]_qx^{n-1}$ and
$\int_{x, q} u^n\d u:= \frac{x^{n+1}}{[n+1]_q}$.
Then the $(q_L, q_R)$-analogue of the pumping function is given by
\begin{equation}
\Bxy_q(f, g) = \int_{x,q_L}\int_{y,q_R}
\partial_{x,q_L}g(u,v)\cdot
\partial_{y,q_R}f(u,v)
\diff u\;\diff v.
\end{equation}
We also define recursively $\Bx_q(B)$ by
$\Bx_{q}\left(\NodeBT{L}{R}\right) = \Bxy_q\left(\Bx_q(L), \Bx_q(R)\right)\,$, with initial conditions $\Bx_q(\emptyset_R):= x$ and $\Bx_q(\emptyset_L):=y$. Then
the main idea is to go through a bilinear function on permutations. We
write $\operatorname{\mathbb{Q}}\SG$ the vector space of formal sums of permutations. For any
permutation $\sigma\in\SG_n$ we write $\int\sigma=\sigma[n+1]$ the
permutation in $\SG_{n+1}$ obtained by adding $n+1$ at the end. Again we
extend $\int$ by linearity.
\begin{defi}
The bilinear map
$\BSG:\operatorname{\mathbb{Q}}\SG\times\operatorname{\mathbb{Q}}\SG\mapsto\operatorname{\mathbb{Q}}\SG$ is defined
for $\sigma\in\SG_m$ and $\mu\in\SG_n$ by
$$ \BSG(\sigma,\mu) =
\sum_{\substack{uv\in\SG_{m+n+1} \\ \std(u)=\int\sigma \\ \std(v)=\mu}} uv\,.$$
\end{defi}
For example
$\BSG(21,12)=21345+21435+21534+31425+31524+
41523+32415+32514+42513+43512$.
\begin{lem}
For two non-empty non-ambiguous trees $C$ and $D$, we have
\begin{align*}
\sum_{T\in\BNAT(C, D)} \sigma_L(T) &=
\BSG(\sigma_L(C),\sigma_L(D)) \\
\text{and} \qquad
\sum_{T\in\BNAT(C, D)} \sigma_R(T) &=
\BSG(\sigma_R(D),\sigma_R(C)).
\end{align*}
\end{lem}
\begin{proof}
Let $T$ be a NAT appearing in $\BNAT(C, D)$, and let
its shape be the binary tree $B$.
We may write $\sigma_L(T)=uv$ where $u$ is the word given
by the left postfix reading of the left labels of the left
subtree of $T$, and $v$ the one of its right subtree.
By definition, we have $\std(u)=\int\sigma_L(C)$ and $\std(v)=\sigma_L(D)$. The symbol $\int$ in the first equality comes from the fact that the root of $C$ becomes a left
label in $T$.
Moreover, the choice of a $v$ such that $\std(v)=\sigma_L(D)$
is equivalent to the choice of a labelling of the left vertices
of the right subtree $B_R$ of $B$, such that the order
of the labels in $D$ is respected. The same holds for $u$,
and the corresponding root $C$.
This proves the first equality.
The second one is obtained in the same way.
We have to take care of two things:
that the right postfix reading
reads first the right subtree,
and that the root of $D$ here becomes a right label in $T$.
The order of $C$ and $D$ is thus reversed in $\BSG$.
\end{proof}
Let us define a bilinear map $\Phi_S$. For two permutations
$\tau\in\SG_{m}$ and $\pi\in\SG_{n}$
\begin{equation}
\Phi_S((\tau,\pi)):=
q_L^{S(\tau)}\frac{x^{m+1}}{[m+1]_{q_L}!}\,
q_R^{S(\pi)} \frac{y^{n+1}}{[n+1]_{q_R}!}\,
\end{equation}
We shall use the same notation
for NATs: $\Phi_S(T)= \Phi_S(\sigma_L(T),\sigma_R(T))$.
The main result, analogous to Proposition \ref{prop:commutation_phi} is the following.
\begin{prop}\label{prop:qhook-commut}
For any non-empty binary tree $B$, we have the relation
$$\Phi_S(\NAT(B)) = \Bx_q(B).$$
\end{prop}
As for Proposition \ref{prop:commutation_phi},
Proposition \ref{prop:qhook-commut} is derived
recursively.
The analogue of Lemma \ref{lem:com_phi}
is the following.
\begin{lem}\label{lem:com_phi_q}
For $(T_1, T_2)$ a pair of NATs in $\operatorname{\mathbb{Q}}\NAT_L\times\operatorname{\mathbb{Q}}\NAT_R$, one has
$$\Phi_S(\BNAT(T_1, T_2)) = \Bxy_q(\Phi_S(T_1),\Phi_S(T_2)).$$
\end{lem}
To prove it, we first need a technical result.
\begin{lem}\label{lem:technical-q}
For a statistic $S\in\{\inv,\imaj\}$,
and for $\tau\in\SG_{m}$ and $\pi\in\SG_{n}$, we have:
$$
\sum_{\substack{\theta=uv\in\SG_{n+m+1} \\ \std(u)=\int\tau \\ \std(v)=\pi}}
q^{S(\theta)}
=
q^{S(\tau)+S(\pi)}\binom{m+n+1}{m+1}_{q}.
$$
\end{lem}
\begin{proof}
The case $S=\inv$ is easier to prove. The $q$-binomial consists in choosing a
permutation $\theta$ such that $\theta(1)<\cdots<\theta(m+1)$ and
$\theta(m+2)<\cdots<\theta(m+n+1)$. The term $q^{S(\tau)+S(\pi)}$ comes from the reordering of the $\theta(i)$ in order to have $\std(\theta(1)\cdots\theta(m+1))=\int\tau$ and $\std(\theta(m+2)\cdots\theta(m+n+1))=\pi$.
In order to prove the case $S=\imaj$, we first obtain the following two equations
\begin{equation}\label{eq:imaj_lemma_1}
\sum_{\substack{\theta=uv\in\SG_{n+m+1} \\ n+m+1\in u \\ \std(u)=\int\tau \\ \std(v)=\pi}}
q^{\imaj(\theta)}
=
q^{\imaj(\tau)+\imaj(\pi)+n}\binom{m+n}{n}_{q},
\end{equation}
and
\begin{equation}\label{eq:imaj_lemma_2}
\sum_{\substack{\theta=uv\in\SG_{n+m+1} \\ n+m+1\in v \\ \std(u)=\int\tau \\ \std(v)=\pi}}
q^{\imaj(\theta)}
=
q^{\imaj(\tau)+\imaj(\pi)}\binom{m+n}{n-1}_{q}.
\end{equation}
Equation~\eqref{eq:imaj_lemma_1} is a consequence of Equation~36 in
\cite{HivNovThib08}. We prove Equation~\eqref{eq:imaj_lemma_2} by induction on
$n+m$ and by distinguishing the two cases $n+m\in u$ and $n+m\in v$.
In the first case, by using \eqref{eq:imaj_lemma_1} we have
$$
\sum_{\substack{\theta=uv\in\SG_{n+m+1} \\ n+m+1\in v \\
n+m\in u\\ \std(u)=\int\tau \\ \std(v)=\pi}}
q^{\imaj(\theta)}
=
\sum_{\substack{uv'\in\SG_{n+m+1} \\
n+m\in u\\ \std(u)=\int\tau \\ \std(v')=\pi'}}
q^{\imaj(uv')}
=
q^{\imaj(\tau)+\imaj(\pi')+n-1}\binom{m+n-1}{n-1}_{q},
$$
where $\pi'$ is obtained from $\pi$ by deleting the entry $n$.
Let us set $\epsilon$ equal to $1$ if $n$ is to the left of $n-1$
in $\pi$ and $0$ otherwise. Then we may write:
$\imaj(\pi)=\imaj(\pi')+\epsilon(n-1)$, whence
$$
\sum_{\substack{\theta=uv\in\SG_{n+m+1} \\ n+m+1\in v \\
n+m\in u\\ \std(u)=\int\tau \\ \std(v)=\pi}}
q^{\imaj(\theta)}
=
q^{\imaj(\tau)+\imaj(\pi')+(1-\epsilon)(n-1)}\binom{m+n-1}{n-1}_{q}.
$$
In the second case, by using \eqref{eq:imaj_lemma_1} by induction, we get:
\begin{align*}
\sum_{\substack{\theta=uv\in\SG_{n+m+1} \\ n+m+1\in v \\
n+m\in v\\ \std(u)=\int\tau \\ \std(v)=\pi}}
q^{\imaj(\theta)}
&=q^{\epsilon(m+n)}\sum_{\substack{uv'\in\SG_{n+m+1} \\
n+m\in v'\\ \std(u)=\int\tau \\ \std(v')=\pi'}}
q^{\imaj(uv')}\\
&=q^{\imaj(\tau)+\imaj(\pi')+\epsilon(m+n)}\binom{m+n-1}{n-2}_{q}\\
&=q^{\imaj(\tau)+\imaj(\pi)+\epsilon(m+1)}\binom{m+n-1}{n-2}_{q}.
\end{align*}
Then for any of $\epsilon$, we have
$$
q^{\epsilon(m+1)}\binom{m+n-1}{n-2}_{q}
+
q^{(1-\epsilon)(m+1)}\binom{n-1}{n-1}_{q}
=
\binom{m+n}{n-1}_{q}.
$$
Since \eqref{eq:imaj_lemma_1} is trivially true when $\pi$ is the unique permutation of size $1$, the proof by induction is complete.
To conclude the proof of the lemma, we add Equations~\eqref{eq:imaj_lemma_1} and \eqref{eq:imaj_lemma_2}
and use the identity
$$
\binom{a+b+1}{a}_{q}=
\binom{a+b}{a}_{q}
+
\binom{a+b}{a-1}_{q}.$$
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:com_phi_q}]
Let us set
$$
u=\Phi_S(T_1)=
q_L^{S(\sigma_L(T_1))}\frac{x^{w_L(T_1)}}{[w_L(T_1)]_{q_L}!}\,
q_R^{S(\sigma_R(T_1))} \frac{y^{w_R(T_1)}}{[w_R(T_1)]_{q_R}!}\,
$$
and
$$
v=\Phi_S(T_2)=
q_L^{S(\sigma_L(T_2))}\frac{x^{w_L(T_2)}}{[w_L(T_2)]_{q_L}!}\,
q_R^{S(\sigma_R(T_1))} \frac{y^{w_R(T_1)}}{[w_R(T_1)]_{q_R}!}\,.
$$
Thanks to Lemma \ref{lem:technical-q}, we have to prove that
\begin{align*}
\Bxy_q(u,v)
&=
q_L^{S(\sigma_L(T_1))+S(\sigma_L(T_2))}
\binom{w_L(T_1)+w_L(T_2)-1}{w_L(T_1)}_{q_L}\\
&\times
q_R^{S(\sigma_R(T_1))+S(\sigma_R(T_2))}
\binom{w_R(T_1)+w_R(T_2)-1}{w_R(T_2)}_{q_R}\\
&\times
\frac
{x^{w_L(T_1)+w_L(T_2)}\,y^{w_R(T_1)+w_R(T_2)}}
{[w_L(T_1)+w_L(T_2)]_{q_L}!\,[w_R(T_1)+w_R(T_2)]_{q_R}!}
.
\end{align*}
This is done by a simple computation, as
in the proof of Lemma \ref{lem:com_phi}.
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop:qhook-commut}]
The proof is exactly the same as for Proposition \ref{prop:commutation_phi}. We just have to replace $\Phi$
by $\Phi_S$ and $\Bx$ by $\Bx_q$.
\end{proof}
We are now in a position to prove Theorem \ref{thm-q-hook}.
\begin{proof}[Proof of Theorem \ref{thm-q-hook}]
By definition, we have for any NAT $T$: $\Phi_S(T)=\omega_S(T)\times
\frac{x^{w_L(T)}}{[w_L(T)]_{q_L}!}\frac{y^{w_R(T)}}{[w_R(T)]_{q_L}!}$.
Proposition \ref{prop:qhook-commut} implies that
for a binary tree $B$, $\Phi_S([B])=\Bx_q(B)$, and
we may compute $\Bx_q(B)$ recursively.
Let us suppose that neither $B_L$ nor $B_R$ is empty.
We then have:
\begin{align*}
\Bx_q(B)
&=\Bxy_q(\Bx_q(B_L),\Bx_q(B_R))\\
&=\int_{x,q_L}\int_{y,q_R}
\partial_{x,q_L}\Bx_q(B_L)(u,v)\cdot
\partial_{y,q_R}\Bx_q(B_R)(u,v)
\diff u\;\diff v\\
&=\frac{[w_R(B_L)]_{q_L}[w_L(B_R)]_{q_R}}{[w_L(B)]_{q_L}[w_R(B)]_{q_R}}\Bx_q(B_L)\Bx_q(B_R).
\end{align*}
The same recursive computation as in the proof of Proposition \ref{prop_hook_nat} leads to
$$
\frac{\sum_{T\in\NAT(B)}\omega_S(T)}{[w_L(B)]_{q_L}![w_R(B)]_{q_R}!}
=
\frac{1}{
\displaystyle\prod_{U: \text{left child}}[\EL(U)]_{q_L} \cdot
\prod_{U: \text{right child}}[\ER(U)]_{q_R}
}
$$
which gives Theorem \ref{thm-q-hook}.
\end{proof}
We conclude this section by an example, using the same tree as in Example \ref{exple_hook}.
\begin{exple}
Let $B = \scalebox{0.5}
{ \newcommand{\node (a) {(\color{red}$4$,\color{blue}$5$)}{\node[draw,fill,circle] (a) {$$};}
\newcommand{\node (b) {\color{red}$3$}{\node[draw,fill,circle] (b) {$$};}
\newcommand{\node (c) {\color{red}$1$}{\node[draw,fill,circle] (c) {$$};}
\newcommand{\node (d) {\color{blue}$2$}{\node[draw,fill,circle] (d) {$$};}
\newcommand{\node (e) {\color{blue}$4$}{\node[draw,fill,circle] (e) {$$};}
\newcommand{\node (f) {\color{red}$2$}{\node[draw,fill,circle] (f) {$$};}
\newcommand{\node (g) {\color{blue}$3$}{\node[draw,fill,circle] (g) {$$};}
\newcommand{\node (h) {\color{blue}$1$}{\node[draw,fill,circle] (h) {$$};}
\begin{tikzpicture}[scale=0.5,baseline=(current bounding box.center),
every node/.style={inner sep=2pt}]
\matrix[column sep=.15cm, row sep=.15cm,ampersand replacement=\&]{
\& \& \& \node (a) {(\color{red}$4$,\color{blue}$5$) \& \& \& \& \& \\
\& \node (b) {\color{red}$3$ \& \& \& \& \& \& \node (e) {\color{blue}$4$ \& \\
\node (c) {\color{red}$1$ \& \& \node (d) {\color{blue}$2$ \& \& \& \node (f) {\color{red}$2$ \& \& \& \node (h) {\color{blue}$1$ \\
\& \& \& \& \& \& \node (g) {\color{blue}$3$ \& \& \\
};
\path[thick] (b) edge (c) edge (d)
(f) edge (g)
(e) edge (f)
(a) edge (b) edge (e)
(e) edge (h);
\end{tikzpicture}}$.
The $q$- hook formula is given by:
\begin{align*}
\sum_{T\in\NAT(\B)} \omega_{S}(T) &=
\frac{ [3]_{q_L}! [4]_{q_R}!}{
\left([1]_{q_L}[2]_{q_L}[1]_{q_L} \right) \cdot
\left( [1]_{q_R}[1]_{q_R}[3]_{q_R}[1]_{q_R}\right)
} \\
&=(q_R^3 + q_R^2 + q_R + 1)(q_L^2 +
q_L + 1)(q_R + 1) \,.
\end{align*}
Expanding this expression, one finds that the coefficient of
$q_R^2q_L$ is $2$.
For the $\imaj$ statistic it corresponds to the two following non-ambiguous
trees which are shown with their associated left and right permutations:
\[
{\newcommand{\node (a) {(\color{red}$4$,\color{blue}$5$)}{\node (a) {(\color{red}$4$,\color{blue}$5$)}
;}\newcommand{\node (b) {\color{red}$3$}{\node (b) {\color{red}$3$}
;}\newcommand{\node (c) {\color{red}$1$}{\node (c) {\color{red}$2$}
;}\newcommand{\node (d) {\color{blue}$2$}{\node (d) {\color{blue}$2$}
;}\newcommand{\node (e) {\color{blue}$4$}{\node (e) {\color{blue}$4$}
;}\newcommand{\node (f) {\color{red}$2$}{\node (f) {\color{red}$1$}
;}\newcommand{\node (g) {\color{blue}$3$}{\node (g) {\color{blue}$3$}
;}\newcommand{\node (h) {\color{blue}$1$}{\node (h) {\color{blue}$1$}
;}\begin{tikzpicture}[baseline=(current bounding box.center)]
\matrix[column sep=1.5mm, row sep=1.5mm,ampersand replacement=\&]{
\& \& \& \node (a) {(\color{red}$4$,\color{blue}$5$) \& \& \& \& \& \\
\& \node (b) {\color{red}$3$ \& \& \& \& \& \& \node (e) {\color{blue}$4$ \& \\
\node (c) {\color{red}$1$ \& \& \node (d) {\color{blue}$2$ \& \& \& \node (f) {\color{red}$2$ \& \& \& \node (h) {\color{blue}$1$ \\
\& \& \& \& \& \& \node (g) {\color{blue}$3$ \& \& \\
};
\path (b) edge (c) edge (d)
(f) edge (g)
(e) edge (f) edge (h)
(a) edge (b) edge (e);
\end{tikzpicture}} \left((2, 3, 1), (1, 3, 4, 2)\right),\]
\[{ \newcommand{\node (a) {(\color{red}$4$,\color{blue}$5$)}{\node (a) {(\color{red}$4$,\color{blue}$5$)}
;}\newcommand{\node (b) {\color{red}$3$}{\node (b) {\color{red}$3$}
;}\newcommand{\node (c) {\color{red}$1$}{\node (c) {\color{red}$2$}
;}\newcommand{\node (d) {\color{blue}$2$}{\node (d) {\color{blue}$2$}
;}\newcommand{\node (e) {\color{blue}$4$}{\node (e) {\color{blue}$4$}
;}\newcommand{\node (f) {\color{red}$2$}{\node (f) {\color{red}$1$}
;}\newcommand{\node (g) {\color{blue}$3$}{\node (g) {\color{blue}$1$}
;}\newcommand{\node (h) {\color{blue}$1$}{\node (h) {\color{blue}$3$}
;}\begin{tikzpicture}[baseline=(current bounding box.center)]
\matrix[column sep=1.5mm, row sep=1.5mm,ampersand replacement=\&]{
\& \& \& \node (a) {(\color{red}$4$,\color{blue}$5$) \& \& \& \& \& \\
\& \node (b) {\color{red}$3$ \& \& \& \& \& \& \node (e) {\color{blue}$4$ \& \\
\node (c) {\color{red}$1$ \& \& \node (d) {\color{blue}$2$ \& \& \& \node (f) {\color{red}$2$ \& \& \& \node (h) {\color{blue}$1$ \\
\& \& \& \& \& \& \node (g) {\color{blue}$3$ \& \& \\
};
\path (b) edge (c) edge (d)
(f) edge (g)
(e) edge (f) edge (h)
(a) edge (b) edge (e);
\end{tikzpicture}}
\left((2, 3, 1), (3, 1, 4, 2)\right).\]
For the $\inv$ statistic it corresponds to the two following non-ambiguous
trees which are shown with their associated left and right permutations:
\[
{\newcommand{\node (a) {(\color{red}$4$,\color{blue}$5$)}{\node (a) {(\color{red}$4$,\color{blue}$5$)}
;}\newcommand{\node (b) {\color{red}$3$}{\node (b) {\color{red}$3$}
;}\newcommand{\node (c) {\color{red}$1$}{\node (c) {\color{red}$1$}
;}\newcommand{\node (d) {\color{blue}$2$}{\node (d) {\color{blue}$3$}
;}\newcommand{\node (e) {\color{blue}$4$}{\node (e) {\color{blue}$4$}
;}\newcommand{\node (f) {\color{red}$2$}{\node (f) {\color{red}$2$}
;}\newcommand{\node (g) {\color{blue}$3$}{\node (g) {\color{blue}$1$}
;}\newcommand{\node (h) {\color{blue}$1$}{\node (h) {\color{blue}$2$}
;}\begin{tikzpicture}[baseline=(current bounding box.center)]
\matrix[column sep=1.5mm, row sep=1.5mm,ampersand replacement=\&]{
\& \& \& \node (a) {(\color{red}$4$,\color{blue}$5$) \& \& \& \& \& \\
\& \node (b) {\color{red}$3$ \& \& \& \& \& \& \node (e) {\color{blue}$4$ \& \\
\node (c) {\color{red}$1$ \& \& \node (d) {\color{blue}$2$ \& \& \& \node (f) {\color{red}$2$ \& \& \& \node (h) {\color{blue}$1$ \\
\& \& \& \& \& \& \node (g) {\color{blue}$3$ \& \& \\
};
\path (b) edge (c) edge (d)
(f) edge (g)
(e) edge (f) edge (h)
(a) edge (b) edge (e);
\end{tikzpicture}} \left((1,3,2), (2,1,4,3)\right),\]
\[{ \newcommand{\node (a) {(\color{red}$4$,\color{blue}$5$)}{\node (a) {(\color{red}$4$,\color{blue}$5$)}
;}\newcommand{\node (b) {\color{red}$3$}{\node (b) {\color{red}$3$}
;}\newcommand{\node (c) {\color{red}$1$}{\node (c) {\color{red}$1$}
;}\newcommand{\node (d) {\color{blue}$2$}{\node (d) {\color{blue}$2$}
;}\newcommand{\node (e) {\color{blue}$4$}{\node (e) {\color{blue}$4$}
;}\newcommand{\node (f) {\color{red}$2$}{\node (f) {\color{red}$2$}
;}\newcommand{\node (g) {\color{blue}$3$}{\node (g) {\color{blue}$3$}
;}\newcommand{\node (h) {\color{blue}$1$}{\node (h) {\color{blue}$1$}
;}\begin{tikzpicture}[baseline=(current bounding box.center)]
\matrix[column sep=1.5mm, row sep=1.5mm,ampersand replacement=\&]{
\& \& \& \node (a) {(\color{red}$4$,\color{blue}$5$) \& \& \& \& \& \\
\& \node (b) {\color{red}$3$ \& \& \& \& \& \& \node (e) {\color{blue}$4$ \& \\
\node (c) {\color{red}$1$ \& \& \node (d) {\color{blue}$2$ \& \& \& \node (f) {\color{red}$2$ \& \& \& \node (h) {\color{blue}$1$ \\
\& \& \& \& \& \& \node (g) {\color{blue}$3$ \& \& \\
};
\path (b) edge (c) edge (d)
(f) edge (g)
(e) edge (f) edge (h)
(a) edge (b) edge (e);
\end{tikzpicture}}
\left((1,3,2), ( 1,3, 4, 2)\right).\]
\end{exple}
\section{Non-ambiguous trees in higher dimension}
\label{gnat}
In this section we give a generalisation of NATs to higher dimensions.
NATs are defined as binary trees whose vertices are
embedded in $\mathbb{Z}^2$, and edges are objects of dimension 1 (segments).
Let $d \ge k \ge 1$ be two integers.
In higher dimension, binary trees are replaced by $\binom{d}{k}$-ary trees
embedded in $\mathbb{Z}^d$
and edges are objects of dimension $k$.
As in Section~\ref{differential} we obtain differential equations for these objects.
\subsection{Definitions}
\label{def_natdk}
We call \emph{$(d,k)$-direction} a subset of cardinality $k$ of $\{1, \ldots, d\}$.
The set of $(d,k)$-directions is denoted by $\edir$.
A $(d,k)$-tuple is a $d$-tuple of $(\mathbb{N}\cup\{\bullet\})^d$,
in which $k$ entries are integers and $d-k$ are $\bullet$.
For instance,
$
(\bullet, \ 1, \ \bullet, \ 5,\ 2, \ \bullet, \ \bullet, \ 3, \ \bullet)
$
is a $(9,4)$-tuple.
The direction of a $(d,k)$-tuple $U$ is the set of indices of $U$
corresponding to entries different from $\bullet$.
For instance, the direction of our preceding example is $\{2,4,5,8\}$.
\begin{defi}
A \emph{$\binom{\d}{\k}$-ary} tree $\M$ is a tree in which the children of a given vertex
are indexed by a $(\d,\k)$-direction.
\end{defi}
A $\binom{\d}{\k}$-ary tree will be represented as an ordered tree where the
children of a vertex $S$ are drawn from left to right with respect to the
lexicographic order of their indices.
If a vertex $S$ has no child associated to an index $\dir$, we draw an half
edge in this direction.
Two examples are drawn in Figure~\ref{fig_gnat}.
As for binary trees, for each $(d,k)$-direction $\dir$ we consider that there is
a $\binom{\d}{\k}$-ary tree of size 0: the \emph{empty $\binom{\d}{\k}$-ary
tree of direction $\dir$} noted $\emptyset_\dir$.
\begin{defi}\label{def_gnat}
A \emph{non-ambiguous tree of dimension $(d,k)$} is a labelled
\emph{$\binom{d}{k}$-ary} tree such that:
\begin{enumerate}[ref={\thecor.\arabic*}]
\item\label{gnat_dk_tuple} a child of index $\dir$ is labelled with a
$(d,k)$-tuple of direction $\dir$ and \label{gnat_root} the root is
labelled with a $(d,d)$-tuple;
\item\label{gnat_growth} for any descendant $U$ of $V$, if the $i$-th
component of $U$ and $V$ are different from $\bullet$, then the $i$-th
component of $V$ is strictly greater than the $i$-th component of $U$;
\item\label{gnat_distinct} for each $i \in \i{1}{\d}$, all the $i$th
components, different from $\bullet$, are pairwise distinct;
\item\label{gnat_interval} the set of $i$th components different from
$\bullet$ of every vertices in the tree is an interval whose minimum is
$1$.
\end{enumerate}
The set of non-ambiguous trees of dimensions $(d,k)$ is denoted by $\NATdk$.
\end{defi}
\begin{rque}
The usual NATs are non-ambiguous trees of dimension $(2,1)$.
\end{rque}
We write \nat for a non-ambiguous tree (of dimensions $(d,k)$).
Figure~\ref{fig_gnat} gives an example of a \nat[3][1] and a \nat[3][2].
\begin{figure}[h]
\begin{center}
\resizebox{\textwidth}{!}{
\includegraphics[scale=0.3]{NAT31.pdf}
\includegraphics[scale=0.5]{NAT31gem.pdf}}
\hspace{1cm}
\includegraphics[scale=0.5]{NAT32.pdf}
\includegraphics[scale=0.5]{NAT32gem.pdf}
\end{center}
\caption{A NAT of dimension $(3,1)$ and its geometric representation (above) and a NAT of dimension $(3,2)$ and its geometric representation
(below).}
\label{fig_gnat}\label{ex_kdtree}
\end{figure}
\begin{defi}\label{gnat_gs}
The \emph{geometric size} of a \nat is the $d$-tuple of integers
$(w_1, \ldots, w_d)$ which labels the root of the \nat, it is denoted by $\gs$.
The \emph{$\dir$-size} of a \nat is the number of vertices in the tree of
direction $\dir$, the set of such vertices is denoted by $\DV$.
\end{defi}
Proposition~\ref{prop_relation_sizes} gives the relation between the
geometric size and the $\dir$-size of a non-ambiguous trees.
\begin{prop}
\label{prop_relation_sizes}
Let $\M$ be a $\binom{d}{k}$-ary tree, the root label is constant on elements of
$\NATdk$ of shape $M$ ($\NATdk(\M)$):
$$
\wi_i(M)
:=
\wi_i
=
\sum_{\dir \in \edir \ \mid\ i \in \dir} |\DV(\M)| +1
.
$$
\end{prop}
\subsection{Associated differential equations}
As in section~\ref{differential}, we define a weight $\Phi$, a
$\binom{d}{k}$-linear map $\BNAT$ on generalised non-ambiguous trees of dimension
$(d,k)$ and $\binom{d}{k}$-linear map $\Bxy$ on multivariate power series such
that ``$\Phi\circ\BNAT=\Bxy\circ\Phi$''. From this identity, we obtain a hook formula
for the number of generalised non-ambiguous trees with a fixed underlying tree
and a differential equation satisfied by the generating function of
generalised non-ambiguous trees of dimension $(d,k)$.
Let $\emptyset_\dir$ be the empty $\binom{d}{k}$-ary tree of direction $\dir$, by convention
\[
\wi_i(\emptyset_\dir)=
\left \{
\begin{array}{ll}
0 &\text{ if } i\in\dir,\\
1 &\text{ else.}
\end{array}
\right.
\]
The weight of a generalised non-ambiguous tree $\N$ is given by
\[
\Phi(\N) :=
\prod_{i=1}^d
\frac{
x_i^{\wi_i(\N)}
}{
\wi_i(\N) !
}.
\]
Let $\operatorname{\mathbb{Q}}\NATdk$ be the $\operatorname{\mathbb{Q}}$-vector space generated by generalised non-ambiguous
trees of dimension $(d,k)$, and $\operatorname{\mathbb{Q}}\NATdkp$ its subspace generated by non-empty
generalised non-ambiguous trees and $\emptyset_\dir$. Let $\BNAT$ be the
multilinear map
\[
\BNAT:
\displaystyle\prod_{\dir\in\edir}\mathbb{Q}\NATdkp
\rightarrow
\mathbb{Q}\NATdk
\]
such that
\(
\BNAT((NAT_\dir)_{\dir\in\edir})
\)
is equal to the formal sum of all the generalised non-ambiguous trees such that
the root's child of direction $\pi$ is $NAT_\dir$.
To define $\Bxy$, we need the following notations
\[
\partial_\dir:=\partial_{i_1}\cdots\partial_{i_k}
,\hspace{.5cm}
\int_{\dir}:=
\int_0^{x_{i_1}}
\cdots
\int_0^{x_{i_k}}
\hspace{.5cm}
\text{and}
\hspace{.5cm}
\d x_\dir:=\d x_{i_1}\cdots\d x_{i_k}.
\]
with $\dir=\{i_1,\ldots,i_k\}$. We denote by $\Bxy$ the $\binom{d}{k}$-linear map
\[
\operatorname{\mathbb{Q}}\llbracket x_1,\ldots,x_d \rrbracket^{\edir}
\rightarrow
\operatorname{\mathbb{Q}}\llbracket x_1,\ldots,x_d \rrbracket
\]
such that
\[
\Bxy((f_\dir)_{\dir\in\edir}):=
\mathlarger{\int}_{\llbracket 1,d \rrbracket}
\left(
\prod_{\dir\in\edir}
\partial_{\llbracket 1,d \rrbracket\setminus\dir}(f_\dir)
\right)
\;\d x_{\llbracket 1,d \rrbracket}.
\]
Similarly to the dimension $(2,1)$, by recursion on $\binom{d}{k}$-ary trees, we prove that
\begin{equation}\label{eq:commutative_diag}
\Bxy(\Phi(NAT_\dir)_{\dir\in\edir})
=
\Phi(\BNAT((NAT_\dir)_{\dir\in\edir})).
\end{equation}
As for non-ambiguous trees (Proposition \ref{prop_hook_nat}), there is a hook
formula for the number of non-ambiguous trees of dimension $(d,k)$ with fixed
underlying tree. Let $\M$ be a $\binom{d}{k}$-ary tree, for each vertex $U$ we
denote by $\E_i(U)$ the number of vertices, of the subtree whose root is $U$
(itself included in the count), whose direction contains $i$. Then, by
recursively using the previous equation we obtain that
$$
|\NATdk(M)|
=
\frac{
\displaystyle
\prod_{i=1}^d \left( \wi_i( \M ) - 1 \right) !
}{
\displaystyle
\prod_{
\substack{
U: \text{ child of direction}\\ \text{ containing } i
}
}{
\E_i(U)
}
}
.
$$
Let $\LGFNdk(x_1,\ldots,x_d)$ be the exponential generating function of
non-empty generalised non-ambiguous trees
\[
\LGFNdk(x_1,\ldots,x_d)
:=
\sum_{\N \in \NATdk^*}\Phi(\N).
\]
and $\GFNdk(x_1,\ldots,x_d)$ its derivative
\[
\GFNdk(x_1,\ldots,x_d)
:=
\partial_{\{1,\ldots,d\}}\LGFNdk(x_1,\ldots,x_d).
\]
There is a $(d,k)$-dimensional analogue of the fixed point differential
Equation~\ref{equ_gfn}. Similarly to the notations at the beginning of the
section, we define $x_{\{i_1,\ldots,i_k\}}$ as the product $x_{i_1}\cdots
x_{i_k}$, in particular, $\Phi(\emptyset_\dir) = x_{\llbracket 1,d \rrbracket\setminus\dir}$.
\begin{prop}
\label{prop_equ_diff_natdk}
The exponential generating function $\LGFNdk$ of generalised non-ambiguous trees
satisfies the following differential equation
\begin{equation}
\label{equ_lgfndk}
\LGFNdk
=
\sum_{\dir\in\edir}x_{\llbracket 1,d \rrbracket\setminus\dir}
+
\int_{\llbracket 1,d \rrbracket}
\prod_{\dir\in\edir}
\partial_{\llbracket 1,d \rrbracket\setminus\dir}(\LGFNdk)
\;\d x_{\llbracket 1,d \rrbracket},\\
\end{equation}
and its derivative $\GFNdk$ satisfies
\begin{equation}
\label{equ_gfndk}
\GFNdk
=
\prod_{\dir \in \edir} \left( 1 + \int_{\dir} \GFNdk \right).
\end{equation}
\end{prop}
\begin{proof}
The summation of Equation~\ref{eq:commutative_diag} over
$\displaystyle\prod_{\dir\in\edir}\mathbb{Q}\NATdkp$ gives us
\[
\LGFNdk
=
\sum_{\dir\in\edir}x_{\llbracket 1,d \rrbracket\setminus\dir}
+
\int_{\llbracket 1,d \rrbracket}
\prod_{\dir\in\edir}
\partial_{\llbracket 1,d \rrbracket\setminus\dir}(\LGFNdk-\sum_{\dir'\in(\edir\setminus\dir)}x_{
\llbracket 1,d \rrbracket\setminus\dir'})
\;\d x_\dir.\\
\]
Since $\partial_{\llbracket 1,d \rrbracket\setminus\dir} (x_{\llbracket 1,d
\rrbracket\setminus\dir'})= 0$ for $\dir\neq\dir'$, the
Equation~\ref{equ_lgfndk} is proven. Equation~\ref{equ_gfndk} is obtained by
using the identity
\[
\GFNdk(x_1,\ldots,x_d)
=
\partial_{\{1,\ldots,d\}}\LGFNdk(x_1,\ldots,x_d).
\]
\end{proof}
In the generic case, we are not able to solve those differential equations.
We know that setting a variable $x_d$ to $0$ gives the generating function
of NATs of lower dimension.
\begin{prop}
Let $d > k \ge 1$, then
$
\left.
\GFN_{d,k}
\right|_{x_d=0}
=
\GFN_{d-1,k}
.
$
\end{prop}
For some specific values of $d$ and $k$ we have (at least partial) results.
\begin{prop}
\label{prop_sol_d_d-1}
Let $k=d-1$, if we know a particular solution $s( x_1, \ldots, x_d)$ for
$$
\partial_1 \ldots \partial_d \LGFNdk[d][d-1]
=
\partial_{1} \LGFNdk[d][d-1] \times \ldots \times \partial_{d} \LGFNdk[d][d-1]
$$
then, for any function $s_1(x_1), \ldots, s_d(x_d)$, the function
$s( s_1(x_1), \ldots, s_d(x_d))$ is also a solution.
\end{prop}
\begin{prop}
\label{prop_sol_d_1}
Some non trivial rational functions are solutions of
$$
\partial_1 \ldots \partial_d \LGFNdk[d][1]
=
\prod_{\dir \in \edir[d][d-1]}
\partial_{\dir} \LGFNdk[d][1]
.
$$
\end{prop}
\begin{proof}[Proof (sketch)]
We generalise the first part of the proof of
Proposition~\ref{prop:gf_expressions}.
We define
$
\LGFN_{(i)} = \partial_{\dir} \LGFNdk[d][1]
$
where
$i \in \i{1}{d}$
and
$\dir = \i{1}{d} \setminus \{i\}$.
We get the relation
$
\partial_i \LGFN_{(i)}
=
\prod_{j=1}^d \LGFN_{(j)}
$
and then
$
\prod_{i=1}^d
\partial_i \LGFN_{(i)}
=
\prod_{i=1}^d \LGFN_{(i)}^d
.
$
To obtain a particular solution, we just need to identify, in the previous
equation, the term $\partial_i \LGFN_{(i)}$ to the term $\LGFN_{(i)}^d.$
We thus obtain some non trivial solutions for our equation,
which are rational functions.
\end{proof}
Since dimension $(2,1)$ is the unique case where Proposition~\ref{prop_sol_d_d-1}
and Proposition~\ref{prop_sol_d_1} can be applied at the same time,
and the computation of $\GFNdk[\d][\d]$ is straightforward,
we have the following proposition.
\begin{prop}
We have the closed formulas:
$$
\GFNdk[2][1] = \GFN \qquad\text{and}\qquad
\GFNdk[d][d]=\sum_{n \ge 0} \frac{(x_1\cdot \ldots \cdot x_d)^n}{(n!)^d}.
$$
\end{prop}
We see $\GFNdk[d][d]$ as is a kind of generalised Bessel
function because
\[
\GFNdk[2][2](x/2, -x/2) = J_0(x)
\]
where $J_\alpha$ is the classical Bessel function. This supports our feeling
that the general case leads to serious difficulties.
\subsection{Geometric interpretation}
\label{geom_gnat}
As for non-ambiguous trees, we can give a geometric definition
of non-ambiguous trees of dimensions $(d,k)$ as follows.
We denote by $(e_1,\dots,e_d)$ the canonical basis of $\mathbb{R}^\d$ and
$(X_1,\ldots,X_\d)$ its dual basis, i.e. $X_i$ is $\mathbb{R}$-linear
and $X_i(e_i)=\delta_{i, j}$.
Let $P\in\mathbb{R}^\d$ and $\dir=\{i_1,\ldots,i_k\}$ a $(d,k)$-direction,
we call \emph{cone of origin $P$ and direction $\dir$} the set of points
$C(P, \dir):= \{P + a_1e_{i_1} + \cdots + a_ke_{i_k}\ |\ (a_1,\ldots,a_k)\in\mathbb{N}^k\}.$
\begin{defi}\label{def_gnat_geo}
A \emph{geometric non-ambiguous tree} of dimension $(\d,\k)$ and box $\gs$
is a non-empty set $\mathcal{V}$ of points of $\mathbb{N}^\d$ such that:
\begin{enumerate}[ref={\thecor.\arabic*}]
\item\label{gnat_geo_box} $\mathcal{V}$ is contained in $\i{1}{\wi_1}\times\cdots\times\i{1}{\wi_\d}$.
\item\label{gnat_geo_root} $\mathcal{V}$ contains the point $(\wi_1,\ldots,\wi_\d)$,
which is called the \emph{root},
\item\label{gnat_geo_cone} For $P\in \mathcal{V}$ different from the root, there exists
a unique $(d,k)$-direction $\dir=\{i_1,\ldots,i_k\}$ such that the cone
$c(P, \dir)$ contains at least one point different from $P$. We say that
$P$ is of type $\dir$.
\item\label{gnat_geo_distinct_and_interval} For each $i\in\i{1}{\d}$ and for each
$l\in\i{1}{\wi_i-1}$, the affine hyperplane $\{x_i=l\}$ contains exactly
one point of type $\dir$ that contains $i$.
\item\label{gnat_geo_affine} For $P$ and $P'$ two points of $\mathcal{V}$ belonging to a
same affine space of direction $\operatorname{Vect}(e_{i_1},\ldots,e_{i_k})$, then,
either $\forall j\in\i{1}{\k}\text{, }X_{i_j}(P)>X_{i_j}(P')$, or $\forall
j\in\i{1}{\k}\text{, }X_{i_j}(P')>X_{i_j}(P).$
\end{enumerate}
\end{defi}
Let us compare the original definition (\cite{AvaBouBouSil14}) of non-ambiguous
trees recalled in Section~\ref{definitions} with Definition~\ref{def_gnat_geo}. Both
have a condition for the existence of a root \ref{condition_1_ana} and
\ref{gnat_geo_root}. The existence and uniqueness of a parent for a non-root
point are given by conditions \ref{condition_2_ana} and \ref{gnat_geo_cone},
moreover when $k\geq2$ we also need the condition \ref{gnat_geo_affine}.
Finally, the compactness is given by conditions \ref{condition_3_ana},
\ref{gnat_geo_box} and \ref{gnat_geo_distinct_and_interval}.
\begin{prop}
There is a simple bijection between the set of geometric non-ambiguous trees of
box $\gs$ and the set of non-ambiguous trees of
geometric size $\gs$.
\end{prop}
An example of the correspondence is given in Figure~\ref{fig_gnat}.
\begin{proof}
If $k=d$, $\mathcal{V}$ is of the form
\[
\{(\wi,\dots,\wi),(\wi-1,\dots,\wi-1),\dots,(1,\dots,1)\},
\]
which corresponds exactly to non-ambiguous trees of dimension $(\d,\d)$ defined
in $\ref{def_gnat}$.
Let us now suppose that $k<d$.
\textbf{Definition \ref{def_gnat} implies Definition \ref{def_gnat_geo}:}
Let $\N$ be a non-ambiguous tree of dimension $(d,k)$ as defined in
Definition \ref{def_gnat} and let $\gs$ be its geometric size.
The first step is to define the coordinates, also called completed label, of a vertex $U$
by replacing the $\bullet$ by integers in all the labels of vertices.
Let $U$ be a vertex of $\N$ such that its $i$th component is a $\bullet$. We replace the $i$th component of $U$
with the $i$th component of the first ancestor of $U$ with a $i$th component
different from $\bullet$. Such an ancestor exists
since the root has no $\bullet$ component.
As a consequence, using Point \ref{gnat_growth} of Definition \ref{def_gnat}, we have
for a vertex $V$ of completed label $(v_1,\dots,v_\d)$ that
if $V$ has a child $U$ indexed by a $(\d,\k)$-direction $\dir$
and of completed label $(u_1,\dots,u_\d)$, then for $i\in\dir$,
$v_i>u_i$ and for $i\not\in\dir$, $v_i=u_i$.
Moreover, \ref{gnat_distinct} and the definition of completed labels implies
that, for $U$ and $V$ two vertices, if there exists $i$ such that $u_i=v_i$,
then they have a common ancestor $W$ such that $w_i\neq\bullet$ and
$w_i=u_i=v_i$. We denote by $\mathfrak{a}_i(U)$, or equivalently
$\mathfrak{a}_i(V)$, the vertex $W$.
Let $V$ be a vertex of $\N$ of completed label $(u_1,\dots,u_\d)$.
We denote by $P_V\in\mathbb{N}^\d$ the point $(u_1,\dots,u_\d)$.
Let $\mathcal{V}$ be the set of points $\{P_V\ |\ V\text{ vertex of }\N\}$.
By definition, each vertex has a different completed label.
Let us prove that $\mathcal{V}$ satisfies the conditions of \ref{def_gnat_geo}.
\begin{enumerate}
\item It is a consequence of \ref{gnat_interval}, and Definition~\ref{gnat_gs}.
\item $\mathcal{V}$ contains $(\wi_1,\dots,\wi_\d)$, since $(\wi_1,\dots,\wi_\d)$
is the label of the root of $\N$.
\item Let $P_U$ be a point of $\mathcal{V}$ different from $(\wi_1,\dots,\wi_\d)$. Since
$U$ is not the root, it is a child indexed by a $(\d,\k)$-direction $\dir$,
of a vertex $V$. Hence for $i\not\in\dir$, $X_i(P_V)=X_i(P_U)$ and for
$i\in\dir$, $X_i(P_V)>X_i(P_U)$. So $P_V$ is in the cone of origin $P_U$
and direction $\dir$, in particular, $P_U$ is of type $\dir$. Let us prove
the uniqueness by contradiction. Suppose there is another $(\d,\k)$-direction
$\dir'$ such that the cone of origin $P_U$ and direction $\dir'$ contains a
point $P_{V'}$ different from $P_U$. Let $i\in\dir\setminus\dir'$. Then, by
definition of the cone we have $X_i(P_U)=X_i(P_V')$. Since
$\mathfrak{a}_i(U)=V$, then $\mathfrak{a}_i(V')=V$. Let
$i'\in\dir'\setminus\dir$ then $X_i'(P_{V'})>X_i(P_U)=X_i(P_V)$, which is not
possible since $V$ is an ancestor of $V'$. In particular the type of $P_U$
corresponds to the type of $U$.
\item Let $i\in\llbracket1,d\rrbracket$ and $l\in\llbracket1,\wi-1\rrbracket$,
by \ref{gnat_interval}, there exists $U$ such that $u_i\neq\bullet$ and
$u_i=l$. Let $\dir$ be the index of $u$ then $i\in\dir$, hence, $P_U$
satisfies \ref{gnat_geo_distinct_and_interval}. Suppose there exists
another point $P_V$ satisfying \ref{gnat_geo_distinct_and_interval}, let
$\dir'$ be its type. Then $V$ is indexed by $\dir'$ and $i\in\dir'$. Thus, $V$
is another vertex such that $v_i\neq\bullet$ and $v_i=l$ which is in
contradiction with \ref{gnat_distinct}.
\item Let $\mathcal{F}$ be an affine space of direction $\operatorname{Vect}(e_{i_1},\ldots,e_{i_k})$
containing a point $P_U$. We denote by $\dir$ the set
$\{e_{i_1},\ldots,e_{i_k}\}$. If $U$ is the root, then
\ref{gnat_geo_affine} is satisfied. Else, let $(V_0,V_1,\dots,V_m)$ be the
sequence of ancestors of $U$, i.e $V_0=U$ for all $j$, $V_j$ is a child of
$V_{j+1}$ and $V_m$ is the root of $\N$. Let $l$ be the index such that
$V_{l}$ is a child of $V_{l+1}$ not indexed by $\dir$ and $\forall
j\in\i{0}{l-1}$, $V_j$ is the child indexed by $\dir$ of $\mathcal{V}_{j+1}$, $V_l$
will be denoted $V$. Let $\dir'$ be the $(d,k)$-direction indexing $V$ and
let $i\in\dir'\setminus\dir$. If $V$ is the root then $\dir'=\llbracket 1,d
\rrbracket$. Let $P_{U'}$ be another point of $\mathcal{F}$, since
$i\not\in\dir$, then $u'_i=u_i$. Hence, we have
$\mathfrak{a}_i(U')=\mathfrak{a}_i(U)=V$.
Since for all $j\not\in\dir$, $u'_j=v_j$, the path from $U'$ to $V$ contains
only vertices indexed by $\dir$. Hence, by definition of $V$, $U$ is an
ancestor of $U'$ or the converse, which proves \ref{gnat_geo_affine}.
\end{enumerate}
\textbf{Definition \ref{def_gnat_geo} implies Definition \ref{def_gnat}:}
Let $\mathcal{V}$ be a non-ambiguous tree defined with Definition \ref{def_gnat_geo}. We
start by constructing the underlying $\binom{d}{k}$-ary tree $\M$ of $\mathcal{V}$. The
vertices of $\M$ correspond to the points of $\mathcal{V}$. In particular, the root of
$\M$ corresponds to the root of $\mathcal{V}$. Let $P$ be a point of $\mathcal{V}$, we denote by
$V_P$ the corresponding vertex of $\M$. Let $P$ be a point of $\mathcal{V}$ different
from the root, let $\dir=\{e_{i_1},\ldots,e_{i_k}\}$ be the $(d,k)$-direction
defined by \ref{gnat_geo_cone}. Using \ref{gnat_geo_affine}, we can define
without ambiguity the parent of $V_P$ as the vertex $V_{P'}$ such that $P'$ is the
closest point to $P$ belonging to $c(P,\dir)$. $V_P$ is the child of $V_{P'}$
indexed by $\dir$, moreover, for all $i\in\dir$, $X_i(P)<X_i(P')$ and for all
$i\not\in\dir$, $X_i(P)=X_i(P')$. The labelling is done as follows. For each
point $P$ of $\mathcal{V}$ we label the vertex $V_P$ with the coordinates of $P$. Then, given a vertex
$V$ of type $\dir$, for all $i\not\in\dir$, we replace the $i$th
component of its label with $\bullet$. Thus, if the $i$th component of a vertex
$V_{P'}$ is equal to $l$, then for each descendant $V_P$ of $V_{P'}$ we have
$X_i(P)\leqslant l$ and if $X_i(P)=l$ then the $i$th component of $V_{P'}$
is $\bullet$. Let us prove that the conditions of Definition~\ref{def_gnat} are
satisfied.
\begin{enumerate}
\item By construction of the labels.
\item Let $P$ and $P''$ be two points such that $V_P$ is a descendant of
$V_{P''}$ indexed by $\dir$ such that $i$ belongs to $\dir$. Let $P'$ be
the father of $P$, then $X_i(P)<X_i(P')\leqslant X_i(P'')$.
\item Let $V_P$ be a vertex of $\M$ such that its $i$th component is different
from $\bullet$. If $V_P$ is the root, then all the vertices of $\M$ are its
descendants, hence its $i$th label appears only once. Else, $V_P$ is a
child indexed by $\dir$ of a vertex $V_{P'}$. In particular, $\dir$
contains $i$ since the $i$th component of $V_P$ is not $\bullet$. Hence, by
\ref{gnat_geo_distinct_and_interval}, the $i$th component of $V_P$ is
unique.
\item \ref{gnat_geo_distinct_and_interval} implies that for each
$i\in\i{1}{\d}$, for all $l\in\i{1}{\wi_i-1}$ there is a point $P$ of type
$\dir$ such that $X_i(P)=l$ and $i\in\dir$, so that the $i$th component of
$V_P$ is equal to $l$. Moreover, the coordinates of the root are
$(\wi_1,\dots,\wi_\d)$ and $\mathcal{V}$ is contained in the box
$\i{1}{\wi_1}\times\cdots\times\i{1}{\wi_\d}$. Therefore, the set of $i$th
components, different from $\bullet$, is the interval $\i{1}{\wi_i}$.
\end{enumerate}
\end{proof}
\section{A new statistic on binary trees: the hook statistic}\label{hook}
We present in this section a bijection between binary trees and ordered trees, sending the vertices to edges and the hook statistic defined in Definition \ref{def_hook} to the number of vertices having at least a child which is a leaf, what we will call the \emph{child-leaf statistic}.
The corresponding integer series appears as \cite[A127157]{oeis} in OEIS.
We denote by $\Bp$ and $\Op$ respectively the exponential generating series of
these trees, with these statistics, the variable $x$ indexing the number of
vertices in $\Bp$ and the number of edges in $\Op$, and $t$ the statistic.
Then, these generating series satisfy:
\begin{prop}
The generating series of binary trees with hook statistic and ordered trees with the child-leaf statistic are given by the following functional equations:
\begin{equation*}
\Bp= 1+xt\times\left(\frac{1}{1-x\Bp} \right)^2
\end{equation*}
\begin{equation*}
\Op=\frac{1}{1-x(\Op-1))}\times\left(1+xt\times\frac{1}{1-x\Op} \right)
\end{equation*}
These generating series are equal.
\end{prop}
\begin{proof}
The first functional equation is obtained by considering the vertices in the hook of the root: there can be none or there is a root, a list of left descendant (whose right child is a binary tree) and a list of right descendant (whose left child is a binary tree).
The second functional equation is obtained by considering, if the ordered tree
is not reduced to a vertex, the first leaf of the root from left to right, if
it exists. Then, on the left side of this leaf, there is a list of
ordered trees not reduced to a vertex and on the right side a list of ordered trees, if there is a leaf.
Then, by multiplying the preceding equations by $1-x\Bp$ and $1-x(\Op-1))$ respectively, they are equivalent to:
\begin{equation*}
\Bp -x\Bp^2= 1-x\Bp + xt \frac{1}{1-x\Bp}
\end{equation*}
\begin{equation*}
\Op -x \Op^2 +x \Op = 1 +xt\frac{1}{1-x\Op}.
\end{equation*}
\end{proof}
Let us now exhibit a bijection between these two objects. This bijection comes from the following equation:
\begin{equation*}
(\Bp-1) -x (\Bp-1)-x (\Bp-1)^2 = xt \frac{1}{1-x\Bp}.
\end{equation*}
This equation can be viewed as considering only binary trees whose root has no left descendants or ordered trees such that the leftmost child of the root is a leaf.
We obtain the following bijection:
\begin{prop}
The map $\zeta$ sends a binary tree $\B$ to an ordered tree $\O$ by mapping:
\begin{itemize}
\item the leftmost descendant of the root, if it is a leaf, to an edge between the root and its only child
\item the leftmost descendant of the root $v$ to an edge between the root of the tree associated with the descendants of $v$ and the root of the tree obtained from what is left
\item the set of right descendants of the root to the set of children of the root.
\end{itemize}
It is a bijection between binary trees and ordered trees, sending the vertices to edges and the hook statistic to the child-leaf statistic.
\end{prop}
The bijection is described in Figure~\ref{zeta} and the explicit correspondance
for the sizes one, two and three is given in Table~\ref{tableau_zeta} the first
terms in the bijection. Another way of describing recursively the bijection
$\zeta$ is given in Figure~\ref{zeta2}, the empty binary tree is still send to
the ordered tree reduced to one vertex.
\begin{figure}
\scalebox{0.75}{
\begin{tikzpicture}
\draw (-0.5,0) -- (0,0.75) -- (0.5,0) -- cycle;
\draw[fill] (0,0.75) circle(0.1);
\draw (-0.5,0)--(-1,-0.75);
\draw[fill] (-1,-0.75) circle(0.1);
\draw (1,0) node{$\Leftrightarrow$};
\draw (1,-0.75) -- (1.5,0) -- (2,-0.75) -- cycle;
\draw[fill] (1.5,0) circle(0.1);
\draw (1.5,0)--(1.5,0.75);
\draw[fill] (1.5,0.75) circle(0.1);
\end{tikzpicture} }
\hspace{1cm}
\scalebox{0.75}{
\begin{tikzpicture}
\draw (-0.5,0) -- (0,0.75) -- (0.5,0) -- cycle;
\draw (0,0.25) node{$1$};
\draw[fill] (0,0.75) circle(0.1);
\draw (-0.5,0)--(-1,-0.75);
\draw[fill] (-1,-0.75) circle(0.1);
\draw (-1,-2.25) -- (-0.5,-1.5) -- (0,-2.25) -- cycle;
\draw (-0.5,-2) node{$2$};
\draw[fill] (-0.5,-1.5) circle(0.1);
\draw (-0.5,-1.5)--(-1,-0.75);
\draw (1,-0.75) node{$\Leftrightarrow$};
\draw (2,-1) -- (2.5,-0.25) -- (3,-1) -- cycle;
\draw (2.5,-0.75) node{$2$};
\draw[fill] (2.5,-0.25) circle(0.1);
\draw (2.5,-0.25)--(1.5,-0.75);
\draw[fill] (1.5,-0.75) circle(0.1);
\draw (1,-1.5) -- (1.5,-0.75) -- (2,-1.5) -- cycle;
\draw (1.5,-1.25) node{$1$};
\end{tikzpicture} }
\hspace{1cm}
\scalebox{0.75}{
\begin{tikzpicture}
\draw (0,2.5)--(2.25,1);
\draw[fill] (0,2.5) circle(0.1);
\draw[fill] (0.75,2) circle(0.1);
\draw[fill] (1.5,1.5) circle(0.1);
\draw[fill] (9/4,1) circle(0.1);
\draw (0.75,1.5) node{$1$};
\draw (1.5,1) node{$2$};
\draw (9/4,0.5) node{$3$};
\draw (0.30,1.25) -- (0.75,2) -- (1.20,1.25) -- cycle;
\draw (1.05,0.75) -- (1.5,1.5) -- (1.95,0.75) -- cycle;
\draw (1.80,0.25) -- (2.25,1) -- (2.70,0.25) -- cycle;
\draw (3,1.25) node{$\Leftrightarrow$};
\draw[fill] (4,2) circle(0.1);
\draw[fill] (3.5,1) circle(0.1);
\draw[fill] (4.5,1) circle(0.1);
\draw[fill] (5.5,1) circle(0.1);
\draw[fill] (6.5,1) circle(0.1);
\draw (4.5,0.5) node{$1$};
\draw (5.5,0.5) node{$2$};
\draw (6.5,0.5) node{$3$};
\draw (4,2)--(3.5,1);
\draw (4,2)--(4.5,1);
\draw (4,2)--(5.5,1);
\draw (4,2)--(6.5,1);
\draw (4.05,0.25) -- (4.5,1) -- (4.95,0.25) -- cycle;
\draw (5.05,0.25) -- (5.5,1) -- (5.95,0.25) -- cycle;
\draw (6.05,0.25) -- (6.5,1) -- (6.95,0.25) -- cycle;
\end{tikzpicture}}
\caption{Bijection $\zeta$}\label{zeta}
\end{figure}
\begin{table}
\begin{tabular}{|c|cccccccc|}
\hline
binary trees &
\begin{tikzpicture}
\node[blue] (r) at (0,0) {$\bullet$};
\end{tikzpicture}&
\begin{tikzpicture}
\node[blue] (r) at (0,0) {$\bullet$};
\node[blue](fg) at (-0.5,-0.75){$\bullet$};
\draw[blue] (r) --(fg);
\end{tikzpicture}&
\begin{tikzpicture}
\node[blue] (r) at (0,0) {$\bullet$};
\node[blue](fd) at (0.5,-0.75){$\bullet$};
\draw[blue] (r) --(fd);
\end{tikzpicture}&
\begin{tikzpicture}
\node[blue] (r) at (0,0) {$\bullet$};
\node[blue](fg) at (-0.5,-0.75){$\bullet$};
\node[blue](fd) at (0.5,-0.75){$\bullet$};
\draw[blue] (fd)--(r) --(fg);
\end{tikzpicture} &
\begin{tikzpicture}
\node[blue] (r) at (0,0) {$\bullet$};
\node[blue](fg) at (-0.5,-0.75){$\bullet$};
\node[blue](fgg) at (-1,-2*0.75){$\bullet$};
\draw[blue] (r) --(fg)--(fgg);
\end{tikzpicture}&
\begin{tikzpicture}
\node[blue] (r) at (0,0) {$\bullet$};
\node[blue](fd) at (0.5,-0.75){$\bullet$};
\node[blue](fdd) at (1,-2*0.75){$\bullet$};
\draw[blue] (r) --(fd)--(fdd);
\end{tikzpicture}&
\begin{tikzpicture}
\node[blue] (r) at (0,0) {$\bullet$};
\node[blue](fg) at (-0.5,-0.75){$\bullet$};
\node[purple](fgg) at (0,-2*0.75){$\bullet$};
\draw[blue] (r) --(fg);
\draw (fg)--(fgg);
\end{tikzpicture}&
\begin{tikzpicture}
\node[blue] (r) at (0,0) {$\bullet$};
\node[blue](fd) at (0.5,-0.75){$\bullet$};
\node[purple](fdd) at (0,-2*0.75){$\bullet$};
\draw[blue] (r) --(fd);
\draw (fd)--(fdd);
\end{tikzpicture}
\\ \hline
ordered trees &
\begin{tikzpicture}
\node[blue](r) at (0,0){$\bullet$};
\node(f) at (0,-0.75){$\bullet$};
\draw (r)--(f);
\end{tikzpicture}&
\begin{tikzpicture}
\node(r) at (0,0){$\bullet$};
\node[blue](f) at (0,-0.75){$\bullet$};
\node(ff) at (0,-2*0.75){$\bullet$};
\draw (r)--(f)--(ff);
\end{tikzpicture}&
\begin{tikzpicture}
\node[blue] (r) at (0,0) {$\bullet$};
\node(fg) at (-0.5,-0.75){$\bullet$};
\node(fd) at (0.5,-0.75){$\bullet$};
\draw (fd)--(r) --(fg);
\end{tikzpicture} &
\begin{tikzpicture}
\node(rr) at (0,0.75){$\bullet$};
\node[blue] (r) at (0,0) {$\bullet$};
\node(fg) at (-0.5,-0.75){$\bullet$};
\node(fd) at (0.5,-0.75){$\bullet$};
\draw (fd)--(r) --(fg);
\draw (r)--(rr);
\end{tikzpicture} &
\begin{tikzpicture}
\node(r) at (0,0){$\bullet$};
\node(f) at (0,-0.75){$\bullet$};
\node[blue](ff) at (0,-2*0.75){$\bullet$};
\node(fff) at (0,-3*0.75){$\bullet$};
\draw (r)--(f)--(ff)--(fff);
\end{tikzpicture}&
\begin{tikzpicture}
\node[blue] (r) at (0,0) {$\bullet$};
\node(fg) at (-0.5,-0.75){$\bullet$};
\node(fd) at (0.5,-0.75){$\bullet$};
\node(fm) at (0,-0.75){$\bullet$};
\draw (fd)--(r) --(fg);
\draw (r)--(fm);
\end{tikzpicture} &
\begin{tikzpicture}
\node[blue] (r) at (0,0) {$\bullet$};
\node(fg) at (-0.5,-0.75){$\bullet$};
\node[purple](fd) at (0.5,-0.75){$\bullet$};
\node(f) at (0.5,-2*0.75){$\bullet$};
\draw (f)--(fd)--(r) --(fg);
\end{tikzpicture} &
\begin{tikzpicture}
\node[blue] (r) at (0,0) {$\bullet$};
\node[purple] (fg) at (-0.5,-0.75){$\bullet$};
\node(f) at (-0.5,-2*0.75){$\bullet$};
\node(fd) at (0.5,-0.75){$\bullet$};
\draw (fd)--(r) --(fg)--(f);
\end{tikzpicture}
\\ \hline
\end{tabular}
\caption{First terms of the bijection $\zeta$}\label{tableau_zeta}
\end{table}
\begin{figure}
\scalebox{1}{
\begin{tikzpicture}
\draw[blue] (0,2.5)--++ (-3,-1.5);
\draw[blue] (0,2.5)--++ (3,-1.5);
\draw (-3,1) --++ (.5,-.75);
\draw (-2,1.5) --++ (.5,-.75);
\draw (-1,2) --++ (.5,-.75);
\draw (1,2) --++ (-.5,-.75);
\draw (2,1.5) --++ (-.5,-.75);
\draw (3,1) --++ (-.5,-.75);
\draw[fill,blue] (-3,1) circle(0.1);
\draw[fill,blue] (-2,1.5) circle(0.1);
\draw[fill,blue] (-1,2) circle(0.1);
\draw[fill,blue] (0,2.5) circle(0.1);
\draw[fill,blue] (1,2) circle(0.1);
\draw[fill,blue] (2,1.5) circle(0.1);
\draw[fill,blue] (3,1) circle(0.1);
\draw (-2.5,-.25) node{$a$};
\draw (-1.5,.25) node{$b$};
\draw (-.5,.75) node{$c$};
\draw (.5,.75) node{$1$};
\draw (1.5,.25) node{$2$};
\draw (2.5,-.25) node{$3$};
\draw (-2.5,.25)--++ (-.45,-.75) --++ (.9,0) -- cycle;
\draw (-1.5,.75)--++ (-.45,-.75) --++ (.9,0) -- cycle;
\draw (-.5,1.25)--++ (-.45,-.75) --++ (.9,0) -- cycle;
\draw (.5,1.25)--++ (-.45,-.75) --++ (.9,0) -- cycle;
\draw (1.5,0.75)--++ (-.45,-.75) --++ (.9,0) -- cycle;
\draw (2.5,0.25)--++ (-.45,-.75) --++ (.9,0) -- cycle;
\draw (4.5,1.25) node{$\Longleftrightarrow$};
\end{tikzpicture}
\hspace{5mm}
\begin{tikzpicture}
\draw (4,2)--++(2.25,2.25);
\draw (6.25,3.75) node{\Small$\zeta(a)$};
\draw (5.5,3) node{\Small$\zeta(b)$};
\draw (4.75,2.25) node{\Small$\zeta(c)$};
\draw (4.5,0.5) node{\Small$\zeta(1)$};
\draw (5.5,0.5) node{\Small$\zeta(2)$};
\draw (6.5,0.5) node{\Small$\zeta(3)$};
\draw (6.25,4.25) --++ (-.45,-.75) --++ (.9,0) -- cycle;
\draw (5.5,3.5) --++ (-.45,-.75) --++ (.9,0) -- cycle;
\draw (4.75,2.75) --++ (-.45,-.75) --++ (.9,0) -- cycle;
\draw (4,2)--(3.5,1);
\draw (4,2)--(4.5,1);
\draw (4,2)--(5.5,1);
\draw (4,2)--(6.5,1);
\draw[fill] (6.25,4.25) circle(0.1);
\draw[fill] (5.5,3.5) circle(0.1);
\draw[fill] (4.75,2.75) circle(0.1);
\draw[fill,blue] (4,2) circle(0.1);
\draw[fill] (3.5,1) circle(0.1);
\draw[fill] (4.5,1) circle(0.1);
\draw[fill] (5.5,1) circle(0.1);
\draw[fill] (6.5,1) circle(0.1);
\draw (4.05,0.25) -- (4.5,1) -- (4.95,0.25) -- cycle;
\draw (5.05,0.25) -- (5.5,1) -- (5.95,0.25) -- cycle;
\draw (6.05,0.25) -- (6.5,1) -- (6.95,0.25) -- cycle;
\end{tikzpicture}
}
\caption{Bijection $\zeta$, alternative description.}\label{zeta2}
\end{figure}
\section*{Perspectives}
In this work, we give new results about NATs, and we generalise the
definition of NATs to higher dimensions with a choice on the dimension of the
edges. It gives rise to several questions that we now detail.
We exhibit nice formulas for the generating function and the generating polynomial
of non-ambiguous trees that take into account the number of points in the first
column and in the first row. Those two parameters correspond to the parameters
$\alpha$ and $\beta$ of the PASEP. These formulas have been obtained with two different
technics: by solving a differential equation and by decomposing non-ambiguous
trees into hooks. In the context of the PASEP, it raises natural questions. Is it
possible to introduce the parameter $q$ of the PASEP in either one of them ? What
can we deduce from the hook decomposition of tree-like tableaux ?
We give a polynomial analogue for the hook formula enumerating the NATs with a
fixed binary tree, by adapting the methods of \cite{HivNovThib08}. Since there is a hook formula also for NATs of higher dimension, we could extend the work
to higher dimensions to get polynomial analogue.
The generating functions of generalised NATs satisfy differential equations
similar to the case of NATs. While we give a solution for the case of NATs and
get a nice closed form, we have not been able to tackle the general case. It
would be interesting to find a generic way for solving this type of
differential equations.
By generalising the NATs to higher dimension, we answer a question raised
in the perspectives of \cite{AvaBouBouSil14}. In this last paper, the authors also study NATs with a complete underlying binary tree, and
obtain nice combinatorial identities. It would be interesting to see if
the same happens in higher dimension.
Finally, as mentioned in the introduction, NATs correspond to tree-like
tableaux of rectangular shape. The question of the generalisation of tree-like tableaux to
higher dimension is raised by our work and we hope we will obtain again an insertion algorithm, which is a key
property of tree-like tableaux.
\ \\
{\bf Acknowledgement.}
The authors thank Samanta Socci for fruitful discussions which were
the starting point of the generalisation of non-ambiguous trees.
This research was driven by computer exploration using the open-source software
\texttt{Sage}~\cite{sage} and its algebraic combinatorics features developed by
the \texttt{Sage-Combinat} community~\cite{Sage-Combinat}.
The authors also thank the referees for their helpful suggestions.
\nocite{*}
\bibliographystyle{alpha}
|
\section{Introduction}
Collisions of protons with nuclei offer opportunities to study the production and interaction of heavy quarks inside the nucleus. Charm-quark production in hadron collisions is sensitive to the gluon content of colliding hadrons, and can be used to probe modifications of the parton distributions inside the nucleus \cite{Vogt:2004dh}. While traversing the nucleus, heavy quarks are also subject to energy loss that can lead to the suppression of bound states \cite{Arleo:2012rs}. Once the heavy-quark pair exits the nucleus, late-stage interactions with co-moving hadrons can disrupt fully formed quarkonium states \cite{Capella:1996va}. Measurements in proton-nucleus collisions also give an experimental baseline for the interpretation of quarkonium suppression in nucleus-nucleus collisions, where color screening in a deconfined quark-gluon plasma is expected to be a dominant effect \cite{Matsui:1986dk}.
Studies of quarkonium suppression in {\ensuremath{p}}Pb\xspace collisions revealed that
the excited states, such as the charmonium {\ensuremath{\Ppsi{(2S)}}}\xspace state or the bottomonium $\Y2S$ and $\Y3S$ states, show a different suppression pattern compared to the {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace and $\PUpsilon(1{\rm S})$ states (see \cite{LHCB-PAPER-2017-014,LHCb-PAPER-2015-058,Adare:2016psx,LHCb-PAPER-2018-035,Acharya:2020wwy,Aaboud:2017cif,Sirunyan:2018pse} and references therein). Such a difference cannot be explained by processes taking place during the initial stages of the collision, \mbox{\itshape i.e.}\xspace acting on the quark--antiquark pair. Instead, the processes must occur after the hadronization of the heavy-quark pair into a final state, \mbox{\itshape e.g.}\xspace through dissociation due to interactions with the co-moving matter created at the collision point \cite{Ferreiro:2014bia,Ferreiro:2018wbd}. Currently, the ${\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace$ and {\ensuremath{\Ppsi{(2S)}}}\xspace mesons are the only charmonium states which have been measured in collisions of protons with nuclei at the Large Hadron Collider (LHC).
The {\ensuremath{\Pchi_{\cquark J}}}\xspace states, with $J=0,1,2$ denoting the total angular momentum, comprise a triplet of orbitally excited 1P charmonia. They are typically studied in collider experiments via their radiative decay \decay{{\ensuremath{\Pchi_{\cquark J}}}\xspace}{{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace\Pgamma}, with a subsequent decay \decay{{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace}{\ensuremath{\ell^+ \ell^-}\xspace}, where $\ell$ denotes electron or muon.
A selection of recent measurements in \ensuremath{pp}\xspace and \ensuremath{p{\overline p}}\xspace collisions can be found in Refs. \cite{Abulencia:2007bra,LHCB-PAPER-2011-019,LHCB-PAPER-2013-028,Chatrchyan:2012ub,Adare:2011vq}.
The binding energies of {\ensuremath{\Pchi_{\cquark J}}}\xspace states are significantly smaller than that of the {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace state and greater than the binding energy of {\ensuremath{\Ppsi{(2S)}}}\xspace state \cite{Satz:2005hx}. The small difference in the binding energies of {\ensuremath{\Pchi_{\cquark 1}}}\xspace and {\ensuremath{\Pchi_{\cquark 2}}}\xspace charmonia makes the ratio of their production cross-sections, $\sigma({\ensuremath{\Pchi_{\cquark 2}}}\xspace) / \sigma({\ensuremath{\Pchi_{\cquark 1}}}\xspace)$, a useful tool to study their sensitivity to final-state nuclear effects, which are expected to be similar for both states. The ${\ensuremath{\Pchi_{\cquark J}}}\xspace$ states also form an important feed-down contribution to ${\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace$ production, so measurements of nuclear effects on ${\ensuremath{\Pchi_{\cquark J}}}\xspace$ states can clarify interpretation of the ${\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace$ data. Moreover, various efficiency factors and sources of uncertainty cancel out in the ratio, allowing for a more precise measurement.
In nuclear collisions, the
{\ensuremath{\Pchi_{\cquark J}}}\xspace states have been measured by the HERA-B \cite{Abt:2008ed} and PHENIX collaborations \cite{Adare:2013ezl}. To date, no measurement has been reported at the LHC energies.
Here we present the first measurement of the cross-section ratio of promptly produced ${\ensuremath{\Pchi_{\cquark 2}}}\xspace$ and ${\ensuremath{\Pchi_{\cquark 1}}}\xspace$ states, $\sigma({\ensuremath{\Pchi_{\cquark 2}}}\xspace) / \sigma({\ensuremath{\Pchi_{\cquark 1}}}\xspace)$, in nuclear collisions at the LHC.
The measurement is performed using data collected by the LHCb collaboration in {\ensuremath{p}}Pb\xspace collisions, at the center-of-mass energy per nucleon pair \ensuremath{\protect\sqrt{s_{\scriptscriptstyle\text{NN}}}}\xspace~=~8.16~TeV, in 2016.
\section{Experimental apparatus}
The \mbox{LHCb}\xspace detector~\cite{LHCb-DP-2008-001,LHCb-DP-2014-002} is a single-arm forward spectrometer covering the pseudorapidity range $2 < \eta < 5$, designed for the study of particles containing {\ensuremath{\Pb}}\xspace or {\ensuremath{\Pc}}\xspace quarks. The detector consists of a high-precision silicon-strip vertex locator (VELO) surrounding the interaction region, a set of four planar tracking stations coupled to a dipole magnet with a 4 Tm\xspace bending power, a pair of ring-imaging Cherenkov detectors to discriminate between different types of charged hadrons, followed by calorimetric and muon systems that are of particular importance in this measurement.
The calorimetric system allows for identification of electrons and photons and consists of a scintillating pad detector (SPD), a pre-shower system (PS), an electromagnetic (ECAL) calorimeter, and a hadronic (HCAL) calorimeter.
The SPD and PS are designed to discriminate between signals from photons and electrons, while ECAL and HCAL provide the energy measurement and identify electromagnetic radiation and neutral hadrons.
Muons are identified by a system composed of alternating layers of iron and multiwire proportional chambers.
The {\ensuremath{p}}Pb\xspace data were collected with the LHCb experiment in two distinct beam configurations.
In the forward configuration, the particles produced in the direction of the proton beam are measured in a center-of-mass rapidity region $1.5 < y^* < 4.0$, while in the backward configuration, particles produced in the lead-beam direction are measured at center-of-mass rapidity $-5.0 < y^* < -2.5$. The forward (backward) data sample corresponds to an integrated luminosity of about $ 14 \ensuremath{\,\upmu\nospaceunit{b}}\xspace^{-1}$ ($21 \ensuremath{\,\upmu\nospaceunit{b}}\xspace^{-1}$).
\section{Data selection}
The analyzed events are selected by a set of triggers designed to record collisions containing the decay \decay{{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace}{{\ensuremath{\Pmu^+\Pmu^-}}\xspace}.
The {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace candidates are reconstructed from a pair of oppositely charged muons with momentum component transverse to the beam, \ensuremath{p_{\mathrm{T}}}\xspace, larger than $700 \mbox{ \ensuremath{\aunit{Me\kern -0.1em V\!/}c}\xspace}$, originating from a common vertex and an invariant mass within $\pm 42 ~ \ensuremath{\aunit{Me\kern -0.1em V\!/}c^2}\xspace$ of the known {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace mass \cite{PDG2020} (corresponding to three times the dimuon mass resolution). The {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace candidates are combined with a photon candidate to form a {\ensuremath{\Pchi_{\cquark 1,2}}}\xspace candidate.
Photons used in this analysis are classified in two mutually exclusive types: those that converted in the detector material upstream of the dipole magnet and of which the electron and positron tracks were reconstructed in the tracking system (\textit{converted photons}), or those reconstructed through their energy deposits in the calorimetric system (\textit{calorimetric photons}).
The calorimetric photon sample is about an order of magnitude larger than the converted photon sample but has worse mass resolution.
Converted photons are reconstructed from a pair of oppositely
charged electron candidates and are required to have a transverse momentum $\ensuremath{p_{\mathrm{T}}}\xspace$~$> 600$~\ensuremath{\aunit{Me\kern -0.1em V\!/}c}\xspace and a good-quality conversion vertex \decay{\Pgamma}{{\ensuremath{\Pe^+\Pe^-}}\xspace}.
Calorimetric photons are identified
using the ratio of their energy deposited in the hadronic and electromagnetic calorimeters
and a pair of likelihood-based classifiers that discriminate photons from electrons and hadrons \cite{LHCb-DP-2018-001,Beteta:2020gui}.
Calorimetric photons accepted for analysis are required to have $\ensuremath{p_{\mathrm{T}}}\xspace$~$> 1$~\ensuremath{\aunit{Ge\kern -0.1em V\!/}c}\xspace.
The two measurements discussed here are independent given the different reconstruction between the converted and the calorimetric photons.
The selected {\ensuremath{\Pmu^+\Pmu^-\Pgamma}}\xspace combinations, which comprise the {\ensuremath{\Pchi_{\cquark 1,2}}}\xspace candidates, are required to be reconstructed within the pseudorapidity window $2 < \eta < 4.5$ and in the transverse momentum range of $3 < \ensuremath{p_{\mathrm{T}}}\xspace < 15$ \ensuremath{\aunit{Ge\kern -0.1em V\!/}c}\xspace for the converted and $5 < \ensuremath{p_{\mathrm{T}}}\xspace < 15$ \ensuremath{\aunit{Ge\kern -0.1em V\!/}c}\xspace for the calorimetric candidates.
In order to select the {\ensuremath{\Pchi_{\cquark 1,2}}}\xspace candidates produced \textit{promptly} at the primary-collision vertex and to suppress \textit{nonprompt} production from $b$-hadron decays occurring away from the primary vertex, an upper limit is imposed on the pseudo-decay time of the candidates, defined as
\begin{equation}
t_z = \frac{ \left( z_{\rm decay} - z_{\rm PV} \right) \times M_{{\ensuremath{\Pchi_{\cquark 1}}}\xspace} }{p_z},
\label{eq:pseudo_propertime}
\end{equation}
where $z_{\rm decay} - z_{\rm PV}$ is the difference between the positions of the reconstructed vertex of the {\ensuremath{\Pchi_{\cquark 1,2}}}\xspace candidate and the primary proton-nucleus collision vertex along the beam axis, $p_z$ is the longitudinal component of the {\ensuremath{\Pchi_{\cquark 1,2}}}\xspace candidate momentum and $M_{{\ensuremath{\Pchi_{\cquark 1}}}\xspace}$ is the known mass of the {\ensuremath{\Pchi_{\cquark 1}}}\xspace meson \cite{PDG2020}.
The pseudo-decay time is limited to $t_z < 0.1 \text{ ps}$.
The {\ensuremath{\Pchi_{\cquark 1}}}\xspace and {\ensuremath{\Pchi_{\cquark 2}}}\xspace candidates originating from decays of short-lived resonances, such as {\ensuremath{\Ppsi{(2S)}}}\xspace produced at the interaction point, are also considered in the analysis.
The effects of the detector acceptance as well as of the reconstruction and selection efficiencies are investigated with simulated events. The {\ensuremath{\Pchi_{\cquark 1,2}}}\xspace signal is generated in \mbox{\textsc{Pythia}}\xspace~\cite{Sjostrand:2007gs} with an \mbox{LHCb}\xspace specific configuration \cite{LHCb-PROC-2010-056}. The {\ensuremath{\Pchi_{\cquark 1}}}\xspace and {\ensuremath{\Pchi_{\cquark 2}}}\xspace states are generated assuming unpolarized production.
The underlying minimum bias forward and backward {\ensuremath{p}}Pb\xspace collisions are generated using the \mbox{\textsc{Epos}}\xspace event generator configured for the LHC \cite{Pierog:2013ria}.
Unstable particles are decayed via \mbox{\textsc{EvtGen}}\xspace \cite{Lange:2001uf}. The \decay{{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace}{{\ensuremath{\Pmu^+\Pmu^-}}\xspace} decays are corrected for final-state electromagnetic radiation using \mbox{\textsc{Photos}}\xspace~\cite{davidson2015photos}.
The response of the detector to the interactions of the generated particles is implemented using the \mbox{\textsc{Geant4}}\xspace toolkit~\cite{Allison:2006ve,*Agostinelli:2002hh}; for a detailed description see Ref.~\cite{LHCb-PROC-2011-006}.
\section{Data analysis}
This paper aims at measuring the ratio of the cross sections for prompt {\ensuremath{\Pchi_{\cquark 1}}}\xspace and {\ensuremath{\Pchi_{\cquark 2}}}\xspace production.
The cross-section ratio is defined as
\begin{equation}
\frac{ \sigma({\ensuremath{\Pchi_{\cquark 2}}}\xspace) }{ \sigma({\ensuremath{\Pchi_{\cquark 1}}}\xspace) } =
\frac{ N_{{\ensuremath{\Pchi_{\cquark 2}}}\xspace} }{ N_{{\ensuremath{\Pchi_{\cquark 1}}}\xspace} }
\frac{ \varepsilon_{{\ensuremath{\Pchi_{\cquark 1}}}\xspace} }{ \varepsilon_{{\ensuremath{\Pchi_{\cquark 2}}}\xspace} }
\frac{ \mathcal{B}(\decay{{\ensuremath{\Pchi_{\cquark 1}}}\xspace}{{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace\Pgamma})}{ \mathcal{B}(\decay{{\ensuremath{\Pchi_{\cquark 2}}}\xspace}{{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace\Pgamma}) }.
\label{eq:cross-section-ratio_def}
\end{equation}
Here, $N_{{\ensuremath{\Pchi_{\cquark 2}}}\xspace}$ and $N_{{\ensuremath{\Pchi_{\cquark 1}}}\xspace}$ represent the signal yields of the $\chi_{c2}$ and $\chi_{c1}$ states, respectively, and $\varepsilon_{{\ensuremath{\Pchi_{\cquark 2}}}\xspace}$ and $\varepsilon_{{\ensuremath{\Pchi_{\cquark 1}}}\xspace}$ denote the efficiencies to reconstruct and select the corresponding state.
The branching fractions for the {\ensuremath{\Pchi_{\cquark 1,2}}}\xspace decays are $\mathcal{B}(\decay{{\ensuremath{\Pchi_{\cquark 1}}}\xspace}{{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace\Pgamma}) = \left( 34.3 \pm 1.0 \right) \%$ and $\mathcal{B}(\decay{{\ensuremath{\Pchi_{\cquark 2}}}\xspace}{{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace\Pgamma}) = \left( 19.0 \pm 0.5 \right) \%$ \cite{PDG2020}.
The {\ensuremath{\Pchi_{\cquark 1}}}\xspace and {\ensuremath{\Pchi_{\cquark 2}}}\xspace signal yields are determined by performing a binned maximum-likelihood fit to the spectra of the difference between the invariant mass of the ${\ensuremath{\Pmu^+\Pmu^-}}\xspace \Pgamma$ candidate and that of the {\ensuremath{\Pmu^+\Pmu^-}}\xspace pair, $\Delta M \equiv M({\ensuremath{\Pmu^+\Pmu^-}}\xspace \gamma) - M({\ensuremath{\Pmu^+\Pmu^-}}\xspace)$.
The fit function comprises a Gaussian shape for the {\ensuremath{\Pchi_{\cquark 1}}}\xspace and {\ensuremath{\Pchi_{\cquark 2}}}\xspace resonances and a background component described with a second-order Chebyshev polynomial.
In the fit, the difference between the values of the $\chi_{c1}$ and $\chi_{c2}$ masses is set to the known mass difference \cite{PDG2020}. The widths of the $\chi_{c1}$ and $\chi_{c2}$ peaks are set to be equal, following expectations from simulation, and left as a free parameter. The $\chi_{c0}$ peak is also included in the fit, however no significant $\chi_{c0}$ yield is observed.
The fit to the spectra of converted candidates is performed in the range $200 < \Delta M < 800 \ (850)$ MeV/$c^2$ at forward (backward) rapidity. For the calorimetric candidates, the invariant-mass difference spectrum is fitted between $250 < \Delta M < 650$ MeV/$c^2$ in the two rapidity intervals.
The mass-difference spectra of the converted and calorimetric samples are shown, together with the fit components, in Figs. \ref{fig:signal_extraction_conv} and \ref{fig:signal_extraction_calo}, respectively.
In the converted samples, the yield ratio $N_{{\ensuremath{\Pchi_{\cquark 2}}}\xspace}/N_{{\ensuremath{\Pchi_{\cquark 1}}}\xspace}$ is determined to be $0.51 \pm 0.23$ at forward and $0.56 \pm 0.26$ at backward rapidity, where the uncertainties are statistical.
In the calorimetric samples, these ratios are found to be $0.63 \pm 0.08$ at forward and $0.67 \pm 0.10$ at backward rapidity. Individual yields as well as their corresponding significance are listed in Table~\ref{tab:zbytecna_tabulka}.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=0.49\linewidth]{Fig1a.pdf}
\includegraphics[width=0.49\linewidth]{Fig1b.pdf}
\vspace*{-0.5cm}
\end{center}
\caption{
Mass-difference spectra of converted {\ensuremath{\Pchi_{\cquark 1,2}}}\xspace candidates in forward (left) and backward (right) configuration data. The data are superimposed with a fit (solid blue line) comprising {\ensuremath{\Pchi_{\cquark 1}}}\xspace and {\ensuremath{\Pchi_{\cquark 2}}}\xspace signals and combinatorial background (dashed black line).
}
\label{fig:signal_extraction_conv}
\end{figure}
\begin{figure}[tb]
\begin{center}
\includegraphics[width=0.49\linewidth]{Fig2a.pdf}
\includegraphics[width=0.49\linewidth]{Fig2b.pdf}
\vspace*{-0.5cm}
\end{center}
\caption{
Mass-difference spectra of calorimetric {\ensuremath{\Pchi_{\cquark 1,2}}}\xspace candidates in forward (left) and backward (right) data. The data are superimposed with a fit result (solid blue line) comprising ${\ensuremath{\Pchi_{\cquark 1}}}\xspace$ and ${\ensuremath{\Pchi_{\cquark 2}}}\xspace$ signals and combinatorial background (dashed black line).
}
\label{fig:signal_extraction_calo}
\end{figure}
Since the kinematics of ${\ensuremath{\Pchi_{\cquark 1}}}\xspace$ and ${\ensuremath{\Pchi_{\cquark 2}}}\xspace$ decays are nearly identical, various detector effects such as tracking and particle-identification efficiencies cancel out in the ratio, so that the efficiency ratio in Eq.~(\ref{eq:cross-section-ratio_def}) can be expressed as
$\displaystyle{\frac{ \varepsilon_{{\ensuremath{\Pchi_{\cquark 1}}}\xspace} }{ \varepsilon_{{\ensuremath{\Pchi_{\cquark 2}}}\xspace} } = \frac{ \varepsilon^{\rm acc}_{{\ensuremath{\Pchi_{\cquark 1}}}\xspace} }{ \varepsilon^{\rm acc}_{{\ensuremath{\Pchi_{\cquark 2}}}\xspace} } \frac{ \varepsilon^{\rm reco}_{{\ensuremath{\Pchi_{\cquark 1}}}\xspace} }{ \varepsilon^{\rm reco}_{{\ensuremath{\Pchi_{\cquark 2}}}\xspace} }}$.
The factor $\varepsilon^{\rm acc}$ expresses the geometrical acceptance of the decay products to fall within the LHCb acceptance,
while the factor $\varepsilon^{\rm reco}$ represents the efficiency of selection and reconstruction of the signal candidates.
These correction factors are computed from dedicated simulated events.
\begin{table}[t]
\caption{
Yields of {\ensuremath{\Pchi_{\cquark 1}}}\xspace and {\ensuremath{\Pchi_{\cquark 2}}}\xspace signals with statistical uncertainties and corresponding significance (given in standard deviations).
}
\begin{center}\begin{tabular}{@{} c c c c c c @{} }
\toprule
\multicolumn{2}{c}{
Data sample} & $N_{{\ensuremath{\Pchi_{\cquark 1}}}\xspace}$ & Significance & $N_{{\ensuremath{\Pchi_{\cquark 2}}}\xspace}$ & Significance \\
\midrule
\multirow{2}{2.3cm}{\centering Converted photons} & $1.5 < y^* < 4.0$ & $41 \pm 9$ & $6.0$ & $21 \pm 8$ & $3.1$ \\
& $-5.0 < y^* < -2.5$ & $38 \pm 9$ & $4.4$ & $21 \pm 8$ & $3.0$\\
\midrule
\multirow{2}{2.3cm}{\centering Calorimetric photons} & $1.5 < y^* < 4.0$ & $1151 \pm 69$ & $15.7$ & $721 \pm 76$ & $9.8$ \\
& $-5.0 < y^* < -2.5$ & $1004 \pm 73$ & $13.3$ & $676 \pm 82$ & $8.5$\\
\bottomrule
\end{tabular}\end{center}
\label{tab:zbytecna_tabulka}
\end{table}
\section{Systematic uncertainties}
The systematic uncertainties on the cross-section ratios are determined as follows.
A systematic uncertainty on the signal extraction is determined by varying the models used in the mass-difference fits.
Several different signal and background models are tested. The signal shapes are varied between Gaussian functions and Voigtian functions (a convolution of a Breit-Wigner and a Gaussian function), and the background shape is varied between second- and third-order Chebyshev polynomials. The natural widths of the {\ensuremath{\Pchi_{\cquark 1}}}\xspace and {\ensuremath{\Pchi_{\cquark 2}}}\xspace states are narrow compared to the resolution, the Breit-Wigner widths are therefore fixed to the known values~\cite{PDG2020}.
The fit range is varied between $100 \ (150) < \Delta M < 900~\ensuremath{\aunit{Me\kern -0.1em V\!/}c^2}\xspace$ and $200 < \Delta M < 800 \ (850)~\ensuremath{\aunit{Me\kern -0.1em V\!/}c^2}\xspace$ for the converted candidates at forward (backward) rapidity.
For the calorimetric candidates, the fit range is varied between $250 < \Delta M < 650~\ensuremath{\aunit{Me\kern -0.1em V\!/}c^2}\xspace$ and $300 < \Delta M < 600~\ensuremath{\aunit{Me\kern -0.1em V\!/}c^2}\xspace$ in the two rapidity intervals.
The various choices of signal shape, background parametrization, and range give a total of eight fits to each of the mass-difference spectra in each rapidity interval.
In all cases, the $\chi_{c0}$ peak is also included in the fit; however, no significant $\chi_{c0}$ yield is observed.
The systematic uncertainty on the yield ratios due to the fitting procedure is assigned as the standard deviation between the values returned by the eight individual fits.
For the converted sample, this systematic uncertainty amounts to $4.9\%$ ($3.2\%$) at forward (backward) rapidity. For the calorimetric sample it is $2.6\%$ ($6.8\%$) at forward (backward) rapidity. The residual background from the nonprompt $\chi_{c1,2}$ production is verified as negligible and shown to cancel out in the ratio, hence no related uncertainty is assigned.
The systematic uncertainty on the acceptance and efficiency corrections includes contributions from the limited size of the simulated samples used to compute the $\varepsilon^{\rm acc}$ and $\varepsilon^{\rm reco}$ factors, and the uncertainty due to the discrepancy of the $\chi_{c1,2}$ and photon properties between data and simulation.
The latter is estimated using simulated samples, weighted to reproduce the kinematic distributions of $\chi_{c1,2}$ and photons in background-subtracted data, and obtained using the \mbox{\em sPlot}\xspace technique, with $\Delta M$ as the discriminating variable~\cite{Pivk:2004ty}.
The weights are extracted by comparing the transverse momentum and rapidity dependent ratios of the simulated counts $N_{{\ensuremath{\Pchi_{\cquark 1}}}\xspace} / N_{{\ensuremath{\Pchi_{\cquark 2}}}\xspace}$ with those in data. The simulated ${\ensuremath{\Pchi_{\cquark 1}}}\xspace$ samples are then weighted event-by-event
and the uncertainty is assessed as the difference between the efficiency ratios computed from simulated samples prior to and after weighting.
In the case of calorimetric photons, an additional weighting process is required in order to recover kinematic distributions of final-state photons observed in the data as well, in a similar event-by-event process as the weights obtained from $\chi_{c1,2}$ kinematic distributions.
The effect of the photon-identification selection and the reproducibility of relevant variables in simulation are also taken into account. For the converted $\chi_{c1,2}$ sample, the total systematic uncertainty on the acceptance and efficiency equals $9.6\%$ at forward and $14.9\%$ at backward rapidity, while for the calorimetric sample the uncertainty is $8.1\%$ at forward rapidity and $12.4\%$ at backward rapidity.
The ratio of the branching fractions of the ${\ensuremath{\Pchi_{\cquark 1,2}}}\xspace \ensuremath{\rightarrow}\xspace {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace\gamma$ decays contributes with an uncertainty of $3.9\%$.
A summary of contributions to the
statistical and
systematic uncertainties of each analyzed sample is given in Table \ref{tab:systematic_uncertainty}.
\begin{table}[t]
\caption{
Statistical and systematic uncertainties on the cross-section ratio, $\sigma({\ensuremath{\Pchi_{\cquark 2}}}\xspace) / \sigma({\ensuremath{\Pchi_{\cquark 1}}}\xspace)$. The total systematic uncertainty is also quoted.}
\begin{center}\begin{tabular}{@{} c c c c @{} }
\toprule
Analyzed sample & Source & $1.5 < y^* < 4.0$ & $-5.0 < y^* < -2.5$ \\
\midrule
\multirow{6}{*}{Converted photons} & Signal extraction & $4.9 \%$ & $ 3.2 \%$ \\
& Limited simulation sample size & $ 5.6 \% $ & $ 6.5 \% $ \\
& Efficiency correction & $ 7.7 \% $ & $ 13.4 \% $ \\
& Branching fraction ratio & $3.9\%$ & $3.9\%$ \\
\cmidrule{2-4}
& Total systematic uncertainty & $11.4 \%$ & $15.7 \%$ \\
& Statistical uncertainty & $45.2 \%$ & $ 47.0 \%$ \\
\midrule
\multirow{6}{*}{Calorimetric photons} & Signal extraction & $2.6 \%$ & $ 6.8 \%$ \\
& Limited simulation sample size & $ 2.5 \% $ & $ 2.8 \% $ \\
& Efficiency correction & $ 7.7 \% $ & $ 12.1 \% $ \\
& Branching fraction ratio & $3.9\%$ & $3.9\%$ \\
\cmidrule{2-4}
& Total systematic uncertainty & $9.3 \%$ & $14.7\%$ \\
& Statistical uncertainty & $12.2 \%$ & $ 14.2 \%$ \\
\bottomrule
\end{tabular}\end{center}
\label{tab:systematic_uncertainty}
\end{table}
\section{Results}
The prompt-production cross-section ratio $\sigma({\ensuremath{\Pchi_{\cquark 2}}}\xspace) / \sigma({\ensuremath{\Pchi_{\cquark 1}}}\xspace)$ in {\ensuremath{p}}Pb\xspace collisions at the center-of-mass energy per nucleon pair \ensuremath{\protect\sqrt{s_{\scriptscriptstyle\text{NN}}}}\xspace~=~8.16~TeV is shown for the two rapidity regions in Fig. \ref{fig:cross_section_ratio}.
The ratio measured from converted photons amounts to
\begin{align*}
\frac{ \sigma({\ensuremath{\Pchi_{\cquark 2}}}\xspace) }{ \sigma({\ensuremath{\Pchi_{\cquark 1}}}\xspace) }
&= 0.92 \pm 0.42 \mbox{ (stat.) } \pm 0.11 \mbox{ (syst.) } && \mbox{ for } 1.5 < y^* < 4.0, \\
\frac{ \sigma({\ensuremath{\Pchi_{\cquark 2}}}\xspace) }{ \sigma({\ensuremath{\Pchi_{\cquark 1}}}\xspace) }
&= 0.98 \pm 0.46 \mbox{ (stat.) } \pm 0.15 \mbox{ (syst.) } && \mbox{ for } -5.0 < y^* < -2.5.
\end{align*}
The ratio measured from calorimetric photons is found to be
\begin{align*}
\frac{ \sigma({\ensuremath{\Pchi_{\cquark 2}}}\xspace) }{ \sigma({\ensuremath{\Pchi_{\cquark 1}}}\xspace) }
&= 1.11 \pm 0.14 \mbox{ (stat.) } \pm 0.10 \mbox{ (syst.) } && \mbox{ for } 1.5 < y^* < 4.0, \\
\frac{ \sigma({\ensuremath{\Pchi_{\cquark 2}}}\xspace) }{ \sigma({\ensuremath{\Pchi_{\cquark 1}}}\xspace) }
&= 1.14 \pm 0.16 \mbox{ (stat.) } \pm 0.17 \mbox{ (syst.) } && \mbox{ for } -5.0 < y^* < -2.5.
\end{align*}
The cross-section ratios for both converted and calorimetric samples are consistent with unity in both rapidity regions.
The significantly larger yield of the calorimetric sample allows more precise conclusions on the observed trend to be drawn.
The cross-section ratio obtained in {\ensuremath{p}}Pb\xspace data is compared with the corresponding ratio measured in \ensuremath{pp}\xspace collisions at \ensuremath{\protect\sqrt{s}}\xspace~=~7~TeV by the LHCb collaboration \cite{LHCB-PAPER-2013-028}.
The two measurements are consistent within two standard deviations. While the ratio in the \ensuremath{pp}\xspace data was measured at a lower center-of-mass energy than that of {\ensuremath{p}}Pb\xspace collisions, results show that the relative cross-section of different charmonium states is independent of energy at the LHC energy scale \cite{Acharya:2017hjh}.
Thus, the only aspect to consider in a direct comparison between the shown {\ensuremath{p}}Pb\xspace and \ensuremath{pp}\xspace data is the rapidity range, where the {\ensuremath{p}}Pb\xspace results are shifted by $-0.5$ in rapidity.
Bearing that in mind, we can express the relative suppression of $\chi_{c2}$ and $\chi_{c1}$ states via the ratio of their nuclear modification factors
\begin{equation}
\mathcal{R} \equiv \frac{\sigma({\ensuremath{\Pchi_{\cquark 2}}}\xspace) / \sigma({\ensuremath{\Pchi_{\cquark 1}}}\xspace)\scalebox{1.5}{$\vert$}_{p{\rm Pb}} }{\sigma({\ensuremath{\Pchi_{\cquark 2}}}\xspace) / \sigma({\ensuremath{\Pchi_{\cquark 1}}}\xspace)\scalebox{1.5}{$\vert$}_{\ensuremath{pp}\xspace} }.
\label{eq:rppb_ratio}
\end{equation}
Using the more precise calorimetric {\ensuremath{p}}Pb\xspace results, the ratio of nuclear-modification factors amounts to
$\mathcal{R} = 1.41 \pm 0.21 \mbox{ (stat.) } \pm 0.18 \mbox{ (syst.)}$ at forward and
$\mathcal{R} = 1.44 \pm 0.24 \mbox{ (stat.) } \pm 0.25 \mbox{ (syst.)}$ at backward rapidity,
showing no significant change relative to the \ensuremath{pp}\xspace ratio in either rapidity region.
The measured cross-section ratio and ratio of nuclear-modification factors suggest that the nuclear effects have the same impact on both ${\ensuremath{\Pchi_{\cquark 1}}}\xspace$ and ${\ensuremath{\Pchi_{\cquark 2}}}\xspace$ states within uncertainties, independent of rapidity.
\begin{figure}[tb]
\begin{center}
\includegraphics[width=0.7\linewidth]{Fig3.pdf}
\vspace*{-0.5cm}
\end{center}
\caption{
Cross-section ratio, $\sigma({\ensuremath{\Pchi_{\cquark 2}}}\xspace) / \sigma({\ensuremath{\Pchi_{\cquark 1}}}\xspace)$ as a function of center-of-mass rapidity $y^*$, for the ${\ensuremath{\Pchi_{\cquark 2}}}\xspace$ and ${\ensuremath{\Pchi_{\cquark 1}}}\xspace$ promptly produced in {\ensuremath{p}}Pb\xspace collisions measured using converted photons (red circles) and calorimetric photons (blue squares).
The error bars correspond to the total uncertainties.
Blue points and vertical uncertainties are shifted horizontally to improve visibility. The {\ensuremath{p}}Pb\xspace data are compared with results of the converted sample in \ensuremath{pp}\xspace collisions at \ensuremath{\protect\sqrt{s}}\xspace~=~7~TeV \cite{LHCB-PAPER-2013-028} (yellow triangles).
}
\label{fig:cross_section_ratio}
\end{figure}
\section{Summary}
In summary, we present the first measurement of $\chi_{c1,2}$ charmonium production in nuclear collisions at the LHC. The cross-section ratio $\sigma({\ensuremath{\Pchi_{\cquark 2}}}\xspace) / \sigma({\ensuremath{\Pchi_{\cquark 1}}}\xspace)$ is consistent with unity for both forward and backward rapidity regions. Moreover, comparison with the ratio measured in \ensuremath{pp}\xspace collisions hints at a suppression pattern between the two states, which is comparable within uncertainties. This suggests that the final-state nuclear effects impact the ${\ensuremath{\Pchi_{\cquark 1}}}\xspace$ and ${\ensuremath{\Pchi_{\cquark 2}}}\xspace$ states similarly within the achieved precision.
\section*{Acknowledgements}
\noindent We express our gratitude to our colleagues in the CERN
accelerator departments for the excellent performance of the LHC. We
thank the technical and administrative staff at the LHCb
institutes.
We acknowledge support from CERN and from the national agencies:
CAPES, CNPq, FAPERJ and FINEP (Brazil);
MOST and NSFC (China);
CNRS/IN2P3 (France);
BMBF, DFG and MPG (Germany);
INFN (Italy);
NWO (Netherlands);
MNiSW and NCN (Poland);
MEN/IFA (Romania);
MSHE (Russia);
MICINN (Spain);
SNSF and SER (Switzerland);
NASU (Ukraine);
STFC (United Kingdom);
DOE NP and NSF (USA).
We acknowledge the computing resources that are provided by CERN, IN2P3
(France), KIT and DESY (Germany), INFN (Italy), SURF (Netherlands),
PIC (Spain), GridPP (United Kingdom), RRCKI and Yandex
LLC (Russia), CSCS (Switzerland), IFIN-HH (Romania), CBPF (Brazil),
PL-GRID (Poland) and NERSC (USA).
We are indebted to the communities behind the multiple open-source
software packages on which we depend.
Individual groups or members have received support from
ARC and ARDC (Australia);
AvH Foundation (Germany);
EPLANET, Marie Sk\l{}odowska-Curie Actions and ERC (European Union);
A*MIDEX, ANR, Labex P2IO and OCEVU, and R\'{e}gion Auvergne-Rh\^{o}ne-Alpes (France);
Key Research Program of Frontier Sciences of CAS, CAS PIFI, CAS CCEPP,
Fundamental Research Funds for the Central Universities,
and Sci. \& Tech. Program of Guangzhou (China);
RFBR, RSF and Yandex LLC (Russia);
GVA, XuntaGal and GENCAT (Spain);
the Leverhulme Trust, the Royal Society
and UKRI (United Kingdom).
\addcontentsline{toc}{section}{References}
\ifx\mcitethebibliography\mciteundefinedmacro
\PackageError{LHCb.bst}{mciteplus.sty has not been loaded}
{This bibstyle requires the use of the mciteplus package.}\fi
\providecommand{\href}[2]{#2}
\begin{mcitethebibliography}{10}
\mciteSetBstSublistMode{n}
\mciteSetBstMaxWidthForm{subitem}{\alph{mcitesubitemcount})}
\mciteSetBstSublistLabelBeginEnd{\mcitemaxwidthsubitemform\space}
{\relax}{\relax}
\bibitem{Vogt:2004dh}
R.~Vogt, \ifthenelse{\boolean{articletitles}}{\emph{{Shadowing and absorption
effects on {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace production in dA collisions}},
}{}\href{https://doi.org/10.1103/PhysRevC.71.054902}{Phys.\ Rev.\
\textbf{C71} (2005) 054902},
\href{http://arxiv.org/abs/hep-ph/0411378}{{\normalfont\ttfamily
arXiv:hep-ph/0411378}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Arleo:2012rs}
F.~Arleo and S.~Peign\'{e},
\ifthenelse{\boolean{articletitles}}{\emph{{Heavy-quarkonium suppression in
p-A collisions from parton energy loss in cold QCD matter}},
}{}\href{https://doi.org/10.1007/JHEP03(2013)122}{JHEP \textbf{03} (2013)
122}, \href{http://arxiv.org/abs/1212.0434}{{\normalfont\ttfamily
arXiv:1212.0434}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Capella:1996va}
A.~Capella, A.~Kaidalov, A.~Kouider~Akil, and C.~Gerschel,
\ifthenelse{\boolean{articletitles}}{\emph{{{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace and $\psi'$ suppression in
heavy ion collisions}},
}{}\href{https://doi.org/10.1016/S0370-2693(96)01650-4}{Phys.\ Lett.\
\textbf{B393} (1997) 431},
\href{http://arxiv.org/abs/hep-ph/9607265}{{\normalfont\ttfamily
arXiv:hep-ph/9607265}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Matsui:1986dk}
T.~Matsui and H.~Satz, \ifthenelse{\boolean{articletitles}}{\emph{{{\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace
suppression by quark-gluon plasma formation}},
}{}\href{https://doi.org/10.1016/0370-2693(86)91404-8}{Phys.\ Lett.\
\textbf{B178} (1986) 416}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCB-PAPER-2017-014}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Prompt and nonprompt {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace\
production and nuclear modification in {\ensuremath{\Pp}}\xspace{}Pb collisions at $\ensuremath{\protect\sqrt{s_{\scriptscriptstyle\text{NN}}}}\xspace =
$8.16\aunit{Te\kern -0.1em V}\xspace}}, }{}\href{https://doi.org/10.1016/j.physletb.2017.09.058}{Phys.\
Lett.\ \textbf{B774} (2017) 159},
\href{http://arxiv.org/abs/1706.07122}{{\normalfont\ttfamily
arXiv:1706.07122}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2015-058}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Study of {\ensuremath{\Ppsi{(2S)}}}\xspace production
cross-sections and cold nuclear matter effects in {\ensuremath{\Pp}}\xspace{}Pb collisions at
$\ensuremath{\protect\sqrt{s_{\scriptscriptstyle\text{NN}}}}\xspace = $5\aunit{Te\kern -0.1em V}\xspace}}, }{}\href{https://doi.org/10.1007/JHEP03(2016)133}{JHEP
\textbf{03} (2016) 133},
\href{http://arxiv.org/abs/1601.07878}{{\normalfont\ttfamily
arXiv:1601.07878}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Adare:2016psx}
PHENIX collaboration, A.~Adare {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Measurement of the relative
yields of $\psi(2S)$ to $\psi(1S)$ mesons produced at forward and backward
rapidity in $p$$+$$p$, $p$$+$Al, $p$$+$Au, and $^{3}$He$+$Au collisions at
$\sqrt{s_{_{NN}}}=200$ GeV}},
}{}\href{https://doi.org/10.1103/PhysRevC.95.034904}{Phys.\ Rev.\
\textbf{C95} (2017) 034904},
\href{http://arxiv.org/abs/1609.06550}{{\normalfont\ttfamily
arXiv:1609.06550}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PAPER-2018-035}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Study of {\ensuremath{\PUpsilon}}\xspace production
in {\ensuremath{\Pp}}\xspace{}Pb collisions at $\ensuremath{\protect\sqrt{s_{\scriptscriptstyle\text{NN}}}}\xspace = $8.16\aunit{Te\kern -0.1em V}\xspace}},
}{}\href{https://doi.org/10.1007/JHEP11(2018)194}{JHEP \textbf{11} (2018)
194}, \href{http://arxiv.org/abs/1810.07655}{{\normalfont\ttfamily
arXiv:1810.07655}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Acharya:2020wwy}
ALICE collaboration, S.~Acharya {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Measurement of nuclear effects on
$\psi\rm{(2S)}$ production in p-Pb collisions at $\sqrt{\textit{s}_{\rm NN}}
= 8.16$ TeV}}, }{}\href{https://doi.org/10.1007/JHEP07(2020)237}{JHEP
\textbf{07} (2020) 237},
\href{http://arxiv.org/abs/2003.06053}{{\normalfont\ttfamily
arXiv:2003.06053}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Aaboud:2017cif}
ATLAS collaboration, M.~Aaboud {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Measurement of quarkonium
production in proton\textendash{}lead and proton\textendash{}proton
collisions at $5.02~\mathrm {TeV}$ with the ATLAS detector}},
}{}\href{https://doi.org/10.1140/epjc/s10052-018-5624-4}{Eur.\ Phys.\ J.\
\textbf{C78} (2018) 171},
\href{http://arxiv.org/abs/1709.03089}{{\normalfont\ttfamily
arXiv:1709.03089}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Sirunyan:2018pse}
CMS collaboration, A.~M. Sirunyan {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Measurement of prompt $\psi$(2S)
production cross sections in proton-lead and proton-proton collisions at
$\sqrt{s_{_\mathrm{NN}}}=$ 5.02 TeV}},
}{}\href{https://doi.org/10.1016/j.physletb.2019.01.058}{Phys.\ Lett.\ B
\textbf{790} (2019) 509},
\href{http://arxiv.org/abs/1805.02248}{{\normalfont\ttfamily
arXiv:1805.02248}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Ferreiro:2014bia}
E.~G. Ferreiro, \ifthenelse{\boolean{articletitles}}{\emph{{Excited charmonium
suppression in proton\textendash{}nucleus collisions as a consequence of
comovers}}, }{}\href{https://doi.org/10.1016/j.physletb.2015.07.066}{Phys.\
Lett.\ \textbf{B749} (2015) 98},
\href{http://arxiv.org/abs/1411.0549}{{\normalfont\ttfamily
arXiv:1411.0549}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Ferreiro:2018wbd}
E.~G. Ferreiro and J.-P. Lansberg,
\ifthenelse{\boolean{articletitles}}{\emph{{Is bottomonium suppression in
proton-nucleus and nucleus-nucleus collisions at LHC energies due to the same
effects?}}, }{}\href{https://doi.org/10.1007/JHEP10(2018)094}{JHEP
\textbf{10} (2018) 094},
\href{http://arxiv.org/abs/1804.04474}{{\normalfont\ttfamily
arXiv:1804.04474}}, [Erratum: JHEP \textbf{03}, 063 (2019)]\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Abulencia:2007bra}
CDF collaboration, A.~Abulencia {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Measurement of
$\sigma_{\chi_{c2}}{\cal B}(\chi_{c2} \ensuremath{\rightarrow}\xspace J/\psi \gamma)/\sigma_{\chi_{c1}}
{\cal B}(\chi_{c1} \ensuremath{\rightarrow}\xspace J/\psi \gamma)$ in $p \bar{p}$ Collisions at
$\sqrt{s}$ = 1.96-TeV}},
}{}\href{https://doi.org/10.1103/PhysRevLett.98.232001}{Phys.\ Rev.\ Lett.\
\textbf{98} (2007) 232001},
\href{http://arxiv.org/abs/hep-ex/0703028}{{\normalfont\ttfamily
arXiv:hep-ex/0703028}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCB-PAPER-2011-019}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Measurement of the cross-section
ratio $\sigma({\ensuremath{\Pchi_{\cquark 2}}}\xspace)/\sigma({\ensuremath{\Pchi_{\cquark 1}}}\xspace)$ for prompt $\chi_c$ production at
\mbox{$\ensuremath{\protect\sqrt{s}}\xspace=$7\aunit{Te\kern -0.1em V}\xspace}}},
}{}\href{https://doi.org/10.1016/j.physletb.2012.06.077}{Phys.\ Lett.\
\textbf{B714} (2012) 215},
\href{http://arxiv.org/abs/1202.1080}{{\normalfont\ttfamily
arXiv:1202.1080}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCB-PAPER-2013-028}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Measurement of the relative rate
of prompt ${\ensuremath{\Pchi_{\cquark 0}}}\xspace$, ${\ensuremath{\Pchi_{\cquark 1}}}\xspace$ and ${\ensuremath{\Pchi_{\cquark 2}}}\xspace$ production at
\mbox{$\ensuremath{\protect\sqrt{s}}\xspace=$7\aunit{Te\kern -0.1em V}\xspace}}}, }{}\href{https://doi.org/10.1007/JHEP10(2013)115}{JHEP
\textbf{10} (2013) 115},
\href{http://arxiv.org/abs/1307.4285}{{\normalfont\ttfamily
arXiv:1307.4285}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Chatrchyan:2012ub}
CMS collaboration, S.~Chatrchyan {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Measurement of the relative
prompt production rate of $\chi_{c2}$ and $\chi_{c1}$ in $pp$ Collisions at
$\sqrt{s}=7$ TeV}},
}{}\href{https://doi.org/10.1140/epjc/s10052-012-2251-3}{Eur.\ Phys.\ J.\
\textbf{C72} (2012) 2251},
\href{http://arxiv.org/abs/1210.0875}{{\normalfont\ttfamily
arXiv:1210.0875}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Adare:2011vq}
PHENIX collaboration, A.~Adare {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Ground and excited charmonium
state production in $p+p$ collisions at $\sqrt{s}=200$ GeV}},
}{}\href{https://doi.org/10.1103/PhysRevD.85.092004}{Phys.\ Rev.\
\textbf{D85} (2012) 092004},
\href{http://arxiv.org/abs/1105.1966}{{\normalfont\ttfamily
arXiv:1105.1966}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Satz:2005hx}
H.~Satz, \ifthenelse{\boolean{articletitles}}{\emph{{Colour deconfinement and
quarkonium binding}},
}{}\href{https://doi.org/10.1088/0954-3899/32/3/R01}{J.\ Phys.\ \textbf{G32}
(2006) R25}, \href{http://arxiv.org/abs/hep-ph/0512217}{{\normalfont\ttfamily
arXiv:hep-ph/0512217}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Abt:2008ed}
HERA-B collaboration, I.~Abt {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Production of the charmonium
states $\chi_{c1}$ and $\chi_{c2}$ in proton nucleus interactions at
$\sqrt{s} = 41.6\aunit{Ge\kern -0.1em V}\xspace$}},
}{}\href{https://doi.org/10.1103/PhysRevD.79.012001}{Phys.\ Rev.\
\textbf{D79} (2009) 012001},
\href{http://arxiv.org/abs/0807.2167}{{\normalfont\ttfamily
arXiv:0807.2167}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Adare:2013ezl}
PHENIX collaboration, A.~Adare {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Nuclear modification of $\psi'$,
$\chi_c$, and {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace production in d+Au collisions at $\sqrt{s_{NN}}
=200\aunit{Ge\kern -0.1em V}\xspace$}}, }{}\href{https://doi.org/10.1103/PhysRevLett.111.202301}{Phys.\
Rev.\ Lett.\ \textbf{111} (2013) 202301},
\href{http://arxiv.org/abs/1305.5516}{{\normalfont\ttfamily
arXiv:1305.5516}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-DP-2008-001}
LHCb collaboration, A.~A. Alves~Jr.\ {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{The \mbox{LHCb}\xspace detector at the LHC}},
}{}\href{https://doi.org/10.1088/1748-0221/3/08/S08005}{JINST \textbf{3}
(2008) S08005}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-DP-2014-002}
LHCb collaboration, R.~Aaij {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{LHCb detector performance}},
}{}\href{https://doi.org/10.1142/S0217751X15300227}{Int.\ J.\ Mod.\ Phys.\
\textbf{A30} (2015) 1530022},
\href{http://arxiv.org/abs/1412.6352}{{\normalfont\ttfamily
arXiv:1412.6352}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{PDG2020}
Particle Data Group, P.~A. Zyla {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{\href{http://pdg.lbl.gov/}{Review
of particle physics}}}, }{}\href{https://doi.org/10.1093/ptep/ptaa104}{Prog.\
Theor.\ Exp.\ Phys.\ \textbf{2020} (2020) 083C01}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-DP-2018-001}
R.~Aaij {\em et~al.}, \ifthenelse{\boolean{articletitles}}{\emph{{Selection and
processing of calibration samples to measure the particle identification
performance of the LHCb experiment in Run 2}},
}{}\href{https://doi.org/10.1140/epjti/s40485-019-0050-z}{Eur.\ Phys.\ J.\
Tech.\ Instr.\ \textbf{6} (2019) 1},
\href{http://arxiv.org/abs/1803.00824}{{\normalfont\ttfamily
arXiv:1803.00824}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Beteta:2020gui}
C.~Abell\'{a}n~Beteta {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Calibration and performance of
the LHCb calorimeters in Run 1 and 2 at the LHC}},
}{}\href{http://arxiv.org/abs/2008.11556}{{\normalfont\ttfamily
arXiv:2008.11556}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Sjostrand:2007gs}
T.~Sj\"{o}strand, S.~Mrenna, and P.~Skands,
\ifthenelse{\boolean{articletitles}}{\emph{{A brief introduction to PYTHIA
8.1}}, }{}\href{https://doi.org/10.1016/j.cpc.2008.01.036}{Comput.\ Phys.\
Commun.\ \textbf{178} (2008) 852},
\href{http://arxiv.org/abs/0710.3820}{{\normalfont\ttfamily
arXiv:0710.3820}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PROC-2010-056}
I.~Belyaev {\em et~al.}, \ifthenelse{\boolean{articletitles}}{\emph{{Handling
of the generation of primary events in Gauss, the LHCb simulation
framework}}, }{}\href{https://doi.org/10.1088/1742-6596/331/3/032047}{J.\
Phys.\ Conf.\ Ser.\ \textbf{331} (2011) 032047}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Pierog:2013ria}
T.~Pierog {\em et~al.}, \ifthenelse{\boolean{articletitles}}{\emph{{EPOS LHC:
Test of collective hadronization with data measured at the CERN Large Hadron
Collider}}, }{}\href{https://doi.org/10.1103/PhysRevC.92.034906}{Phys.\ Rev.\
\textbf{C92} (2015) 034906},
\href{http://arxiv.org/abs/1306.0121}{{\normalfont\ttfamily
arXiv:1306.0121}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Lange:2001uf}
D.~J. Lange, \ifthenelse{\boolean{articletitles}}{\emph{{The EvtGen particle
decay simulation package}},
}{}\href{https://doi.org/10.1016/S0168-9002(01)00089-4}{Nucl.\ Instrum.\
Meth.\ \textbf{A462} (2001) 152}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{davidson2015photos}
N.~Davidson, T.~Przedzinski, and Z.~Was,
\ifthenelse{\boolean{articletitles}}{\emph{{PHOTOS interface in C++:
Technical and physics documentation}},
}{}\href{https://doi.org/https://doi.org/10.1016/j.cpc.2015.09.013}{Comp.\
Phys.\ Comm.\ \textbf{199} (2016) 86},
\href{http://arxiv.org/abs/1011.0937}{{\normalfont\ttfamily
arXiv:1011.0937}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Allison:2006ve}
Geant4 collaboration, J.~Allison {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Geant4 developments and
applications}}, }{}\href{https://doi.org/10.1109/TNS.2006.869826}{IEEE
Trans.\ Nucl.\ Sci.\ \textbf{53} (2006) 270}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Agostinelli:2002hh}
Geant4 collaboration, S.~Agostinelli {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Geant4: A simulation toolkit}},
}{}\href{https://doi.org/10.1016/S0168-9002(03)01368-8}{Nucl.\ Instrum.\
Meth.\ \textbf{A506} (2003) 250}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{LHCb-PROC-2011-006}
M.~Clemencic {\em et~al.}, \ifthenelse{\boolean{articletitles}}{\emph{{The
\mbox{LHCb}\xspace simulation application, Gauss: Design, evolution and experience}},
}{}\href{https://doi.org/10.1088/1742-6596/331/3/032023}{J.\ Phys.\ Conf.\
Ser.\ \textbf{331} (2011) 032023}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Pivk:2004ty}
M.~Pivk and F.~R. Le~Diberder,
\ifthenelse{\boolean{articletitles}}{\emph{{sPlot: A statistical tool to
unfold data distributions}},
}{}\href{https://doi.org/10.1016/j.nima.2005.08.106}{Nucl.\ Instrum.\ Meth.\
\textbf{A555} (2005) 356},
\href{http://arxiv.org/abs/physics/0402083}{{\normalfont\ttfamily
arXiv:physics/0402083}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\bibitem{Acharya:2017hjh}
ALICE collaboration, S.~Acharya {\em et~al.},
\ifthenelse{\boolean{articletitles}}{\emph{{Energy dependence of
forward-rapidity {\ensuremath{{\PJ\mskip -3mu/\mskip -2mu\Ppsi}}}\xspace and {\ensuremath{\Ppsi{(2S)}}}\xspace production in pp collisions at the LHC}},
}{}\href{https://doi.org/10.1140/epjc/s10052-017-4940-4}{Eur.\ Phys.\ J.\
\textbf{C77} (2017) 392},
\href{http://arxiv.org/abs/1702.00557}{{\normalfont\ttfamily
arXiv:1702.00557}}\relax
\mciteBstWouldAddEndPuncttrue
\mciteSetBstMidEndSepPunct{\mcitedefaultmidpunct}
{\mcitedefaultendpunct}{\mcitedefaultseppunct}\relax
\EndOfBibitem
\end{mcitethebibliography}
\newpage
\centerline
{\large\bf LHCb collaboration}
\begin
{flushleft}
\small
R.~Aaij$^{32}$,
C.~Abell{\'a}n~Beteta$^{50}$,
T.~Ackernley$^{60}$,
B.~Adeva$^{46}$,
M.~Adinolfi$^{54}$,
H.~Afsharnia$^{9}$,
C.A.~Aidala$^{85}$,
S.~Aiola$^{25}$,
Z.~Ajaltouni$^{9}$,
S.~Akar$^{65}$,
J.~Albrecht$^{15}$,
F.~Alessio$^{48}$,
M.~Alexander$^{59}$,
A.~Alfonso~Albero$^{45}$,
Z.~Aliouche$^{62}$,
G.~Alkhazov$^{38}$,
P.~Alvarez~Cartelle$^{55}$,
S.~Amato$^{2}$,
Y.~Amhis$^{11}$,
L.~An$^{48}$,
L.~Anderlini$^{22}$,
A.~Andreianov$^{38}$,
M.~Andreotti$^{21}$,
F.~Archilli$^{17}$,
A.~Artamonov$^{44}$,
M.~Artuso$^{68}$,
K.~Arzymatov$^{42}$,
E.~Aslanides$^{10}$,
M.~Atzeni$^{50}$,
B.~Audurier$^{12}$,
S.~Bachmann$^{17}$,
M.~Bachmayer$^{49}$,
J.J.~Back$^{56}$,
S.~Baker$^{61}$,
P.~Baladron~Rodriguez$^{46}$,
V.~Balagura$^{12}$,
W.~Baldini$^{21,48}$,
J.~Baptista~Leite$^{1}$,
R.J.~Barlow$^{62}$,
S.~Barsuk$^{11}$,
W.~Barter$^{61}$,
M.~Bartolini$^{24}$,
F.~Baryshnikov$^{82}$,
J.M.~Basels$^{14}$,
G.~Bassi$^{29}$,
B.~Batsukh$^{68}$,
A.~Battig$^{15}$,
A.~Bay$^{49}$,
M.~Becker$^{15}$,
F.~Bedeschi$^{29}$,
I.~Bediaga$^{1}$,
A.~Beiter$^{68}$,
V.~Belavin$^{42}$,
S.~Belin$^{27}$,
V.~Bellee$^{49}$,
K.~Belous$^{44}$,
I.~Belov$^{40}$,
I.~Belyaev$^{41}$,
G.~Bencivenni$^{23}$,
E.~Ben-Haim$^{13}$,
A.~Berezhnoy$^{40}$,
R.~Bernet$^{50}$,
D.~Berninghoff$^{17}$,
H.C.~Bernstein$^{68}$,
C.~Bertella$^{48}$,
A.~Bertolin$^{28}$,
C.~Betancourt$^{50}$,
F.~Betti$^{20,d}$,
Ia.~Bezshyiko$^{50}$,
S.~Bhasin$^{54}$,
J.~Bhom$^{35}$,
L.~Bian$^{73}$,
M.S.~Bieker$^{15}$,
S.~Bifani$^{53}$,
P.~Billoir$^{13}$,
M.~Birch$^{61}$,
F.C.R.~Bishop$^{55}$,
A.~Bitadze$^{62}$,
A.~Bizzeti$^{22,k}$,
M.~Bj{\o}rn$^{63}$,
M.P.~Blago$^{48}$,
T.~Blake$^{56}$,
F.~Blanc$^{49}$,
S.~Blusk$^{68}$,
D.~Bobulska$^{59}$,
J.A.~Boelhauve$^{15}$,
O.~Boente~Garcia$^{46}$,
T.~Boettcher$^{64}$,
A.~Boldyrev$^{81}$,
A.~Bondar$^{43}$,
N.~Bondar$^{38,48}$,
S.~Borghi$^{62}$,
M.~Borisyak$^{42}$,
M.~Borsato$^{17}$,
J.T.~Borsuk$^{35}$,
S.A.~Bouchiba$^{49}$,
T.J.V.~Bowcock$^{60}$,
A.~Boyer$^{48}$,
C.~Bozzi$^{21}$,
M.J.~Bradley$^{61}$,
S.~Braun$^{66}$,
A.~Brea~Rodriguez$^{46}$,
M.~Brodski$^{48}$,
J.~Brodzicka$^{35}$,
A.~Brossa~Gonzalo$^{56}$,
D.~Brundu$^{27}$,
A.~Buonaura$^{50}$,
C.~Burr$^{48}$,
A.~Bursche$^{27}$,
A.~Butkevich$^{39}$,
J.S.~Butter$^{32}$,
J.~Buytaert$^{48}$,
W.~Byczynski$^{48}$,
S.~Cadeddu$^{27}$,
H.~Cai$^{73}$,
R.~Calabrese$^{21,f}$,
L.~Calefice$^{15,13}$,
L.~Calero~Diaz$^{23}$,
S.~Cali$^{23}$,
R.~Calladine$^{53}$,
M.~Calvi$^{26,j}$,
M.~Calvo~Gomez$^{84}$,
P.~Camargo~Magalhaes$^{54}$,
A.~Camboni$^{45,84}$,
P.~Campana$^{23}$,
A.F.~Campoverde~Quezada$^{6}$,
S.~Capelli$^{26,j}$,
L.~Capriotti$^{20,d}$,
A.~Carbone$^{20,d}$,
G.~Carboni$^{31}$,
R.~Cardinale$^{24,h}$,
A.~Cardini$^{27}$,
I.~Carli$^{4}$,
P.~Carniti$^{26,j}$,
L.~Carus$^{14}$,
K.~Carvalho~Akiba$^{32}$,
A.~Casais~Vidal$^{46}$,
G.~Casse$^{60}$,
M.~Cattaneo$^{48}$,
G.~Cavallero$^{48}$,
S.~Celani$^{49}$,
J.~Cerasoli$^{10}$,
A.J.~Chadwick$^{60}$,
M.G.~Chapman$^{54}$,
M.~Charles$^{13}$,
Ph.~Charpentier$^{48}$,
G.~Chatzikonstantinidis$^{53}$,
C.A.~Chavez~Barajas$^{60}$,
M.~Chefdeville$^{8}$,
C.~Chen$^{3}$,
S.~Chen$^{27}$,
A.~Chernov$^{35}$,
V.~Chobanova$^{46}$,
S.~Cholak$^{49}$,
M.~Chrzaszcz$^{35}$,
A.~Chubykin$^{38}$,
V.~Chulikov$^{38}$,
P.~Ciambrone$^{23}$,
M.F.~Cicala$^{56}$,
X.~Cid~Vidal$^{46}$,
G.~Ciezarek$^{48}$,
P.E.L.~Clarke$^{58}$,
M.~Clemencic$^{48}$,
H.V.~Cliff$^{55}$,
J.~Closier$^{48}$,
J.L.~Cobbledick$^{62}$,
V.~Coco$^{48}$,
J.A.B.~Coelho$^{11}$,
J.~Cogan$^{10}$,
E.~Cogneras$^{9}$,
L.~Cojocariu$^{37}$,
P.~Collins$^{48}$,
T.~Colombo$^{48}$,
L.~Congedo$^{19,c}$,
A.~Contu$^{27}$,
N.~Cooke$^{53}$,
G.~Coombs$^{59}$,
G.~Corti$^{48}$,
C.M.~Costa~Sobral$^{56}$,
B.~Couturier$^{48}$,
D.C.~Craik$^{64}$,
J.~Crkovsk\'{a}$^{67}$,
M.~Cruz~Torres$^{1}$,
R.~Currie$^{58}$,
C.L.~Da~Silva$^{67}$,
E.~Dall'Occo$^{15}$,
J.~Dalseno$^{46}$,
C.~D'Ambrosio$^{48}$,
A.~Danilina$^{41}$,
P.~d'Argent$^{48}$,
A.~Davis$^{62}$,
O.~De~Aguiar~Francisco$^{62}$,
K.~De~Bruyn$^{78}$,
S.~De~Capua$^{62}$,
M.~De~Cian$^{49}$,
J.M.~De~Miranda$^{1}$,
L.~De~Paula$^{2}$,
M.~De~Serio$^{19,c}$,
D.~De~Simone$^{50}$,
P.~De~Simone$^{23}$,
J.A.~de~Vries$^{79}$,
C.T.~Dean$^{67}$,
D.~Decamp$^{8}$,
L.~Del~Buono$^{13}$,
B.~Delaney$^{55}$,
H.-P.~Dembinski$^{15}$,
A.~Dendek$^{34}$,
V.~Denysenko$^{50}$,
D.~Derkach$^{81}$,
O.~Deschamps$^{9}$,
F.~Desse$^{11}$,
F.~Dettori$^{27,e}$,
B.~Dey$^{73}$,
P.~Di~Nezza$^{23}$,
S.~Didenko$^{82}$,
L.~Dieste~Maronas$^{46}$,
H.~Dijkstra$^{48}$,
V.~Dobishuk$^{52}$,
A.M.~Donohoe$^{18}$,
F.~Dordei$^{27}$,
A.C.~dos~Reis$^{1}$,
L.~Douglas$^{59}$,
A.~Dovbnya$^{51}$,
A.G.~Downes$^{8}$,
K.~Dreimanis$^{60}$,
M.W.~Dudek$^{35}$,
L.~Dufour$^{48}$,
V.~Duk$^{77}$,
P.~Durante$^{48}$,
J.M.~Durham$^{67}$,
D.~Dutta$^{62}$,
M.~Dziewiecki$^{17}$,
A.~Dziurda$^{35}$,
A.~Dzyuba$^{38}$,
S.~Easo$^{57}$,
U.~Egede$^{69}$,
V.~Egorychev$^{41}$,
S.~Eidelman$^{43,v}$,
S.~Eisenhardt$^{58}$,
S.~Ek-In$^{49}$,
L.~Eklund$^{59,w}$,
S.~Ely$^{68}$,
A.~Ene$^{37}$,
E.~Epple$^{67}$,
S.~Escher$^{14}$,
J.~Eschle$^{50}$,
S.~Esen$^{32}$,
T.~Evans$^{48}$,
A.~Falabella$^{20}$,
J.~Fan$^{3}$,
Y.~Fan$^{6}$,
B.~Fang$^{73}$,
S.~Farry$^{60}$,
D.~Fazzini$^{26,j}$,
P.~Fedin$^{41}$,
M.~F{\'e}o$^{48}$,
P.~Fernandez~Declara$^{48}$,
A.~Fernandez~Prieto$^{46}$,
J.M.~Fernandez-tenllado~Arribas$^{45}$,
F.~Ferrari$^{20,d}$,
L.~Ferreira~Lopes$^{49}$,
F.~Ferreira~Rodrigues$^{2}$,
S.~Ferreres~Sole$^{32}$,
M.~Ferrillo$^{50}$,
M.~Ferro-Luzzi$^{48}$,
S.~Filippov$^{39}$,
R.A.~Fini$^{19}$,
M.~Fiorini$^{21,f}$,
M.~Firlej$^{34}$,
K.M.~Fischer$^{63}$,
C.~Fitzpatrick$^{62}$,
T.~Fiutowski$^{34}$,
F.~Fleuret$^{12}$,
M.~Fontana$^{13}$,
F.~Fontanelli$^{24,h}$,
R.~Forty$^{48}$,
V.~Franco~Lima$^{60}$,
M.~Franco~Sevilla$^{66}$,
M.~Frank$^{48}$,
E.~Franzoso$^{21}$,
G.~Frau$^{17}$,
C.~Frei$^{48}$,
D.A.~Friday$^{59}$,
J.~Fu$^{25}$,
Q.~Fuehring$^{15}$,
W.~Funk$^{48}$,
E.~Gabriel$^{32}$,
T.~Gaintseva$^{42}$,
A.~Gallas~Torreira$^{46}$,
D.~Galli$^{20,d}$,
S.~Gambetta$^{58,48}$,
Y.~Gan$^{3}$,
M.~Gandelman$^{2}$,
P.~Gandini$^{25}$,
Y.~Gao$^{5}$,
M.~Garau$^{27}$,
L.M.~Garcia~Martin$^{56}$,
P.~Garcia~Moreno$^{45}$,
J.~Garc{\'\i}a~Pardi{\~n}as$^{26,j}$,
B.~Garcia~Plana$^{46}$,
F.A.~Garcia~Rosales$^{12}$,
L.~Garrido$^{45}$,
C.~Gaspar$^{48}$,
R.E.~Geertsema$^{32}$,
D.~Gerick$^{17}$,
L.L.~Gerken$^{15}$,
E.~Gersabeck$^{62}$,
M.~Gersabeck$^{62}$,
T.~Gershon$^{56}$,
D.~Gerstel$^{10}$,
Ph.~Ghez$^{8}$,
V.~Gibson$^{55}$,
H.K.~Giemza$^{36}$,
M.~Giovannetti$^{23,p}$,
A.~Giovent{\`u}$^{46}$,
P.~Gironella~Gironell$^{45}$,
L.~Giubega$^{37}$,
C.~Giugliano$^{21,f,48}$,
K.~Gizdov$^{58}$,
E.L.~Gkougkousis$^{48}$,
V.V.~Gligorov$^{13}$,
C.~G{\"o}bel$^{70}$,
E.~Golobardes$^{84}$,
D.~Golubkov$^{41}$,
A.~Golutvin$^{61,82}$,
A.~Gomes$^{1,a}$,
S.~Gomez~Fernandez$^{45}$,
F.~Goncalves~Abrantes$^{63}$,
M.~Goncerz$^{35}$,
G.~Gong$^{3}$,
P.~Gorbounov$^{41}$,
I.V.~Gorelov$^{40}$,
C.~Gotti$^{26}$,
E.~Govorkova$^{48}$,
J.P.~Grabowski$^{17}$,
R.~Graciani~Diaz$^{45}$,
T.~Grammatico$^{13}$,
L.A.~Granado~Cardoso$^{48}$,
E.~Graug{\'e}s$^{45}$,
E.~Graverini$^{49}$,
G.~Graziani$^{22}$,
A.~Grecu$^{37}$,
L.M.~Greeven$^{32}$,
P.~Griffith$^{21,f}$,
L.~Grillo$^{62}$,
S.~Gromov$^{82}$,
B.R.~Gruberg~Cazon$^{63}$,
C.~Gu$^{3}$,
M.~Guarise$^{21}$,
P. A.~G{\"u}nther$^{17}$,
E.~Gushchin$^{39}$,
A.~Guth$^{14}$,
Y.~Guz$^{44,48}$,
T.~Gys$^{48}$,
T.~Hadavizadeh$^{69}$,
G.~Haefeli$^{49}$,
C.~Haen$^{48}$,
J.~Haimberger$^{48}$,
T.~Halewood-leagas$^{60}$,
P.M.~Hamilton$^{66}$,
Q.~Han$^{7}$,
X.~Han$^{17}$,
T.H.~Hancock$^{63}$,
S.~Hansmann-Menzemer$^{17}$,
N.~Harnew$^{63}$,
T.~Harrison$^{60}$,
C.~Hasse$^{48}$,
M.~Hatch$^{48}$,
J.~He$^{6,b}$,
M.~Hecker$^{61}$,
K.~Heijhoff$^{32}$,
K.~Heinicke$^{15}$,
A.M.~Hennequin$^{48}$,
K.~Hennessy$^{60}$,
L.~Henry$^{25,47}$,
J.~Heuel$^{14}$,
A.~Hicheur$^{2}$,
D.~Hill$^{49}$,
M.~Hilton$^{62}$,
S.E.~Hollitt$^{15}$,
J.~Hu$^{17}$,
J.~Hu$^{72}$,
W.~Hu$^{7}$,
W.~Huang$^{6}$,
X.~Huang$^{73}$,
W.~Hulsbergen$^{32}$,
R.J.~Hunter$^{56}$,
M.~Hushchyn$^{81}$,
D.~Hutchcroft$^{60}$,
D.~Hynds$^{32}$,
P.~Ibis$^{15}$,
M.~Idzik$^{34}$,
D.~Ilin$^{38}$,
P.~Ilten$^{65}$,
A.~Inglessi$^{38}$,
A.~Ishteev$^{82}$,
K.~Ivshin$^{38}$,
R.~Jacobsson$^{48}$,
S.~Jakobsen$^{48}$,
E.~Jans$^{32}$,
B.K.~Jashal$^{47}$,
A.~Jawahery$^{66}$,
V.~Jevtic$^{15}$,
M.~Jezabek$^{35}$,
F.~Jiang$^{3}$,
M.~John$^{63}$,
D.~Johnson$^{48}$,
C.R.~Jones$^{55}$,
T.P.~Jones$^{56}$,
B.~Jost$^{48}$,
N.~Jurik$^{48}$,
S.~Kandybei$^{51}$,
Y.~Kang$^{3}$,
M.~Karacson$^{48}$,
M.~Karpov$^{81}$,
N.~Kazeev$^{81}$,
F.~Keizer$^{55,48}$,
M.~Kenzie$^{56}$,
T.~Ketel$^{33}$,
B.~Khanji$^{15}$,
A.~Kharisova$^{83}$,
S.~Kholodenko$^{44}$,
K.E.~Kim$^{68}$,
T.~Kirn$^{14}$,
V.S.~Kirsebom$^{49}$,
O.~Kitouni$^{64}$,
S.~Klaver$^{32}$,
K.~Klimaszewski$^{36}$,
S.~Koliiev$^{52}$,
A.~Kondybayeva$^{82}$,
A.~Konoplyannikov$^{41}$,
P.~Kopciewicz$^{34}$,
R.~Kopecna$^{17}$,
P.~Koppenburg$^{32}$,
M.~Korolev$^{40}$,
I.~Kostiuk$^{32,52}$,
O.~Kot$^{52}$,
S.~Kotriakhova$^{38,30}$,
P.~Kravchenko$^{38}$,
L.~Kravchuk$^{39}$,
R.D.~Krawczyk$^{48}$,
M.~Kreps$^{56}$,
F.~Kress$^{61}$,
S.~Kretzschmar$^{14}$,
P.~Krokovny$^{43,v}$,
W.~Krupa$^{34}$,
W.~Krzemien$^{36}$,
W.~Kucewicz$^{35,t}$,
M.~Kucharczyk$^{35}$,
V.~Kudryavtsev$^{43,v}$,
H.S.~Kuindersma$^{32}$,
G.J.~Kunde$^{67}$,
T.~Kvaratskheliya$^{41}$,
D.~Lacarrere$^{48}$,
G.~Lafferty$^{62}$,
A.~Lai$^{27}$,
A.~Lampis$^{27}$,
D.~Lancierini$^{50}$,
J.J.~Lane$^{62}$,
R.~Lane$^{54}$,
G.~Lanfranchi$^{23}$,
C.~Langenbruch$^{14}$,
J.~Langer$^{15}$,
O.~Lantwin$^{50,82}$,
T.~Latham$^{56}$,
F.~Lazzari$^{29,q}$,
R.~Le~Gac$^{10}$,
S.H.~Lee$^{85}$,
R.~Lef{\`e}vre$^{9}$,
A.~Leflat$^{40}$,
S.~Legotin$^{82}$,
O.~Leroy$^{10}$,
T.~Lesiak$^{35}$,
B.~Leverington$^{17}$,
H.~Li$^{72}$,
L.~Li$^{63}$,
P.~Li$^{17}$,
Y.~Li$^{4}$,
Y.~Li$^{4}$,
Z.~Li$^{68}$,
X.~Liang$^{68}$,
T.~Lin$^{61}$,
R.~Lindner$^{48}$,
V.~Lisovskyi$^{15}$,
R.~Litvinov$^{27}$,
G.~Liu$^{72}$,
H.~Liu$^{6}$,
S.~Liu$^{4}$,
X.~Liu$^{3}$,
A.~Loi$^{27}$,
J.~Lomba~Castro$^{46}$,
I.~Longstaff$^{59}$,
J.H.~Lopes$^{2}$,
G.H.~Lovell$^{55}$,
Y.~Lu$^{4}$,
D.~Lucchesi$^{28,l}$,
S.~Luchuk$^{39}$,
M.~Lucio~Martinez$^{32}$,
V.~Lukashenko$^{32}$,
Y.~Luo$^{3}$,
A.~Lupato$^{62}$,
E.~Luppi$^{21,f}$,
O.~Lupton$^{56}$,
A.~Lusiani$^{29,m}$,
X.~Lyu$^{6}$,
L.~Ma$^{4}$,
R.~Ma$^{6}$,
S.~Maccolini$^{20,d}$,
F.~Machefert$^{11}$,
F.~Maciuc$^{37}$,
V.~Macko$^{49}$,
P.~Mackowiak$^{15}$,
S.~Maddrell-Mander$^{54}$,
O.~Madejczyk$^{34}$,
L.R.~Madhan~Mohan$^{54}$,
O.~Maev$^{38}$,
A.~Maevskiy$^{81}$,
D.~Maisuzenko$^{38}$,
M.W.~Majewski$^{34}$,
J.J.~Malczewski$^{35}$,
S.~Malde$^{63}$,
B.~Malecki$^{48}$,
A.~Malinin$^{80}$,
T.~Maltsev$^{43,v}$,
H.~Malygina$^{17}$,
G.~Manca$^{27,e}$,
G.~Mancinelli$^{10}$,
R.~Manera~Escalero$^{45}$,
D.~Manuzzi$^{20,d}$,
D.~Marangotto$^{25,i}$,
J.~Maratas$^{9,s}$,
J.F.~Marchand$^{8}$,
U.~Marconi$^{20}$,
S.~Mariani$^{22,g,48}$,
C.~Marin~Benito$^{11}$,
M.~Marinangeli$^{49}$,
P.~Marino$^{49,m}$,
J.~Marks$^{17}$,
P.J.~Marshall$^{60}$,
G.~Martellotti$^{30}$,
L.~Martinazzoli$^{48,j}$,
M.~Martinelli$^{26,j}$,
D.~Martinez~Santos$^{46}$,
F.~Martinez~Vidal$^{47}$,
A.~Massafferri$^{1}$,
M.~Materok$^{14}$,
R.~Matev$^{48}$,
A.~Mathad$^{50}$,
Z.~Mathe$^{48}$,
V.~Matiunin$^{41}$,
C.~Matteuzzi$^{26}$,
K.R.~Mattioli$^{85}$,
A.~Mauri$^{32}$,
E.~Maurice$^{12}$,
J.~Mauricio$^{45}$,
M.~Mazurek$^{36}$,
M.~McCann$^{61}$,
L.~Mcconnell$^{18}$,
T.H.~Mcgrath$^{62}$,
A.~McNab$^{62}$,
R.~McNulty$^{18}$,
J.V.~Mead$^{60}$,
B.~Meadows$^{65}$,
C.~Meaux$^{10}$,
G.~Meier$^{15}$,
N.~Meinert$^{76}$,
D.~Melnychuk$^{36}$,
S.~Meloni$^{26,j}$,
M.~Merk$^{32,79}$,
A.~Merli$^{25}$,
L.~Meyer~Garcia$^{2}$,
M.~Mikhasenko$^{48}$,
D.A.~Milanes$^{74}$,
E.~Millard$^{56}$,
M.~Milovanovic$^{48}$,
M.-N.~Minard$^{8}$,
L.~Minzoni$^{21,f}$,
S.E.~Mitchell$^{58}$,
B.~Mitreska$^{62}$,
D.S.~Mitzel$^{48}$,
A.~M{\"o}dden~$^{15}$,
R.A.~Mohammed$^{63}$,
R.D.~Moise$^{61}$,
T.~Momb{\"a}cher$^{15}$,
I.A.~Monroy$^{74}$,
S.~Monteil$^{9}$,
M.~Morandin$^{28}$,
G.~Morello$^{23}$,
M.J.~Morello$^{29,m}$,
J.~Moron$^{34}$,
A.B.~Morris$^{75}$,
A.G.~Morris$^{56}$,
R.~Mountain$^{68}$,
H.~Mu$^{3}$,
F.~Muheim$^{58,48}$,
M.~Mukherjee$^{7}$,
M.~Mulder$^{48}$,
D.~M{\"u}ller$^{48}$,
K.~M{\"u}ller$^{50}$,
C.H.~Murphy$^{63}$,
D.~Murray$^{62}$,
P.~Muzzetto$^{27,48}$,
P.~Naik$^{54}$,
T.~Nakada$^{49}$,
R.~Nandakumar$^{57}$,
T.~Nanut$^{49}$,
I.~Nasteva$^{2}$,
M.~Needham$^{58}$,
I.~Neri$^{21}$,
N.~Neri$^{25,i}$,
S.~Neubert$^{75}$,
N.~Neufeld$^{48}$,
R.~Newcombe$^{61}$,
T.D.~Nguyen$^{49}$,
C.~Nguyen-Mau$^{49,x}$,
E.M.~Niel$^{11}$,
S.~Nieswand$^{14}$,
N.~Nikitin$^{40}$,
N.S.~Nolte$^{48}$,
C.~Nunez$^{85}$,
A.~Oblakowska-Mucha$^{34}$,
V.~Obraztsov$^{44}$,
D.P.~O'Hanlon$^{54}$,
R.~Oldeman$^{27,e}$,
M.E.~Olivares$^{68}$,
C.J.G.~Onderwater$^{78}$,
A.~Ossowska$^{35}$,
J.M.~Otalora~Goicochea$^{2}$,
T.~Ovsiannikova$^{41}$,
P.~Owen$^{50}$,
A.~Oyanguren$^{47}$,
B.~Pagare$^{56}$,
P.R.~Pais$^{48}$,
T.~Pajero$^{29,m,48}$,
A.~Palano$^{19}$,
M.~Palutan$^{23}$,
Y.~Pan$^{62}$,
G.~Panshin$^{83}$,
A.~Papanestis$^{57}$,
M.~Pappagallo$^{19,c}$,
L.L.~Pappalardo$^{21,f}$,
C.~Pappenheimer$^{65}$,
W.~Parker$^{66}$,
C.~Parkes$^{62}$,
C.J.~Parkinson$^{46}$,
B.~Passalacqua$^{21}$,
G.~Passaleva$^{22}$,
A.~Pastore$^{19}$,
M.~Patel$^{61}$,
C.~Patrignani$^{20,d}$,
C.J.~Pawley$^{79}$,
A.~Pearce$^{48}$,
A.~Pellegrino$^{32}$,
M.~Pepe~Altarelli$^{48}$,
S.~Perazzini$^{20}$,
D.~Pereima$^{41}$,
P.~Perret$^{9}$,
K.~Petridis$^{54}$,
A.~Petrolini$^{24,h}$,
A.~Petrov$^{80}$,
S.~Petrucci$^{58}$,
M.~Petruzzo$^{25}$,
T.T.H.~Pham$^{68}$,
A.~Philippov$^{42}$,
L.~Pica$^{29,n}$,
M.~Piccini$^{77}$,
B.~Pietrzyk$^{8}$,
G.~Pietrzyk$^{49}$,
M.~Pili$^{63}$,
D.~Pinci$^{30}$,
F.~Pisani$^{48}$,
A.~Piucci$^{17}$,
Resmi ~P.K$^{10}$,
V.~Placinta$^{37}$,
J.~Plews$^{53}$,
M.~Plo~Casasus$^{46}$,
F.~Polci$^{13}$,
M.~Poli~Lener$^{23}$,
M.~Poliakova$^{68}$,
A.~Poluektov$^{10}$,
N.~Polukhina$^{82,u}$,
I.~Polyakov$^{68}$,
E.~Polycarpo$^{2}$,
G.J.~Pomery$^{54}$,
S.~Ponce$^{48}$,
D.~Popov$^{6,48}$,
S.~Popov$^{42}$,
S.~Poslavskii$^{44}$,
K.~Prasanth$^{35}$,
L.~Promberger$^{48}$,
C.~Prouve$^{46}$,
V.~Pugatch$^{52}$,
H.~Pullen$^{63}$,
G.~Punzi$^{29,n}$,
W.~Qian$^{6}$,
J.~Qin$^{6}$,
R.~Quagliani$^{13}$,
B.~Quintana$^{8}$,
N.V.~Raab$^{18}$,
R.I.~Rabadan~Trejo$^{10}$,
B.~Rachwal$^{34}$,
J.H.~Rademacker$^{54}$,
M.~Rama$^{29}$,
M.~Ramos~Pernas$^{56}$,
M.S.~Rangel$^{2}$,
F.~Ratnikov$^{42,81}$,
G.~Raven$^{33}$,
M.~Reboud$^{8}$,
F.~Redi$^{49}$,
F.~Reiss$^{13}$,
C.~Remon~Alepuz$^{47}$,
Z.~Ren$^{3}$,
V.~Renaudin$^{63}$,
R.~Ribatti$^{29}$,
S.~Ricciardi$^{57}$,
K.~Rinnert$^{60}$,
P.~Robbe$^{11}$,
A.~Robert$^{13}$,
G.~Robertson$^{58}$,
A.B.~Rodrigues$^{49}$,
E.~Rodrigues$^{60}$,
J.A.~Rodriguez~Lopez$^{74}$,
A.~Rollings$^{63}$,
P.~Roloff$^{48}$,
V.~Romanovskiy$^{44}$,
M.~Romero~Lamas$^{46}$,
A.~Romero~Vidal$^{46}$,
J.D.~Roth$^{85}$,
M.~Rotondo$^{23}$,
M.S.~Rudolph$^{68}$,
T.~Ruf$^{48}$,
J.~Ruiz~Vidal$^{47}$,
A.~Ryzhikov$^{81}$,
J.~Ryzka$^{34}$,
J.J.~Saborido~Silva$^{46}$,
N.~Sagidova$^{38}$,
N.~Sahoo$^{56}$,
B.~Saitta$^{27,e}$,
D.~Sanchez~Gonzalo$^{45}$,
C.~Sanchez~Gras$^{32}$,
R.~Santacesaria$^{30}$,
C.~Santamarina~Rios$^{46}$,
M.~Santimaria$^{23}$,
E.~Santovetti$^{31,p}$,
D.~Saranin$^{82}$,
G.~Sarpis$^{59}$,
M.~Sarpis$^{75}$,
A.~Sarti$^{30}$,
C.~Satriano$^{30,o}$,
A.~Satta$^{31}$,
M.~Saur$^{15}$,
D.~Savrina$^{41,40}$,
H.~Sazak$^{9}$,
L.G.~Scantlebury~Smead$^{63}$,
S.~Schael$^{14}$,
M.~Schellenberg$^{15}$,
M.~Schiller$^{59}$,
H.~Schindler$^{48}$,
M.~Schmelling$^{16}$,
B.~Schmidt$^{48}$,
O.~Schneider$^{49}$,
A.~Schopper$^{48}$,
M.~Schubiger$^{32}$,
S.~Schulte$^{49}$,
M.H.~Schune$^{11}$,
R.~Schwemmer$^{48}$,
B.~Sciascia$^{23}$,
A.~Sciubba$^{23}$,
S.~Sellam$^{46}$,
A.~Semennikov$^{41}$,
M.~Senghi~Soares$^{33}$,
A.~Sergi$^{24,48}$,
N.~Serra$^{50}$,
L.~Sestini$^{28}$,
A.~Seuthe$^{15}$,
P.~Seyfert$^{48}$,
D.M.~Shangase$^{85}$,
M.~Shapkin$^{44}$,
I.~Shchemerov$^{82}$,
L.~Shchutska$^{49}$,
T.~Shears$^{60}$,
L.~Shekhtman$^{43,v}$,
Z.~Shen$^{5}$,
V.~Shevchenko$^{80}$,
E.B.~Shields$^{26,j}$,
E.~Shmanin$^{82}$,
J.D.~Shupperd$^{68}$,
B.G.~Siddi$^{21}$,
R.~Silva~Coutinho$^{50}$,
G.~Simi$^{28}$,
S.~Simone$^{19,c}$,
N.~Skidmore$^{62}$,
T.~Skwarnicki$^{68}$,
M.W.~Slater$^{53}$,
I.~Slazyk$^{21,f}$,
J.C.~Smallwood$^{63}$,
J.G.~Smeaton$^{55}$,
A.~Smetkina$^{41}$,
E.~Smith$^{14}$,
M.~Smith$^{61}$,
A.~Snoch$^{32}$,
M.~Soares$^{20}$,
L.~Soares~Lavra$^{9}$,
M.D.~Sokoloff$^{65}$,
F.J.P.~Soler$^{59}$,
A.~Solovev$^{38}$,
I.~Solovyev$^{38}$,
F.L.~Souza~De~Almeida$^{2}$,
B.~Souza~De~Paula$^{2}$,
B.~Spaan$^{15}$,
E.~Spadaro~Norella$^{25,i}$,
P.~Spradlin$^{59}$,
F.~Stagni$^{48}$,
M.~Stahl$^{65}$,
S.~Stahl$^{48}$,
P.~Stefko$^{49}$,
O.~Steinkamp$^{50,82}$,
S.~Stemmle$^{17}$,
O.~Stenyakin$^{44}$,
H.~Stevens$^{15}$,
S.~Stone$^{68}$,
M.E.~Stramaglia$^{49}$,
M.~Straticiuc$^{37}$,
D.~Strekalina$^{82}$,
F.~Suljik$^{63}$,
J.~Sun$^{27}$,
L.~Sun$^{73}$,
Y.~Sun$^{66}$,
P.~Svihra$^{62}$,
P.N.~Swallow$^{53}$,
K.~Swientek$^{34}$,
A.~Szabelski$^{36}$,
T.~Szumlak$^{34}$,
M.~Szymanski$^{48}$,
S.~Taneja$^{62}$,
F.~Teubert$^{48}$,
E.~Thomas$^{48}$,
K.A.~Thomson$^{60}$,
M.J.~Tilley$^{61}$,
V.~Tisserand$^{9}$,
S.~T'Jampens$^{8}$,
M.~Tobin$^{4}$,
S.~Tolk$^{48}$,
L.~Tomassetti$^{21,f}$,
D.~Torres~Machado$^{1}$,
D.Y.~Tou$^{13}$,
M.~Traill$^{59}$,
M.T.~Tran$^{49}$,
E.~Trifonova$^{82}$,
C.~Trippl$^{49}$,
G.~Tuci$^{29,n}$,
A.~Tully$^{49}$,
N.~Tuning$^{32,48}$,
A.~Ukleja$^{36}$,
D.J.~Unverzagt$^{17}$,
E.~Ursov$^{82}$,
A.~Usachov$^{32}$,
A.~Ustyuzhanin$^{42,81}$,
U.~Uwer$^{17}$,
A.~Vagner$^{83}$,
V.~Vagnoni$^{20}$,
A.~Valassi$^{48}$,
G.~Valenti$^{20}$,
N.~Valls~Canudas$^{45}$,
M.~van~Beuzekom$^{32}$,
M.~Van~Dijk$^{49}$,
E.~van~Herwijnen$^{82}$,
C.B.~Van~Hulse$^{18}$,
M.~van~Veghel$^{78}$,
R.~Vazquez~Gomez$^{46}$,
P.~Vazquez~Regueiro$^{46}$,
C.~V{\'a}zquez~Sierra$^{48}$,
S.~Vecchi$^{21}$,
J.J.~Velthuis$^{54}$,
M.~Veltri$^{22,r}$,
A.~Venkateswaran$^{68}$,
M.~Veronesi$^{32}$,
M.~Vesterinen$^{56}$,
D.~~Vieira$^{65}$,
M.~Vieites~Diaz$^{49}$,
H.~Viemann$^{76}$,
X.~Vilasis-Cardona$^{84}$,
E.~Vilella~Figueras$^{60}$,
P.~Vincent$^{13}$,
G.~Vitali$^{29}$,
A.~Vollhardt$^{50}$,
D.~Vom~Bruch$^{10}$,
A.~Vorobyev$^{38}$,
V.~Vorobyev$^{43,v}$,
N.~Voropaev$^{38}$,
R.~Waldi$^{76}$,
J.~Walsh$^{29}$,
C.~Wang$^{17}$,
J.~Wang$^{5}$,
J.~Wang$^{4}$,
J.~Wang$^{3}$,
J.~Wang$^{73}$,
M.~Wang$^{3}$,
R.~Wang$^{54}$,
Y.~Wang$^{7}$,
Z.~Wang$^{50}$,
H.M.~Wark$^{60}$,
N.K.~Watson$^{53}$,
S.G.~Weber$^{13}$,
D.~Websdale$^{61}$,
C.~Weisser$^{64}$,
B.D.C.~Westhenry$^{54}$,
D.J.~White$^{62}$,
M.~Whitehead$^{54}$,
D.~Wiedner$^{15}$,
G.~Wilkinson$^{63}$,
M.~Wilkinson$^{68}$,
I.~Williams$^{55}$,
M.~Williams$^{64,69}$,
M.R.J.~Williams$^{58}$,
F.F.~Wilson$^{57}$,
W.~Wislicki$^{36}$,
M.~Witek$^{35}$,
L.~Witola$^{17}$,
G.~Wormser$^{11}$,
S.A.~Wotton$^{55}$,
H.~Wu$^{68}$,
K.~Wyllie$^{48}$,
Z.~Xiang$^{6}$,
D.~Xiao$^{7}$,
Y.~Xie$^{7}$,
A.~Xu$^{5}$,
J.~Xu$^{6}$,
L.~Xu$^{3}$,
M.~Xu$^{7}$,
Q.~Xu$^{6}$,
Z.~Xu$^{5}$,
Z.~Xu$^{6}$,
D.~Yang$^{3}$,
S.~Yang$^{6}$,
Y.~Yang$^{6}$,
Z.~Yang$^{3}$,
Z.~Yang$^{66}$,
Y.~Yao$^{68}$,
L.E.~Yeomans$^{60}$,
H.~Yin$^{7}$,
J.~Yu$^{71}$,
X.~Yuan$^{68}$,
O.~Yushchenko$^{44}$,
E.~Zaffaroni$^{49}$,
K.A.~Zarebski$^{53}$,
M.~Zavertyaev$^{16,u}$,
M.~Zdybal$^{35}$,
O.~Zenaiev$^{48}$,
M.~Zeng$^{3}$,
D.~Zhang$^{7}$,
L.~Zhang$^{3}$,
S.~Zhang$^{5}$,
Y.~Zhang$^{5}$,
Y.~Zhang$^{63}$,
A.~Zhelezov$^{17}$,
Y.~Zheng$^{6}$,
X.~Zhou$^{6}$,
Y.~Zhou$^{6}$,
X.~Zhu$^{3}$,
V.~Zhukov$^{14,40}$,
J.B.~Zonneveld$^{58}$,
S.~Zucchelli$^{20,d}$,
D.~Zuliani$^{28}$,
G.~Zunica$^{62}$.\bigskip
{\footnotesize \it
$^{1}$Centro Brasileiro de Pesquisas F{\'\i}sicas (CBPF), Rio de Janeiro, Brazil\\
$^{2}$Universidade Federal do Rio de Janeiro (UFRJ), Rio de Janeiro, Brazil\\
$^{3}$Center for High Energy Physics, Tsinghua University, Beijing, China\\
$^{4}$Institute Of High Energy Physics (IHEP), Beijing, China\\
$^{5}$School of Physics State Key Laboratory of Nuclear Physics and Technology, Peking University, Beijing, China\\
$^{6}$University of Chinese Academy of Sciences, Beijing, China\\
$^{7}$Institute of Particle Physics, Central China Normal University, Wuhan, Hubei, China\\
$^{8}$Univ. Grenoble Alpes, Univ. Savoie Mont Blanc, CNRS, IN2P3-LAPP, Annecy, France\\
$^{9}$Universit{\'e} Clermont Auvergne, CNRS/IN2P3, LPC, Clermont-Ferrand, France\\
$^{10}$Aix Marseille Univ, CNRS/IN2P3, CPPM, Marseille, France\\
$^{11}$Universit{\'e} Paris-Saclay, CNRS/IN2P3, IJCLab, Orsay, France\\
$^{12}$Laboratoire Leprince-Ringuet, CNRS/IN2P3, Ecole Polytechnique, Institut Polytechnique de Paris, Palaiseau, France\\
$^{13}$LPNHE, Sorbonne Universit{\'e}, Paris Diderot Sorbonne Paris Cit{\'e}, CNRS/IN2P3, Paris, France\\
$^{14}$I. Physikalisches Institut, RWTH Aachen University, Aachen, Germany\\
$^{15}$Fakult{\"a}t Physik, Technische Universit{\"a}t Dortmund, Dortmund, Germany\\
$^{16}$Max-Planck-Institut f{\"u}r Kernphysik (MPIK), Heidelberg, Germany\\
$^{17}$Physikalisches Institut, Ruprecht-Karls-Universit{\"a}t Heidelberg, Heidelberg, Germany\\
$^{18}$School of Physics, University College Dublin, Dublin, Ireland\\
$^{19}$INFN Sezione di Bari, Bari, Italy\\
$^{20}$INFN Sezione di Bologna, Bologna, Italy\\
$^{21}$INFN Sezione di Ferrara, Ferrara, Italy\\
$^{22}$INFN Sezione di Firenze, Firenze, Italy\\
$^{23}$INFN Laboratori Nazionali di Frascati, Frascati, Italy\\
$^{24}$INFN Sezione di Genova, Genova, Italy\\
$^{25}$INFN Sezione di Milano, Milano, Italy\\
$^{26}$INFN Sezione di Milano-Bicocca, Milano, Italy\\
$^{27}$INFN Sezione di Cagliari, Monserrato, Italy\\
$^{28}$Universita degli Studi di Padova, Universita e INFN, Padova, Padova, Italy\\
$^{29}$INFN Sezione di Pisa, Pisa, Italy\\
$^{30}$INFN Sezione di Roma La Sapienza, Roma, Italy\\
$^{31}$INFN Sezione di Roma Tor Vergata, Roma, Italy\\
$^{32}$Nikhef National Institute for Subatomic Physics, Amsterdam, Netherlands\\
$^{33}$Nikhef National Institute for Subatomic Physics and VU University Amsterdam, Amsterdam, Netherlands\\
$^{34}$AGH - University of Science and Technology, Faculty of Physics and Applied Computer Science, Krak{\'o}w, Poland\\
$^{35}$Henryk Niewodniczanski Institute of Nuclear Physics Polish Academy of Sciences, Krak{\'o}w, Poland\\
$^{36}$National Center for Nuclear Research (NCBJ), Warsaw, Poland\\
$^{37}$Horia Hulubei National Institute of Physics and Nuclear Engineering, Bucharest-Magurele, Romania\\
$^{38}$Petersburg Nuclear Physics Institute NRC Kurchatov Institute (PNPI NRC KI), Gatchina, Russia\\
$^{39}$Institute for Nuclear Research of the Russian Academy of Sciences (INR RAS), Moscow, Russia\\
$^{40}$Institute of Nuclear Physics, Moscow State University (SINP MSU), Moscow, Russia\\
$^{41}$Institute of Theoretical and Experimental Physics NRC Kurchatov Institute (ITEP NRC KI), Moscow, Russia\\
$^{42}$Yandex School of Data Analysis, Moscow, Russia\\
$^{43}$Budker Institute of Nuclear Physics (SB RAS), Novosibirsk, Russia\\
$^{44}$Institute for High Energy Physics NRC Kurchatov Institute (IHEP NRC KI), Protvino, Russia, Protvino, Russia\\
$^{45}$ICCUB, Universitat de Barcelona, Barcelona, Spain\\
$^{46}$Instituto Galego de F{\'\i}sica de Altas Enerx{\'\i}as (IGFAE), Universidade de Santiago de Compostela, Santiago de Compostela, Spain\\
$^{47}$Instituto de Fisica Corpuscular, Centro Mixto Universidad de Valencia - CSIC, Valencia, Spain\\
$^{48}$European Organization for Nuclear Research (CERN), Geneva, Switzerland\\
$^{49}$Institute of Physics, Ecole Polytechnique F{\'e}d{\'e}rale de Lausanne (EPFL), Lausanne, Switzerland\\
$^{50}$Physik-Institut, Universit{\"a}t Z{\"u}rich, Z{\"u}rich, Switzerland\\
$^{51}$NSC Kharkiv Institute of Physics and Technology (NSC KIPT), Kharkiv, Ukraine\\
$^{52}$Institute for Nuclear Research of the National Academy of Sciences (KINR), Kyiv, Ukraine\\
$^{53}$University of Birmingham, Birmingham, United Kingdom\\
$^{54}$H.H. Wills Physics Laboratory, University of Bristol, Bristol, United Kingdom\\
$^{55}$Cavendish Laboratory, University of Cambridge, Cambridge, United Kingdom\\
$^{56}$Department of Physics, University of Warwick, Coventry, United Kingdom\\
$^{57}$STFC Rutherford Appleton Laboratory, Didcot, United Kingdom\\
$^{58}$School of Physics and Astronomy, University of Edinburgh, Edinburgh, United Kingdom\\
$^{59}$School of Physics and Astronomy, University of Glasgow, Glasgow, United Kingdom\\
$^{60}$Oliver Lodge Laboratory, University of Liverpool, Liverpool, United Kingdom\\
$^{61}$Imperial College London, London, United Kingdom\\
$^{62}$Department of Physics and Astronomy, University of Manchester, Manchester, United Kingdom\\
$^{63}$Department of Physics, University of Oxford, Oxford, United Kingdom\\
$^{64}$Massachusetts Institute of Technology, Cambridge, MA, United States\\
$^{65}$University of Cincinnati, Cincinnati, OH, United States\\
$^{66}$University of Maryland, College Park, MD, United States\\
$^{67}$Los Alamos National Laboratory (LANL), Los Alamos, United States\\
$^{68}$Syracuse University, Syracuse, NY, United States\\
$^{69}$School of Physics and Astronomy, Monash University, Melbourne, Australia, associated to $^{56}$\\
$^{70}$Pontif{\'\i}cia Universidade Cat{\'o}lica do Rio de Janeiro (PUC-Rio), Rio de Janeiro, Brazil, associated to $^{2}$\\
$^{71}$Physics and Micro Electronic College, Hunan University, Changsha City, China, associated to $^{7}$\\
$^{72}$Guangdong Provencial Key Laboratory of Nuclear Science, Institute of Quantum Matter, South China Normal University, Guangzhou, China, associated to $^{3}$\\
$^{73}$School of Physics and Technology, Wuhan University, Wuhan, China, associated to $^{3}$\\
$^{74}$Departamento de Fisica , Universidad Nacional de Colombia, Bogota, Colombia, associated to $^{13}$\\
$^{75}$Universit{\"a}t Bonn - Helmholtz-Institut f{\"u}r Strahlen und Kernphysik, Bonn, Germany, associated to $^{17}$\\
$^{76}$Institut f{\"u}r Physik, Universit{\"a}t Rostock, Rostock, Germany, associated to $^{17}$\\
$^{77}$INFN Sezione di Perugia, Perugia, Italy, associated to $^{21}$\\
$^{78}$Van Swinderen Institute, University of Groningen, Groningen, Netherlands, associated to $^{32}$\\
$^{79}$Universiteit Maastricht, Maastricht, Netherlands, associated to $^{32}$\\
$^{80}$National Research Centre Kurchatov Institute, Moscow, Russia, associated to $^{41}$\\
$^{81}$National Research University Higher School of Economics, Moscow, Russia, associated to $^{42}$\\
$^{82}$National University of Science and Technology ``MISIS'', Moscow, Russia, associated to $^{41}$\\
$^{83}$National Research Tomsk Polytechnic University, Tomsk, Russia, associated to $^{41}$\\
$^{84}$DS4DS, La Salle, Universitat Ramon Llull, Barcelona, Spain, associated to $^{45}$\\
$^{85}$University of Michigan, Ann Arbor, United States, associated to $^{68}$\\
\bigskip
$^{a}$Universidade Federal do Tri{\^a}ngulo Mineiro (UFTM), Uberaba-MG, Brazil\\
$^{b}$Hangzhou Institute for Advanced Study, UCAS, Hangzhou, China\\
$^{c}$Universit{\`a} di Bari, Bari, Italy\\
$^{d}$Universit{\`a} di Bologna, Bologna, Italy\\
$^{e}$Universit{\`a} di Cagliari, Cagliari, Italy\\
$^{f}$Universit{\`a} di Ferrara, Ferrara, Italy\\
$^{g}$Universit{\`a} di Firenze, Firenze, Italy\\
$^{h}$Universit{\`a} di Genova, Genova, Italy\\
$^{i}$Universit{\`a} degli Studi di Milano, Milano, Italy\\
$^{j}$Universit{\`a} di Milano Bicocca, Milano, Italy\\
$^{k}$Universit{\`a} di Modena e Reggio Emilia, Modena, Italy\\
$^{l}$Universit{\`a} di Padova, Padova, Italy\\
$^{m}$Scuola Normale Superiore, Pisa, Italy\\
$^{n}$Universit{\`a} di Pisa, Pisa, Italy\\
$^{o}$Universit{\`a} della Basilicata, Potenza, Italy\\
$^{p}$Universit{\`a} di Roma Tor Vergata, Roma, Italy\\
$^{q}$Universit{\`a} di Siena, Siena, Italy\\
$^{r}$Universit{\`a} di Urbino, Urbino, Italy\\
$^{s}$MSU - Iligan Institute of Technology (MSU-IIT), Iligan, Philippines\\
$^{t}$AGH - University of Science and Technology, Faculty of Computer Science, Electronics and Telecommunications, Krak{\'o}w, Poland\\
$^{u}$P.N. Lebedev Physical Institute, Russian Academy of Science (LPI RAS), Moscow, Russia\\
$^{v}$Novosibirsk State University, Novosibirsk, Russia\\
$^{w}$Department of Physics and Astronomy, Uppsala University, Uppsala, Sweden\\
$^{x}$Hanoi University of Science, Hanoi, Vietnam\\
\medskip
}
\end{flushleft}
\end{document} |
\section{Introduction and Main Results}
\label{sec:introduction_and_main_results}
In this article we study aspects related to the thermodynamic formalism for systems having some hyperbolicity, as the time-one map of an hyperbolic flow, or an Anosov diffeomorphism. This is one of the most important pieces in smooth ergodic theory, encompassing several examples and general theorems. Although this part of the theory is fairly complete, the results discussed here are not present in the literature (in this generality, as far as we know), and therefore the present project has a two-fold purpose; on the one hand, introducing the aforementioned results, and on the other serve a (crude) blueprint for future research in more general cases.
We will rely on our previous work \cite{EqStatesCenter}, and in fact the present manuscript can be seen as a complement to this other one. Nevertheless, we will give a self-contained presentation and no a priori knowledge of this other work is necessary.
By an hyperbolic flow we mean a ($\mathcal{C}^1$) flow $\Phi=\{\Phi_t:M\to M\}_{t\in\Real}$ of a closed manifold satisfying: there exists a Riemannian metric on $M$, a splitting of the form $TM=E^{s}\oplus E^c\oplus E^{u}$, and $0<\lambda<1$ such that the following conditions are true.
\begin{enumerate}
\item $E^c$ is generated by the vector field tangent to the flow.
\item $D_x\Phi_tE^{s/u}(x)=E^{s/u}(\Phi_t(x))$, for all $t\in\Real$, for every $x\in M$.
\item $\sup_{x\in M}\norm{D_x\Phi_t|E^{s}(x)}, \sup_{x\in M}\norm{D_x\Phi_{-t}|E^{u}(x)}\leq \lambda^t$, for all $t\geq 0$.
\end{enumerate}
We call $E^{s}, E^{u}$ the strong stable and unstable bundles, respectively, and $E^c$ will be referred as the center bundle. An hyperbolic flow is said to be of codimension-one if $\dim E^{u}=1$. The prototypical example of codimension-one hyperbolic flow is the geodesic flow corresponding to a (closed) surface of negative sectional curvate; the reader can safely assume this being the case for the rest of the article, since this is probably the most interesting case where our results apply.
The concept of hyperbolic diffeomorphism is analogous; we say that a diffeomorphism $f:M\to M$ of an closed manifold $M$ is hyperbolic (or Anosov) if there exists a Riemannian metric on $M$, a splitting of the form $TM=E^{s}\oplus E^{u}$, and $0<\lambda<1$ such that the following conditions are true.
\begin{enumerate}
\item $D_xfE^{s/u}(x)=E^{s/u}(f(x))$, for every $x\in M$.
\item $\sup_{x\in M}\norm{D_xf|E^{s}(x)}, \sup_{x\in M}\norm{D_xf|E^{u}(x)}\leq \lambda$.
\end{enumerate}
\begin{remark}
For the rest of the article we will work only with (codimension-one) hyperbolic flows, leaving the direct modifications for the diffeomorphism case to the reader.
\end{remark}
Now fix a codimension-one hyperbolic flow; by the classical Hadamard-Perron's stable manifold theorem, it follows in this setting that $E^u$ integrates, provided that it is orientable, to a (in general, only transversally continuous) flow $\Phi^u=\{\Phi^u_t:M\to M\}_{t\in \Real}$, called the horocyclic flow. It turns out that if we further assume transitivity of $\Phi$, meaning the existence of a dense orbit, there are only two possibilities \cite{AnosovFlow}; either
\begin{enumerate}
\item $M$ is a fiber bundle over $S^1$, with fibers transverse to the flow line (the suspension case), or
\item $\Phi^u$ is minimal, that is, every orbit of $\Phi^u$ is dense in $M$
\end{enumerate}
The geodesic flow corresponding to a negatively curved surface falls into this last category.
\smallskip
\noindent\textbf{Convention.} For the rest of this article we assume that the unstable bundle of any considered hyperbolic flow is orientable. Modulo a two-fold covering, this not a significative loss of generality.
\smallskip
We then have the following.
\begin{theorem}
If $\Phi$ is a (transitive) codimension one Anosov flow that is not a suspension, then $\Phi^u$ is uniquely ergodic. That is, there exists only one (probability) measure invariant under $\Phi^u$.
\end{theorem}
Transverse and invariant measures are in one-to-one correspondence for flows, so the above theorem says that there is only one transverse invariant measure for the flow foliation. The previous theorem is an important result originally due to H. Furstenberg \cite{Furstenberg1973} for the geodesic flow corresponding to hyperbolic surfaces, and later generalized by B. Marcus \cite{Marcus1975} to the variable curvature case, and in fact to arbitrary codimension-one Anosov flows (that are not suspension, that is). The most general result is due to R. Bowen and B. Marcus, where they do not assume one-dimensionality of the unstable bundle \cite{Bowen1977}.
Here we generalize the above result. Let us introduce some definitions.
\begin{definition}\label{def:transverse measure}
Let $\Psi$ be a flow on $M$. A (local) transversal to the flow is a sub-manifold $D\subset M$ such that is transverse to the flow direction, for every $x\in D$, and a pairwise disjoint family $\mathscr{D}=\{D\}_{\scriptstyle D\ \mathrm{transversal}}$ is a global transversal if given $x\in M$ there exists $t\in\Real$ and $D\in \mathscr{D}$ such that $\Phi_t(x)\in D$. The global transversal $\mathscr{D}$ is full if for every $x\in M$ there exists $D\in\mathscr{D}$ with $x\in D$.
A transverse measure for $\Phi$ is a family of measures $\nu=\{\nu_D: D\in\mathscr{D}\}$ where $\mathscr{D}$ is a global transversal to $\Phi$ and for each $D\in \mathscr{D}$, $\nu_D$ is a Radon measure on $D$: $\mathscr{D}$ is the support of $\nu$. We will assume that the support of the transverse measures considered is full\footnote{This is not loss of generality for our purposes.}
\end{definition}
If $\nu=\{\nu_D: D\in\mathscr{D}\}$ is a transverse measure to $\Phi$ we also write $\nu=\{\nu_x\}_{x\in M}$ where
$\nu_x=\nu_{D}$.
\begin{definition}
Let $\Psi$ be a flow on $M$ and $\{\nu_x\}_x$ be a transverse measure for $\Psi$. We say that $\{\nu_x\}_x$ is quasi-invariant if there exists a family of positive functions $\mathrm{Jac}_{\nu}=\{\mathrm{Jac}_{x_0,y_0}:M\to \Real_{>0}:y_0=\Psi_{t_0}x_0, x_0\in M, t_0\in\Real\}$ and $C>0$ such that
\[
\Psi_{-t_0}\nu_{y_0}= C\mathrm{Jac}_{x_0,y_0}\nu_{x_0}.
\]
In this case $\mathrm{Jac}_{\nu}$ is the Jacobian of the quasi-invariant measure, whereas $C$ is the associated normalization constant.
\end{definition}
An invariant (transverse) measure corresponds to $\mathrm{Jac}_{\nu}\equiv 1$. We will be interested in quasi-invariant transverse measures for $\Phi^u$ having well-behaved Jacobian, and having some relation with the dynamics $\Phi$. As a convenient family of transversals for the horocyclic flow we consider the leaves of the center stable foliation $\Fcs=\{W^{cs}(x)\}_{x\in M}$, whose leaves are tangent to $E^{cs}=E^{c}\oplus E^s$. This choice of transversals is implicit in what follows, and in particular transverse measures to $\Phi^u$ are assumed to be supported in center stable leaves. The existence and basic properties of invariant foliations associated to $\Phi$ is recalled in Section 2.
We first state the following.
\begin{proposition}[Proposition 2.10 in \cite{EqStatesCenter}]\label{pro:quasinvariancecocycle}
Suppose that $\mu^{cs}=\{\mu^{cs}_x\}$ is a transverse measure for $\Phi^u$ such that $\forall x\in M$
\[
\Phi_{-1}\mu^{cs}_{\Phi_1(x)}=C h\mu_{x}^{cs},
\]
where $C>0$ and $h:M\to\Real_{+}$ is H\"older. Then $\mu^{cs}$ is $\Phi^u$-quasi invariant, with Jacobian
\[
\Jacu[x_0,y_0](x)=\prod_{j=1}^{\oo}\frac{h(\Phi_{-j}\circ \Phi^u_{t(x)} x)}{h(\Phi_{-j}(x))}.
\]
where for $x\in\Wcs{x_0}, \Phi^u_{t(x)}(x)\in\Wcs{y_0}$.
\end{proposition}
Consider the set of H\"older multiplicative cocycles
\begin{equation}\label{eq:cocicles}
\Coc=\{h:M\to\Real_{+}: h(x)=e^{\int_0^1 \varphi(\Phi_t x)dt}\text{ for some H\"older function }\varphi\}.
\end{equation}
This is a natural class of functions for $\Phi$, related to its equilibrium states; see below. In the next section we recall the by now classical fact that any equilibrium state of the flow $\Phi$ corresponding to a H\"older potential $\varphi$ determines $h\in\Coc$ and a $\Phi^u$-transverse quasi invariant measure $\mu^{cs}$ satisfying the hypotheses of the previous proposition. Given $h\in\Coc$ define the family of functions
\begin{equation}\label{eq:hjacobiano}
\mathscr{h}=\{H_{x_0,y_0}:x_0,y_0\in M, y_0\in\Wu{x_0}\}
\end{equation}
with
\begin{equation}
H_{x_0,y_0}(x)=\prod_{j=1}^{\oo}\frac{h(\Phi_{-j}\circ \Phi^u_{t(x)} x)}{h(\Phi_{-j}(x))},\quad x\in\Wcs{x_0}, \Phi^u_{t(x)}(x)\in\Wcs{y_0}.
\end{equation}
Our first main result is as follows.
\begin{theoremA}
Let $\Phi$ be a codimension-one Anosov flow of class $\mathcal{C}^2$ that is not a suspension, and consider $h\in\Coc$. Assume that the unstable bundle of $\Phi$ is orientable and denote by $\Phi^u$ the induced horocyclic flow. Then there exists $\mu^{cs}=\{\mu_x^{cs}\}_x$ a transverse measure for $\Phi^u$ such that $\mu^{cs}$ is the unique quasi-invariant measure with Jacobian given by the family $\mathscr{h}$ determined by $h$.
\end{theoremA}
Given a flow $\Psi=\{\psi_t\}_t$ on $M$ and $J$ a non-negative multiplicative cocycle over $\Psi$, a (Borel) measure $\mu$ is said to be $J$-conformal if $\psi_t^{-1}\mu=J_t\cdot \mu$; the related definition for diffeomorphism is similar. The notion was introduced by Patterson \cite{Patterson1976} in the context of limit sets for Fuchsian groups, and has a important role in geometry and ergodic theory. The reader is directed to the book \cite{FeliksPrzytycki2011} for an extensive discussion of conformal measures in ergodic theory, and to the survey \cite{overviewpattersonsullivan} for a discussion of applications in geometry, in particular of the Patterson-Sullivan theory.
We single out the following remarkable existence and uniqueness result for conformal measures of A. Douady and J-C. Yoccoz \cite{Douady1999}.
\begin{theorem}
Let $f:\Tor\to\Tor$ be a $\mathcal{C}^2$ diffeomorphism of the circle with irrational rotation number. Then for every $s\in \Real$ there exists a unique $s\cdot Df$-conformal measure.
\end{theorem}
The following is a re-writing of Theorem A, which in view of the importance of the concept of conformal measures we isolate it as a corollary.
\begin{corollaryA}
In the same hypotheses of Theorem A, the horocyclic flow $\Phi^u$ has a unique $\mathscr{h}$-conformal measure.
\end{corollaryA}
The previous result was established originally by M. Babillot and F. Ledrappier \cite{geodesicbabillotledrappier} in the case of (Abelian covers of) the geodesic flow in a hyperbolic closed manifold, and later proven by a more geometrical argument by B. Schapira in \cite{SCHAPIRA2004} in the same setting. Our method has some similarities with Schapira's, but are less dependent on non-trivial geometrical considerations for the geodesic flow, and are more direct. This allow us to consider general (codimension one) Anosov flows and diffeomorphisms. Of course, the tailored arguments for the geodesic flow give additional information in that case, and are generalizable for some non-compact hyperbolic manifolds; see \cite{schapirafrench}, \cite{Paulin2015}. Applicability of the ideas in this paper to the non-compact case remains to be investigated.
The previous two results are evidence that parabolic systems may not have many more than one conformal measure; we can ask
\begin{question}
If $\Phi$ is a minimal parabolic flow and $J$ is non-negative multiplicative cocycle, does there exist a unique $J$-conformal measure.
\end{question}
This question is particularly interesting for unipotent flows, in view of the very general rigidity results due to Ratner \cite{Ratner1991}, and others derived from her work.
\paragraph{Equilibrium States}
The proof of the above Theorem A relies on the study of equilibrium states for the time-one map of the corresponding hyperbolic flow. Let us recall some basic definitions and refer the reader to \cite{EqStatesCenter} and references therein for a more throughout discussion.
For a continuous endomorphism of a compact metric space $f:M\to M$ we denote by $\PTM{f}{M}$ the set of $f$-invariant probabiity measures, equipped with the $\omega^{\ast}$-topology. Given a continuous real valued map $\varphi:M\to \Real$ (the potential), the topological pressure of the system is given as \cite{WaltersPres}
\begin{equation}\label{eq:variationalprinciple}
\Ptop(\varphi)=\sup_{\mathclap{\nu\in \PTM{f}{M}}}\{h_{\nu}(f)+\int \varphi d\nu\}
\end{equation}
where $h_{\nu}(f)$ denotes the metric entropy of $f$ with respect to $\nu$. We denote
\begin{equation}
\mathscr{Eq}_f=\{\mu\in\PTM{f}{M}:\Ptop(\varphi)=h_{\mu}(f)+\int \varphi d\mu\};
\end{equation}
elements in $\mathscr{Eq}_f$ are called \emph{equilibrium states} (for the system $(f,\varphi)$).
Similarly, if $\Phi=(\Phi_t)_t$ is a flow we denote $\PTM{\Phi}{M}=\cap_{t}\PTM{\Phi_t}{M}$ the set of its invariant measures; the topological pressure of the system $(\Phi,\varphi)$ is given as
\begin{equation}
\Ptop(\Phi, \varphi)= \sup_{\mathclap{\nu\in \PTM{\Phi}{M}}}\{h_{\nu}(\Phi_1)+\int \varphi d\nu\},
\end{equation}
and we define $\mathscr{Eq}(\Phi,\varphi)$ analogously. One can show (cf. proposition $6.2$ in \cite{EqStatesCenter}) that for a (transitive) Anosov flow one has $\mathscr{Eq}(\Phi,\varphi)=\mathscr{Eq}(\Phi_1,\tilde{\varphi})$ where
\[
\tilde{\varphi}(x)=\int_0^1 \varphi(\Phi_t x) dt.
\]
The next theorem is a central piece in smooth ergodic theory.
\begin{theorem*}[\cite{SinaiMarkov,SRBattractor,Bowen1974}]
Let $\Phi$ be a transitive Anosov flow of class $\mathcal{C}^1$ and $\varphi:M\to\Real$ be a H\"older function,
Then $\mathscr{Eq}(\Phi,\varphi)$ consists of a unique element $\muf[(\Phi,\varphi)]$.
\end{theorem*}
Here we are able to give the following strong characterization of $\muf[\Phi,\varphi]$ in terms of conditional measures along unstables. Let us recall that given a probability measure $\mathrm{m}$ on $M$ a measurable partition (in the sense of Rohklin \cite{Rokhlin}) $\xi$ is said to be
\begin{itemize}
\item increasing if $f\xi<\xi$ (i.e.\ for $\mathrm{m}\aep(x)$ the atom $f(\xi(x))$ consists of atoms of $\xi$);
\item subordinated to $\Fuu$ is for $\mathrm{m}\aep(x)$ it holds $\xi(x)\subset \Wuu{x}$;
\item a SLY partition if it is increasing, subordinated to $\Fuu$, and for $\mathrm{m}\aep(x)$ the atom $\xi(x)$ contains a relative neighborhood of $x$ in $\Wuu{x}$.
\end{itemize}
It is by now classical (and in much more generality that the case the we are considering here) that for Anosov flows there exist SLY partitions or arbitrarily small mesh.
We can prove the following.
\begin{theoremB}
Assume the same hypotheses of Theorem A. There exists a family of measures $\{\nu^u_x\}_{x\in M}$ where each $\nu^u_x$ is a Radon measure on $\Wu{x}$ such that if $\mathrm{m}$ is any probability measure (not necessarily invariant under any $\Phi_t$) such that if $\xi$ is a $\mathrm{m}$ SLY partition such that its conditionals are of the form
\[
\mathrm{m}^{\xi}_x=\frac{\nu^u_x}{\nu^u_x(\xi(x))}
\]
then necessarily $\mathrm{m}=\muf[\Phi,\varphi]$.
\end{theoremB}
\paragraph*{Acknowledgments} The authors thank Barbara Schapira for bringing to our attention the references \cite{geodesicbabillotledrappier,SCHAPIRA2004,schapirafrench,Paulin2015}.
\section{Measures along strong unstable leaves and characterization of equilibrium states}\label{sec:Basic_theory}
In this part we give the preliminaries necessary for the proof Theorems A and B.
\smallskip
From now on $\Phi=(\Phi_t:M\to M)_{t\in\Real}$ denotes a transitive hyperbolic flow of codimension-one that is not a suspension, with associated invariant decomposition $TM=E^{s}\oplus E^c\oplus E^{u}$, with $E^{u}$ tangent to the flow lines of the horocyclic flow $\Phi^u$. The bundles $E^{s}, E^{u}, E^{cs}=E^{c}\oplus E^{s}, E^{cs}=E^c\oplus E^{u}$ are all integrable to flow-invariant foliations $\F^{\ast}$, where the superscript $\ast$ coincides with the one of the bundle that it integrates. In particular, $\Fu$ is the foliation by orbits of $\Phi^u$. Each leaf of $\Fcs,\Fcu$ is saturated by leaves of $\Fs, \Fc$ and $\Fu, \Fc$, respectively.
\smallskip
\noindent\textbf{Notation:} If $A\subset M, \ep>0$ and $\F^{\ast}=\{W^{\ast}(x)\}_{x\in M}$ is one of the invariant foliations, let
\[
W^{\ast}(A,\ep):=\{y\in M:\exists x\in A/ d_{\mathcal{\F^{\ast}}}(x,y)<\ep\},
\]
where $d_{\mathcal{\F^{\ast}}}$ is the intrinsic distance in the corresponding leaf. If $A=\{x\}$ we write $W^{\ast}(x,\ep)=W^{\ast}(\{x\},\ep)$.
\smallskip
Since the foliations $\Fu, \Fcs$ are transverse, there exists some $\mathrm{c}_{\scriptstyle geo}$ such that
\[
x,y\in M, d(x,y)<\ep\leq\mathrm{c}_{\scriptstyle geo}\Rightarrow \# \Wu{x,2\ep}\cap \Wcs{y,2\ep}=1.
\]
For a submanifold $L\subset M$ denote $\RM[L]$ the set of Radon measures (for the induced topology) on $L$, and for $\F^{\ast}$ consider
\begin{align*}
&\mathrm{Rad}(\F^{\ast}):=\bigsqcup_{L\in\F^{\ast}} \RM[L]\\
&\mathrm{Meas}(\F^{\ast}):=\{\nu:M\rightarrow \mathrm{Rad}:\nu_x:=\nu(x)\in \RM[W^{\ast}(x)]\forall\ x\in M\}.
\end{align*}
We have the following.
\begin{proposition}[Proposition $6.1$ in \cite{EqStatesCenter}]\label{pro:medidasinestables}
There exists $\mu^u\in\mathrm{Meas}(\Fu)$ satisfying:
\begin{enumerate}
\item[a)] $\supp(\mu^u_x)=\Wu{x}$ and is non-atomic.
\item[b)] $\forall t\in\Real, x\in M$, it holds
\[f^{-t}\mu^u_{f^t(x)}=e^{\Ptop(\varphi) t-\int_0^t \varphi(f^s(x))ds}\mu^u_x.
\]
\item[c)] The map $x\to \mu^u_x$ is weakly continuous, meaning: given $y\in \Wu{x,\mathrm{c}_{\scriptstyle geo}}$ and $\hs_{y,x}:\Wu{y}\to\Wu{x}$ the locally defined Poincare' map determined by the center unstable foliation, it follows that for any $A\subset \Wu{x}$ relatively open and pre-compact it holds
\[
\mu_{y}^u(\hs_{y,x}(A))\xrightarrow[y\to x]{}\mu^u_x(A).
\]
If moreover $A\subset \Wu{x, \mathrm{c}_{\scriptstyle geo}}$, then the convergence is uniform in $x$.
\end{enumerate}
\end{proposition}
With this we were able to establish the next theorem.
\begin{theorem}[Theorem $6.3$ in \cite{EqStatesCenter}]\label{thm:Bforrankone}
If $\Phi$ is $\mathcal{C}^2$ Anosov flow with minimal unstable foliation, then the family $\mu^u$ of the previous theorem satisfies: if $\mathrm{m}\in\PM[M]$ satisfies
\begin{enumerate}
\item $\mathrm{m}$ is $f^t$-invariant for some $t\neq \emptyset$, and
\item $\mathrm{m}$ has conditionals absolutely continuous with respect to $\mu^u$,
\end{enumerate}
then $\mathrm{m}=\muf[(\Phi,\varphi)]$.
\end{theorem}
From now on we denote $f=\Phi_1$ and fix $\varphi:M\to\Real$ together with its corresponding measures $\mu^u$ given in \cref{pro:medidasinestables}. We denote
\[
h_{\varphi}(x)=e^{\tilde{\varphi}(x)},
\]
and given $x\in M$ define $\Delta_x^u:\Wu{x}\to\Real_{>0}$,
\begin{equation}\label{eq:JacobianoDelta}
y\in\Wu{x}\Rightarrow \Delta_x^u(y):=\prod_{k=1}^{\oo}\frac{h_{\varphi}(f^{-k}y)}{h_{\varphi}(f^{-k}x)}.
\end{equation}
Since $\tilde{\varphi}$ is H\"older it follows that $\Delta_x^u$ is continuous, and converges to one as $y\mapsto x$.
\begin{definition}\label{def:medidasproyectivas}
For $x\in M$ let $\nu^u_x=\Delta_x^u\ \mux$.
\end{definition}
Clearly $\nu^u_x\in\RM[\Wu{x}]$ and satisfies the following properties.
\begin{enumerate}
\item[a)] If $y\in \Wu{x}$ we have $\nu^u_y=c(y,x)\cdot \nux$ for some function $c(y,x)=\Delta_y(x)>0$, hence $\{\nu^u_y\}_{y\in\Wu{x}}$ defines a projective class of measures $[\nu^u_x]$ in $\Wu{x}$.
\item[b)] $f^{-1}\nux[fx]=e^{\Ptop(\varphi)-\tilde{\varphi(x)}}\nux$.
\end{enumerate}
\smallskip
\begin{definition}\label{def:compatiblemeasures}
A family of measures $\zeta^{cs}\in\mathrm{Meas}(\Fcs)$ is said to be compatible with $(\varphi,\mu^u)$ if its quasi-invariant for $\Phi^u$ with Jacobian given by the family $\mathscr{h_{\varphi}}$ (cf.\@ \cref{eq:hjacobiano}).
\end{definition}
We proved before (Section $3.2$ of \cite{EqStatesCenter}) that given $\zeta^{cs}$ compatible with $(\varphi,\mu^u)$
one can construct a probability measure $\mathrm{m}$ on $M$ that is given locally as follows. For $W=\Wcs{x_0,\ep}, 0<\ep\leq\mathrm{c}_{\scriptstyle geo}$,
\[
\mathrm{m}\approx \int_W \nux[x] d\zeta^{cs}_{x_0}(x).
\]
Compatibility is used to show that the definition does not depend on the chosen center stable disc, and therefore one can glue these local measures into $\mathrm{m}$.
\begin{remark}
Moreover, if $\zeta^{cs}$ is quasi-invariant measure for $f$ as in \cref{pro:quasinvariancecocycle}, then the resulting $\mathrm{m}$ is an equilibrium state for $(f, \tilde{\varphi})$, and therefore the unique equilibrium state for the system $(\Phi,\varphi)$. This does not require codimension-one and was noted in our previous article.
\end{remark}
We point out that there exists a family $\mu^{cs}\in\mathrm{Meas}(\Fcs)$ satisfying the quasi-invariance condition refereed in the previous remark. The equilibrium state $\muf[\Phi,\varphi]$ can be constructed using this family.
\section{Unique ergodicity for the conditional measures}
\label{sec:unique_ergodicity}
Let $\mathrm{c}_{\scriptstyle int}>0$ be much smaller than $\mathrm{c}_{\scriptstyle geo}$, and for $x\in M$ define
\begin{equation}
\label{eq:Ix} I_x:=\Phi^u_{[-\mathrm{c}_{\scriptstyle int},\mathrm{c}_{\scriptstyle int}]}(x).
\end{equation}
The set $I_x$ will be called a $u$-interval. Since the conditionals $\nux$ are non-atomic (cf.\@ \cref{pro:medidasinestables} , we deduce:
\begin{lemma}\label{lem:fronteraIxesnula}
It holds $\forall x\in M, \nu^u_x(\partial I_x)=0$.
\end{lemma}
Let $\mathcal{I}=\{I_x\}_{x\in M}$ and observe that the family $\mathcal{I}$ has the following continuity property:
\[
I_y\xrightarrow[y\to x]{\mathrm{Haus}} I_x
\]
where the above convergence is in the Hausdorff topology. We also have the following.
\begin{lemma}\label{lem:continuidadnuxIx}
The functions $x\to \mux(I_x)$ and $x\to\nux(I_x)$ are continuous.
\end{lemma}
\begin{proof}
By \cref{eq:JacobianoDelta} it suffices to prove the first part, which follows by $c)$ of \cref{pro:medidasinestables}, the lemma above, and Alexandrov's theorem.
\end{proof}
Recall that $f=\Phi_1$ denotes the time-one map of a (topologically mixing) codimension-one Anosov flow. For $\mathfrak{h}\in \CM[M]$ we define the operator
\begin{equation}\label{eq:Rnh}
R_n\mathfrak{h}(x)=\frac{1}{\nux[f^n(x)](f^n(I_x))}\int_{f^n(I_x)}\mathfrak{h}(t)d\nux[f^n(x)](t)=\frac{1}{\nux(I_x)}\int_{I_x}\mathfrak{h}\circ f^n(t)d\nux(t).
\end{equation}
From now on we fix $\muf$ an equilibrium state for $(f,\tilde{\varphi})$ constructed as explained in the previous section.
\medskip
\noindent\textbf{Notation.} We write $W^{\ast}_{\scriptstyle \mathrm{loc}}(x)=W^{\sigma}(x,\mathrm{c}_{\scriptstyle geo})$. If $B$ is foliation box of diameter less than $\mathrm{c}_{\scriptstyle geo}$, the connected component of $W^{\ast}_{\scriptstyle \mathrm{loc}}(x)\cap B$ that contains $x$ is denoted by $W^{\ast}(x, B)$.
\smallskip
We start by establishing the following.
\begin{proposition}\label{pro:equicontinuidad}
For every $\mathfrak{h}\in \CM$ the family $\{R_n\mathfrak{h}\}_n$ is equicontinuous and uniformly bounded.
\end{proposition}
\begin{proof}
As the family $\{R_n\mathfrak{h}\}_n$ is clearly bounded by $\normC{\mathfrak{h}}{0}$ we only need to show that it is equicontinuous. To this end, given two points $x,y$ with $d(x,y)<\mathrm{c}_{\scriptstyle geo}$ we can chose $z$ and $w$ so that $z\in \Wuloc{x}\cap \Wcsloc{y}$ and $w\in \Wcloc{z}\cap \Wsloc{y}$. See \cref{fig:comparaumeasure} below. To show equicontinuity we will make comparisons between $x$ and $z$, $z$ and $w$ and $w$ and $y$.
\begin{figure}[H]
\centering
\includegraphics[scale=0.9]{comparaumeasure.pdf}
\caption{Diagram for the proof.}
\label{fig:comparaumeasure}
\end{figure}
Since $z\in\Wu{x}$ we have $\nux=c(x,z)\nux[z]$, where $c(x,z)$ converges uniformly to one as $z$ approaches $x$ (cf.\@ the remark after the definition of the measures $\nux$). Therefore
\begin{align*}
|R_n\mathfrak{h}(x)-R_n\mathfrak{h}(z)|&=\left|\frac{1}{\nux(I_x)}\int_{I_x} \mathfrak{h}\circ f^n(t)d\nux(t)-\frac{1}{\nux(I_z)}\int_{I_z} \mathfrak{h}\circ f^n(t)d\nux(t)\right|\\
&\leq\norm{\mathfrak{h}}_{C^0}\int\left|\frac{\one_{I_x}(t)}{\nux(I_x)}-\frac{\one_{I_z}(t)}{\nux(I_z)}\right|d\nux(t)\leq \frac{2\norm{\mathfrak{h}}_{C^0}}{\nux(I_x)}\cdot\nux(I_x\triangle I_z)
\end{align*}
and the later term can be taken arbitrarily small if $z$ is close enough to $x$, by \cref{lem:continuidadnuxIx,lem:fronteraIxesnula}.
The comparison between $w$ and $y$ follows similar lines, using that $\Jacu[w,y]$ converges uniformly to $1$ as $w$ approaches $y$ (cf.\@ \cref{pro:quasinvariancecocycle}). Finally, for the comparison between $z$ and $w$ we observe that the center holonomy (i.e.\@ the flow) sends $\nu_z^u$ to $g_{z,w}\nu^u_w$, where $g_{z,w}$ is a continuous positive function that converges uniformly to one as $z\mapsto w$. Therefore, we get
\begin{align*}
&\left|\frac{1}{\nux(I_z)}\int_{I_z}\mathfrak{h}\circ f^n(t)d\nux[z](t)-\frac{1}{\nux[y](I_w)}\int_{I_w}\mathfrak{h}\circ f^n(t)d\nux[w](t)\right|\\
&=\left|\frac{1}{\nu^u_z(I_z)}\int\Big(\one_{I_z}(t)\mathfrak{h}\circ f^n(t)-\frac{\nux[z](I_z)}{\nux[w](I_w)}\one_{\hc_{z,w}(\tilde{I}_z)}(t)\mathfrak{h}\circ f^n(\hc(t))\cdot g_{z,w}(t)\Big)d\nux[w](t)\right|
\end{align*}
As $z\mapsto w$ we have $\frac{\nux[z](I_z)}{\nux[w](I_z)}\rightrightarrows 1$ as well. Since $f$ is part of the flow, we conclude that $|\mathfrak{h}\circ f^n(\hc(t))-\mathfrak{h}\circ f^n(t)|$ can be taken uniformly small for $z$ close to $w$, independently of $n$. Altogether we have shown that for every $n$, $|R_n\mathfrak{h}(z)-R_n\mathfrak{h}(w)|$ is small whenever $w\in\Wc{z}$ is close to $z$, and this concludes the proof of equicontinuity.
\end{proof}
\subsection{SOT convergence of \texorpdfstring{$(R_n)_n$}{Rn}.} Now we will show the central part of our argument: the family of positive operators $(R_n:\CM\to\CM)_n$ converges in the strong operator topology to some multiple of the identity. More precisely, there exists some probability measure $\mathrm{p}^u$ such that
\[
\mathfrak{h}\in\CM\Rightarrow \normC{R_n\mathfrak{h} -\int \mathfrak{h} d\mathrm{p}^u}{0}\xrightarrow[n\to\oo]{}0.
\]
Due to Riesz-Markov representation theorem, this amounts to establishing the following.
\begin{theorem}\label{thm:uniformconvergenceofRnh}
For every $\mathfrak{h}\in \CM$ the sequence $(R_n\mathfrak{h})_n$ converges uniformly to a constant $a$.
\end{theorem}
We first determine the value $a$. Define $\Omega$ as the set of all sequences
$\{\Theta^m_n\}_{m,n\geq 0}$ of functions $\Theta^m_n:M\rightarrow \mathrm{Rad}(\Fu)$ satisfying
\begin{enumerate}
\item $\Theta^m_{n,x}:=\Theta^{m,n}(x)$ is a probability measure on $f^m(I_x)$, and
\item given $\epsilon>0$ there exists $N_{\epsilon}>0$ such that for $n,m\geq N_{\epsilon}$ it holds for every $x\in M$
\begin{equation}\label{eq:Thetameasures}
R_{n+m}\mathfrak{h}(x)=\int R_{n}\mathfrak{h}(y)d\Theta^m_{n,x}(y)+E_{n,m}(x)
\end{equation}
where $|E_{n,m}(x)|<\epsilon$.
\end{enumerate}
Assume for the moment that $\Omega$ is non-empty and take $\{\Theta^m_n\}_{n,m\geq 0}\in \Omega$. Define
\[
a_{n}:=\inf_{x}R_n\mathfrak{h}(x)
\]
and observe that by \eqref{eq:Thetameasures}
\begin{align*}
a_{n+m}\geq a_n-\sup_{x}|E_{n,m}(x)|.
\end{align*}
Let $a:=\limsup_n a_n,b=\liminf_n a_n$; these are finite numbers in virtue of \cref{pro:equicontinuidad}. Given $\epsilon>0$ take $N_{\epsilon}$ such that $n,m\geq N_{\epsilon}$ implies $\sup_{x}|E_{n,m}(x)|<\epsilon$: since any neighborhood of $a$ contains points of $(a_n)$, by the previous formula we deduce that $|b-a|<2\epsilon$, and as $\epsilon$ is arbitrary we have that $a=b$, therefore
\[
a=\lim_n a_n.
\]
Now that we have our candidate for constant function, we will proceed as follows: consider $\mathfrak{g}\in \CM$ an accumulation point of some subsequence $(R_{n_k}\mathfrak{h})_{k}$ (which exists by equicontinuity of the family). Given an open set $U\subset M$, we will find $\Theta\in\PM$ with $U\subset \supp(\Theta)$ and $\int_U (\mathfrak{g}-a) d\Theta=0$. Since $\mathfrak{g}(x)\geq a$ for every $x$, this will imply the existence of some $x_U\in U$ such that $g(x_U)=a$ on $U$. As $U$ can be chosen in a basis of the topology on $M$, we will deduce that $g=a$, thus showing $\lim_n \normC{R_n\mathfrak{h}-a}{0}=0$.
The remainder of this part is devoted to establishing this result.
\smallskip
Take $U\subset M$ open box of the form $U\approx \Wc{\Ws{x_0,\delta},\delta}\times \Wu{x_0,\delta}$ with $\delta$ much smaller than $\mathrm{c}_{\scriptstyle geo}$.
\medskip
\noindent\textbf{Notation.} For $0<\rho<< \mathrm{c}_{\scriptstyle geo}$ and $S\subset M$ we denote by $\partial_{\rho} S$ the $\rho$-neighborhood of the boundary of $S$. If $U\approx \Wc{\Wu{x_0,\delta},\delta}\times \Ws{x_0,\delta}$ is a box then its $cu$-boundary of size $\rho$ is the set $\partial^{cu}_\rho U$ corresponding to $\Wc{\Wu{x_0,\delta},\delta}\times \partial_{\rho} \Ws{x_0,\delta}$.
\smallskip
\begin{lemma}\label{lem:boundU}
If $\muf(\partial U)=0$ there exists $\mathrm{C}_U>0$ such that for every $x\in M$
\[
\liminf_{m\mapsto\oo}\frac{\nu_{f^mx}(U\cap f^m(I_x))}{\nu_{f^mx}(f^m(I_x))}>\mathrm{C}_U>0.
\]
\end{lemma}
\begin{remark}
Since $(f,\muf)$ is a Kolmogorov system, the above inequality holds for $\muf\aep(x)$, however for our purposes it is important to establish the validity for all $x\in M$.
\end{remark}
\begin{proof}
Start observing that $\frac{\nu_{f^mx}(U\cap f^m(I_x))}{\nu_{f^mx}(f^m(I_x))}=\frac{\nu_{x}(f^{-m}U\cap (I_x))}{\nu_{x}((I_x))}$.
Assume by means of contradiction that the conclusion is not true. Since $\inf_x \nu^u_x(I_x)>0$ we have that for every $\tau>0$ there exists $x_{\tau}$ and $n_{\tau}$ such that for every $n\geq n_{\tau}$ it holds
\[
\nu_{x_{\tau}}(I_{x_{\tau}}\cap f^{-n}U)\leq \tau.
\]
Choose $\rho>0$ so that $\muf(\partial^{cu}_\rho U)$ is much smaller than $\tau$, set $d_U=\diam U$ and define
\[
B_x=\bigcup_{y\in \Wc{I_x, 5d_U}}I_y
\]
We want to compare the measures $\muf(B_x\cap f^{-n}U)$ and $\nu^u_x(I_x\cap f^{-n}U)$. Divide the connected components of $B_x\cap f^{-n}U$ into two types, $\mathrm{Good}_x^n$ and $\mathrm{Bad}_x^n$ where $Z\in \mathrm{Good}_x^n\Leftrightarrow$ for every $z\in Z$ it holds
\begin{enumerate}
\item $f^n(\Ws{z, Z})\subset \Ws{f^nz, U}$.
\item $f^n(\Wu{z, Z}) \supset \Wu{f^nz, U}$.
\end{enumerate}
Observe that as $f$ preserves center lengths the second condition above implies
\begin{enumerate}
\item[3.] $f^n(\Wcu{z, Z}) \supset \Wcu{f^nz, U}$
\end{enumerate}
It follows that if $Z\in \mathrm{Bad}^n_x$ then either
\begin{enumerate}
\item $Z\subset \partial_{\rho}^{cu} U$, or
\item $f^n(I_x\cap f^{-n}U\cap Z)\not \supset \Wu{f^nz, U}$.
\end{enumerate}
Observe that we can write,
\begin{align*}
\MoveEqLeft\frac{\muf(f^{-n}U\cap B_{x})}{\nux(f^{-n}\cap I_x)}=\frac{\muf(\bigcup_{Z\in \mathrm{Good}_x^n} Z)}{\nux(f^{-n}\cap I_x)}\left(1+\frac{\muf(\bigcup_{Z\in \mathrm{Bad}_x^n}Z)}{\muf(\bigcup_{Z\in \mathrm{Good}_x^n}) Z}\right)
\shortintertext{which by the local product structure of $\muf$ implies, for some uniform $\mathrm{c}_{\scriptstyle prod}$,}
&\frac{\muf(\bigcup_{Z\in \mathrm{Good}_x^n} Z)}{\nux(f^{-n}\cap I_x)}\left(1+\frac{\muf(\partial^{cu}_{\rho} U)+\mathrm{c}_{\scriptstyle prod}\cdot\nux(\partial_{\lambda^{-n}}2\mathrm{c}_{\scriptstyle int})}{\muf(\bigcup_{Z\in \mathrm{Good}_x^n}) Z}\right).
\end{align*}
Using again the local product structure we deduce the existence of $\tilde{\mathrm{d}}>0$ satisfying: for every $x$ there exists $m_x$ so that for $n\geq m_x$,
\[
\muf(f^{-n}U\cap B_{x})\leq \tilde{\mathrm{d}}\cdot \nux(I_x\cap f^{-n}U).
\]
Putting all the pieces together we finally conclude that for every $\epsilon>0$ there exist $x_{\ep}\in M$ and $n_{\ep}'\in\Nat$ such that if $n\geq n_{\ep}$ then
\begin{equation*}
\muf(B_{x_{\tau}}\cap f^{-n}U)<\epsilon.
\end{equation*}
Since the system is mixing, we then deduce that for $n\geq n_{\ep}''$,
\[
\muf(B_{x_{\tau}})\cdot \muf(U)< 2\ep.
\]
We will show below that $\inf_{x} \muf(B_{x})>0$: but then we get that $\muf(U)=0$, contradicting the fact that $\muf$ has full support. To finish the proof we establish the following.
\smallskip
\noindent\textbf{Claim:} If $\lambda$ is a probability on $M$ with full support, then $\inf_{x}\lambda(B_x)>0$.
\smallskip
Suppose not: then we can find a converging sequence $x_n\xrightarrow[n\mapsto\oo]{}x$ such that the corresponding sets $B_{x_n}$ converge to $B_x$ in the Hausdorff topology, and furthermore
\[
\lambda(B_{x_n})\xrightarrow[n\mapsto\oo]{} 0.
\]
Take $D(x,r)\subset B_x$ open disc of radius $r$, centered at $x$, and such that $\lambda(\partial D(x,r))=0$: as
\begin{align*}
D(x_n,r)\xrightarrow[n\mapsto\oo]{\mathrm{Haus}}D(x,r)
\end{align*}
it follows that
\[
\lambda(D(x_n,r))\xrightarrow[n\mapsto\oo]{}\lambda(D(x,r)),
\]
and hence $\lambda(D(x,r))=0$, contrary to our assumption that $\lambda$ has full support.
\end{proof}
We want to sub-divide $f^m(I_x)$ in intervals that are well positioned with respect to the connected components of $f^m(I_x)\cap U$; for this we will use the following lemma (cf.\@ lemma $3.2$ in \cite{Marcus1975}).
\begin{lemma}\label{lem:Marcus}
\label{lem:marcus} Given $\ep>0$ there exists $n_{\ep}\in \Nat$ such that for every $x\in M$, the set
\[
\{\Phi^u_{t 2\mathrm{c}_{\scriptstyle int}}(x):t=0,\cdots, n_{\ep}\}
\]
is $\ep$-dense.
\end{lemma}
To cover $f^m(I_x)$ we use the following algorithm. Denote $\mathrm{d}_U=\diam U$ and choose $0<\ep< \frac{\mathrm{d}_U}{5}$. For $x\in M$ consider $\{C_{x}^{(m)}(l)\}_{l=0}^{o(x,m)}$ the set of connected components of $f^m(I_x)\cap U$, ordered according to the orientation of $f^m(I_x)$; note that since $\mathrm{d}_U$ is small with respect to $\mathrm{c}_{\scriptstyle int}$, if $y\in C_{x}^{(m)}(l)$ then $I_y\cap U=C_{x}^{(m)}(l)$. Choose $y_{x,0}^{(m)}\in C_{x}^{(m)}(0)$ and for $n\in \Z$ let $y_{x,n}^{(m)}=\Phi^u_{2\mathrm{c}_{\scriptstyle int}}$. Observe that by applying the above lemma to the family $\{y_{x,n}^{(m)}\}_{n\geq 0}$ we get
\begin{equation}\label{eq:proporcionplacafmU}
\forall l, \#\{l\leq n\leq l+2n_{\ep}: y_{x,n}^{(m)}\in U\}\geq n_{\ep}.
\end{equation}
After choosing $y_{x,0}^{(m)}$ consider $\overline{k}_x^{m}=\overline{k}_x^m(y_{x,1}^m), \underline{k}_x^m=\underline{k}_x^m(y_{x,1}^m)$ the largest natural numbers such that $I_{y_{x,\overline{k}_x^m}^{(m)}}\subset f^m(I_x), I_{y^{(m)}_{x,-\underline{k}_x^m}}\subset f^m(I_x)$ and let
\begin{align}
&\mathrm{Good}^m_x:=\bigcup_{l=-\underline{k}_x^m}^{\overline{k}_x^m} I_{y_{x,l}^{(m)}}\\
&\mathrm{Bad}^m_x:=f^m(I_x)\setminus\mathrm{Good}^m_x
\end{align}
Note that
\begin{equation}\label{eq:medidamalos}
\frac{\nu^u_{f^mx}(\mathrm{Bad}^m_x)}{\nu^u_{f^m(x)}(f^m(I_x))}\leq \frac{\nu^u_{f^m(x)}(\partial_{4\mathrm{c}_{int}} f^m(I_x))}{\nu^u_{f^m(x)}(f^m(I_x))}=\frac{\nu^u_x(f^{-m}(\partial_{4\mathrm{c}_{int}} f^m(I_x))}{\nu^u_{x}(I_x)}.
\end{equation}
\begin{lemma}\label{lem:medidamalos}
$\frac{\nu^u_x(f^{-m}(\partial_{4\mathrm{c}_{int}} f^m(I_x))}{\nu^u_{x}(I_x)}\xrightarrow[m\to\oo]{}0$ uniformly in $x$.
\end{lemma}
\begin{proof}
Assume that this were not the case: then we could find $\zeta>0$, a converging sequence $x_k\xrightarrow[k\mapsto\oo]{} x$ and corresponding $m_k\xrightarrow[k\mapsto\oo]{}\oo$ such that
\begin{itemize}[leftmargin=*]
\item $\frac{\nu^u_{x_k}(f^{-m_k}\partial_{4\mathrm{c}_{int}} I_{f^{m_k}(x_k)})}{\nu^u_{z_k}(I_{x_k})}\geq \zeta$.
\item $\clo{I_{x_k}}\xrightarrow[k\mapsto\oo]{\mathrm{Haus}}\clo{I_x}$.
\end{itemize}
Since the measures $\nu^u_{x_k}$ and $\nu^u_{x}$ are comparable, we deduce the existence of $\zeta'>0$
and a sequence $0<\delta_k\xrightarrow[k\mapsto\oo]{}0$ such that for all sufficiently large $k$ it holds
\[
\nu^u_{x}(\partial_{\delta_k} I_{x})\geq \zeta'.
\]
This contradicts the fact that $\nu^u_x(\partial I_x)=0$.
\end{proof}
On the other hand, by \cref{lem:boundU,eq:proporcionplacafmU} and since the the sizes of $\{\nux(I_x)\}_{x\in M}$ are uniformly comparable, we deduce the following.
\begin{lemma}
\label{lem:medidaplacasUIx} There exists $\mathrm{c}_{U}'>0$ such that
\[
\forall x\in M,\ \liminf_{m\to\oo} \frac{\nux[f^mx](\{Z\in \mathrm{Good}^m_x:Z\cap U\neq \emptyset\})}{\nux[f^mx](\{Z\in \mathrm{Good}^m_x\})}\geq \mathrm{c}_U'>0.
\]
\end{lemma}
Now we prove that $\Omega\neq \emptyset$.
\begin{proposition}\label{pro:existenciathetameasurecasoequicontinuo}
There exist $\{\Theta^m_n\}_{m,n\geq 0}\in \Omega$ and $N(U)\in\Nat, \mathrm{C}_U>0$ such that for $n,m\geq N(U),x\in M$ it holds
\[
\Theta^m_n(U)\geq \mathrm{C}_U.
\]
\end{proposition}
\begin{proof}
Let us start by computing
\begin{align*}
&R_{n+m}\mathfrak{h}(x)=\frac{1}{\nux(I_x)}\int_{I_x}\mathfrak{h}\circ f^{n+m}(t)d\nux(t)=\frac{1}{\nux[f^mx](f^m(I_x))}\int_{f^m(I_x)}\mathfrak{h}\circ f^{n}(t)d\nu^u_{f^m(x)}(t)\\
&=\frac{1}{\nux[f^m(x)](f^m(I_x))}\sum_{l=-\underline{k}_x^m}^{\overline{k}_x^m}\int_{I_{y_{x,l}^{(m)}}}\mathfrak{h}\circ f^nd\nu^u_{f^m(x)}
+\frac{1}{\nux[f^m(x)](f^m(I_x))}\int_{\mathrm{Bad}^m_x}\mathfrak{h}\circ f^nd\nux[f^m(x)]\\
&=\frac{1}{\nux[f^m(x)](f^m(I_x))}\sum_{l=-\underline{k}_x^m}^{\overline{k}_x^m} \frac{\nux[f^m(x)](I_{y_{x,l}^{(m)}})}{\nux[y_{x,l}^{(m)}](I_{y_{x,l}^{(m)}})} \int_{I_{y_{x,l}^{(m)}}}\mathfrak{h}\circ f^nd\nux[y_{x,l}^{(m)}]\\
&+\frac{1}{\nux[f^m(x)](f^m(I_x))}\int_{\mathrm{Bad}_x^m}\mathfrak{h}\circ f^nd\nux[f^m(x)].
\end{align*}
Denote $c^{(m)}_l(x)=\frac{\nux[f^m(x)](I_{y_{x,l}^{(m)}})}{\nux[f^m(x)](f^m(I_x))}$ and define
\[
\Theta^m_{n,x}=\sum_{l=-\underline{k}_x^m}^{\overline{k}_x^m} c^{(m)}_l(x)\delta_{y_{x,l}^{(m)}}+\nux[f^m(x)](\ \cdot\ |\mathrm{Bad}_x^m);
\]
this is a probability measure supported on $f^m(I_x)$, and by the computation above
\[
R_{n+m}\mathfrak{h}(x)=\int R_n\mathfrak{h}(t)d\Theta^m_{n,x}(t)+H^{(n)}_m(x)
\]
with $H^{(n)}_m(x)=\frac{1}{\nux[f^m(x)](f^m(I_x))}\int_{\mathrm{Bad}_x^m}\left(\mathfrak{h}\circ f^n-R_n\mathfrak{h}\right)d\nux[f^m(x)]$, which by \eqref{eq:medidamalos} and \cref{pro:equicontinuidad} satisfies
\begin{align*}
|H^{(m)}_{n}(x)|\leq 2\normC{\mathfrak{h}}{0}\frac{\nu^u_{f^mx}(\mathrm{Bad}^m_x)}{\nu^u_{f^m(x)}(f^m(I_x))}\xrightarrow[m\to\oo]{}0
\end{align*}
uniformly in $x,n$. It remains to show that $\Theta^m_{n,x}(U)$ is uniformly bounded from below: we compute
\begin{align*}
\Theta^m_{n,x}(U)&\geq \sum_{l=-\underline{k}_x^m}^{\overline{k}_x^m} c^{(m)}_l(x)\shortintertext{which by \cref{lem:boundU,lem:medidaplacasUIx} satisfies for $m$ large}
&\geq \mathrm{C}_U:=C_Uc_U'.
\end{align*}
\end{proof}
We are ready to establish the SOT convergence of $\{R_n\}_n$.
\begin{proof}[Proof of \cref{thm:uniformconvergenceofRnh}]
We will use the notation of the previous proposition. Letting $a_n=\inf R_n\mathfrak{h}(x)$ it follows, since $\Omega\neq \emptyset$ that $\exists \lim_n a_n=\sup a_n= a$. Proposition \ref{pro:equicontinuidad} and Arzel\'{a}-Ascoli's theorem imply that $\{R_nh\}_n$ is pre-compact; we will show that it only accumulates on the constant function $a$, thus establishing the theorem. To this end, let $(n_k)_{k}$ be any subsequence such that $\{R_{n_k}h\}_k$ is convergent to some function $\mathfrak{g}\in \CMc[M]$: necessarily $\min \mathfrak{g}=a$.
Observe that for every $m$ it holds
\[
R_{n_k+m}\mathfrak{h}(x)-a=
\sum_{l=-\underline{k}_x^m}^{\overline{k}_x^m} c^{(m)}_l(x) \left(R_{n_k}\mathfrak{h}(y_{x,l}^{(m)})-a\right)+H^{(m)}_{n_k}(x)-a\left(1-\sum_{l=-\underline{k}_x^m}^{\overline{k}_x^m} c^{(m)}_l(x) \right).
\]
Take $x_{n_k+m}$ such that $R_{n_k+m}\mathfrak{h}(x_{n_k+m})=a$: then
\begin{equation}\label{eq:medidastheta1}
0=\sum_{l} c^{(m)}_l(x_{n_k+m})\left(R_{n_k}\mathfrak{h}(z_{k,l}^{(m)})-a\right)+ H^{(m)}_{n_k}(x_{n_k+m})-a\left(1-\sum_l c^{(m)}_l(x_{n_k+m}) \right).
\end{equation}
Here the $z_{k,l}^{(m)}$ are associated to the points $x_{n_k+m}$. Now consider the sequence of measures
\[
\Theta^m_{(n_k)}=\Theta^m_{n,x_{n_k+m}},
\]
Since $c^{(m)}_l(x)\to 0$ uniformly in $x$ as $m\to\infty$ we get that
\[
H^{(m)}_{n_k}(x_{n_k+m})-a\left(1-\sum_l c^{(m)}_l(x_{n_k+m}) \right)\to 0
\]
as $m\to\infty$, uniformly for $n_k\to\infty$. By using the diagonal argument we can find sub-sequences $\{n_{k'}\}_{k'}$ of $\{n_k\}_k$, $\{m_j\}\subset \mathbb{N}$ and a measure $\Theta$ such that
\[
\lim_{k'\mapsto\oo}\lim_{j\mapsto\oo}\Theta^{(m_j)}_{n_{k'}}=\Theta.
\]
Note that by the previous proposition, $\Theta(U)\geq \mathrm{c}_U'>0$. Take limit as $m_j\to\infty$ and $n_k\to\infty$ simultaneously in equation \eqref{eq:medidastheta1} to get, on the one hand
\[
\int \left(R_{n_k}\mathfrak{h}-a\right) d\Theta^{(m_j)}_{n_k}\to 0
\]
and on the other,
\[\int \left(R_{n_k}\mathfrak{h}-a\right) d\Theta^{(n_k)}_{m_j}\to \int \left(\mathfrak{g}-a\right) d\Theta,
\]
so that
\[\int \left(\mathfrak{g}-a\right) d\Theta=0.
\]
Finally, since $\min \mathfrak{g}=a$, $\mathfrak{g}-a\geq 0$ and thus
\[
0=\int \left(\mathfrak{g}-a\right) d\Theta\geq \int_U \left(\mathfrak{g}-a\right) d\Theta\geq \inf_U (\mathfrak{g}-a)\widetilde{C}(U)
\]
which in turn implies that $\mathfrak{g}-a$ has a zero in $U$. Observe that $U$ can be taken arbitrarily small (as long as $\muf(U)=0$), therefore we deduce that $\mathfrak{g}\equiv a$ on $M$.
We have shown that the only accumulation point of $(R_n\mathfrak{h})_n$ is the constant function $a$, hence $\lim_n \normC{R_n\mathfrak{h}-a}{0}=0$ as we wanted to show.
\end{proof}
\section{Proof of uniqueness: theorems A and B}
\label{sec:proof_of_uniqueness_theorems_a_and_b}
Now we fix $\xi$ a SLY partition for the flow, and let
\begin{equation}
\mathcal{N}:=\bigcup_{n\geq 0}f^{-n}\bigcup_{x\in M} \partial\xi(x)
\end{equation}
By \cref{thm:Bforrankone} it holds that $\mathrm{m}(\mathcal{N})=0$ for any equilibrium state $\mathrm{m}$ of the system, in particular for $\mathrm{m}=\muf$. For $n\in \Nat$ we consider also $\xi^n=f^n\xi$ (i.e.\@ $x\in\mathcal{N}, \xi^n(x)=f^n\xi(f^{-n}x)$). It is also no loss of generality to assume that for every $x\not\in N$ the atom $\xi(x)$ contains a neighborhood of $x$ inside $\Wu{x}$.
Given $\mathfrak{h}\in \CM[M], x\in M\setminus \mathcal{N}$ define
\begin{equation}\label{eq:Enh}
E_n\mathfrak{h}(x)=\frac{1}{\nu^u_x(\xi^n(x))}\int_{\xi^n(x)} \mathfrak{h}(t) d\nu^u_x(t)=\frac{1}{\nu^u_{f^{-n}x}(\xi(f^{-n}x))}\int_{\xi(f^{-n}x)} \mathfrak{h}\circ f^n(t) d\nu^u_{f^{-n}x}(t).
\end{equation}
It follows that
\begin{enumerate}
\item $\sup_x |E_n\mathfrak{h}(x)|\leq \normC{\mathfrak{h}}{0}$;
\item for every $\mathrm{m}\in \PM[M]$ having conditionals given by $\{\nux\}_{x\in M}$ (in particular, for every equilibrium state by \cref{thm:Bforrankone}), $E_n\mathfrak{h}$ is a version of the conditional expectation $\ie{\mathrm{m}}{\mathfrak{h}|\xi^n}$.
\end{enumerate}
We now argue similarly as we did in the previous part: given $x\in M\setminus \mathcal{N}$ and $m$ large, cover $f^m(\xi(x))$ as follows: consider $I_{z^{m}_{0,x}},\cdots, I_{z^{m}_{o(x,m),x}}\subset f^m(\xi(x))$ a maximal disjoint family, and call $\mathrm{Good}^m_x=\bigcup_{l=0}^{o(x,m)}I_{z^{m}_{0,x}}$, $\mathrm{Bad}^m_x=f^m(I_x)\setminus \mathrm{Good}^m$.
Observe that $\mathrm{Bad}^m_x\subset \partial_{2\mathrm{c}_{\scriptstyle int}}f^m\xi(x)$, and since $\nux(\partial \xi(x))=0$ we can argue as in \cref{lem:medidamalos} and deduce
\begin{equation}\label{eq:medidamalospuntual}
\lim_{m\to \oo}\frac{\nux[f^mx](\cup_{Z\in \mathrm{Bad}^m_x}Z)}{\nux[f^mx](\cup_{Z\in \mathrm{Good}^m_x}Z)}=0.
\end{equation}
Now compute, as in the proof \cref{pro:existenciathetameasurecasoequicontinuo},
\begin{align*}
E_{n}\mathfrak{h}(x)=\frac{1}{\nux(\xi^n(x))}\int_{\xi^n(x)} \mathfrak{h}(t) d\nu^u_x(t)=\sum_{l=0}^{o(x,n)} \frac{\nux(I_{z^{n}_{l,x}})}{\nux(\xi^n(x))}R_n\mathfrak{h}(z^{n}_{l,x})+\frac{1}{\nux(\xi^n(x))}\int_{\mathrm{Bad}^n_x}h(t) d\nux(t)\\
=\int R_n\mathfrak{h}(t)d\Theta_{n,x}(t)+T_{n,x}
\end{align*}
where $\Theta_{n,x}$ is a probability supported on $\xi^{n}(x)$ and $T_{n,x}$ converges to zero as $n\mapsto \oo$, due to \cref{eq:medidamalospuntual}. Taking an accumulation point of $\{\Theta_{n,x}\}$ and using the uniform convergence of $\{R_n\mathfrak{h}\}_{n\geq 0}$ we deduce
\begin{equation}\label{eq:converEE}
\lim_n E_n\mathfrak{h}(x)=a\quad \forall x\not\in\mathcal{N}.
\end{equation}
\begin{proof}[Proof of Theorem B]
Let $\mathrm{m}\in \PM[M]$ be such that its conditionals in $\xi$ are given by $\{\nux\}_{x}$, i.e.
\[
\mathrm{m}^{\xi}_x=\nux(\cdot | \xi(x))\quad \mathrm{m}\aep(x).
\]
In this case, $\mathrm{m}^{\xi^n}=\nux(\cdot | \xi^n(x))$. Now for $\mathfrak{h}\in\CM$ we can compute
\begin{align*}
\int \mathfrak{h} d \mathrm{m}=\int E_n\mathfrak{h}(x) d\mathrm{m}(x)\xrightarrow[n\to\oo]{}\int a d\mathrm{m}(x)=a
\end{align*}
where we have used \eqref{eq:converEE} and the fact $\mathrm{m}(M\setminus\mathcal{N})=1$. Since $a$ does not depend on $\mathrm{m}$ we conclude
\[
\int \mathfrak{h} d\mathrm{m}=\int \mathfrak{h} d\muf\quad \forall \mathfrak{h}\in\CM
\]
and therefore, $\mathrm{m}=\muf$.
\end{proof}
\begin{proof}[Proof of Theorem A]
Let $\tilde{\mu}^{cs}=\{\tilde{\mu}^{cs}_x\}_{x\in M}$ be a quasi-invariant measure for $\Phi^u$ with Jacobian determined by $h\in\Coc$. Construct the corresponding probability measure $\mathrm{m}$ on $M$ as explained in the last part of the Second section. By construction $\mathrm{m}$ has conditionals given by the family $\{\nux\}$, hence by Theorem B $\mathrm{m}=\muf[\Phi,\varphi]$. From here we deduce, using the local product structure of $\mathrm{m}$ that $\tilde{\mu}^{cs}=\mu^{cs}$.
\end{proof}
\newpage
\printbibliography
\end{document} |
\section{Introduction}
In recent years, research in deep neural networks (DNNs) has been fueled by new available computational resources, which have brought a wide variety of new techniques for visual object recognition, object detection and speech recognition among many others~\cite{lecun2015deep}.
The rise of DNNs in many applications, such as medicine \cite{zeleznik2021deep,de2018clinically}, climate \cite{waldmann2019mapping}, wildlife ecology \cite{norouzzadeh2018automatically}, physics \cite{udrescu2020ai,kwon2020magnetic} or sustainability \cite{vinuesa_et_al_2020}, has not been overlooked in fluid-mechanics research \cite{kutz2017deep,brunton2020machine}.
Some of the outstanding applications of DNNs in fluid mechanics are the improvement of Reynolds-averaged Navier--Stokes simulations \cite{ling2016reynolds}, the extraction of turbulence theory for two-dimensional flow \cite{jimenez2018machine}, prediction of temporal dynamics~\cite{srinivasan2019predictions,eivazi2021recurrent} or the embedding of physical laws in DNN predictions \cite{raissi_et_al}.
Generative adversarial networks (GANs), firstly introduced in Ref.~\onlinecite{goodfellow2014generative}, are one of the latest advances in DNN research.
Based on game theory, GANs are composed of two competing networks: a generator that tries to produce an artificial output which mimics reality; and a discriminator, which is in charge of distinguishing between reality and artificial outputs.
During training, the generator network makes its output more realistic by improving the features that the discriminator identified as artificial.
Among the different areas in which GANs have been applied successfully, their use to enhance image resolution stands out \cite{ledig2017photo,wang2018esrgan}.
In fluid-mechanics research, they have been successfully applied to recover high-resolution fields in different types of flow, such as the wake behind one or two side-by-side cylinders \cite{deng2019super} or volumetric smoke data \cite{werhahn2019multi}.
While in these works the training has been carried out with a supervised approach i.e., with paired high- and low-resolution flow fields, GANs have been recently applied with an unsupervised approach to enhance the resolution of homogeneous turbulence and channel flows \cite{kim2021unsupervised}.
GANs are now challenging other resolution-enhancement strategies based on convolutional neural networks (CNNs), which showed to be successful for the cases of the flow around a cylinder, two-dimensional decaying isotropic turbulence \cite{fukami2019super} and channel flows \cite{liu2020deep}.
More recently, Ref.~\onlinecite{fukami2021machine} has proposed a methodology based on CNNs to recover high-resolution sequences of flow fields in homogeneous isotropic and wall turbulence from the low-resolution fields at the beginning and end of the sequence.
CNNs have also been used successfully to estimate flow fields using field measurements of wall shear and/or pressure.
Several methods have been proposed, such as the direct reconstruction of the flow field from the wall quantities using fully-convolutional networks (FCNs) ~\cite{guastoni,guastoni2020convolutional}, or the use of proper orthogonal decomposition\cite{lumley1967structure} (POD) in combination with CNNs \cite{guemes2019sensing} and FCNs \cite{guastoni2020convolutional}.
Moreover, Ref.~\onlinecite{guemes2019sensing} studied the effect of the wall-resolution measurements on the predictions accuracy, showing that their architecture was able to continue providing predictions of similar accuracy for downsampling factors up to 4.
When a limited number of sensors is available, shallow neural networks (SNNs) offer another option for this task.
Ref.~\onlinecite{erichson2020shallow} compared SNNs with POD for the reconstruction of a circular cylinder wake, sea surface temperature, and decaying homogeneous isotropic turbulence, showing that the new data-driven approach outperforms the traditional one.
In the first part of the present work, a GAN-based methodology is proposed to recover high-resolution fields of wall measurements.
Because the results are very positive when performing this task, and it has already been shown that GANs can be used successfully in enhancing turbulent-flow resolution\cite{kim2021unsupervised}, the second part of this work extends their use to reconstruct high-resolution wall-parallel flow fields from coarse wall measurements.
This method is compared with the FCN-POD architecture proposed in Ref.~\onlinecite{guastoni2020convolutional}.
The choice is based on the proven capability of this network to deal with low-resolution input information \cite{guemes2019sensing}.
The paper is organized as follows: \S\ref{sec:metho} outlines the details of the numerical database used for this study and presents the different DNNs employed for that purpose; the main results for wall-resolution enhancement are provided in \S\ref{sec:wall}, while the flow-reconstruction results are reported in \S\ref{sec:flow}.
To close the paper, \S\ref{sec:concl} presents the main conclusions of the work.
\section{Methodology}\label{sec:metho}
This section presents the details of the numerical database employed for this study, as well as the DNN architectures and the training methodology with which they have been optimized.
Throughtout the paper $x$, $y$, and $z$ denote the streamwise, wall-normal, and spanwise directions respectively, with $u$, $v$, and $w$ referring to their corresponding instantaneous velocity fluctuations.
Streamwise and spanwise wall-shear-stress fluctuations are referred to as $\tau_{w_x}$ and $\tau_{w_z}$ respectively, with $p_w$ denoting the pressure fluctuations at the wall.
\subsection{Dataset description}
The methodology proposed in this work has been tested with a direct numerical simulation (DNS) of a turbulent open-channel flow generated with the pseudo-spectral code SIMSON \citep{chevalier}.
The simulation domain extends $4\pi h \times h \times 2\pi h$ (where $h$ is the channel height) in the streamwise, wall-normal and spanwise directions respectively, with the flow represented by 65 Chebyshev modes in the wall-normal direction and with 192 Fourier modes in the streamwise and spanwise directions.
The simulation is characterized by a friction Reynolds number $Re_{\tau}=180$, which is based on $h$ and the friction velocity $u_{\tau}=\sqrt{\tau_w/\rho}$ (where $\tau_w$ is the magnitude of the wall-shear stress and $\rho$ is the fluid density).
The superscript '+' denotes inner-scaled quantities, using $u_{\tau}$ for the velocities and the viscous length $\ell^*=\nu/u_{\tau}$ (where $\nu$ is the fluid kinematic viscosity) for the distances.
DNNs have been trained with 50,400 samples separated by $\Delta t^+=5.08$, while 3,125 samples with time separation $\Delta t^+=1.69$ have been used for testing them.
For further simulation details, see Ref.~\onlinecite{guastoni2020convolutional}.
Wall information, used as input to reconstruct wall-parallel fluctuating velocity fields, is composed of streamwise and spanwise shear stress, as well as pressure fluctuations.
To assess the capability of our methodology to reconstruct turbulent wall and flow fields from coarse wall measurements, three different sets of downsampled wall fields have been generated, with downsampling factors $f_d=[4,8,16]$.
Note that $f_d$ is defined as the resolution reduction in each direction, thus a downsampling factor of $f_d$ yields a downsampled field with a number of points equal to $f_d^{-2}$ times the original.
It has to be noted that $f_d$ values of 2 and 4 were considered in Ref.~\onlinecite{guemes2019sensing}, although for a test case with larger $Re_{\tau}$.
The reconstruction of the fluctuating velocity fields is evaluated at four different inner-scaled wall-normal distances: $y^+=[15,30,50,100]$.
\subsection{Super-resolution generative adversarial networks}
Super-resolution GAN (SRGAN) is proposed as a method to reconstruct turbulent wall-measurement fields.
Additionally, SRGANs are explored also for direct estimation of velocity fields in wall-parallel planes.
A typical SRGAN architecture consists of two networks: a generator ($G$) and a discriminator ($D$); $G$ is in charge of generating a high-resolution artificial image $\widetilde{H}_R$ from its low-resolution counterpart $L_R$, whereas $D$ is in charge of distinguishing between high-resolution real images $H_R$ and artificial ones.
Note that the purpose of this work is not to generate a custom architecture to tackle fluid-mechanics cases, since these types of DNNs are already available in the literature~\cite{deng2019super,werhahn2019multi,kim2021unsupervised}.
Therefore, the architecture presented in Ref.~\onlinecite{ledig2017photo} was used in this study.
It uses a CNN as generator, where the main core is composed of 16 residual blocks \cite{he2016deep}, and the resolution increase is carried out at the end of the network by means of $\log_2(f_d)$ sub-pixel convolution layers \citep{shi2016real}.
In the case of flow-field reconstruction with full-resolution wall data as input, the sub-pixel convolution layers are removed.
For the discriminator, convolution layers are also used before adding two fully-connected layers, using a sigmoid activation in the last one to obtain a probability to discern whether the high-resolution input is real or not.
A schematic view of the generator network is shown in Figure~\ref{fig:01}a) and the rest of details can be found in Ref.~\onlinecite{ledig2017photo}.
The discriminator loss is defined as:
\begin{equation}
\mathcal{L}_D=-\mathbb{E}[\log D(H_R)] - \mathbb{E}[\log(1-D(G(L_R)))].
\end{equation}
For the generator loss, we have used the perceptual loss \cite{ledig2017photo}, where the content loss is evaluated with the pixel-based mean-squared error between $H_R$ and $\widetilde{H}_{R}$, leading to:
\begin{equation}
\mathcal{L}_G=\frac{1}{N_xN_z}\sum^{Nx}_{i=1}\sum^{Nz}_{j=1}|G(L_R)_{i,j} - H_R{_{i,j}}|^2 - \lambda \mathcal{L}_D,
\end{equation}
\noindent where $N_x$ and $N_z$ are the number of grid points in the streamwise and spanwise directions for the high-resolution images (192 for both of them in our case) and $\lambda$ is a scalar to weight the value of the adversarial loss, set to $10^{-3}$.
The weights of the model for each downsampling case have been optimized for 30 epochs using the Adam algorithm~\cite{kingmaba} with learning rate $10^{-4}$.
\begin{figure*}
\centerline{
\includegraphics[width=\linewidth]{figs/figure01}}
\caption{Schematic view of the DNN architectures for a) generator network in SRGAN, and b) FCN-POD. The colour coding for each layer is: 2D-convolution \sy{conv}{b}, parametric-ReLU-activation \sy{prel}{b}, batch-normalization \sy{batc}{b}, sub-pix-convolution \sy{subp}{b}, ReLU-activation \sy{relu}{b}, and max-pooling \sy{pool}{b} layers. The kernel size and the number of filters are shown at the bottom of the convolution layers.}
\label{fig:01}
\end{figure*}
\subsection{POD-based fully-convolutional networks}
The baseline method for assessing the quality of flow reconstructions achieved by SRGAN is the FCN-POD approach \cite{guemes2019sensing,guastoni2020convolutional}.
This method divides the turbulent flow fields into $N_s$ two-dimensional subdomains of $N_p\times N_p$ grid points, and POD is performed on each of these subdomains.
The number of subdomains is chosen based on $Re_{\tau}$, with the purpose of ensuring that ~90\% of the flow kinetic energy is contained within $\mathcal{O}(10^2)$ POD modes which can be represented by convolutional filters.
For the $Re_{\tau}=180$ case, each field is divided into $12\times12$ subdomains, each of them with $16\times16$ grid points.
The proposed architecture will reconstruct this three-dimensional tensor of POD coefficients from the wall quantities; this tensor is later converted into the flow field by projecting the POD coefficients of each subdomain into its corresponding basis.
Note that this method does not ensure continuity between subdomains; nevertheless, the convolutional layers have been shown to provide reasonably smooth flow fields\cite{guastoni2020convolutional}.
For each wall-normal distance a different model has been used, the weights of which have been optimized for 30 epochs using the Adam optimizer \cite{kingmaba} with $\epsilon=0.1$, learning rate $10^{-3}$ and an exponential decay starting from epoch 10.
A schematic representation of the architecture is shown in Figure \ref{fig:01}b), and the rest of the implementation details can be found in Ref.~\onlinecite{guastoni2020convolutional}.
For the case of coarse input data, a modified version of the FCN-POD model has been used.
To deal with the different sizes of the input and output tensors, $\log_2(f_d)$ pooling layers have been removed from the original model.
\section{Assessment of resolution enhancement for wall measurements}\label{sec:wall}
The quality of the resolution enhancement of the wall fields is evaluated first.
Figure~\ref{fig:02} shows an instantaneous field of the streamwise and spanwise wall-shear-stress and pressure fluctuations for the DNS reference and the SRGAN predictions.
While the reconstructions from fields with $f_d=4$ and $f_d=8$ recover almost all the flow features present in the DNS references, the instantaneous field for $f_d=16$ exhibits loss of small-scale details.
Moreover, it appears that the high-intensity regions are attenuated for the latter case.
Note however, that the locations and sizes of the largest flow structures are very well represented even for $f_d=16$.
\begin{figure*}
\begin{center}
\includegraphics[width=\linewidth]{figs/pof_fig02.pdf}
\end{center}
\caption{Comparison of the wall-quantity fluctuating fields at $Re_{\tau} = 180$, scaled with their corresponding standard deviation. Reference DNS is reported at left panel, while the six-row panels report the different $f_d$ cases, covering $f_d=4$ (left), $f_d=8$ (center), and $f_d=16$ (right). Odd rows refer to low-resolution inputs, and even ones to the SRGAN predictions. Top two-row panels report streamwise wall-shear stress, middle ones report spanwise wall-shear stress and bottom ones refer to pressure fluctuations.}
\label{fig:02}
\end{figure*}
The first observations on the resolution-enhancement performance with respect to $f_d$ obtained from the inspection of instantaneous fields are confirmed when analyzing the mean-squared-error of those fields.
The errors, normalized with the standard deviation of each quantity, are reported in Table~\ref{tab:01}. SRGANs show excellent results for $f_d=4$ in the three wall quantities, and confirm the performance decay between $f_d=8$ and $f_d=16$.
When assessing the performance differences among wall quantities, it is clear that with larger downsampling factors the errors in the streamwise wall-shear-stress fields are lower than for the other two wall quantities.
This behaviour can be ascribed to the spatial organization of streawise wall-shear-stress fluctuations, which exhibit a characteristic alignment in the streamwise direction.
\begin{table}
\caption{\label{tab:01}Mean-squared-error in the instantaneous wall fields scaled with their corresponding standard deviations.}
\begin{ruledtabular}
\begin{tabular}{lccc}
$f_d$ & $\tau_{w_x}$ & $\tau_{w_z}$ & $p_w$\\
\hline
4 & 0.0187 & 0.0244 & 0.0153 \\
8 & 0.2240 & 0.3041 & 0.2741 \\
16 & 0.6531 & 0.7732 & 0.7461 \\
\end{tabular}
\end{ruledtabular}
\end{table}
The pre-multiplied two-dimensional inner-scaled spectra for the three wall quantities are reported in Figure~\ref{fig:03}.
The high-energy peak containing 90\% or more of $\tau_{w_x}$ is well captured by the predictions with $f_d=4$ and $f_d=8$, while for $f_d=16$ this is not recovered, even showing a significant attenuation of 50\% of the total energy content.
The energy attenuation is even stronger for $\tau_{w_z}$ and $p_w$, where the predictions of $f_d=4$ are the only ones capturing the energy distribution for both quantities.
In the case of $f_d=16$, the attenuation is so significant that even the 50\% energy-content level is not recovered.
The distribution of scales over a larger range also explains why the $\tau_{w_x}$ error is smaller than that of the other two wall quantities, since for the first one the SRGAN architecture deals with a lower parametric space.
\begin{figure}
\centerline{\includegraphics[width=\columnwidth]{figs/pof_fig03.pdf}}
\caption{Pre-multiplied two-dimensional power-spectral densities for a) fluctuating streamwise wall-shear-stress, b) fluctuating spanwise wall-shear-stress, and c) wall-pressure fluctuations. The contour levels contain 10\%, 50\% and 90\% of the maximum DNS power-spectral density. Shaded contours refer to the reference DNS data, while coloured lines denote $f_d=4$ \lcap{-}{fd04}, $f_d=8$ \lcap{-}{fd08}, and $f_d=16$ \lcap{-}{fd16}.}
\label{fig:03}
\end{figure}
Although the scope of this work is not to develop a customized SRGAN architecture for wall turbulence, here we briefly compare our results with those of other studies in the literature.
For example, Ref.~\onlinecite{kim2021unsupervised} used an unsupervised GAN to reconstruct wall-parallel velocity fields at $y^+=15$ and $y^+=100$ in a turbulent channel flow at $Re_{\tau}=1000$ with $f_d=8$.
They report good resolution-enhancing results in terms of instantaneous fields, turbulence statistics and one-dimensional spectra, similar to ours for the same $f_d$.
With respect to the spectra, their work and ours coincide in identifying the small-scale structures as those most difficult to recover.
Because of the different $Re_{\tau}$ in both studies, it is important to highlight that $f_d$ is a pixel ratio between the high- and low-resolution fields, and it does not take into account how many viscous lengths are contained in a single pixel.
For a fair comparison in turbulent flows, we propose the following normalized downsampling factor:
\begin{equation}
\tilde{f}_{d}=f_d\sqrt{\Delta x^{{+}^2} + \Delta z^{{+}^2}},
\label{eq:03}
\end{equation}
\noindent where $\Delta x^{+}$ and $\Delta z^{+}$ are the inner-scaled grid spacing in physical space for $x$ and $z$ respectively.
Using equation~(\ref{eq:03}) yields a normalized downsampling factor $\tilde{f}_{d}\approx 105$, while the work of Ref.~\onlinecite{kim2021unsupervised} tackles a problem with $\tilde{f}_{d}\approx 109$, therefore showing that the comparison is fair.
\section{Prediction of turbulent flow fields \\from coarse wall measurements}\label{sec:flow}
This section presents the reconstruction performance of wall-parallel velocity fields from wall measurements.
Reconstructions are performed at four different wall-normal distances: $y^+=[15,30,50,100]$, considering 4 downsampling cases for the input wall measurements: $f_d=[1,4,8,16]$.
Note that $f_d=1$ means that no information is lost at the wall with respect to the DNS reference.
Figure~\ref{fig:04} shows instantaneous fields of the streamwise velocity fluctuations at the four wall-normal distances of interest in this study.
Predictions generated with SRGAN and FCN-POD networks are compared to the DNS reference.
Note that the FCN-POD predictions are only reported for $f_d=1$ case, and they are analogous to the results presented in Ref.~\onlinecite{guastoni2020convolutional}.
Inspecting the fields, it can be seen that the best results are obtained closer to the wall, with the lowest downsamplings.
When moving away from the wall or reducing the information provided by the wall, the small-scale fluctuations in the fields start to disappear, and one of the networks recover the high-intensity fluctuating regions of the flow.
However, there is a clear performance difference between FCN-POD and SRGAN predictions.
While the loss of small-scale fluctuations is clearly observable at $y^+=30$ for the FCN-POD predictions, SRGAN is able to capture most of them at $y^+=50$.
It is not until $y^+=100$ that it is clearly observed that small-scale fluctuations are not recovered by the SRGAN architecture.
Nonetheless, the results of $f_d=8$ and $f_d=16$ at $y^+=15$ are successful in capturing most of the flow features present in the DNS reference, and the same can be said for $f_d=8$ at $y^+=30$.
Since most of the flow-control techniques actuate over this region \cite{choi1994active,lee1997application,bai2014active}, these results indicate that equally-distributed probes would be sufficient to feed flow information to these control techniques, instead of using image-based acquisition systems, which are more expensive and difficult to implement.
\begin{figure*}
\centerline{\includegraphics[width=\textwidth]{figs/pof_fig04.pdf}}
\caption{Countour map for the streamwise velocity fluctuation fields scaled with the corresponding standard deviation. From top to bottom, rows denote FCN-POD, reference DNS, and SRGAN predictions with $f_d=[1,4,8,16]$ for the wall information. From left to right, columns indicate $y^+=15$, $y^+=30$, $y^+=50$ and $y^+=100$.}
\label{fig:04}
\end{figure*}
A global view of the flow-reconstruction performance is provided in terms of mean-squared-error.
Figure~\ref{fig:05} reports the evolution of the error with respect to the wall-normal distance for the three flow quantities, the four $f_d$ values and the two reconstruction techniques.
There are two aspects to analyze: the performance difference between the two networks, and the evolution of the error with respect to $f_d$.
When comparing the error evolution for both networks, it can be seen that SRGAN outperforms FCN-POD predictions for all $f_d=4$ and $f_d=8$ cases, where the errors for the predictions generated with SRGAN are better than when using FCN-POD approach.
However, for the $f_d=16$ case both errors collapse, and therefore the benefit of using SRGAN disappears.
This deterioration of the flow reconstruction can be ascribed to the low amount of information contained by the coarse wall measurements.
In the case of $f_d=1$, the wall data contains information of the fluctuations with characteristic lengths as small as $\sim10\ell^*$ in the streamwise direction, while for $f_d=16$ this increases up to $\sim160\ell^*$.
While $f_d=4$ and $f_d=8$ recover the small scales present in the DNS reference, $f_d=16$ does not succeed in this task.
In any case, it is important to remark the significant accuracy improvement of SRGAN for $f_d=1$ with respect to the FCN-POD method, especially in the wall-normal and spanwise components.
Although not presented here, this improvement is noticeable also if compared with the FCN method used in Ref.~\onlinecite{guastoni2020convolutional}.
\begin{figure}
\centerline{\includegraphics[width=\columnwidth]{figs/pof_fig05.pdf}}
\caption{Mean-squared-error in the instantaneous fields of a) streamwise, b) wall-normal, and c) spanwise velocity fluctuations scaled with their corresponding standard deviations. Line styles refer to \lcap{-}{fd01} SRGAN, and \lcap{--}{fd01} FCN-POD predictions, respectively. Colours and symbols denote $f_d=1$ \sy{fd01}{o*}, $f_d=4$ \sy{fd04}{t*}, $f_d=8$ \sy{fd08}{dt*}, and $f_d=16$ \sy{fd16}{s*}.}
\label{fig:05}
\end{figure}
The second factor to analyze is the performance decay of the predictions when increasing $f_d$.
In a previous study\cite{guemes2019sensing}, the effect of $f_d$ when reconstructing the large-scale structures present in wall-parallel flow fields from wall measurements on a turbulent channel flow of $Re_{\tau}=1000$ was analyzed.
The analyzed effect of $f_d=[1,2,4]$ reported only a weak deterioration effect due to the increase of $f_d$.
However, the results presented in Figure~\ref{fig:06} show a clear dependency between $f_d$ and the mean-squared-error. Once again, the question arises whether $f_d$ is adequate to characterize the downsampling effect in wall turbulence.
If we used the normalized downsampling factor proposed in equation~(\ref{eq:03}), $f_d=4$ becomes $\tilde{f}_{d}\approx44$ for Ref.~\onlinecite{guemes2019sensing}, while in our case it is $\tilde{f}_{d}\approx52$, increasing to 105 and 210 for $f_d$ values equal to 8 and 16 respectively.
Therefore, it can be argued that in this work we are facing a more challenging wall-information loss.
Furthermore, it must be recalled that the flow scales to be predicted also affect the performance of the method.
Ref.~\onlinecite{guemes2019sensing} only targeted the flow scales in the first 10 POD modes, while this work targets the entire energy spectra.
The first 10 POD modes of Ref.~\onlinecite{guemes2019sensing} refer to the most energetic structures present in the flow.
Large coherent structures are more persistent over time, with lives proportional to their scale\cite{lozano2014time}.
These characteristic length and time scales make them less sensitive to the changes in the resolution of the wall data.
However, small-scale structures are affected, both because they are smaller than the scales contained in the coarse wall data and because the modulation effect of large scales~\cite{hutchins2007large,dogan_modulation} is also hidden by the low-resolution data.
The pre-multiplied two-dimensional energy spectra of the flow quantities at the four wall-normal locations discussed above are shown in Figure~\ref{fig:07}.
As reported in Ref.~\onlinecite{guastoni2020convolutional}, the amount of energy captured by the predictions decreases moving farther from the wall.
Moreover, it is important to note that the FCN-POD method is able to recover scales larger than the subdomain size, although a discontinuity in the spectra can be observed at that wavelength, especially in the wall-normal and spanwise components.
With respect to the effect of using SRGAN as a reconstruction method, the findings presented above are corroborated by the spectra.
The predictions generated with SRGAN recover a wider range of energetic scales in both the streamwise and spanwise wavelengths for the three velocity fluctuations, even up to the case $f_d=8$, while for $f_d=16$ both methods have been shown to provide the same mean-squared error.
Nonetheless, it is also important to mention that for $f_d=16$ at $y^+=100$ no energetic scales above the 10\% of the DNS reference has been recovered in the wall-normal and spanwise spectra.
This also occurs in the spanwise fluctuation spectra at $y^+=50$, but only for the predictions generated without SRGAN.
\begin{figure*}
\centerline{\includegraphics[width=\textwidth]{figs/pof_fig06.pdf}}
\caption{Pre-multiplied two-dimensional power-spectral densities for streamwise (first row), wall-normal (second row), and spanwise (third row) velocity fluctuations. From left to right, columns refer to inner-scaled wall distance $y^+$ equal to 15, 30, 50, and 100. The contour levels contain 10\%, 50\% and 90\% of the maximum DNS power-spectral density. Shaded contours refer to the reference DNS data, while contour lines refer to \lcap{-}{fd01} SRGAN-FCN-POD, and \lcap{--}{fd01} FCN-POD predictions, respectively. Colours denote $f_d=1$ \sy{fd01}{s*}, $f_d=4$ \sy{fd04}{s*}, $f_d=8$ \sy{fd08}{s*}, and $f_d=16$ \sy{fd16}{s*}.}
\label{fig:06}
\end{figure*}
While Figure~\ref{fig:04} shows a large error for the extreme cases i.e., those reconstructions farther from the wall or with high $f_d$ values, a visual inspection of Figure~\ref{fig:05} reveals that even in these case the SRGAN predictions are able to capture the large-scale organization of the instantaneous flow field.
Therefore, the large observed errors can be ascribed to the attenuation of the velocity fluctuations, which makes it necessary to define a metric that evaluates the error based on the scale wavelengths.
Following Ref.~\onlinecite{encinar2019logarithmic}, a spectral fractional error can be defined as:
\begin{equation}
R_{ab}(k_x,y,k_z) = \frac{\mathcal{R}e\langle(a-a^{\dagger})(b-b^{\dagger})^*\rangle(k_x,y,k_z)}{\mathcal{R}e\langle ab^*\rangle(k_x,y,k_z)},
\label{eq:04}
\end{equation}
\noindent where $k_x$ and $k_z$ are the wave numbers in the streamwise and spanwise directions respectively, superscripts `*' and `$\dagger$' refer to complex conjugate and estimated quantities respectively, $\mathcal{R}e$ denotes real part, while $a$ and $b$ stand for either $u$, $v$, or $w$.
Note that equation~(\ref{eq:04}) is related to the linear cohrence spectrum \cite{baars2016spectral,encinar2019logarithmic,tanarro2020effect}.
Figure~\ref{fig:07} shows the iso-contours of $R_{ab}=0.5$ for each case.
The reconstruction performance of $f_d=1$ and $f_d=4$ in the viscous region, which covers entirely the wavelengths above 10\% energy content of the streamwise fluctations, is particularly remarkable.
It can also be observed that the farther from the wall the fewer small scales are recovered in the reconstruction.
A similar behaviour is observed when increasing the downsampling factor $f_d$.
\begin{figure*}[!t]
\centerline{\includegraphics[width=\textwidth]{figs/pof_fig07.pdf}}
\caption{Fractional spectral error for streamwise (first row), wall-normal (second row), and spanwise (third row) velocity fluctuations. From left to right, columns refer to inner-scaled wall-normal locations $y^+$ equal to 15, 30, 50, and 100. The contour level corresponds to $R_{ab}=0.5$. Contour lines refer to \lcap{-}{fd01} SRGAN, and \lcap{--}{fd01} FCN-POD predictions, respectively. Colours denote $f_d=1$ \sy{fd01}{s*}, $f_d=4$ \sy{fd04}{s*}, $f_d=8$ \sy{fd08}{s*}, and $f_d=16$ \sy{fd16}{s*}. Shaded contours refer to pre-multiplied two-dimensional power-spectral densities for the reference DNS data.}
\label{fig:07}
\end{figure*}
Since the SRGAN predictions in the most challenging configurations exhibit some similarities with the filtered fields of the DNS reference, it is of interest to conduct the comparison.
Low-pass filtering has been applied to the DNS reference, where the cut-off lengths are adjusted to retain those scale with $R_{ab}<0.5$.
Figure~\ref{fig:08} shows this comparison for the case at $y^+=50$ with $f_d=8$ at the wall input data.
While the SRGAN prediction does not yield small-scale details, it exhibits a remarkable resemblance in terms of the streak patterns when compared with the filtered DNS.
Note that for the case of Figure~\ref{fig:08}, the cut-off wavelengths are set to $\lambda_x^+\approx500$ and $\lambda_z^+\approx100$.
If the mean-squared error displayed in Figure~\ref{fig:05} is computed with this filtered reference, the error reduces from 0.603 to 0.317 for the case of Figure~\ref{fig:08}.
\begin{figure}
\centerline{\includegraphics[width=\columnwidth]{figs/pof_fig08.pdf}}
\caption{Countour map of a sample streamwise velocity fluctuation field scaled with the corresponding standard deviation at $y^+=50$. Top and bottom panels denote reference filtered DNS and SRGAN prediction with $f_d=8$, respectively.}
\label{fig:08}
\end{figure}
\section{Summary and conclusions}\label{sec:concl}
The reconstruction of wall-parallel velocity fields from coarse measurements at the wall in a wall-bounded turbulent flow has been evaluated in this work, together with the resolution enhancement of the wall measurements.
For that purpose, SRGAN has been proposed for both tasks.
In the case of flow reconstruction from wall measurements, this architecture has been compared with the FCN-POD method proposed by Ref.~\onlinecite{guastoni2020convolutional}.
The resolution enhancement of wall fields from their coarse counterparts has been carried out with donwsampling factors $f_d=[4,8,16]$.
In the case of flow reconstruction, the methods have been evaluated at the following wall-normal locations: $y^+=[15,30,50,100]$ with wall downsampling factors $f_d=[1,4,8,16]$.
SRGAN is shown to provide accurate reconstructions for the case of resolution enhancement of wall fields at $f_d=[4,8]$.
In the most challenging case $(f_d=16)$, it can be observed that small-scale contributions are not recovered in the reconstruction, but the large-scale footprint of the flow at the wall is very well represented.
With respect to the flow reconstruction with full resolution at the wall, SRGAN is shown to provide a significant improvement with respect to the baseline FCN-POD method\cite{guastoni2020convolutional}.
The effect of increasing $f_d$ is also evaluated, showing a clear performance decrease unlike in the work of Ref.~\onlinecite{guemes2019sensing}, where only a weak effect is reported.
This difference is ascribed to two reasons: First, the range of scales targeted in Ref.~\onlinecite{guemes2019sensing} only cover the large wavelengths, while this study does it for the entire spectrum.
Small-scale structures have characteristic time and length scales smaller than that of the filtering bandwidth from the coarse measurements, thus losing relevant information for the reconstruction.
Second, $f_d$ is not an adequate parameter to compare different databases of wall-bounded turbulent flows.
To overcome this issue, we propose to use $\tilde{f}_{d}$, which takes into account the fraction of viscous length covered by a pixel.
With this parameter, the effect of the downsampling is homogenized among the various works, showing a clear trend between the results of Ref.~\onlinecite{guemes2019sensing} and the ones presented here.
To the authors' knowledge this is the first study where DNNs are used to reconstruct flow fields from coarse wall measurements in a turbulent flow, and this approach has great potential in the context of closed-loop control.
In any case, it is observed that the accuracy improvements of SRGAN start to decrease when increasing $f_d$.
The capability of SRGAN to recover the large-scale structures present in the flow is also evaluated by means of the fractional spectral error\cite{encinar2019logarithmic} and assessment of filtered instantaneous fields.
It is shown that the SRGAN predictions are in good agreement with large-scale patterns obtained from the filtered DNS reference.
The present study has used high-resolution DNS data to train the proposed GAN network.
However, the computational cost and requirements for producing this data increase with $Re_{\tau}$, being impossible to obtain for Reynolds numbers from real-life applications.
Consequently, it is advisable to look for alternatives that allow the use of the proposed network in real-life scenarios.
For instance, transfer learning could be explored as in Guastoni et al. (2021), to confirm that GANs are able to generalize from one $Re_{\tau}$ to another.
Another option could be to rely on experimental data obtained from those real-life scenarios.
However, experimental data might be contaminated by noise and have lower spatial and temporal resolution than well-resolved DNS data.
Therefore, future investigations should focus on how to combine different neural networks with incomplete or noisy turbulent data.
\section*{Acknowledgements}
RV acknowledges the support by the G\"oran Gustafsson Foundation.
SD and AI acknowledge the support by the European Research Council, under the COTURB grant ERC-2014.AdG-669505.
HA acknowledges the support by Wallenberg AI, Autonomous Systems, and Software Program (WASP-AI).
We would also like to acknowledge Hampus Tober for useful discussions throughout the present study.
\section*{Data Availability Statement}
The data that support the findings of this study are available from the corresponding author upon reasonable request.
The codes that support the findings of this study are openly available in GitHub at https://doi.org/10.5281/zenodo.5067426.
|
\section{Introduction}
Since the beginning of consumption data collection, the comparison and disaggregation of load curves have been two popular domains of study. While the former focuses on clustering rather than extracting meaningful features, the latter consists in decomposing a load curve as a sum of curves each representing a particular consumption pattern. In both cases, it is assumed that a panel of load curves
$\set{X_j(u)}{u \in \mathcal{U}, \; 1 \leq j \leq J}$
is observed, where $\mathcal{U}$ is an interval of $\mathbb{R}$ (e.g. $\mathcal{U} = [0,24)$) and $j$ is an index (e.g. the day, the customer, the sensor, etc). Numerous clustering methods have been applied to compare such curves (see \cite{CHICCO2012,WEI20181027,Rajabi17} for an overview). On the other hand, load curves disaggregation is usually done in the context of Non Intrusive Load Monitoring (NILM) where the goal is to decompose the globally observed consumption into a sum of the consumption of several devices. Over the past decade, Blind source separation methods such as Nonnegative matrix factorization (NMF) have gained in popularity in the NILM community (see e.g. \cite{Henriet19NILM} and the references therein). The idea is to decompose the load curve of the $r$-th device as a \emph{signature} curve $a_r(u) \geq 0$ which is modulated across observations by an \emph{activation} $b_{j,r} \geq 0$, that is
$
X_j(u) \approx \sum_{r=1}^R a_r(u) b_{j,r} \; .
$
In a multi-site context, disaggregating the load curves can be a way to extract features to describe the load profiles and cluster the sites. Assume we observe a panel of daily load curves
$\set{X_{j,n}(u)}{u \in [0,24), 1 \leq j \leq J, \, 1 \leq n \leq N}$
where $u$ represents the intra-day time, $j$ the observed day and $n$ the site. Generalizing the NMF model of NILM gives the Nonnegative Tensor Factorization (NTF) model,
$
X_{j,n}(u) \approx \sum_{r=1}^R a_r(u) b_{j,r} c_{n,r}
$,
where $a_r(u) \geq 0$ is the \emph{signature} which is modulated across days by the \emph{day activation} $b_{j,r} \geq 0$ and across sites by the \emph{site activation} $c_{n,r} \geq 0$. The signature and activations are refered to as the \emph{factors}.
Tensor factorization models are very popular in chemometrics and psychometrics but their use for electrical load curves analysis is still recent \cite{Figueiredo2014ExploringTP,Figueiredo15NTFNilm, SANDOVAL2020106431}. Multi-sites load curves disaggregation has been proposed in \cite{SANDOVAL2020106431} using PARAFAC, which is the same as NTF without positivity constraints (see \cite{KoBa09,Cichocki-NTF}). In particular, the authors use the site activations for clustering but do not give any interpretation of the factors nor the clusters. In this paper, we propose several modifications of the NTF model in order to orient the decomposition into more interpretable factors. We also use additional knowledge on the outside temperature and different consumption regimes. The model is presented in \Cref{sec:our-model} and is validated on simulated and real data in \Cref{sec:expe}.
The notations used throughout this paper are the following. For all integer $N \geq 1$, $\integersfromoneto{N} = \{1, \cdots, N\}$. For tensor operations, we use the notations of \cite{Cichocki-NTF}, that is $\norm{\cdot}_F$ is the Froebenius norm, $[\cdot]_+$ the positive part, $\oast$ the Hadamard product, $\otimes$ the Kronecker product, $\odot$ the Khatri-Rao product and $\circ$ the tensor product. Note that ${\bf A}^{\oast 2} = {\bf A} \oast {\bf A}$. The unfolding ${\bf X}_{(k)}$ of a tensor ${\bf X}$ is defined in \cite{KoBa09}.
\section{Proposed model}\label{sec:our-model}
The proposed model is based on NTF with the additional assumption that the consumption $X_{j,n}(u)$ only depends on the day of the year $j$ through two variables : the daily temperature $T_{j,n}$ and the consumption regime $\epsilon_{j,n}$ (e.g. business and non-business days). This leads to the following model
\begin{equation}\label{eq:spline-ntf}
X_{j,n}(u) \approx \sum_{r=1}^R a_r(u) b_{r}(T_{j,n}) c_{n,r}^{(\epsilon_{j,n})}\; ,
\end{equation}
where $a_r(u) \geq 0, b_r(t) \geq 0, c_{n,r}^{(\epsilon)} \geq 0$. We also assume that the functions $a_r$ and $b_r$ are smooth and that $a_r$ is periodic with a period of $24$ hours (since it represents intra-day behavior). By analogy to the NMF and NTF cases we call the functions $b_r$ the \emph{thermal activations}.
\subsection{The optimization problem}
Let us consider two grids $(u_1, \cdots, u_I) \subset [0,24) $ and $(t_1, \cdots, t_K) \subset \mathbb{R}$ which contain all the observed intra-day times and temperatures. Let also $E$ be the number of consumption regimes so that $\epsilon_{j,n} \in \integersfromoneto{E}$ for all $j,n$. Then the factors $a_r, b_r$, $c_r^{(\epsilon)}$, are estimated by solving
\begin{align}
&\min \left\lbrace F + P \right\rbrace \nonumber \\
&\text{such that for all } r, i, t, n, \epsilon, \label{eq:optim-problem}\\
&\begin{cases}
a_r(u_i) \geq 0\,, b_r(t_k) \geq 0\, , c_{n,r}^{(\epsilon)} \geq 0\, \\
\int_0^{24} a_r = \int_{t_1}^{t_K} b_r = 1 \,, \\
a_r(0^+) = a_r(24^-) \\
a_r'(0^+) = a_r'(24^-)\\
a_r''(0^+) = a_r''(24^-) \, ,
\end{cases} \nonumber
\end{align}
and
\begin{align*}
F &= \sum_{i,j,n} \left(X_{j,n}(u_i) - \sum_{r=1}^R a_r(u_i) b_r(T_{j,n}) c_{n,r}^{(\epsilon_{j,n})}\right)^2 \, , \\
P &= \alpha \sum_{r=1}^R \int_0^{24} (a_r'')^2 + \beta \sum_{r=1}^R \int_{t_1}^{t_K} (b_r'')^2 \; .
\end{align*}
The scaling constraints prevent the factors from diverging since the error $F$ is not affected by multiplying one of the factors by a constant as soon as the other factors are scaled accordingly. These scaling constraints also make sure that we can compare the site activations.
Finally, the penalizations on the $L^2$-norm of the second derivatives imply that the solutions of Problem \eqref{eq:optim-problem} are necessarily smooth spline functions. Namely, for all $r \in \integersfromoneto{R}$, the function $a_r$ must be a $24$-periodic cubic spline and the function $b_r$ must be a natural cubic spline. Since spline functions are characterized by their sample points, we can reformulate the problem as a weighted NTF.
\subsection{Formulation as a weighted NTF problem}
Classical results on cubic splines imply that there exist $\mathbf{v}_1 \in \mathbb{R}^I$ and ${\bf Q}_1 \in \mathbb{R}^{I\times I}$ positive definite such that for any 24-period cubic spline $a$ on $[0,24)$, we have $\int_0^{24} a = \mathbf{v}_1^\top {\bf a}$ and $\int_0^{24} (a'')^2 = {\bf a}^\top {\bf Q}_1 {\bf a}$, with ${\bf a} = [a(u_1), \cdots, a(u_I)]^\top$. Similarly there exist $\mathbf{v}_2 \in \mathbb{R}^K$ and ${\bf Q}_2\in\mathbb{R}^{K\times K}$ positive definite such that for any cubic spline $b$ on $[t_1, t_K]$, we have $\int_{t_1}^{t_K} b = \mathbf{v}_2^\top \mathbf{b}$, and $\int_{t_1}^{t_K} (b'')^2 = \mathbf{b}^\top {\bf Q}_2 \mathbf{b}$, with $\mathbf{b} = [b(t_1), \cdots, b(t_K)]$.
We now define
${\bf A} \in \mathbb{R}_+^{I \times R}$, ${\bf B} \in \mathbb{R}_+^{K \times R}$
and ${\bf C} \in \mathbb{R}_+^{EN \times R}$ by ${\bf A}_{i,r} = a_r(u_i), \; {\bf B}_{k,r} = b_r(t_k), \; {\bf C}_{(\epsilon-1) N + n,r} = c_{n,r}^{(\epsilon)}$,
and ${\bf W} \in \mathbb{R}^{I \times K \times EN}$, ${\bf X} \in \mathbb{R}^{I \times K \times EN}$ by
$
{\bf W}_{i, k,(\epsilon-1) N + n} = \varsqrt{ \sum_{j=1}^J \mathbbm{1}_{T_{j,n} = t_k} \mathbbm{1}_{\epsilon(j,n) = \epsilon}}
$
and
$$
{\bf X}_{i,k,(\epsilon-1) N + n} = \frac{ \sum_{j=1}^J \mathbbm{1}_{T_{j,n} = t_k} \mathbbm{1}_{\epsilon_{j,n} = \epsilon} X_{j,n}(u_i)}{{\bf W}_{i, k,(\epsilon-1) N + n }^2} \;,
$$
with the convention that $0/0 = 0$.
Then Problem \eqref{eq:optim-problem} is equivalent to
\begin{equation}\label{eq:optim-problem-samples}
\begin{split}
&\min_{{\bf A} \geq 0 , {\bf B} \geq 0, {\bf C} \geq 0} f_{\bf W}({\bf A}, {\bf B}, {\bf C}) \\
& \text{such that for all } r, \mathbf{v}_1^\top {\bf a}_r = \mathbf{v}_2^\top \mathbf{b}_r = 1 \,,
\end{split}
\end{equation}
where ${\bf a}_r, \mathbf{b}_r$ and ${\bf c}_r$ are the $r$-th columns of ${\bf A}, {\bf B}$ and ${\bf C}$ respectively, and
$$
f_{\bf W}({\bf A}, {\bf B}, {\bf C}) = L_{\bf W}({\bf A}, {\bf B}, {\bf C}) + P({\bf A},{\bf B}) \; ,
$$
with
$$
L_{{\bf W}}({\bf A}, {\bf B}, {\bf C}) = \norm{{\bf W} \oast \left({\bf X} - \sum_{r=1}^R {\bf a}_r \circ \mathbf{b}_r \circ {\bf c}_r\right)}_F^2 \; ,
$$
and
$$
P({\bf A},{\bf B}) = \alpha \mathrm{Tr}({\bf A}^\top {\bf Q}_1 {\bf A}) + \beta \mathrm{Tr}({\bf B}^\top {\bf Q}_2 {\bf B}) \; .
$$
\subsection{Fast HALS algorithm}
We solve Problem \eqref{eq:optim-problem-samples} using a Fast HALS algorithm. The idea of HALS is to minimize $f_{\bf W}$ alternatively in the columns of ${\bf A}$, ${\bf B}$, ${\bf C}$. The update in ${\bf a}_r$ is obtained by solving
$$
\min_{\substack{{\bf a}_r \geq 0 \\ \mathbf{v}_1^\top {\bf a}_r = 1}} \norm{{\bf W} \oast ({\bf X}^{(r)} - {\bf a}_r \circ \mathbf{b}_r \circ {\bf c}_r)^2}_F^2 + \alpha {\bf a}_r^T {\bf Q}_1 {\bf a}_r \, ,
$$
where ${\bf X}^{(r)} = {\bf X} - \sum_{s \neq r} {\bf a}_s \circ \mathbf{b}_s \circ {\bf c}_s$. This leads to the following update steps
\begin{align*}
{\bf a}_r &\leftarrow {\bf M}_r^{-1} \left({\bf W}_{(1)} \oast {\bf X}^{(r)}_{(1)}\right)\left({\bf c}_r \otimes \mathbf{b}_r\right) \\
{\bf a}_r &\leftarrow \frac{[{\bf a}_r]_+}{\mathbf{v}_1^\top [{\bf a}_r]_+} \; ,
\end{align*}
where ${\bf M}_r := {\rm diag}({\bf W}_{(1)}^{\oast 2} ({\bf c}_r \otimes \mathbf{b}_r)^{\oast 2}) + \alpha {\bf Q}_1$.
The Fast HALS algorithm uses a clever way to write ${\bf X}^{(r)}$ and ${\bf M}_r$ to avoid repeating computations unnecessarily. The updates in $\mathbf{b}_r$ and ${\bf c}_r$ are obtained similarly.
\section{Experimental results}\label{sec:expe}
We validate our model on two datasets. The first was extracted from energy demand data simulated for $1$ year and with hourly rate by the Office of Energy Efficiency \& Renewable Energy (EERE)\footnote{Available at \url{https://openei.org/doe-opendata/dataset/commercial-and-residential-hourly-load-profiles-for-all-tmy3-locations-in-the-united-states}.}. We took a total of $775$ sites in California, Arizona, Nevada, Utah, Oregon, Idaho and Washington, gathering $5$ different building types. The second dataset consists of energy demand collected by EDF from $108$ supermarkets across France over a period of $1$ year with a sample rate of $10$ minutes. The average external temperature of each day is observed for each site\footnote{EERE's data can be obtained from the TMY3 weather stations using the \texttt{eeweather} Python package.}, EERE's data have one consumption regime and EDF's data have two consumption regimes (closing and opening days). Because of space constraints, we will not discuss the detection of regimes. In order to compare sites of different sizes we scale the observed load curves by the average daily consumption.
For both NTF and our model, we use the Fast HALS updates and the algorithm is stopped when the relative improvement of the loss reaches $10^{-5}$. Each factors was initialized by taking the positive part of the singular vectors of the corresponding unfolding of ${\bf X}$ (e.g. ${\bf a}_r = [{\bf u}_r]_+$ where ${\bf u}_r$ is the $r$-th singular vector of ${\bf X}_{(1)}$). Since there is no foolproof method to select the number of components $R$ (see \cite{Timmerman00-ncomp-parafac,Bro03-ncomp-parafac,Ceulemans06-ncomp-parafac} for ad-hoc methods) and since cross-validation is not straightforward with tensor data (see \cite{Bro08-cv,Owen09-cv}), we take $R = 6$ and $\alpha=\beta=3000$ for EERE's data and $\alpha=\beta=50$ for EDF's data. This choice gives a good balance between goodness of fit and interpretation of the factors.
For clustering, we run $K$-means on the sites activations for all regimes. This means that site $n$ is represented by the feature vector $(c_{n,1}^{(1)}, \cdots, c_{n,R}^{(1)}, \cdots, c_{n,1}^{(E)}, \cdots, c_{n,R}^{(E)})$ where we recall that $E$ is the number of regimes.
\subsection{Results on EERE's data}
The factors obtained by NTF and our model are represented in \Cref{fig:factors-eere} where the colors correspond to the building types. The dependence on the temperature is justified by the fact that the day activations of NTF mainly indicate season changes. Moreover, the advantage of smoothing the signatures is that only the most important peaks are kept, which is valuable for interpretation. The radar plots of the site activations presented in \Cref{fig:radar-eere} show a better separation of the building types with our model (especially between the restaurants). To quantify this observation, we ran K-means with $5$ clusters and compared the clusters with the true labels using the adjusted random index. Our model gives a perfect fit with an adjusted random index of $1$ compared to $0.75$ for NTF which mostly fails to separate the two types of restaurants. In our model, both hotels have a high site activation in Component 3 whose signature is typical of a hotel (high for breakfast and dinner and medium for lunch). Small Hotels tend to heat at night (Component 4) while Large Hotels tend to heat during the day (Component 1). Apartments have a high site activation in Component 4 whose signature and thermal activation are typical of heating in residential buildings. Components 5 and 6 also characterize the Apartments and can be interpreted as holidays. Indeed, the thermal activation of Component 5 is high when it is very cold and very hot (winter and summer holidays) where people are more at home in the middle of the day (where the signature is high) while the thermal activation of Component 6 peaks for medium temperatures and its signature presents a typical working day profile. Finally, the signature of Component 2, which peaks before lunch and dinner (and a bit before breakfast), is characteristic of restaurants. The two types of restaurants differ by the importance of lunch (more important for Quick Service Restaurants) and the amount of heating used (Component 1).
\begin{figure*}
\captionsetup[subfigure]{aboveskip=-8pt,belowskip=-5pt}
\begin{center}
\begin{subfigure}{0.45\textwidth}
\begin{center}
\includegraphics[width=\textwidth]{eere-ntf-factors.png}
\end{center}
\caption{NTF}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\begin{center}
\includegraphics[width=\textwidth]{eere-spline-ntf-factors.png}
\end{center}
\caption{Our model}
\end{subfigure}
\end{center}
\caption{EERE's Dataset : factors \label{fig:factors-eere}}
\end{figure*}
\begin{figure}
\captionsetup[subfigure]{aboveskip=-5pt,belowskip=-1pt}
\begin{center}
\begin{subfigure}{0.45\textwidth}
\begin{center}
\includegraphics[width=\textwidth]{eere-ntf-radar.png}
\end{center}
\caption{NTF}
\end{subfigure}
\begin{subfigure}{0.46\textwidth}
\begin{center}
\includegraphics[width=\textwidth]{eere-spline-ntf-radar.png}
\end{center}
\caption{Our model}
\end{subfigure}
\end{center}
\caption{EERE's Dataset : site activations \label{fig:radar-eere}}
\end{figure}
\subsection{Results on EDF's data}
\begin{figure*}
\captionsetup[subfigure]{aboveskip=-8pt,belowskip=-5pt}
\begin{center}
\begin{subfigure}{0.45\textwidth}
\begin{center}
\includegraphics[width=\textwidth]{super-ntf-factors.png}
\end{center}
\caption{NTF}
\end{subfigure}
\begin{subfigure}{0.45\textwidth}
\begin{center}
\includegraphics[width=\textwidth]{super-spline-ntf-factors.png}
\end{center}
\caption{Our model}
\end{subfigure}
\end{center}
\caption{EDF's Dataset : factors \label{fig:factors-edf}}
\end{figure*}
\begin{figure}
\begin{center}
\includegraphics[width=0.4\textwidth]{super-spline-ntf-radar.png}
\end{center}
\caption{EDF's Dataset : site activations of our model \label{fig:radar-edf}}
\end{figure}
The results obtained for EDF's data are presented in \Cref{fig:factors-edf} where the colors correspond to the clusters obtained by K-means. The number of clusters was selected by the silhouette index (between $2$ and $9$ clusters). The two consumption regimes can be seen in the day activations of NTF which justifies taking them into account in our model. The site activations are not very diverse in NTF and the two clusters obtained mainly differ from each other by Components 1 and 5 which respectively represent the standard profiles of opening and closing days. On the contrary, our model seems to extract more variable site activations (see also \Cref{fig:radar-edf}) thus exhibiting more clusters. The interpretation of the components is also easier because of smoothness and of the dependence on temperature. For example, the first and last components are independent of the temperature and represent some constant behavior. A high site activation in the first component implies small difference between opening and closing hours (and therefore is high for regime 1 which represents closing days) and a high site activation in Component 6 implies big difference between opening and closing hours. The other components represent various heating (Components 2 and 3) and cooling (Components 4 and 5) profiles. The signature of Component 2 is characteristic of heater since it peaks in the morning and then slowly decays because of inertia. The signature and thermal activation of Component 5 indicate that it represents air conditioning which is activated for high temperatures and in the middle of the day. Finally, the thermal activation of Component 4 indicates that it represents food refrigeration which is activated for lower temperatures than air conditioning and reaches a saturation level for high temperatures. The clusters can be interpreted as follows.
Cluster 1 : Small variation between opening and closing hours (Component 1) and small impact of air conditioning (Components 5).
Cluster 2 : Use a lot of air conditioning (Components 5) and medium heating profile (Component 3).
Cluster 3 : High heating (Component 2) and food refrigeration (Component 4).
Cluster 4 : Not much impacted by the temperature but the difference between opening and closing hours is high (Components 6).
Giving a physical meaning to the factors and clusters is highly valuable for monitoring or maintenance purpose. For example, sites in Cluster 2 should reduce their use of air conditioning while for sites in Cluster 3, reducing heating is a priority.
\section{Conclusion}\label{sec:discussion}
We proposed a model based on a functional formulation of the nonnegative tensor factorization model and an associated optimization algorithm for the disaggregation of multi-sites load curves. By taking into account additional information such as the outside temperature and smoothness of the factors, we showed that this model exhibits more meaningful features and clusters than previously used NTF models.
|
\section*{Introduction}
\addcontentsline{toc}{section}{Introduction}
Supposons donnés un corps de nombres $K$, un premier $\ell$ et une $\Zl$-extension $K_\%$ de $K$.\smallskip
Le résultat emblématique de la théorie d'Iwasawa (cf. e.g. \cite {Ser,Was}) affirme que les ordres respectifs $\ell ^{x(n)}$ des $\ell$-groupes de classes d'idéaux $\,\Cl(K_n)$ attachés aux étages finis $K_n$ de la tour $K_\infty/K$, de degrés respectifs $[K_n:K]=\ell^n$ sont donnés pour $n$ assez grand par la formule:\smallskip
\centerline{$x(n) \ =\ \mu \ell ^n \ + \ \lambda n \ + \ \nu$,}\smallskip
\noindent où $\nu$ est un entier relatif, et où $\lambda$ et $\mu$ sont des entiers naturels déterminés par la pseudo-décomposition de la limite projective (pour les applications normes) $\,\C(K_\%) = \varprojlim \, \Cl(K_n)$, regardée comme module de torsion sur l'algèbre d'Iwasawa $\Lambda = \Zl [[\gamma -1]]$ construite sur un générateur topologique $\gamma$ du groupe procyclique $\Gamma = \Gal(K_\infty/K)$. Il est alors commode de réécrire l'égalité précédente sous une forme ne faisant intervenir que ces deux derniers paramètres:\smallskip
\centerline{$x(n) \ \approx \ \mu \ell ^n \ + \ \lambda n$,}\smallskip
\noindent en convenant de tenir pour équivalentes deux suites d'entiers dont la différence est ultimement constante. L'identité obtenue vaut alors identiquement si l'on remplace les $\ell$-groupes $\,\Cl(K_n)$ par leurs quotients respectifs d'exposant $\ell ^n$ (ou $\ell^{n+k}$, pour $k$ fixé), comme expliqué dans \cite{J18}.\medskip
Soient maintenant $S$ et $T$ deux ensembles finis disjoints de places de $K$; et soit $\Cl^{\si{S}}_{\si{T}}(K_n)$ le pro-$\ell$-groupe des $T$-classes $S$-infinitésimales de $K_n$. Ce pro-$\ell$-groupe correspond, par la théorie $\ell$-adique du corps de classes (cf. \cite{J31}), à la pro-$\ell$-extension abélienne maximale de $K_n$ qui est non-ramifiée en dehors des places divisant celles de $S$ et totalement décomposée aux places au-dessus de celles de $T$; et c'est en particulier un $\Zl$-module de type fini. Son quotient d'exposant $\ell^n$, disons ${}^{\ell^n}\!\Cl_{\si{T}}^{\si{S}}(K_n)$, est ainsi un $\ell$-groupe; et on s'attend à ce que la $\ell$-valuation ${x_{\si{T}}^{\si{S}}(n)}$ de son ordre s'exprime asymptotiquement de façon simple à partir des invariants structurels du module d'Iwasawa:\smallskip
\centerline{$\C_{\si{T}}^{\si{S}}(K_\%)= \varprojlim \,\Cl_{\si{T}}^{\si{S}}(K_n)$.}\smallskip
C'est le programme initié dans \cite{J18}, puis développé dans \cite{J40}. La formule attendue\smallskip
\centerline{$x_{\si{T}}^{\si{S}}(n) \ \simeq \ \rho_{\si{T}}^{\si{S}} n\ell^n \ + \ \mu_{\si{T}}^{\si{S}} \ell ^n \ + \ \lambda_{\si{T}}^{\si{S}} n$,}\smallskip
\noindent qui fait intervenir la quantité $\rho_{\si{T}}^{\si{S}}=\dim_\Lambda\,\C_{\si{T}}^{\si{S}}(K_\%)$ (i.e. la dimension sur le corps des fractions $\Phi$ de $\Lambda$ du $\Phi$-espace $\Phi \otimes_\Lambda \C_{\si{T}}^{\si{S}}(K_\%)$) ainsi que la $\ell$-valuation $\mu_{\si{T}}^{\si{S}}$ et le degré $\lambda_{\si{T}}^{\si{S}}$ du polynôme caractéristique du sous-module de $\Lambda$-torsion $\T_{\si{\,T}}^{\si{S}}(K_\%)$ de $\, \C_{\si{T}}^{\si{S}}$ et $ \mu_{\si{T}}^{\si{S}}$, coïncident bien avec les invariants structurels, mais ce n'est pas toujours le cas pour les paramètres $ \lambda_{\si{T}}^{\si{S}}$. Précisons cela:
\begin{Def*} Étant données $(a_n)_{n\in\NN}$ et $(b_n)_{n\in\NN}$ deux suites d'entiers relatifs, convenons d'écrire:\smallskip
\centerline{$a_n \simeq b_n$, lorsque la différence $a_n-b_n$ est bornée; et $a_n \approx b_n$, lorsqu'elle est ultimement constante.}
\end{Def*}
Le résultat principal de \cite{J52}, corrigeant \cite{J40}, est le suivant: soit $e$ le plus petit entier $n$ tel que les places non contenues dans $S$ qui se ramifient dans $K_\%/K$ le font totalement dans $K_\%/K_n$ et les places de $T$ finiment décomposées dans $K_\%/K_n$ ne se décomposent pas dans $K_\%/K_n$. Alors:
\begin{Th*}[Jaulent--Maire--Perbet]\label{JMB}
Il existe un entier relatif $\kappa_{\si{T}}^{\si{S}}$ tel que l'on ait le paramétrage:\smallskip
\centerline{$x_{\si{T}}^{\si{S}}(n) \ \simeq \ \rho_{\si{T}}^{\si{S}} n\ell^n \ + \ \mu_{\si{T}}^{\si{S}} \ell ^n \ + \ (\lambda_{\si{T}}^{\si{S}} -\kappa_{\si{T}}^{\si{S}}) n$,}\smallskip
\begin{itemize}
\item[(i)] Dans le cas spécial où l'extension $K_\infty/K$ est elle-même $S$-ramifiée et $T$-décomposée, on a $\kappa_{\si{T}}^{\si{S}}=-1$ ainsi que l'estimation asymptotique stricte:
\centerline{$x_{\si{T}}^{\si{S}}(n) \ \approx \ \rho_{\si{T}}^{\si{S}} \,n\ell^n \ + \ \mu_{\si{T}}^{\si{S}} \,\ell ^n \ + \ (\lambda_{\si{T}}^{\si{S}} +1)\,n$.}\smallskip
\item[(ii)] Dans tous les autres cas, on a: $0 \leqslant \kappa_{\si{T}}^{\si{S}} \leqslant \ell^e \rho_{\si{T}}^{\si{S}}$. En particulier, le paramètre $\kappa_{\si{T}}^{\si{S}}$ est nul dès que l'invariant structurel $\rho_{\si{T}}^{\si{S}}$ l'est, autrement dit dès que $\,\C_{\si{T}}^{\si{S}}(K_\%)$ est un $\Lambda$-module de torsion. Et, dans cette dernière situation, on a même l'estimation asymptotique stricte:\smallskip
\centerline{$x_{\si{T}}^{\si{S}}(n) \ \approx \ \rho_{\si{T}}^{\si{S}} \,n\ell^n \ + \ \mu_{\si{T}}^{\si{S}} \,\ell ^n \ + \ \lambda_{\si{T}}^{\si{S}}\,n$.}
\end{itemize}
\end{Th*}
Le but de la présente note est de rectifier les formules défectueuses de \cite{J31}, énoncées en termes de caractères, et de donner en particulier une formulation correcte des identités du miroir de Gras pour les caractères $\lambda_{\si{T}}^{\si{S}}$ attachés à la $\Zl$-extension cyclotomique d'un corps de nombres. Puis, dans un second temps, d'étudier le caractère de défaut $\kappa_{\si{T}}^{\si{S}}$ correspondant et d'en donner une interprétation arithmétique simple. Enfin, au moins sous certaines hypothèses, de le déterminer explicitement.\smallskip
Pour ne pas alourdir cette étude, nous nous limitons au cas au cas $\ell\ne 2$, techniquement plus facile, mais le cas $\ell=2$ relève essentiellement des mêmes méthodes. Nous supposons que $K$ est une extension abélienne contenant $\mmu_\ell^\ph$ d'un sous-corps totalement réel $F$ , de groupe de Galois $\Delta$ d'ordre étranger à $\ell$. Nous nous intéressons à la $\Zl$-extension cyclotomique $K_\%$ de $K$ et nous définissons le défaut $\kappa_{\si{T}}^{\si{S}}$ correspondant comme un caractère de $\Delta$. Notant alors $\chi \mapsto \chi^*$ l'involution du miroir et $\chi=\chi^{\si{\oplus}}+\chi^{\si{\ominus}}$ la décomposition en composantes réelle et imaginaire, nous prouvons:
\begin{TP}
Soient $\ell$ un premier impair, $K/F$ une extension abélienne contenant $\mmu_\ell^\ph$ d'un sous-corps totalement réel, de groupe de Galois $\Delta$ d'ordre étranger à $\ell$, et $K_\%=KF_\%$ la $\Zl$-extension cyclotomique de $K$. Soient enfin $S$ et $T$ deux ensembles finis disjoints de places finies de $F$ dont la réunion contient l'ensemble $L$ des places au-dessus de $\ell$.
\smallskip
Dans le cas spécial $S \supset L$, $T=\emptyset$, le défaut: $\kappa_{\si{\emptyset}}^{\si{S}}=-1$ est l'opposé du caractère unité. Il suit:\smallskip
\centerline{$\lambda^{\si{S}}_{\si{\emptyset}}=(\lambda_{\si{L}}^{\si{\emptyset}}+\chi_{\si{S}}-1)^*$.}\smallskip
Hors le cas spécial $S \supset L$, $T=\emptyset$, le défaut $\kappa_{\si{T}}^{\si{S}}$ est imaginaire: $\kappa_{\si{T}}^{\si{S\,\oplus}}=0$ et $\kappa_{\si{T}}^{\si{S}}=\kappa_{\si{T}}^{\si{S\,\ominus}}\geqslant 0$ .\smallskip
$(i)$ Si $T$ contient $L$, on a de plus: $\kappa_{\si{T}}^{\si{S\,\ominus}}=0$ et le défaut $\kappa_{\si{T}}^{\si{S}}$ est nul. Il suit:\smallskip
\centerline{$\lambda_{\si{T}}^{\si{S\,\ominus}}=\lambda_{\si{L}}^{\si{S\,\ominus}}=\lambda_{\si{L}}^{\si{\emptyset\,\ominus}}+(\chi_{\si{S}}^{\si{\,\oplus}}-1)^*$,}\smallskip
\noindent où $\chi_{\si{S}}=\sum_{\p\in S}\ell^{n_{\si{\p}}}\chi_{\p}$ est la somme des induits $\chi_\p=\Ind^\Delta_{\Delta_\p}1_{\Delta_\p}$ des caractères unités des sous-groupes de décomposition respectifs $\Delta_\p$ des $\p$ de $S$ dans $K/F$ comptés avec une multiplicité égale à leur indice de décomposition $\ell^{n_{\si{\p}}}$ dans $K_\%/K$.
\smallskip
$(ii)$ Si $S$ contient $L$, on a: $\kappa_{\si{T}}^{\si{S\,\ominus}}=\chi_{\si{T}}^{\si{L}}$, où $\chi_{\si{T}}^{\si{L}}$ est le caractère de l'image semi-locale $p_{\si{L}}(\E_{\si{T}}^{\si{\ominus}})$ du $\Zl$-module construit sur les $T$-unités imaginaires du corps $K_\%$.
Sous la conjecture de Leopoldt dans $K_\%$ et pour $T$ contenant $ L$, il vient ainsi (par inversion les rôles de $S$ et de $T$):\smallskip
\centerline{$\lambda_{\si{T}}^{\si{S\,\oplus}}=\lambda_{\si{L}}^{\si{S\,\ominus}}=\lambda_{\si{L}}^{\si{\emptyset\,\oplus}}+(\chi_{\si{S}}^{\si{\,\ominus}}-\chi_{\si{S}}^{\si{L}})^*$.}\smallskip
$(iii)$ Pour $F=\QQ$ enfin, et $S$ ne contenant pas $\ell$, le caractère de défaut $\chi_{\si{S}}^{\si{L}}$ est donné par:\smallskip
\centerline{$\chi_{\si{T}}^{\si{L}} = (\chi_{\si{S}}\wedge \ell^{\,\max_{\si{p\in S}}\{n_{\si{p}}\}} \chi_\reg)^{\si{\ominus}}= \sum^{\si{\ominus}}_{\varphi}\,[\, \ell^{\,\max_{\si{p\in S_{\varphi}}}\{n_{\si{p}}\}}]\varphi$,}\smallskip
\noindent où $\varphi$ décrit les caractères irréductibles imaginaires de $\Delta$ et $S_{\varphi}=\{p\in S\;|\; {\varphi}(\Delta_p)=1\}$. Il suit:\smallskip
\centerline{$\lambda_{\si{T}}^{\si{S\,\oplus}}=\lambda_{\si{L}}^{\si{S\,\oplus}}=\lambda_{\si{L}}^{\si{\emptyset\,\oplus}}+ \sum^{\si{\ominus}}_{\varphi}\,[\,( \sum_{p \in S_{\varphi} }\ell^{n_{\si{p}}} ) - \ell^{\,\max_{\si{p\in S_{\varphi}}}\{n_{\si{p}}\}}]\varphi^*$.}\smallskip
En particulier, $\lambda_{\si{T}}^{\si{S\,\oplus}}$ et $\lambda_{\si{L}}^{\si{\emptyset\,\oplus}} = \lambda_{\si{\emptyset}}^{\si{\emptyset\,\oplus}}$ ont même $\varphi^*$-composante dès que $S_{\varphi}$ a au plus 1 élément.
\end{TP}
La dernière assertion recoupe et prolonge les résultats de Itoh, Mizusawa et Ozaki \cite{IMO, Ito} sur le $\Zl$-rang des modules d'Iwasawa modérément ramifiés.\smallskip
Une conséquence {\em a priori} surprenante est que, même dans le cas le plus simple où $F$ est le corps des rationnels, le caractère de défaut $\kappa_{\si{T}}^{\si{S\,\ominus}}$ s'avère ainsi généralement non-trivial.
\newpage
\section{Énoncé du théorème des paramètres en termes de caractères}
Supposons désormais que $K/F$ est une extension abélienne de corps de nombres, de degré $d$ étranger à $\ell$, de groupe de Galois $\Delta$.
L'algèbre $\ell$-adique $\Zl [\Delta ]$ est alors une algèbre semi-locale, produit direct des extensions non ramifiées $Z_\varphi=\Zl[\Delta]e_\varphi$ de $\Zl$; et les idempotents primitifs $e_\varphi$ sont donnés à partir des caractères $\ell$-adiques irréductibles $\varphi $ de $\Delta $ par les formules classiques:\smallskip
\centerline{$e_\varphi\, =\, \frac{1}{d}\sum_{\tau\in \Delta}\varphi(\tau^{-1})\tau$.}\smallskip
De façon toute semblable, l'algèbre de groupe $\Lambda[\Delta]$ s'écrit canoniquement comme produit:\smallskip
\centerline{$\Lambda[\Delta] = \oplus_\varphi \,\Lambda[\Delta] e_\varphi = \oplus_\varphi\, \Lambda_\varphi\ $.}\smallskip
\noindent Et, pour chaque caractère irréductible $\varphi$ du groupe $\Delta$, la $\varphi$-composante $\Lambda_\varphi = \Lambda[\Delta] e_\varphi$ associée à l'idempotent $e_\varphi$ s'identifie à l'algèbre des séries formelles $Z_\varphi [[\gamma -1]]$ en l'indéterminée $\gamma -1$.\medskip
Plus généralement, par action des idempotents primitifs $e_\varphi$ tout $\Lambda[\Delta]$-module noethérien $X$ se décompose naturellement comme somme directe de ses $\varphi$-composantes $X_\varphi = X^{e_\varphi}$, chaque $X_\varphi$ étant pseudo-isomorphe, comme $\Lambda_\varphi$-module noethérien à un unique $\Lambda_\varphi$ module élémentaire:\smallskip
\centerline{$X_\varphi \sim \Lambda_\varphi^{\rho_\varphi} \oplus \big( \oplus_{i=0}^{s_\varphi}\Lambda_\varphi /f_{\varphi,i}\Lambda_\varphi \big) \oplus \big( \oplus_{j=0}^{t_\varphi}\Lambda_\varphi / \ell^{m_{\varphi,i}}\Lambda_\varphi \big)$.}\smallskip
\noindent Notant alors $P_\varphi = \prod_{j=0}^{t_\varphi}\ell^{m_{\varphi,j}}\prod_{i=0}^{s_\varphi}f_{\varphi,i} \ \in \ZZ_\varphi[\gamma-1]$ le polynôme caractéristique du sous-module de $\Lambda_\varphi$-torsion de $X_\varphi$, on obtient ainsi les trois invariants structurels $\rho_\varphi=\dim_{\Lambda_\varphi}X_\varphi$, $\mu_\varphi=\nu_\ell(P_\varphi)$ et $\lambda_\varphi=\deg((P_\varphi)$,
qu'il est commode de coder globalement en introduisant les caractères structurels:\smallskip
\centerline{$\rho = \sum_\varphi \rho_\varphi \ \varphi \ , \qquad
\mu = \sum_\varphi \mu_\varphi \ \varphi \ , \qquad
\lambda = \sum_\varphi \lambda_\varphi \ \varphi $.}\smallskip
Dans ce contexte équivariant, le Théorème principal de \cite{J52} se généralise comme suit:
\begin{Th} [Théorème des paramètres]\label{TP}
Soit $K/F$ une extension abélienne de corps de nombres, de degré $d$ étranger à $\ell$, puis $F_\%=\bigcup F_n$ une $\Zl$-extension arbitraire de $F$ et $K_\%=\bigcup K_n$, avec $[K_n:K]=[F_n:F]=\ell^n$, de sorte qu'on a: $\Delta=\Gal(K_\%/F_\%)\simeq\Gal(K/F)$. Soit enfin $\gamma$ un générateur topologique du groupe procyclique $\Gamma=\Gal(K_\%/K)\simeq\Gal(F_\%/F)$ et $\Lambda=\Zl[[\gamma-1]]$ l'algèbre d'Iwasawa associée.\par
Deux ensembles finis disjoints $S$ et $T$ de places de $F$ étant donnés, notons $\rho^{\si{S}}_{\si{T}}$, $\mu^{\si{S}}_{\si{T}}$ et $\lambda^{\si{S}}_{\si{T}}$ les caractères structurels pour sa structure de $\Lambda[\Delta]$-module attachés à la limite projective (pour la norme) $\,\C^{\si{S}}_{\si{T}}(K_\%)=\varprojlim\,\Cl^{\si{S}}_{\si{T}}(K_n)$ des (pro)-$\ell$-groupes de $T$-classes $S$-infinitésimales des corps $K_n$.\smallskip
Il existe alors un caractère $\ell$-adique virtuel $\kappa^{\si{S}}_{\si{T}}$ de $\Delta$ tel que la $\ell$-valuation $x^{\si{S}}_{\si{T}}(n)_\varphi$ de la $\varphi$-composante du quotient d'exposant $\ell^n$ de $\,\Cl^{\si{S}}_{\si{T}}(K_n)$ soit donnée asymptotiquement par:\smallskip
\centerline{$x^{\si{S}}_{\si{T}}(n)_\varphi \ \approx \ <\rho^{\si{S}}_{\si{T}}, \varphi>n\ell^n \ + \ <\mu^{\si{S}}_{\si{T}}, \varphi>\ell^n \ + \ <\lambda^{\si{S}}_{\si{T}}-\kappa^{\si{S}}_{\si{T}}, \varphi>n $.}\smallskip
\begin{itemize}
\item[(i)] Dans le cas spécial où l'extension $F_\infty/F$ est elle-même $S$-ramifiée et $T$-décomposée, on a $\kappa^{\si{S}}_{\si{T}}=-1$ (opposé du caractère unité) ainsi que l'estimation asymptotique stricte:\smallskip
\centerline{$x^{\si{S}}_{\si{T}}(n) \ \approx \ <\rho^{\si{S}}_{\si{T}}, \varphi>n\ell^n \ + \ <\mu^{\si{S}}_{\si{T}}, \varphi>\ell^n \ + \ <\lambda^{\si{S}}_{\si{T}}+1, \varphi>n $.}\smallskip
\item[(ii)] Dans tous les autres cas, on a: $0 \leqslant \kappa^{\si{S}}_{\si{T}} \leqslant \ell^e\,\rho^{\si{S}}_{\si{T}} $, où $e$ est le plus petit entier $n$ tel que les places ramifiées dans $F_\%/F$ le soient totalement dans $F_\%/F_n$ et les places de $T$ finiment décomposées dans $F_\%/F_n$ ne se décomposent pas dans
$F_\%/F_n$.
En particulier, pour $\varphi$ fixé $<\kappa^{\si{S}}_{\si{T}},\varphi>$ est nul dès que $<\rho^{\si{S}}_{\si{T}},\varphi>$ l'est; autrement dit dès que la $\varphi$-composante de $\,\C^{\si{S}}_{\si{T}}(K_\%)$ est un $\Lambda_\varphi$-module de torsion; auquel cas on a l'estimation asymptotique stricte:\smallskip
\centerline{$x^{\si{S}}_{\si{T}}(n) \ \approx \ <\rho^{\si{S}}_{\si{T}}, \varphi>n\ell^n \ + \ <\mu^{\si{S}}_{\si{T}}, \varphi>\ell^n \ + \ <\lambda^{\si{S}}_{\si{T}}, \varphi>n $.}
\end{itemize}
\end{Th}
\Preuve Le Théorème s'obtient tout simplement en appliquant {\em mutatis mutandis} le résultat de \cite{J52} rappelé plus haut aux $\varphi$-composantes des groupes $\,\Cl^{\si{S}}_{\si{T}}(K_n)$, ce qui revient à remplacer l'anneau des scalaires $\Zl$ par son extension non-ramifiée $Z_\varphi$.\par
Il vient ainsi corriger le Théorème 2.2 de \cite{J43} qui se trouve en défaut lorsque le caractère $\kappa^{\si{S}}_{\si{T}}$ n'est pas nul.
\Remarque Le produit scalaire $<\varphi,\varphi>$ est le degré $\deg\varphi=[Z_\varphi:\Zl]$ du caractère $\ell$-adique irréductible $\varphi$. Il vaut $1$ si et seulement si $\varphi$ est absolument irréductible; par exemple pour $d \mid (\ell-1)$.
\newpage
\section{Conséquence des identités du miroir de Gras}
Nous nous restreignons dans cette section au cas où $\ell$ est impair, $K$ contient les racines $\ell$-ièmes de l'unité et $F_\%$ est la $\Zl$-extension cyclotomique de $F$.\par
Dans ce contexte les identités du miroir de Gras (cf. \cite{Gra1,J43}) mettent en reflet les sous-groupes de $\ell^n$-torsion respectifs des pro-$\ell$-groupes de $S$-classes $T$-infinitésimales et de $T$-classes $S$-infinitésimales attachés aux étages finis $K_n$ de la tour cyclotomique.\smallskip
Précisons quelques notations: pour chaque place $\p_\%$ de $F_\%$, désignons par $\Delta_{\p_\%}$ le sous-groupe de décomposition de $\p_\%$ dans $\Delta=\Gal(K_\%/F_\%)$ (qui ne dépend que de la place $\p$ de $F$ au-dessous de $\p_\%$); notons $\chi_{\p_\%}$ l'induit à $\Delta$ du caractère unité de $\Delta_{\p_\%}$; posons enfin:\smallskip
\centerline{$\chi^\ph_{\si{S}}=\sum_{\p_\%\in S_\%}\chi_{\p_\%} \qquad\&\qquad \chi^\ph_{\si{T}}=\sum_{\p_\%\in T_\%}\chi_{\p_\%}$,}\smallskip
\noindent où la somme porte sur les places de $F_\%$ au-dessus de $S$ et $T$ respectivement.\smallskip
Introduisons le caractère de Teichmüller $\omega$, défini ici comme le caractère de l'action de $\Delta$ sur le module de Tate $\Tl=\varprojlim \mmu_{\ell^n}$; notons $\chi^{\si{-1}}$ le contragrédient d'un caractère $\chi$, donné par $\sigma\mapsto\chi(\sigma^{\si{-1}})$; et écrivons
\centerline{$\chi\mapsto\chi^*=\omega\chi^{\si{-1}}$}\smallskip
\noindent l'involution du miroir. Cela étant, il vient:
\begin{Th}[Théorème de réflexion]
Si le nombre premier $\ell$ est impair, $K$ contient le groupe $\mmu_\ell$ des racines $\ell$-ièmes de l'unité, et si la réunion $S\cup T$ contient l'ensemble $L$ des places de $F$ au-dessus de $\ell$, les caractères intervenant dans le Théorème des paramètres satisfont les identités:\smallskip
\centerline{$\lambda^{\si{T}}_{\si{S}}-\kappa^{\si{T}}_{\si{S}} +(\chi_{\si{S}}-1)\,=\,\big(\lambda^{\si{S}}_{\si{T}}-\kappa^{\si{S}}_{\si{T}} +(\chi_{\si{T}}-1)\big)^*$.}
\end{Th}
\Preuve C'est exactement l'assertion $(iii)$ du Théorème 2.6 de \cite{J43}, une fois corrigés les caractères structurels $\lambda$ des défauts $\kappa$, conformément au Théorème des paramètres plus haut.
\begin{Cor}\label{C1}
Si $S$ contient $L$, il vient directement: $\lambda_{\si{T}}^{\si{S}}=\lambda^{\si{S}}_{\si{\emptyset}}=\big(\lambda_{\si{L}}^{\si{\emptyset}}+(\chi_{\si{S}}-1)\big)^*$.
\end{Cor}
\Preuve Les places modérées (i.e. ne divisant pas $\ell$) étant sans inertie au-dessus de $K_\%$, il vient $\,\C_{\si{T}}^{\si{S}}(K_\%)=\C_{\si{\emptyset}}^{\si{S}}(K_\%)$, donc $\lambda_{\si{T}}^{\si{S}}=\lambda^{\si{S}}_{\si{\emptyset}}$; puis, en vertu du cas spécial $\kappa_{\si{\emptyset}}^{\si{S}}=-1$:\smallskip
\centerline{$\lambda^{\si{S}}_{\si{\emptyset}}= \lambda^{\si{S}}_{\si{\emptyset}} -\kappa^{\si{S}}_{\si{\emptyset}} +(\chi_{\si{\emptyset}}-1)=\big(\lambda_{\si{S}}^{\si{\emptyset}} -\kappa_{\si{S}}^{\si{\emptyset}} +(\chi_{\si{S}}-1)\big)^*$.}\smallskip
Et, $\,\C^{\si{\emptyset}}_{\si{S}}(K_\%)$ étant un $\Lambda$-module de torsion, on a: $\rho_{\si{S}}^{\si{\emptyset}}=0$, donc $\kappa_{\si{S}}^{\si{\emptyset}}=0$; d'où le résultat.
\begin{Cor}\label{Cor}
Si, au contraire, $T$ contient $L$, il vient: $\lambda^{\si{S}}_{\si{T}}=\lambda^{\si{S}}_{\si{L}}=\lambda^{\si{\emptyset}}_{\si{L}}+(\chi_{\si{S}}-1-\kappa^{\si{L}}_{\si{S}})^*$.\smallskip
Et le caractère $\kappa_{\si{S}}^{\si{L}}$ vérifie les inégalités: $0 \leqslant \kappa_{\si{S}}^{\si{L}} \leqslant \ell^e\rho^{\si{L}}_{\si{S}} = \ell^e \rho^{\si{L}}_{\si{\emptyset}}$.
\end{Cor}
\Preuve $\,\C^{\si{S}}_{\si{L}}(K_\%)$ est un $\Lambda$-module de torsion; d'où $\rho_{\si{L}}^{\si{S}}=0$ et $\kappa_{\si{L}}^{\si{S}}=0$; puis, comme plus haut:\smallskip
\centerline{$\lambda^{\si{S}}_{\si{T}}+(\chi_{\si{L}}-1)=\lambda^{\si{S}}_{\si{L}}+(\chi_{\si{L}}-1)=\lambda^{\si{S}}_{\si{L}}-\kappa^{\si{S}}_{\si{L}}+(\chi_{\si{L}}-1)= \big(\lambda^{\si{L}}_{\si{S}}-\kappa^{\si{L}}_{\si{S}}+(\chi_{\si{S}}-1)\big)^*$.}\smallskip
\noindent Les places modérées étant sans inertie au-dessus de $K_\%$, il vient $\,\C_{\si{S}}^{\si{L}}(K_\%)=\C_{\si{\emptyset}}^{\si{L}}(K_\%)$ donc $\lambda^{\si{T}}_{\si{S}}=\lambda^{\si{T}}_{\si{\emptyset}}$:\smallskip
\centerline{$\lambda^{\si{S}}_{\si{T}}+(\chi_{\si{L}}-1)= \big(\lambda^{\si{L}}_{\si{\emptyset}} -\kappa^{\si{L}}_{\si{\emptyset}}-(\kappa^{\si{L}}_{\si{S}}-\kappa^{\si{L}}_{\si{\emptyset}})+(\chi_{\si{S}}-1)\big)^*$,}
\noindent c'est à dire:
\centerline{$\lambda^{\si{S}}_{\si{T}}+(\chi_{\si{L}}-1)= \big(\lambda^{\si{L}}_{\si{\emptyset}}-\kappa^{\si{L}}_{\si{\emptyset}}-(\chi_{\si{\emptyset}}-1)-\kappa_{\si{S}}^{\si{L}}+(\chi_{\si{S}}-1)\big)^*$,}\smallskip
\noindent via $\kappa_{\si{\emptyset}}^{\si{L}}=-1$ en vertu du cas spécial; et finalement, par $\kappa^{\si{\emptyset}}_{\si{L}}=\rho^{\si{\emptyset}}_{\si{L}}=0$:\smallskip
\centerline{$\lambda^{\si{S}}_{\si{T}}+(\chi_{\si{L}}-1)=\lambda^{\si{\emptyset}}_{\si{L}}+(\chi_{\si{L}}-1)+(\chi_{\si{S}}-1-\kappa_{\si{S}}^{\si{L}})^*$.}\smallskip
Enfin, le caractère $\kappa_{\si{S}}^{\si{L}}$ vérifie: $0 \leqslant \kappa_{\si{S}}^{\si{L}} \leqslant \ell^e\rho^{\si{L}}_{\si{S}}$; et l'on a $\rho^{\si{L}}_{\si{S}}= \rho^{\si{L}}_{\si{\emptyset}}$, toujours parce que les places modérées non-ramifiées au-dessus de $K_\%$ sont complètement décomposées.\medskip
\begin{Def}
Nous disons que $\kappa_{\si{S}}^{\si{L}}=(\chi_{\si{S}}-1)-(\lambda_{\si{L}}^{\si{S}}-\lambda_{\si{L}}^{\si{\emptyset}})^*$ est le caractère de défaut attaché à l'ensemble de places modérées $S$ dans l'involution du miroir.
\end{Def}
Tout le problème est alors d'évaluer précisément $\kappa_{\si{S}}^{\si{L}}$ et de l'interpréter arithmétiquement.
\newpage
\section{Étude en présence de conjugaison complexe}
Pour aller plus loin, plaçons-nous dans le cas où le corps de base $F$ est totalement réel et notons $\bar\tau\in\Delta$ la conjugaison complexe. Il est habituel dans ce cadre de dire qu'un caractère $\ell$-adique est {\em réel} lorsque tous ses facteurs {\em absolument} irréductibles prennent la valeur $+1$ en $\bar\tau$; qu'il est {\em imaginaire} lorsque tous prennent la valeur $-1$; et de décomposer chaque caractère $\ell$-adique comme somme $\chi=\chi^{\si{\oplus}}+\chi^{\si{\ominus}}$ de ses composantes réelles et imaginaires.\smallskip
Par exemple, il résulte de la finitude bien connue du défaut de Leopoldt dans la $\Zl$-extension cyclotomique (cf. e.g. \cite{Was}) que les caractères $\rho^{\si{S}}_{\si{T}}$ sont imaginaires. En particulier, hors le cas spécial, il suit: $0\leqslant\kappa^{\si{S\,\oplus}}_{\si{T}}\leqslant\ell^e\rho^{\si{S\,\oplus}}_{\si{T}}=0$, donc, en vertu du Corollaire \ref{Cor}:
\begin{Prop}\label{Cas imaginaire}
En dehors du cas spécial $S \supset L\;\&\;T=\emptyset$, où l'on a $\kappa^{\si{S}}_{\si{T}}=-1$, la composante réelle du caractère de défaut est triviale: $\kappa^{\si{S\,\oplus}}_{\si{T}}=0$.\par
En particulier, pour $T\supset L$, convenant d'écrire $\lambda_{\si{L}}$ pour $\lambda_{\si{L}}^{\si{\emptyset}}$, on a:\par
\centerline{$\lambda_{\si{T}}^{\si{S\,\ominus}}=\lambda_{\si{L}}^{\si{S\,\ominus}}=\lambda_{\si{L}}^{\si{\ominus}}+(\chi_{\si{S}}^{\si{\oplus}}-1)^*$.}
\end{Prop}
Pour les composantes réelles, c'est plus compliqué, car le défaut $\kappa^{\si{S\,\ominus}}_{\si{T}}$ peut être non-trivial:
\begin{Prop}
Pour $T\supset L$ et $S$ non-vide, on a seulement $\lambda_{\si{T}}^{\si{S\,\oplus}}=\lambda_{\si{L}}^{\si{S\,\oplus}}$ avec:\smallskip
\centerline{$ 0 \leqslant \lambda_{\si{L}}^{\si{S\,\oplus}} - \lambda_{\si{L}}^{\si{\oplus}}= \big( \chi_{\si{S}}^*-\kappa_{\si{L}}^{\si{S}\,*}\big)^{\si{\oplus}} \leqslant \chi_{\si{S}}^{*\,\si{\oplus}}$.}
\end{Prop}
La Proposition assure en particulier que le caractère de défaut $\kappa^{\si{S\,\ominus}}_{\si{T}}$ est indépendant des places modérées intervenant dans $T$. C'est également le cas des places sauvages, au moins sous la conjecture de Leopoldt. Nous avons, en effet, généralisant un résultat de Greenberg (\cite{Grb1}, Prop. 1):
\begin{Th}\label{GéGb}
Pour tout corps de nombres totalement réel $\bar K$, si la conjecture de Leopoldt est satisfaite pour le premier $\ell$ à chaque étage fini de la $\Zl$-tour cyclotomique $\bar K_\%/ \bar K$, le sous-module du pro-$\ell$-groupe $\,\C^{\si{\,S}}_{\bar K_\%}=\varprojlim \,\Cl^{\si{\,S}}_{\bar K_n}$ construit sur les places au-dessus de $\ell$ est fini.\smallskip
En particulier, dans le contexte de cette note, on a l'égalité: $\lambda_{\si{L}}^{\si{S\,\oplus}}=\lambda_{\si{\,\emptyset}}^{\si{S\,\oplus}}$.
\end{Th}
\Preuve La propriété annoncée étant asymptotique, ce n'est pas restreindre la généralité que de supposer (pour cette démonstration) que les places au-dessus de $\ell$ se ramifient totalement dans $\bar K_\%/\bar K$. Or, sous cette hypothèse, les classes de rayons modulo $S$ des places de $\bar K_n$ au-dessus de $\ell$ sont invariantes par le groupe de Galois $\Gamma_{\!n}=\Gal(\bar K_n/\bar K)$. Tout revient donc à vérifier que les classes ambiges de rayons modulo $S$ restent bornées lorsqu'on monte la tour $\bar K_\%/\bar K$.\smallskip
La formule des classes ambiges, écrite pour les $\ell$-groupes de classes S-infinitésimales (cf. \cite{J18}, Cor. II.2.35, ou l'appendice {\em infra} pour plus de détails) donne ici:
$$
|\,(\,\Cl^{\si{\,S}}_{\bar K_n})^{\Gamma_{\!n}}|= |\,\Cl^{\si{\,S}}_{\bar K}|\;\frac{\prod e_\p(\bar K_n/\bar K)}{[\bar K_n:\bar K]\,\big(\E^{\si{S}}(\bar K):\E^{\si{S}}(\bar K)\cap N_{\bar K{\si{n\!}}/\!\bar K}(\R_{\bar K_{\si{n}}})\big)}
$$
Dans celle-ci le produit des indices de ramification $e_\p(\bar K_n/\bar K)$ au numérateur est tout simplement l'indice normique $\big(\,\U_{\si{L}}(\bar K): N_{ K_{\si{n\!}}/\!\bar K}(\U_{\si{L}}(\bar K_{\si{n}}))\big)$, où $\,\U_L$ désigne le groupe des unités semi-locales attachées aux places de $L$; et $\,\E^{\si{S}}(\bar K)$ au dénominateur est le $\ell$-groupe des unités $S$-infinitésimales.\smallskip
Maintenant, sous la conjecture de Leopoldt, $\,\E^{\si{S}}(\bar K)$ s'identifie par le morphisme canonique de semi-localisation à un sous-module d'indice fini du noyau $\,\U_{\si{L}}^*(\bar K)$ de la norme $N_{\bar K/\QQ}$ dans $\,\U_{\si{L}}(\bar K)$. Nous pouvons donc, à un borné près, remplacer le quotient à droite par
$$
\frac{\big(\,\U_{\si{L}}(\bar K): N_{\bar K_{\si{n\!}}/\!\bar K}(\U_{\si{L}}(\bar K_{\si{n}}))\big)}{[\bar K_n:\bar K]\,\big(\,\U^*_{\si{L}}(\bar K): N_{\bar K_{\si{n\!}}/\!\bar K}(\U^*_{\si{L}}( R_{\si{n}}))\big)}=1.
$$
\begin{Cor}\label{kappa}
En dehors du cas spécial, sous la conjecture de Leopoldt dans $K_\%$, le caractère de défaut $\kappa_{\si{S}}^{\si{L}}$ est donné par:
\centerline{$\kappa_{\si{S}}^{\si{L}} = [\chi_{\si{S}}-(\lambda^{\si{S}}_{\si{L}}-\lambda^{\si{\emptyset}}_{\si{L}})^*]^{\si{\ominus}}= [\chi_{\si{S}}-(\lambda^{\si{S}}_{\si{\emptyset}}-\lambda^{\si{\emptyset}}_{\si{\emptyset}})^*]^{\si{\ominus}}$,}\smallskip
\noindent où le caractère structurel $\lambda^{\si{\emptyset}}_{\si{\emptyset}}$ est attaché à la limite projective des $\ell$-groupes de classes d'idéaux $\,\Cl_{\bar K_n}$ tandis que le caractère $\lambda^{\si{S}}_{\si{\emptyset}}$ l'est à celle des $\ell$-groupes $ \,\Cl^{\si{\,S}}_{\bar K_n}$ de classes de rayons modulo $S$.
\end{Cor}
\newpage
\section{Interprétation arithmétique du caractère de défaut}
Pour interpréter $\kappa^{\si{S}}$, appuyons-nous sur la théorie $\ell$-adique du corps de classes (cf. \cite{J31}).
Rappelons que, pour tout ensemble fini $\Sigma$ de places d'un corps de nombres $N$, le pro-$\ell$-groupe d'idèles associé à la pro-$\ell$-extension abélienne $\Sigma$-ramifiée maximale $N^{\si{\Sigma}}$ de $N$ est le produit $\U^{\si{\Sigma}}\R$, avec\smallskip
\centerline{$\U^{\si{\Sigma}}=\prod_{\p\notin \Sigma}\,\U_\p$ \quad et \quad $\R=\Zl\otimes_\ZZ N^\times$}\smallskip
\noindent où $\,\U_\p=\varprojlim U_\p^{\ell^{\si{k}}}$ est le $\ell$-groupe des unités en $\p$ et $\R$ le groupe des idèles principaux.
Posant alors:\smallskip
\centerline{$\U_{\si{\Sigma}}=\prod_{\p\in \Sigma}\,\U_\p$.}
\noindent on obtient:
\centerline{$\Gal(N^{\si{LS}}/N^{\si{S}}) \simeq \,\U^{\si{S}}\R/\U^{\si{LS}}\R \simeq \,\U_{\si{L}}/(\,\U_{\si{L}}\cap\,\U^{\si{LS}}\R)
\simeq \,\U_{\si{L}}/p_{\si{L}}(\E^{\si{S}})$,}\smallskip
\noindent où $p_{\si{L}}$ est le morphisme canonique de semi-localisation, $\E=\Zl\otimes_\ZZ E$ le $\ell$-adifié du groupe des unités et $\,\E^{\si{S}}=\E\cap\,\U^{\si{S}}$ son sous-groupe $S$-infinitésimal. De façon semblable, il vient:\smallskip
\centerline{$\Gal(N^{\si{LS}}/N^{\si{L}}) \simeq \,\U^{\si{L}}\R/\U^{\si{LS}}\R \simeq \,\U_{\si{S}}/(\,\U_{\si{S}}\cap\,\U^{\si{LS}}\R)
\simeq \,\U_{\si{S}}/p_{\si{S}}(\E^{\si{L}}) \simeq \,\U_{\si{S}}$,}\smallskip
\noindent avec ici $\,\E^{\si{L}}=1$, dès que le corps $N$ vérifie la conjecture de Leopoldt en $\ell$ (cf. \cite{J31}, \S2.3). Et enfin:\smallskip
\centerline{$\Gal(N^{\si{LS}}/N^{\si{L}}N^{\si{S}}) \simeq (\,\U^{\si{L}}\R\cap\,\U^{\si{S}}\R/\U^{\si{LS}}\R)
\simeq (\,\U_{\si{S}}\cap \,\U^{\si{S}}\R)/(\,\U_{\si{S}}\cap\,\U^{\si{LS}}\R)
\simeq p_{\si{S}}(\E)$,}\smallskip
\noindent toujours sous la conjecture de Leopoldt dans $N$.\medskip
Appliquons cela aux divers étages $ \bar K_n$ de la $\Zl$-extension cyclotomique du sous-corps réel $ \bar K$ de $K$ en supposant satisfaite la conjecture de Leopoldt dans $ \bar K_\%$. Par passage à la limite projective pour les applications normes, nous obtenons les isomorphismes de $\Lambda[\Delta]$-modules:\smallskip
\centerline{$\Gal ( \bar K_\%^{\si{LS}} /\bar K_\%^{\si{L}} ) \simeq \overset{_\leftarrow}{\U}_{\si{S}},\qquad
\Gal ( \bar K_\%^{\si{LS}} / \bar K_\%^{\si{L}} \bar K_\%^{\si{S}}) \simeq p_{\si{S}}(\overset{_\leftarrow}{\E}),\qquad
\Gal ( \bar K_\%^{\si{LS}} /\bar K_\%^{\si{S}} ) \simeq \overset{_\leftarrow}{\U}_{\si{L}}/p_{\si{L}}(\overset{_\leftarrow}{\E}{}^{\si{S}})$,}\smallskip
\noindent où $\overset{_\leftarrow}{\E} = \varprojlim \,\E( \bar K_n)$ est la limite projective des groupes d'unités; $\overset{_\leftarrow}{\U}_{\si{L}} = \varprojlim \,\U_{\si{L}}( \bar K_n)$ celle des groupes d'unités locales attachées aux places au-dessus de $\ell$
; et $\,\overset{_\leftarrow}{\U}_{\si{S}} = \varprojlim \,\U_{\si{S}}( \bar K_n) \simeq \prod_{\p_\%\in S}\,\varprojlim\, \mu_{\p_n}$ est un $\Zl[\Delta]$-module projectif de caractère $\omega\chi_{\si{S}}^\ph =\omega\chi_{\si{S}}^{\si{-1}} =\chi_{\si{S}}^*$.\smallskip
Or, nous avons directement, par projection via l'idempotent $\frac{1}{2}(1+\bar\tau)$ (et avec $ \bar K^{\si{L}}_\% \cap \bar K_\%^{\si{S}} = \bar K^{\si{\emptyset}}_\%$):\smallskip
\centerline{$\Gal ( \bar K^{\si{L}}_\% / K_\%) \simeq \,\C^{\si{L\;\oplus}}_{K_\%}{},\qquad
\Gal ( \bar K^{\si{L}}_\%\cap \bar K_\%^{\si{L}} / \bar K_\%) \simeq \,\C^{\si{\emptyset\;\oplus}}_{K_\%},\qquad
\Gal ( \bar K^{\si{S}}_\% / \bar K_\% ) \simeq \,\C^{\si{S\;\oplus}}_{K_\%}$.}\smallskip
L'ensemble de ces résultats est donc résumé dans le diagramme galoisien:
\begin{center}
\setlength{\unitlength}{2pt}
\begin{picture}(104,108)(0,0)
\thinlines
\put(48,102){$ \bar K^{\si{LS}}_\%$}
\put(46,82){$ \bar K^{\si{L}}_\% \bar K^{\si{S}}_\%$}
\put(8,52){$ \bar K^{\si{L}}_\%$}
\put(91,52){$ \bar K^{\si{S}}_\%$}
\put(35,22){$ \bar K^{\si{\emptyset}}_\% = \bar K^{\si{L}}_\% \cap \bar K^{\si{S}}_\%$}
\put(51,0){$ \bar K_\%$}
\bezier{60}(46,102)(18,92)(13,62)
\bezier{60}(56,102)(87,92)(93,62)
\bezier{60}(13,47)(18,22)(46,4)
\bezier{60}(93,47)(88,22)(60,4)
\put(53,99){\line(0,-1){10}}
\put(53,19){\line(0,-1){10}}
\put(49,79){\line(-3,-2){32}}
\put(58,79){\line(3,-2){33}}
\put(50,28){\line(-3,2){33}}
\put(58,28){\line(3,2){33}}
\put(54,92){$p_{\si{S}}(\overset{_\leftarrow}{\E})$}
\put(12,82){$\overset{_\leftarrow}{\U}_{\si{S}}$}
\put(87,82){$\overset{_\leftarrow}{\U}_{\si{L}}/p_{\si{L}}(\overset{_\leftarrow}{\E}{}^{\si{S}})$}
\put(15,22){$\C^{\si{L}}_{ \bar K_\%}$}
\put(85,22){$\C^{\si{S}}_{ \bar K_\%}$}
\put(54,12){$\C_{ \bar K_\%}$}
\end{picture}
\end{center}
Et chacun des six groupes de Galois précédents est un $\Zl|\Delta]$-module noethérien. Ainsi, de l'isomorphisme $\Gal ( \bar K^{\si{S}}_\% / ( \bar K^{\si{L}}_\% \cap \bar K^{\si{S}}_\%)) \simeq \Gal ( \bar K^{\si{L}}_\% \bar K^{\si{S}}_\%/ \bar K^{\si{L}}_\%) \simeq \overset{_\leftarrow}{\U}_{\si{S}}/p_{\si{S}}(\overset{_\leftarrow}{\E})$, nous tirons:
\begin{Prop}\label{kappa*}
Le reflet $\kappa_{\si{L}}^{\si{S}*}$ du caractère de défaut $\kappa^{\si{S}}_{\si{L}}$ est le caractère du $\Zl[\Delta]$-module projectif $\Gal(\bar K^{\si{LS}}_\%/\bar K^{\si{L}}_\%\bar K^{\si{S}}_\%)\simeq p_{\si{S}}(\overset{_\leftarrow}{\E})$.
\end{Prop}
\newpage
\section{Interprétation kummérienne du caractère de défaut}
Plaçons-nous maintenant du point de vue de la Théorie de Kummer (cf. \cite{J18}, I.2 ou \cite{J23}).
Par le miroir, le radical $\bar\r_\%=\Rad(\bar K^\ab_\%/\bar K_\%)$ de la pro-$\ell$-extension abélienne maximale $\bar K^\ab_\%$ du sous-corps réel de $K_\%$ est la composante imaginaire du radical universel $\r_\%=(\Ql/\Zl)\otimes_\ZZ K_\%^\times$.\smallskip
Notons $\,\E_{\si{L}}^{\si{\ominus}}$ le $\Zl$-module construit sur les $L$-unités imaginaires de $K_\%$; puis $\,\E_{\si{S}}^{\si{\ominus}}$ et $\,\E_{\si{LS}}^{\si{\ominus}}$ ceux construits respectivement sur les $L$ et les $LS$-unités imaginaires; et $\,\E_{\si{S}}^{\si{L\,\ominus}}$ le sous-module de $\,\E_{\si{S}}^{\si{\ominus}}$ formé des $\varepsilon$ dont l'image semi-locale $p_{\si{L}}(\varepsilon)$ dans $\,\U_{\si{L}}=\prod_{\l_\%\in L}\U_{\l_\%}$ tombe dans $\mu_L=\prod_{\l_\%\in L}\mu^\ph_{\l_\%}$.
Et considérons le diagramme où apparaissent les pro-$\ell$-extension abéliennes maximales respectivement non-ramifiée $\bar K_\%^{\si{\emptyset}}$, $L$-ramifiée $\bar K_\%^{\si{L}}$, $S$-ramifiée $\bar K_\%^{\si{S}}$ et $LS$-ramifiée $\bar K_\%^{\si{LS}}$ de $\bar K_\%$, en convenant d'écrire $N_\%\asymp N'_\%$ lorsque les corps $N_\%$ et $N'_\%$ coïncident à une extension de degré fini près:
\begin{displaymath}
\xymatrix{
& & \!\!\!\!\!\!\!\!\!\!\!\!\!\! \bar K^{\si{LS}}_\% \asymp \bar K^{\si{\emptyset}}_\%[\!\sqrt[\ell^\%]{\E_{\si{LS}}^{\si{\,\ominus}}}\,] \!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \ar@{-}[dl]_{\kappa^{\si{L}*}_{\si{S}}} \ar@{-}[dr] & &\\
& \bar K^{\si{L}}_\% \bar K^{\si{S}}_\% \ar@{-}[dl] \ar@{-}[dr] & & \!\!\!\!\!\!\!\!\!\bar K^{\si{S}}_\%[\!\sqrt[\ell^\%]{\E_{\si{S}}^{\si{\,\ominus}}}\,] \ar@{-}[dl] \ar@{-}[dr] & \\
\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \bar K^{\si{L}}_\% \asymp \bar K^{\si{\emptyset}}_\%[\!\sqrt[\ell^\%]{\E_{\si{L}}^{\si{\,\ominus}}}\,]\!\!\!\!\!\!\!\!\!\! \ar@{-}[dr] & & \!\!\!\!\!\!\!\!\!\!\!\!\! \bar K^{\si{S}}_\% \asymp \bar K^{\si{\emptyset}}_\%[\!\sqrt[\ell^\%]{\E_{\si{S}}^{\si{L\,\ominus}}}\,]\!\!\!\!\!\!\!\!\!\!\!\!\!\!\! \ar@{-}[dl] \ar@{-}[dr] & & \!\!\!\!\!\!\!\!\!\! \bar K_\%[\!\sqrt[\ell^\%]{\E_{\si{S}}^{\si{\,\ominus}}}\,] \ar@{-}[dl]^{\chi^{\si{L}*}_{\si{S}}} \\
& \bar K^{\si{\emptyset}}_\% \ar@{-}[dr] & & \!\!\!\!\!\!\!\!\!\!\! \bar K_\%[\!\sqrt[\ell^\%]{\E_{\si{S}}^{\si{L\,\ominus}}}\,] \ar@{-}[dl] & \\
& & \bar K_\% & &
}
\end{displaymath}\smallskip
\noindent Par exemple, la suite exacte courte canonique $1 \to (\Ql/\Zl)\otimes_\Zl \E_{\si{L}}^{\si{\ominus}} \to \Rad(\bar K_\%^{\si{L}}/\bar K_\%) \to \Cl_{\si{L}}^{\si{\ominus}} \to 1$, où le $\ell$-groupe des $L$-classes $\Cl_{\si{L}}^{\si{\ominus}}$ s'identifie au radical $\Rad(\bar K^\cd_\%/\bar K_\%)$ attaché à la pro-$\ell$-extension abélienne partout complètement décomposée sur $\bar K_\%$, fournit l'identité: $ \bar K^{\si{L}}_\% \asymp \bar K^{\si{\emptyset}}_\%[\!\sqrt[\ell^\%]{\E_{\si{L}}^{\si{\,\ominus}}}\,]$, puisque $K^{\si{\emptyset}}_\%$ est de degré fini sur $K^\cd_\%$ sous la conjecture de Leopoldt par \cite{Grb1}, Prop. 1.\smallskip
Pour chacun des neuf corps $N_\%$ représentés, le groupe de Galois $\Gal(N_\%/\bar K_\%)$ est un $\Lambda[\Delta]$-module noethérien de $\Lambda$-torsion déterminé à pseudo-isomorphisme près par ses invariants $\lambda$ et $\mu$. Le caractère $\mu$ est nul pour les trois corps $\bar K_\%$, $\bar K_\%[\!\sqrt[\ell^\%]{\E_{\si{S}}^{\si{L\,\ominus}}}\,]$ et $\bar K_\%[\!\sqrt[\ell^\%]{\E_{\si{S}}^{\si{\,\ominus}}}\,]$; il est égal à $\mu=\mu^{\si{\emptyset}}_{\si{\emptyset}}$, pour les six autres (donc conjecturalement nul). De plus, le diagramme est pseudo-exact en ce sens que, pour chacun des cinq losanges galoisiens représentés, l'intersection des sommets latéraux est de degré fini sur le sommet du bas; et le sommet du haut est de degré fini sur leur compositum.\smallskip
Par exemple, l'intersection $\bar K^{\si{L}}_\% \cap \bar K_\%[\!\sqrt[\ell^\%]{\E_{\si{S}}^{\si{\,\ominus}}}\,]$, qui est non-ramifiée aux places au-dessus de $S$, est de degré fini sur $\bar K_\%$; de sorte que $\Gal(\bar K^{\si{L}}_\% \bar K_\%[\!\sqrt[\ell^\%]{\E_{\si{S}}^{\si{\,\ominus}}}\,]/\bar K_\%) \sim \Gal(\bar K^{\si{L}}_\%/ \bar K_\%)\times\Gal(\bar K_\%[\!\sqrt[\ell^\%]{\E_{\si{S}}^{\si{\,\ominus}}}\,]/\bar K_\%)$ est un $\Lambda[\Delta]$-module de caractères $\lambda^{\si{L}}+\chi^*_{\si{L}}$ et $\mu$, tout comme $\Gal( \bar K^{\si{LS}}_\%/\bar K_\%)$; et que $\bar K^{\si{LS}}_\%$ est donc de degré fini sur $\bar K^{\si{L}}_\% \bar K_\%[\!\sqrt[\ell^\%]{\E_{\si{S}}^{\si{\,\ominus}}}\,] \asymp \bar K^{\si{\emptyset}}_\%[\!\sqrt[\ell^\%]{\E_{\si{LS}}^{\si{\,\ominus}}}\,]$. Il suit de là que la sous-extension maximale $\bar K_\%^{\si{S}}$ de $\bar K_\%^{\si{LS}} \asymp \bar K^{\si{\emptyset}}_\%[\!\sqrt[\ell^\%]{\E_{\si{LS}}^{\si{\,\ominus}}}\,]$ qui est non-ramifiée aux places au-dessus de $L$ est de degré fini sur $\bar K^{\si{\emptyset}}_\%[\!\sqrt[\ell^\%]{\E_{\si{S}}^{\si{L\,\ominus}}}\,]$.\smallskip
En particulier, du pseudo-isomorphisme $\Gal(\bar K^{\si{LS}}_\%/\bar K^{\si{L}}_\%\bar K^{\si{S}}_\%)\sim\Gal(\bar K_\%[\!\sqrt[\ell^\%]{\E_{\si{S}}^{\si{\,\ominus}}}\,] / \bar K_\%[\!\sqrt[\ell^\%]{\E_{\si{S}}^{\si{L\,\ominus}}}\,])$, on tire d'après la Proposition \ref{kappa*}:
\begin{Th}
Sous la conjecture de Leopoldt dans $K_\%$, le caractère de défaut $\kappa^{\si{S}}_{\si{L}}$ est le caractère $\chi^{\si{S}}_{\si{L}}$ du $\Zl[\Delta]$-module $\,\E_{\si{S}}^{\si{\ominus}}/\E_{\si{S}}^{\si{L\,\ominus}}$; i.e. de l'image semi-locale $p_{\si{L}}(\E_{\si{S}}^{\si{\ominus}})$ modulo torsion du $\Zl[\Delta]$-module $\E_{\si{S}}^{\si{\ominus}}$ des $S$-unités imaginaires de $K_\%$.
\end{Th}
\Remarque Un intérêt essentiel de ce résultat est que le pro-$\ell$-groupe $\E_{\si{S}}^{\si{\ominus}}$ des $S$-unités imaginaires de $K_\%$, et donc son image $p_{\si{L}}(\E_{\si{S}}^{\si{\ominus}})$, se lisent à un étage fini $K_{n_{\si{0}}}$ de la tour cyclotomique; en fait dès que celles des places au-dessus de $S$ qui sont décomposées par la conjugaison complexe ne se décomposent pas dans $K_\%/K_{n_{\si{0}}}$.
\newpage
\section{Calcul du caractère de défaut pour $F=\QQ$}
Pour illutrer les résultats précédents, regardons le cas le plus simple $F=\QQ$, ce qui nous permettra de retrouver et de préciser un résultat d'Itoh, Mizusawa et Ozaki \cite{IMO}.
\smallskip
Désignons par $K_{n_{\si{0}}}$ le plus petit étage de la $\Zl$-tour $K_\%$ au-dessus duquel les places de $S$ ne se décomposent plus et par $G_0$ le groupe $\Gal (K_{n_{\si{0}}}/\QQ)$. Le $\Zl$-module $\,\E_{\si{S}}^{\si{\,\ominus}}$ des $S$-unités imaginaires,regardé modulo son sous-groupe de torsion, est alors un $\Zl[G_0]$-module projectif de caractère $\psi_{\si{S}}^{\si{\ominus}} =\sum_{p\in S} \psi_p^{\si{\ominus}}$, où $\psi_p =\Ind^{G_{\si{0}}}_{G_p}1_{G_p}$ est l'induit à $G_{\si{0}}$ du caractère unité du sous-groupe de décomposition $G_p$ de $p$ dans $K_{n_{\si{0}}}/\QQ$.\smallskip
Et la conjecture générale d'indépendance $\ell$-adique établie dans \cite{J10} pour les corps abéliens affirme que le caractère de l'image semi-locale $p_{\si{L}}(\E_{\si{S}}^{\si{\ominus}})$ (toujours modulo torsion) est le plus grand caractère $\psi_{\si{S}}^{\si{L\,\ominus}}=\psi_{\si{S}}^{\si{\ominus}}\wedge \psi^\reg_G$ de $G$ qui est contenu dans $\psi_{\si{S}}^{\si{\ominus}}$ comme dans le caractère régulier $\psi^\reg_G$.\smallskip
Par restriction à $\Delta = \Gal(K/F)$, il vient ainsi:
\begin{Th}\label{Calcul}
Soient $\ell$ un nombre premier impair, $K\supset\QQ[\mmu_\ell]$ une extension abélienne du corps des rationnels de degré étranger à $\ell$ et $S$ un ensemble fini de nombres premiers $p\ne\ell$. Pour chaque $p\in S$ , notons $K_{n_{\si{p}}}$ le sous-corps de décomposition de $p$ dans la tour cyclotomique $K_\%/K$ et $\Delta_p$ son sous-groupe de décomposition dans $\Delta=\Gal(K/\QQ$).\smallskip
Le caractère du $\Zl[\Delta]$-module $p_{\si{L}}(\E_{\si{S}}^{\si{\ominus}})$ est alors: $\chi_{\si{S}}^{\si{L}}=\sum^{\si{\ominus}}_{\varphi}\ell^{\,\max_{\si{p\in S_{\varphi}}}\{n_{\si{p}}\}}\, \varphi$.\smallskip
Dans cette formule $\varphi$ décrit les caractères $\ell$-adiques irréductibles imaginaires de $\Delta$ et, pour $\varphi$ fixé, $p$ décrit le sous-ensemble $S_{\varphi}=\{p\in S\;|\; {\varphi}(\Delta_p)=1\}$ des $p$ qui vérifient ${\varphi} \leqslant \chi_p$.
\end{Th}
\begin{Cor}
Sous les hypothèses du théorème, et pour tout $T \supset L$ on a: $\lambda_{\si{T}}^{\si{S\,\oplus}}=\lambda_{\si{\emptyset}}^{\si{S\,\oplus}}$ et:\smallskip
\centerline{$\lambda_{\si{T}}^{\si{S\,\oplus}}=\lambda_{\si{L}}^{\si{S\,\oplus}}=\lambda_{\si{L}}^{\si{\emptyset\,\oplus}}+ \sum^{\si{\ominus}}_{\varphi}\,[\,( \sum_{p \in S_{\varphi} }\ell^{n_{\si{p}}} ) - \ell^{\,\max_{\si{p\in S_{\varphi}}}\{n_{\si{p}}\}}]\varphi^*$.}\smallskip
En particulier, $\lambda_{\si{T}}^{\si{S\,\oplus}}$ et $\lambda_{\si{L}}^{\si{\emptyset\,\oplus}} = \lambda_{\si{\emptyset}}^{\si{\emptyset\,\oplus}}$ ont même $\varphi^*$-composante dès que $S_{\varphi}$ a au plus 1 élément.
\end{Cor}
\Exemple Prenant $K=\QQ[\mmu_\ell]$ et $\varphi^*=1$, on obtient la valeur donnée dans \cite{IMO}:\smallskip
\centerline{$\lambda^{\si{S}}(\QQ_\%)=\big(\sum_{p \in S_\omega}\ell^{n_{\si{p}}}\big)-\ell^{\,\max_{\si{p\in S_\omega}} \{n_{\si{p}}\}}$.}\smallskip
Ici la somme et le maximum portent sur les seuls premiers $p$ de $S$ pour lesquels $\omega=1^*$ est représenté dans $\chi_p$, i.e. sur les premiers $p$ de $S$ complètement décomposés dans $\QQ[\mmu_\ell]/\QQ$.
\medskip
\Remarque Pour les composantes imaginaires, on a directement, en vertu du Corollaire \ref{Cor}:\smallskip
\centerline{$\lambda^{\si{S\,\ominus}}_{\si{T}}=\lambda^{\si{S\,\ominus}}_{\si{L}}=\lambda^{\si{\emptyset\,\ominus}}_{\si{L}}+(\chi^{\si{\oplus}}_{\si{S}}-1)^*$.}\medskip
\Preuve Le Corollaire résulte du Corollaire \ref{Cor}, compte tenu de l'égalité $\kappa^{\si{S\,\ominus}}_{\si{L}}=\chi^{\si{S\,\ominus}}_{\si{L}}$ donnée par le Théorème \ref{kappa} et de l'expression de $\chi^{\si{S\,\ominus}}_{\si{L}}$ donnée par le Théorème ci-dessus.\par
Et lorsque $S_\varphi$ est soit vide soit un singleton $\{p\}$, le terme correctif $\big(\sum_{p \in S_\omega}\ell^{n_{\si{p}}}\big)-\ell^{\,\max_{\si{p\in S_\omega}}}\{n_{\si{p}}\}$ est nul. C'est évidemment toujours le cas lorsque $S$ lui-même est un singleton.
\medskip
\Remarque
En cohérence avec l'ensemble de la note, nous avons supposé dans cette section que le degré $[K:F]$ de l'extension abélienne considérée est étranger à $\ell$. Mais cette restriction n'est nullement nécessaire pour établir le Théorème \ref{Calcul}: dans le diagramme kummérien de la section précédente, les divers groupes de Galois qui interviennent sont, en effet, des $\Zl$-modules noethériens. Or, pour un tel module $\X$, l'invariant $\lambda$ d'Iwasawa n'est autre que la dimension du $\Ql$-espace vectoriel $V_\X=\Ql\otimes_\Zl\X$.\par
Il est donc encore loisible de décomposer $V_\X$ comme somme directe de ses composantes isotypiques $V_\X^{e_{\si{\varphi}}}$ indexées par les caractères $\ell$-adiques irréductibles de $\Delta$ et de définir le {\em caractère} $\lambda$ de $\Delta$ attaché à $\X$ par la formule $\lambda= \sum_\varphi\, \lambda_\varphi \varphi$, avec $\lambda_\varphi =\dim_{\Ql} V_\X^{e_{\si{\varphi}}}$, quand bien même l'hypothèse $\ell\nmid [K:F]$ serait en défaut. De ce fait, les formules obtenues pour les composantes réelles des invariants $\lambda$ sont encore valides dans ce cadre plus large.\par
On retrouve ainsi, par exemple, le fait que pour un corps abélien réel $\bar K$, on a $\lambda_{\si{T}}^{\si{S}}=\lambda_{\si{\emptyset}}^{\si{\emptyset}}$ lorsque $S$ est un singleton $\{p\}$ avec $p\ne\ell$, quel que soit $T$ fini ne contenant pas $p$.
\newpage
\section*{Appendice: Suite exacte des classes $S$-infinitésimales ambiges}
\addcontentsline{toc}{section}{Appendice: Identité des classes $S$-infinitésimales ambiges}
Nous reproduisons dans cet appendice pour la commodité du lecteur une preuve succincte de la formule des classes ambiges dans le cas particulier des $\ell$-classes $S$-infinitésimales qui nous intéresse ici. Nous renvoyons à \cite{J18}, II.2 pour une étude équivariante plus générale.\smallskip
Les données sont les suivantes: $\ell$ est un nombre premier impair; $L/K$ est une $\ell$-extension cyclique de groupe $\Gamma$, et $S$ est un ensemble fini de places finies de $K$ étrangères à $\ell$.\par
Pour chacun des corps ci-dessus, par exemple $L$, le $\ell$-groupe des classes $S$-infinitésimales $\,\Cl_L^{\,\si{S}}$ n'est autre que le $\ell$-adifié $\Zl\otimes_\ZZ Cl_L^{\,\m}$ du groupe des classes de rayons modulo $\m_{\si{L}}^{\si{S}}=\prod_{\p_{\si{L}}\mid S}\p_{\si{L}}$ défini comme le quotient $Cl_L^{\,\m}=D_L^{\,\m}/P_L^{\,\m}$ du groupe $D_L^{\,\m}$ des diviseurs étrangers à $\m_{\si{L}}^{\si{S}}$ par le sous-groupe $P_L^{\,\m}$ des diviseurs principaux engendrés par les $x$ de $L^\times$ qui vérifient $x\equiv 1 \;{\rm mod}^\times \m_{\si{L}}^{\si{S}}$.\smallskip
Il vient: $\Cl_L^{\,\si{S}}=\D_L^{\,\si{S}}/\P_L^{\,\si{S}}$ avec $\D_L^{\,\si{S}}=\Zl\otimes_\ZZ D_L^{\,\si{S}}$ et $\P_L^{\,\m}=\{(x)\in \D_L^{\,\m} \,|\, p_{\si{S}}(x)=1\}$, puisqu'aux places étrangères à $\ell$, les $\ell$-adifiés $\,\U_{\p^\ph_{\si{L}}}\!$ des groupes d'unités locales $U_{\p^\ph_{\si{L}}}\!$ se réduisent aux $\ell$-groupes $\mu_{\p^\ph_{\si{L}}}\!$ de racines de l'unité, de sorte que les éléments de $\R_N=\Zl\otimes_\ZZ N^\times$ construits sur les $x$ de $L^\times$ qui vérifient la congruence précédente sont précisément ceux d'image locale triviale aux places au-dessus de $S$; i.e. les éléments du sous-groupe $S$-infinitésimal $\R_L^{\,\m}=\{x\in \R_L \,|\, p_{\si{S}}(x)=1\}$.
\begin{Th*}[Classes $S$-infinitésimales ambiges]
Soient $\ell$ un nombre premier impair, $L/K$ une $\ell$-extension cyclique de groupe $\Gamma$ et $S$ un ensemble fini de places finies de $K$ étrangères à $\ell$.\par
Alors le nombre de $\ell$-classes de $\,\Cl^{\,\si{S}}_L$ qui sont invariantes par $\Gamma$ est donné par:
$$
|(\Cl^{\,\si{S}}_L)^\Gamma| = |\Cl^{\,\si{S}}_K|\;\frac{\prod_{\p_{\si{K}}\notin S} e_{\p_{\si{K\!}}}(L/K) } { [L:K]\;\big(\E^{\,\si{S}}_K:\E^{\,\si{S}}_K\cap N_{L/K}(\R^\ph_L)\big)}
$$
\noindent où $e_{\p^\ph_{\si{K\!}}}(L/K)$ est l'indice de ramification de $\p_{\si{K}}$ et $\,\E_K^{\,\si{S}}$ le groupe des unités $S$ -infinitésimales.
\end{Th*}
\Preuve Elle est essentiellement identique à celle de la formule de Chevalley (\cite{Chv}, pp. 402--406).\smallskip
$(i)$ Comparaison des classes ambiges et des classes d'ambiges: on dispose d'un isomorphisme\smallskip
\centerline{$(\Cl^{\,\si{S}}_L)^\Gamma/cl^{\si{S}}(\D^{\,\si{S}}_L)^\Gamma)\simeq \E^{\,\si{S}}_K/\E^{\,\si{S}}_K\cap N_{L/K}(\R^\ph_L)$,}\smallskip
\noindent obtenu en prenant un générateur arbitraire $\sigma$ de $\Gamma$ et en envoyant la classe d'un idéal $\a$ qui vérifie $\a^{\sigma-{\si{1}}}=(\alpha)$ sur celle de l'élément $\varepsilon = N_{L/K}(\alpha)$. D'où l'identité:\smallskip
\centerline{$\big((\Cl^{\,\si{S}}_L)^\Gamma:cl^{\si{S}}(\D^{\,\si{S}}_L)^\Gamma)\big) = \frac{\big(\E^{\,\si{S}}_K: N_{L/K}(\E^{\,\si{S}}_L)\big) }{\big(\E^{\,\si{S}}_K \cap N_{L/K}(\R^\ph_L) : N_{L/K}(\E^{\,\si{S}}_L) \big)}$}\smallskip
$(ii)$ Comparaison des classes d'ambiges et des classes étendues: on a l'égalité immédiate\smallskip
\centerline{$|cl^{\si{S}}\big((\D^{\,\si{S}}_L)^\Gamma\big)| =\big((\D^{\,\si{S}}_L)^\Gamma:(\P^{\,\si{S}}_L)^\Gamma\big) = \frac{\big((\D^{\,\si{S}}_L)^\Gamma:\D^{\,\si{S}}_K)\big) \; \big(\D^{\,\si{S}}_K:\P^{\,\si{S}}_K)\big) }{\big((\P^{\,\si{S}}_L)^\Gamma:\P^{\,\si{S}}_K)\big) } $}\smallskip
\noindent avec, au numérateur, $\big((\D^{\,\si{S}}_L)^\Gamma:\D^{\,\si{S}}_K)\big) = \prod_{\p_{\si{K}}\notin S} e_{\p_{\si{K\!}}}(L/K)$ et
$ \big(\D^{\,\si{S}}_K:\P^{\,\si{S}}_K\big) = |\Cl^{\,\si{S}}_K|$.\medskip
$(iii)$ Interprétation cohomologique du dénominateur: $\big((\P^{\,\si{S}}_L)^\Gamma:\P^{\,\si{S}}_K)\big)=|H^1(\Gamma,\E_L^{\,\si{S}})|$\smallskip
Partant de la suite exacte courte $1 \to \E^{\si{S}}_L \to \R^{\si{S}}_L \to \P^{\si{S}}_L \to 1$, prenant ensuite la suite longue de cohomologie et la comparant à la suite de départ écrite pour $K$, on obtient la suite exacte\smallskip
\centerline{$1 \to \P^{\,\si{S}}_K \to (\P^{\,\si{S}}_L)^\Gamma \to H^1(\Gamma,\E^{\si{S}}_L) \to H^1(\Gamma,\R^{\si{S}}_L)$}\smallskip
\noindent Or, prenant la suite de localisation $1 \to \R_L^{\si{S}} \to \R_L \to \prod_{\p\in S} \R_{L_\p} \to 1$, puis la cohomologie, on a:\smallskip
\centerline{$1 \to \R_K^{\si{S}} \to \R_K \to \prod_{\p\in S} \R_{K_\p} \to H^1(\Gamma, \R_L^{\si{S}}) \to H^1(\Gamma, \R_L)=1$}\smallskip
\noindent et le terme de droite est trivial en vertu du Théorème 90 de Hilbert; d'où: $H^1(\Gamma, \R_L^{\si{S}})=1$.\medskip
$(iv)$ Utilisation du quotient de Herbrand $q(\Gamma,E_L)=\frac{|H^2(\Gamma,E_L)|}{|H^1(\Gamma,E_L)|}=\frac{1}{[L:K]}$ (cf. Herbrand \cite{He1,He2}):\smallskip
Observant que $\,\E^{\si{S}}_L$ est d'indice fini dans $\,\E_L$ on a: $q(\Gamma,\E^{\si{S}}_L)=q(\Gamma,\E_L)=q(\Gamma,E_L)=\frac{1}{[L:K]}$.\medskip
Récapitulant le tout, on obtient le résultat annoncé.
\newpage
\section*{\normalsize\sc Addendum}
L'erreur sur le module de défaut, apparue initialement dans \cite{J18}, a été reproduite sans changement dans \cite{J40} puis dans \cite{J43} avant d'être repérée par Salle \cite{Sal} et corrigée, en collaboration avec Maire et Perbet, dans \cite{J52}. Comme expliqué dans l'introduction, le but initial de la présente note est de préciser cette correction en termes de caractères et de donner une formulation exacte des identités du miroir de Gras pour les modules d'Iwasawa.\smallskip
Les résultats obtenus recoupent ceux de Itoh, Mizusawa et Ozaki \cite{IMO} ainsi que ceux de Itoh \cite{Ito} sur les modules d'Iwasawa modérément ramifiés. L'approche d'Itoh, totalement différente de la nôtre, repose sur le théorème de Kronecker-Weber et la description explicite des annulateurs pour les modules d'Iwasawa dans les tours cyclotomiques. Accessoirement elle utilise en outre les résultats de Khare et Wintemberger \cite{KW1,KW2} sur certains radicaux de Kummer.\smallskip
Je remercie enfin tout particulièrement Ch. Maire et G. Gras pour leur lecture critique.
\bigskip
\def\refname{\normalsize{\sc Références}}
\addcontentsline{toc}{section}{Bibliographie}
{\footnotesize
|
\section{Introduction}
The Kadomtsev-Petviashvili (KP) hierarchy is an archetypal infinite hierarchy
of compatible nonlinear differential equations with
infinitely many independent (time) variables
${\bf t}=\{t_1, t_2, t_3, \ldots \}$.
In the Lax-Sato formalism, the main object
is the Lax operator which is a pseudo-differential operator of the form
\begin{equation}\label{int0}
{\cal L}=\partial_x +u_1\partial_x^{-1}+u_2 \partial_x^{-2}+\ldots
\end{equation}
The coefficient functions $u_i$ are dependent variables.
Among all solutions to equations of the hierarchy,
of special interest are solutions which have a finite number of poles
in the variable $x$ in a fundamental domain of
the complex plane. Most general solutions of this type
are those for which the coefficient functions $u_i$
are elliptic (double-periodic in the complex plane)
functions of $x$ with poles depending on the times ${\bf t}$.
The study of singular solutions to nonlinear integrable
equations and dynamics of their poles was initiated in the pioneering papers
\cite{AMM77,Krichever78,Krichever80,CC77}. Now it
is a well known subject in the theory of integrable systems.
The remarkable result is
that the poles of solutions to the KP equation
as functions of the time $t_2$ move
as particles of the integrable Calogero-Moser many-body system
\cite{Calogero71,Calogero75,Moser75,OP81} which is known to be integrable, i.e.
having a large number of integrals of motion in involution. Elliptic,
trigonometric and rational solutions
correspond respectively to the Calogero-Moser systems with
elliptic, trigonometric and rational potentials.
In the work \cite{Shiota94}, Shiota has shown that in the case of rational solutions
the correspondence between the KP equation and the rational Calogero-Moser
system can be extended to the whole KP hierarchy.
Namely, the evolution of poles with respect to the higher time $t_m$ was considered
and it was shown that it is described by
the higher Hamiltonian flow
of the rational Calogero-Moser system with the Hamiltonian
$H_m=\mbox{tr}\, L^m$, where $L$ is the Lax matrix depending on the coordinates and momenta
in a special way.
Recently this remarkable correspondence was generalized to trigonometric
and elliptic solutions to the KP hierarchy (see respectively
\cite{Haine07,Z19a} and \cite{PZ21}). However, in the elliptic case this correspondence
is no longer formulated in terms of traces of the Lax matrix (which in this case depends
on a spectral parameter). Instead, the Hamiltonian $H_m$ which governs the dynamics of poles
with respect to $t_m$ is shown to be obtained by expansion of the Calogero-Moser
spectral curve near a distinguished marked point. It was also shown in \cite{PZ21}
that for trigonometric
and rational degenerations of elliptic solutions this construction gives the results
which agree with the previously
obtained ones for trigonometric and rational solutions.
There exists a matrix generalization of the KP hierarchy (matrix KP hierarchy).
In matrix KP hierarchy, the coefficient functions $u_i$ in the Lax operator (\ref{int0})
are $n\! \times \! n$ matrices.
Like the KP hierarchy, it
is an infinite set of compatible nonlinear differential equations with
infinitely many independent variables
${\bf t}$ and matrix dependent variables.
It is a subhierarchy of a more general multi-component ($n$-component) KP hierarchy
\cite{DJKM81,KL93,TT07,Teo11}, which has an extended set of independent variables
$
\{{\bf t}_1, {\bf t}_2, \ldots , {\bf t}_n\}$,
${\bf t}_{\alpha}=\{t_{\alpha , 1}, t_{\alpha , 2}, t_{\alpha , 3}, \ldots \, \}$,
$\alpha = 1, \ldots , n.
$
The matrix KP hierarchy is obtained by the restriction $t_{\alpha , m}=t_m$ for all
$\alpha , m$.
The elliptic, trigonometric and rational solutions to the matrix KP
equation were investigated in \cite{KBBT95}. In the matrix case, the data
of singular solutions include not only the positions of poles $x_i$ but also
some ``internal degrees of freedom'' which are matrix residues at the poles
(they were fixed in the scalar case).
In the work \cite{KBBT95} it was shown that the dynamics of the data of such solutions
with respect to the time $t_2$ is isomorphic to the dynamics of a
spin generalization of the Calogero-Moser
system which is also known as the Gibbons-Hermsen model \cite{GH84}. It is a system of $N$
particles with coordinates $x_i$ and with internal degrees of freedom
represented by $n$-dimensional column vectors ${\bf a}_i, {\bf b}_i$ with components
$a_i^{\alpha}$, $b_i^{\alpha}$, $\alpha =1, \ldots , n$.
The rank 1 matrices
$\rho_i=
{\bf a}_i{\bf b}_i^T$, where ${\bf b}_i^T$ is the row vector obtained from the vector
${\bf b}_i$ by transposition, represent matrix residues at the poles $x_i$.
The particles pairwise interact with each other.
The Hamiltonian of the elliptic model is
\begin{equation}\label{int2}
H=\sum_{i=1}^{{\cal N}}p_i^2-\sum_{i\neq k}
({\bf b}_i^T{\bf a}_k)({\bf b}_k^T{\bf a}_i)\wp (x_i-x_k),
\end{equation}
where $\wp (x)$ is the Weierstrass $\wp$-function which is the elliptic function with the
only second order pole at $x=0$ in the fundamental domain.
The non-vanishing Poisson brackets are
$
\{x_i, p_k\}=\delta_{ik}, \, \{a_i^{\alpha}, b_k^{\beta}\}=
\delta_{\alpha \beta}\delta_{ik}
$.
The model is known to be integrable and possessing the Lax representation with the Lax matrix
$L(\lambda )$ depending on a spectral parameter $\lambda$ lying on an elliptic curve.
The extension of the isomorphism between rational and trigonometric solutions of the matrix
KP equation and the Gibbons-Hermsen system to the whole hierarchy was recently made in
\cite{PZ18} for rational solutions and in \cite{PZspintrig} for trigonometric ones.
In this paper we generalize these results to elliptic solutions of the matrix
KP hierarchy.
Our main result is that
the dynamics of poles $x_i$ and vectors ${\bf a}_i$, ${\bf b}_i$
which parametrize matrix residues at the poles
with respect to all higher times $t_m$ of the matrix KP hierarchy
is Hamiltonian, with the corresponding Hamiltonians being higher Hamiltonians
of the spin elliptic Calogero-Moser model. We find them in terms of expansion of the
spectral curve
\begin{equation}\label{int3}
\det_{N\times N} \Bigl ((z+\zeta (\lambda ))I-L(\lambda )\Bigr )=0
\end{equation}
($\zeta (\lambda )$ is the Weierstrass $\zeta$-function)
near some distinguished marked points at infinity. The spectral curve is a covering
of the elliptic curve, where the spectral parameter $\lambda$ lives. We show that above
the point $\lambda =0$, there are $n$ points at infinity $P_{\alpha}=(\infty_{\alpha}, 0)$,
where $z=\infty$, so that there are $n$ distinguished sheets of the covering (neighborhoods
of the points $P_{\alpha}$). In a neighborhood of the point $\lambda =0$ $n$ different
branches of the function $z(\lambda )$
such that $z(\lambda )\to \infty$ as $\lambda \to 0$ are defined
by the equation of the spectral curve. Let us denote them by $z_{\alpha}(\lambda )$ and let
$\lambda_{\alpha}(z)$ be inverse functions. Our main result is that the sum
over all branches
$\displaystyle{\sum_{\alpha =1}^n \lambda_{\alpha}(z)}$ is the generating function
for the Hamiltonians $H_m$:
\begin{equation}\label{int4}
\sum_{\alpha =1}^n \lambda_{\alpha}(z)=-Nz^{-1}-\sum_{m\geq 1}z^{-m-1}H_m
\end{equation}
or $H_m=-\displaystyle{\sum_{\alpha =1}^n\mathop{\hbox{res}}\limits_{\infty}\Bigl (z^m \lambda_{\alpha}(z)\Bigr )}$.
We also show that the degeneration of this construction to the rational and
trigonometric cases allows one to reproduce the results of the papers \cite{PZ18}
and \cite{PZspintrig}.
The organization of the paper is as follows. In section 2 we remind the reader
the main facts about the multi-component and matrix KP hierarchies. We recall both
Lax-Sato approach based on Lax equations and the bilinear (Hirota) approach based on
the bilinear relation for the tau-function. In section 3 we introduce elliptic solutions
and discuss the corresponding double-Bloch solutions for the wave function. Section 4
contains derivation of the dynamics of poles and residues with respect to the time $t_2$.
Following \cite{KBBT95}, we derive the equations of motion together with their Lax
representation. In section 5 we discuss properties of the spectral curve and define the
distinguished branches of the function $z(\lambda )$ around the point $\lambda =0$.
Sections 6 and 7 are devoted to derivation of the Hamiltonian dynamics of
respectively poles and residues in the higher times $t_m$. In section 8 we find explicitly
the first two Hamiltonians using the expansion (\ref{int4}) and identify them with the
Hamiltonians of the spin generalization of the Calogero-Moser system. Finally, in section 9
we consider the rational and trigonometric degenerations of our construction and show that
the results of the previous works are reproduced by the new approach.
\section{The matrix KP hierarchy}
Here we briefly review the main facts about the multi-component and matrix
KP hierarchies following \cite{TT07,Teo11}.
We start from the more general multi-component KP hierarchy.
The independent variables are $n$ infinite sets of continuous ``times''
$$
{\bf t}=\{{\bf t}_1, {\bf t}_2, \ldots , {\bf t}_n\}, \qquad
{\bf t}_{\alpha}=\{t_{\alpha , 1}, t_{\alpha , 2}, t_{\alpha , 3}, \ldots \, \},
\qquad \alpha = 1, \ldots , n.
$$
It is convenient to introduce also the variable $x$ such that
\begin{equation}\label{multi2}
\partial_x =\sum_{\alpha =1}^n \partial_{t_{\alpha , 1}}.
\end{equation}
The hierarchy is an infinite set of evolution equations in the times ${\bf t}$ for matrix
functions of the variable $x$.
In the Lax-Sato formalism, the main object
is the Lax operator which is a pseudo-differential operator of the form
\begin{equation}\label{multi3}
{\cal L}=\partial_x +u_1\partial_x^{-1}+u_2 \partial_x^{-2}+\ldots
\end{equation}
where the coefficients $u_i=u_i(x, {\bf t})$ are $n\! \times \! n$ matrices.
The coefficient functions $u_k$ depend on $x$ and also on all the times:
$$
u_k(x, {\bf t})=u_k(x+t_{1,1}, x+t_{2,1}, \ldots , x+t_{n,1};
t_{1,2}, \ldots , t_{n,2}; \ldots ).
$$
Besides, there are
other matrix pseudo-differential operators ${\cal R}_1, \ldots , {\cal R}_n$
of the form
\begin{equation}\label{multi5}
{\cal R}_{\alpha}=E_{\alpha}+ u_{\alpha , 1}\partial_x^{-1}+u_{\alpha , 2}\partial_x^{-2}+\ldots ,
\end{equation}
where $E_{\alpha}$ is the $n \! \times \! n$ matrix with the $(\alpha , \alpha )$ element
equal to 1 and all other components equal to $0$, and $u_{\alpha , i}$ are also
$n \! \times \! n$ matrices. The operators ${\cal L}$, ${\cal R}_1, \ldots , {\cal R}_n$
satisfy the conditions
\begin{equation}\label{multi6}
{\cal L}{\cal R}_{\alpha}={\cal R}_{\alpha}{\cal L}, \quad
{\cal R}_{\alpha}{\cal R}_{\beta}=\delta_{\alpha \beta}{\cal R}_{\alpha}, \quad
\sum_{\alpha =1}^n {\cal R}_{\alpha}=I,
\end{equation}
where $I$ is the unity matrix.
The Lax equations of the hierarchy which define evolution in the times read
\begin{equation}\label{multi8}
\partial_{t_{\alpha , k}}{\cal L}=[B_{\alpha , k}, \, {\cal L}], \quad
\partial_{t_{\alpha , k}}{\cal R}_{\beta}=[B_{\alpha , k}, \, {\cal R}_{\beta}],
\quad B_{\alpha , k} = \Bigl ({\cal L}^k{\cal R}_{\alpha}\Bigr )_+,
\quad k=1,2,3, \ldots ,
\end{equation}
where $(\ldots )_+$ means the differential part of a pseudo-differential operator,
i.e. the sum of all terms with $\partial_x^k$, where $k\geq 0$.
Let us introduce the matrix pseudo-differential ``wave operator'' ${\cal W}$ with matrix elements
\begin{equation}\label{m113}
{\cal W}_{\alpha \beta} = \delta_{\alpha \beta}+\sum_{k\geq 1}
\xi^{(k)}_{\alpha \beta}(x,{\bf t})\partial_x^{-k},
\end{equation}
where $\xi^{(k)}_{\alpha \beta}(x,{\bf t})$ are the
some matrix functions. The operators ${\cal L}$ and ${\cal R}_{\alpha}$ are obtained
from the ``bare'' operators $I\partial_x $ and $E_{\alpha}$ by `dressing''
by means of the wave operator:
\begin{equation}\label{multi12}
{\cal L}={\cal W}\partial_x {\cal W}^{-1}, \qquad
{\cal R}_{\alpha}={\cal W}E_{\alpha} {\cal W}^{-1}.
\end{equation}
Clearly, there is an ambiguity in the definition of the dressing operator: it can be multiplied
from the right by any pseudo-differential operator with constant coefficients
commuting with $E_{\alpha}$ for any $\alpha$.
A very important role in the theory is played by the wave function $\Psi$ and its adjoint
$\Psi^{\dag}$ (hereafter, ${}^{\dag}$ does not mean Hermitean conjugation).
The wave function
is defined as a result of action of the wave operator
to the exponential function:
\begin{equation}\label{m113a}
\Psi (x,{\bf t}; z)={\cal W}
\exp \Bigl (xzI+\sum_{\alpha =1}^n E_{\alpha}\xi ({\bf t}_{\alpha}, z)\Bigr ),
\end{equation}
where we use the standard notation
$$
\xi ({\bf t}_{\alpha}, z)=\sum_{k\geq 1}t_{\alpha , k}z^k.
$$
By definition, the operators $\partial_x^{-k}$ with negative powers act
to the exponential function as $\partial_x^{-k}e^{xz}=z^{-k}e^{xz}$.
The wave function depends on the spectral parameter $z$ which does not enter the auxiliary
linear problems explicitly.
The adjoint wave function is introduced by the formula
\begin{equation}\label{m113b}
\Psi ^{\dag} (x,{\bf t}; z)=\exp \Bigl (-xzI-\sum_{\alpha =1}^n
E_{\alpha}\xi ({\bf t}_{\alpha}, z)\Bigr )
{\cal W}^{-1}.
\end{equation}
Here we use the convention that
the operators $\partial_x$
which enter ${\cal W}^{-1}$ act to the left rather than to the right,
the left action being defined as
$f\stackrel{\leftarrow}{\partial_x} \equiv -\partial_x f$. Clearly, the expansion of the wave function
as $z\to \infty$ is as follows:
\begin{equation}\label{multi13}
\Psi_{\alpha \beta} (x,{\bf t}; z)=e^{xz+\xi ({\bf t}_{\beta}, z)}
\Bigl (\delta_{\alpha \beta}+\xi^{(1)}_{\alpha \beta}z^{-1}+
\xi^{(2)}_{\alpha \beta}z^{-2}+\ldots \Bigr ).
\end{equation}
As is proved in \cite{Teo11}, the wave function
satisfies the linear
equations
\begin{equation}\label{m13c}
\partial_{t_{\alpha , m}}\Psi (x,{\bf t}; z)=B_{\alpha ,m} \Psi (x,{\bf t}; z),
\end{equation}
where $B_{\alpha ,m}$ is the differential operator (\ref{multi8}), i.e.
$
B_{\alpha ,m}= \Bigl ({\cal W} E_{\alpha}\partial_x^m {\cal W}^{-1}\Bigr )_+
$
and the adjoint wave function satisfies the transposed equations
\begin{equation}\label{m13e}
-\partial_{t_{\alpha , m}}\Psi^{\dag} (x,{\bf t}; z)=\Psi^{\dag} (x,{\bf t}; z) B_{\alpha ,m} .
\end{equation}
Again, the operator
$B_{\alpha , m}$ here acts to the left.
In particular, it follows from
(\ref{m13c}), (\ref{m13e}) at $m=1$ that
\begin{equation}\label{m13d}
\sum_{\alpha =1}^{n}\partial_{t_{\alpha , 1}}\Psi (x,{\bf t}; z)=\partial_x \Psi (x,{\bf t}; z),
\qquad
\sum_{\alpha =1}^{n}\partial_{t_{\alpha , 1}}\Psi^{\dag} (x,{\bf t}; z)=\partial_x \Psi^{\dag} (x,{\bf t}; z),
\end{equation}
so the vector field $\partial_x$ can be naturally identified with the vector field
$\displaystyle{\sum_{\alpha }\partial_{t_{\alpha , 1}}}$.
Another approach to the multi-component KP hierarchy is provided by the
bilinear formalism.
In the bilinear formalism,
the dependent variables are the tau-function $\tau (x,{\bf t})$ and tau-functions
$\tau _{\alpha \beta}(x,{\bf t})$ such that $\tau _{\alpha \alpha}(x,{\bf t})=
\tau (x, {\bf t})$ for any $\alpha$.
The $n$-component KP hierarchy is the infinite set of bilinear equations
for the tau-functions which are encoded in the basic bilinear
relation
\begin{equation}\label{m5}
\sum_{\nu =1}^n \epsilon_{\alpha \nu}\epsilon_{\beta \nu}
\oint_{C_{\infty}}\! dz \,
z^{\delta_{\alpha \nu}+\delta_{\beta \nu}-2}
e^{\xi ({\bf t}_{\nu}-{\bf t}_{\nu}', \, z)}
\tau _{\alpha \nu} \left (x,{\bf t}-[z^{-1}]_{\nu}\right )
\tau _{\nu \beta}\left (x,{\bf t}'+[z^{-1}]_{\nu}\right )=0
\end{equation}
valid for any ${\bf t}$, ${\bf t}'$. Here $\epsilon_{\alpha \beta}$ is a sign factor:
$\epsilon_{\alpha \beta}=1$ if $\alpha \leq \beta$, $\epsilon_{\alpha \beta}=-1$
if $\alpha >\beta$.
In (\ref{m5}) we use the following standard
notation:
$$
\left ({\bf t}\pm [z^{-1}]_{\gamma}\right )_{\alpha k}=t_{\alpha , k}\pm
\delta_{\alpha \gamma} \frac{z^{-k}}{k}.
$$
The integration contour $C_{\infty}$ is a big circle around $\infty$.
The tau-functions are universal dependent variables of the hierarchy.
All other objects including
the coefficient functions $u_i$ of the Lax operator and the wave functions can be
expressed in terms of them. In particular, for the wave function and its adjoint we have:
\begin{equation}\label{m2}
\begin{array}{l}
\displaystyle{\Psi_{\alpha \beta}(x,{\bf t};z)=
\epsilon_{\alpha \beta}\,
\frac{\tau_{\alpha \beta} \left (x,
{\bf t}-[z^{-1}]_{\beta}\right )}{\tau (x,{\bf t})}\,
z^{\delta_{\alpha \beta}-1}e^{\xi ({\bf t}_{\beta}, z)},
}
\\ \\
\displaystyle{\Psi_{\alpha \beta}^{\dag}(x,{\bf t};z)=
\epsilon_{\beta \alpha}\,
\frac{\tau _{\alpha \beta}\left (x,
{\bf t}+[z^{-1}]_{\alpha}\right )}{\tau (x,{\bf t})}\,
z^{\delta_{\alpha \beta}-1}e^{-\xi ({\bf t}_{\alpha}, z)}
}
\end{array}
\end{equation}
Note that the bilinear relation (\ref{m5}) can be written in the form
\begin{equation}\label{m3}
\oint_{C_{\infty}}\! dz \, \Psi (x,{\bf t};z)\Psi^{\dag} (x,{\bf t}';z)=0.
\end{equation}
The coefficient $\xi_{\alpha \beta}^{(1)}(x,{\bf t})$
plays an important role in what follows. Equations (\ref{m2}) imply that this coefficient
is expressed through the tau-functions as
\begin{equation}\label{m2a}
\xi_{\alpha \beta}^{(1)}(x,{\bf t})=\left \{
\begin{array}{l}
\displaystyle{\epsilon_{\alpha \beta}\,
\frac{\tau_{\alpha \beta}(x, {\bf t})}{\tau (x, {\bf t})}, \quad \alpha \neq \beta ,}
\\ \\
\displaystyle{
-\frac{\partial_{t_{\beta}}\tau (x, {\bf t})}{\tau (x, {\bf t})}, \quad \,\, \alpha =\beta .}
\end{array}
\right.
\end{equation}
Let us point out a useful corollary of the bilinear
relation (\ref{m5}).
Differentiating it with respect to $t_{\kappa , m}$ and putting ${\bf t}'={\bf t}$ after this, we
obtain:
\begin{equation}\label{m11b}
\frac{1}{2\pi i}
\oint_{C_{\infty}}dz \, z^m \Psi_{\alpha \kappa}(x,{\bf t}; z)
\Psi^{\dag}_{\kappa \beta}(x,{\bf t}; z)=-\partial_{t_{\kappa , m}}\xi_{\alpha \beta}^{(1)}(x,{\bf t})
\end{equation}
or, equivalently,
\begin{equation}\label{m11c}
\mathop{\hbox{res}}\limits_{\infty}\Bigl (z^m \Psi_{\alpha \kappa}\Psi^{\dag}_{\kappa \beta}
\Bigr ) =-\partial_{t_{\kappa , m}}\xi_{\alpha \beta}^{(1)}.
\end{equation}
The integrand in (\ref{m11b}) should be regarded as a Laurent series in $z$ and the
residue at infinity is defined according to the convention
$\mathop{\hbox{res}}\limits_{\infty}\, (z^{-k})=\delta_{k1}$.
The matrix KP hierarchy is a subhierarchy of the
multi-component KP one. It is obtained by
the following restriction of the independent
variables:
$t_{\alpha , m}=t_m$ for each $\alpha$ and $m$.
The corresponding vector fields are related as
$\displaystyle{\partial_{t_m}=\sum_{\alpha =1}^n \partial_{t_{\alpha , m}}}$.
As is clear from (\ref{multi13}),
the wave function for the matrix KP hierarchy has the expansion
\begin{equation}\label{m7}
\Psi_{\alpha \beta}(x,{\bf t};z)=\left (\delta_{\alpha \beta}+
\xi_{\alpha \beta}^{(1)}({\bf t})z^{-1}+O(z^{-2})\right )
e^{xz+\xi ({\bf t}, z)},
\end{equation}
where $\displaystyle{\xi ({\bf t}, z)=\sum_{k\geq 1}t_kz^k}$.
Equations (\ref{m13c}) imply that the
wave function of the matrix KP hierarchy and its adjoint
satisfy the linear
equations
\begin{equation}\label{m13a}
\partial_{t_m}\Psi ({\bf t}; z)=B_m \Psi ({\bf t}; z),
\qquad
-\partial_{t_m}\Psi^{\dag} ({\bf t}; z)=\Psi^{\dag} ({\bf t}; z) B_m , \qquad m\geq 1,
\end{equation}
where $B_m$ is the differential operator
$
B_m= \Bigl ({\cal W} \partial_x^m {\cal W}^{-1}\Bigr )_+.
$
At $m=1$ we have $\partial_{t_1}\Psi =\partial_x \Psi$,
so we can identify
$\displaystyle{
\partial_x = \partial_{t_1}=\sum_{\alpha =1}^N \partial_{t_{\alpha , 1}}}
$. This means that
the evolution in the time $t_1$ is simply a shift of
the variable $x$:
$
\xi^{(k)}(x, t_1, t_2, \ldots )=\xi^{(k)}(x+t_1, t_2, \ldots ).
$
At $m=2$ equations (\ref{m13a}) turn into the linear problems
\begin{equation}\label{m14}
\partial_{t_2}\Psi = \partial_x^2\Psi +2V(x,{\bf t})\Psi ,
\end{equation}
\begin{equation}\label{m14a}
-\partial_{t_2}\Psi^{\dag} = \partial_x^2\Psi^{\dag} +2\Psi^{\dag}V(x,{\bf t})
\end{equation}
which have the form of the matrix non-stationary Schr\"odinger equations with
\begin{equation}\label{m15}
V(x,{\bf t})=-\partial_x \xi^{(1)}(x,{\bf t}).
\end{equation}
Summing (\ref{m11b}) over $\kappa$, we obtain an analog of (\ref{m11b}) for the
matrix KP hierarchy:
\begin{equation}\label{m11a}
\frac{1}{2\pi i}
\sum_{\nu =1}^n \oint_{C_{\infty}}dz \, z^m \Psi_{\alpha \nu}(x,{\bf t}; z)
\Psi^{\dag}_{\nu \beta}(x,{\bf t}; z)=-\partial_{t_m}\xi_{\alpha \beta}^{(1)}(x,{\bf t}).
\end{equation}
Below we will use equations (\ref{m11b}) and (\ref{m11a})
for derivation of dynamics of poles and residues of elliptic solutions
in higher times.
\section{Elliptic solutions of the matrix KP hierarchy and double-Bloch functions}
Our aim is to study solutions to the matrix KP hierarchy which are elliptic functions
of the variable $x$ (and, therefore, $t_1$).
For the elliptic solutions, we take the tau-function in the form
\begin{equation}\label{t1}
\tau (x,{\bf t})= C\prod_{i=1}^{{\cal N}} \sigma (x-x_i({\bf t})),
\end{equation}
where
$$
\sigma (x)=\sigma (x |\, \omega , \omega ')=
x\prod_{s\neq 0}\Bigl (1-\frac{x}{s}\Bigr )\, e^{\frac{x}{s}+\frac{x^2}{2s^2}},
\quad s=2\omega m+2\omega ' m' \quad \mbox{with integer $m, m'$}
$$
is the Weierstrass
$\sigma$-function with quasi-periods $2\omega$, $2\omega '$ such that
${\rm Im} (\omega '/ \omega )>0$.
It is connected with the Weierstrass
$\zeta$- and $\wp$-functions by the formulas $\zeta (x)=\sigma '(x)/\sigma (x)$,
$\wp (x)=-\zeta '(x)=-\partial_x^2\log \sigma (x)$.
The monodromy properties of the function $\sigma (x)$
are
\begin{equation}\label{e1a}
\sigma (x+2\omega )=-e^{2\zeta (\omega ) (x+\omega )}\sigma (x), \quad
\sigma (x+2\omega ' )=-e^{2\zeta (\omega ' ) (x+\omega ' )}\sigma (x),
\end{equation}
where the constants $\zeta (\omega )$, $\zeta (\omega ' )$ are related by
$\zeta (\omega )\omega ' -\zeta (\omega ') \omega =\pi i /2$.
The $N$ zeros $x_i$ of (\ref{t1}) are assumed to be all distinct.
We also assume that the tau-functions $\tau_{\alpha \beta}$ at $\alpha \neq \beta$
have the form
\begin{equation}\label{t1a}
\tau_{\alpha \beta} (x,{\bf t})= C_{\alpha \beta}\prod_{i=1}^{{\cal N}}
\sigma (x-x_i^{(\alpha \beta )}({\bf t}))
\end{equation}
with
\begin{equation}\label{t1b}
\sum_i x_i({\bf t}) = \sum_i x_i^{(\alpha \beta )}({\bf t})
\quad \mbox{for all $\alpha , \beta$}.
\end{equation}
The consistency of this assumption is justified below.
Equation (\ref{m2a}) together with the condition (\ref{t1b}) implies that
$V(x, {\bf t})=-\partial_x \xi^{(1)}$ in the linear problem (\ref{m14}) is an
elliptic function of $x$. Therefore, one can find solutions to (\ref{m14})
which are {\it double-Bloch functions}.
The double-Bloch function satisfies the monodromy properties
$\Psi_{\alpha \beta} (x+2\omega )=B_{\beta}\Psi_{\alpha \beta} (x)$,
$\Psi_{\alpha \beta} (x+2\omega ' )=B_{\beta}'\Psi_{\alpha \beta} (x)$ with some
Bloch multipliers $B_{\beta}$, $B_{\beta}'$.
The Bloch multipliers of the wave function
(\ref{m2}) are:
\begin{equation}\label{e3}
\begin{array}{l}
\displaystyle{
B_{\beta}=\exp \Bigl (2\omega z-2\zeta (\omega )\sum_i (e^{-D_{\beta}(z)}\! -\! 1)x_i )\Bigr ),}
\\ \\
\displaystyle{
B_{\beta}'=\exp \Bigl (2\omega ' z -2\zeta (\omega ' )\sum_i (e^{-D_{\beta}(z)}\! -\! 1)x_i)\Bigr ),}
\end{array}
\end{equation}
where the differential operator $D_{\beta}(z)$ is
\begin{equation}\label{D(z)}
D_{\beta}(z)=\sum_{k\geq 1}\frac{z^{-k}}{k}\, \partial_{t_{\beta ,k}}.
\end{equation}
Since the right hand side of (\ref{m11a}) is an elliptic function of $x$,
the Bloch multipliers of the adjoint wave function should be $1/B_{\alpha}$,
$1/B_{\alpha}'$: $\Psi^{\dag}_{\alpha \beta} (x+2\omega )=(B_{\alpha})^{-1}
\Psi^{\dag}_{\alpha \beta} (x)$, $\Psi^{\dag}_{\alpha \beta} (x+2\omega ' )=(B'_{\alpha})^{-1}
\Psi^{\dag}_{\alpha \beta} (x)$.
Any non-trivial double-Bloch function (i.e. the one which is not just an exponential
function) must have at least one pole in $x$ in the fundamental domain.
Let us introduce the
elementary double-Bloch function $\Phi (x, \lambda )$ having just one pole
in the fundamental domain and defined as
\begin{equation}\label{Phi}
\Phi (x, \lambda )=\frac{\sigma (x+\lambda )}{\sigma (\lambda )\sigma (x)}\,
e^{-\zeta (\lambda )x}
\end{equation}
(here $\zeta (\lambda )$ is the Weierstrass $\zeta$-function).
The monodromy properties of the function $\Phi$ follow from (\ref{e1a}):
$$
\Phi (x+2\omega , \lambda )=e^{2(\zeta (\omega )\lambda - \zeta (\lambda )\omega )}
\Phi (x, \lambda ),
$$
$$
\Phi (x+2\omega ' , \lambda )=e^{2(\zeta (\omega ' )\lambda - \zeta (\lambda )\omega ' )}
\Phi (x, \lambda ).
$$
We see that it is indeed a double-Bloch function.
It has a single simple pole in the fundamental domain
at $x=0$ with residue 1:
$$
\Phi (x, \lambda )=\frac{1}{x} -\frac{1}{2}\, \wp (\lambda )x +\ldots , \qquad
x\to 0,
$$
It is easy to see that $\Phi (x, \lambda )$ is an elliptic function of $\lambda$.
The expansion as $\lambda \to 0$ is
\begin{equation}\label{Phi1}
\begin{array}{c}
\Phi (x, \lambda )=\left (\lambda^{-1}+ \zeta (x)+\frac{1}{2} \, (\zeta^2(x)-\wp (x))\lambda +
O(\lambda^2)\right )e^{-x/\lambda}.
\end{array}
\end{equation}
We will also need the $x$-derivatives
$\Phi '(x, \lambda )=\partial_x \Phi (x, \lambda )$,
$\Phi ''(x, \lambda )=\partial^2_x \Phi (x, \lambda )$.
It is clear from (\ref{m2}) that the wave functions $\Psi$, $\Psi^\dag$
(and thus the coefficient $\xi^{(1)}$),
as functions of $x$,
have simple poles at $x=x_i$.
It is shown in \cite{PZ18} that the residues
at these poles are matrices of rank $1$. We parametrize the residues
of $\xi^{(1)}$ through the
column vectors
${\bf a}_i =(a_i^1, a_i^2, \ldots , a_i^n)^T$,
${\bf b}_i =(b_i^1, b_i^2, \ldots , b_i^n)^T$ ($T$ means transposition):
\begin{equation}\label{t3a}
\xi_{\alpha \beta}^{(1)}=S_{\alpha \beta}-\sum_i a_i^{\alpha}b_i^{\beta}\zeta (x-x_i),
\end{equation}
where $S_{\alpha \beta}$ does not depend on $x$.
Therefore,
\begin{equation}\label{t3b}
V(x, {\bf t})=-\sum_i a_i^{\alpha}b_i^{\beta}\wp (x-x_i).
\end{equation}
The components of the vectors ${\bf a}_i$, ${\bf b}_i$ are going to be
spin variables of the elliptic Gibbons-Hermsen model.
One can expand the wave functions
using the elementary double-Bloch functions as follows:
\begin{equation}\label{t3}
\Psi_{\alpha \beta}=e^{k_{\beta} x+\xi ({\bf t}, z)}\sum_i a_i^{\alpha}c_i^{\beta}
\Phi (x-x_i, \lambda_{\beta}),
\end{equation}
\begin{equation}\label{t4}
\Psi^{\dag}_{\alpha \beta}=e^{-k_{\alpha} x-\xi ({\bf t}, z)}\sum_i
c_i^{*\alpha}b_i^{\beta}\Phi (x-x_i, -\lambda_{\alpha}),
\end{equation}
where $c_i^{\alpha}$, $c_i^{*\alpha}$ are components of some $x$-independent vectors
${\bf c}_i=(c_i^1, \ldots , c_i^n)^T$,
${\bf c}^{*}_i=(c^{*1}_i, \ldots , c^{*n}_i)^T$.
This is similar to expansion of a rational function
in a linear combination of simple fractions.
One can see that (\ref{t3}) is
a double-Bloch function with Bloch multipliers
\begin{equation}\label{e6}
B_{\beta}=e^{2\omega (k_{\beta}-\zeta (\lambda _{\beta})) +
2\zeta (\omega )\lambda_{\beta} }, \qquad
B_{\beta} '=e^{2\omega ' (k_{\beta}-\zeta (\lambda_{\beta} )) +
2\zeta (\omega ' )\lambda_{\beta} }
\end{equation}
and (\ref{t4}) has Bloch multipliers $(B_{\alpha})^{-1}$ and $(B_{\alpha}')^{-1}$.
These Bloch multipliers
should coincide with (\ref{e3}).
Therefore, comparing (\ref{e3}) with (\ref{e6}), we get
$$
2\omega \Bigl (k_{\beta}-z-\zeta (\lambda_{\beta} )\Bigr )+
2\zeta (\omega )\Bigl (\lambda_{\beta} +(e^{-D_{\beta}(z)}-1)\sum_i x_i\Bigr )=
2\pi i n,
$$
$$
2\omega ' \Bigl (k_{\beta}-z-\zeta (\lambda_{\beta} )\Bigr )+
2\zeta (\omega ' )\Bigl (\lambda_{\beta} +(e^{-D_{\beta}(z)}-1)\sum_i x_i\Bigr )=
2\pi i n'
$$
with some integer $n, n'$. These equations can be regarded as a linear system.
The solution is
$$
k_{\beta}-z-\zeta (\lambda_{\beta})=2n'\zeta (\omega)-2n \zeta (\omega '),
$$
$$
\lambda_{\beta} +(e^{-D_{\beta}(z)}-1)\sum_i x_i=2n\omega ' -2n'\omega .
$$
Shifting $\lambda_{\beta}$ by a suitable vector of the lattice spanned by $2\omega$, $2\omega '$,
we can
represent the connection between the spectral parameters $k_{\beta},z, \lambda_{\beta}$
in the form
\begin{equation}\label{e7}
\left \{\begin{array}{l}
k_{\beta}=z+\zeta (\lambda_{\beta} ),
\\ \\
\displaystyle{\lambda_{\beta} = (1-e^{-D_{\beta}(z)}) \sum_i x_i}.
\end{array} \right.
\end{equation}
These two equations for three spectral parameters $k_{\beta}, z, \lambda_{\beta}$
determine the spectral curve, with the index $\beta$ numbering different sheets of it.
Another description of the same spectral curve is obtained below as
the spectral curve of the spin generalization of the Calogero-Moser system
(it is given by the characteristic polynomial of the Lax matrix
$L(\lambda )$ for the spin Calogero-Moser system).
As we shall see below, it has the form $R(k,\lambda)=0$, where
$R(k,\lambda)$ is a polynomial in $k$ whose coefficients are elliptic functions of
$\lambda$.
These coefficients are integrals of motion in involution. The spectral curve
in the form $R(k,\lambda)=0$ appears if one excludes $z$ from the equations (\ref{e7}).
Equivalently, one can represent the spectral curve as a relation connecting
$z$ and $\lambda_{\beta}$:
\begin{equation}\label{e7a}
R(z+\zeta (\lambda_{\beta} ), \lambda_{\beta} )=0.
\end{equation}
The function $z(\lambda )$ defined by this equation is multivalued, $z_{\beta}(\lambda )$
being different branches of this function. Then the function $\lambda_{\beta}(z)$ is the inverse
function to the $z_{\beta}(\lambda )$.
Using the same arguments as in \cite{PZ21}, one can see that
the second equation in (\ref{e7}) can be written as
\begin{equation}\label{e8}
\lambda_{\beta}(z) = D_{\beta}(z)\sum_i x_i=\sum_{j\geq 1}\frac{z^{-j}}{j}\, V^{(\beta )}_j,
\quad V^{(\beta )}_j=\partial_{t_{\beta , j}}\sum_i x_i,
\end{equation}
where $\lambda_{\beta}(z)$ should be understood as expansion of the
$\beta$-th branch of the function $\lambda (z)$ in negative powers of $z$ near $z=\infty$.
\section{Dynamics of poles and residues in $t_2$}
We first consider the dynamics of the poles and residues
with respect to the time $t_2$. Following Krichever's approach,
we consider the linear problems
(\ref{m14}), (\ref{m14a}),
and substitute the pole ansatz (\ref{t3}), (\ref{t4}) for the wave functions.
Consider first the equation for $\Psi$.
After the substitution, we see that the expression has poles at $x=x_i$
up to the third order.
Equating coefficients at the poles of different orders at $x=x_i$,
we get the conditions:
\begin{itemize}
\item
At $\frac{1}{(x-x_i)^3}$: \phantom{a} $b_i^{\nu}a_i^{\nu}=1$;
\item
At $\frac{1}{(x-x_i)^2}$: \phantom{a} $\displaystyle{
-\frac{1}{2}\, \dot x_i c_i^{\beta}-\sum_{j\neq i}
b_i^{\nu}a_j^{\nu}c_j^{\beta}\Phi (x_i\! -\! x_j, \lambda_{\beta})=
k_{\beta}c_i^{\beta}}$;
\item
At $\frac{1}{x-x_i}$: $\phantom{aaaa} \partial_{t_2}(
a_i^{\alpha}c_i^{\beta})=(k_{\beta}^2-z^2
+\wp (\lambda_{\beta} )) a_i^{\alpha}c_i^{\beta}$
$$
\phantom{aaaaaaaaaaaaa}
-2\sum_{j\neq i}a_i^{\alpha}b_i^{\nu}a_{j}^{\nu}c_j^{\beta}\Phi '(x_i-x_j, \lambda_\beta)
-2c_i^{\beta}\sum_{j\neq i}a_i^{\nu}b_j^{\nu}a_j^{\alpha}\wp (x_i-x_j),
$$
\end{itemize}
where dot means $t_2$-derivative. Here and below summation over repeated Greek indices
numbering components of vectors from 1 to $n$ is implied, unless otherwise stated.
Similar calculations for the linear problem for $\Psi^{\dag}$ lead to the conditions
\begin{itemize}
\item
At $\frac{1}{(x-x_i)^3}$: \phantom{a} $b_i^{\nu}a_i^{\nu}=1$ (the same as above);
\item
At $\frac{1}{(x-x_i)^2}$: \phantom{a} $\displaystyle{
-\frac{1}{2}\, \dot x_i c_i^{*\alpha}- \sum_{j\neq i}
c_j^{*\alpha}b_j^{\nu}a_i^{\nu}\Phi (x_j\! -\! x_i, \lambda_{\alpha})
=k_{\alpha}c_i^{*\alpha}}$;
\item
At $\frac{1}{x-x_i}$:
$\phantom{aaaa} \partial_{t_2}(
c_i^{*\alpha}b_i^{\beta})=-(k_{\beta}^2-z^2
+\wp (\lambda_{\alpha} )) c_i^{*\alpha}b_i^{\beta}$
$$
\phantom{aaaaaaaaaaaaa}
+2\sum_{j\neq i}c_j^{*\alpha}b_j^{\nu}a_{i}^{\nu}b_i^{\beta}\Phi '(x_j-x_i, \lambda_\alpha )
+2c_i^{*\alpha}\sum_{j\neq i}b_i^{\nu}a_j^{\nu}b_j^{\beta}\wp (x_i-x_j).
$$
\end{itemize}
Here we have used the obvious property $\Phi (x, -\lambda )=-\Phi (-x, \lambda )$.
The conditions coming from the
third order poles are constraints on the vectors ${\bf a}_i$, ${\bf b}_i$.
The other conditions can be written in the matrix form
\begin{equation}\label{t6}
\left \{
\begin{array}{l}
(k_{\beta}I-L(\lambda_{\beta})){\sf c}^{\beta}=0
\\ \\
\dot {\sf c}^{\beta}=M(\lambda_{\beta}){\sf c}^{\beta},
\end{array}
\right.
\end{equation}
\begin{equation}\label{t7}
\left \{
\begin{array}{l}
{\sf c}^{*\alpha}(k_{\alpha}I-L(\lambda_{\alpha}))=0
\\ \\
\dot {\sf c}^{*\alpha}={\sf c}^{*\alpha}M^{*}(\lambda_{\alpha})
\end{array}
\right.
\end{equation}
(no summation over $\alpha , \beta$),
where ${\sf c}^{\beta}=(c^{\beta}_1, \ldots c^{\beta}_{N})^T$,
${\sf c}^{*\alpha}=(c^{*\alpha}_1, \ldots c^{*\alpha}_N)$,
are $N$-dimensional vectors, $I$ is the unity matrix, and
$L(\lambda )$, $M(\lambda )$, $M^{*}(\lambda )$
are $N\! \times \! N$ matrices of the form
\begin{equation}\label{t8}
L_{ij}(\lambda )=-\frac{1}{2}\, \dot x_i \delta_{ij}-(1-\delta_{ij})
b_i^{\nu}a_j^{\nu}\Phi (x_i-x_j, \lambda ),
\end{equation}
\begin{equation}\label{t9}
M_{ij}(\lambda )=(k^2-z^2+\wp (\lambda ) -\Lambda_i)\delta_{ij}-2(1-\delta_{ij})
b_i^{\nu}a_j^{\nu}\Phi '(x_i-x_j, \lambda ),
\end{equation}
\begin{equation}\label{t10}
M^{*}_{ij}(\lambda )=-(k^2-z^2+\wp (\lambda ) -\Lambda^{*}_i)\delta_{ij}+2(1-\delta_{ij})
b_i^{\nu}a_j^{\nu}\Phi '(x_i-x_j, \lambda ).
\end{equation}
Here
\begin{equation}\label{t11}
\Lambda_i=\frac{\dot a_i^{\alpha}}{a_i^{\alpha}}+2
\sum_{j\neq i}\frac{a_j^{\alpha}b_j^{\nu}a_i^{\nu}}{a_i^{\alpha}}\, \wp (x_i-x_j),
\quad
-\Lambda_i^*=\frac{\dot b_i^{\alpha}}{b_i^{\alpha}}-2
\sum_{j\neq i}\frac{ b_i^{\nu}a_j^{\nu}b_j^{\alpha}}{b_i^{\alpha}}\, \wp (x_i-x_j)
\end{equation}
do not depend on the index $\alpha$ (there is summation over $\nu$
but no summation over $\alpha$). In fact one can see that $\Lambda_i=\Lambda_i^*$,
so that $M^{*}(\lambda )=-M(\lambda )$.
Indeed, multiplying equations (\ref{t11}) by $a_i^{\alpha}b_i^{\alpha}$
(no summation here!), summing over
$\alpha$ and summing the two equations, we get $\Lambda_i-\Lambda_i^*=\partial_{t_2}(
a_i^{\alpha}b_i^{\alpha})=0$ by virtue of the constraint
$a_i^{\alpha}b_i^{\alpha}=1$.
Differentiating the first equation in (\ref{t6}) by $t_2$, we get
the compatibility condition of equations (\ref{t6}):
\begin{equation}\label{t12}
(\dot L+[L,M]){\sf c}^{\beta}=0.
\end{equation}
One can see, taking into account equations (\ref{t11}), which we write here in the form
\begin{equation}\label{t13}
\dot a_i^{\alpha}=\Lambda_i a_i^{\alpha}-2
\sum_{j\neq i}a_j^{\alpha}b_j^{\nu}a_i^{\nu}\wp (x_i-x_j),
\end{equation}
\begin{equation}\label{t13a}
\dot b_i^{\alpha}=-\Lambda_i b_i^{\alpha}+2
\sum_{j\neq i}b_i^{\nu}a_j^{\nu}b_j^{\alpha}\wp (x_i-x_j)
\end{equation}
(in this form they are equations of motion for the spin degrees of freedom)
that the off-diagonal elements of the matrix $\dot L+[L,M]$ are equal to zero.
Vanishing of the diagonal elements yields equations of motion for the poles $x_i$:
\begin{equation}\label{t14}
\ddot x_i=4\sum_{j\neq i}
b_i^{\mu}a_k^{\mu}
b_k^{\nu}a_i^{\nu}\wp '(x_i-x_j).
\end{equation}
The gauge transformation $a_i^{\alpha}\to a_i^{\alpha}q_i$,
$b_i^{\alpha}\to b_i^{\alpha}q_i^{-1}$ with
$\displaystyle{q_i=\exp \Bigl (\int^{t_2}\Lambda_i dt\Bigr )}$ eliminates
the terms with $\Lambda_i$ in (\ref{t13}), (\ref{t13a}), so we can put
$\Lambda_i=0$. This gives the equations of motion
\begin{equation}\label{t15}
\dot a_i^{\alpha}=-2
\sum_{j\neq i}a_j^{\alpha}b_j^{\nu}a_i^{\nu}\wp (x_i-x_j),
\quad
\dot b_i^{\alpha}=2
\sum_{j\neq i}b_i^{\nu}a_j^{\nu}b_j^{\alpha}\wp (x_i-x_j).
\end{equation}
Together with (\ref{t14}) they are equations of motion of the elliptic
Gibbons-Hermsen model.
Their Lax representation is given by the matrix equation $\dot L=[M,L]$. It states that
the time evolution of the Lax matrix is an isospectral transformation. It follows
that the quantities $\mbox{tr}\, L^m(\lambda )$ are integrals of motion. In particular,
\begin{equation}\label{t16}
H_2=\sum_{i=1}^{N}p_i^2-\sum_{i\neq j}
b_i^{\mu}a_j^{\mu} b_j^{\nu}a_i^{\nu}\wp (x_i-x_j)=
\mbox{tr}\, L^2(\lambda ) +\mbox{const}
\end{equation}
is the Hamiltonian of the elliptic Gibbons-Hermsen model. Equations of motion (\ref{t14}),
(\ref{t15}) are equivalent to the Hamiltonian equations
\begin{equation}\label{t17}
\dot x_i=\frac{\partial H_2}{\partial p_i}, \quad \dot p_i=-\frac{\partial H_2}{\partial x_i},
\quad
\dot a_i^{\alpha}=\frac{\partial H_2}{\partial b_i^{\alpha}}, \quad
\dot b_i^{\alpha}=-\frac{\partial H_2}{\partial a_i^{\alpha}}.
\end{equation}
We see that $\dot x_i =2p_i$ and the Lax matrix is expressed through the momenta as follows:
\begin{equation}\label{t18}
L_{ij}(\lambda )=-p_i \delta_{ij}-(1-\delta_{ij})b_i^{\nu}a_j^{\nu}\Phi (x_i-x_j, \lambda ).
\end{equation}
As we shall see, the higher time flows are also Hamiltonian with the Hamiltonians
being linear combinations of spectral invariants of the Lax matrix, i.e. linear combinations
of traces of its powers $\mbox{tr}\, L^j(\lambda )$. It is not difficult to see that
\begin{equation}\label{t19}
G^{\alpha \beta}=\sum_i a_i^{\alpha}b_i^{\beta}
\end{equation}
are integrals of motion for all time flows: $\partial_{t_m}G^{\alpha \beta}=0$. Indeed, we have
$$
\partial_{t_m}\Bigl (\sum_i a_i^{\alpha}b_i^{\beta}\Bigr )=
\sum_i \left ( b_i^{\beta}\frac{\partial H_m}{\partial b_i^{\alpha}}-
a_i^{\alpha}\frac{\partial H_m}{\partial a_i^{\beta}}\right )
$$
and this is zero because $H_m$ is a linear combination of
$\mbox{tr}\, L^j(\lambda )$ and
$$
\sum_i \left (b_i^{\beta}\mbox{tr}\, \Bigl (\frac{\partial L}{\partial b_i^{\alpha}}\, L^{j-1}\Bigr )
-a_i^{\alpha}\, \mbox{tr}\, \Bigl (\frac{\partial L}{\partial a_i^{\beta}}\, L^{j-1}\Bigr )\right )
$$
$$
=\sum_i \sum_{l,k}\left (b_i^{\beta}\, \frac{\partial L_{lk}}{\partial b_i^{\alpha}}\, L_{kl}^{j-1}-
a_i^{\alpha}\, \frac{\partial L_{lk}}{\partial a_i^{\beta}}\, L_{kl}^{j-1}\right )
$$
$$
=\sum_i \sum_{l\neq k}(\delta_{ik}-\delta_{il})b_l^{\beta}a_k^{\alpha}\Phi (x_l-x_k)
L_{kl}^{j-1}=0.
$$
A simple lemma from linear algebra states that eigenvalues $\nu_{\alpha}$ of the
$n\times n$ matrix $G$ (\ref{t19}) coincide with nonzero eigenvalues of the rank $n$
$N\times N$ matrix $F$ with matrix elements
\begin{equation}\label{t20}
F_{ij}=b_i^{\nu}a_j^{\nu}
\end{equation}
(we assume that $n\leq N$). Indeed, consider the rectangular $N\times n$ matrix
${\bf A}_{i\alpha}=
a_i^{\alpha}$ and ${\bf B}_{i \beta}=b_i^{\beta}$, then
$G={\bf A}^T {\bf B}$, $F={\bf B}{\bf A}^T$ and a straightforward verification shows that
traces of all powers of these matrices coincide: $\mbox{tr}\, G^m=\mbox{tr}\, F^m$ for all
$m\geq 1$. This means that their nonzero eigenvalues also coincide. Note that
$\displaystyle{\mbox{tr}\, G=\sum_{\alpha =1}^n \nu_{\alpha}=N}$.
\section{The spectral curve}
The first of the equations (\ref{t6}) determines a connection between
the spectral parameters $k=k_{\beta}, \lambda=\lambda_{\beta}$ which
is the equation of the spectral curve:
\begin{equation}\label{spec1}
R(k, \lambda ):=\det \Bigl (kI-L(\lambda )\Bigr )=0.
\end{equation}
As we already mentioned, the spectral curve
is an integral of motion.
The matrix
$L=L(\lambda )$ has an essential singularity at $\lambda =0$. It can be
represented in the form $L=V\tilde L V^{-1}$, where
$V$ is the diagonal matrix $V_{ij}=\delta_{ij}
e^{-\zeta (\lambda )x_i}$.
Matrix elements of
$\tilde L$ do not have any
essential singularity in $\lambda$. We conclude that
$$
R(k, \lambda )=\sum_{m=0}^{N}R_m(\lambda )k^m,
$$
where the coefficients $R_m(\lambda )$ are elliptic functions of $\lambda$ with poles
at $\lambda =0$.
They can be represented as linear combinations of the $\wp$-function and
its derivatives, coefficients
of this expansion being integrals of motion. Fixing their values, we obtain
an algebraic curve $\Gamma$ which is an
$N$-sheet covering of the initial elliptic curve ${\cal E}$ realized as a factor
of the complex plane with respect to the lattice generated by $2\omega$, $2\omega '$.
In a neighborhood of the point $\lambda =0$
the matrix $\tilde L(\lambda)$
can be represented as
$$
\tilde L(\lambda )=\lambda ^{-1}(I-F)+O(1),
$$
where
$F$ is the rank $n$ matrix (\ref{t20}) (recall that
$n\leq N$). This matrix has $N-n$ vanishing eigenvalues and $n$ nonzero eigenvalues
$\nu_{\alpha}$, $\alpha =1, \ldots , n$. They are time-independent quantities because
as we have shown above they coincide with eigenvalues of the matrix $G$ (\ref{t19}) which is
an integral of motion.
Therefore, we can write
$$
\det \Bigl (kI-L(\lambda )\Bigr )=\prod_{\alpha =1}^n \Bigl (k-(1-\nu_{\alpha})\lambda^{-1}-
h_{\alpha}(\lambda )\Bigr )\prod_{j=n+1}^N \Bigl (k-\lambda^{-1}-
h_{j}(\lambda )\Bigr ),
$$
where $h_{\alpha}$, $h_{j}$ are regular functions of $\lambda$ near $\lambda =0$.
This means that the function $k$
has simple poles on all sheets at the points of the curve
$\Gamma$ located
above $\lambda =0$. Now,
recalling the connection between $k$ and $z$ given by the first equation in (\ref{e7}), we
have
\begin{equation}\label{spec2}
\det \Bigl ((z+\zeta (\lambda ))I-L(\lambda )\Bigr )=\prod_{\alpha =1}^n \Bigl (z+
\nu_{\alpha}\lambda^{-1}-
h_{\alpha}(\lambda )\Bigr )\prod_{j=n+1}^N \Bigl (z-
h_{j}(\lambda )\Bigr ).
\end{equation}
We see that $n$ sheets of the curve $\Gamma$ lying above a neighborhood of the point
$\lambda =0$ are distinguished. There are $n$ points at infinity above $\lambda =0$:
$P^{(\infty )}_1=(\infty _1, 0), \ldots P^{(\infty )}_n=(\infty _n, 0)$.
In the vicinity of the point $P^{(\infty )}_{\alpha}$
the function $\lambda =\lambda_{\alpha}(z)$ has the following expansion:
\begin{equation}\label{spec3}
\lambda =\lambda_{\alpha}(z)=-\nu_{\alpha}z^{-1} +O(z^{-2}).
\end{equation}
As it is shown in \cite{KBBT95}, the points $P^{(\infty )}_{\alpha}\in \Gamma$ are the
marked points, where the Baker-Akhiezer function on the spectral curve
has essential singularities.
With the expansion (\ref{spec3}) at hand, we can make a more detailed identification
of the wave function (\ref{m2}) with the expansion (\ref{m7})
and the wave function (\ref{t3}). The expansion of the function (\ref{t3})
as $\lambda_{\beta}\to 0$ yields
$$
\Psi_{\alpha \beta}=e^{zx+\xi ({\bf t}, z)}\sum_i
\Bigl (a_i^{\alpha}b_i^{\beta}\nu_{\beta}^{-1} +\lambda_{\beta}\nu_{\beta}^{-1}
(a_i^{\alpha}d_{i}^{\beta} +a_i^{\alpha}b_i^{\beta}\zeta (x-x_i)) +O(\lambda_{\beta}^2)\Bigr ),
$$
where we took into account that the identification implies the expansion
\begin{equation}\label{spec4}
c_i^{\beta}=\nu_{\beta}^{-1}\lambda_{\beta}^{-1}e^{-x_i\zeta (\lambda )}
\Bigl (b_i^{\beta}+\lambda_{\beta} d_i^{\beta}+O(\lambda_{\beta}^2)\Bigr ),
\quad \lambda_{\beta}\to 0.
\end{equation}
Therefore, taking into account (\ref{spec3}), we can write
$$
\Psi_{\alpha \beta}=e^{zx+\xi ({\bf t}, z)}\left (\sum_i
a_i^{\alpha}b_i^{\beta}\nu_{\beta}^{-1} +
z^{-1}\Bigl (S_{\alpha \beta}\! -\! \sum_i
a_i^{\alpha}b_i^{\beta}\zeta (x-x_i)\Bigr ) +O(z^{-2})\right ).
$$
Comparing with (\ref{m7}), we conclude that
\begin{equation}\label{spec5}
\sum_i a_i^{\alpha}b_i^{\beta}=\nu_{\alpha}\delta_{\alpha \beta}.
\end{equation}
It is easy to see that the Hamiltonian and the Lax matrix are invariant with respect to the
gauge transformation
\begin{equation}\label{spec6}
{\bf a}_i\longrightarrow W^{-1} {\bf a}_i, \qquad
{\bf b}_i^T\longrightarrow {\bf b}_i^T W
\end{equation}
with arbitrary non-degenerate $n\times n$ matrix $W$. Therefore,
after the transformation $G\to W^{-1}gW$ the matrix $G$ can always
be regarded as diagonal matrix, as in (\ref{spec5}), with the eigenvalues being the same
as nonzero eigenvalues $\nu_{\alpha}$ of the $N\times N$ matrix $F$.
\section{Dynamics of poles in the higher times}
Our basic tool is equation (\ref{m11b}). Substituting $\Psi$, $\Psi^{\dag}$ in the form
(\ref{t3}), (\ref{t4}) and $\xi^{(1)}$ in the form (\ref{t3a}), we have:
\begin{equation}\label{ht1}
\begin{array}{l}
\displaystyle{
\frac{1}{2\pi i}\oint_{C_{\infty}} dz \, z^m \sum_{i,j}
a_i^{\alpha}c_i^{\nu}c_j^{*\nu}b_j^{\beta}\Phi (x-x_i, \lambda_{\nu})
\Phi (x-x_j, -\lambda_{\nu})}
\\ \\
\displaystyle{\phantom{aaaaaaaaaaaaaaa}
=\sum_i \partial_{t_{\nu, m}}x_i a_i^{\alpha}b_i^{\beta}
\wp (x-x_i)+\sum_i \partial_{t_{\nu , m}}(a_i^{\alpha}b_i^{\beta})\zeta (x-x_i)}
\end{array}
\end{equation}
(no summation over $\nu$ here!).
Equating the coefficients in front of the second order poles at $x=x_i$, we get the
relation
\begin{equation}\label{ht2}
\partial_{t_{\nu, m}}x_i=\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m c_i^{*\nu}c_i^{\nu}\Bigr )=
\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m {\sf c}^{*\nu}E_i {\sf c}^{\nu}\Bigr ),
\end{equation}
where $E_i$ is the diagonal $N\times N$ matrix with matrix elements
$(E_i)_{jk}=\delta_{ij}\delta_{ik}$ (again, no summation over $\nu$).
Summing over $i$, we get
\begin{equation}\label{ht3}
\partial_{t_{\nu, m}}\sum_i x_i=\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m {\sf c}^{*\nu}{\sf c}^{\nu}\Bigr )
\end{equation}
Comparing with equation (\ref{e8}), we conclude that
\begin{equation}\label{ht4}
{\sf c}^{*\alpha}{\sf c}^{\alpha}=-\nu_{\alpha}z^{-2}+
\sum_{m\geq 2}z^{-m-1}\partial_{t_{\alpha, m}}\sum_i x_i=
-\lambda '_{\alpha}(z)
\end{equation}
(no summation over $\alpha$!).
Now, we note that $E_i=-\partial_{p_i}L$ and continute (\ref{ht2}) as the following
chain of equalities, using (\ref{t6}), (\ref{t7}), (\ref{ht4}):
$$
\partial_{t_m}x_i=\sum_{\nu}\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m {\sf c}^{*\nu}E_i {\sf c}^{\nu}\Bigr )=-
\sum_{\nu}\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m {\sf c}^{*\nu}\partial_{p_i}L(\lambda_{\nu}) {\sf c}^{\nu}\Bigr )
$$
$$
=-\sum_{\nu}\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m\partial_{p_i}\Bigl (
{\sf c}^{*\nu}L(\lambda_{\nu}) {\sf c}^{\nu}\Bigr )\Bigr )+
\sum_{\nu}\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m (\partial_{p_i}
{\sf c}^{*\nu})L(\lambda_{\nu}) {\sf c}^{\nu}\Bigr )+
\sum_{\nu}\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m
{\sf c}^{*\nu}L(\lambda_{\nu}) \partial_{p_i}{\sf c}^{\nu}\Bigr )
$$
$$
=-\sum_{\nu}\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m\partial_{p_i}\Bigl (
{\sf c}^{*\nu}L(\lambda_{\nu}) {\sf c}^{\nu}\Bigr )\Bigr )+
\sum_{\nu}\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m (\partial_{p_i}
{\sf c}^{*\nu})k_{\nu}{\sf c}^{\nu}\Bigr )+
\sum_{\nu}\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m
{\sf c}^{*\nu}k_{\nu} \partial_{p_i}{\sf c}^{\nu}\Bigr )
$$
$$
=-\sum_{\nu}\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m\partial_{p_i}\Bigl (
{\sf c}^{*\nu}k_{\nu} {\sf c}^{\nu}\Bigr )\Bigr )+
\sum_{\nu}\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m k_{\nu}
\partial_{p_i}({\sf c}^{*\nu} {\sf c}^{\nu})\Bigr )
$$
$$
=\sum_{\nu}\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m \lambda'_{\nu}(z)\partial_{p_i}k_{\nu}\Bigr ).
$$
Regarding $z$ as an independent variable, we apply the same argument as in
\cite{PZ21} to obtain
\begin{equation}\label{ht5}
\partial_{t_m}x_i=-\sum_{\nu}\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m \partial_{p_i}\lambda_{\nu}(z)\Bigr ).
\end{equation}
In this way we obtain the first half of the higher Hamiltonian equations for poles
\begin{equation}\label{ht6}
\partial_{t_m}x_i =\frac{\partial H_m}{\partial p_i}
\end{equation}
with the Hamiltonian
\begin{equation}\label{ht7}
H_m=\sum_{\alpha=1}^n \mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m\lambda_{\alpha}(z)\Bigr ).
\end{equation}
The second half of the Hamiltonian equations for poles can be obtained by
taking the $t_2$-derivative of (\ref{ht2}) and using (\ref{t6}), (\ref{t7}).
In this way we obtain:
\begin{equation}\label{ht8}
\partial_{t_{\nu, m}}\dot x_i=\mathop{\hbox{res}}\limits_{\infty}\Bigl (z^m {\sf c}^{*\nu}[E_i, M(\lambda_{\nu})]
{\sf c}^{\nu}\Bigr ).
\end{equation}
A straightforward verification shows that
\begin{equation}\label{ht9}
[E_i, M(\lambda )]=2\partial_{x_i}L(\lambda ).
\end{equation}
Recalling also that $\dot x_i=2p_i$, we rewrite (\ref{ht8}) as
\begin{equation}\label{ht10}
\partial_{t_{\nu, m}}p_i=\mathop{\hbox{res}}\limits_{\infty}\Bigl (z^m {\sf c}^{*\nu}\partial_{x_i}L(\lambda_{\nu} )
{\sf c}^{\nu}\Bigr )
\end{equation}
(no summation over $\nu$!). With the relation (\ref{ht10}) at hand, one can repeat
the chain of equalities after equation (\ref{ht4}) with the change $\partial_{p_i}\to \partial_{x_i}$
to obtain
\begin{equation}\label{ht11}
\partial_{t_m}p_i=\sum_{\nu}\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m \partial_{x_i}\lambda_{\nu}(z)\Bigr ),
\end{equation}
so that
\begin{equation}\label{ht6a}
\partial_{t_m}p_i =-\frac{\partial H_m}{\partial x_i}
\end{equation}
with the same Hamiltonian (\ref{ht7}).
Let us make some comments on a more general case when the tau-function for elliptic
solutions has a slightly more general form
\begin{equation}\label{pt17}
\tau (x, {\bf t})=e^{Q(x, {\bf t})}\prod_{i=1}^N \sigma (x-x_i({\bf t})),
\end{equation}
where
\begin{equation}\label{pt18}
Q(x, {\bf t})=c(x+t_1)^2 +(x+t_1)\sum_{j\geq 2}a_j t_j +b(t_2, t_3, \ldots )
\end{equation}
with some constants $c$, $a_j$ and a function $b(t_2, t_3, \ldots )$.
Repeating the arguments leading to (\ref{ht5}), one can see that now the
first equation in (\ref{e7}) will be modified as
\begin{equation}\label{pt19}
k_{\beta}=z -\alpha (z)+\zeta (\lambda_{\beta}), \qquad
\alpha (z)=2cz^{-1}+\sum_{j\geq 2}\frac{a_j}{j}\, z^{-j}.
\end{equation}
Instead of (\ref{ht5}) we will have
\begin{equation}\label{pt20}
\partial_{t_m}x_i=-\sum_{\nu}\mathop{\hbox{res}}\limits_{\infty} \Bigl (z^m \partial_{p_i}\lambda_{\nu}(z)(1-\alpha '(z))\Bigr ),
\end{equation}
so the Hamiltonian for the $m$-th flow will be a linear combination of $H_m$ and $H_j$ with
$1\leq j<m$.
\section{Dynamics of spin variables in the higher times}
The Hamiltonian dynamics of spin variables in the higher times can be derived
by analysis of first order poles in (\ref{ht1}). Equating coefficients in front of
first order poles, we get the relation
$$
\partial_{t_{\nu , m}}(a_i^{\alpha}b_i^{\beta})=\mathop{\hbox{res}}\limits_{\infty}\left (
z^m \sum_{j\neq i} a_{i}^{\alpha}c_i^{\nu}c_j^{*\nu}b_j^{\beta}\Phi (x_i\! -\! x_j, -\lambda_{\nu})
+z^m\sum_{j\neq i} a_{j}^{\alpha}c_j^{\nu}c_i^{*\nu}b_i^{\beta}\Phi (x_i\! -\! x_j, \lambda_{\nu})
\right )
$$
which can be rewritten as
$$\hspace{-2cm}
a_i^{\alpha}\left [\partial_{t_{\nu , m}}b_i^{\beta} +\mathop{\hbox{res}}\limits_{\infty}\Bigl (
z^m c_i^{\nu}\sum_{j\neq i}c_j^{*\nu}b_j^{\beta}
\Phi (x_j\! -\! x_i, \lambda_{\nu})\Bigr )\right ]
$$
$$\hspace{2cm}
+b_i^{\beta}\left [\partial_{t_{\nu , m}}a_i^{\alpha} -\mathop{\hbox{res}}\limits_{\infty}\Bigl (
z^m c_i^{*\nu}\sum_{j\neq i}c_j^{\nu}a_j^{\alpha}
\Phi (x_i\! -\! x_j, \lambda_{\nu})\Bigr )\right ]=0.
$$
Now we notice that
\begin{equation}\label{spin1}
\begin{array}{l}
\displaystyle{
\frac{\partial L_{jk}(\lambda )}{\partial a_i^{\alpha}}=-\delta_{ik}(1-\delta_{jk})
b_j^{\alpha}\Phi (x_j-x_i, \lambda ),}
\\ \\
\displaystyle{
\frac{\partial L_{jk}(\lambda )}{\partial b_i^{\beta}}=-\delta_{ij}(1-\delta_{jk})
a_k^{\beta}\Phi (x_i-x_k, \lambda ),}
\end{array}
\end{equation}
so the equation above can be written as
\begin{equation}\label{spin2}
\begin{array}{l}
\displaystyle{a_i^{\alpha}\left [\partial_{t_{\nu , m}}b_i^{\beta} -\mathop{\hbox{res}}\limits_{\infty}\Bigl (
z^m {\sf c}^{*\nu}\frac{\partial L(\lambda_{\nu} )}{\partial a_i^{\beta}}\, {\sf c}^{\nu}\Bigr )\right ]}
\displaystyle{
+b_i^{\beta}\left [\partial_{t_{\nu , m}}a_i^{\alpha} +\mathop{\hbox{res}}\limits_{\infty}\Bigl (
z^m {\sf c}^{*\nu}\frac{\partial L(\lambda_{\nu} )}{\partial b_i^{\alpha}}\, {\sf c}^{\nu}\Bigr )\right ]=0}.
\end{array}
\end{equation}
Having this equation at hand, one can repeat
the chain of equalities after equation (\ref{ht4}) with the changes $\partial_{p_i}\to
\partial / \partial a_i^{\beta}$, $\partial_{p_i}\to
\partial / \partial b_i^{\alpha}$
to obtain
\begin{equation}\label{spin3}
a_i^{\alpha}P_i^{\beta}-b_i^{\beta}Q_i^{\alpha}=0,
\end{equation}
where
\begin{equation}\label{spin4}
P_i^{\beta}=-\partial_{t_m}b_i^{\beta}+\sum_{\nu} \mathop{\hbox{res}}\limits_{\infty}\Bigl (
z^m \frac{\partial}{\partial a_i^{\beta}}\, \lambda_{\nu}(z)\Bigr )=-\partial_{t_m}b_i^{\beta}-
\frac{\partial H_m}{\partial a_i^{\beta}},
\end{equation}
\begin{equation}\label{spin5}
Q_i^{\alpha}=\partial_{t_m}a_i^{\alpha}+\sum_{\nu} \mathop{\hbox{res}}\limits_{\infty}\Bigl (
z^m \frac{\partial}{\partial b_i^{\alpha}}\, \lambda_{\nu}(z)\Bigr )=\partial_{t_m}a_i^{\alpha}-
\frac{\partial H_m}{\partial b_i^{\alpha}}.
\end{equation}
It follows from (\ref{spin3}) that
$$
\frac{Q_i^{\alpha}}{a_i^{\alpha}}=\frac{P_i^{\beta}}{b_i^{\beta}}=\Lambda_i^{(m)},
$$
and the equations (\ref{spin4}), (\ref{spin5}) acquire the form
\begin{equation}\label{spin6}
\partial_{t_m}a_i^{\alpha}=a_i^{\alpha}\Lambda_i^{(m)}+\frac{\partial H_m}{\partial b_i^{\alpha}},
\end{equation}
\begin{equation}\label{spin6a}
\partial_{t_m}b_i^{\beta}=-b_i^{\beta}\Lambda_i^{(m)}-\frac{\partial H_m}{\partial a_i^{\beta}}.
\end{equation}
The gauge transformation $a_i^{\alpha}\to a_i^{\alpha} q_i^{(m)}$,
$b_i^{\alpha}\to b_i^{\alpha} (q_i^{(m)})^{-1}$ with
$\displaystyle{q_i^{(m)}=\exp \left (\int^{t_m}\Lambda_i^{(m)}dt\right )}$
eliminates the terms with $\Lambda_i^{(m)}$, so we can put $\Lambda_i^{(m)}=0$.
We obtain the Hamiltonian equations of motion for spin variables
in the higher times:
\begin{equation}\label{spin7}
\partial_{t_m}a_i^{\alpha}=\frac{\partial H_m}{\partial b_i^{\alpha}}, \qquad
\partial_{t_m}b_i^{\alpha}=-\frac{\partial H_m}{\partial a_i^{\alpha}}.
\end{equation}
with $H_m$ given by (\ref{ht7}).
\section{How to obtain the first two Hamiltonians}
In order to find the Hamiltonians, we need to expand the spectral curve near
$\lambda =0$. Using the expansion (\ref{Phi1}), we represent the equation of the
spectral curve as
\begin{equation}\label{H1}
\det \Bigl (zI+F\lambda^{-1} +Q +S\lambda +O(\lambda^2)\Bigr )=0,
\end{equation}
where the matrices $Q$, $S$ are
\begin{equation}\label{H2}
Q_{ij}=p_i\delta_{ij}+(1-\delta_{ij})F_{ij}\zeta (x_i-x_j),
\end{equation}
\begin{equation}\label{H3}
S_{ij}=\frac{1}{2}\, (1-\delta_{ij})\,F_{ij}\Bigl (\zeta^2(x_i-x_j)-\wp (x_i-x_j)\Bigr ).
\end{equation}
We set
\begin{equation}\label{H5}
z=-\frac{\omega}{\lambda},
\end{equation}
then the equation (\ref{H1}) acquires the form
\begin{equation}\label{H1a}
\det \Bigl (\omega I-F -Q\lambda -S\lambda^2 +O(\lambda^3)\Bigr )=0.
\end{equation}
This equation has $n$ roots $\omega_{\alpha}$ such that
\begin{equation}\label{H4}
\omega_{\alpha}=\omega_{\alpha}(\lambda )=\nu_{\alpha}+\omega_1^{(\alpha )}\lambda +
\omega_2^{(\alpha )}\lambda^2 +O(\lambda^3)
\end{equation}
and $N-n$ roots which are $O(\lambda )$. These roots are eigenvalues of the matrix
$F +Q\lambda +S\lambda^2 +O(\lambda^3)$. Expressing $\lambda$ through $z$ from equation
(\ref{H5}) and expanding in powers of $z^{-1}$, we have
\begin{equation}\label{H6}
\lambda_{\alpha}=-\frac{\nu_{\alpha}}{z}+\nu_{\alpha}\omega_1^{(\alpha )}z^{-2}-
(\nu^2_{\alpha}\omega_2^{(\alpha )}+\nu_{\alpha}(\omega_1^{(\alpha )})^2)z^{-3}+O(z^{-4}).
\end{equation}
Then
\begin{equation}\label{H7}
\begin{array}{l}
\displaystyle{
H_1=-\sum_{\alpha}\nu_{\alpha}\omega_1^{(\alpha )}}
\\ \\
\displaystyle{
H_2=\sum_{\alpha}(\nu^2_{\alpha}\omega_2^{(\alpha )}+\nu_{\alpha}(\omega_1^{(\alpha )})^2)}.
\end{array}
\end{equation}
We regard the matrix $Q\lambda +S\lambda^2$ as a small variation of the matrix $F$.
The idea is to find the variation of the eigenvalues (the corrections
$\omega_1^{(\alpha )}\lambda +
\omega_2^{(\alpha )}\lambda^2$ in (\ref{H4})) using first two orders of the
perturbation theory.
Let $\psi^{(j)}$ be a basis in the $N$-dimensional space and $\tilde \psi^{(j)}$
be the dual basis such that $(\tilde \psi^{(i)}\psi^{(j)})=0$ at $i\neq j$. We take first
$n$ vectors to be
$$
\psi^{(\alpha )}_i=b^{\alpha}_i , \qquad
\tilde \psi^{(\alpha )}_i=a^{\alpha}_i,
$$
then
$$
(\tilde \psi^{(\alpha )}\psi^{(\beta )})=\sum_ia^{\alpha}_ib^{\beta}_i=
\nu_{\alpha}\delta_{\alpha \beta}, \quad \alpha , \beta =1, \ldots , n.
$$
These vectors are eigenvectors of the (non-perturbed) matrix $F$ with nonzero eigenvalues:
\begin{equation}\label{H8}
F\psi^{(\alpha )}=\nu_{\alpha}\psi^{(\alpha )}, \quad
\tilde \psi^{(\alpha )}F=\nu_{\alpha}\tilde \psi^{(\alpha )}.
\end{equation}
The other $N-n$ vectors are chosen to be orthonormal:
$$
(\tilde \psi^{(i)}\psi^{(j)})=\delta_{ij}, \quad i,j=n+1, \ldots , N.
$$
In the first order of the perturbation theory we have:
\begin{equation}\label{H9}
\omega_1^{(\alpha )}=
\frac{(\tilde \psi^{(\alpha )}Q\psi^{(\alpha )})}{(\tilde \psi^{(\alpha )}\psi^{(\alpha )})}.
\end{equation}
The next coefficient, $\omega_2^{(\alpha )}$, is obtained in the second order of the
perturbation theory as
\begin{equation}\label{H10}
\omega_2^{(\alpha )}=\frac{(\tilde \psi^{(\alpha )}S\psi^{(\alpha )})}{(\tilde \psi^{(\alpha )}\psi^{(\alpha )})}+
\sum_{j\neq \alpha}\frac{(\tilde \psi^{(\alpha )}Q\psi^{(j )})\,
(\tilde \psi^{(j )}Q\psi^{(\alpha )})}{(\tilde \psi^{(\alpha )}\psi^{(\alpha )})
(\tilde \psi^{(j )}\psi^{(j )})(\nu_{\alpha}-\nu_j)}.
\end{equation}
In the denominator of the last term $\nu_j =\nu_{\beta}$ at $j=\beta$, $\beta =1, \ldots , n$ and
$\nu_j=0$ at $j=n+1, \ldots , N$.
Using these formulas, we have:
\begin{equation}\label{H11}
\sum_{\alpha}\nu_{\alpha}\omega_1^{(\alpha )}=\sum_{\alpha}
(\tilde \psi^{(\alpha )}Q\psi^{(\alpha )})=
\sum_{\alpha}\sum_{i,j}a_i^{\alpha}Q_{ij}b_j^{\alpha}=
\sum_{i,j}F_{ji}Q_{ij}=\mbox{tr}\, (FQ),
\end{equation}
$$
\sum_{\alpha}\Bigl (\nu^2_{\alpha}\omega_2^{(\alpha )}+\nu_{\alpha}(\omega_1^{(\alpha )})^2\Bigr )
=\sum_{\alpha \neq \beta}\frac{\nu_{\alpha}(\tilde \psi^{(\alpha )}Q\psi^{(\beta )})
(\tilde \psi^{(\beta )}Q\psi^{(\alpha )})}{\nu_{\beta} (\nu_{\alpha}-\nu_{\beta})}+
\sum_{j=n+1}^N \sum_{\alpha}(\tilde \psi^{(\alpha )}Q\psi^{(j )})
(\tilde \psi^{(j )}Q\psi^{(\alpha )})
$$
$$\phantom{aaaaaaaaaaa}+\sum_{\alpha}\nu_{\alpha}^{-1}
(\tilde \psi^{(\alpha )}Q\psi^{(\alpha )})^2 +\sum_{\alpha}\nu_{\alpha}
(\tilde \psi^{(\alpha )}S\psi^{(\alpha )})
$$
$$
=\sum_{\alpha , \beta}\nu_{\alpha}^{-1}(\tilde \psi^{(\alpha )}Q\psi^{(\beta )})
\tilde \psi^{(\beta )}Q\psi^{(\alpha )})+
\sum_{j=n+1}^N \sum_{\alpha}(\tilde \psi^{(\alpha )}Q\psi^{(j )})
(\tilde \psi^{(j )}Q\psi^{(\alpha )})
+\sum_{\alpha}\nu_{\alpha}
(\tilde \psi^{(\alpha )}S\psi^{(\alpha )})
$$
$$
=\sum_{ijkl}\Bigl (\sum_{\alpha}\nu_{\alpha}^{-1}a_i^{\alpha}b_l^{\alpha}+
\sum_{r=n+1}^N \tilde \psi^{(r)}_i\psi^{(r)}_l\Bigr )Q_{ij}F_{jk}Q_{kl}+
\sum_{\alpha}\nu_{\alpha}a_i^{\alpha}S_{ij}b_j^{\alpha}.
$$
But
$$
\sum_{\alpha}\nu_{\alpha}^{-1}a_i^{\alpha}b_l^{\alpha}+
\sum_{r=n+1}^N \tilde \psi^{(r)}_i\psi^{(r)}_l=\delta_{il}
$$
(the completeness relation),
and so finally we obtain
\begin{equation}\label{H12}
\sum_{\alpha}\Bigl (\nu^2_{\alpha}\omega_2^{(\alpha )}+
\nu_{\alpha}(\omega_1^{(\alpha )})^2\Bigr )=
\mbox{tr}\, (QFQ)+\mbox{tr}\, (FSF).
\end{equation}
From (\ref{H11}) we obtain
\begin{equation}\label{H13}
H_1=-\mbox{tr}\, (FQ)=-\sum_i p_i F_{ii}-\sum_{i\neq j}F_{ij}F_{ji}\zeta (x_i-x_j)=
-\sum_ip_i,
\end{equation}
which is indeed the first Hamiltonian.
The calculation of (\ref{H12}) is more involved. We have, after some cancellations:
$$
\mbox{tr}\, (QFQ)=\sum_i p_i^2+
\sum_{k\neq i}\sum_{j\neq i}F_{ij}F_{jk}F_{ki}\zeta (x_i-x_j)\zeta (x_k-x_i),
$$
$$
\mbox{tr}\, (FSF)=\frac{1}{2}\sum_l \sum_{i\neq j}F_{li}F_{ij}F_{jl}
\Bigl (\zeta ^2(x_i-x_j)-\wp (x_i-x_j)\Bigr ).
$$
Therefore,
\begin{equation}\label{H14}
H_2=\sum_i p_i^2 -\sum_{i\neq j}F_{ij}F_{ji}\wp (x_i-x_j)+{\cal F},
\end{equation}
where
\begin{equation}\label{H15}
{\cal F}=\sum {}^{'}\! F_{ij}F_{jk}F_{ki}\zeta (x_i-x_j)\zeta (x_k-x_i)+
\frac{1}{2}\sum {}^{'}\! F_{ij}F_{jk}F_{ki}\Bigl (\zeta ^2(x_i-x_j)-\wp (x_i-x_j)\Bigr )=0.
\end{equation}
Here $\sum {}^{'}$ means summation over all distinct indices $ijk$. The proof of identity
(\ref{H15}) is given in Appendix A. To conclude, we have reproduced the correct Hamiltonians
$H_1$ and $H_2$ within our approach.
\section{Rational and trigonometric limits}
In the rational limit $\omega , \omega ' \to \infty$, $\sigma (\lambda )=\lambda$,
$\Phi (x, \lambda )=(x^{-1}+\lambda^{-1})e^{-x/\lambda}$ and the equation of the spectral
curve becomes
\begin{equation}\label{r1}
\det \Bigl (zI-L_{\rm rat} +\lambda^{-1}F\Bigr )=0,
\end{equation}
where
\begin{equation}\label{r2}
(L_{\rm rat})_{ij}=-\delta_{ij}p_i -(1-\delta_{ij})\frac{b_i^{\nu}a_j^{\nu}}{x_i-x_j}
\end{equation}
is the Lax matrix of the spin generalization of the rational Calogero-Moser model.
Let us rewrite the equation of the spectral curve in the form
\begin{equation}\label{r3}
\det \left ( \lambda I+F \frac{1}{zI-L_{\rm rat}}\right )=0.
\end{equation}
Expanding the determinant, we have:
\begin{equation}\label{r4}
\lambda^N +\sum_{j=1}^{n}D_j(z) \lambda^{N-j}=0, \qquad D_1(z)=\mbox{tr}\,
\Bigl (F \frac{1}{zI-L_{\rm rat}} \Bigr ),
\end{equation}
where we took into account that rank of $F$ is equal to $n\leq N$. Let us note that the
functions $\lambda_{\alpha}(z)$ are different nonzero roots of equation (\ref{r4}) and
the sum of these roots is equal to $-D_1(z)$. Therefore, we can write
\begin{equation}\label{r5}
H_m=-\sum_{\nu}\mathop{\hbox{res}}\limits_{\infty}\Bigl (z^m\lambda_{\nu}(z)\Bigr )=
\mathop{\hbox{res}}\limits_{\infty}\Bigl (z^m \mbox{tr}\, \Bigl (
F \frac{1}{zI-L_{\rm rat}} \Bigr )\Bigr )=\mbox{tr}\Bigl (FL_{\rm rat}^m\Bigr ).
\end{equation}
It is straightforward to check the commutation relation
\begin{equation}\label{r6}
[X,L_{\rm rat}]=F-I, \qquad X=\mbox{diag} \, (x_1, \ldots , x_N).
\end{equation}
Substituting it into (\ref{r5}), we see that
\begin{equation}\label{r7}
H_m=\mbox{tr}L_{\rm rat}^m.
\end{equation}
This is the result of paper \cite{PZ18} obtained there by another method.
We now pass to the trigonometric limit. We choose the
period of the trigonometric (or hyperbolic) functions to be $\pi i/\gamma$, where
$\gamma$ is some complex constant (real for hyperbolic functions and purely imaginary
for trigonometric functions).
The second period tends to infinity.
The Weierstrass functions in this limit become
$$
\sigma (x)=\gamma^{-1}e^{-\frac{1}{6}\, \gamma^2x^2}\sinh (\gamma x),
\quad
\zeta (x)=\gamma \coth (\gamma x)-\frac{1}{3}\, \gamma^2 x.
$$
The tau-function for trigonometric solutions is \cite{PZspintrig}
\begin{equation}\label{trig1}
\tau =\prod_{i=1}^N \Bigl (e^{2\gamma x}-e^{2\gamma x_i}\Bigr ),
\end{equation}
so we should consider
\begin{equation}\label{trig2}
\tau =\prod_{i=1}^N \sigma (x-x_i)e^{\frac{1}{6}\, \gamma^2
(x-x_i)^2 +\gamma (x+x_i)}.
\end{equation}
Similarly to the KP case \cite{PZ21},
equation (\ref{e7}) with this choice acquires the form
\begin{equation}\label{trig3}
k_{\beta}=z+\gamma \coth (\gamma \lambda_{\beta} ).
\end{equation}
The trigonometric limit of the function $\Phi (x, \lambda )$ is
$$
\Phi (x, \lambda )=\gamma \Bigl (\coth (\gamma x)+\coth (\gamma \lambda )\Bigr )
e^{-\gamma x \coth (\gamma \lambda )}.
$$
For further calculations it is convenient to pass to the variables
\begin{equation}\label{trig4}
w_i=e^{2\gamma x_i}
\end{equation}
and introduce the diagonal matrix $W=\mbox{diag}\, (w_1, w_2, \ldots , w_N)$.
In this notation, the equation of the spectral curve acquires the form
\begin{equation}\label{trig5}
\det \Bigl (W^{1/2}(zI-(L_{\rm trig}-\gamma I))W^{-1/2}+\gamma (\coth (\gamma \lambda )-1)F
\Bigr )=0,
\end{equation}
where $L_{\rm trig}$ is the Lax matrix of the spin Calogero-Moser model with matrix elements
\begin{equation}\label{trig6}
(L_{\rm trig})_{ij}=-p_i\delta_{ij}-\frac{(1-\delta_{ij})
\gamma F_{ij}}{\sinh (\gamma (x_i-x_j))}=
-p_i\delta_{ij}-2(1-\delta_{ij})\frac{\gamma w_i^{1/2}w_j^{1/2}F_{ij}}{w_i-w_j}.
\end{equation}
Some simple transformations allow one to bring the equation of the spectral curve
to the form
\begin{equation}\label{trig7}
\det \left (\omega I+2\gamma W^{-1/2}FW^{1/2}\, \frac{1}{zI\! -\!
(L_{\rm trig}-\gamma I)}\right )=0,
\qquad \omega =e^{2\gamma \lambda}-1.
\end{equation}
Expanding the determinant, we have:
\begin{equation}\label{trig8}
\omega^N +\sum_{j=1}^{n}K_j(z) \omega^{N-j}=0,
\end{equation}
where we took into account that rank of $F$ is equal to $n\leq N$. In particular,
$$
K_1=\mbox{tr}\, Y, \quad K_2=\frac{1}{2} (\mbox{tr}^2 Y-\mbox{tr}\, Y^2),
$$
where $Y$ is the matrix
$$
Y=2\gamma W^{-1/2}FW^{1/2}\, \frac{1}{zI\! -\!
(L_{\rm trig}-\gamma I)}.
$$
The coefficients
$K_j$ are expressed through the
elementary symmetric polynomials $e_j=e_j(\omega_1, \ldots , \omega_n)$
of nonzero roots $\omega_\nu=\omega_{\nu}(z)$ of this equation
as $K_j=(-1)^j e_j(\omega_1, \ldots , \omega_n)$. Therefore,
$$
\sum_{\nu}\lambda_{\nu}(z)=\frac{1}{2\gamma}\sum_{\nu}\log (1+\omega_{\nu} (z))
=\frac{1}{2\gamma}\log \prod_{\nu} (1+\omega_{\nu} (z))
$$
$$
=\frac{1}{2\gamma}\log \Bigl (\sum_{j=0}^n e_j (\omega_1, \ldots , \omega_n)\Bigr )=
\frac{1}{2\gamma}\log \Bigl (\sum_{j=0}^n (-1)^jK_j \Bigr ).
$$
From this we conclude that
\begin{equation}\label{trig9}
\sum_{\nu}\lambda_{\nu}(z)=
\frac{1}{2\gamma}\, \log \det \left [I-2\gamma
W^{-1/2}FW^{1/2}
\frac{1}{zI-(L_{\rm trig}-\gamma I)}\right ].
\end{equation}
Starting from this point, one can literally repeat the corresponding calculation
from \cite{PZ21} with the change of the rank 1 matrix $E$ to the rank $n$ matrix $F$ and
using the easily proved relation
\begin{equation}\label{trig10}
[L_{\rm trig},W]=2\gamma (W^{1/2}FW^{1/2}-W).
\end{equation}
The result is
\begin{equation}\label{trig11}
\begin{array}{c}
\displaystyle{\sum_{\nu}\lambda_{\nu}(z) =\frac{1}{2\gamma}\, \mbox{tr} \Bigl
(\log (I-z^{-1}(L_{\rm trig}+\gamma I))-
\log (I-z^{-1}(L_{\rm trig}-\gamma I))\Bigr )}
\\ \\
\displaystyle{=-\frac{1}{2\gamma}\, \mbox{tr} \sum_{m\geq 1}\frac{z^{-m}}{m}
\Bigl ((L_{\rm trig}+\gamma I)^m -(L_{\rm trig}-\gamma I)^m \Bigr )}
\end{array}
\end{equation}
and
\begin{equation}\label{trig12}
H_m=\frac{1}{2\gamma (m+1)}\, \mbox{tr} \Bigl ((L_{\rm trig}+\gamma I)^{m+1} -
(L_{\rm trig}-\gamma I)^{m-1} \Bigr )
\end{equation}
which agrees with the result of paper \cite{PZspintrig}.
\section*{Appendix A: Proof of identity (\ref{H15})}
\defA\arabic{equation}{A\arabic{equation}}
\setcounter{equation}{0}
\addcontentsline{toc}{section}{\hspace{6mm}Appendix A}
Here we prove identity (\ref{H15}) ${\cal F}=0$, where
$$
{\cal F}=\sum {}^{'}\! F_{ij}F_{jk}F_{ki}\zeta (x_i-x_j)\zeta (x_k-x_i)+
\frac{1}{2}\sum {}^{'}\! F_{ij}F_{jk}F_{ki}\Bigl (\zeta ^2(x_i-x_j)-\wp (x_i-x_j)\Bigr ).
$$
and $\sum {}^{'}$ means summation over all distinct indices $ijk$. Using the behavior
of the $\zeta$-function under shifts by periods
$$
\zeta (x+2\omega )=\zeta (x)+2\eta , \quad \zeta (x+2\omega ' )=\zeta (x)+2\eta ',
$$
one can see that ${\cal F}$ is a double-periodic function of any of $x_i$. Consider,
for example, the shift of $x_1$ by $2\omega$. The terms in
${\cal F}(x_1+2\omega )-{\cal F}(x_1)$ proportional to $\eta^2$ are:
$$
-(2\eta )^2\!\!\sum_{j\neq k\neq 1} \! F_{1j}F_{jk}F_{k1}+\frac{1}{2}\, (2\eta )^2
\!\!\sum_{j\neq k\neq 1} \!F_{1j}F_{jk}F_{k1}+\frac{1}{2}\, (2\eta )^2
\!\! \sum_{i\neq k\neq 1}\! F_{1k}F_{ki}F_{i1}=0.
$$
The terms proportional to $\eta$ are:
$$
-2\eta \!\!\sum_{j\neq k\neq 1} \! F_{1j}F_{jk}F_{k1}\zeta (x_1-x_k)-
2\eta \!\!\sum_{j\neq k\neq 1} \! F_{1j}F_{jk}F_{k1}\zeta (x_1-x_j)
$$
$$
+2\eta \!\!\sum_{j\neq k\neq 1} \! F_{j1}F_{1k}F_{kj}\zeta (x_j-x_k)+
2\eta \!\!\sum_{j\neq k\neq 1} \! F_{j1}F_{1k}F_{kj}\zeta (x_k-x_j)
$$
$$
+2\eta \!\!\sum_{j\neq k\neq 1} \! F_{k1}F_{1j}F_{jk}\zeta (x_1-x_j)
+2\eta \!\!\sum_{j\neq k\neq 1} \! F_{j1}F_{1k}F_{kj}\zeta (x_1-x_j)=0.
$$
Therefore, we see that ${\cal F}(x_1+2\omega )={\cal F}(x_1)$.
The double-periodicity in all other arguments is established in the same way.
Next, the function ${\cal F}$ as a function of $x_1$ may have poles only at the points
$x_i$, $i=2, \ldots , N$. The second order poles cancel identically in the obvious way.
We find the residue at the
simple pole at $x_1=x_2$ as follows:
$$
-\sum_{k\neq 1,2}F_{12}F_{2k}F_{k1}\zeta (x_1-x_k)-
\sum_{j\neq 1,2}F_{1j}F_{j2}F_{21}\zeta (x_1-x_j)
$$
$$
+\sum_{k\neq 1,2}F_{21}F_{1k}F_{k2}\zeta (x_1-x_k)
+\sum_{j\neq 1,2}F_{2j}F_{j1}F_{12}\zeta (x_1-x_j)=0.
$$
Vanishing of the residues in all other points and for all other variables
can be proved in the same way.
We see that the function ${\cal F}$ is a regular elliptic function and, therefore,
it must be a constant. To find this constant, we set $x_j=j\varepsilon$
and tend $\varepsilon$ to 0.
Thanks to the fact that $\zeta (x)=x^{-1}+O(x^3)$,
$\wp (x)=x^{-2}+O(x^2)$ as $x\to 0$, we find:
$$
{\cal F}=\frac{1}{\varepsilon^2}\sum {}^{'}\!\frac{F_{ij}F_{jk}F_{ki}}{(i-j)(k-i)}+O(\varepsilon^2)
$$
Making the cyclic changes of the summation variables $(ijk)\to (jki)$ and
$(ijk)\to (kij)$, we have:
$$
{\cal F}=\frac{1}{3\varepsilon^2}\sum {}^{'}\! F_{ij}F_{jk}F_{ki}
\left (\frac{1}{(i-j)(k-i)}+ \frac{1}{(j-k)(i-j)}+\frac{1}{(k-i)(j-k)}\right )+O(\varepsilon^2)
$$
$$
=\frac{1}{3\varepsilon^2}\sum {}^{'}\! F_{ij}F_{jk}F_{ki}\,
\frac{(j-k)+(k-i)+(i-j)}{(i-j)(j-k)(k-i)}+O(\varepsilon^2)=O(\varepsilon^2).
$$
Therefore, we conclude that ${\cal F}=0$ and the identity (\ref{H15}) is proved.
\section*{Acknowledgments}
\addcontentsline{toc}{section}{\hspace{6mm}Acknowledgments}
The research of A.Z. has been funded within the framework of the
HSE University Basic Research Program and the Russian Academic Excellence Project '5-100'.
|
\section{Introduction}
\label{sec:Introduction}
Concepts of quantum information theory play an increasingly important role for describing and understanding complex quantum systems. Starting point is usually the von Neumann or quantum entropy associated to a state or density matrix $\rho$ \cite{vonNeumann1955,Nielsen2010,Wilde2013,Bengtsson2017},
\begin{equation}
S (\rho) = - \text{Tr} \left\{ \rho \ln \rho\right\}.
\label{eq:defvNEnt}
\end{equation}
It quantifies missing information or uncertainty encoded in the state $\rho$ \cite{Shannon1948,Jaynes1963}. Despite its importance in quantum information theory, the von Neumann entropy $S(\rho)$ has a major drawback: it can typically not be accessed directly in experiments. A computation of the von Neumann entropy requires detailed knowledge of the quantum state $\rho$, which is oftentimes out of reach. While there are experimental approaches to reconstruct the full state $\rho$ through quantum state tomography \cite{Nielsen2010}, the effort for these techniques scales unfavorably with system size. Alternatively, quantum entropies can be measured in (relatively small) quantum systems with finite dimensional Hilbert space $\text{dim} \mathcal{H} < \infty$ by considering multiple copies of the quantum system and using quantum many-body interference \cite{Abanin2012,Islam2015}.
For infinite dimensional quantum states as they arise for example in continuous variable systems or quantum field theory it is substantially more difficult to access or constrain quantum entropies experimentally.
In this work, we investigate a somewhat different approach and consider the Wehrl entropy $S_{\text{W}} (\rho)$, which is an entropy of the Husimi $Q$-distribution, as a quasi-classical measure for missing information in quantum mechanical phase space \cite{Wehrl1978,Wehrl1979}. While the density operator $\rho$ and the Husimi $Q$-distribution contain both the full information about the state of the quantum system (see e.\ g.\ ref.\ \cite{Schleich2001}), the latter has the advantage that it can be accessed experimentally (see e.\ g.\ \cite{Kunkel2019} for recent developments in spinor Bose-Einstein condensates). Therefore, we suggest a more regular use of the Wehrl entropy $S_\text{W} (\rho)$ when analyzing tabletop experiments from an information theoretic perspective.
In this work, we focus on two major branches of modern quantum information theory, where the Wehrl entropy reveals its strength.
First, we discuss the role of Wehrl entropy in the context of entropic uncertainty relations (EURs). The first entropic uncertainty relation was introduced by Białynicki-Birula and Mycielski \cite{Bialynicki-Birula1975} based on previous works by Everett \cite{Everett1957}, Hirschmann \cite{Hirschman1957} and Beckner \cite{Beckner1975}. It is an uncertainty relation for a single pair of conjugate variables, namely position $x$ and momentum $p$, which may be distributed according to the probability density functions $f(x)=\braket{x | \rho | x}$ and $g(p)=\braket{p | \rho | p}$. This corresponds to the so-called homodyne detection protocol, where both variables are recorded separately by measuring in their corresponding eigenbases. Then, the Białynicki-Birula and Mycielski (BBM) entropic uncertainty relation reads \cite{Bialynicki-Birula2011,Coles2017,Hertz2019}
\begin{equation}
h (f) + h (g) \geq \ln e \pi,
\label{eq:BirulaEUR}
\end{equation}
where $h (f)$ is the classical differential entropy associated to the distribution $f(x)$ \cite{Cover2006}
\begin{equation}
h (f) = - \int \text{d} x \, f(x) \ln f(x).
\end{equation}
It should be noted that the bound on the right-hand side of the inequality \eqref{eq:BirulaEUR} is state \textit{independent} whereas the left-hand side increases for mixed states due to concavity. Thus, one can expect the bound to be rather loose for mixed states. Nevertheless, the relation \eqref{eq:BirulaEUR} turns out to be stronger than and implies the well-known formulation of an uncertainty relation in terms of variances \cite{Robertson1929,Kennard1927,Heisenberg1927,Schroedinger1930}, which is discussed in detail for example in ref.\ \cite{Hertz2019}.
Another entropic uncertainty relation for continuous variables was found by Frank and Lieb (FL) \cite{Frank2012} based on previous work by Rumin \cite{Rumin2011}
\begin{equation}
h (f) + h (g) \geq \ln 2 \pi + S(\rho).
\label{eq:FrankEUR}
\end{equation}
Therein the von Neumann entropy $S(\rho)$ accounts for the \textit{mixedness} of the system's state $\rho$. Obviously $\ln 2 \pi$ alone is a less tight bound than the one in \eqref{eq:BirulaEUR}, but for highly mixed states one expects the bound in \eqref{eq:FrankEUR} to be closer to equality.
Also for the Wehrl entropy there exists an entropic uncertainty relation, which is known as the Wehrl-Lieb inequality (WL) \cite{Lieb1978,Carlen1991,Lieb2005,Grabowski1984,Ohya1993}
\begin{equation}
S_{\text{W}} (\rho) \ge 1,
\label{eq:WehrlLiebInequality1}
\end{equation}
where equality only holds for pure coherent state projectors $\rho_\alpha = \ket{\alpha} \bra{\alpha}$.
It was shown that the latter relation does also imply the variance based uncertainty relation by Heisenberg and an entropic uncertainty relation of the form \eqref{eq:BirulaEUR} based on differential entropies of smeared-out distributions (the marginals of the Husimi $Q$-distribution, which are not the true marginals $f$ and $g$) \cite{Grabowski1984}.
One main concern of this work will be to investigate in which regimes the Wehrl-Lieb entropic uncertainty relation \eqref{eq:WehrlLiebInequality1} is tighter (closer to equality) than the relations in terms of differential entropies \eqref{eq:BirulaEUR} and \eqref{eq:FrankEUR}; this will be addressed in \autoref{sec:Monopartite}.
The second question we study is how the Wehrl entropy $S_{\text{W}} (\rho)$ can be used to make statements about entanglement in continuous variable systems. In general, detecting bipartite entanglement for a state $\rho$ which lives in an infinite dimensional Hilbert space $\mathcal{H}$ can be quite challenging \cite{Horodecki2009}. Thus, many criteria are concerned with analyzing the second-order moments \cite{Simon2000,Giovannetti2003,Guehne2004,Hyllus2006}, which is a powerful approach if the state $\rho$ is of Gaussian form (see e.\ g.\ refs. \cite{Weedbrook2012,Serafini2017,Berges2018} for a general discussion of Gaussian states). Incorporating also higher-order moments allows to define a sufficient condition for bipartite entanglement, but can be unmanageable in practice \cite{Schukin2005}.
Another approach is to witness entanglement by violating conditions based on differential entropies \cite{Walborn2009} or differential conditional entropies \cite{Walborn2011,Frank2013,Furrer2014}, which are closely related to entropic uncertainty relations (see also ref.\ \cite{Huang2013,Bergh2021a,Bergh2021b}). They turn out to be successful in detecting many classes of entangled states, especially in the pure state case.
What these approaches have in common is that they rely on a homodyne detection protocol. In \autoref{sec:Bipartite} we will present an approach based on a Wehrl mutual information\footnote{An approach solely based on Wehrl entropy is discussed in ref.\ \cite{Mintert2004}.}, which presumes a heterodyne detection scheme and turns out to witness every pure entangled state and is there furthermore able to bound the entanglement entropy from below. Moreover we discuss the role of the Wehrl conditional entropy, which differs from its fully quantum analog.
\textit{This paper is organized as follows}. We begin introducing coherent states and the Wehrl entropy of an $N$ mode system in \autoref{sec:Basics}. Then, we discuss the relations between the three entropic uncertainty relations \eqref{eq:WehrlLiebInequality1}, \eqref{eq:BirulaEUR} and \eqref{eq:FrankEUR} in \autoref{sec:Monopartite}. In particular, we consider common classes of states, such as number eigenstates or thermal states. This analysis allows us to show that the Wehrl-Lieb inequality is stronger than the relation by Białynicki-Birula and Mycielski almost everywhere in the examined regimes. In \autoref{sec:Bipartite} we consider a bipartite system of $N+M$ modes and construct a Wehrl mutual information from a Wehrl relative entropy, which witnesses every pure entangled bipartite state. The concept is applied to some well-known examples like the two-mode squeezed state or the class of N00N-states. Additionally, we discuss the Wehrl conditional entropy. Finally, we summarize our results and provide an outlook in \autoref{sec:Conclusion}.
\textit{Notation}. In this paper we work with natural units, i.e. $\hbar=k_\text{B}=1$ and we drop operator hats. Rather, we employ capital letters for observables $X$ and small letters for their eigenvalues $x$ and eigenvectors $\ket{x}$. Also, we use the symbol $S$ for the von Neumann entropy, $h$ for classical differential entropies and $S_{\text W}$ for Wehrl entropies with the underlying probability distribution $f(x)$ (or state $\rho$) as argument $h = h(f)$ (or $S = S(\rho)$). We denote conditional entropy by $S(A|B)$ and mutual information by $I(A:B)$.
\section{Coherent states and Wehrl entropy}
\label{sec:Basics}
We start from a set of $N$ continuous quantum variables $X_j$ and conjugate momenta $P_j$ and the (bosonic) commutation relation
\begin{equation}
[X_j, P_k] = i \delta_{jk}.
\label{eq:commuationRelationXP}
\end{equation}
While we use a notation as for positions of particles, the variables $X_j$ could also denote other quantities such as amplitudes of field modes. Note that we make no detailed assumptions about the Lagrangian or Hamiltonian at this point.
We also introduce the annihilation operators
\begin{equation}
a_j = \frac{1}{\sqrt{2}} (X_j + i P_j),
\end{equation}
and creation operators $a_j^\dagger = (a_j)^\dagger$ such that eq.\ \eqref{eq:commuationRelationXP} reads $[a_j, a_k^\dagger] = \delta_{jk}$. Coherent fields are defined as usual as eigenstates of the annihilation operators (see e.\ g.\ \cite{Schleich2001,Mandel2013}),
\begin{equation}
a_j \ket{\vec{\alpha}} = \alpha_j \ket{\vec\alpha},
\end{equation}
with eigenvalues $\alpha_j \in \mathbb{C}$. They can be expressed in terms of number eigenstates (with ground state $| 0 \rangle$)
\begin{equation}
\ket{\vec n} = \prod_{j=1}^N \left\{ \frac{1}{\sqrt{n_j!}}(a_j^\dagger)^{n_j} \right\} |0 \rangle,
\end{equation}
through
\begin{equation}
\ket{\vec \alpha} = \prod_{j=1}^N \left\{ \sum_{n_j=0}^{\infty} e^{- \frac{1}{2} \abs{\alpha_j}^2} \frac{\alpha_j^{n_j}}{\sqrt{n_j!}} \right\} \ket{\vec n}.
\end{equation}
For our purposes, it will often be convenient to parameterize the complex vector of eigenvalues $\vec{\alpha}$ in terms of position $\vec{x}$ and momentum $\vec{p}$,
\begin{equation}
\vec{\alpha} = \frac{1}{\sqrt{2}} \left(\vec{x} + i \vec{p} \right).
\label{eq:PhaseToPositionMomentum}
\end{equation}
Coherent states have many interesting properties. They are not orthogonal,
\begin{equation}
\abs{ \braket{\vec{\alpha} | \vec{\beta}} }^2 = e^{- \abs{\vec{\alpha} - \vec{\beta}}^2},
\end{equation}
and span an (overcomplete) basis,
\begin{equation}
\mathds{1} = \int \frac{\text{d}^{2N} \alpha}{\pi^N} \ket{\vec{\alpha}} \bra{\vec{\alpha}}.
\label{eq:CoherentStatesIdentity}
\end{equation}
Moreover, they minimize the uncertainty principle, which suggests that they can be regarded as a favorable measurement basis for simultaneous measurements of $x$ and $p$. Note that there are other states minimizing the uncertainty principle (squeezed states). Coherent states are the only such states symmetric in the quadratures. In fact, we can construct a positive operator-valued measure (POVM), see e.\ g.\ ref.\ \cite{Holevo2011}, based on pure coherent state projectors
\begin{equation}
E_{\vec{\alpha}} = \ket{\vec{\alpha}} \bra{\vec{\alpha}},
\end{equation}
which resolve the identity according to \eqref{eq:CoherentStatesIdentity}. The probability density for the outcome $\vec\alpha$ is given by
the Husimi $Q$-distribution \cite{Husimi1940,Schleich2001,Mandel2013},
\begin{equation}
Q^{\rho} (\vec{\alpha}) = \text{Tr} \left\{ \rho \, E_{\vec{\alpha}} \right\} = \braket{\vec{\alpha} | \rho | \vec{\alpha}},
\end{equation}
such that the probability to find $\vec \alpha$ in some interval is given by $Q^{\rho} (\vec{\alpha}) d^{2N}\alpha/\pi^N$.
Based on the Husimi $Q$-distribution one can formally define a state after measurement, without recording the outcome,
\begin{equation}
\rho_{\vec{\alpha}} = \int \frac{\text{d}^{2N} \alpha}{\pi^N} Q^{\rho} (\vec{\alpha}) \ket{\vec{\alpha}} \bra{\vec{\alpha}},
\end{equation}
and it is of Glauber $P$-form.
The Husimi $Q$-distribution is a \textit{quasi}-probability distribution\footnote{For general literature on quasi-probability distributions see e.g. refs. \cite{Mandel2013,Zhang1990,Wigner1931,Glauber1963,Sudarshan1963,Lee1995}.} due to the non-orthogonality of coherent states, which causes the Husimi $Q$-distribution to violate Kolmogorov's third axiom, namely $\sigma$-additivity. In other words, two values $Q^{\rho} (\vec \alpha)$ and $Q^{\rho} (\vec \alpha')$ for $\vec \alpha \neq \vec \alpha'$ do not represent probabilities for mutually exclusive events. Nevertheless, $Q^{\rho} (\vec \alpha)$ is normalized to unity in the sense
\begin{equation}
\int \frac{\text{d}^{2N} \alpha}{\pi^N} Q^{\rho} (\vec \alpha) = \text{Tr} \{ \rho\} = 1.
\label{eq:HusimiQNormalization}
\end{equation}
What we have described here is known as the heterodyne detection protocol \cite{Schleich2001}. Intuitively, it corresponds to jointly measuring position $x$ and momentum $p$ with the highest precision allowed by Heisenberg's uncertainty relation.
In contrast to the Wigner $W$-distribution, the Husimi $Q$-distribution is non-negative and bounded \cite{Cartwright1976,Lee1995}
\begin{equation}
0 \le Q^{\rho} (\vec \alpha) \le 1,
\end{equation}
which allows to associate a (differential) entropy to it. This is known as the Wehrl entropy \cite{Wehrl1978,Wehrl1979}
\begin{equation}
\begin{split}
S_\text{W} (\rho) & = - \int \frac{\text{d}^{2N} \alpha}{\pi^N} Q^{\rho} (\vec \alpha) \ln Q^{\rho} (\vec \alpha) \\
& = - \int \frac{\text{d}^N x\, \text{d}^N p}{(2\pi)^N} Q^\rho(\vec x, \vec p) \ln Q^{\rho} (\vec x, \vec p).
\end{split}
\label{eq:WehrlEntropy}
\end{equation}
In the second equation we used the parametrization introduced in eq.\ \eqref{eq:PhaseToPositionMomentum}.
The Wehrl entropy can be seen as a quasi-classical differential entropy in quantum mechanical phase space. It is bounded from below by the true von Neumann entropy (as expected for a coarse-grained notion of an entropy), but also bounded from above by the von Neumann entropy of the measured state\footnote{The second statement can be proven using the well-known Berezin-Lieb inequality \cite{Berezin1972}, see e.g. ref.\ \cite{Wehrl1979}.}
\begin{equation}
S(\rho) \le S_{\text{W}} (\rho) \le S(\rho_{\vec \alpha}).
\end{equation}
Furthermore, it is uniquely minimized by any pure coherent state projector $\rho = \ket{\vec \alpha} \bra{\vec \alpha}$, in which case its value is $S_{\text{W}} = N$. This result is known as the Wehrl-Lieb inequality \cite{Lieb1978,Carlen1991,Lieb2014},
\begin{equation}
S_{\text{W}} (\rho) \ge S_\text{W} \left(\ket{\vec \alpha}\bra{\vec \alpha} \right) = N.
\label{eq:WehrlLiebInequality}
\end{equation}
Rather recently, improvements of this bound were discussed in refs.\ \cite{dePalma2018a,dePalma2018b}.
Another property of the Wehrl entropy $S_{\text{W}} (\rho)$ is important when it comes to entanglement. In contrast to the von Neumann entropy $S(\rho)$, it is monotonous under partial trace, i.e. if we consider a bipartite quantum system $\mathcal{H} = \mathcal{H}_A \otimes \mathcal{H}_B$ with the reduced density operator $\rho_A = \text{Tr}_B \{ \rho \}$, we have
\begin{equation}
S_{\text{W}} (\rho_A) \le S_{\text{W}} (\rho).
\label{eq:WehrlMonotonicity}
\end{equation}
This highlights the classical nature of Wehrl entropy. The analogous statement is oftentimes not true for the von Neumann entropy $S(\rho)$ due to entanglement \cite{Wehrl1978,Wehrl1979}. This already indicates that capturing entanglement with a Wehrl entropy may require a different line of reasoning compared to a von Neumann entropy.
\section{Monopartite quantum system}
\label{sec:Monopartite}
At first we consider a monopartite quantum system. In the following, we want to compare the differential entropies of the two phase space distributions $Q^\rho(x,p)/(2\pi)$ and $f(x)g(p)$ associated with the heterodyne and homodyne measurement protocol respectively. Let us first state the following relations
\begin{equation}
\begin{split}
h\left(\frac{Q^\rho}{2\pi}\right) &= -\int \text d x\, \text d p \frac{Q^\rho(x,p)}{2\pi} \ln \left(\frac{Q^\rho(x,p)}{2\pi}\right) \\
&= S_{\text W}(\rho) + \ln (2\pi), \\
h(fg) &= \int \text d x\, \text d p f(x) g(p) \ln(f(x)g(p)) \\
&= h(f) + h(g).
\end{split}
\end{equation}
The two differential entropies are related up to additive constants with $S_{\text W}(\rho)$ and $h(f) + h(g)$, for which we know entropic uncertainty relations.
Henceforth, our goal is to compare the uncertainty deficit, i.e. the distance to the bound of an entropic uncertainty relation, of these relations as an estimator for the strength of the statement. We refer to a lower uncertainty deficit as the \textit{tightness} of an uncertainty relation. It is clear that being close to equality is desirable for inequalities, however it should be noted here that this does not immediately lead to a formal implication between different uncertainty relations. In fact, the Wehrl-Lieb inequality is tight on a different set of states than the inequality by Białynicki-Birula and Mycielski.
To properly compare them, we rearrange the three inequalities such that their right-hand sides coincide. Thus, we have the following three relations
\begin{equation}
\begin{rcases}
\text{WL} &S_{\text{W}}(\rho) + \ln \pi \\
\text{BBM} &h(f) + h(g) \\
\text{FL} &h(f) + h(g) - S(\rho) + \ln e/2
\end{rcases}
\ge \ln e \pi.
\label{eq:EURSummary}
\end{equation}
We begin our analysis with pure number eigenstates, then we investigate a mixture of the lowest two number eigenstates and finally consider thermal states. This analysis will allow us to state a conjecture about their relations for a general state $\rho$. If possible, results will be given analytically. Furthermore, the Husimi $Q$-distribution will be written down in terms of position $x$ and momentum $p$, i.e. using the parametrization \eqref{eq:PhaseToPositionMomentum}, in order to make the different (quasi)-probability distributions comparable.
\subsection{Number eigenstates}
\label{sec:NumberEigenstates}
The density matrix of a pure number eigenstate
\begin{equation}
\rho_n = \ket{n} \bra{n},
\end{equation}
has vanishing quantum entropy, $S(\rho_n) = 0$. As a consequence, the Frank-Lieb bound will be less tight than the Białynicki-Birula and Mycielski bound.
Furthermore, the probability density functions $f(x)$ and $g(p)$ are of equal shape,
\begin{equation}
\begin{split}
f_n (x) &= \abs{\psi_n (x)}^2 = \frac{1}{\sqrt{\pi} \, 2^n n!} \, H_n^2 (x) \, e^{-x^2},
\label{eq:ProbabilitesFockStates}
\end{split}
\end{equation}
where $H_n (x)$ are the commonly-known physicist's Hermite polynomials. The Husimi $Q$-distribution is given by
\begin{equation}
\begin{split}
Q_n(x,p) &= \abs{\braket{n | \alpha}}^2 \\
&= \frac{1}{2^n n!} \, \left( x^2 + p^2\right)^n \, e^{-\frac{1}{2}(x^2+p^2)}.
\end{split}
\label{eq:HusimiQFockStates}
\end{equation}
From these expressions all appearing entropies can be calculated in the following.
The differential entropies $h(f)$ and $h(g)$ appearing in the Białynicki-Birula and Mycielski as well as the Frank-Lieb uncertainty relation were investigated in refs.\ \cite{Dehesa1994,Majernik1996,Ruiz1997,Toranzo2019}. The main challenge in this computation is due to the Hermite polynomials $H_n (x)$. In principle the differential entropy is given by
\begin{equation}
h(f_n) = \ln \left(\sqrt{\pi} \, 2^n n! \right) + n + \frac{1}{2} + \frac{1}{\sqrt{\pi} \, 2^n n!} E (H_n),
\end{equation}
where $E (H_n)$ is the so-called entropy of Hermite polynomial
\begin{equation}
E (H_n) = -\int_{- \infty}^{+ \infty} d x \, H_n^2(x) \ln \left( H_n^2 (x) \right) e^{-x^2}.
\end{equation}
After a tedious and lengthy calculation one finds \cite{Ruiz1997,Toranzo2019}
\begin{equation}
\begin{split}
h(f_n) =& \ln \left(\sqrt{\pi} \, 2^n n! \right) + n + \frac{1}{2} + n \gamma \\
& - 2 \sum_{i=1}^n x_{n,i}^2 \, {}_2 F_2 \left(1,1; \frac{3}{2}, 2; -x_{n,i}^2 \right) \\
& + \sum_{k=1}^n \binom{n}{k} \frac{(-1)^k 2^k}{k} \sum_{i=1}^n {}_1 F_1 \left(k, \frac{1}{2}, - x_{n,i}^2 \right).
\end{split}
\end{equation}
Therein ${}_2 F_2 (a,b;c,d;z)$ denotes a generalized hypergeometric function, ${}_1 F_1 (a,b;z)$ is Kummer's confluent hypergeometric function for $z \in \mathbb{C}$ and $x_{n,i}$ are the roots of the $n$-th Hermite polynomial $H_n (x^2)$. Also, we have introduced the Euler-Mascheroni constant $\gamma \approx 0.577$.
The latter expression can be evaluated numerically for all $n \in \mathbb{N}$. Furthermore, for $n \gg 1$ one can give a simple approximate formula for the differential entropy \cite{Ruiz1997}
\begin{equation}
h (f_n) \approx \frac{1}{2} \left(-2 + \ln 2 \pi^2 n \right).
\label{eq:FockDifferentialEntropyApprox}
\end{equation}
In contrast, the Wehrl entropy can be computed analytically due to the simple form of the Husimi $Q$-distribution for number eigenstates \eqref{eq:HusimiQFockStates}. Although this is a rather technical issue, one might appreciate the fact that computing a Wehrl entropy is much easier than computing a differential entropy in many cases.
\begin{figure*}[t!]
\centering
\includegraphics[height=0.23\textwidth]{figures/QHOEURPureFockStates.pdf}
\includegraphics[height=0.22\textwidth]{figures/QHOEURPureFockStatesLargen.pdf}
\caption{Left panel: Values of the three entropic uncertainty relations (EURs, cf. \eqref{eq:EURSummary}) as a function of the energy level $n$. One can see that the Białynicki-Birula and Mycielski (BBM) as well as the Wehrl-Lieb (WL) inequality are equivalent and tight for the ground state $n=0$. For $n=1$ we have $S_{\text{W}} (\rho_1) + \ln \pi \approx 2.72 > h (f_1) + h(g_1) \approx 2.69$, whereas for $n=2$ we get $S_{\text{W}} (\rho_2) + \ln \pi \approx 2.992 < h (f_2) + h(g_2) \approx 2.997$. Thus, the Wehrl-Lieb uncertainty relation is the strongest for $n \in \mathbb{N} \setminus \{1 \}$, i.e. almost everywhere. For pure states the bound of the Frank-Lieb (FL) relation becomes state-independent. In that case it is less tight than the relation by Białynicki-Birula and Mycielski. Right panel: Behavior of the Wehrl-Lieb and the Białynicki-Birula and Mycielski uncertainty relations for large $n$ (the latter was investigated numerically in \cite{Majernik1996}). The result for the Wehrl entropy is already valid for small $n$, whereas the result for the differential entropies is only valid for $n \gg 1$. This is again due to the fact that the Wehrl entropy is of such simple form.}
\label{fig:QHOEURPureStates}
\end{figure*}
Starting from \eqref{eq:HusimiQFockStates}, one can compute the Wehrl entropy for number eigenstates. After a straight forward exercise in Gaussian integration one finds (see also ref.\ \cite{Orlowski1999})
\begin{equation}
S_{\text{W}} (\rho_n) = \ln{n!} + n + 1 + n\gamma - n \eta_n,
\label{eq:FockWehrlEntr2}
\end{equation}
where $\eta_n$ is the $n$-th harmonic number
\begin{equation}
\eta_n = \sum_{k=1}^{n} \frac{1}{k},
\end{equation}
with $\eta_0 = 0$.
The result can be further simplified when considering large $n \gg 1$ and using the Stirling approximation formula
\begin{equation}
n! \approx \sqrt{2\pi n}\left(\frac{n}{e}\right)^n,
\end{equation}
which leads to
\begin{equation}
S_{\text{W}}(\rho_n) \approx \frac{1}{2} \left( 1 + \ln{2\pi n} \right), \label{eq:FockWehrlApprox}
\end{equation}
for $n \gg 1$.
For the sake of a visual comparison we present all three bounds as a function of $n$ in the left panel of \autoref{fig:QHOEURPureStates}. The main result is that the Wehrl-Lieb relation is tightest for $n>1$. Moreover, we obtain the expected result for $n=0$, i.e. the Wehrl-Lieb as well as the Białynicki-Birula and Mycielski relations are tight and equivalent. In general, the Frank-Lieb relation is worst since we consider pure states.
Furthermore, it is interesting to compare the Wehrl-Lieb relation with the Białynicki-Birula and Mycielski relation for large $n \gg 1$ (cf. right panel of \autoref{fig:QHOEURPureStates}). From the asymptotic expressions \eqref{eq:FockDifferentialEntropyApprox} and \eqref{eq:FockWehrlApprox} we can read off the scaling behaviors for large $n \gg 1$, which gives $h(f_n) + h(g_n) \propto \ln n$ and $S_{\text{W}} (\rho_n) \propto 1/2 \, \ln n$. Therefore, the Wehrl-Lieb relation grows much slower than the Białynicki-Birula and Mycielski relation. Also, we observe that the Stirling approximation \eqref{eq:FockWehrlApprox} is close to the true result \eqref{eq:FockWehrlEntr2} also for small $n \sim 1$, while $\eqref{eq:FockDifferentialEntropyApprox}$ becomes a reasonable approximation only for $n \gg 1$.
\subsection{Mixture of number eigenstates}
Next, we analyze some particular mixtures of number eigenstates. For simplicity we concentrate on a mixed state of the first two number eigenstates, as well as a thermal state. We will again compare different entropic uncertainty relations and investigate how close they are to being violated.
\subsubsection{Mixture of first two number eigenstates}
A relatively simple mixed state is the following superposition of the first two number eigenstates,
\begin{equation}
\rho_{0 1} = q \ket{0} \bra{0} + (1-q) \ket{1} \bra{1}.\label{eq:mixure01}
\end{equation}
The resulting entropies involve now a sum inside a logarithm. We use numerical methods to solve the integrals. The resulting curves for the entropic uncertainty relations are shown in \autoref{fig:QHOEURMixedStates} (left panel). Therein, one can see that the Wehrl-Lieb relation is tighter than the Białynicki-Birula and Mycielski relation except in a small region around $q=0$, where \eqref{eq:mixure01} corresponds to the first excited state. Thus we observe a similar situation as for pure states. Also, the Wehrl-Lieb relation turns out to be less concave than that of Białynicki-Birula and Mycielski. In contrast, the Frank-Lieb relation becomes tightest for more mixed states.
\begin{figure*}[t!]
\centering
\includegraphics[height=0.22\textwidth]{figures/QHOEURMixedFockStates01.pdf}
\includegraphics[height=0.22\textwidth]{figures/QHOEURThermalState.pdf}
\caption{Left panel: Entropic uncertainty relations (EURs, cf. \eqref{eq:EURSummary}) for a mixture of the ground state with probability $q$ and the first excited state with probability $1-q$. The Wehrl entropy decreases monotonically for larger $q$, i.e. if the ground state contributes more to the total state. Right panel: Entropic uncertainty relations for the thermal state. The Wehrl-Lieb relations is again tighter than the Białynicki-Birula and Mycielski relation and both become tight for zero temperature $\beta \to \infty$. In contrast the Frank-Lieb (FL) relation becomes tight in the infinite temperature limit $\beta \to 0$, but is less tight for lower temperatures.}
\label{fig:QHOEURMixedStates}
\end{figure*}
\subsubsection{Thermal state}
At last, we consider a thermal state $\rho_T$ under the assumption that the Hamiltonian is given by $H=\omega(n+1/2)$. The thermal state is then the state with maximal entropy for a fixed energy expectation value \cite{Jaynes1957,Jaynes19572,Jaynes1968,Landau1980}. It reads
\begin{equation}
\rho_T = \frac{1}{Z} e^{- \beta H} = \sum_{n=0}^{\infty} p_n \ket{n} \bra{n},
\end{equation}
where $p_n$ is the Boltzmann distribution
\begin{equation}
p_n = \frac{1}{Z} e^{- \beta \omega (n + \frac{1}{2})},
\end{equation}
and
\begin{equation}
Z = \sum_{n=0}^{\infty} e^{- \beta \omega (n + \frac{1}{2})} = \frac{e^{- \beta \omega/2}}{1-e^{-\beta \omega}} = \frac{1}{2 \sinh(\beta \omega / 2)},
\end{equation}
is the canonical partition function. To calculate the probability density $f_T (x)$ we make use of Mehler's formula \cite{Mehler1866}, which leads to
\begin{equation}
f_T (x) = \frac{1}{Z} \sum_{n=0}^{\infty} \abs{\psi_n (x)}^2 e^{- \beta \omega (n + \frac{1}{2})} = \frac{1}{\sqrt{2 \pi \sigma^2}} e^{- \frac{x^2}{2 \sigma ^2}},
\end{equation}
where we used the abbreviation
\begin{equation}
\frac{1}{2 \sigma^2} = \tanh \left( \frac{\beta \omega}{2} \right).
\end{equation}
Thus, the distribution $f_T (x)$ is a Gaussian centered around $x=0$ with variance $\sigma^2$, which depends non-linearly on the inverse temperature $\beta$.
The associated differential entropy is
\begin{equation}
h (f) = \ln \left(\sqrt{2 \pi e \sigma^2} \right).
\end{equation}
A similar consideration for the momentum distribution leads to the following expression for the uncertainty relation of Białynicki-Birula and Mycielski
\begin{equation}
h (f_T) + h (g_T) = 1 + \ln \pi - \ln \left( \tanh \left( \frac{\beta \omega}{2} \right) \right).
\end{equation}
We see that for $\beta \to \infty$, i.\ e.\ in the zero temperature limit, the inequality becomes tight, as expected based on the discussion in section \ref{sec:NumberEigenstates}.
Since thermal states are highly mixed we should have a closer look at the Frank-Lieb relation. Additionally to the computation of the differential entropies, we need to calculate the von Neumann entropy of the thermal state $\rho_T$. It is given by
\begin{equation}
S(\rho_T) = - \ln(1 - e^{- \beta \omega}) + \frac{\beta \omega}{e^{\beta \omega} - 1}.
\end{equation}
Thus, the left-hand side of the Frank-Lieb relation reads (again we rearrange everything such that the right-hand side equals $\ln e \pi$)
\begin{equation}
\begin{split}
&h (f_T) + h (g_T) - S(\rho_T) + 1 - \ln 2\\
&= 2 + \ln \left( \frac{\pi}{2} \frac{1 - e^{- \beta \omega}}{\tanh( \beta \omega / 2)} \right) - \frac{\beta \omega}{e^{\beta \omega} - 1}.
\end{split}
\end{equation}
Also, we need to calculate the Husimi $Q$-distribution, which yields
\begin{equation}
\begin{split}
Q_T (x,p) &= \frac{1}{Z} \sum_{n=0}^{\infty} Q_n (x,p) e^{- \beta E_n}\\
&= \frac{1}{Z} e^{- \frac{1}{2} (x^2 + p^2) - \frac{\beta \omega}{2}} \sum_{n=0}^{\infty} \frac{1}{2^n n!} (x^2 + p^2)^n e^{- \beta \omega n} \\
&= \frac{1}{Z} e^{- \frac{1}{2} (x^2 + p^2) \left(1-e^{- \beta \omega} \right) - \frac{\beta \omega}{2}}.
\end{split}
\end{equation}
For the left-hand side of the Wehrl-Lieb uncertainty relation we then find
\begin{equation}
S_{\text{W}} (\rho_T) + \ln \pi = 1 + \frac{\beta \omega}{2} + \ln \left(\frac{\pi}{2} \text{csch} (\beta \omega/2) \right).
\end{equation}
We show a comparison of the three entropic uncertainty relations in \autoref{fig:QHOEURMixedStates} (right panel). The Wehrl-Lieb relation is again stronger than the Białynicki-Birula and Mycielski relation. This is due to the fact that for a thermal state the ground state is populated most and higher energy eigenstates are suppressed exponentially. Moreover, both become tight for the temperature tending to zero, which corresponds to the system approaching the ground state. Interestingly, the relation by Frank and Lieb behaves differently, which was already noted by the authors themselves \cite{Frank2012,Coles2017}. The uncertainty relation becomes tight in the \textit{infinite} temperature limit in contrast to the other two inequalities and it is more loose for low temperatures.
\subsubsection{General comment}
What we have observed is that the Frank-Lieb relation \eqref{eq:FrankEUR} is tightest for highly mixed states, while the Wehrl-Lieb relation \eqref{eq:WehrlLiebInequality} is tightest in the number eigenstate regime (except for $n=1$). Furthermore, we have seen that the Wehrl-Lieb relation is less concave under mixing than the Białynicki-Birula and Mycielski relation (cf. \autoref{fig:QHOEURMixedStates}). Thus, we can conjecture that the Wehrl-Lieb entropic uncertainty relation is tighter than the Białynicki-Birula and Mycielski entropic uncertainty relation in most situations. Furthermore, we have seen that the Wehrl entropy is very often easy to compute, especially compared to differential entropies of marginal distributions.
\section{Bipartite quantum system}
\label{sec:Bipartite}
In the second part of this work we consider a bipartite quantum system $AB$ with partitions $A$ and $B$ of $N$ and $M$ modes, respectively. Moreover, we associate the sets of coherent states $\ket{\vec \alpha}$ to $A$, $\ket{\vec \beta}$ to B and $\ket{\vec \alpha \vec \beta} = \ket{\vec \alpha} \otimes \ket{\vec \beta}$ to $AB$. Then, the local Husimi $Q$-distribution $Q^{\rho}_{A} (\vec \alpha)$ can be obtained from the global one $Q^\rho (\vec \alpha, \vec \beta)$ by integrating out system $B$
\begin{equation}
Q^{\rho}_{A} (\vec \alpha) = \int \frac{\text{d}^{2M} \beta}{\pi^M} Q^{\rho} (\vec \alpha, \vec \beta).
\end{equation}
Note that this agrees with the Husimi $Q$-distribution obtained from the reduced density matrix $\rho_A = \text{Tr}_B\{\rho\}$. Similarly, one can define a local Husimi $Q$-distribution $Q^{\rho}_{B} (\vec \beta)$ for subsystem $B$.
It is also useful to consider a situation where a heterodyne measurement is performed first only on subsystem $B$. Assuming that the outcome is $\vec\beta$ leads to a Husimi $Q$-distribution for subsystem $A$ that we call the \textit{conditional} Husimi $Q$-distribution,
\begin{equation}
\begin{split}
Q^{\rho}(\vec \alpha | \vec \beta) = & \text{Tr} \left\{\frac{\braket{\vec \beta | \rho | \vec \beta}}{\braket{\vec \beta | \rho_B | \vec \beta}} E_{\vec \alpha} \right\} = \frac{\langle \vec \alpha \vec\beta | \rho | \vec\alpha \vec\beta \rangle }{\langle \vec \beta |\rho_B| \vec \beta \rangle} \\
= & \frac{Q^{\rho}(\vec \alpha, \vec \beta) }{Q^{\rho}_{B} (\vec \beta)}.
\end{split}
\label{eq:ConditionalHusimiDef}
\end{equation}
Note that these relations are in complete analogy to classical probability densities.
\subsection{Wehrl relative entropy}
In the following we will consider different quantities used in classical and quantum information theory for Husimi $Q$-distributions defined in analogy to the Wehrl entropy $S_{\text{W}} (\rho)$. We begin with defining the Wehrl relative entropy, which will also be helpful to define a Wehrl conditional entropy and Wehrl mutual information. For any state $\rho$ and some model state $\sigma$ on an $N$-mode system it reads (in complete analogy to the standard definition by Kullback and Leibler \cite{Kullback1951,Kullback1959})
\begin{equation}
\begin{split}
S_{\text{W}} (\rho \| \sigma)
= \int \frac{\text{d}^{2N} \alpha}{\pi^N} Q^\rho (\vec \alpha) (\ln Q^{\rho} (\vec \alpha) - \ln Q^\sigma (\vec \alpha)).\label{eq:generalWehrlRelativeEntropy}
\end{split}
\end{equation}
The usual support condition for the states $\text{supp}(\rho) \subseteq \text{supp} (\sigma)$ then translates into the support condition for the Husimi $Q$-distributions $\text{supp}(Q^{\rho}) \subseteq \text{supp} (Q^{\sigma})$. If this condition is violated we set $S_{\text{W}} (\rho \| \sigma) = + \infty$. Intuitively, \eqref{eq:generalWehrlRelativeEntropy} can be interpreted as a measure for the \textit{distinguishability} of the true Husimi $Q$-distribution $Q^{\rho}$ and some model distribution $Q^{\sigma}$. Furthermore, it is a non-negative quantity being zero if and only if the two distributions agree\footnote{See appendix \ref{app:WehrlRelativeEntropy} for a simple proof.}. However, it should be emphasized that the most important property of the quantum relative entropy, namely its monotonicity under quantum channels, does \textit{not} hold for the Wehrl relative entropy. This is due to the fact the Wehrl entropy is not invariant under general unitary transformations. In fact, the Wehrl entropy is only invariant for transformations which map coherent states to coherent states \cite{Wehrl1979}.
\subsection{Wehrl conditional entropy}
Let us recall first some properties of quantum conditional entropies.
In the finite dimensional case there exist several ways of defining this quantity, which are all equivalent. For an infinite dimensional Hilbert space this is not necessarily the case. An adequate definition for both cases comprises the quantum relative entropy and requires $S(\rho_A) < \infty$ (see ref.\ \cite{Kuznetsova2011}). Then, we can define the quantum conditional entropy as
\begin{equation}
S^\rho (A | B) = S (\rho_A) - S(\rho \| \rho_A \otimes \rho_B),
\label{eq:QuantumConditionalEntropy}
\end{equation}
which reduces to
\begin{equation}
S^\rho (A | B) = S (\rho) - S (\rho_B),
\end{equation}
provided that the von Neumann entropy of subsystem $B$ is also finite $S(\rho_B) < \infty$.
The quantum conditional entropy $S^\rho (A | B)$ allows for a simple separability criterion. In particular, one can easily prove that all separable states\footnote{Note that for an infinite dimensional Hilbert space a separable state may also be written as a convex integral over pure product states. This is particularly interesting for mixed Gaussian states.}
\begin{equation}
\rho = \sum_i p_i \, \rho^i = \sum_i p_i \left(\rho_A^i \otimes \rho_B^i \right),
\label{eq:SeparableState}
\end{equation}
where $p_i \ge 0$ and $\sum_i p_i = 1$ is a probability distribution and $\rho_A^i$ and $\rho_B^i$ can be taken to be pure, fulfill
\begin{equation}
S^\rho (A | B) \ge 0.
\label{eq:QuantumConditionalEntropySeparable}
\end{equation}
Thus, the violation of the latter inequality is a sufficient condition for entanglement \cite{Cerf1997,Cerf1999}. One may easily construct examples where this bound is violated, e.\ g.\ a pure entangled state.
In complete analogy to \eqref{eq:QuantumConditionalEntropy} we define the Wehrl conditional entropy for $S_{\text{W}} (\rho_A) < \infty$ as
\begin{equation}
S^\rho_{\text{W}} (A | B) = S_{\text{W}} (\rho_A) - S_{\text{W}} (\rho \| \rho_A \otimes \rho_B).
\label{eq:WehrlConditionalEntropy}
\end{equation}
Analogously, we get
\begin{equation}
S^\rho_{\text{W}} (A | B) = S_{\text{W}} (\rho) - S_{\text{W}} (\rho_B)
\end{equation}
if $S_{\text{W}} (\rho_B) < \infty$.
Note that one may also write this with the conditional Husimi $Q$-distribution in \eqref{eq:ConditionalHusimiDef} as
\begin{equation}
\begin{split}
S^\rho_\text{W}(A | B) = & - \int \frac{d^{2M}\beta}{\pi^M} Q^\rho_{B}(\vec \beta) \\ & \times \int \frac{d^{2N}\alpha}{\pi^N} Q^\rho(\vec \alpha | \vec \beta ) \ln Q^\rho(\vec \alpha | \vec \beta),
\end{split}
\end{equation}
similar as for a classical conditional entropy. Because the conditional Husimi $Q$-distribution is itself a Husimi $Q$-distribution, the Wehrl-Lieb inequality implies
\begin{equation}
S^\rho_\text{W}(A | B) \geq N,
\label{eq:ConditionalWL}
\end{equation}
which is a refined version of the monotonocity property \eqref{eq:WehrlMonotonicity} since it can be reformulated as
\begin{equation}
S_{\text{W}} (\rho_B) + N \le S_{\text{W}} (\rho).
\label{eq:WehrlRefinedMonotonicity}
\end{equation}
Also, the relation \eqref{eq:ConditionalWL} can be interpreted as an entropic uncertainty relation in the presence of classical memory (see \cite{dePalma2018c} for an entropic uncertainty relation with quantum memory). In contrast to the relation discussed in ref.\ \cite{dePalma2018c}, where the analysis was restricted to quasi-classical states and corresponding entropies, \eqref{eq:ConditionalWL} shows that a fully classical Wehrl conditional entropy can not serve as an entanglement witness analogous to eq.\ \eqref{eq:QuantumConditionalEntropySeparable}. In this sense, it behaves like a classical conditional entropy.
\subsection{Wehrl mutual information}
To quantify the correlations between the two subsystems $A$ and $B$ we introduce the Wehrl mutual information
\begin{equation}
I^\rho_{\text{W}} (A : B) = S^\rho_{\text{W}} (\rho \| \rho_A \otimes \rho_B).
\label{eq:WehrlMutualInformation}
\end{equation}
Most importantly, its non-negativity is inherited from the Wehrl relative entropy \eqref{eq:generalWehrlRelativeEntropy}. Also, it is zero if and only if the state $\rho$ is a product state. We can express \eqref{eq:WehrlMutualInformation} in terms of other Wehrl entropies provided that they are finite
\begin{equation}
\begin{split}
I^\rho_{\text{W}} (A : B) &= S_{\text{W}} (\rho_A) + S_{\text{W}} (\rho_B) - S_{\text{W}} (\rho) \\
&= S_{\text{W}} (\rho_A) - S^\rho_{\text{W}} (A | B).
\end{split}
\end{equation}
In general, a mutual information can not distinguish between classical and quantum correlations. Thus, let us now consider the case where some state $\rho$ is known to be pure. In this case the Wehrl mutual information $I^\rho_{\text{W}} (A:B) $ quantifies only the quantum correlations between $A$ and $B$. Moreover, it is zero if and only if the state is a product state $\rho = \rho_A \otimes \rho_B$, which follows directly from the according property of the Wehrl relative entropy. Therefore, we can conclude that the Wehrl mutual information $I^\rho_{\text{W}} (A : B) $ is a perfect witness for pure state entanglement.
It should be noted that the positive partial transpose (PPT) criterion already provides a necessary and sufficient condition for pure state entanglement (see e.g. \cite{Horodecki2009}), but requires the knowledge of the state $\rho$, which is typically inaccessible in experiments. Therefore, the advantage of the Wehrl mutual information $I^\rho_{\text{W}} (A : B) $ lies in its measurability and in this sense it can be regarded as a measurable equivalent to the PPT criterion for pure states. Also, one should note that not every classical mutual information is a perfect witness for pure state entanglement. For example, any marginal mutual information does not contain information about correlations in or with other variables.
Additionally, it is interesting to check whether $I^\rho_{\text{W}} (A:B) $ fulfills the requirements for an entanglement \textit{measure} or an entanglement \textit{monotone}. Since the Wehrl mutual information does not reduce to the entanglement entropy, it is not an entanglement measure. Also, the Wehrl mutual information does \textit{not} fulfill a minimum requirement, namely invariance under local unitaries
\begin{equation}
I^{\rho^\prime}_{\text{W}} (A:B) \neq I^\rho_{\text{W}} (A : B),
\end{equation}
with $\rho^\prime = (U_A \otimes U_B) \, \rho\, (U_A^\dag \otimes U_B^\dag)$ and $U_{A,B}$ unitary. A simple counterexample is given by $U_A = \mathds{1}$ and $U_B = S(\kappa)$, where $S(\kappa)$ is the squeezing operator on subsystem $B$. Therefore, the Wehrl mutual information should be considered as a witness for pure state entanglement and not as an entanglement monotone.
Nevertheless, there is a non-trivial relation between the Wehrl mutual information and the quantum mutual information. Following \cite{Lieb2005}, we have the relation
\begin{equation}
I^\rho_{\text{W}} (A : B) \le I^\rho (A : B).
\label{eq:MutualInformationBound}
\end{equation}
For globally pure states the latter relation reduces to
\begin{equation}
\frac{1}{2} I^\rho_{\text{W}} (A:B) \le S(\rho_A) = S(\rho_B),
\end{equation}
which is an interesting and useful relation when it comes to quantifying entanglement in an actual experiment. If the Husimi $Q$-distribution of a pure state $\rho$ is measured, one can not only decide if the state $\rho$ is entangled or not, one can also give a lower bound on the entanglement entropy $S(\rho_A)$. Thus, the measurable Wehrl mutual information $I^\rho_{\text{W}} (A : B)$ allows to make a statement about how much two subsystems are entangled \textit{at least}. If the global state $\rho$ is not pure, one can still make a statement about how much the two subsystems are correlated at least.
It should be noted that the left-hand side of the relation \eqref{eq:MutualInformationBound} is not restricted to the Wehrl mutual information, but holds true for all classical mutual informations constructed from positive operator-valued measures in subsystems $A$ and $B$ (cf. corollary 2 in ref.\ \cite{Lieb2005}).
Also, we want to point out that the relation \eqref{eq:MutualInformationBound} is similar to an open conjecture stated in ref.\ \cite{Schneeloch2014}, which is based on a homodyne detection protocol. It reads
\begin{equation}
I^{\rho} (f_A : f_B) + I^{\rho} (g_A : g_B) \overset{?}{\le} I^{\rho} (A:B),
\end{equation}
where $f_A = f(x_A), f_B=f(x_B)$ are probability densities of the positions and $g_A = g(p_A)$, $g_B=g(p_B)$ are the momentum distributions measured in $A$ or $B$, respectively. Note that the measured mutual informations in the latter relation do not contain any information about correlations between positions and momenta, which are naturally included in the Wehrl mutual information $I^\rho_{\text{W}} (A : B) $.
\subsection{Entanglement of common pure states}
\label{sec:EntanglementExamples}
In the following, we compute the Wehrl mutual information $I^\rho_{\text{W}} (A : B)$ and the Wehrl conditional entropy $S^\rho_{\text{W}} (A | B)$ for some well-known pure states. To that end we use the parametrization \eqref{eq:PhaseToPositionMomentum} for each subsystem.
\subsubsection{Gaussian states}
We begin with a general Gaussian state $\rho$. Such a state is characterized by a global Husimi $Q$-distribution of the form\footnote{Without loss of generality we assume that the expectation values of the quadratures vanish.}
\begin{equation}
Q^{\rho} (\vec{r}) = \frac{1}{Z} \exp \left(- \frac{1}{2} \, \vec{r}^T \cdot C \cdot \vec{r} \right),
\label{eq:HusimiQGauss}
\end{equation}
where the bilinear form in the exponent contains all four phase space vectors $\vec x_A, \vec p_A, \vec x_B, \vec p_B$ combined into a single vector
\begin{equation}
\vec{r} = \left(\vec{r}_A, \vec{r}_B \right)^T = \left(\vec x_A, \vec p_A, \vec x_B, \vec p_B \right)^T,
\label{eq:rVector}
\end{equation}
and the symmetric matrix $C$ has the form
\begin{equation}
C =
\begin{pmatrix}
C_A & C_M \\
C_M^T & C_B
\end{pmatrix}.
\label{eq:MMatrix}
\end{equation}
Note that $C$ is related to the covariance $V$ in symplectic phase space,
\begin{equation}
V_{i j} = \frac{1}{2} \text{Tr} \left\{\rho \left(r_i r_j + r_j r_i \right) \right\},
\end{equation}
by
\begin{equation}
C = \left(V + \frac{1}{2} \, \mathds{1} \right)^{-1}.
\end{equation}
This follows from convolving the corresponding Gaussian Wigner $W$-distribution with a Gaussian distribution of covariance $(1/2) \mathds{1}$.
Also, $Z$ is a normalization constant, such that \eqref{eq:HusimiQNormalization} is fulfilled. Then, the expression \eqref{eq:HusimiQGauss} can be rewritten as
\begin{equation}
\begin{split}
&Q^{\rho} (\vec{r}_A, \vec{r}_B)\\
&= \frac{1}{Z} \exp \left(- \frac{1}{2} \vec{r}_A^T C_A \vec{r}_A - \frac{1}{2} \vec{r}_B^T C_B \vec{r}_B - \vec{r}_A^T C_M \vec{r}_B \right),
\end{split}
\end{equation}
where the correlations are encoded in the mixing term $\vec{r}_A^T C_M \vec{r}_B$. If this term would be absent, the Husimi $Q$-distribution for the joint system $A B$ would factorize into the individual Husimi $Q$-distributions.
\begin{figure*}[t!]
\centering
\includegraphics[height=0.20\textwidth]{figures/QHOWehrlMutualInformationTwoModeSqueezed.pdf}
\includegraphics[height=0.20\textwidth]{figures/QHOWehrlMutualInformationNOON.pdf}
\caption{Left panel: Wehrl mutual information $I^\rho_{\text{W}} (A : B)$ and Wehrl conditional entropy $S^\rho_{\text{W}} (A | B)$ together with the quantum mutual information $I^\rho (A : B)$ of the two-mode squeezed state as a function of the squeezing parameter $\lambda$. For $\lambda = 0$ one recovers the product state of two vacua, while for $\lambda \to 1$ the state approaches the fully correlated Einstein-Podolsky-Rosen (EPR) state. Right panel: Wehrl mutual information $I^\rho_{\text{W}} (A : B)$, Wehrl conditional entropy $S^\rho_{\text{W}} (A | B)$ and quantum mutual information $I^\rho (A : B)$ for the N00N states up to $N=10$ (the curves represent interpolations). Higher N00N states have larger Wehrl mutual information, although the value for $N=1$ is slightly smaller than for $N=2$.}
\label{fig:QHOWehrlMutualInformation}
\end{figure*}
Performing a Gaussian integration, we arrive at the joint Wehrl entropy
\begin{equation}
S_{\text{W}} (\rho) = - \frac{1}{2} \ln \det C + N + M.
\label{eq:WehrlEntropyGaussianAB}
\end{equation}
Note that the Wehrl-Lieb inequality for $N+M$ modes then implies that
\begin{equation}
\det C \le 1 \Leftrightarrow \det \left(V + \frac{1}{2} \, \mathds{1} \right) \ge 1, \label{eq:CovarianceUncertainty}
\end{equation}
which is an uncertainty relation in terms of the covariance matrix $V$ for the special case of Gaussian states.
Unlike the Robertson-Schrödinger uncertainty relation for Gaussian states $\det V \geq 1/2^{N+M}$ (see for example \cite{Hertz2019,Berges2018}), \eqref{eq:CovarianceUncertainty} is not invariant under similarity transformations of $V$ like symplectic transformations (e.g. squeezing). This shows a difference between the Robertson-Schrödinger uncertainty relation and the Wehrl-Lieb inequality. The former introduces a minimum for the uncertainty product while the latter is only tight for coherent states, i.e. equal uncertainty in position and momentum.
The next step is to compute the reduced Husimi $Q$-distribution for one subsystem. Without loss of generality we consider subsystem $B$ and end up with
\begin{equation}
\begin{split}
&Q^\rho_{B} (\vec{r}_B) \\
&= \frac{1}{(2 \pi)^N} \int \text{d}^{2N} r_A \, Q^{\rho} (\vec{r}) \\
&= \sqrt{\frac{\det C}{\det C_A}} \, \exp \left(- \frac{1}{2} \vec{r}_B^T \left(C_B - C_M^T C_A^{-1} C_M \right) \vec{r}_B \right).
\end{split}
\end{equation}
Calculating the local Wehrl entropy gives
\begin{equation}
S_{\text{W}} (\rho_B ) = - \frac{1}{2} \ln \det C + \frac{1}{2} \ln \det C_A + M,
\label{eq:WehrlEntropyGaussianB}
\end{equation}
where we used the identity
\begin{equation}
\det C = \det C_A \, \det \left(C_B - C_M^T C_A^{-1} C_M \right),
\end{equation}
which holds provided that $\det C_A \neq 0$. Here, the Wehrl-Lieb inequality \eqref{eq:WehrlLiebInequality} requires
\begin{equation}
\det C \le \det C_A.
\end{equation}
Assuming that all Wehrl entropies are finite allows to calculate the two quantities of our interest. Both appear to have a simple form
\begin{equation}
\begin{split}
S^\rho_{\text{W}} (A | B) &= N - \frac{1}{2} \ln \det C_A, \\
I^\rho_{\text{W}} (A : B) &= \frac{1}{2} \ln \frac{\det C_A \det C_B}{\det C}.
\end{split}
\label{eq:PureGaussianStateEntanglement}
\end{equation}
It can be seen easily that in the pure state case that $I^\rho_\text{W} (A : B) = 0$ is a necessary and sufficient separability condition for Gaussian states as it corresponds to the Wehrl relative entropy between the full Husimi $Q$-distribution and its factorization. The latter is equivalent to $C_M=0$, which itself is equivalent to $V_M = 0$. This already implies separability via the positive partial transpose (PPT) criterion for pure Gaussian states \cite{Lami2018,Serafini2017}. For the special case $N=M=1$, which is the setup of the Simon criterion \cite{Simon2000}, we show that the Simon criterion indeed implies $C_M = 0$ for pure separable Gaussian states in appendix \ref{app:SimonCriterion}. At least in this special case, the criterion of ref. \cite{Simon2000} and our condition for separability, $I^\rho_\text{W} (A : B) = 0$, are therefore equivalent.
Interestingly, we find also that the Wehrl conditional entropy $S^\rho_{\text{W}} (A | B)$ is independent of the correlations encoded in $C_M$. Furthermore, applying \eqref{eq:ConditionalWL} to \eqref{eq:PureGaussianStateEntanglement} leads to local uncertainty relations of the form
\begin{equation}
\det C_A \le 1 \hspace{0.5cm} \text{and} \hspace{0.5cm} \det C_B \le 1.
\end{equation}
\subsubsection{Two-mode squeezed state}
As a more concrete example, we consider the two-mode squeezed state for $N=M=1$, which is defined by \cite{Weedbrook2012}
\begin{equation}
\ket{\psi} = \sqrt{1 - \lambda^2} \sum_{n=0}^{\infty} (- \lambda)^n \ket{n_A,n_B},
\label{eq:TwoModeSqueezedState}
\end{equation}
where the squeezing parameter $\lambda \in [0,1]$ is related to the squeezing parameter $r \in [0, \infty)$ by
\begin{equation}
\lambda = \tanh r.
\end{equation}
Since the two-mode squeezed state is a pure Gaussian state, the corresponding Husimi $Q$-distribution is of Gaussian form \eqref{eq:HusimiQGauss}. Determining the matrix $C$ yields
\begin{equation}
C_A = C_B = \mathds{1}_2 \hspace{0.5cm} \text{and} \hspace{0.5cm} C_M = \text{diag} (\lambda, - \lambda).
\end{equation}
As expected, the correlation matrix $C_M$ becomes zero for $\lambda = 0$, in which case the state \eqref{eq:TwoModeSqueezedState} is separable, such that the global Husimi $Q$-distribution factorizes. For the Wehrl conditional entropy and the Wehrl mutual information (cf. Eq. \eqref{eq:PureGaussianStateEntanglement}) we find
\begin{equation}
\begin{split}
S^\rho_{\text{W}} (A | B) &= 1, \\
I^\rho_{\text{W}} (A : B) &= - \ln \left(1 - \lambda^2 \right).
\end{split}
\label{eq:TwoModeSqueezedStateEntanglement}
\end{equation}
Both quantities allow for a straight-forward classification in an actual experiment: since the Wehrl conditional entropy is independent of quantum correlations encoded in the squeezing parameter $\lambda$, any measured deviation from $1$ is sufficient to show that the state exhibits additional classical correlations. For example, the squeezing parameter $\lambda$ may fluctuate in subsequent realizations of an experiment, which increases the Wehrl conditional entropy due to its concavity. Note that this statement is a consequence of the Wehrl conditional entropy \eqref{eq:PureGaussianStateEntanglement} being independent of the correlations encoded in $C_M$. Also, it is interesting to note that the lower bound \eqref{eq:ConditionalWL} \textit{is} attained for the two-mode squeezed state (see \autoref{fig:QHOWehrlMutualInformation}, left panel, red dotted curve), to indicate the amount of quantum correlations captured by the Wehrl mutual information.
Furthermore, the Wehrl mutual information allows to set a lower bound on the amount of bipartite entanglement between $A$ and $B$. The latter is monotonically increasing for an increasing squeezing parameter (see \autoref{fig:QHOWehrlMutualInformation}, left panel, blue straight curve). In particular, it vanishes if and only if the state is separable ($\lambda \to 0$) and tends to infinity if the state approaches the Einstein-Podolsky-Rosen state ($\lambda \to 1$) representing perfect correlations $x_A = x_B$ and $p_A = - p_B$. Furthermore, we see how the bound \eqref{eq:MutualInformationBound} behaves as we have also plotted the quantum mutual information (see \autoref{fig:QHOWehrlMutualInformation}, left panel, green dotted curve), which can be calculated easily following refs.\ \cite{Serafini2003,Weedbrook2012}.
\subsubsection{N00N states}
Another interesting class of pure entangled states are the N00N states, which are of the form (the normalization constant is chosen such that the following expression is normalized to unity also for $N=0$)
\begin{equation}
\ket{\psi_N} = \frac{1}{\sqrt{2 (1 + \delta_{0 N})}} \left(\ket{N,0} + \ket{0,N} \right),
\end{equation}
for $1+1$ modes (note that here $N$ denotes the $N$th excitation instead of the $N$th mode). These states are separable if and only if $N=0$, in which case we end up with two vacua again. Furthermore, for $N=1$ one recovers one of the four Bell states as a special case.
The N00N states are known to be hard to detect if considered in the context of a continuous variables quantum system \cite{Weedbrook2012}. For example, the strong entropic separability criterion in ref.\ \cite{Walborn2009} only detects N00N states up to $N=5$. Since the Wehrl mutual information is a perfect witness, all N00N states are witnessed.
Calculating the global and one local Husimi $Q$-distribution gives
\begin{equation}
\begin{split}
Q^{\rho} (\vec{r}) = & \frac{1}{2^{N+1} \, N! \, (1 + \delta_{0 N})} \, e^{-(r_A^2 + r_B^2) /2} \\
& \times \left( (x_A - i p_A)^N + (x_B - i p_B)^N \right) \\
& \times \left( (x_A + i p_A)^N + (x_B + i p_B)^N \right), \\
Q^\rho_{B} (\vec{r_B}) = & \frac{1}{2^{N+1} \, N!} \, e^{-r_B^2 /2} \, \left( r_B^{2 N} + 2^N N! \right),
\end{split}
\end{equation}
where we used the abbreviations $r_i^2 = \abs{\vec{r}_i}^2 = x_i^2 + p_i^2$ for $i \in \{ A, B \}$. The Wehrl conditional entropy and the Wehrl mutual information can be computed numerically (see \autoref{fig:QHOWehrlMutualInformation}, right panel, same color scheme as before). We observe that the Wehrl mutual information increases monotonically for $N \ge 2$. It can be seen that it approaches the quantum mutual information, which is simply $I^{\rho} (A:B) = 2 \ln 2$ for all $N \in \mathbb{N}$.
\section{Conclusion and outlook}
\label{sec:Conclusion}
In summary, we have investigated here the relation between the Wehrl-Lieb entropic uncertainty relation for the Husimi $Q$-distribution and other relations for marginal distributions. Moreover, we have discussed to which extent the Wehrl entropy and related quantities are able to witness entanglement of continuous variables.
We found that the Wehrl-Lieb inequality is tighter than the Białynicki-Birula and Mycielski relation almost everywhere. In particular, we have shown its superior tightness for number eigenstates $\rho_n = \ket{n} \bra{n}$, where $n=1$ turned out to be an exception. Also, the calculation showed that Wehrl entropies are technically simpler to handle compared to differential entropies of marginal distributions.
In the second half, we have introduced analogs of well-known quantum information theoretic quantities in terms of Wehrl entropies. It turned out that especially the Wehrl mutual information is of great use when it comes to entanglement witnessing. Although it is neither an entanglement measure, nor an entanglement monotone, it is a \textit{perfect} witness for pure state entanglement, i.e. it detects \textit{all} pure entangled states. Furthermore, it gives a lower bound on the entanglement entropy, which allows to quantify to which extent some pure state $\rho = \ket{\psi} \bra{\psi}$ is entangled \textit{at least}.
To exemplify the potential of the Wehrl mutual information we discussed several examples. For the Gaussian state case, after a straight forward exercise in Gaussian integration, we found a simple formula allowing for a direct classification. Here it would be of further interest to relate the Wehrl mutual information as perfect witness for pure Gaussian states to existing criteria like for example the Simon criterion \cite{Simon2000}. Additionally, we have seen that the Wehrl-Lieb inequality for $N+M$ mode Gaussian states implies a novel uncertainty relation in terms of the covariance matrix, which may be studied further in the future. Another interesting example we considered was the class of $N00N$ states, which are particularly hard to detect, especially for large $N$. Also in this case the Wehrl mutual information was a reasonable entanglement witness, in fact it is a witness for all $N \in \mathbb{N}$.
Besides the Wehrl mutual information we also considered the Wehrl conditional entropy, which turned out to be bounded from below similar to the Wehrl entropy itself. This statement can be interpreted as an entropic uncertainty relation in the presence of classical memory as well as a refined version of the monotonicity property of the Wehrl entropy under partial trace. Furthermore, for the special case of Gaussian states, we found that the Wehrl conditional entropy is independent of the correlations encoded in the off-diagonal blocks of the covariance matrix.
Overall, we want to emphasize again that the significance of the Wehrl entropy and all related quantities is that they are \textit{measurable} when making use of the heterodyne detection scheme. Thus, the Wehrl mutual information and the Wehrl conditional entropy can be applied immediately to actual experiments, where they serve as a perfect witness for pure state entanglement and as an indicator for non-Gaussianity, respectively.
One question left over for future work concerns a more general notion of coherent states, namely $SU(N)$ coherent states. It should be checked whether the presented approach to witness pure state entanglement can be generalized to $SU(N)$ coherent states. This question is of particular interest for e.g. experiments with spinor Bose-Einstein condensates \cite{Kunkel2018,Bergh2021a,Bergh2021b}.
It is noteworthy, that the Wehrl-Lieb inequality holds also for generalized coherent states (see refs.\ \cite{Davies1976,Grabowski1984}), which may be another direction for future investigations.
Other directions for subsequent research are the generalization of the presented quantities to quantify entanglement also in the multipartite case and to other interesting states, e.\ g.\ Schrödinger cat states (see \cite{Miranowicz2001} for results on the Wehrl entropy in this case).
Furthermore, it would be of great interest to generalize the concept of an entropy in quantum mechanical phase space to a quantum field theory. In this way, it should be possible to formulate entropic uncertainty relations for conjugate quantum fields with a Wehrl (relative) entropy and to constrain measurable (quantum) correlations analogous to what we have discussed here.
\section*{Acknowledgements}
The authors thank Martin Gaerttner and Markus Oberthaler for valuable discussions. This work is supported by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany's Excellence Strategy EXC 2181/1 - 390900948 (the Heidelberg STRUCTURES Excellence Cluster), SFB 1225 (ISOQUANT) as well as FL 736/3-1.
|
\section{Introduction}
Improvements in high pressure and temperature experiments mean that the topic of liquid-liquid phase transitions has attracted significant attention recently. In particular, there are debates about whether a change in liquid structure can be regarded as a true phase transition, or a gradual crossover.
Determining this is challenging for either experiment or simulation. In a diamond anvil cell experiment it is near-impossible to observe phase coexistence and be confident that the system has reached thermodynamic equilibrium. Indeed, many if not most high pressure experiments report phase coexistence across a range of pressures - something which is forbidden in equilibrium thermodynamics. The situation is no different in simulations - typical electronic structure calculations are done at a given pressure and temperature and ``discontinuities" are inferred by extrapolation or, at best, hysteresis.
The melt curve for most materials has a positive slope on a PT phase diagram. This means that the liquid is less dense than the solid.
There are exceptions, notably, water is denser than ice, and other examples among elements include silicon, gallium, carbon.
These textbook exceptions at ambient pressure can generally be understood as due to the partial breakdown of a network of well defined covalent or hydrogen bonds causing the atoms in the liquid to have a higher coordination than the solid.
Another group of materials which have a negative Clapeyron slope at high pressure are the alkali metals\cite{Tsuji1990,Tsuji1996,falconi2006X-ray,mcbride2015one,marques2009potassium,gatti2010sodium,woolman2018structural,Marques2011,marques2011optical,zhao2019commensurate,Miao2014,frost2018equation,frost2019high,schaeffer2012high,guillaume2011cold,gregoryanz2008structural,eshet2012microscopic,hong2019reentrant,zha1985melting,narygina2011melting,kartoon2021structural}. Here, the slope is typically positive at ambient pressure, reaching a maximum, then becoming negative in a pressure region where the solid phase is typically close packed. At still higher pressures, there is a minimum in the melt curve before the slope becomes positive again.
The solid phase in the region of negative slope is close packed, so the densification on melting cannot come from a collapsing open network.
Curiously, hydrogen has a remarkably similar phase diagram to the alkalis which can be explained by competition between free rotors and quadrupole interactions\cite{magduau2017simple,zong2020understanding,van2020quadrupole}
Density functional theory can reproduce the negative slope\cite{hernandez2010first,feng2015nuclear,robinson2019chain}. It also shows some anomalous behaviour in liquid heat capacity, compressibility, viscosity, and thermal expansion\cite{zongK}. This implies that there is some significant change in the liquid binding, though whether it is a true transformation or a crossover remains unclear.
As a consequence, there is renewed interest is analytic equation of state which can be fitted to data. For single phases, functional forms such as the Vinet equation of state work well, but many interesting phenomena occur where the equation of state is concave or discontinuous.
The thermodynamically stable state is the one with the lowest Gibbs free energy:
\[ G(P,T) = U + PV - TS \]
taking $P$ and $T$ as the independent variables. Evidently $G$ depends on three quantities, energy, density and entropy. Any attempt to relate microscopic to macroscopic properties needs to consider all three, and how they vary between phases.
There are a wide range of approaches to describe complex high pressure structures. Those based on electronic-structure include electride\cite{marques2009potassium}, two-band\cite{ackland2003two,ackland2006two}, Fermi-surface\cite{Degtyareva2003,ackland2004origin,degtyareva2006simple}, s-p\cite{jr:Yanming} or s-d\cite{ReedBa2000} transfer, Mott transitions\cite{johansson1974alpha} or pairing\cite{jr:Pairing}, molecularisation\cite{ashcroft2000hydrogen,katayama2000first,monaco2003nature}, polymerisation, \cite{ross2006polymerization} and ``simple-complex"\cite{gorelli2018simple} transition types.
Other approaches based on interatomic forces include soft-core\cite{ogura1977computer,young1977melting}, and associating particles\cite{Jackson_etal_1988,Hopkins_etal_2006}.
Despite this huge variety of microscopic models, we are unaware of a simple, analytic thermodynamic model for the melting point maxima\cite{rapoport1968melting,gregoryanz2005melting,kechin2001melting} and liquid-liquid transformation\cite{mukherjee2007high, Cadien2013}. Rapoport\cite{Rapoport1967} implies that Klement built such a model, but it was never published - Rapoport's own analysis of Klement's model does not show a melting point maximum. A number of lattice-based approaches have been tried \cite{dijkstra1994evidence}, but for obvious reasons their applicability to the fluid state is debatable.
We note that most of the microscopic models are based on a trade-off between two types of interaction, one which has lower energy, the other lower volume.
The purpose of this paper is to lay out the minimal requirements for an analytic model of a discontinuous liquid-liquid transformation and a melting point maximum based only on only the idea that a material can adopt two different states.
The paper is structured as follows - we start by deriving thermodynamic results for heat capacity, expansivity and compressibility in a convenient analytic form. We then present a mixing model between two thermodynamic states, demonstrating the Widom lines. A microscopic model inspired by the electride transition\cite{dye1990electrides,ahulwalia1999structural,marques2009potassium,pickard2009dense,gatti2010sodium,raty2007electronic,rousseau2011exotic,li2015metallic,woolman2018structural,Marques2011,marques2011optical,zhao2019commensurate,Miao2014,miao2015high,yu2018optical,frost2018equation,elatresh2019high,paul2020thermal,ayrinhac2020high}, where the states differ only in volumes, is worked through in detail for both solid and liquid cases.
It is shown that this model is sufficient to obtain the melt curve maximum, and can support a discontinuous phase transition in the liquid.
A parameterization for potassium is presented.
\section{Thermodynamic model}
The theory derived here is of very general applicability. However, we found it helpful to have a concrete microscopic model in mind as it is developed.
\subsection{Motivation from simple metals at pressure}
We propose that the structure of the high-pressure alkali metals can be modeled as a mixture of two distinct electronic states: a low-pressure s-type free electrons, and a high-pressure ``electride’’ state, with electrons localized in interstitial pockets, referred to as pseudoanions. In the case of fcc, we can imagine that the octahedral site is the pseudoanion, so the electride has a rocksalt structure.
This should not be taken too literally because in reality, the electride transition is accompanied by a crystal structural transformation. Similar evolution happens in a liquid, but here the transition is continuous because differently-sized pseudoanion sites are available, and there is no symmetry.
This microscopic picture can be related to a macroscopic one by considering the energy, volume and entropy of the two states:
\begin{itemize}
\item The electride has higher energy because the electron is confined away from the positively charged ion.
\item The electride has small volume, because it can occupy the interstitial site between ions, leading to more efficient packing.
\item A mixture of the two states gives higher entropy.
\end{itemize}
In addition to the electride transition, we may also compare solid and liquid phases for which
the solid has lower entropy and enthalpy, independent of the electride fraction.
The need to describe $U$, $TS$ and $PV$ for each phase means that even the simplest model inevitably has several parameters.
\subsection{Thermodynamics}
In a general two-state model, a Gibbs free energy is written as $G(x,P,T)$ where $x$ is the fraction of one of the two states. The equilibrium value for $G(P,T)$ is obtained by minimising $G(x,P,T)$ with respect to $x$. So for all $P,T$ we have
\begin{equation} G(P,T)= \min_x G(x,P,T) \label{eq:Gx}
\end{equation}
A necessary, but not sufficient requirement for equilibrium is:
\begin{equation} \left (\frac{\partial G}{\partial x}\right )_{P,T}=\left (\frac{\partial H}{\partial x}\right )_{P,T} -T \left (\frac{\partial S}{\partial x}\right )_{P,T} =0 \label{eq:eqm} \end{equation}
Simply solving that equation will also generate unphysical free energy maxima, and metastable states.
Thermodynamic properties are obtained as derivatives of the free energy. Although the calculus is routine, we present the results here because of the additional terms which arise due to the $x$ factor, and the fact that some derivative cannot be written analytically because of the requirement to minimise $x$.
\subsubsection{Heat capacity $C_p$}
The standard thermodynamic definitions of the heat capacity are
\begin{equation} C_p = \left (\frac{\partial H}{\partial T}\right )_P
= T\left (\frac{\partial S}{\partial T}\right )_P=T\left
(\frac{\partial^2 }{\partial T^2}Min_{x} [G(P,T,x)]\right )_P
\end{equation}
Note that $x$ is not an independent variable, and changes in $x$ contribute to the heat capacity.
\[ C_P =\left ( \frac{\partial H}{\partial T}\right )_{P,x} +
\left ( \frac{\partial H}{\partial x}\right )_P\left ( \frac{\partial x}{\partial T}\right )_P\]
The quantity $\frac{\partial T}{\partial x}$ can be awkward to evaluate, so to eliminate it, we consider
\begin{equation}\left (\frac{\partial}{\partial T}\right )\left (\frac{\partial G}{\partial x}\right ) = -\left (\frac{\partial S}{\partial x}\right ) + \left (\frac{\partial x}{\partial T}\right ) \left[ \left (\frac{\partial^2 G}{\partial x^2}\right )
\right]
\end{equation}
dropping the subscripts for clarity. Using the equilibrium condition (Eq.\ref{eq:eqm}), this gives
\begin{equation} \left ( \frac{\partial T}{\partial x}\right )_P
= T \left (\frac{\partial^2 G}{\partial x^2}\right )_P / \left (\frac{\partial H}{\partial x}\right )_P \label{eq:Tx}
\end{equation}
From which the expression for the heat capacity becomes:
\begin{equation} C_P =\left ( \frac{\partial H}{\partial T}\right )_{P}
= \left ( \frac{\partial H}{\partial T}\right )_{P,x} + \frac{1}{T}\left (\frac{\partial H}{\partial x}\right )^2_P / \left (\frac{\partial^2 G}{\partial x^2}\right )_P
\end{equation}
\subsubsection{Isothermal Compressibility}
The standard thermodynamic definitions of compressibility are
\begin{equation} K = -\frac{1}{V}\left ( \frac{\partial V}{\partial P}\right )_T
= -\frac{1}{V}\left (\frac{\partial^2 G}{\partial P^2}\right )_T
\end{equation}
including the internal variable $x$.
\[ \left (\frac{\partial^2 G}{\partial P^2}\right )_T =
\left (\frac{\partial^2 G}{\partial P^2}\right )_{T,x} +
2
\left (\frac{\partial^2 G}{\partial P\partial x}\right )_T
\left (\frac{\partial x}{\partial P}\right )_{T} + \cancelto{0}{
\left (\frac{\partial G}{\partial x} \right )_T\left (\frac{\partial^2 x}{\partial P^2}\right )_T}
\]
The equilibrium condition $\left ( \frac{dG}{dx}\right ) =0$ ensures that the third term is zero.
Again, there is no convenient relationship between $P$ and $x$, but following a similar argument to Eq.\ref{eq:Tx} we find
\[\left ( \frac{\partial x}{\partial P}\right )_T
=-\left (\frac{\partial^2 G}{\partial P\partial x}\right )_T/\left (\frac{\partial^2 G}{\partial x^2}\right )_T
\]
and
\[\left ( \frac{\partial^2 x}{\partial P^2}\right )_T
=\left (\frac{\partial^3 G}{\partial P^2\partial x}\right )_T/\left (\frac{\partial^2 G}{\partial x^2}\right )_T
-\left (\frac{\partial^3 G}{\partial P\partial x^2}\right )^2_T/\left (\frac{\partial^2 G}{\partial x^2}\right )^2_T
\]
\subsubsection{Thermal Expansion}
The standard thermodynamic definitions of compressibility are
\begin{equation} \alpha = \frac{1}{V}\left ( \frac{\partial V}{\partial T}\right )_T
= \frac{1}{V}\left (\frac{\partial^2 G}{\partial T \partial P}\right )
\end{equation}
\begin{equation}
\frac{\partial^2 G}{\partial T \partial P} = \left (\frac{\partial^2 G}{\partial T \partial P}\right )_x +
\left (\frac{\partial^2 G}{\partial P\partial x}\right )_T
\left (\frac{\partial x}{\partial T}\right )_{P}
+\left (\frac{\partial^2 G}{\partial T\partial x}\right )_P
\left (\frac{\partial x}{\partial P}\right )_{T}
\end{equation}
Again, using the equilibrium condition
$\left (\frac{\partial G}{\partial x} \right )_T=0$.
\subsection{Linear combination model with ideal solution\label{model:SS}}
In a slightly more specific model, a system is described by particles in two possible thermodynamic states A (x=1) and B (x=0). When mixed in an ideal solution, the Gibbs free energy is given by:
\begin{equation}
G(P,T) = x G(1,P,T) + (1-x) G(0,P,T) + RT[x \ln{x} + (1-x)\ln{(1-x)}]
\label{Gibbs:SS}
\end{equation}
where $x$ is the fraction of particles in state A, $G_A=G(1,P,T)$ and $G_B=G(0,P,T)$ are the Gibbs free energies of pure A and B states.
The equilibrium value for $x$ is found by minimising $G(P,T)$:
\begin{equation}
x(P,T) = \frac{e^{-G_A/RT}}{e^{-G_A/RT}+e^{-G_B/RT}}
= \frac{1}{1+e^{-\Delta G/RT}}
\\
\label{xp:SS}
\end{equation}
with $\Delta G= G_B-G_A$.
We can find the volume
\begin{equation}
V(P,T) = \left ( \frac{\partial G}{\partial P}\right )_T = xV_A(P,T) + (1-x)V_B(P,T)
\label{volume:SS}
\end{equation}
and entropy
\begin{equation}
S(P,T) = -\left ( \frac{\partial G}{\partial T}\right )_P = xS_A(P,T) + (1-x)S_B(P,T) + R[x \log{x} + (1-x)\log{(1-x)}]
\label{entropy:SS}
\end{equation}
always remembering that $x=x(P,T)$.
We further derive analytic expressions for compressibility
\begin{equation}
\kappa_T = -\frac{1}{V}\left (\frac{\partial V}{\partial P}\right )_T= x \kappa_{T,A} + (1-x)\kappa_{T,B} +
\frac{\Delta V}{V} \left ( \frac{\partial x}{\partial P}\right )_T
\label{kappa:SS}
\end{equation}
with $\Delta V$=$V_B-V_A$
and thermal expansion
\begin{equation}
\alpha = \frac{1}{V}\left( \frac{\partial V}{\partial T}\right )_P = x \alpha_{T,A} + (1-x)\alpha_{T,B} +
\frac{\Delta V}{V} \left ( \frac{\partial x}{\partial T}\right )_P
\label{alpha:SS}
\end{equation}
Both of which have an anomalous component arising from the conversion of material between the two states, in addition to the weighted average.
For the heat capacity there is an additional anomalous term from the mixing entropy
\begin{equation}
C_P = T \left ( \frac{\partial S}{\partial T}\right )_P = x C_{P,A} + (1-x)C_{P,B} +\left (
\Delta S + R\ln{\frac{x}{1-x}} \right ) \left ( \frac{\partial x}{\partial T}\right)_P
\label{cp:SS}
\end{equation}
with $\Delta S$=$S_B-S_A$.
From Equation \ref{xp:SS} we immediately see that there is no discontinuity in $x$, from which it follows that this model cannot describe a phase transition, only a crossover. We also observe that the ideal solution entropy ensures that mathematically, as well as intuitively, $0<x<1$.
\subsection{Non-ideal solution solid model}
A small rephrasing of the Bragg-Williams\cite{bragg1934effect} (BW) model can be used to extend the model from section \ref{model:SS} to describe a discontinuous transition within a single solid phase. BW is a mean field approximation to the Ising model, where for high-pressure applications the spins are mapped to ``electride" and ``s-electron" states, and the ``field" is mapped to the enthalpy difference between the two states. Although the model has wider applicability, will use the electride terminology here.
The enthalpy in this case is
\[ H=x(\Delta U_e + P\Delta V_e) +Jx(1-x)\]
where $x$ is the electride fraction, $\Delta U_e$ and $\Delta V_e$ are the change in energy and volume when an electron moves to an electride pseudoanion site, both assumed positive, $J$ is a local coupling between electride and free electron.
A high pressure phase transition at $T=0$ occurs when the field/enthalpy difference changes sign ($P_T=\Delta U/\Delta V$).
Including entropy, the Gibbs Free Energy is:
\begin{equation}
G_{BW} = x(\Delta U_e - P \Delta V_e -T \Delta S_e) + RT[x\ln{x}+(1-x)\ln(1-x)] +Jx(1-x)
\label{gibbs:BW}
\end{equation}
We find that the $x$-dependent contributions are
\[ V = -x\Delta V_e \]
\[ U = x \Delta U_e + Jx(1-x) \]
\[ S = R[x\ln{x}+(1-x)\ln(1-x)] +\Delta S_e
\]
Obtaining these results by differentiating $G$ is not completely trivial, as they rely on
the stationary property of $G(x)$ at equilibrium (Eq.\ref{eq:eqm}).
We now find
\begin{equation}
\frac{\partial H}{\partial x} = (\Delta U_e - P \Delta V_e) +J(1-2x)
\end{equation}
\begin{equation}
\frac{\partial G}{\partial x} = (\Delta U_e - P \Delta V_e - T\Delta S_e) + RT\ln[{x/(1-x)}] +J(1-2x)
\end{equation}
\begin{equation}
\frac{\partial^2 G}{\partial x^2} = \frac{RT}{x(1-x)} -2J
\end{equation}
\begin{equation} C_P(x) = \frac{1}{T}
\frac{\left (
\Delta U_e - P \Delta V_e +J(1-2x)\right )^2
}{ \frac{RT}{x(1-x)} -2J }
\end{equation}
We can see immediately that the heat capacity has a discontinuity if $RT/2J= x(1-x)$,
and since $x(1-x)$ must lie between 0 and 1/4, a discontinuous phase transition occurs for any $T<J/2R$
at $P=\Delta U_e/\Delta V_e$. Interestingly, along a line above the critical point, the contribution to $C_P$ goes to zero
\begin{figure}[!htb
\centering
\includegraphics[width=0.98\columnwidth]{freex.pdf}
\includegraphics[width=0.98\columnwidth]{xvsPressure.pdf}
\caption{(top) Variation of free energy with x for liquid model with $\Delta V_0=1$, $\Delta E_0=1$, J=1.25 at T=0.1. Lines correspond to different pressures with selected values around the transition at P=1 highlighted. The equilibrium value of $x$ corresponds to the free energy minimum. These conditions permit a first order transformation, and a metastable state can be seen for P=0.98.
(bottom) Variation of fraction of two states ($x$) for solid and liquid model, with identical energy and volume differences. At reduced T=0.4 the solid already exhibits a discontinuous phase transition, while the liquid does not (T=0.1 shown). If J is increased from 1.0 to 1.25, the liquid model also exhibits a phase transition.
The chosen values of $\Delta E_0=1$; $\Delta V_0=1$ mean that the transition pressure is at $P=1$ in either model.
Inset - schematic showing the different volume changes available in liquid (Orange) compared with unique value in solid (blue, delta function).\label{fig:BWx}}
\end{figure}
If it seems odd that $V$ and $U$ go to zero, at high and low pressures. remember that the
full free energy of the system will include terms independent of $x$, representing the equation of state of a reference $(x=0)$ material.
To compare with a real system, one needs to add an $x$-independent free energy $G_{ref}(P,T)$ to Eq.~\ref{gibbs:BW} which adds a smoothly varying additional term to all quantities.
Bragg and Williams considered an atomic level system\cite{bragg1934effect}, so assumed that the two sites have equal entropy; In applications such as polymerization or atomic-molecular where the number of independent objects changes a further term $T\Delta S_e=0$ could be added to relax this assumption. This introduces a slope to the phase boundary and additional tilt to the Widom line, but does not change the general picture.
\subsection{Two site electride liquid model}
Our liquid more differs from the solid in just one detail.
Because of the symmetry in a crystal, the available electride sites are equivalent.
In a liquid, we assume there are a range of different possible electride sites, each entailing different volume changes $\Delta V$ (see fig.\ref{fig:BWx}).
The site offering the largest volume reduction will be occupied first. For simplicity, we assume sites are linearly distributed and range from $-2 \Delta V_e $ to 0. Note that a positive $\Delta V$ implies that the electride would increase the volume, so such sites will never be occupied.
With finite electride fraction $x$, the total volume therefore changes by:
\begin{equation} \Delta V = \int_0^x 2\Delta V_e (x' - 1) dx' = \Delta V_e (x^2-2x)
\end{equation}
The excess Gibbs Free Energy is thus:
\begin{equation}
G_{EL}(x) = x\Delta U_e - x(2-x)P \Delta V_e + k_B T[x\ln{x}+(1-x)\ln(1-x)] +Jx(1-x)
\end{equation}
There are now nonlinearities in energy, entropy and density.
\[ V = -x(2-x)\Delta V_e \]
\[ U = x \Delta U + Jx(1-x) \]
\[ S = R[x\ln{x}+(1-x)\ln(1-x)]
\]
We now find
\begin{equation}
\frac{\partial H}{\partial x} = (\Delta U_e - (2-2x)P \Delta V_e) +J(1-2x)
\end{equation}
\begin{equation}
\frac{\partial G}{\partial x} = (\Delta U_e - (2-2x)P \Delta V_e) + RT\ln[{x/(1-x)}] +J(1-2x)
\end{equation}
\begin{equation}
\frac{\partial^2 G}{\partial x^2} = 2P\Delta V_e + \frac{RT}{x(1-x)} -2J
\end{equation}
\begin{equation} C_P = \frac{1}{T}
\frac{\left [
\Delta U_e - (2-2x) P \Delta V_e +J(1-2x)\right ]^2
}{ \frac{RT}{x(1-x)} -2J +2P\Delta V_e }
\end{equation}
This model does not necessarily have a critical point: the entropic and volume terms are always convex, so only the demixing $J$ term can drive phase separation. Whether this happens depends on the value of $P$ at the putative phase boundary $x=\frac{1}{2}$, giving $P=\Delta U_e/\Delta V_e$.
These quantities are plotted in Figure~\ref{fig:BW}, where it is again clear that the model predicts a peak in specific heat and compressibility, along with a dip in the thermal expansivity. These extrema trace out the Widom lines of the phase diagram (Figure~\ref{fig:Widom}). It is important to note that this phase diagram includes only the two-site Hamiltonian: the underlying free energy of the $x=1$ and $x=0$ states is ignored.
Free energy variation with $x$ is shown in Figure~\ref{fig:BWx}, for a range of pressures around the phase transition. Below $T_c$ there are two minima, degenerate at $P=\Delta U_e/ \Delta V_e$, indicating a first order phase transition. An analytic estimate of $T_c$ can be obtained from $\frac{\partial^2 G}{\partial x^2}=0$ or from setting $x=1/2/$. Note that the existence of the critical point requires two nonlinear terms in $G$, coming here from the entropy and the interaction energy. Figure \ref{fig:BWx}(lower) showing that $x$ changes discontinuously along an isotherm at the transformation, in either solid or liquid model. Notice that, for equivalent parameters, the critical point in the liquid falls at a lower $T_c$ than for the solid.
If the liquid structure cannot accommodate enough potential electride sites, the model can be extended to a maximum electride fraction $f$. This would result in a change of the $P\Delta V_e$ prefactor from $x(2-x)$ to $x(2-x/f)$, but this additional complication makes no difference to the general argument, so hereinafter we take $f=1$.
Positive $J$ generates a first order transition with a critical point. The phase line is vertical (at $P=\Delta U_e/ \Delta V_e$) and ends at the critical temperature $Tc$. Note that the high pressure phase transition we are describing corresponds to the Ising spin-up $\rightarrow$ spin-down transition, not the usual BW paramagnetic one.
Above the critical temperature, there are anomalies in several observables, as shown in Figure~\ref{fig:BW}. The extreme values (Widom lines) for various properties do not fall in the same place: any definition of the supercritical transition pressure depends on which property is considered.
\begin{figure}[ht
\centering
\includegraphics[width=0.98\columnwidth]{WidomBWliq.pdf}
\caption{Phase diagram for the liquid model with parameters as shown ($\Delta S=0$ - if non-zero then the phase transition line has a slope. Points are the calculated maxima and minimum of the thermodynamic anomaly in compressibility, expansivity and heat capacity. The first order transition between high and low $x$ liquids ends in a critical point: in application of the model to real materials, this critical point may lie below the melt line. Above the critical temperature, the extrema of the thermodynamic properties trace out the Widom lines which converge and end at the critical point.}
\label{fig:Widom}
\end{figure}
\clearpage
\subsection{Entropy-driven transformation}
So far we have considered models where the difference between the two phases is in the enthalpy. In other cases, such as the molecular-atomic transition in hydrogen, there is a significant change in entropy between the two states - in the hydrogen case because the number of particles doubles.
The addition of an $x$-dependent entropy term gives a slope to the phase boundary, and a similar change of slope to the Widom lines: some of which can even have the opposite slope to the phase boundary.
From the Clausius-Clapeyron equation, the slope of the phase boundary is $\frac{dP}{dT} = \Delta S/\Delta V$. Exactly similar to the volume change, a linear dependence of entropy with $x$ does not create a first order transition, the lowest order term which can do so is $x(1-x)\Delta S$. Such entropic demixing occurs in models with hard-core cubes and spheres \cite{dijkstra1994phase,dijkstra1994evidence,dijkstra1998phase}, and has been claimed experimentally in supercooled water \cite{holten2012entropy}.
The heat capacity model with $\Delta S=2$ is illustrated in Figure~\ref{fig:heatcapacity}, showing the lambda profile of the discontinuous transition changing to the broad peak above the critical point. The gradient of the Clapeyron slope is evident from the shift of the lambda peak to higher pressure. The inset shows the peak in heat capacity in the low-$x$ phase.
\subsection{Two site model for liquid-solid transformation}
We can extend the two-site model to compare liquid and solid phases and calculate a melt line. This requires us to consider the $x$-independent contributions to the free energy. A full equation of state is required for the non-anomalous contributions to $C_p$, $\alpha$ and $\kappa$, however, to calculate the phase boundary, we need only know the free energy {\it difference} of $x$-independent contributions to the solid-liquid free energy $\Delta G_{sl}$.
Thus we have an equation for the phase boundary
\begin{equation} \Delta G_{sl}(P,T)=G^l(x_l,P,T)-G^s(x_s,P,T)\end{equation}
where $x_l(P,T)$ and $x_s(P,T)$ are the equilibrium values of $x$ in liquid and solid respectively
In figure \ref{fig:melt} we show an illustrative example
with a zero pressure melting point at T=0.4 and a positive Clapeyron slope of 0.8, in reduced units.
To illustrate the model, we use the same parameters $x$-dependent parameters $\Delta V=1$, $\Delta U=1$, $\Delta S=0$, $J=1$ in both liquid and solid.
This means that $x$-dependent terms in free energy for solid and liquid models are equal in the $x\rightarrow 0$ and $x\rightarrow 1$ limits.
For the $x$-independent terms, we assume that energy, entropy and density differences between solid and liquid are constant.
This choice of parameters means that the solid-solid phase line is vertical ($\Delta S=0$) and there is no discontinuous liquid-liquid transition. This is similar to the case of the simple metals. A significantly larger value of $J$ would be needed to extend the phase boundary into the liquid region, as shown in figure 3.
A non-zero $\Delta S$ leads to a slope in the phase boundary, but does not change the general picture.
The figure also shows how $x$ varies across the phase diagram - gradually in the liquid, but discontinuously in the solid.
\subsection{Example - application to potassium}
The high pressure crystallography and re-entrant melt curve of potassium has been determined experimentally\cite{Lundegaard2009b,mcbride2015one,jr:Reentrantincommensuratepotassium}. DFT calculations show an electride transition in potassium as in other simple metals\cite{falconi2006ab,marques2009potassium,gatti2010sodium,woolman2018structural,Marques2011,marques2011optical,zhao2019commensurate,Miao2014}. Liquid potassium calculations suggest a number of irregularities\cite{zongK} in the thermodynamic properties which cannot be fitted by smoothly varying models\cite{li2019regularities}.
Figure \ref{fig:KLF} shows an {\it ab oculo} parameterization of the liquid-solid transformation model to this data, with a simple linear model for $x$-independent terms. The reduced units of the model, correspond to 20GPa and 1000K for potassium.
While the overall shape is reproduced with a linear fit, the low pressure melt curve appears parabolic and the high pressure line is not sufficiently steep. The fit can be significantly improved by introducing a non-linear equation of state, such that $P^*\rightarrow P^{3/2}$, and is reduced by a factor of 5 above the transition.
Curiously, the unadjusted high-pressure melt line of the linear model follows the chain-melting line, in which the guest atoms in the solid phase III melt\cite{chainmeltingpotassium,mcbride2015one}.
It is notable that the melting line minimum is coincident with the triple point of the solid-solid transformation from fcc to host-guest structure, which has been associated with the electride transition\cite{marques2009potassium,woolman2018structural,robinson2019chain}. The melting point maximum has no such association, which casts doubt on the extrapolation of the fcc-bcc line to the melt curve maximum, which has been drawn and copied without evidence in, e.g. lithium\cite{guillaume2011cold,ackland2017quantum,Marques2011,jr:Lithium,Hanfland2000,schaeffer2012high}. In fact, the 180$\deg$ rule means that it is thermodynamically impossible for a solid-solid phase line to intercept the melt curve at a point of negative curvature such as a maximum.
\section{Discussion}
We have presented a simple analytic model which explains the anomalous shape of the melt line observed in many high pressure systems. The key features required are
\begin{itemize}
\item a microscopic mechanism by which the atoms can reduce their volume, at the expense of increasing their energy. \item disorder in the liquid leading to wider range of possible atomic environments compared to the solid.
\end{itemize}
We have shown that a discontinuous phase transformation can be driven by a repulsion between the two states, analogous to the $Jx(1-x)$ term in the Bragg-Williams model. This repulsion may be either enthalpic or entropic, but much introduce negative curvature to $G(x)$: terms linear in $x$ cannot result in a discontinuous transition.
The discontinuous transitions in the model do not depend on changes in crystal symmetry; In reality, it is likely that a discontinuous change in the type of electronic binding of a solid will also be accompanied by a symmetry change. Thus even in principle the solid-solid critical point can occur only for isostructural transitions such as hydrogen and cerium\cite{ji2019ultrahigh,ackland2020structures,johansson1974alpha}.
Above the critical point, the model predicts a series of experimentally measurable "Widom lines" associated with anomalies of thermodynamic properties. These occur for all parameterizations, even where there is no critical point, or there is a liquid-liquid critical point which lies below the melt line.
By comparing free energy models for crystal and liquid phases, we constructed a melt line from this model. This has a characteristic minimum at the point where the two-state mixing entropy is maximised ($x=\frac{1}{2}$), coincident with the solid-solid phase transformation. Combined with a positive slope at low pressures, this means that there must also be a melting temperature maximum which, curiously, does not appear to be coincident with other features in the phase diagram.
The model has been applied to the melt curve of potassium, using a very simple linear fitting scheme. More accurate fitting to other materials would be straightforward, and the model framework has broad application for producing equations of state for any material with a complex liquid-liquid transformation.
\begin{figure}[ht
\centering
\includegraphics[width=0.48\columnwidth]{BW0.4.pdf}
\includegraphics[width=0.48\columnwidth]{BW0.6.pdf}
\includegraphics[width=0.48\columnwidth]{BWL0.4.pdf}
\includegraphics[width=0.48\columnwidth]{BWL0.6.pdf}
\caption{Thermodynamic anomalies due to Bragg-Williams-type solid model (top) and liquid model (bottom) for T=0.4 and T=0.6. with parameters set to unity, so that the solid critical temperature is 0.5 and the phase line is vertical at P=1 }
\label{fig:BW}
\end{figure}
\begin{figure}[ht
\centering
\includegraphics[width=0.8\columnwidth]{HeatCapacity.pdf}
\caption{Thermodynamic anomalies in the heat capacity for solid model with $\Delta V=1$, $\Delta U=1$, $J=1$, $\Delta S=2$. Inset shows expanded view of low pressure region. }
\label{fig:heatcapacity}
\end{figure}
\begin{figure}[ht
\centering
\includegraphics[width=0.92\columnwidth]{Kfit.pdf}
\includegraphics[width=0.98\columnwidth]{K-ELF-cropped.pdf}
\includegraphics[width=0.98\columnwidth]{KliquidELF.pdf}
\caption{Model fitted to potassium. (upper) The phase diagram based on two state model used parameters $\Delta U=1.15$, $\Delta V=1$, $J=1$ , $\Delta S=-0.5$, the same in both solid and liquid phases. The $x$-independent free energy difference between solid and liquid phases are
$\Delta G_{sl}= 0.12+0.24P-0.33T$.
(lower) Sequence of images showing the transition to electride state in fcc and liquid potassium. Figure shows electron localisation function\cite{schmider2000chemical} (ELF) calculated using CASTEP density functional code\cite{jr:CASTEP}.
Red shows region of high ELF and localised charge (see scale bar for values). The solid electride site is located at the ($\frac{1}{2},\frac{1}{2},\frac{1}{2})$ position - centre of the figure. The "liquid" image is a slice from a molecular dynamics snapshot\cite{zongK} cutting through several atomic sites (red circles at 0GPa - not all lie in the plane.), and then rerun at several densities using the same fractional atomic positions. the intersitial regions, initially green (free electronlike) become increasing red and blue as the electrons localise.
\label{fig:KLF}}
\end{figure}
\begin{figure}[ht
\includegraphics[width=0.88\columnwidth]{phaseline.pdf}
\caption{Phase diagram for the combined solid-liquid two-site model with $\Delta V=1$, $\Delta U =1$, $\Delta S =0$, J=1 in both phases, and a linear free energy difference $\Delta G_{sl}=0.02+0.04P-0.05T$
for the non-anomalous contributions. Data was collected by scanning a dense grid in P-T space and plotting points where the free energy difference was less than 0.0005, in reduced units. Inset shows variation of x across the phase diagram, from blue (x=0) to red (x=1) }
\label{fig:melt}
\end{figure}
\acknowledgements{HZ and GJA acknowledge the ERC project HECATE for funding. HZ thanks the National Natural Science Foundation of China (51931004 and 51871177). Computing resources were obtained via the UKCP EPSRC grant EP/P022561/1. We thank X. Ding and H Ehteshami for useful discussions.}
\clearpage
\bibliographystyle{apsrev}
|
\section{Introduction} \label{sec:intro}
The complexity of the electrical energy system is increasing with higher numbers of distributed devices for energy consumption, generation and storage. As an example, between 2007 and 2017, the European installed capacity of renewable energy almost doubled from 258 GW to 512 GW, primarily coming from new photovoltaic and wind generators \cite{Child2019}.
The centralised architecture of electricity dispatch, generation and distribution is transitioning toward a more distributed orchestration of control, optimization and market services for the balancing of supply and demand. The Internet of Things (IoT) technology has seen massive deployment of inter-connected sensing and control devices through residential, commercial and industrial settings.
The growing volume and heterogeneity of data enables novel data services to support the need for automated, data-driven decision making that is ultimately required to coordinate such increasingly complex energy system.
Modelling the electrical grid is the foundation of many of the required smart data services for energy utilities. An energy system model provides for a digital, mathematical representation of the collective behavior of the complex set of interconnected assets, from consumers and generators up to distribution transformers and electrical substations through power lines and power-flow control devises. Traditionally, this model is obtained from physics-based sets of power flow equations \cite{Abur2004} and informs several processes. Predictions (simulations) of the impact on the grid assets of observed and predicted (or desired) energy consumption and generation profiles can be assessed. Furthermore, the full state of the grid, that is the value of all electrical quantities at the grid assets, can be inferred from any available sensor data through a model inversion process known as state estimation \cite{Monticelli2000}. Asset and sensor data anomalies can also be diagnosed based on statistical analysis of the deviation between the model inference and the sensor observations, a process known as residual analysis \cite{Abur2004}.
The increasing complexity of energy systems, however, makes the derivation of a sufficiently detailed and accurate mathematical model from first principles, with well-defined values for its electrical parameters, an often prohibitive task. The issue is particularly evident for medium- to low-voltage distribution grids, where the growing need for accurate modelling is more and more critical and, at the same time, up-to-date physical models and detailed electrical parameters are often either not available or inaccurate. At the same time, the increasing availability of sensor data from smart metering infrastructure or IoT devices, makes data-driven modelling approaches based on machine learning more appealing. Among others, solutions based on feed-forward neural networks and auto-encoders have been studied to solve power systems state-estimation and prediction \cite{Barbeiro2014,Zamzam2019}. Not accounting for the domain knowledge often accompanying the sensor-data, in the form of grid connectivity of physics constraints, is however a strong limiting factor in the accuracy, transparency and, ultimately, practical applicability of these methods \cite{Zhang2019}. Knowledge-based systems have recently emerged to enrich IoT sensor data with a semantic reasoning layer that supports explicitly embedding domain knowledge into several machine learning tasks such as data exploration, feature engineering and model creation\cite{Schumilin2017,Zhang2017,Eck2019,Chen2018,Ploennigs2017}.
A promising research direction to create data-driven models for power systems that can be explicitly informed from knowledge-based systems is the use of probabilistic graphs \cite{Hu2011,Cosovic2016a,Fusco2018}. Probabilistic graphical models provide for a natural way to represent structural relationships between the variables of a complex system based on domain knowledge such as electrical grid connectivity, known localised correlation structure or even physical equations. In order to overcome limitations of the existing approaches, requiring costly belief propagation algorithms often designed based on ad-hoc heuristics to solve the inference problem for specific structures of the graph or functional form of the probability densities, a fully data-driven approach is proposed here, based on the recently emerged framework of graph neural networks (GNNs) \cite{Scarselli2009,Casas2020}. The proposed GNN model was developed in order to enable congestion management and market bidding data services for a distribution system operator (DSO) participating in energy flexibility markets.
After an overview, in Section \ref{sec:architecture}, of the data-analytic services that were developed in the context a smart-grid research project\cite{goflex}, Section \ref{sec:model} details the proposed modelling approach based on GNNs, underlying these services. Results from experiments during the live deployment of the data services at a pilot demonstration of the flexibility market are then discussed in Section \ref{sec:results}. Conclusions and final remarks are given in Section \ref{sec:conclusion}.
\section{Data Services for Energy Flexibility markets} \label{sec:architecture}
A set of smart-grid technologies to enable an energy flexibility market was developed within the context of the research project GOFLEX, funded by the European Union and involving a consortium of energy utilities,
technology providers and research institutions across Europe \cite{goflex,Eck2019eenergy}.
The flexibility market enabled residential or industrial electrical prosumers (consumers and producers) to actively participate in the energy system by offering to sell the flexibility in their energy production and/or consumption processes. Distribution system operators (DSOs) offered to buy the available flexibility from the market in order to solve potential issues that might be caused by excessive distributed renewable energy generation, thus increasing the grid capacity margin without the need for expensive capital investments.
A set of specific data-analytic services, as shown in Fig. \ref{fig:data-services}, were developed in order to enable the participation of energy utilities, such as DSOs, in the flexibility markets. Grid congestion prediction services estimate the future behavior, over a 24-48 hours window, of a number of electrical quantities of interest at important grid assets, along with the likelihood of them operating outside some user-defined threshold. Where instances of congestions are flagged, market bidding services estimate the corresponding amount of energy flexibility required at different points of the grid in order to avoid the predicted congestions. These specific use-cases of grid prediction and bidding data services are built on top of core machine-learning functionalities, including a model of the electrical grid as well as distributed energy forecasting models of demand and renewable generation. External data services, such as high-resolution weather predictions and grid sensor data from IoT sensors or traditional SCADA systems ultimately drive the data services. The sensor data are enriched by the semantic layer of a knowledge-based system that incorporates available domain expertise in the form, for example, of known relations between grid asset, variables and the sensors.
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{data-services-v1}
\caption{Energy data services based on knowledge- and data-driven machine-learning in the context of energy flexibility markets. }
\label{fig:data-services}
\end{figure}
A cloud-based architecture was designed to enable the data services of Fig. \ref{fig:data-services}, in the context of the smart-grid demonstration pilots of the project in \cite{goflex}. Following a micro-services design pattern, the knowledge-based time-series management micro-service was hosted on a combination of relational and graph database cloud storage services. The machine-learning modelling services for energy forecasting and grid modelling relied on a combination of containerised (orchestration and time-consuming tasks such as model training) and serverless (frequent, bursty workloads such as computing model predictions) cloud-computing infrastructure . Data ingestion and communication, both internally and with external consumers through high-level data services (e.g. the congestion predictions and market bidding) relied on asynchronous messaging based on Message Queuing Telemetry Transport (MQTT) and Advanced Message Queuing Protocol (AMQP) and on serverless computing. Further architecture details were provided in \cite{Eck2019,Chen2018}.
The benefits of knowledge-based time-series systems for managing and automating several aspects of the deployment of energy forecasting machine-learning models in large industrial-scale deployments were detailed in \cite{Eck2019}. In the following, the paper focuses on the specific development of the grid modelling services, which were based on a novel machine-learning model of the electrical grid that can factor in both the sensor data as well as the available domain knowledge, specifically in the form of power network connectivity, by leveraging the graph neural network framework.
\section{Proposed Graph neural network model} \label{sec:model}
A power systems model can be generally expressed as \cite{Abur2004}:
\begin{equation}
y_t = f(x_t) + \varepsilon_t ,
\label{eq:powerflow}
\end{equation}
where: $y_t$ is a set of electrical quantities of interest at a given time $t$, such as active/reactive power, current magnitude, voltage magnitude and angles, etc.; $x_t$ is the state variable and it denotes the minimum independent set of variables that fully describe the electrical system (a typical choice for the state variables in physics-based models is the set of voltage magnitudes and angles at all nodes of the grid, but it might be considered as a latent variable in a machine-learning-based model); $f(\cdot)$ is the set of, generally non-linear, power-flow equation relating the state variable to all electrical quantities of interest; $\varepsilon_t$ is an error term that quantifies modelling uncertainty or sensor noise.
By taking into account the structure of the energy system, the joint density of the model variables in (\ref{eq:powerflow}), can be factorized as follows:
\begin{equation}
p(y,x) = \prod_{i=1}^n \phi_i (y_i | x_i) \prod_{j=1}^{\mathcal{N}(i)} \psi_{ij} (x_i,x_j),
\label{eq:factorized}
\end{equation}
where $y_i$, $x_i$, for $i=1,\ldots n$, are $n$ subsets of the system and state variables, $\phi_i(\cdot)$ are conditional densities and $\psi_{ij}(\cdot)$ are joint densities between each state variable $x_i$ and its neighbours $\mathcal{N}(i)$.
Under Gaussian assumption, the densities $\phi(\cdot)$ and $\psi(\cdot)$ can be further specified as:
\begin{equation}
\begin{aligned}
&\phi_i(y_i|x_i) \propto exp\left(\frac{1}{2}\left[y_i-g_i(x_i)\right]^T\Sigma_i\left[y_i-g_i(x_i)\right]\right) \\
&\psi(x_i,x_j) \propto exp\left(\frac{1}{2}\left[x_{ij}-h_{ij}(x_i,x_j)\right]\Omega_{ij}\left[x_{ij}-h_{ij}(x_i,x_j)\right]\right),
\end{aligned}
\label{eq:factorized_gaussian}
\end{equation}
where $g_i(x_i)$ and $\Sigma_i$ denote, respectively, the conditional mean and covariance of $y_i$ with respect to $x_i$, while $h_{ij}(x_i,x_j)$ and $\Omega_{ij}$ denote the mean and covariance of the combined state variable $x_{ij}^T=[x_i^T; x_j^T]$.
The factorization in (\ref{eq:factorized}) effectively defines a probabilistic graph where the nodes are the variables $y_i$, $x_i$ and the edges are defined by both the conditional (direct edges) and joint densities (indirect edges) \cite{Koller2009}. In the specific case of energy grids, the structure of the graph can reflect knowledge about the physical network connectivity but also known localised correlation structure between the system variables. First-principle physical equations from (\ref{eq:powerflow}) can also be incorporated into the graphical model, for example as conditional mean function $g_i(x_i)$ of the Gaussian density in (\ref{eq:factorized_gaussian}).
Solving prediction, simulation or inversion problems on the graphical model (\ref{eq:factorized})-(\ref{eq:factorized_gaussian}) involves solving the inference problem based respectively on predictions, assumptions or observations of $y_i$. Inference algorithms on such graphical models take the form of message-passing belief propagation, where information is iteratively exchanged between the nodes across the edges of the graph. In the particular case, considered here, of Gaussian graphical models, message passing takes the form of the sum-product algorithm where updates involve matrix summations and multiplications propagating knowledge about mean and covariance of the distributions. Different forms of the sum-product algorithms have been proposed, depending on the specific format of the conditional or joint densities and even on the structure of the graph \cite{Cosovic2016a,Fusco2018}. By adopting the most general representation based on factor graphs \cite{Loelinger2007, Koller2009}, the probabilistic inference of physics-based power flow equations using Gaussian belief propagation \cite{Hu2011,Cosovic2016a} as well as the data-driven learning of localised relationships using neural network as nodes in the graph \cite{Fusco2018} has been investigated. Although Gaussian belief propagation, in the form of the sum-product message passing algorithms, solves probabilistic inference more efficiently than centralised problems not accounting for the graph structure \cite{Fusco2018}, it cannot be completely parallelised and might require expensive iterations for solving loopy graphs \cite{Cosovic2016a}. Furthermore, convergence of the belief propagation in the presence of non linear nodes and loops in the graph is not guaranteed and might require ad-hoc heuristic adjustments to the sum-product algorithm, which might depend on the specific structure of the model \cite{Cosovic2016a}.
\subsection{Message-passing graph neural networks} \label{sec:model_mpnn}
The general framework of graph neural networks (GNNs) incorporates graph structured information in a supervised machine learning model by encoding topological relationships among the nodes of the graph \cite{Scarselli2009}. In particular, each node encodes information about some concept, which is ultimately defined by its features and related concepts, whereby the relationships are encoded by the edges of the graph. By defining, for each node $k$ of the graph, a state vector, $x_k \in \mathbb{R}^p$, observations about a concept and its features, $y_k \in \mathbb{R}^q$, and an output of interest, $o_k \in \mathbb{R}^r$, a GNN model can be generally written as \cite{Scarselli2009}:
\begin{align}
x_k &= f_k(y_k,x_{\mathcal{N}(k)},y_{\mathcal{N}(k)}) \label{eq:gnn_local_transition}\\
o_k &= g_k(x_k,y_k), \label{eq:gnn_local_output}
\end{align}
for $k=1,\ldots n$ nodes in the graph. Based on (\ref{eq:gnn_local_transition}), the state vector at a node is related to the data available at the node itself and to both the state and data at neighbour nodes $\mathcal{N}(k)$ through a local transition function $f_k(\cdot)$. In (\ref{eq:gnn_local_output}), the desired target at a node is related to both the data and state vectors at the same node through a local ourput function $g_k(\cdot)$.
Among the many possibilities of further specifying the GNN model in (\ref{eq:gnn_local_transition})-(\ref{eq:gnn_local_output}), as reviewed in \cite{Battaglia2018}, message passing neural networks (MPNN) provide a suitable framework to solve inference and learning problems on the probabilistic graph in (\ref{eq:factorized_gaussian}) such to emulate principled belief-propagation algorithms \cite{Gilmer2017}. In a MPNN, the local transition function in (\ref{eq:gnn_local_transition}) is expressed as the following iteration:
\begin{align}
x_k^{(0)} &= f_k^i(y_k) \label{eq:gnn_encode} \\
m_{kj}^{(t)} &= f_{kj}^e (x_k^{(t)},x_j^{(t)}) \quad \forall j \in\mathcal{N}(k) \label{eq:gnn_edge} \\
x_k^{(t+1)} &= f_k^n (m_{kj}, \forall j \in \mathcal{N}(k)) \label{eq:gnn_agg},
\end{align}
where $f_k^i(y_k)$ is a node encoding function, $m_{kj}^{(t)}$ is the message, at iteration step $t$, exchanging information between nodes $k$, $j$ through the edge functions $f_{kj}^e(\cdot)$, and $f_k^n(\cdot)$ is a node aggregation function that updates the state vector at node $k$ based on all incoming messages from neighbouring nodes. The iterative scheme in (\ref{eq:gnn_edge})-(\ref{eq:gnn_agg}) defines a finite-steps message passing algorithm that propagates information through the graph neural network. Note that with respect to the general expression in (\ref{eq:gnn_local_transition}), the explicit dependency of the message passing algorithm on the node concept and features $y_k$ was dropped. It is assumed, in (\ref{eq:gnn_encode}), that the state vector fully encodes the information available in the observations and features available at the node, through the encoding function $f_k^i(\cdot)$. By implementing the sequence of encoding step (\ref{eq:gnn_encode}) and message passing steps (\ref{eq:gnn_edge})-(\ref{eq:gnn_agg}) with neural networks one can obtain a fully differentiable function that can be conveniently trained through back propagation and can offer arbitrarily high capacity of functional approximation.
\subsection{Proposed architecture} \label{sec:model_specific}
The Gaussian graphical model for a networked energy system, defined in (\ref{eq:factorized_gaussian}), is expressed here within the MPNN model framework outlined in Section \ref{sec:model_mpnn}, such that both the non-linear conditional and joint density functions, as well as the belief propagation algorithm, can be jointly modelled with neural networks and trained from standard back propagation.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{mpnn}
\caption{Inference in the proposed MPNN, based on (\ref{eq:gnn_encode}) through to (\ref{eq:my_encoder}).}
\label{fig:mpnn}
\end{figure}
The local output functions in (\ref{eq:gnn_local_output}) are used to express both the conditional mean, $\mu_{y_k}$, and variance, $\Sigma_{y_k}$, of the variables at a node $k$ of the graph, namely $y_k$, as function of the state vector at the node:
\begin{equation}
\begin{aligned}
\mu_{y_k} &= g^\mu_k(x_k) \\
\Sigma_{y_k} &=g^\Sigma_k(x_k),
\end{aligned}
\label{eq:my_decoder}
\end{equation}
where $g^\mu_k$ and $g^\Sigma_k$ are both fully differentiable parametric functions implemented as neural networks. In (\ref{eq:my_decoder}), only the variances of the individual components of $y_k$ are modelled with the underlying assumption that its covariance matrix is purely diagonal. Note that the output function as written in (\ref{eq:my_decoder}) can also be interpreted as a \emph{decoding} function mapping the internal state representation space to the target space. Accordingly, the node \emph{encoding} function in (\ref{eq:gnn_encode}) can be more specifically written as:
\begin{equation}
x_k^{(0)} = f_k^i (\mu_{y_k},\Sigma_{y_k}),
\label{eq:my_encoder}
\end{equation}
which provides for the initial iteration of the message passing algorithm in (\ref{eq:gnn_edge})-(\ref{eq:gnn_agg}).
Figure \ref{fig:mpnn} visualises the overall inference algorithm in the proposed MPNN model as a chain of neural network functions $f_k^i$, $f_{kj}^e$, $f_k^n$,$g_k^\mu$, $g_k^\Sigma$. Note that no iterative algorithm is required, as opposed to principled belief propagation, and only one feedforward pass through the chain generates the required inference estimate. The parameters of the individual functions can be trained by minimising, with any gradient-based algorithm, the following negative log-likelihood objective:
\begin{equation}
\mathcal{L} \propto \sum_{t=1}^T \sum_{k=1}^N \dfrac{1}{2}\log|\Sigma_{y_k}^t| + \dfrac{1}{2}(y_k^t-\mu_{y_k}^t)^\top \Sigma_{y_k}^{t^{-1}}(y_k^t-\mu_{y_k}^t),
\label{eq:nll}
\end{equation}
based on the data, $y_k^t$, available at all nodes $k$ an training samples $t$.
Section \ref{sec:model_missingdata} further specifies how the model can be applied in the context of the energy data services described in Section \ref{sec:architecture} in the form of data imputation problems.
\subsection{Handling missing data} \label{sec:model_missingdata}
The modelling approach defined by (\ref{eq:my_decoder}) and (\ref{eq:my_encoder}) can be thought of as graph-structured variational auto-encoder \cite{Boquet2019}, where mean and covariance of the variables of the energy network are modelled as function of a lower-dimensional latent variable. The proposed model structure provides for a general purpose modelling tool that can be used to solve the problems of prediction or state estimation that are typical in the context of electrical power systems, as reviewed above in section \ref{sec:model}, by framing them as missing data imputation problems.
In the specific application proposed in section \ref{sec:results}, the same model serves two different data services. Firstly, the model is used to predict the voltage magnitude at some nodes of the energy system based on predictions about the energy consumption and generation available at some other nodes. In the second use-case, the model is also used to determine the energy consumption or generation variation required at some nodes of the network in order to maintain the voltage at certain nodes at a desired reference. In both cases, the desired prediction or estimation can be obtained by running model inference where only a subset of the data at the variable nodes is assumed known. From a random initialization of the unknown variables, the sequence of encoding, message passing and decoding steps will provide an updated value for all variables. This imputation procedure can be iterated until convergence simulating a Markov chain, which has been shown that converges to the true marginal distribution of missing values given observed values \cite{Boquet2019}. Similarly, the same model can provide for the basis of other typical problems in energy systems such as state estimation of optimal power flow \cite{Barbeiro2014,Zamzam2019}.
\section{Results} \label{sec:results}
The GNN model outlined in section \ref{sec:model} was developed to inform data-analytics services, such as congestion prediction and market bidding, for electrical system operators trading in energy flexibility markets, as overviewed in Section \ref{sec:architecture}.
In the following, the results obtained from a live deployment of the proposed data services at a real-world demonstration site are discussed. Section \ref{sec:results_data} details the available data and the modelling problem, with the chosen GNN architecture and alternative benchmarks. The validation of the model on the grid voltage prediction task are then detailed in section \ref{sec:results_voltage}. An example of the application of the same grid model to predict grid congestions and generate market bids for energy flexibility is then discussed in section \ref{sec:results_bidding}.
\subsection{Available Data and Model Architecture} \label{sec:results_data}
A smart-grid pilot deployment of an energy flexibility market, as part of the research project \cite{goflex}, was available at the Electricity Authority of Cyprus, the operator of the electrical distribution network in Cyprus. As also overviewed in Section \ref{sec:architecture}, the project sought to develop data analytic-services for predicting localised congestions on the grid, in the form of voltage violations due to excessive distributed renewable generation, and eventually prevent them by issuing bids for purchasing energy flexibility on the market \cite{Eck2019eenergy}.
Live data were collected from July 2018 through to the end of 2019. Figure \ref{fig:data-ingestion} shows the rate of data ingestion from January through March 2019, where, on average, 15 million readings were received monthly (nearly 1.4K per hour) from about 500 sensors.
Specifically, the received data consisted of voltage and energy profile data at single- and three-phase electrical prosumers (consumers and producers) collected from smart meters, as well as active power and current load at distribution substation and feeder heads obtained through the utility supervisory control and data acquisition (SCADA) system. Sensor data were received at different rates, both daily and hourly, as it is also clear from the ingestion patterns appearing in Fig. \ref{fig:data-ingestion}. Resampling and integration was applied to obtain 15-minute energy load time series data, as shown in Fig. \ref{fig:sample-load}, from the raw active power data at the distribution feeder and transformers, sampled at 1-minute resolution. Interpolation was used to derive 15-minute voltage time-series data at the prosumers, as in the example shown in Fig. \ref{fig:sample-volt}, from the irregularly sampled raw observations \cite{Eck2019}.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{cyprus-ingestion}
\caption{IoT sensor data ingestion \cite{Eck2019}.}
\label{fig:data-ingestion}
\end{figure}
\begin{figure}
\centering
\subfloat[Voltage magnitude observed at a three-phase prosumer. \label{fig:sample-volt}]{%
\includegraphics[width=0.5\textwidth]{sample-voltage}
}
\hfill
\subfloat[Active power energy load at distribution substation. \label{fig:sample-load}]{%
\includegraphics[width=0.5\textwidth]{sample-load}
}
\caption{Example of available data}
\label{fig:sample-data}
\end{figure}
Ultimately, time-series were available for the energy load at 15 substations and 25 corresponding feeder heads, and for the voltage at 28 prosumers, 7 of which were 3-phase prosumers thus yielding a total of 48 voltage time-series. Historical temperature and solar irradiance observations at 15-minute resolution were also collected at each of the distribution substations, through a high-resolution weather data service provided by The Weather Company \cite{TWC2018}. Based on grid topology information, the GNN model structure as shown in Figure \ref{fig:graph} was derived. Node encoding and decoding functions (\ref{eq:my_encoder}), (\ref{eq:my_decoder}) are defined to model sensor data and features relevant for each specific type of node. Temporal, autoregressive features for each time-series at 24-, 36- and 48-hour lags were considered to account for typical short-term seasonalities in electrical power consumption and generation patterns. Temperature and solar irradiance weather features at the substations nodes, along with autoregressive features at the same lags, are also included based on the availability of the weather data.
As a result, the voltage mean and variance is modelled at the 28 prosumer nodes of dimensionality 8 (24) for single-phase (three-phase) prosumers. The mean and variance of the energy load at the 25 feeders and 15 substations is modelled with feeder nodes of dimensionality 8 and substation nodes of dimensionality 24 (including the weather features). An additional global node was included, with dimensionality 8 representing the mean and variance of the aggregated energy load for the whole DSO pilot.
The GNN model architecture is completed by the message-passing functions along the edges, namely (\ref{eq:gnn_edge})-(\ref{eq:gnn_agg}), which are defined to reflect the grid topology hierarchy going from the global node down to the substations, through their feeders and finally to the prosumers. All encoding, decoding and edge functions are implemented as multi-layer feedforward neural networks. As shown later, one of the benefits of incorporating domain knowledge in the proposed structure of the GNN is that it provides a significant reduction (more than $80\%$) in the parameter space of a centralised neural network model with comparable accuracy.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{gnn-structure}
\caption{Structure of proposed GNN model for the electrical grid. Shown are the decoding functions (\ref{eq:my_decoder}) for the mean and variance at the voltage nodes (electrical prosumers) and at the energy nodes (feeders, substations, global grid node). The model is completed by encoding functions at the nodes, defined as in (\ref{eq:my_encoder}), and by message-passing functions (\ref{eq:gnn_edge})-(\ref{eq:gnn_agg}), at the edges.}
\label{fig:graph}
\end{figure}
The main use-case for the GNN model was the voltage prediction problem, which corresponds to the data imputation problem for the voltage time-series data at a given time $t$. It can be assumed that all autoregressive features, as well as weather and energy data at time $t$ are available.
\subsection{Experimental Set-up and Benchmarks} \label{sec:results_benchmarks}
The GNN, designed as discussed in Section \ref{sec:results_data}, was trained on 1 year of historical data, from July 2018 to the end of June 2019, using the the negative log-likelihood objective function defined in (\ref{eq:nll}). In order to avoid periods of excessive missing data, only samples with at most $10\%$ missing data were included in the training set, resulting in 27137 training samples. The training set was augmented with the same number of samples were the voltage is missing, in order to improve the performance on the voltage prediction task, as suggested in section \ref{sec:model_missingdata}. A gradient-descent training algorithm with a batch size limited to 5000 samples and based on the Adam optimizer \cite{Kingma2015}, with 0.01 learning rate, was utilised. Early stopping was employed to avoid overfitting based on results obtained on a 1-month cross-validation period, in July 2019. Different architecture choices for the number of layers of the encoding/decoding nodes and edge functions, for the size of the latent space at the nodes and for the number of message-passing iterations, were compared. A hyperbolic-tangent activation function was used throughout the experiments as found to give best convergence properties. Final test results were evaluated on data collected in August 2019.
Centralised models based on multi-layer feed-forward neural networks (MLP) and auto-encoding (AE) neural networks were used as reference benchmarks. The same activation function, gradient-descent algorithm and convergence criteria as for the GNN model were used for both model classes. Also, the same set of features and, specifically for the AE, the same overall latent-space dimension as for the GNN model were used. While alternative, more complex neural-network architectures exist, the proposed choice for the benchmarks reflects the current state-of-the art in power systems modelling \cite{Barbeiro2014,Zamzam2019} and, at the same time, is considered sufficient to support the main conclusions of this work.
\subsection{Voltage Prediction} \label{sec:results_voltage}
The GNN grid model, designed as in section \ref{sec:results_data}, was applied to the voltage prediction problem, as in the example of Figure \ref{fig:prediction-detail}. As discussed in section \ref{sec:model_missingdata}, variable prediction using the proposed GNN model corresponds to a data imputation problem, where inference on the graph is run while assuming that the voltage observations at each sample $t$ are unknown. From an initialization of the unknown data with a placeholder value, the GNN inference is iteratively run with the most recent estimate for the unknowns \cite{Boquet2019}. It was observed that after $5$ inference steps, the data imputation procedure consistently converged to negligible updates for the estimate of the unknowns.
Different architectures for the GNN, structured as in Fig. \ref{fig:graph}, were explored by varying the number of layers for the MLPs implementing the node and edge functions, and the number of message-passing (MP) iteration steps. The dimension of the latent space was fixed to $6$ for the voltage and feeder load nodes (with dimensionality $8$) and to $24$ for the substation nodes and three-phase voltage nodes (with dimensionality $24$), as no significant gains were obtained with increasing values. This finding is consistent with practical domain intuition: energy and voltage data are correlated with the 3 lagged features; similarly, some cross-correlation at voltage between different phases and between weather and energy load allows for further compression in latent space of substation load nodes and three-phase prosumer nodes.
\begin{table}[ht]
\centering
\begin{tabular}{l||ccc|ccc}
\hline
\rule{0pt}{2ex}
Model &\# Layers &\# MP &\# Params & MAPE & RMSE \\
\hline
GNN &$2$ &$5$ &$143,464$ &$0.81\%$ & $2.42$ \\
GNN &$3$ &$5$ &$169,144$ &$0.82\%$& $2.45$ \\
GNN &$2$ &$8$ &$142,832$ &$0.80\%$ &$2.40$\\
GNN &$3$ &$8$ &$169,144$ &$0.81\%$& $2.41$ \\
MLP &$2$ &$-$ & $1,271,440$ & $0.80\%$ & $2.40$\\
MLP &$3$ &$-$ & $2,118,760$ & $0.73\%$ & $2.21$\\
AE &$2$ &$-$ & $2,407,088$ & $0.72\%$ & $2.17$ \\
AE &$3$ &$-$ & $3,796,840$ & $0.70\%$ & $2.13$ \\
\hline
\end{tabular}
\vspace{0.2cm}
\caption{Summary test results for GNN, MLP and AE models. Number of layers represent encoding/decoding and edge functions in GNNs and AE. MP denotes message-passing steps in GNNs. }
\label{tab:results_voltage}
\end{table}
Table \ref{tab:results_voltage} reports the error metrics, in terms of mean-absolute-percentage error (MAPE) and root-mean-square error (RMSE), for different choices of the hyper-parameters in the GNN and benchmark models (MLP and AE). It is interesting to see how a comparable accuracy is obtained with respect to the MLP and AE models, but at a more than $80\%$ reduction in the number of neural-network weights. By exploiting sparsity conditions that are readily available from basic types of domain knowledge, such as grid connectivity in this case, GNNs allow for much more efficient machine learning models without sacrificing modelling accuracy. While it is possible to enforcing sparsity in traditional MLP or AE models, for example based on heuristic search over the layers connectivity space, the computational complexity makes this often prohibitive in practice.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{prediction-detail}
\caption{Probabilistic voltage magnitude prediction at one of the prosumers. The shaded area is the $95.4\%$ confidence interval (2 standard deviations) around the predicted mean. }
\label{fig:prediction-detail}
\end{figure}
\begin{table}[ht]
\centering
\begin{tabular}{l||ccccc}
\hline
\rule{0pt}{2ex}
Missing data [$\%$] &$0$ &$0.1$ &$1$ &$5$ &$10$ \\
\hline
Samples &$187$ &$356$ &$1108$ &$1472$ &$192$ \\
MAPE [$\%$] & $0.810$ &$0.817$ &$0.815$ &$0.824$ &$0.854$\\
\hline
\end{tabular}
\vspace{0.2cm}
\caption{Impact of missing data on MAPE of GNN model from Table \ref{tab:results_voltage} with 2-layer MLP functions and 5 message-passing steps. }
\label{tab:results_voltage_missing}
\end{table}
The live data feed during the course of the demonstration project was affected by instances of missing data, often at random, as it also appears from some gaps in the actual data of Fig. \ref{fig:prediction-detail}. The performance of the proposed GNN model on the voltage prediction task was also evaluated with respect to the rate of missing data. Note that since, based on our formulation, the prediction problem is treated as a data imputation problem, no further change to the inference algorithm is required in order to handle missing values for the features. As summarized in Table \ref{tab:results_voltage_missing}, most of the samples in the test period (August 2019) had a missing data rate of $0.1-1\%$ (1108 samples) and $1-5\%$ (1472). While a minor degradation is observed with respect to the $187$ samples with no missing data, the performance remains fairly stable even for instances of $5-10\%$ missing data ratio.
\subsection{Congestions and market bidding} \label{sec:results_bidding}
The same energy system GNN model, validated on the voltage prediction problem in section \ref{sec:results_voltage}, was utilised to inform congestion prediction and bidding data services, as outlined in section \ref{sec:architecture}. The probabilistic voltage predictions generated by the model were used to design statistical tests for determining the likelihood of an occurrence of a voltage congestion based on user-defined criteria. As an example, Fig. \ref{fig:example-bidding} (top) shows examples of congestions flagged as voltage exceeding a threshold of $240\,V$, based on a Z-test with probability 0.8413 (mean estimate exceeds the threshold by 1 standard deviation). The amount of energy variation required at a given point of the grid in order to avoid the congestion is then estimated by running inference on the GNN model with the affected voltage variables set to the desired maximum value and with the energy variables at the corresponding substation and feeder nodes treated as missing points. Figure \ref{fig:example-bidding} (bottom) shows a comparison between the resulting estimate of the energy load and its actual value, their difference representing the amount of energy variation required to maintain the voltage below the desired threshold. Based on this difference, a bid on an energy flexibility market can be placed by the grid operator to purchase a corresponding increase or reduction in energy.
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{example-bidding}
\caption{Example of voltage congestion prediction service (top) and market bidding service, based on the estimate of energy flexibility (bottom) required to avoid the congestion. }
\label{fig:example-bidding}
\end{figure}
Unlike for the voltage prediction service, a quantitative validation of the bidding service was not available based on the smart-grid pilot demonstration set-up. Specifically, a ground-truth value for the estimated energy variation can not be physically measured and the amount of energy flexibility available within the market was too small to generate variations that are visible in the data. It can be, however, noted, based on the example in Fig. \ref{fig:example-bidding}, how the model suggests increasing the energy load (increase demand or reduce generation) to reduce the voltage profile, which is qualitatively consistent with the physics of power flow models.
\section{Conclusion} \label{sec:conclusion}
A novel probabilistic graphical model for energy systems, based on the framework of graph neural networks (GNNs), was introduced. The model was used to inform several data-analytics services for electrical system operators, such as the prediction of grid congestions and the estimation of the amount of energy flexibility required in order to avoid such congestions. Both problems are treated as data imputation inference on the graph. The model was evaluated in the context of a smart-grid demonstration project, to predict and manage voltage congestions due to high levels of distributed solar generation.
With respect to existing probabilistic graphical models, the proposed approach can learn the belief propagation algorithm from the data, instead of relying on ad-hoc, heuristics-based rules that are hard to generalise. Furthermore, the sparsity in the GNN, informed by the domain knowledge, significantly reduces the number of model parameters when compared with traditional machine-learning models with similar accuracy. For example, when compared with a traditional MLP model, a reduction of more than $80\%$ in the number of neural-network weights is achieved.
The proposed GNN model can also efficiently adapt to changes in the electrical systems or in the available sensor data, by adapting the graph structure accordingly and requiring retraining only on the parameters of the affected portions of the graph. These topics will be scope for future research.
\section*{Acknowledgement}
The authors wish to acknowledge Ioannis Papageorgiou from the Electricity Authority of Cyprus for providing access to and understanding of the data used in this study.
This research has received funding from the European
Research Council under the European Unions Horizon 2020
research and innovation programme (grant agreement no.
731232).
\bibliographystyle{IEEEtran}
\balance
|
\section{Introduction}
Geometric packing and covering problems arise in a wide range
of natural applications. They also have a long history of
spawning many demanding (and often still unsolved) mathematical
challenges. These difficulties are also notable from
an algorithmic perspective, as relatively straightforward
one-dimensional variants of packing and covering are already $\NP$-hard \cite{garey1975complexity};
however, deciding whether a given set of one-dimensional segments can be packed
into a given interval can be checked by computing their total length.
This simple criterion is no longer available for two-dimensional, geometric
packing or covering problems, for which the total area often does not suffice
to decide feasibility of a set, making it necessary to provide an explicit packing or covering.
A recent result by Abrahamsen et al.~\cite{till2DER} indicates that these difficulties
have far-reaching consequences: Two-dimensional packing problems are $\exists\mathbb{R}$-hard,
so they are unlikely to even belong to \NP.
We provide a provably optimal answer for a natural and previously unsolved
case of \emph{tight worst-case area bounds}, based on the notion of \emph{critical packing density}:
What is the largest number $\delta\leq 1$, such that any set
of squares with a total area of at most $\delta$ can always be packed (in a not necessarily
axis-parallel fashion) into a disk $C$ of area 1, regardless of the individual sizes of the squares?
We show the following theorem that implies $\delta=\nicefrac{8}{5\pi}\approx0.509$ for squares in a disk.
\begin{restatable}{theorem}{crit}\label{thm:mainAlgo}
Every set of squares with a total area of at most $\nicefrac{8}{5}$ can be packed into the unit disk.
This is worst-case optimal, i.e., for every $A > \nicefrac{8}{5}$ there exists a set of squares with total area $A$ that cannot be packed into the unit disk.
\end{restatable}
This critical density $\delta$ is of mathematical importance, as it settles the last open case of
packing circular or square objects into a circular or square container.
\cref{fig:overview} provides an overview of the critical densities in similar settings, i.e.,
the critical density for packing squares in a square ($\nicefrac{1}{2}$),
disks in a square ($\nicefrac{\pi}{(3+2\sqrt{2})}\approx 0.539$),
and disks in a disk ($\nicefrac{1}{2}$).
\begin{figure}[b!]
\centering
\includegraphics[page=2]{figures/ourANDdisksindisks.pdf}
\caption{Worst-case optimal approaches and matching worst-case instances for packing: \newline
(a) Squares into a square with \textsc{Shelf Packing}\xspace by Moon and Moser~\cite{MM1967some}.\\
(b) Disks into a square by Morr et al.~\cite{morr2017split,fekete:splitpacking}.\\
(c) Disks into a disk by Fekete et al.~\cite{fks-pddow-19}.\\
(d) Squares into a disk [this paper].}
\label{fig:overview}
\end{figure}
This result is also of algorithmic interest, because it provides a simple sufficient criterion for feasibility.
Note that the previous results illustrated in \cref{fig:overview} benefitted from recursive subdivisions of a square container into
subpieces bounded by straight lines, or from recursion based on
the similarity of objects and container when both are disks; neither applies when objects are squares
and the container is a disk.
This gives our approach added methodical significance, as it showcases a general framework
for establishing computer-assis\-ted proofs for difficult packing problems for which concise manual
arguments may be elusive.
A proof of \cref{thm:mainAlgo} consists of (i) a class of instances that
provide the upper bound of~$\nicefrac{8}{5\pi}$ for the critical packing density $\delta$
and (ii) an algorithm that achieves the matching lower bound for $\delta$ by packing any
set of squares with a total area of at
most $\nicefrac{8}{5}$ into the unit disk. The first part is relatively simple:
As shown in \cref{fig:overview}(d), a critical configuration consists of
two squares of side length $s=\nicefrac{2}{\sqrt 5}$ and a disk~$\mathcal D$
of radius 1. It is easy to see that any infinitesimally larger square
(of side length $s+\varepsilon$ for any~$\varepsilon>0$) must contain the
center of $\mathcal D$ in its interior, so two such squares
cannot be packed. \new{Hence, we draw the following conclusion, showing that the bound in \cref{thm:mainAlgo} is tight.}
\new{\begin{lemma}
For every $\epsilon>0$, there exists a set of squares of total area $\nicefrac{8}{5}+\epsilon$ that cannot be packed into the unit square.
\end{lemma}}
The remainder of our paper focuses on the difficult part:
providing a strategy for packing sets of squares into a disk (described in \cref{sec:algorithm}),
and then proving that any set of squares with a total area of at most $\nicefrac{8}{5}$ can indeed be packed
into the unit disk.
This proof is set up with two sets of tools: In \cref{sec:interval-arithmetic-proofs}, we describe
a general technique that we employed for automated parts of our proof, while
\cref{sec:analysis-subroutines} provides a number of helpful lemmas.
\cref{sec:analysis-algorithm} gives the actual analysis of our algorithm.
\subsection{Related work: geometric packing}
Problems of geometric packing have been studied for a long time.
Providing a survey that does justice to the wide range of relevant work goes beyond the scope of this paper;
therefore, we strictly focus on very closely related results, in particular, concerning critical packing density.
We refer to Fejes Tóth~\cite{toth1999recent, toth20172}, Lodi, Martello and Monaci~\cite{LMM2002two},
Brass, Moser and Pach~\cite{brass2006research} and Böröczky~\cite{boroczky2004finite} for more comprehensive surveys.
Even the decision problem whether it is possible to pack a given set of squares into the unit
square was shown to be strongly $\NP$-complete by Leung et
al.~\cite{LTWYC1990packing}, using a reduction from \textsc{3-Partition}.
Already in 1967, Moon and Moser~\cite{MM1967some} proved that it is possible to
pack a set of squares into the unit square if their total area does not exceed
$\nicefrac{1}{2}$. This bound is best possible, because two squares even
infinitesimally larger than the ones shown in \cref{fig:overview}(a) cannot be
packed.
The proof is based on a simple recursive argument.
For the scenario with circular objects, Demaine, Fekete, and
Lang~\cite{DFL2010circle} showed in 2010 that deciding whether a given set of
disks can be packed into a unit square is $\NP$-hard.
Using a recursive procedure for partitioning the container into
triangular pieces, Morr, Fekete and
Scheffer~\cite{fekete:splitpacking,morr2017split} proved that the critical
packing density of disks in a square is $\nicefrac{\pi}{(3+2\sqrt{2})} \approx
0.539$.
More recently, Fekete, Keldenich and Scheffer~\cite{fks-pddow-19} established the critical
packing density of disks into a disk. Employing a number of algorithmic
techniques in combination with some interval arithmetic and computer-assisted case
checking, they proved that the critical packing density of disks in a disk is
$\nicefrac{1}{2}$; they also provide a video including an animated overview~\cite{bfk+-pgoow-19}.
In a similar manner, Fekete et al.~\cite{fgk+-wcocrd-20} established a closed-form description of the
total disk area that is sometimes necessary and always sufficient to cover a rectangle depending on its aspect ratio.
Note that the main objective of this line of research is to compute tight worst-case bounds.
For specific instances, a packing may still be possible, even if the density is higher;
this also implies that proofs of infeasibility for specific instances may be trickier.
However, the idea of using the total item volume for computing packing bounds can still
be applied. See the work by Fekete and Schepers~\cite{fs-ncflb-01,fs-gfbhd-04}, which shows how
a {\em modified} volume for geometric objects can be computed, yielding good lower bounds for one- or higher-dimensional
scenarios.
\subsection{Related work: interval arithmetic and computer-assisted proofs}
Establishing tight worst-case bounds for packing problems needs to overcome two main difficulties.
The first is to deal with the need for accurate computation in the presence of potentially complicated coordinates;
the second is the tremendous size of a full case analysis for a complete proof.
Developing methods for both of these challenges has a long tradition in mathematics.
One of the first instances of interval arithmetic
is Archimedes's classic proof~\cite{pi} that
$\nicefrac{223}{71} \leq \pi \leq \nicefrac{22}{7}$, establishing a narrow interval
for the fundamental constant of geometry.
This entails dealing with inaccurate
computation not merely by giving a close approximation, but by establishing an interval
for the correct value, which can be used for valid lower and upper bounds for subsequent
computations.
Employing electronic devices (e.g., calculators or computer algebra)
for mathematical arguments is a well-established method for eliminating tedious, error-prone manual calculations.
A famous milestone for the role of computers in theorem proving itself
is the confirmation of the Four Color Theorem, a tantalizing open problem for more than 100 years~\cite{wilson}.
While the first pair of papers by Appel and Haken~\cite{AH1,AH2}
was still disputed, the universally accepted proof by Robertson et al.~\cite{RSST} still relies on
extensive use of automated checking.
Another example is the resolution of the Kepler conjecture by Hales et al.~\cite{kepler_17}:
While the first version of the proof~\cite{kepler_05}
was still met with some skepticism, the revised and cleaned up
variant~\cite{kepler_17} fits the mold of a more traditional proof, despite
relying both on combinatorial results and computational case checking.
\new{Note that this proof uses a subdivision technique and interval arithmetic in a manner similar to the one used in this paper.
As in this paper, the result of Hales et al.~\cite{kepler_17} is tight in the numerical sense,
which means that due to the discretization error introduced by subdividing a space over $\mathbb{R}$ into finitely many pieces,
parts of the proof must be carried out by other means.}
Other instances of classic geometric problems that were resolved with the help
of computer-assisted proofs are a tight bound for the {E}rdős-{S}zekeres
problem for the existence of convex paths in planar sets of 17
points~\cite{szekeres:computer}, a precursor by Hass and Schlafly~\cite{hass2000double} to the proof of the double bubble theorem by Hutchings et
al.~\cite{hutchings:bubble} (the shape that encloses and separates two given
volumes and has the minimum possible surface area is a standard double bubble,
i.e., three spherical surfaces meeting at angles of $\nicefrac{2\pi}{3}$ on a common
disk), or the proof of $\NP$-hardness of finding a minimum-weight triangulation
(MWT) of a planar point set by Mulzer and Rote~\cite{mulzer:minimum}.
Further examples in the context of
packing and covering include a branch-and-bound approach for covering of
polygons by not necessarily congruent disks with prescribed centers and a
minimal sum of radii by Bánhelyi et al.\!~\cite{Bnhelyi2015OptimalCC}.
\new{Compared to some more extensive uses of computer assistance in proofs, such as for the four-color theorem,
in this paper, we use a fairly limited amount of computation to prove our results; we mainly use it to replace
an extensive explicit case distinction involving manual analysis and calculus that would be very hard to read
and verify.}
\section{A worst-case optimal algorithm}
\label{sec:algorithm}
Now we describe \textsc{Layer Packing}\xspace,
our worst-case optimal algorithm for packing squares into the unit disk~$\mathcal D$.
The basic idea is to combine refined variants of basic techniques (such as \textsc{Shelf Packing}\xspace) in several
directions, subdividing the packing area into multiple geometric layers and components.
\subsection{Outline of \textsc{Layer Packing}\xspace}
By $s_1,\dots,s_n$, we denote a sequence of squares and simultaneously their side lengths and assume that $s_1 \geq \dots \geq s_n$ is a sorted sequence.
\textsc{Layer Packing}\xspace distinguishes three cases that depend on the sizes of the first few, largest squares;
see \cref{fig:Alg} for illustrations.
\begin{description}
\item[\textbf{(C\;\!1)}\xspace] If $s_1 \leq 0.295$, we place a square of side length $\mathcal{X} = 1.388$ concentric into $\mathcal{D}$
and place one square of side length $\mathcal{X}_i= 0.295$, $i\in\{1,\dots,4\}$, to each side of $\mathcal X$, see \cref{fig:Alg}(a).
The four largest squares $s_1,\ldots,s_4$ are placed in these containers $\mathcal{X}_1,\ldots,\mathcal{X}_4$.
All other squares are packed into $\mathcal{X}$ using \textsc{Shelf Packing}\xspace.
\item[\textbf{(C\;\!2)}\xspace] If $s_1 \leq \nicefrac{1}{\sqrt{2}}$ and $s_1^2+s_2^2+s_3^2+s_4^2\geq \nicefrac{39}{25}$,
let $\mathcal{X}_1,\dots,\mathcal{X}_4$ be four squares of side length $\nicefrac{1}{\sqrt{2}}$ that are placed into $\mathcal{D}$ as depicted in \cref{fig:Alg}(b).
Furthermore, let $\mathcal{X}$ be a square of side length $\nicefrac{\sqrt{2}}{5}$ that can be packed into~$\mathcal{D}$ in addition to $\mathcal{X}_1,\ldots,\mathcal{X}_4$; see \cref{fig:Alg}(b).
For $i\leq 4$, $s_i$ is the only square packed into~$\mathcal{X}_i$; this is possible because $s_i\leq s_1\leq \nicefrac{1}{\sqrt{2}}$.
All other squares are packed into $\mathcal{X}$ using \textsc{Shelf Packing}\xspace.
\item[\textbf{(C\;\!3)}\xspace] In the remaining cases, we make extensive use of a refined shelf packing approach.
Specifically, the largest square $s_1$ is packed into~$\mathcal{D}$ as
high as possible, see Figures~\ref{fig:Alg}(c) and \ref{fig:TopBottom}.
The bottom side of $s_1$ induces a horizontal split of $\mathcal D$ into a \emph{top} and a \emph{bottom} part,
which are then filled by two subroutines called \textsc{Top Packing}\xspace and \textsc{Bottom Packing}\xspace, described in \cref{sec:subroutines}.
For each $i \geq 2$, we then
\begin{description}
\item[\textbf{(C\;\!3a)}\xspace] use \textsc{Top Packing}\xspace to pack $s_i$ if possible,
\item[\textbf{(C\;\!3b)}\xspace] else we use \textsc{Bottom Packing}\xspace to pack $s_i$.
\end{description}
\end{description}
\begin{figure}[h]
\centering
\includegraphics[page=3]{figures/spcase1.pdf}
\caption{
Illustration of the packings (a) in Case~\textbf{(C\;\!1)}\xspace,
(b) in Case~\textbf{(C\;\!2)}\xspace,
and (c) in Case~\textbf{(C\;\!3)}\xspace.}
\label{fig:Alg}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[page=1]{figures/outline.pdf}
\caption{In case \textbf{(C\;\!3)}\xspace, the largest
(hatched) square is packed topmost, inducing a top and a
bottom part of $\mathcal D$.
Subsequent (white) squares are packed into the pockets of the top part
with \emph{\textsc{Top Packing}\xspace} (using \textsc{Refined Shelf Packing}\xspace as a subroutine) if they fit;
if they do not fit, they are shown in gray and packed into the bottom part
with \emph{\textsc{Bottom Packing}\xspace}, which uses horizontal \textsc{SubContainer Slicing}\xspace, and vertical \textsc{Refined Shelf Packing}\xspace
within each slice.}
\label{fig:TopBottom}
\end{figure}
\subsection{Subroutines of \textsc{Layer Packing}\xspace}
\label{sec:subroutines}
\textsc{Layer Packing}\xspace employs a number of different subroutines.
\begin{description}
\item[\textsc{Refined Shelf Packing}\xspace] \label{sec:shelf} The greedy-type packing procedure \textsc{Shelf Packing}\xspace, employed by Moon and Moser~\cite{MM1967some}, packs objects by decreasing size; see the top of \cref{fig:overview}(a).
At each stage, there is a (w.l.o.g.~horizontal) straight cut that separates
the unused portion of the container from a ``shelf'' into which the next square is packed.
The height of a shelf is determined by the first packed object. Subsequent
objects are packed next to each other, until an object no longer
fits into the current shelf; in this case, a new shelf is opened on top of the previous one.
For \textsc{Layer Packing}\xspace, we use two modifications.
(1) Parts of the shelf boundaries may be circular arcs; however, we still have a supporting straight axis-parallel boundary and a second, orthogonal straight boundary.
(2) Our refined shelf packing uses the axis-parallel boundary line of a shelf as a support line
for packing squares; in case of a collision with the circular boundary,
we may move a square towards the middle of a shelf if this allows packing it. Note that this may only occur in shelves containing the horizontal diameter.
\item[\textsc{Top Packing}\xspace] The first and largest square $s_1$ is packed as high as possible into $\mathcal D$; see \cref{fig:segmentpacking}(a).
Then the horizontal line through the bottom of $s_1$ cuts the container into a \emph{top part} that contains $s_1$,
with two congruent empty pockets ${C}_\ell\xspace$ and ${C}_r\xspace$ left and right of~$s_1$; and a \emph{bottom part}.
Each pocket has two straight axis-parallel boundaries, $b_x$ and $b_y$.
By $\sigma$, we denote the largest square that fits into either pocket.
For large $s_1$, the bottom side of $\sigma$ does not lie on the same height as the bottom side of $s_1$;
in that case, we ignore the parts of ${C}_\ell\xspace$ and ${C}_r\xspace$ that lie below $\sigma$; see \cref{fig:segmentpacking}(e).
We use \textsc{Refined Shelf Packing}\xspace with shelves parallel to the shorter boundary among $b_x$ and $b_y$,
as shown in \cref{fig:segmentpacking}(b) and (c).
If a square does not fit into either pocket, it is packed
into the bottom part.
\begin{figure}[htb]
\centering
\includegraphics[page=3]{figures/sn1.pdf}
\caption{(a) Packing $s_1$ topmost into
$\mathcal{D}$ yields the top part of $\mathcal{D}$ with pockets ${C}_\ell\xspace$ and
${C}_r\xspace$, and the bottom part of $\mathcal{D}$. The bottom part is partitioned by \textsc{SubContainer Slicing}\xspace into
subcontainers~$C_i$, with heights corresponding to the first packed square.\\
(b) A pocket ${C}_\ell\xspace$ for which $b_x\leq b_y$ implies horizontal shelf
packing. \\
(c) A pocket ${C}_\ell\xspace$ for which $b_x> b_y$ implies vertical shelf
packing.\\
(d) Within each subcontainer~$C_i$, \textsc{SubContainer Packing}\xspace places squares along vertical shelves, starting from the longer straight
cut of $C_i$.\\
(e) For large $s_1$, we disregard the parts of $C_{\ell}$ and $C_r$ that lie below their inscribed square $\sigma$.}
\label{fig:segmentpacking}
\label{fig:slicing}
\end{figure}
\item [\textsc{Bottom Packing}\xspace]
\label{sec:bottom}
A square that does not fit into the top part of $\mathcal D$ is packed into the bottom part.
For this purpose, we use (horizontal) \textsc{SubContainer Slicing}\xspace, and (vertical)
\textsc{SubContainer Packing}\xspace within each subcontainer; see \cref{fig:TopBottom} for the overall picture.
\item[\textsc{SubContainer Slicing}\xspace]
For packing squares in the bottom part of $\mathcal D$, \textsc{SubContainer Slicing}\xspace subdivides $\mathcal D$ into smaller containers $C_i$,
by using straight horizontal cuts; see \cref{fig:slicing}(a).
The height of a subcontainer is determined by the first square packed into it.
\item [\textsc{SubContainer Packing}\xspace]
Within each subcontainer, we use \textsc{Refined Shelf Packing}\xspace with vertical shelves.
These shelves are packed from the longer of the two horizontal cuts, i.e., to pack~$C_i$, we start from the boundary
that is closer to the disk center; see \cref{fig:slicing}(d).
\end{description}
\section{Proofs based on interval arithmetic}
\label{sec:interval-arithmetic-proofs}
In interval arithmetic, operations like addition, multiplication or taking the
square root are performed on real intervals $[a,b] \subset \mathbb{R}$ instead
of real numbers.
When applied to intervals, an operation $[a_1,b_1] \circ [a_2,b_2]$ results in the smallest
interval that contains all possible values of $x \circ y$ for $x \in [a_1,b_1], y \in [a_2,b_2]$.
In a practical implementation on computers with finite precision,
computing the smallest such interval is not always possible.
However, using appropriate rounding modes or error bounds, it is still possible to compute an interval that over-approximates the resulting interval,
i.e., contains all possible outcomes of the corresponding real operation.
Predicates such as $[a_1,b_1] \leq [a_2,b_2]$ can also be evaluated on intervals.
The result is a subset of $\{\texttt{false},\texttt{true}\}$ containing all possible outcomes of $x \leq y$ for $x \in [a_1,b_1], y \in [a_2,b_2]$.
This allows evaluating quantifier-free formulas on the Cartesian product of intervals in an over-approximative way.
In many cases throughout this paper, we want to prove that a given non-linear system of real constraints over a bounded $k$-dimensional space $\mathcal{R}$ is unsatisfiable.
This space is typically spanned by a set of $k$ real variables.
Conceptually, to do this in an automatic fashion, we subdivide $\mathcal{R}$ into a sufficiently large number of $k$-dimensional cuboids.
Each such cuboid is defined by an interval for each of the $k$ real variables spanning $\mathcal{R}$.
We then apply interval arithmetic to each such cuboid $\mathcal{C}$ to find a set $S$ of constraints that together eliminate all points of $\mathcal{C}$,
thus proving that no point in $\mathcal{C}$ satisfies all our constraints for a counterexample.
\new{We use this simple technique of subdividing into cubes because it is sufficient for our case, relatively simple to manage and implement and thus arguably less error-prone than relying on deeper, more advanced methods. Moreover, interval arithmetic scales relatively well with the complexity and the number of constraints and can handle non-polynomial constraints involving functions such as $\arccos(x)$ that occur in some of our proofs.}
To improve the efficiency of this approach, our implementation of the basic concept is optimized in several ways.
For instance, the subdivision proceeds in a tree-like fashion according to a fixed ordering of the $k$ variables $v_1,\ldots,v_k$ spanning $\mathcal{R}$.
If variables $v_1,\ldots,v_j$ suffice to exclude a part of~$\mathcal{R}$, we do not split $v_{j+1},\ldots,v_k$ on that part.
Furthermore, we adaptively increase the local fineness of our subdivision if a coarser subdivision does not suffice for some part of $\mathcal{R}$.
Overall, this leads to a limited number of automated proofs\footnote{Source code available at \url{https://github.com/phillip-keldenich/squares-in-disk}.};
for some of these, manual checking would also be feasible, but would involve many case distinctions and would be tedious and unsatisfying.
\revised{Instead, we replace these proofs by the automatic procedure outlined above to have a clear structure instead of an otherwise overwhelming set of arguments.}
\revised{Combined, all automatic proofs required for this paper take less than 1.5 hours and less than 300 MB of memory on the 4 physical cores of the 2.3 GHz Intel i5-8259U CPU in one of the authors' laptops. The proofs involve up to $k = 9$ variables.}
\section{Analysis of subroutines}
\label{sec:analysis-subroutines}
In the following, we establish a number of bounds for the subroutines from \cref{sec:subroutines},
which we use to prove the performance guarantee for \textsc{Layer Packing}\xspace.
\new{To prove such a lower bound $\ell$ on the square area packed into some type of container $\mathcal{C}$ by \textsc{Layer Packing}\xspace or some subroutine $\mathcal{A}$,
we typically argue indirectly, by assuming that we have some sequence of squares $s_1,\ldots,s_n$ that $\mathcal{A}$ \emph{fails to pack}.
W.l.o.g., we always assume that such a sequence is \emph{minimal} in the sense that $\mathcal{A}$ successfully packs $s_1,\ldots,s_{n-1}$,
but placing $s_n$ according to $\mathcal{A}$ would result in placing a square such that it intersects the exterior of $\mathcal{C}$ or another square $s_i$.
We then establish our lower bound $\ell$ by showing that the area of $s_1,\ldots,s_n$ must be strictly greater than $\ell$.
In the remainder of the paper, whenever we say that \textsc{Layer Packing}\xspace or a subroutine \emph{fails to pack $s_n$ (or $s_1,\ldots,s_n$)},
we refer to such a sequence of squares, and always assume that $s_n$ is the first square our algorithm cannot pack.}
\subsection{Shelf Packing}
In several places we make use of the following classic result regarding \textsc{Shelf Packing}\xspace.
\begin{restatable}[\hspace{-0.2pt}\cite{MM1967some}]{lemma}{shelfpacking}\label{lem:shelfpacking}
\textsc{Shelf Packing}\xspace packs every sequence $t_1 \geq \dots \geq t_u$ of squares with a total area of at most $\nicefrac{1}{2}\cdot hw$ into an $h \times w$-rectangle with $t_1 \leq h \leq w$.
\end{restatable}
If the side length of the largest square is small compared to the size of the container, one can guarantee a higher packing density.
\begin{restatable}{lemma}{shelfpackingtwo}\label{lem:shelfpackingtwo}
Any finite set of squares with largest square $x_1<\nicefrac{1}{2}$ is packed by
\textsc{Shelf Packing}\xspace into a unit square, provided its total area is at most $\nicefrac{1}{2} + 2(x_1 - \nicefrac{1}{2})^2$.
\end{restatable}
\begin{proof}
We show that if a set of squares cannot be packed, then its area $A$ exceeds the bound of $\nicefrac{1}{2} + 2(x_1 - \nicefrac{1}{2})^2$,
\new{hence using the same type of indirect argument as outlined in the first paragraph of \cref{sec:analysis-subroutines}.}
To this end, we assume that the last square in the sequence cannot be packed by \textsc{Shelf Packing}\xspace, as illustrated in \cref{fig:shelfbound}(a).
%
\begin{figure}[htb]
\centering
\includegraphics{figures/HighDensity}
\caption{Establishing a refined density bound by shelf packing.
(a) The area of the first square in each shelf is accounted with the previous shelf.
(b) Illustration of the lower bound on the area packed in each shelf.}
\label{fig:shelfbound}
\end{figure}
%
We denote the height of the first square in shelf $i$ by~$x_i$.
Let $x_1+x_2+\cdots+x_k>1$ be the total height of the arrangement when the last square of the sequence cannot be placed in
a feasible shelf and is placed in an additional shelf that exceeds the height of the container, i.e., the last square has height $x_k$.
For computing the area of the packed squares, we account the first square of each shelf with the preceding shelf, see \cref{fig:shelfbound}(a). Then, for $i\geq 1$, in shelf $i$ the area of the packed squares is at least
\[x_{i+1}(1-x_i)\geq x_{i+1}(1-x_1)\] as depicted in illustrated by \cref{fig:shelfbound}(b).
We can conclude that the total packed area $A$ is
\[ A > x_1^2 + (x_2+x_3+\dots x_k)(1-x_1) > x_1^2 + (1-x_1)^2 = \nicefrac{1}{2}+ 2(x_1-\nicefrac{1}{2})^2.\hfill\qedhere\]
\end{proof}
\subsection{Top Packing}
\label{sec:anapocketpacking}
We prove the following lower bound on the square area packed by \textsc{Top Packing}\xspace, as long as at least one square fits into the left pocket ${C}_\ell\xspace$.
By $\sigma = \sigma(s_1)$, we denote the side length of the largest square that can be packed into ${C}_\ell\xspace$ or ${C}_r\xspace$; see \cref{fig:t-t1-sigmaApp}(c)--(e).
\begin{restatable}{lemma}{anapocketpacking}\label{lem:anapocketpacking}
Let $s_1 \geq \cdots \geq s_n$ be a sequence of squares for which \textsc{Layer Packing}\xspace fails to pack $s_n$.
If $s_n \leq \sigma$, then \textsc{Top Packing}\xspace packs squares of total area at least $0.83\sigma^2$.
\end{restatable}
Intuitively, the proof makes use of the \textsc{Shelf Packing}\xspace bound on the squares inscribed in ${C}_\ell\xspace$ and ${C}_r\xspace$,
but additionally uses the gaps in ${C}_r\xspace$ and ${C}_\ell\xspace$ to bound the square area packed into each of ${C}_\ell\xspace$ and ${C}_r\xspace$ by $0.415\sigma^2$.
Before presenting its proof, we make some helpful observations. We assume the center of our unit disk $\mathcal{D}$ lies at the origin $(0,0)$ of our coordinate system.
Recall that \textsc{Top Packing}\xspace packs the largest square $s_1$ as high as possible into $\mathcal{D}$. This implies that the center of $s_1$ is on the vertical line $x = 0$.
For some $u \in (-1,1)$, we denote by $T(u)$ the side length of the largest square with center on $x = 0$ and bottom on $y = u$ that fits into $\mathcal{D}$; see \cref{fig:t-t1-sigmaApp}(a).
\begin{figure*}[htb]
\centering
\resizebox{.99\textwidth}{!}{\includegraphics{figures/t_t1_sigma.pdf}}
\caption{(a) The definition of $T(u)$ and its defining equation. (b) The definition of $T^{-1}(s)$ and its equation.
(c) The pockets ${C}_\ell\xspace$ and ${C}_r\xspace$ used by \textsc{Top Packing}\xspace with their inscribed square $\sigma$ and its equation if $s_1 < s_1^*$,
(d) $s_1 = s_1^*$ and (e) $s_1 > s_1^*$. (f) The value $s_1' \neq s_1^*$ for $s_1$ for which both cases for $\sigma$ coincide.}
\label{fig:t-t1-sigmaApp}
\end{figure*}
The inverse function~$T^{-1}(s)$ of $T(u)$ describes the highest possible $y$-coordinate of the bottom side of a square of side length $s$; see \cref{fig:t-t1-sigmaApp}(b). Thus, \textsc{Top Packing}\xspace places the bottom-left corner of $s_1$ at $(-\nicefrac{s_1}{2}, T^{-1}(s_1))$; note that this can be below or above the center of $\mathcal{D}$.
Furthermore, recall that \textsc{Top Packing}\xspace packs the remaining disks into the pockets ${C}_\ell\xspace$ and ${C}_r\xspace$ induced by placing $s_1$; see \cref{fig:t-t1-sigmaApp}(c).
Now, we present explicit formulas. Solving the equations in \cref{fig:t-t1-sigmaApp}(a)--(b), we get
\[
T(u) = \nicefrac{2}{5} \cdot \left(\sqrt{5-u^2} - 2u\right)\quad \text{ and }\quad
T^{-1}(s) = \sqrt{1 - \nicefrac{1}{4}\cdot s^2} - s\text{.}\]
To compute $\sigma(s_1)$, we observe the following.
Below some threshold~$s_1^*$, the bottom side of the inscribed square of ${C}_\ell\xspace$ lies on the horizontal line $y = T^{-1}(s_1)$ and its top left corner touches~$\mathcal{D}$;
see \cref{fig:t-t1-sigmaApp}(c).
\new{In this case, the fact that our disk has radius $1$ implies that $\sigma$ satisfies the equation
\[\left(\nicefrac{s_1}{2} + \sigma\right)^2 + \left(T^{-1}(s_1) + \sigma\right)^2 = 1,\]
which has only one positive real solution $\sigma = \nicefrac{1}{4}\cdot\left(-s_1 - 2T^{-1}(s_1) + \sqrt{8 - (s_1 - 2T^{-1}(s_1))^2}\right)$.
For values $s_1 > s_1^*$ above that threshold, the center of ${C}_\ell\xspace$'s inscribed square lies on $y = 0$ and both left corners touch the disk; see \cref{fig:t-t1-sigmaApp}(e).
In this case, $\sigma$ has to satisfy the equation
\[\left(\nicefrac{s_1}{2} + \sigma\right)^2 + \nicefrac{\sigma^2}{4} = 1,\]
which has only one positive solution $\sigma = \nicefrac{1}{5}\cdot\left(\sqrt{20 - s_1^2} - 2s_1\right)$.
There are exactly two positive real values $s_1$ for which these two cases coincide;
the smaller one yields the threshold value
$s_1^* = \sqrt{\nicefrac{1}{3}\cdot (2+\sqrt{2})} \approx 1.0668$, see \cref{fig:t-t1-sigmaApp}(d);
the larger one is $s_1' = \sqrt{\nicefrac{1}{13}\left(11 + 6\sqrt{3}\right)} \approx 1.2828$ and corresponds to the situation depicted in \cref{fig:t-t1-sigmaApp}(f).
Overall, we obtain the following solution for $\sigma$:}
\begin{equation}\sigma = \begin{cases}
\nicefrac{1}{4}\cdot\left(-s_1 - 2T^{-1}(s_1) + \sqrt{8 - (s_1 - 2T^{-1}(s_1))^2}\right), & \text{if }s_1 \leq s_1^*,\\
\nicefrac{1}{5}\cdot\left(\sqrt{20 - s_1^2} - 2s_1\right), & \text{otherwise.}
\end{cases}
\label{eq:sigma}
\end{equation}
Now we are ready to present a proof of \cref{lem:anapocketpacking}.
\begin{proof}[Proof of \cref{lem:anapocketpacking}]
We begin by observing that $s_n \leq \sigma$ would fit into either ${C}_\ell\xspace$ or ${C}_r\xspace$; as we fail to pack $s_n$, ${C}_\ell\xspace$ and ${C}_r\xspace$ must contain other squares.
In the following, we prove that \textsc{Top Packing}\xspace packs squares of area $A \geq 0.415\sigma^2$ into ${C}_\ell\xspace$.
An analogous argument works for~${C}_r\xspace$, implying an overall bound of $0.83\sigma^2$.
By $\ell_1 \coloneqq \sqrt{1 - T^{-1}(s_1)^2} - \nicefrac{s_1}{2}$, we denote the length of the bottom boundary of ${C}_\ell\xspace$; see Figure~\ref{fig:lemma_interval1}(a).
W.l.o.g., we assume $\ell_1 \leq s_1$; the other case is symmetric.
In other words, we assume that the bottom boundary of ${C}_\ell\xspace$ is shorter than its right boundary,
which means that we are using horizontal shelves that we fill from right to left as depicted in \cref{fig:segmentpacking}(b).
\begin{figure*}[b!]
\centering
\includegraphics{figures/lemma_interval1.pdf}
\caption{Illustration of the proof of \cref{lem:summaryInterval1}. In both parts, the red point is always contained in the disk $\mathcal{D}$. (a) The values occurring in part (1).
(b) The situation for $t_1, t_{u+1}$ of maximum possible size in part (2).}
\label{fig:lemma_interval1}
\end{figure*}
Consider the subsequence $t_1,\ldots,t_u,t_{u+1}$ of $s_1,\ldots,s_n$, where $t_1,\ldots,t_u$ are the squares packed by \textsc{Top Packing}\xspace into ${C}_\ell\xspace$ before height $\sigma$ is (strictly) exceeded,
and $t_{u+1}$ is the next square that we try to pack into ${C}_\ell\xspace$.
We observe that $t_{u+1}$ may or may not be packed into ${C}_\ell\xspace$ by \textsc{Top Packing}\xspace, and that $u \geq 1$ by $t_1 \leq \sigma$, i.e., after placing the first square, height $\sigma$ is not exceeded.
We make use of the following lemma, proved by interval arithmetic.
\begin{lemma}[Automatic Analysis for \textsc{Top Packing}\xspace]\label{lem:summaryInterval1}
Let $\ell_1 \leq s_1$, $x_{+1} = \nicefrac{s_1}{2} + \nicefrac{\sigma}{2\sqrt{2}}$ and
$x_{+2} = \nicefrac{s_1}{2} + 0.645\sigma$.
Furthermore, let
\[ y_{+1} = \begin{cases}
T^{-1}(s_1) + \sigma + \nicefrac{\sigma}{2\sqrt{2}}, &\text{if }s_1 \leq s_1^*,\\
\nicefrac{\sigma}{2} + \nicefrac{\sigma}{2\sqrt{2}}, &\text{otherwise,}
\end{cases}\quad
y_{+2} = \begin{cases}
T^{-1}(s_1) + 2 \cdot 0.645\sigma, &\text{if }s_1 \leq s_1^*,\\
-\nicefrac{\sigma}{2} + 2 \cdot 0.645\sigma, &\text{otherwise;}
\end{cases}
\]
see \cref{fig:lemma_interval1}. Let \(F_{TP_1}(s_1) \coloneqq x_{+1}^2 + y_{+1}^2\) and \(F_{TP_2}(s_1) \coloneqq x_{+2}^2 + y_{+2}^2\).
Then, for all \(0.295 \leq s_1 \leq \sqrt{\nicefrac{8}{5}}\), we have \textup{(1)} $F_{TP_1}(s_1) \leq 1$ and \textup{(2)} $F_{TP_2}(s_1) \leq 1$.
\end{lemma}
If $0.645\sigma \leq t_1$, the packed area inside ${C}_\ell\xspace$ is at least $t_1^2 \geq 0.645^2\sigma^2 > 0.415\sigma^2$.
Thus, in the following, we assume $t_1 < 0.645\sigma$.
Furthermore, if $t_{u+1} \leq \nicefrac{\sigma}{2 \sqrt{2}}$, we can apply \cref{lem:summaryInterval1}~(1), showing that $t_{u+1}$ can be packed into ${C}_\ell\xspace$ by \textsc{Top Packing}\xspace; see Figure~\ref{fig:lemma_interval1}(a).
In particular, $t_{u+1}$ can always be packed into ${C}_\ell\xspace$ such that its bottom side lies on height $\sigma$ and its right side touches $s_1$.
The total area packed by \textsc{Top Packing}\xspace into ${C}_\ell\xspace$ is at least the total area packed by \textsc{Shelf Packing}\xspace into the square of area $\sigma$; here we use the fact that the height of the bottom segment of a pocket and the bottom segment of the contained square $\sigma$ coincide, see also \cref{fig:segmentpacking}(e).
Because packing $t_{u+1}$ exceeds height $\sigma$, \cref{lem:shelfpackingtwo} implies that the total area of $t_{u+1}$ and the squares already packed into ${C}_\ell\xspace$ exceeds $\nicefrac{\sigma^2}{2}$.
Thus, in the following, we assume $\nicefrac{\sigma}{2 \sqrt{2}} < t_{u+1} \leq t_1 < 0.645\sigma$.
If $t_1 \leq \nicefrac{\sigma}{2}$, at least four squares are packed into ${C}_\ell\xspace$ by \textsc{Refined Shelf Packing}\xspace before height $\sigma$ is exceeded.
Consequently, the total packed area is at least $4\left( \nicefrac{\sigma}{2 \sqrt{2}} \right)^2 = \nicefrac{\sigma^2}{2}$.
Thus, in the following we assume $t_1 > \nicefrac{\sigma}{2}$.
Now let us assume that only one shelf is constructed before height $\sigma$ is exceeded.
That shelf has height $t_1$ and thus we must have $t_{u+1} > \sigma - t_1$.
We use \cref{lem:summaryInterval1}~(2) to prove that we can pack $t_{u+1}$ on top of the first shelf, even when assuming that $t_1 = t_{u+1} = 0.645\sigma$ are as large as possible; see Figure~\ref{fig:lemma_interval1}(b).
Thus, the total area packed into ${C}_\ell\xspace$ is at least $t_1^2 + t_{u+1}^2 \geq t_1^2 + (\sigma - t_1)^2 \geq \nicefrac{\sigma^2}{2}$.
Otherwise, at least two shelves are constructed before height $\sigma$ is exceeded.
The first shelf has height $t_1$.
The second shelf contains at least two squares because its height is at most $\sigma - t_1 \leq \nicefrac{\sigma}{2}$, and thus at most half of its width.
Thus, the area packed into ${C}_\ell\xspace$ is at least $t_1^2 + 2t_{u+1}^2 \geq \nicefrac{\sigma^2}{4} + 2(\nicefrac{\sigma}{2\sqrt{2}})^2 = \nicefrac{\sigma^2}{2}$, concluding the proof of \cref{lem:anapocketpacking}.
\end{proof}
\subsection{Subcontainer Packing}
For the analysis of \textsc{SubContainer Packing}\xspace, let $C_1,\dots,C_k$ be the subcontainers constructed by \textsc{Bottom Packing}\xspace and let $R_1,\dots,R_k$ be the maximal rectangles contained in $C_1,\dots,C_k$; see \cref{fig:scp-overview}.
\begin{figure}[ht]
\centering
\begin{minipage}[t]{.29\textwidth}
\centering
\includegraphics[scale=1]{figures/scp_overview.pdf}
\caption{Subcontainers $C_i$, $ i\leq k$, produced by \textsc{SubContainer Slicing}\xspace.
}
\label{fig:scp-overview}
\end{minipage}\hfill
\begin{minipage}[t]{.67\textwidth}
\centering
\includegraphics[scale=1]{figures/functionG.pdf}
\caption{The computation of $X$ for a square of side length~$u$:
(a) in case $b \geq 0$, which is symmetric to $a < 0$;
(b) in case $a > 0 > b$ and $u \leq 2c$;
(c) in case $a > 0 > b$ and $u > 2c$.}
\label{fig:functionG}
\end{minipage}
\end{figure}
For $i = 1,\dots,k$, let $h_i$ and $w_i$ denote the height and the width of $R_i$.
Recall that $h_i$ simultaneously denotes the height of $C_i$ and the first square packed into $C_i$.
Let $z$ be the largest square that could be packed below $C_k$.
We define $h_{k+1} \coloneqq s_n$, so $h_{i+1}$ always denotes the first square that did not fit into $C_i$.
Furthermore, we denote the total area of squares packed into $C_i$ by $\| C_i \|$.
We establish several lower bounds on this area $\| C_i \|$.
One such bound is derived from the following observation.
\begin{observation}\label{lem:betterShelfPacking}
The total area packed by \textsc{SubContainer Packing}\xspace into $C_i$ is at least the total area packed by \textsc{Shelf Packing}\xspace into $R_i$.
\end{observation}
If the width of $R_i$ is at least twice its height, the following lemma improves on this bound.
\begin{restatable}{lemma}{Aone}\label{cor:Aone}
For every sequence of squares $s_1,\ldots,s_n$ for which \textsc{Layer Packing}\xspace constructs at least $i$ subcontainers and fails to pack $s_n$, if $w_i \geq 2 h_i$, the area packed into~$C_i$ is
\begin{align*}
\| C_i \| \geq B_1(h_i,w_i,h_{i+1})
\coloneqq & \max \begin{cases}
\nicefrac{1}{2}\cdot h_i w_i + \nicefrac{1}{4}\cdot h_i^2,\\
h_i^2 + (w_i - h_i - h_{i+1})h_{i+1},\\
\nicefrac{1}{2}\cdot h_i (w_i + h_i) - h_{i+1}^2.
\end{cases}
\end{align*}
\end{restatable}
For better readability, we present the proof of \cref{cor:Aone} in the end of this subsection, namely in \cref{app:scp}.
Moreover, we can extend $B_1$ to the cases where the width of $R_i$ is smaller than twice its height as follows.
\begin{restatable}{lemma}{Aeleven}\label{lem:Aeleven}
For every sequence of squares $s_1,\ldots,s_n$ for which \textsc{Layer Packing}\xspace constructs \new{exactly $j \geq i$ subcontainers and fails to pack $h_{j+1} := s_n$,
the area packed into~$C_i$ is}
\begin{align*}
\| C_i \| &\geq B_{2}(h_i,w_i,h_{i+1})
:= \begin{cases}
h_i^2 & \text{ if } w_i < h_i + h_{i+1}, \phantom{\leq w_i \leq 2 h_{i}}\\
h_i^2 + h_{i+1}^2 & \text{ if }\;\; \phantom{w_i \leq} h_i + h_{i+1} \leq w_i < 2 h_{i}, \\
B_1(h_i,w_i,h_{i+1}) & \text{ if }\; \phantom{w_i \leq h_i + h_{i+1} \leq w_i \leq \ } 2 h_{i} \leq w_i.
\end{cases}
\end{align*}
\end{restatable}
\begin{proof}
We always pack at least the square $h_i$ into $C_i$.
As $h_{i+1} \leq h_i$, if $h_i + h_{i+1} \leq w_i$, we pack at least two squares into $C_i$:
Let $s_j$ be the second square we consider packing into $C_i$.
If $h_i$ and $s_j$ do not fit into $C_i$, then $s_j = h_{i+1}$, as we would open a new subcontainer for $s_j$.
This contradicts $h_i + h_{i+1} \leq w_i$, as $h_i$ and $h_{i+1}$ fit into $R_i$ and thus into $C_i$.
\end{proof}
For the last lemma of this subsection, we introduce some useful notation.
Let $a,b$ be the $y$-coordinates of the upper and the lower side of~$C_i$ and let $c\coloneqq c(a,b) = \min \{ a,-b \}$.
When $C_i$ contains the center of the disk, i.e., $a> 0>b$, $c$ denotes the distance of the origin to the nearer side of $C_i$, see \cref{fig:functionG}(b) and (c).
The maximal $x$-coordinate $X$ of the left side of a square of side length $u$ in $C_i$ is determined by
\begin{align*}
X(a,b,u) \coloneqq&
\left.
\begin{cases}
\sqrt{1-(u+b)^2} - u & \text{if } b\geq 0,\\
\sqrt{1-(u-a)^2} - u & \text{if } a < 0,\\
T^{-1}(u) & \text{else if } u \leq 2 c,\\
\sqrt{1-(u-c)^2} - u & \text{otherwise}
\end{cases}
\right\rbrace
=\begin{cases}
T^{-1}(u) & \text{if } u \leq 2 c,\\
\sqrt{1-(u-c)^2} - u & \text{otherwise}
\end{cases}
\end{align*}
The $x$-coordinate of the right side of the first square $h_i$ packed into subcontainer~$C_i$ is $-\nicefrac{1}{2}\cdot w_i + h_i$.
Thus, as $h_{i+1}$ did not fit into $C_i$, we can lower bound the total width of squares packed into $C_i$ after~$h_i$, see \cref{fig:lowerBoundA2}(a), by
\[Y \coloneqq Y(a,h_i,w_i,h_{i+1}) \coloneqq \nicefrac{1}{2} \cdot w_{i} - h_{i} + X(a,a-h_i,h_{i+1}).\]
\begin{figure*}[b]
\centering
\includegraphics[page=3]{figures/lowerBoundA2.pdf}
\caption{(a) Definition of $Y$.
(b) The lower bound $B_3(a, h_i,w_i,h_{i+1})$ when two squares are packed into $C_i$.
(c) The lower bound $B_3(a, h_i,w_i,h_{i+1})$ when at least three squares packed into~$C_i$.}
\label{fig:lowerBoundA2}
\end{figure*}
\begin{restatable}{lemma}{AtwoBound}\label{lem:AtwoBound}
For every sequence of squares $s_1,\ldots,s_n$ for which \textsc{Layer Packing}\xspace constructs exactly $j \geq i$ subcontainers and fails to pack $h_{j+1} := s_n$,
the area packed into~$C_i$ is at least
\[ \| C_i \| \geq B_3(a, h_i, w_i, h_{i+1}) \coloneqq \max \begin{cases}
h_i^2 + \max\{0,Y(a,h_i,w_i,h_{i+1})\} \cdot h_{i+1}, & \textup{(10.1)}{}\\
h_i^2 + \min\left(\max^2(Y(a,h_i,w_i,h_{i+1}),0), 2h_{i+1}^2\right). & \textup{(10.2)}
\end{cases}\]
\end{restatable}
\begin{proof}
If $Y \leq 0$, both bounds (10.1) and (10.2) simplify to $h_i^2$ and are valid, because $h_i$ is packed into~$C_i$.
Thus, let us assume $Y > 0$.
This implies that there are at least two squares packed into $C_i$; otherwise, $h_{i+1}$ would fit into $C_i$.
As $Y$ is a lower bound on the total width of squares packed into~$C_i$ after $h_i$ and $h_{i+1}$ is a bound on their height,
we obtain bound~(10.1); see \cref{fig:lowerBoundA2}(a).
Furthermore, if exactly two squares are packed into $C_i$, $Y^2$ can be used as lower bound on the area of the second square, see \cref{fig:lowerBoundA2}(b).
Otherwise, at least three squares are packed into~$C_i$, and we can add $2h_{i+1}^2$ to $h_i^2$ to bound their area, see \cref{fig:lowerBoundA2}(c).
Combining these two cases yields bound~(10.2).
\end{proof}
We combine these previous bounds into a general lower bound for $\| C_i \|$.
\begin{restatable}{corollary}{combined}\label{cor:combined}
For every sequence of squares $s_1,\ldots,s_n$ for which \textsc{Layer Packing}\xspace constructs exactly $j \geq i$ subcontainers and fails to pack $h_{j+1} := s_n$,
the area packed into~$C_i$ is bounded by
\[\| C_i \| \geq B_4(a, h_i,w_i,h_{i+1}) := \max \begin{cases}
B_2(h_i, w_i, h_{i+1}), & \text{(\cref{lem:Aeleven})}\\
B_3(a, h_i, w_i, h_{i+1}). & \text{(\cref{lem:AtwoBound})}
\end{cases}\]
\end{restatable}
\subsubsection{Proof of \cref{cor:Aone}}\label{app:scp}
In this \new{sub}section, we provide the proof of \cref{cor:Aone}.
\Aone*
\new{The proof consists of the three \cref{lem:shelfpackingnew,lem:anaFramesOfBasePacking,lem:anaFramesOfBasePackingTwo}, each establishing the validity of one of the three lower bounds combined in $B_1$.
In the setting of \cref{cor:Aone}, we are dealing with a sequence of squares $s_1,\ldots,s_n$, for which \textsc{Layer Packing}\xspace fails to pack $s_n$.
Furthermore, \textsc{Layer Packing}\xspace constructs subcontainer $C_i$; thus, there must be a square $t_1 = h_i$ that is placed into $C_i$ (implying that $h_i\leq w_i$), and a subsequence $t_1,\ldots,t_{u},t_{u+1}$ of squares that \textsc{SubContainer Packing}\xspace tries to pack into $C_i$,
where $t_u$ denotes the last square that is packed into $C_i$, and $t_{u+1}$ does not fit.}
\begin{restatable}{lemma}{shelfpackingnew}\label{lem:shelfpackingnew}
{Let $t_1 = h_i \leq w_i$. If $w_i\geq 2h_i$
then the total area of the squares $t_1,\dots,t_u$ that \textsc{SubContainer Packing}\xspace packs into $C_i$ is $\|C_i\|\geq \nicefrac{1}{2}\cdot h_iw_i + \nicefrac{1}{4}\cdot h_i^2$.}
\end{restatable}
\begin{proof}
If $t_{u+1} \leq \nicefrac{1}{2}\cdot h_i$, we consider the part~$R$ of the rectangle $R_i$ that remains after removing $t_1$.
Because $t_{u+1}$ is not packed by \textsc{SubContainer Packing}\xspace into $C_i$, \cref{lem:shelfpacking,lem:betterShelfPacking} imply that the total area of
$t_2, \dots, t_{u+1}$ is at least $\nicefrac{1}{2}\cdot(w_i-h_i)h_i$.
Consequently, the total area of $t_1,\ldots,t_u$ is at least \[h_i^2 + \nicefrac{1}{2}(w_i-h_i)h_i - t_{u+1}^2 \geq \nicefrac{1}{2} \cdot h_iw_i+ \nicefrac{1}{2}\cdot h_i^2 - \nicefrac{1}{4}\cdot h_i^2 = \nicefrac{1}{2}\cdot h_iw_i+ \nicefrac{1}{4}\cdot h_i^2.\]
Thus, we may restrict ourselves to $t_{u+1} =\nicefrac{1}{2}\cdot h_i + \delta$ for some $\delta > 0$; see \cref{fig:area-bound-overview}.
\begin{figure}[htb]
\centering
\includegraphics{figures/bound_area_overview.pdf}
\caption{Our bound on the area $\|C_i\|$ in the remaining case.}
\label{fig:area-bound-overview}
\end{figure}
Let $y \coloneqq w - t_1 - \ldots - t_u$; note that $y$ may be negative or positive, but we have $y \leq t_{u+1} = \nicefrac{1}{2}\cdot h_i + \delta$, as otherwise, we could have packed $t_{u+1}$.
We can bound the area packed into $C_i$ by
\begin{align*}
\|C_i\| &\geq h_i^2 + (w_i - h_i - y)\frac{h_i}{2} + (w_i - h_i - y)\delta\\
&= \frac{1}{2}h_i^2 + \frac{1}{2}w_ih_i - \frac{1}{2}yh_i + (w_i - h_i - y)\delta\\
&= \frac{1}{2} w_ih_i+ \frac{1}{4}h_i^2 + \underbrace{\frac{1}{4} h_i^2+ - \frac{1}{2}yh_i + (w_i - h_i - y)\delta}_{\text{(I)}}.
\end{align*}
Therefore, it suffices to prove that $\text{(I)} = \frac{1}{4}h_i^2 - \frac{1}{2}yh_i + (w_i - h_i - y)\delta \geq 0$, \new{which is equivalent to} $ (w_i - h_i - y) \delta \geq \frac{1}{2}h_i \left(y - \frac{1}{2}h_i \right)$.
We observe that both factors on the left \new{hand} side and one factor on the right \new{hand} side are non-negative, and distinguish whether $w_i \geq 2 h_i + \delta$ or not.
If $w_i \geq 2 h_i + \delta$, from $y \leq t_{u+1}$ we obtain $w_i - h_i - y \geq \frac{1}{2}h_i$ and $\delta \geq y - \frac{1}{2}h_i$, finishing this case.
Finally, we have to handle the case $2h_i \leq w_i \leq 2h_i + \delta$, which implies that only two squares are packed into $C_i$, i.e., $u = 2$.
The total area of $t_1,t_2$ is lower bounded by
\begin{align*}
t_1^2 + t_2^2 &\geq h_i^2 + t_3^2 = h_i^2 + \left(\frac{1}{2}h_i + \delta \right)^2\\
&= h_i^2 + \frac{1}{4}h_i^2 + h_i\delta + \delta^2 \geq \frac{1}{4}h_i^2 + h_i^2 + \frac{1}{2}h_i\delta\\
&= \frac{1}{4}h_i^2 + \underbrace{(2h_i + \delta)}_{\geq w_i}\frac{1}{2}h_i \geq \frac{1}{2}h_iw_i + \frac{1}{4}h_i^2,
\end{align*}
concluding the proof.
\end{proof}
\new{Now, we establish the second lower bound used in \cref{cor:Aone}.}
\begin{restatable}{lemma}{anaFramesOfBasePacking}\label{lem:anaFramesOfBasePacking}
\new{ For every sequence of squares $s_1,\ldots,s_n$ for which \textsc{Layer Packing}\xspace constructs exactly $j$ subcontainers and fails to pack $s_n$,
the total area of squares that \textsc{SubContainer Packing}\xspace packs into $C_i$ for all $i\leq j$ is $\|C_i\| \geq h_i^2 + h_{i+1}(w_i - h_i - h_{i+1})$,
with $h_{j+1} \coloneqq s_n$.}
\end{restatable}
\begin{proof}
The square that failed to be packed into $R_i$ has a side length of $h_{i+1}$, see \cref{fig:LowerBoundStrippacking}.
W.l.o.g., we assume that we fill the vertical shelves in $R_i$ from the top.
The width of the piece of the top of $R_i$ that is not covered by squares is less than $h_{i+1}$;
otherwise, $h_{i+1}$ would have fit.
Therefore, the total square area packed into $C_i$ after $h_i$ is at least $h_{i+1}(w_i - h_i - h_{i+1})$,
because the height of all squares packed into $C_i$ is at least $h_{i+1}$, concluding the proof.
\end{proof}
\begin{figure}[htb]
\centering
\includegraphics{figures/lowerBoundAreaInFrame.pdf}
\caption{Illustration for the proof of \cref{lem:anaFramesOfBasePacking}: A lower bound (gray) for the area (hatched) packed by \textsc{SubContainer Packing}\xspace into a subcontainer.}
\label{fig:LowerBoundStrippacking}
\end{figure}
\new{Finally, we establish the third lower bound used in \cref{cor:Aone}.}
\begin{restatable}{lemma}{anaFramesOfBasePackingTwo}\label{lem:anaFramesOfBasePackingTwo}
\new{For every sequence of squares $s_1,\ldots,s_n$ for which \textsc{Layer Packing}\xspace constructs exactly $j$ subcontainers and fails to pack $s_n$, if $w_i \geq 2h_i$,
the total area of squares that \textsc{SubContainer Packing}\xspace packs into $C_i$ for all $i\leq j$ is $\|C_i\| \geq \nicefrac{1}{2}\cdot h_i(w_i+h_i) - h_{i+1}^2$, with $h_{j+1} \coloneqq s_n$.}
\end{restatable}
\begin{proof}
We consider the remainder $R_i'$ of rectangle $R_i$ after $h_i$ is packed.
$R_i'$ has height $h_i$ and width $w_i-h_i$. \new{Because $w_i\geq2h_i$ we have that $w_i-h_i \geq h_i$ and hence, at least one square is packed into $R_i'$.}
Consequently, \cref{lem:shelfpacking,lem:betterShelfPacking} imply that the total area packed by \textsc{SubContainer Packing}\xspace into $C_i$ is at least $\nicefrac{1}{2}\cdot h_i(w_i - h_i)-h_{i+1}^2$.
Thus, we have $\|C_i\| \geq h_i^2 + \nicefrac{1}{2}\cdot h_i(w_i - h_i) -h_{i+1}^2= \nicefrac{1}{2}\cdot h_i (w_i +h_i) - h_{i+1}^2$.
\end{proof}
\section{Analysis of the main algorithm}
\label{sec:analysis-algorithm}
In this section, we prove our main result using the tools provided in \cref{sec:interval-arithmetic-proofs,sec:analysis-subroutines}.
On the highest level, the proof consists of three parts corresponding to the three cases that our algorithm distinguishes.
\subsection{Analysis of \textbf{(C\;\!1)}\xspace}
\label{sec:analysis-first-case}
Recall that in case \textbf{(C\;\!1)}\xspace, we place a container square $\mathcal{X}$ of side length $1.388$ into $\mathcal{D}$,
and pack the first four squares into pockets outside $\mathcal{X}$ and all remaining disks into $\mathcal{X}$ using \textsc{Shelf Packing}\xspace; see \cref{fig:Alg}(a).
\begin{restatable}{lemma}{correctnessStepOne}\label{lem:correctnessStepOne}
If \textsc{Layer Packing}\xspace fails to pack a sequence of squares $s_1,\ldots,s_n$ with $s_1 \leq 0.295$, the total area of the squares exceeds $\nicefrac{8}{5}$.
\end{restatable}
\begin{proof}
Consider scaling down all side lengths by a factor of $\nicefrac{1}{1.388}$, such that $\mathcal X$ is the unit square and $s_1 \leq \nicefrac{0.295}{1.388} \approx 0.2125$.
As $s_5$ is the first square packed by \textsc{Shelf Packing}\xspace into $\mathcal X$,
\cref{lem:shelfpackingtwo} implies that the total area packed into the scaled $\mathcal{D}$ is at least
$f(s_5) = 4 s_5^2 + \nicefrac{1}{2} + 2(s_5-\nicefrac{1}{2})^2$ with derivative $f'(s_5)= 12s_5 - 2,$
which is minimized for $s_5 = \nicefrac{1}{6}$, where $f(\nicefrac{1}{6}) = \nicefrac{5}{6}$.
Thus, in the non-scaled configuration, the area packed is at least
$\nicefrac{5}{6}\cdot 1.388^2 = \nicefrac{120409}{75000} \approx 1.605 > \nicefrac{8}{5}$, concluding the proof.
\end{proof}
\subsection{Analysis of \textbf{(C\;\!2)}\xspace}
\label{sec:analysis-second-case}
Recall that in case \textbf{(C\;\!2)}\xspace, we pack the four largest squares into the squares $\mathcal{X}_1,\dots, \mathcal{X}_4$;
all other squares are packed into a square container $\mathcal{X}$ on top of them; see \cref{fig:Alg}(b).
\begin{restatable}{lemma}{correctnessStepTwo}\label{lem:correctnessStepTwo}
If \textsc{Layer Packing}\xspace fails to pack a sequence $s_1,\ldots,s_n$ of squares with $0.295 < s_1 \leq \nicefrac{1}{\sqrt{2}}$ and $s_1^2 + s_2^2 + s_3^2 + s_4^2 \geq \nicefrac{39}{25}$,
the total area of the squares exceeds $\nicefrac{8}{5}$.
\end{restatable}
\begin{proof}
By assumption, the total area of the squares $s_1,s_2,s_3,s_4$ is at least $\nicefrac{39}{25} = \nicefrac{8}{5} - \nicefrac{1}{25}$.
As $\mathcal{X}$ has an area of $\nicefrac{2}{25}$, \cref{lem:shelfpacking} implies that \textsc{Shelf Packing}\xspace (and thus \textsc{Layer Packing}\xspace) only fails to pack all remaining squares into $\mathcal{X}$ if their area exceeds $\nicefrac{1}{25}$. Consequently, the total area of the squares exceeds $\nicefrac{8}{5}$, concluding the proof.
\end{proof}
\subsection{Analysis of \textbf{(C\;\!3)}\xspace}
Recall that $z$ denotes the largest square that could be packed below the last subcontainer $C_k$ constructed by \textsc{Bottom Packing}\xspace,
as illustrated in \cref{fig:scp-overview} or in \cref{fig:formula-z}, where we have reflected the instance along the x-axis.
\begin{figure*}[b]
\centering
\includegraphics[scale=1]{figures/formula_z.pdf}
\caption{(a) Computing $z$ for $k = 2$ using functions $T$ and $T^{-1}$. (b) Computing the width $w_i = w(y_t,h)$ of rectangle $R_i$.}
\label{fig:formula-z}
\end{figure*}
We consider a sequence $s_1,\ldots,s_n$ of squares of total area $S$ that \textsc{Layer Packing}\xspace fails to pack, and assume w.l.o.g.\ that $s_{n-1}$ is packed.
This implies $z < s_n$, where $z$ denotes the side length of the largest square that could be packed below the last subcontainer $C_k$ constructed by \textsc{Bottom Packing}\xspace;
otherwise, a further subcontainer is constructed.
We have $z = T(-T^{-1}(s_1) + \sum_{i=1}^{k} h_i)$; see \cref{fig:formula-z}(a).
By $w(y_t,h) = 2\sqrt{\min\{1 - y_t^2, 1 - (y_t-h)^2\}}$,
we denote the maximum width of a rectangle~$R$ that can be placed in $\mathcal{D}$ with top side at $y = y_t$ and height $h$; see \cref{fig:formula-z}(b).
Thus, we can express the width $w_i$ of the rectangle $R_i$ inscribed in some subcontainer $C_i$ in terms of $s_1,h_1,\ldots,h_i$ as
\[ w_i \coloneqq w\left(\bigg( T^{-1}(s_1)-\sum\limits_{j=1}^{i-1}h_j\bigg), h_i\right).\]
Recall that $\sigma \coloneqq \sigma(s_1)$ denotes the side length of the largest squares that fits into the pockets ${C}_\ell\xspace$ and ${C}_r\xspace$ as illustrated in \cref{fig:t-t1-sigmaApp}(c).
In order to distinguish whether \textsc{Top Packing}\xspace has packed any squares into the pockets, we consider the function
\[
E(s_1, s_n) \coloneqq
\begin{cases}
0.83\cdot \sigma(s_1)^2,& \text{if } s_n \leq \sigma,\\
0,& \text{ otherwise,}
\end{cases}
\]
which describes the total square area that \textsc{Top Packing}\xspace is guaranteed to pack due to \cref{lem:anapocketpacking}.
For the analysis of Case~\textbf{(C\;\!3)}\xspace, we distinguish cases depending on the number $k$ of subcontainers constructed by \textsc{Bottom Packing}\xspace.
Specifically, we consider the cases $k=0,k=1,k\in\{2,3,4\}$, and $k \geq 5$.
\subsubsection{Analysis for no subcontainer}
\label{sec:NoRow}
\begin{restatable}{lemma}{anaNoStrip}\label{lem:anaNoStrip}
If \textsc{Layer Packing}\xspace fails to pack a sequence $s_1,\ldots,s_n$ of squares and \textsc{Bottom Packing}\xspace does not construct a subcontainer, the total area of the squares exceeds $\nicefrac{8}{5}$.
\end{restatable}
\begin{proof}
Because the algorithm fails to construct a first subcontainer in the bottom part, it follows that placing $s_n$ as far to the bottom as possible yields an overlap with $s_1$.
However, the minimum value for $s_1^2 + s_n^2$ for two overlapping squares packed into a disk is attained for $s_1 = s_n$.
This corresponds to the worst-case configuration, implying that the total area of $s_1$ and $s_n$ exceeds $\nicefrac{8}{5}$.
\end{proof}
\subsubsection{Analysis for one subcontainer}
\begin{restatable}{lemma}{anaOneRow}\label{lem:anaOneRow}
If \textsc{Layer Packing}\xspace fails to pack a sequence $s_1,\ldots,s_n$ of squares and \textsc{Bottom Packing}\xspace constructs exactly one subcontainer, the total area of the squares exceeds $\nicefrac{8}{5}$.
\end{restatable}
\begin{proof}
Combining \cref{lem:anapocketpacking,cor:combined} allows us to bound the area of $s_1,\ldots,s_n$ by
\[ S \geq F_{SC_1}(s_1,h_1,s_n) \coloneqq s_1^2 + B_{4}(T^{-1}(s_1), h_1, w_1, s_n) + s_n^2 + E(s_1, s_n),\]
where $
E(s_1, s_n) =
0.83\cdot \sigma(s_1)^2 $ if $s_n \leq \sigma$, and
$E(s_1, s_n)=0$ if $s_n > \sigma$.
Furthermore, we know $0 < z < s_n$, because \textsc{Layer Packing}\xspace fails to pack $s_n$.
Moreover, we claim that at least one of the following conditions must hold:
$s_1 > \nicefrac{1}{\sqrt{2}}$, $w_1 < 2h_1$, or $s_1^2 + h_1^2 + 2s_n^2 < \nicefrac{39}{25}$.
Assume for contradiction that neither of these conditions hold.
By $w_1 \geq 2h_1$, we know that at least two squares are packed into the first subcontainer.
One of these squares has area $h_1^2$, and the other has area at least $s_n^2$.
In particular, this implies that the algorithm packs $s_1,s_2$ and $s_3$.
This implies $s_2 \geq h_1$ and $s_3,s_4 \geq s_n$.
Thus we have $s_1^2 + s_2^2 + s_3^2 + s_4^2 \geq s_1^2 + h_1^2 + 2s_n^2 \geq \nicefrac{39}{25}$,
which together with $s_1 \leq \nicefrac{1}{\sqrt{2}}$ implies that we are in Case~\textbf{(C\;\!2)}\xspace of our algorithm.
This is a contradiction, because we only construct subcontainers in Case~\textbf{(C\;\!3)}\xspace.
Thus the following lemma, proved automatically using interval arithmetic, proves that these conditions are sufficient to ensure $S \geq \nicefrac{8}{5}$.
\begin{lemma}[\textsc{One Subcontainer}, Automatic Analysis for Lemma~\ref{lem:anaOneRow}]\label{lem:summaryInterval2}
Let $z \coloneqq T(T^{-1}(s_1) + h_1)$.
For all $s_1, h_1, s_n$ with $0 < z < s_n \leq h_1 \leq s_1$, $h_1 \leq T^{-1}(s_1) + 1$ and
\[\left(s_1 > \nicefrac{1}{\sqrt{2}}\right) \vee \left(w_1 < 2h_1\right) \vee \left(s_1^2 + h_1^2 + 2s_n^2 < \nicefrac{39}{25}\right),\]
we have $F_{SC_1}(s_1,h_1,s_n) > \nicefrac{8}{5}$.\qedhere
\end{lemma}
\end{proof}
\subsubsection{Analysis for two to four subcontainers}
\begin{restatable}{lemma}{anaTwoThreeFourRows}\label{lem:anaTwoThreeFourRows}
If \textsc{Layer Packing}\xspace fails to pack a sequence $s_1,\ldots,s_n$ of squares and \textsc{Bottom Packing}\xspace constructs $k \in \{2,3,4\}$ subcontainers, the total area of the squares exceeds $\nicefrac{8}{5}$.
\end{restatable}
\begin{proof}
We use similar ideas as in the proof of \cref{lem:anaOneRow}.
We bound the area packed by \textsc{Top Packing}\xspace by $E(s_1, s_n)$ using \cref{lem:anapocketpacking}.
Furthermore, we use \cref{cor:combined} to bound the area packed into each of the $k \in \{2,3,4\}$ subcontainers by
\[ \| C_i \| \geq B_4\left(\bigg(T^{-1}(s_1) - \sum\limits_{j=1}^{i-1} h_i\bigg), h_i, w_i, h_{i+1}\right), 1 \leq i \leq k,\]
where $h_{k+1} \coloneqq s_n$.
We can express $w_i = w(T^{-1}(s_1) - \sum_{j=1}^{i-1}h_j, h_i)$ in terms of $s_1$ and $h_j, 1 \leq j \leq i$.
In total, for $k$ subcontainers, this yields the bound
\[ S \geq F_{SC_k}(s_1,h_1,\ldots,h_{k},s_n) \coloneqq s_1^2 + s_n^2 + E(s_1, s_n) + \sum\limits_{i=1}^{k} B_4\left(\bigg(T^{-1}(s_1) - \sum\limits_{j=1}^{i-1} h_i\bigg), h_i, w_i, h_{i+1}\right).\]
Finally, we know $0 < z < s_n$ because the algorithm fails to pack $s_n$.
Thus, the following lemma, proved automatically using interval arithmetic, suffices to complete the proof of \cref{lem:anaTwoThreeFourRows}.
\begin{lemma}[Automatic Analysis for Lemma~\ref{lem:anaTwoThreeFourRows}]
Let $z_k = T(-T^{-1}(s_1) + \sum_{i=1}^{k}h_i)$.
\begin{description}
\item[$(k = 2)$] For all $s_1,h_1,h_2,s_n$ with $0 < z_2 < s_n \leq h_2 \leq h_1 \leq s_1$ and $0.295 \leq s_1 \leq \sqrt{\nicefrac{8}{5}}$ and $h_1 + h_2 \leq 1 + T^{-1}(s_1)$,
we have $F_{SC_2} > \nicefrac{8}{5}$.
\item[$(k = 3)$] For all $s_1,h_1,h_2,h_3,s_n$ with $0 < z_3 < s_n \leq h_3 \leq h_2 \leq h_1 \leq s_1$
and $0.295 \leq s_1 \leq \sqrt{\nicefrac{8}{5}}$ and $h_1 + h_2 + h_3 \leq 1 + T^{-1}(s_1)$,
we have $F_{SC_3} > \nicefrac{8}{5}$.
\item[$(k = 4)$] For all $s_1,h_1,h_2,h_3,h_4,s_n$ with $0 < z_4 < s_n \leq h_4 \leq h_3 \leq h_2 \leq h_1 \leq s_1$
and $0.295 \leq s_1 \leq \sqrt{\nicefrac{8}{5}}$ and $h_1 + h_2 + h_3 + h_4 \leq 1 + T^{-1}(s_1)$,
we have $F_{SC_4} > \nicefrac{8}{5}$.\qedhere
\end{description}
\end{lemma}
\end{proof}
We defer the analysis for five or more subcontainers to \cref{sec:details-analysis-algorithm}.
This completes the analysis of \textbf{(C\;\!3)}\xspace and thus the proof of our main result.
\subsubsection{Analysis for five or more subcontainers}
\label{sec:details-analysis-algorithm}
In this section, we handle the case where \textsc{Layer Packing}\xspace constructs at least five subcontainers.
\begin{restatable}{lemma}{anaFiveRows}\label{lem:anaFiveRows}
If \textsc{Layer Packing}\xspace fails to pack a sequence $s_1,\ldots,s_n$ of squares and \textsc{Bottom Packing}\xspace constructs $k \geq 5$ subcontainers, the total area of the squares exceeds $\nicefrac{8}{5}$.
\end{restatable}
On the highest level, we distinguish the two cases
$s_n > \sigma$ (\cref{lem:five-containers-above-sigma})
and $s_n \leq \sigma$ (\cref{lem:five-containers-below-sigma}), together establishing \cref{lem:anaFiveRows}.
\paragraph{Proof of \texorpdfstring{\cref{lem:five-containers-above-sigma}}{Claim 22}}
\begin{claim}
\label{lem:five-containers-above-sigma}
If \textsc{Layer Packing}\xspace fails to pack a sequence $s_1,\ldots,s_n$ of squares, $s_n > \sigma$ and
\textsc{Bottom Packing}\xspace constructs $k \geq 5$ subcontainers, the total area of the squares exceeds $\nicefrac{8}{5}$.
\end{claim}
\begin{claimproof}
Because of $0.295 \leq s_1 \leq \sqrt{\nicefrac{8}{5}}$ \new{and \cref{eq:sigma}}, we have $\sigma > 0.231$.
Of the $2$ units of vertical space, $s_1$ takes up $1 - T^{-1}(s_1) > 0.305$.
Therefore, $s_n > \sigma$ implies that only $k \leq \lfloor \nicefrac{2-0.305}{0.231}\rfloor = 7$ subcontainers can be constructed.
Because $s_n > \sigma$, no square is packed into the pockets by \textsc{Top Packing}\xspace.
We use \cref{cor:combined} to derive the same lower bound on the area packed into container $C_i$ as in the proof of \cref{lem:anaTwoThreeFourRows}:
\[ \| C_i \| \geq B_4\left(\bigg(T^{-1}(s_1) - \sum\limits_{j=1}^{i-1} h_i\bigg), h_i, w_i, h_{i+1}\right), 1 \leq i \leq k,\]
where $h_{k+1} \coloneqq s_n$.
We then bound the total area of $s_1,\ldots,s_n$ by
\[S \geq F_{SC_k}(s_1,h_1,\ldots,h_{k},s_n) \coloneqq s_1^2 + s_n^2 + \sum\limits_{i=1}^{k} B_4\left(\bigg(T^{-1}(s_1) - \sum\limits_{j=1}^{i-1} h_i\bigg), h_i, w_i, h_{i+1}\right).\]
Thus, the following lemma proved using interval arithmetic concludes the proof, and thus the proof of \cref{lem:anaFiveRows} for the case $s_n > \sigma$.
\end{claimproof}
\begin{lemma}[Automatic Analysis for \cref{lem:five-containers-above-sigma}]
Let $z_k = T(-T^{-1}(s_1) + \sum_{i=1}^{k}h_i)$.
\begin{description}
\item[$(k = 5)$] For all $s_1,h_1,\ldots,h_5,s_n$ with $0 < z_5,\sigma < s_n \leq h_5 \leq \cdots \leq h_1 \leq s_1$
and $0.295 \leq s_1 \leq \sqrt{\nicefrac{8}{5}}$ and $h_1 + \cdots + h_5 \leq 1 + T^{-1}(s_1)$,
we have $F_{SC_5} > \nicefrac{8}{5}$.
\item[$(k = 6)$] For all $s_1,h_1,\ldots,h_6,s_n$ with $0 < z_6,\sigma < s_n \leq h_6 \leq \cdots \leq h_1 \leq s_1$
and $0.295 \leq s_1 \leq \sqrt{\nicefrac{8}{5}}$ and $h_1 + \cdots + h_6 \leq 1 + T^{-1}(s_1)$,
we have $F_{SC_6} > \nicefrac{8}{5}$.
\item[$(k = 7)$] For all $s_1,h_1,\ldots,h_7,s_n$ with $0 < z_7,\sigma < s_n \leq h_7 \leq \cdots \leq h_1 \leq s_1$
and $0.295 \leq s_1 \leq \sqrt{\nicefrac{8}{5}}$ and $h_1 + \cdots + h_7 \leq 1 + T^{-1}(s_1)$,
we have $F_{SC_7} > \nicefrac{8}{5}$.
\end{description}
\end{lemma}
\paragraph{Proof of \texorpdfstring{\cref{lem:five-containers-below-sigma}}{Claim 24}}
\begin{claim}
\label{lem:five-containers-below-sigma}
If \textsc{Layer Packing}\xspace fails to pack a sequence $s_1,\ldots,s_n$ of squares, $s_n \leq \sigma$ and
\textsc{Bottom Packing}\xspace constructs $k \geq 5$ subcontainers, the total area of the squares exceeds $\nicefrac{8}{5}$.
\end{claim}
In the last remaining case, we have $k \geq 5$ subcontainers and $s_n \leq \sigma$.
To handle the \new{arbitrarily large number} of subcontainers in this case,
we begin by deriving bounds for the total area of squares packed into subcontainers $C_j,\ldots,C_k$
and $s_n$, which we assume, w.l.o.g., to be the first square that \textsc{Layer Packing}\xspace fails to pack.
Let $C_j$ be a subcontainer whose top side is below the center of $\mathcal{D}$, i.e., below the line $y = 0$.
Let $A_{j}$ be the area of the part of the disk below the top side of $C_{j}$,
containing $C_{j},\dots,C_k$; see \cref{fig:rowsUpperBoundApp}(a).
Let $H_j$ be the vertical distance between the top side of $C_j$ and the lowest point of $\mathcal{D}$.
\begin{figure*}[htb]
\centering
\includegraphics[page=3]{figures/rowsUpperBound.pdf}
\caption{
(a)~The definition of $A_j, h_j, H_j$ and $w_j$.
The area of a subcontainer $C_i$, $i\geq j$ can be upper bounded by the area of the smallest enclosing rectangle (gray-white hatched).
The area below $C_k$ is upper bounded by the area of the $z \times w_k$-rectangle (gray-white hatched) below $C_k$,
where $z$ denotes the side length of a largest square fitting below $C_k$.
(b)~Bounding the area below $C_k$ (with bold outline) by $zw_k < h_{k+1}w_k$ requires
showing that the red area does not exceed the gray area.}
\label{fig:rowsUpperBoundApp}
\end{figure*}
We first prove the following lemma.
\begin{lemma}\label{lem:totalAreaPackedByBasePacking}
If $s_n \leq 0.6$, the total area of $s_n$ and the squares packed into $C_j,\ldots,C_k$ by \textsc{Bottom Packing}\xspace is at least
\[B_5(h_j,H_j) \coloneqq A_{j+1} + h_j^2 - H_jh_j.\]
\end{lemma}
\begin{proof}
Applying \cref{lem:anaFramesOfBasePacking} to $C_j,\ldots,C_k$ yields that
the total area of $s_n = h_{k+1}$ and the squares packed into $C_j,\ldots,C_k$ is at least
\begin{align*}
S :=h_{k+1}^2+\sum\limits_{i=j}^{k} \| C_i \| &\geq h_{k+1}^2 + \sum\limits_{i=j}^{k} \big( h_i^2 + h_{i+1}\cdot(w_i - h_i - h_{i+1})\big)\\
&= h_{k+1}^2 + \sum\limits_{i=j}^{k} h_i^2 - \sum\limits_{i=j}^{k}h_{i+1}^2 + \sum\limits_{i=j}^{k} h_{i+1}w_i - \sum\limits_{i=j}^{k} h_{i+1}h_i \\
&= h_{k+1}^2 + h_j^2 - h_{k+1}^2 + \sum\limits_{i=j}^{k} h_{i+1}w_i - \sum\limits_{i=j}^{k} h_{i+1}h_i\\
&= h_j^2 + \sum\limits_{i=j}^{k} h_{i+1}w_i - \sum\limits_{i=j}^{k} h_{i+1}h_i.
\end{align*}
Next, we show that we can bound the disk area $A_{k+1}$ below $C_k$ by $B \coloneqq w_kh_{k+1}$;
see \cref{fig:rowsUpperBoundApp}(b).
We have $A_{k+1} = \theta-t\sqrt{1-t^2} = \theta - \sin\theta \cdot \cos\theta$, where $t \coloneqq \cos\theta$.
We also have $(t+z)^2 + \nicefrac{z^2}{4} = 1$, i.e., $t = \sqrt{1-\nicefrac{z^2}{4}} - z$.
Because $0 < z < s_n \leq 0.6 < 1$, we can bound this by
\[ t > 1 - \nicefrac{1}{4}\cdot z^2- z > 1 - \nicefrac{5}{4}\cdot z \implies z > \nicefrac{4}{5}\cdot(1-t).\]
We have $B = h_{k+1}w_k > zw_k = z \cdot 2\sin\theta$, and thus
$B > \nicefrac{4}{5}\cdot(1-\cos\theta) \cdot 2\sin\theta = \nicefrac{8}{5} \cdot (1-\cos\theta)\sin\theta,$
which implies \[B - A_{k+1} > \nicefrac{8}{5}\cdot\sin\theta - \nicefrac{8}{5}\cdot\sin\theta\cos\theta-\theta+\sin\theta\cos\theta = \underbrace{\nicefrac{8}{5}\sin\theta - \nicefrac{3}{5}\sin\theta\cos\theta - \theta}_{\eqqcolon f(\theta)}.\]
Clearly, $f(0) = 0$.
Moreover, the derivative is
\[ f'(\theta)
= -\nicefrac{2}{5}(3\cos^2\theta-4\cos\theta+1)
= \nicefrac{2}{5}(1-\cos\theta)(3\cos\theta - 1),\]
which is non-negative for $0 \leq \theta \leq \cos^{-1}(\nicefrac{1}{3})$.
Hence, we have $f(\theta) \geq 0$ and thus $B > A_{k+1}$ for $0 \leq \theta \leq \cos^{-1}(\nicefrac{1}{3})$.
Because of $z < s_n \leq 0.6$, the largest angle $\theta$ that occurs is
\[\theta = \underbrace{\cos^{-1}\left(\sqrt{1-\nicefrac{0.6^2}{4}}-0.6\right)}_{\approx 1.209} <
\underbrace{\cos^{-1}\left(\nicefrac{1}{3}\right)}_{\approx 1.230}.\]
We have thus established $A_{k+1} < h_{k+1}w_k$ and can bound
$A_{j+1} \leq \sum_{i=j+1}^{k+1}h_iw_{i-1} = \sum_{i=j}^{k}h_{i+1}w_i$;
see \cref{fig:rowsUpperBoundApp}(a).
Thus, by $h_j \geq h_{j+1} \geq \cdots \geq h_{k+1} = s_n$, we have
\begin{align*}
S &\geq h_j^2 + A_{j+1} - \sum\limits_{i=j}^{k}h_{i+1}h_i \geq h_j^2 + A_{j+1} - h_j\sum\limits_{i=j}^{k}h_{i+1}\\
&\geq h_j^2 + A_{j+1} - h_j\sum\limits_{i=j}^{k}h_{i} \geq h_j^2 + A_{j+1} - h_jH_j\text{,}
\end{align*}
and thus the claimed bound.
\end{proof}
To apply \cref{lem:totalAreaPackedByBasePacking}, we first observe that $\sigma(s_1) \leq 0.6$ for any $s_1$.
Assume for contradiction that $\sigma(s_1) > 0.6$ for some $s_1$.
Because two squares of side length $\sigma$ fit into $\mathcal{D}$ besides $s_1$,
$2\sigma + s_1 < 2$ and thus $s_1 < 0.8$.
However, for $0 \leq s_1 \leq 0.8$, $\sigma(s_1)$ is monotonically increasing with $\sigma(0.8) < 0.4345 < 0.6$, which is a contradiction.
To prove \cref{lem:five-containers-below-sigma},
we distinguish whether the $y$-coordinate $y_3$ of the bottom side of
the third subcontainer $C_3$ is positive or not.
If $y_3 \leq 0$, i.e., $C_3$'s bottom side is below the center of $\mathcal{D}$,
we apply \cref{lem:totalAreaPackedByBasePacking,lem:anapocketpacking,lem:Aeleven} to lower-bound the total square area of $s_1,\ldots,s_n$ by
\[
F_{MSC_1} \coloneqq s_1^2 + 0.83\sigma^2 + B_5(h_4, H_4) + \sum_{i=1}^{3} B_4\left(\bigg(T^{-1}(s_1)-\sum_{\ell=1}^{i-1}h_{\ell}\bigg), h_i, w_i, h_{i+1}\right),
\]
with $H_4 = 1 + T^{-1}(s_1) - h_1 - h_2 - h_3$ and $w_i = w(T^{-1}(s_1) - \sum_{\ell=1}^{i-1}h_{\ell}, h_i)$.
The case $y_3 \leq 0$ can then be handled using the following lemma, proved by interval arithmetic.
\begin{lemma}[Automatic Analysis for \cref{lem:five-containers-below-sigma}, $y_3 \leq 0$]
For all $s_1 \geq h_1 \geq \cdots \geq h_4 > 0$
with $0.295 \leq s_1 \leq \sqrt{\nicefrac{8}{5}}$ and
$0 \leq H_4 \leq 1$, we have $F_{MSC_1} > \nicefrac{8}{5}$.
\end{lemma}
In the last remaining case we have $y_3 > 0$, i.e., the third subcontainer lies strictly above the center of $\mathcal{D}$.
Let $j$ be the maximal index such that the top side of $C_j$ lies strictly above $\mathcal{D}$'s center; we have $j > 3$ due to $y_3 > 0$.
As before, we bound the area packed by \textsc{Top Packing}\xspace by $s_1^2 + 0.83\sigma^2$ with \cref{lem:anapocketpacking} and the area packed by \textsc{Bottom Packing}\xspace into $C_1$ and $C_2$ using \cref{cor:combined}.
Furthermore, we bound the area packed into $C_{j+1},\ldots,C_k$ and $s_n$ using \cref{lem:totalAreaPackedByBasePacking}.
For the total square area $S_c$ packed into $C_3,\ldots,C_j$,
we consider the rectangle $R$ of height $H_R = \sum_{i=3}^{j}h_i$ and
width $W_R = \min\{w_3,\ldots,w_j\}$ inscribed into $C_3 \cup \cdots \cup C_j$;
see \cref{fig:y3-gt-0}.
Next, we prove the following bound on the total square area $S_c$ packed into $C_3,\ldots,C_j$:
\[ S_c \geq B_6(H_R,W_R,h_{j+1}) \coloneqq \nicefrac{1}{2}\cdot H_RW_R + \nicefrac{1}{4}\cdot h_{j+1}H_R.\]
\begin{figure}[htb]
\centering
\includegraphics[scale=1]{figures/y3_gt_0.pdf}
\caption{In the case $y_3 > 0$, we use the rectangle $R$ of height $H_R$ and width $W_R$
inscribed in $C_3,\ldots,C_j$ to bound the total area packed into $C_3,\ldots,C_j$.}
\label{fig:y3-gt-0}
\end{figure}
First, we observe that due to the fact that three subcontainers fit above $\mathcal{D}$'s center and below~$s_1$, $T^{-1}(s_1) - h_1 - h_2 - h_3 > 0$ and thus $h_3 < \nicefrac{1}{3}\cdot T^{-1}(s_1) \leq \nicefrac{1}{3} \cdot T^{-1}(0.295)< 0.232$.
This also bounds the distance $\Delta_y$ between $\mathcal{D}$'s center and the bottom side of $R$ by $0.232$.
We thus obtain $W_R \geq \min(2\sqrt{1-T^{-1}(0.295)^2}, 2\sqrt{1-0.232^2}) > 1.439 > 2h_3$.
This yields $w_i > 2h_i$ for all $i \in \{3,\ldots,j\}$.
This allows us to apply \cref{lem:shelfpackingnew},
which yields \[\|C_i\| \geq \nicefrac{1}{2}\cdot h_iw_i + \nicefrac{1}{4}\cdot h_i^2 \geq \nicefrac{1}{2}\cdot h_iw_i + \nicefrac{1}{4}\cdot h_ih_{j+1}\] for $i \in \{3,\ldots,j\}$.
Summing over $i$ yields
\[ S_c \geq \sum\limits_{i=3}^{j} \frac{1}{2}h_iw_i + \sum\limits_{i=3}^{j} \frac{1}{4}h_ih_{j+1}
\geq \left(\frac{1}{2}W_R + \frac{1}{4}h_{j+1}\right)\sum\limits_{i=3}^{j}h_i = B_6(H_R,W_R,h_{j+1}),\] as claimed.
Overall, we can thus bound the total area $S$ of $s_1,\ldots,s_n$ by
\begin{align*}
S \geq F_{MSC_2} &\coloneqq s_1^2 + 0.83\sigma^2 + B_4(T^{-1}(s_1), h_1, w_1, h_2)\\
& + B_4(T^{-1}(s_1)-h_1,h_2,w_2,h_3) + B_6(H_R, W_R, h_{j+1}) + B_5(h_{j+1}, H_{j+1}).
\end{align*}
We have $H_R = T^{-1}(s_1) - h_1 - h_2 + \Delta_y$ and $W_R = w(T^{-1}(s_1) - h_1 - h_2, H_R)$ and $H_{j+1} = 1 - \Delta_y$.
Therefore, the following lemma, proved using interval arithmetic, concludes the proof of \cref{lem:five-containers-below-sigma} and thus the proof of \cref{lem:anaFiveRows}.
\begin{lemma}[Automatic Analysis for \cref{lem:five-containers-below-sigma}, $y_3 > 0$]
For all $s_1 \geq h_1 \geq h_2 \geq h_3 \geq h_{j+1}$,
$0.295 \leq s_1 < \sqrt{\nicefrac{8}{5}}$ and all $\Delta_y$ with
$T^{-1}(s_1) - h_1 - h_2 - h_3 > 0$ and $\Delta_y \leq h_3$, we have $F_{MSC_2} > \nicefrac{8}{5}$.
\end{lemma}
\section{Conclusion \new{and future directions}}
\label{sec:conc}
We have established the critical density for packing squares into a disk:
Any set of squares of total area at most $\nicefrac{8}{5}$ can be packed into a unit disk. As shown by our lower bound example, this guarantee is best-possible, i.e., it cannot be improved.
The proof is based on an algorithm that subdivides the disk into horizontal subcontainers and uses a refined shelf packing scheme.
The correctness of this algorithm is shown by careful manual analysis, complemented by a computer-assisted part that is based on interval arithmetic.
There is a variety of interesting directions for future research.
Of particular interest is the critical density for packing squares of bounded size into a disk,
which will result in a higher packing density;
a more general problem concerns the critical packing density
for packing other types of objects of bounded size into other types of containers.
Other questions arise from considering questions in three- or even higher-dimensional space.
We are optimistic that many of our techniques will be useful for settling these problems.
|
\section{Conclusions}
\label{sec:concl}
In this paper we presented a multi-layer robotic mapping pipeline able to build
in realtime geometric-semantic representation and complete object instances in
the environment by model matching. We evaluated the proposed method in both
simulation and real environments as well as compared its performance in object
shape-completion against a state-of-the-art{} deep learning method demonstrating how the
models produced by our approach are better suited for robotic applications.
Finally we demonstrated how the proposed method can be used to increase
robustness of navigation. Since the experimental scope was limited to estimating
the extent of a single object category, understanding the scalability of the
approach requires further studies.
In conclusion, the mapping pipeline presented here represents a step toward
mapping methods, where different data sources are combined and artificial
intelligence is employed to integrate and complete missing information. More
work in this direction is still needed, particularly in regards to bridging the
gap between shape completion of synthetic object and scene completion of real
environments as well as to explore the inclusion of agent dynamics in the map.
However, we believe that such mapping approaches will be pivotal to bring the
many recent advances in computer vision to real mobile robot applications.
\section{Discussion}
\label{sec:disc}
Over the past decade, deep learning based image segmentation and shape
completion methods have made significant leaps from a computer vision
perspective. However, our experiments hint at a problem of applying the current
methods in robotics: noisy and incomplete data captured in the wild often
deteriorates their performance considerably. This is particularly true for deep
learning methods that have usually hard time to produce reasonable results on
out-of-distribution data. In our case, the semantic label from Mask R-CNN was
often coarse which led to labels bleeding out of the object boundaries, and
required additional post processing steps. As for the methods like \ac{pcn}, the
network training is performed in isolation with existing databases without
sufficiently taking the real-world factors into account. Hence, when these
models are applied in real world robotic applications, they may fall apart as
seen in the experiments. It is important to note that while visual fidelity and
quality are important from a computer vision and graphics perspective, when
considering robotics, other factors such as precise occupancy and pose may be
more crucial.
When considering the proposed method, though it is scalable to account for more
object classes, it also adds to the requirement of extensive databases of
synthetic models for each object class and the computational overhead to search
through those databases to find a match. The process is time consuming for
larger environments, but can be performed offline. However, despite all these
factors, until deep learning methods improve their ability to transfer, our
experiments confirm model matching as the most reasonable choice for object
completion for robotics applications.
\section{Experiments}
\label{sec:exp}
In this section, we present experiments aiming to validate the proposed method
by answering the following questions:
\begin{enumerate}
\item is the pipeline able to reconstruct environments while navigating in
them?
\item when considering chairs, how reliably can the proposed method locate
them and estimate their full extent?
\item how does the object extent estimated by the proposed method compare
with state-of-the-art{} shape-completion deep learning methods?
\end{enumerate}
In order to answer these questions we conducted multiple experiments in
simulation (\secref{sec:sim}) and on real robotic platforms (\secref{sec:real}).
As a measure to evaluate the method's ability to recognize objects, we used the
F\textsubscript{1}-score{},
\begin{equation}
F_1 = 2\frac{Precision \cdot Recall}{Precision+Recall}.
\end{equation}
All experiments have been run on a laptop with Intel Core i7-8750H, 2.20GHz CPU,
Nvidia GeForce 1050 Ti GPU, and 16 GB RAM. While they are not presented here due
to space limitations, we conducted more experiments that can be found in
\cite{kpthesis2020}.
Finally, in \secref{sec:nav}, we will demonstrate how the proposed method can
improve navigation safety.
\subsection{Simulation experiment}
\label{sec:sim}
In order to evaluate the proposed method, we setup a simulated office
environment \cite{rasouliEffectColorSpace2017} in Gazebo. The environment, shown
in \figref{fig:office}, consists of a few different furnished zones, comprising
a meeting table, a couple of lounge areas, and an a few desks. In total, the
environment contains 19 chairs distributed among the different zones. 8 chairs
around a long meeting table, 6 chairs facing the wall with some objects in
between them, 4 office chairs facing one of the desks, and partially slid
inside, and one chair near a long table facing the room. We spawn in the
environment a simulated model of a Husky robot from Clearpath Robotics, equipped
with a Kinect RGBD camera and a SICK laser scanner.
\begin{table}
\centering
\caption{Experimental results}\label{tab:results}
\begin{tabular}{lccc}
\toprule
Experiment & Precision & Recall & F\textsubscript{1}-score{}\\
\midrule
Simulated office & 0.69 & 0.59 & 0.63\\
Real office & 0.71 & 0.83 & 0.78\\
Real corridor & 0.80 & 1.00 & 0.89\\
\bottomrule
\end{tabular}
\end{table}
The first aspect we wanted to evaluate in simulation was the ability of the
proposed method to correctly identify and then shape-complete the chairs present
in the environment. To this end, we manually teleoperated the robot to construct
a multi-layer map $\mathcal{M}{}$ of the environment using the proposed method. We
empirically set the range for $\lambda$ to $[0.1,0.25]$ m, and $\epsilon = 0.1$
m.
A total of 16 chairs were identified and shape-completed. Out of the 19 chairs
in the environment, 11 were correctly identified and the other 5 were false
positives. The first row of \tabref{tab:results} shows the achieved precision,
recall, and F\textsubscript{1}-score{}. The results are quite modest. Misdetections were caused
by inaccurate segmentations provided by Mask R-CNN, that were not fully
compensated for by the filter described in \secref{sec:completion}.
\figref{fig:false_pos} provides a top view of the scene after object replacement
and the marked objects denote the false positives, and \figref{fig:mm_tab} zooms
over the meeting table, where most of the chairs have been completed properly
except for one false positive arising from one label spreading on the table top.
Mask R-CNN represents current state-of-the-art{} capabilities in terms of object
segmentation, the results presented here are expected to improve with better
object segmentation.
Another investigated aspect was the quality of shape-completion produced by the
proposed method, in comparison to state-of-the-art{} deep learning methods. To this end,
we compared with \ac{pcn} \cite{yuanPCNPointCompletion2018}.
\figref{fig:pcn_perform} shows some examples of the output of \ac{pcn} for
chairs mapped at different level of completion; even when the input is lacking
only few regions, the output produced by the network has visible outliers and
the internal pose estimation of the network fails. When a similar chair in a
different pose as in \figref{fig:sim_3} is subjected to the network, the
confusion of the network is quite evident as it tries to recreate the backrest
resulting in a box-like structure above the legs. \figref{fig:sim_no} and
\ref{fig:sim_1} are the same chairs in different orientations. For the former,
the network is unable to map it to a chair, whereas for the latter it is not
able to infer the extent of legs with one partial leg. All these observations
point to the fact that the network performance does not seem to translate to
realistic partial views that can be obtained through acquisition ``in the
wild''.
\begin{figure}
\centering
\begin{subfigure}[b]{0.49\linewidth}
\centering
\includegraphics[width=\linewidth]{images/pcn_simdata.png}
\caption{Four legs}
\label{fig:sim}
\end{subfigure}%
\begin{subfigure}[b]{0.49\linewidth}
\centering
\includegraphics[width=\linewidth]{images/pcn_bad2.png}
\caption{Three legs}
\label{fig:sim_3}
\end{subfigure}
\begin{subfigure}[b]{.49\linewidth}
\centering
\includegraphics[width=\linewidth]{images/pcn_bad3.png}
\caption{One leg}
\label{fig:sim_1}
\end{subfigure}%
\begin{subfigure}[b]{.49\linewidth}
\centering
\includegraphics[width= \linewidth]{images/pcn_bad.png}
\caption{No legs}
\label{fig:sim_no}
\end{subfigure}
\caption{Performance of \ac{pcn} on chairs with various degree of
completion. Input (left) and output (right) are in the same pose,
misalignment is due to a failure of the network.}\label{fig:pcn_perform}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}[b]{\linewidth}
\centering
\includegraphics[width=.23\linewidth]{images/perfect.png}
\includegraphics[width=.23\linewidth]{images/p_match.png}\hfill
\includegraphics[width=.23\linewidth]{images/good_chair.png}
\includegraphics[width=.23\linewidth]{images/good_chairmatch.png}
\caption{Input has data for backrest, seat, and
leg/legs.}\label{fig:mm_good}
\end{subfigure}
\begin{subfigure}[b]{\linewidth}
\centering
\includegraphics[width=.23\linewidth]{images/office_chair.png}
\includegraphics[width=.23\linewidth]{images/match_off_chair.png}\hfill
\includegraphics[width=.23\linewidth]{images/normal_chair.png}
\includegraphics[width=.23\linewidth]{images/weird_chair.png}
\caption{Input has a key component missing}\label{fig:mm_bad}
\end{subfigure}
\caption{Matched models using the proposed method. Input point-cloud{}s and model
outputs are shown in yellow and blue respectively
(best viewed in color){}.}\label{fig:mm}
\end{figure}
The errors seen with \ac{pcn}, \textit{i.e.},{} noise and outliers, are not present when
using the proposed model matching system. In this case, what influences the
output quality is the region of missing information rather than its quantity
(\figref{fig:mm}). When the input has parts of all the factors that define a
chair which are legs, backrest, seat, and arms a good approximate match is
obtained as seen in \figref{fig:mm_good}. However, when a key component is
completely missing from the input, the match may differ significantly from the
actual object. In \figref{fig:mm_bad}, we can see that a wheeled office chair
lacking arms and the wheeled base in the reconstruction is completed with a
circular base instead of a wheeled base (left), while, when only a partial
extent of one leg is present (right), the corresponding match found has shorter
legs and if this model is placed back in the scene it would hover in the air.
The output also has a set of arms that are not present in the actual object.
From the observations, it is clear that a complete extent of at least one leg is
needed in the input to find a good match that would be properly grounded.
However, even when a different chair model is identified as match, the pose and
overall shape produced by the proposed method are still consistent with the
observed object, which may make these models useful for robotics applications,
as demonstrated in \secref{sec:nav}.
\subsection{Physical experiment}
\label{sec:real}
The tests on simulation provided insights into the performance, the factors
affecting it, and the bottlenecks in the pipeline. With these limitations in
mind, the viability of the pipeline was tested on a Care-O-bot 4
\cite{DBLP:conf/mc/KittmannFSRWH15} robot produced by Fraunhofer IPA. The robot
is equipped with multiple RGBD cameras on the head, neck, and torso to perceive
the area in front of it. It is capable of omnidirectional motion with a maximum
velocity of 1.1 m/s. Three 2D laser scanners mounted in the base allow the robot
to react to static and dynamic obstacles. For the experiments only the RGBD
camera mounted in the neck area was used. We mapped two environments using the
proposed method: an office where we randomly placed some chairs, and a corridor
with chairs along the wall.
\subsubsection{Office}
\label{sec:lab}
\begin{figure}
\centering
\begin{subfigure}[b]{.49\linewidth}
\includegraphics[width=\linewidth]{images/same_area_real_crop.png}
\caption{Office}\label{fig:real_office}
\end{subfigure}
\begin{subfigure}[b]{.49\linewidth}
\includegraphics[width=\linewidth]{images/same_area_crop.png}
\caption{Office with replaced chairs}\label{fig:real_office_res}
\end{subfigure}
\begin{subfigure}[b]{.49\linewidth}
\includegraphics[width=\linewidth]{images/corridor_real_crop.png}
\caption{Corridor}\label{fig:real_corridor}
\end{subfigure}
\begin{subfigure}[b]{.49\linewidth}
\includegraphics[width=\linewidth]{images/corridor_crop.png}
\caption{Corridor with replaced chairs}\label{fig:corridor_res}
\end{subfigure}
\caption{Pipeline performance}\label{fig:real}
\end{figure}
7 chairs of multiple designs were laid out in a chaotic manner in the room. The
robot was teleoperated through the environment for one minute and the
environment was mapped with the proposed pipeline using data captured through
the RGBD camera mounted in the neck. \figref{fig:real_office} shows the
environment as reconstructed by the pipeline (layer $\matr{G}{}$), while
\figref{fig:real_office_res} shows the augmented scene $\matr{A}{}$ with the
complete object models.
As shown in \tabref{tab:results}, the setup yielded an F\textsubscript{1}-score{} of 0.78,
significantly better than the simulation. Of the five correctly detected chairs,
two (chairs 1 and 4) have good reconstruction quality, and three are reasonable
but incomplete: chair 3 because of noise in depth measurements, chair 6 because
of inaccurate semantic segmentation, and chair 5 because of heavy occlusion in
data. Two false positives were caused by inaccurate semantic segmentation of
chair 6. The wheeled base, backrest, and seat were registered whereas the stem
connecting the seat with the base was not registered, causing a false positive
seen below chair 6. Another false positive was the result of label spreading of
Mask R-CNN by which part of the wall behind chair 6 got mislabeled.
\subsubsection{Corridor}
\label{sec:corridor}
4 chairs were laid out in a narrow corridor and the robot was immobile in a
single observation position. \figref{fig:corridor_res} depicts the result of the
pipeline. The F\textsubscript{1}-score{} of the reconstruction quality was 0.89
(\tabref{tab:results}), which shows a level of performance above the previous
cases. Out of four correct detections, three were accurately reconstructed while
one was incomplete due to sensor noise. There was one false positive due to
label spreading.
In summary, the inherent shortcomings of Mask R-CNN and label spreading were
present also in the real environment. Additionally, it was observed that major
occlusions and noise in the depth information influence the quality. However,
even if the geometric shape of distant or heavily occluded chairs was
inaccurate, their pose was matched quite accurately. Correct pose estimation is
a crucial property in robotics, as we will show in the demonstration that
follows.
\subsection{Application: Shape-aware autonomous navigation}
\label{sec:nav}
\begin{figure}
\centering
\begin{subfigure}[b]{0.49\linewidth}
\centering
\includegraphics[width=.9\linewidth]{images/coffeeroom_real_crop.jpg}
\caption{Coffee room}
\label{fig:coffeeroom}
\end{subfigure}
\begin{subfigure}[b]{.49\linewidth}
\centering
\includegraphics[width=.9\linewidth]{images/corridor_real_crop.jpg}
\caption{Corridor}
\label{fig:corridor}
\end{subfigure}
\begin{subfigure}[b]{.49\linewidth}
\centering
\includegraphics[width=.9\linewidth]{images/coffeeroom_slam_crop.png}
\caption{Coffee Room: \ac{slam}}
\label{fig:coffeeroom_slam}
\end{subfigure}
\begin{subfigure}[b]{.49\linewidth}
\centering
\includegraphics[width=.9\linewidth]{images/corridor_slam_crop.png}
\caption{Corridor: \ac{slam}}
\label{fig:corridor_slam}
\end{subfigure}
\begin{subfigure}[b]{0.49\linewidth}
\centering
\includegraphics[width=.9\linewidth]{images/coffeeroom_multimap_crop.png}
\caption{Coffee Room: Multimap}
\label{fig:coffeeroom_multimap}
\end{subfigure}
\begin{subfigure}[b]{.49\linewidth}
\centering
\includegraphics[width=.9\linewidth]{images/corridor_multimap_crop.png}
\caption{Corridor: Multimap}
\label{fig:corridor_multimap}
\end{subfigure}
\caption{Navigation experiment results over two scenarios. By navigating
using \ac{slam} map only, \subfigref{fig:coffeeroom_slam} and
\subfigref{fig:corridor_slam}, the robot collides with chairs, while it
avoids all chairs by navigating on the multimap,
\subfigref{fig:coffeeroom_multimap} and \subfigref{fig:corridor_multimap}
(best viewed in color){}.}\label{fig:navigation}
\end{figure}
As a demonstration of practical usefulness of the multi-layer mapping framework
proposed in this work, we explore its use for robotic navigation. Most robots
navigate using 2D occupancy maps built using 2D lidar-based \ac{slam}, however
those maps only represent occupancy at one specific height and fail to capture
the occupancy of complex obstacles. In this demonstration, we custom fitted a
Turtlebot3 Waffle Pi robot with a Kinect RGBD camera mounted on a short pole to
improve its field-of-view. We setup two environments, shown in
\figref{fig:coffeeroom} and \figref{fig:corridor}, where the robot is on one
side of some chairs and has to navigate to the other side. The chair seats
provide obstacles for the robot such that if the robot plans a trajectory
through the legs of a chair, the pole will collide with a seat.
First, the robot navigated using a 2D lidar-based map and, as can be seen in
\figref{fig:coffeeroom_slam} and \figref{fig:corridor_slam}, the resulting
trajectories passed through the chairs and caused a collision. Then, estimated
object occupancies were projected down to build another 2D costmap, by
projecting all points in the object layer having $z \in [0.1, h]$, where $h$ was
the height of the robot. Using this map, the robot navigated avoiding all
obstacles and reached its destination (\figref{fig:coffeeroom_multimap} and
\figref{fig:corridor_multimap}).
This demonstration serves first of all to illustrate the usefulness of
multi-layer maps that allow integration of information across sensors and map
layers. Secondly, it shows how the knowledge maintained by proposed mapping
framework can enable reasoning on multiple representations to increase robotic
safety, autonomy, and reliability.
\section{Introduction}
\label{sec:intro}
In recent years, we have witnessed great advances in computer vision which have
given robots the ability to understand the world around them like never before
\cite{heMaskRCNN2017, yuanPCNPointCompletion2018}. At the same time, advances in
control of mobile platforms have enabled robots from quadrupedal robot dogs to
drones to move in different environments and react to unpredicted circumstances
with high reliability.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{images/cover_version2.png}
\caption{An environment representation built using the proposed
framework.}\label{fig:map}
\end{figure}
However, when considering a mobile robot’s ability to perform advanced
high-level tasks in unstructured human-inhabited environments, there is a
crucial bottleneck stifling the ability of these sensing and acting
breakthroughs to lead to the desired level of autonomy, performance, and
interaction: the way robots reason on this rich perceptual knowledge is still
very limited. When considering navigation for example, ground robots still rely
on 2D occupancy maps built using \ac{slam} algorithms for path planning and
obstacle avoidance. These maps are built using only laser scanners and the rich
information coming from the variety of visual sensors available to modern robots
is mostly ignored. The combination of these sensors would allow the robot to
formulate a better understanding of its surroundings and, in the case of
navigation, to take more intelligent choices to plan its route, \textit{e.g.},{} by
avoiding obstacles invisible to the laser scanner. Similarly, while interacting
with humans, understanding object semantics is crucial for understanding
commands in everyday language \cite{pandey2014towards}; having a richer
understanding of the environment opens the door for complex capabilities such as
mobile manipulation and natural user interaction.
Though methods exist in literature to construct a meaningful 3D representation
of the environment \cite{newcombeKinectFusionRealtimeDense2011,
oleynikovaVoxbloxIncremental3D2017, mccormacSemanticFusionDense3D2017}, the
acquisition of a full 3D map of an environment is a time-consuming process. In
most realistic scenarios, the underlying geometry reconstructed by these methods
is often incomplete, due to time concerns, occlusions, or environmental
traversability limitations affecting the perception of the scene during the
mapping process. However, partial geometry is a considerable drawback in some
applications; for example, when a robot has to interact with an object, the lack
of knowledge of complete geometry can pose safety risks either to the object,
the surroundings, or the robot itself. Moreover, complete objects model are also
essential for user immersion and safe navigation in telepresence applications
\cite{cesta2012into}.
In this paper, we propose a mapping pipeline to construct 3D metric-semantic
environment representations which include estimates of the full extent of
objects in the environment through object shape completion. An example of this
representation is shown in \figref{fig:map}. We also propose a novel
deterministic approach based on matching (and potential replacement) of a
partial object with a similar synthetic model from a known database, which
avoids the unpredictability of recent deep-learning-based shape completion
approaches. In the experiments, we validate our approach by comparing it to a
state-of-the-art{} deep-learning-based approach for shape completion of chairs.
To summarize, the contributions of this paper are:
\begin{itemize}
\item A pipeline to generate a 3D metric-semantic mesh representation of an
environment that completes partial objects with similar complete synthetic
models.
\item Design and evaluation of a novel method to compare and match partial
observations of 3D object instances against a database of synthetic models
and to replace the partial 3D mesh with the matched model in the
reconstructed 3D mesh of the environment.
\item Evaluation and comparison of a learning-based approach to complete
partial observations of 3D object instances against our proposed method.
\item A demonstration of benefits of the proposed environment representation
in a 2D navigation case on a real TurtleBot 3 Waffle{}.
\end{itemize}
\section{Method}
\label{sec:method}
The proposed pipeline, shown in \figref{fig:full_pipe}, is split into
\emph{Realtime Reconstruction} and \emph{Scene Augmentation}. \emph{Realtime
Reconstruction} includes modules to build the geometric and semantic
representations. These reconstructions form the input to the second half of the
pipeline, that performs \emph{Scene Augmentation} on the input to deliver the
final geometric representation of the scene having objects with complete
geometry. All representations are kept aligned to a 2D occupancy map $\matr{M}{}$
built through \ac{slam}. The proposed pipeline consists of the combination of
existing software components along with novel components to produce the final
representation. While developing the pipeline, the emphasis was kept on the flow
of data rather than on the individual component implementations. Hence, each
component is independent from the others, thereby easily replaceable with any
current or future work that can ensure the same data flow.
\subsection{Realtime reconstruction}
\label{sec:recon}
The purpose of the reconstruction module is to generate the 3D geometric
representation $\matr{G}{}$ and the semantic-geometric representation $\matr{S}{}$
online. Both layers will be represented as 3D colored meshes; in $\matr{G}{}$, each
face's color will represent the color information obtained from the camera,
while in $\matr{S}{}$ the color will map to a semantic object class. Semantic
information is gathered through a deep learning based pixel-level instance
segmentation method, the output of which along with the respective depth image
and pose information is used to generate the semantically annotated 3D mesh. In
this work, we use Voxblox \cite{oleynikovaVoxbloxIncremental3D2017} and
Kimera-Semantics \cite{rosinolKimeraOpenSourceLibrary2020} to construct the
geometric and semantic-geometric representation respectively, employing Mask
R-CNN \cite{heMaskRCNN2017} as the deep learning algorithm to obtain the object
semantic segmentation from images.
This stage outputs $\matr{G}{}$ and $\matr{S}{}$, \textit{i.e.},{} a geometric and a
semantic-geometric representation of the environment available in a mesh format
which are converted to point-cloud{}s due to ease of processing for the latter stage of
the pipeline.
\begin{figure}
\centering
\begin{subfigure}[b]{.49\linewidth}
\centering
\includegraphics[width=\linewidth]{images/chair_wall_final.png}
\caption{Spread onto wall and floor}\label{fig:wall}
\end{subfigure}
\begin{subfigure}[b]{.49\linewidth}
\centering
\includegraphics[width=\linewidth]{images/pcn_table_final_crop.png}
\caption{Spread onto a table}\label{fig:table}
\end{subfigure}
\caption{Discrepancies caused by inaccurate object mask. In both pictures
the marked regions represent erroneous spread of a label
(best viewed in color).}\label{fig:maskrcnn_desc}
\end{figure}
\subsection{Object shape completion and scene augmentation}
\label{sec:completion}
In order to obtain the object instances to shape-complete, the semantic point
cloud is segmented into individual clusters, with each cluster representing a
partial view $\tilde{o}$ of an object $o$. Object instances are identified using
differences of normals. Then, given a semantic class $c$, for each partial view
$\tilde{o}$ for which $\matr{S}{}(\tilde{o}) = c$, we propose an object
shape-completion method $f_c$ to estimate its full extent $o^*$. Formally, $f_c:
\tilde{o} \to o^*$. In this work, we demonstrate the augmentation by replacing
chairs, but the proposed methodology can be applied to other classes of objects
as well.
\subsubsection{Shape Filter}
\label{sec:filter}
After clustering, it can happen that some clusters $\tilde{o}$ may be
incorrectly classified to class $c$. This can happen \textit{e.g.},{} for issues of time
synchronization between RGB and depth images, or because the mask obtained from
Mask R-CNN may contain parts that do not belong to the object. The marked
regions in \figref{fig:maskrcnn_desc} are examples of such cases. To avoid
completing the shapes of mislabeled objects, we pass each cluster through two
filters aiming at recognizing erroneous labeling.
The first filter aims at recognizing label-bleeding along walls and floor by
identifying planarity of a cluster point-cloud{} by its covariance matrix. A cluster is
discarded if the covariance matrix shows values close to zeros for at least one
of the axes.
The second filter aims at identifying mislabeled objects. Objects of a certain
class tend to have similar shape and size, and these can be used to identify
objects which do not conform to the class model. In practice, we determined the
distance of the farthest point to the object centroid $\lambda$ and removed
objects for which this measure was out of range for any sort of chair commonly
found in an office or home environment. Despite the filtering, some false
positives may remain for later stages.
\subsubsection{Shape matching}
In order to estimate the full extent of an object, we match its partial shape
$\tilde{o}$ to a similar synthetic model $o^*$. Formally, given $\tilde{o}$, a
partial view of an object of class $c$, and $\matr{C} = \{o_i\}_{i=1}^N$, a
database of $N$ object models $o_i$ of class $c$, we want to find
\begin{equation}\label{eq:match}
o^* = \argmin_{o_i \in \matr{C}} \delta(o_i, \tilde{o})
\end{equation}
for some distance metric $\delta$. In this study, a custom point-cloud{} database of
chairs was created from the 6778 3D models available in the ShapeNet dataset
\cite{shapenet2015}.
In order to compute the distance $\delta$, $o_i$ and $\tilde{o}$ need to first
be registered to the same pose.
\begin{figure*}
\centering
\begin{subfigure}[b]{.37\linewidth}
\centering
\includegraphics[height=5.5cm]{images/office.png}
\caption{Top view from Gazebo}
\label{fig:office}
\end{subfigure}
\begin{subfigure}[b]{.37\linewidth}
\centering
\includegraphics[height=5.5cm]{images/aug_scene.jpg}
\caption{Augmented scene}\label{fig:false_pos}
\end{subfigure}
\begin{subfigure}[b]{.24\linewidth}
\centering
\includegraphics[height=5.5cm]{images/zoomed_in_meeting.png}
\caption{Meeting table detail}\label{fig:mm_tab}
\end{subfigure}
\caption{Simulated office scene \subfigref{fig:office}. In the 3D geometric
model \subfigref{fig:false_pos}, chair partial observations have been
replaced with the complete models and each area marked in red in
\subfigref{fig:false_pos} denotes a false positive. A detail of the meeting
table in the North-West corner of the environment is shown zoomed in
\subfigref{fig:mm_tab} (best viewed in color){}.}\label{fig:performance}
\end{figure*}
\subsubsection*{Pose matching}
\label{sec:pose}
Inferring the pose of an object directly from its raw partial point-cloud{}
representation is quite difficult. Instead, it is easier to find a
transformation that can be applied to the model that will align it with the
object through point-cloud{} registration methods through \ac{icp}. However, \ac{icp} is
a computationally expensive algorithm and is susceptible to getting stuck at
local minima, so it is best if any initial estimate of the transformation is
available to loosely align the source (database models) with the target (partial
point-cloud{}s). \ac{icp} can then fine-tune the coarse transformation in very few
iterations to tighten the alignment. Given that the natural pose of any grounded
objects will always be around the $z$-axis, to find a coarse transformation, a
random model from the dataset is selected and a set of uniformly sampled
rotations are performed around its $z$-axis. The transformation matrix for the
rotation yielding the minimum average point-to-point distance is used to
initialize the \ac{icp} algorithm. \ac{icp} fine-tunes this rotation over the
$z$-axis to provide the final transformation matrix. The advantage of this
approach is that it is applicable to any grounded object.
\subsubsection*{Model matching}
\label{sec:match}
Finding a match involves searching through the database. The final
transformation refined after \ac{icp} method is applied to each model and the
corresponding point-to-point distance with the partial point-cloud{} is recorded. This
is used as a measure of model distance $\delta$. The hypothesis is that the
model that most closely resembles the actual object should return the smaller
distance. The matched model $o^*$ in the local frame of the partial object is
re-scaled and translated back into the world frame and added to the object
representation layer $\matr{O}{}$.
\subsubsection{Scene augmentation}
Finally, after having built $\matr{G}{}$, $\matr{S}{}$, and $\matr{O}{}$, we build an
augmented scene, \textit{i.e.},{} a representation where we replace objects in $\matr{G}{}$
with their complete counterparts in $\matr{O}{}$. While not constituting one of the
layers of $\mathcal{M}$, this additional \emph{virtual} representation can be used
for planning and navigation, as we will show in \secref{sec:nav}. To this end,
we construct a merged point-cloud{} $\matr{A}{} = (\matr{G}{} \setminus \matr{S}) \cup
\matr{O}{}$, where $\matr{S} \subset \matr{G}{}$ is the subset of points of $\matr{G}{}$
that are within a point-to-point distance threshold $\epsilon$ from any of the
object models in $\matr{O}{}$.
\section{Problem formulation}
\label{sec:prob}
The aim of this work is to build rich map representations where geometric
information of the environment is maintained together with the object semantics.
Additionally, we want to maintain an estimate over the full extent of each
object.
Formally, we want to build a hierarchical multi-layer map $\mathcal{M}{} = \{\matr{M}{},
\matr{G}{}, \matr{S}{}, \matr{O}{}\}$, composed of the following layers:
\begin{enumerate}
\item 2D geometry $\matr{M}{}$: a representation of traversable area, often used
for navigation and obstacle avoidance;
\item 3D geometry $\matr{G}{}$: a representation of the environment where
objects can be recognized by their appearance;
\item 3D semantics $\matr{S}{}$: a representation where semantic information
over objects in the environment is maintained;
\item Object instances $\matr{O}{} = \{o^*\}$: for each object in the
environment, we want to estimate its full extent $o^*$ from a partial view
$\tilde{o}$.
\end{enumerate}
In the next section we will discuss the proposed pipeline, which builds all the
aforementioned maps iteratively by means of visual and range information
obtained by a robotic platform exploring the environment.
\section{Related Work}
\label{sec:related}
Building environment representations where complete models of objects are
estimated requires first reconstructing the 3D geometry of the environment,
understanding its semantics, and then shape-completing each object instance. In
this section, we will review state-of-the-art{} relevant to these fields, together with
relevant works in multi-layer mapping for robots.
\paragraph*{Reconstruction of 3D geometry}
The availability of inexpensive RGB-D cameras such as Kinect has resulted in
profound advances in developing 3D reconstruction methods. KinectFusion
\cite{newcombeKinectFusionRealtimeDense2011} popularised \ac{tsdf}-based
reconstructions as capable of real-time reconstruction of small environments.
The main drawback of KinectFusion is the use of a fixed-size voxel grid which
requires the map size to be known and a large amount of memory. Multiple
extensions have been proposed to address these shortcomings \cite{Whelan12rssw,
article}. Among these, Nießner \textit{et al.}{} \cite{10.1145/2508363.2508374} propose a
spatial hashing scheme to compress space while allowing real-time access and
updates of underlying surface data. This spatial hashing was faster than the
hierarchical grid data structure used in other methods, however, it relies
heavily on GPUs for real-time performance. By exploiting the spatial hashing
technique employed in \cite{10.1145/2508363.2508374}, Voxblox
\cite{oleynikovaVoxbloxIncremental3D2017} focused on improving memory efficiency
and real-time performance on the CPU by building incremental \acp{esdf} from
\acp{tsdf}. This makes Voxblox able to reconstruct large scale scenes with
reasonable computational costs. In this work, since we are interested in
building 3D model of complete environments, we employ Voxblox in the geometric
reconstruction module.
\paragraph*{Semantic reconstruction}
The recent innovations in image segmentation methods enabled semantic knowledge
to be incorporated into 3D reconstructions into so called semantic-metric
representations. Among the methods to perform 3D semantic mapping,
SemanticFusion \cite{mccormacSemanticFusionDense3D2017} combines the real-time
SLAM system ElasticFusion \cite{Whelan2015ElasticFusionDS} and a CNN for object
detection with a Bayesian update scheme for semantic label integration,
Voxblox++ \cite{grinvaldVolumetricInstanceAwareSemantic2019}, an extension of
Voxblox, combines geometric segmentation of depth data with instance
segmentation by Mask R-CNN \cite{heMaskRCNN2017}. A data association strategy
keeps track of labels to ensure global consistency. On the same line,
Kimera-Semantics \cite{rosinolKimeraOpenSourceLibrary2020} uses Voxblox allowing
however flexibility for choice for the 2D instance segmentation algorithm. It
runs faster than \cite{grinvaldVolumetricInstanceAwareSemantic2019} and shows
accurate reconstruction when compared against ground-truth. In this work, we
employ Kimera-Semantics with Mask R-CNN to build semantic representations.
\paragraph*{Shape completion}
Shape completion is the process for inferring a full 3D model of an object based
on a partial measurement. Two main approaches to the problem exist:
model-matching and generative. Model-matching methods attempt at matching a
partial view to a similar model. The most recent example of this class of
methods is Scan2CAD \cite{Avetisyan_2019_CVPR} which uses a 3D CNN trained on a
custom dataset to learn joint embedding between real and synthetic 3D models to
predict accurate correspondence heatmaps between model and scan. It requires
entire 3D scan of a scene as input and tries to find the alignment of models of
all the objects in the scene. The method, however, does not utilize any semantic
information of the scene. Recently, generative deep learning methods for shape
completion have attracted increasing interest in the community. Although methods
working on voxel data exist \cite{daiShapeCompletionUsing2017}, point-cloud{}s are
better suited to represent objects at different resolution. Among the methods
working on point-cloud{}s \cite{huangPFNetPointFractal2020a, DBLP:conf/iclr/ChenCM20,
Tchapmi_2019_CVPR}, one of the best performing ones is \ac{pcn}
\cite{yuanPCNPointCompletion2018}, a shape completion method based on the
PointNet architecture. It is capable of generating high-resolution completions
and shows generalization over unseen objects and real-world data. However, most
of the results presented in works on deep learning shape completion only pertain
synthetic data and their ability to transfer from synthetic-partial data to real
data has been questioned \cite{DBLP:conf/iclr/ChenCM20}. Moreover these methods
often do not account for much input noise while training. The problem of scene
completion (\textit{i.e.},{} shape-completing all objects in a scene) is less explored and
in that context, which we target in this work, errors in semantic segmentation
or geometric reconstruction of the scene can further increase the transfer gap
and lessen the performance of deep learning methods even more. In this work, we
explore the ability of \ac{pcn} to transfer to noisy real data as well as
propose a novel model-matching approach for scene completion and augmentation.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{images/pipeline_final.png}
\caption{Detailed overview of the pipeline}\label{fig:full_pipe}
\end{figure*}
\paragraph*{Multi-layer mapping}
Multi-layer maps for robotics are not a new concept. Many works have proposed
hierarchical mapping methods where a 2D occupancy map is maintained together
with traversability graphs and topology \cite{zender_conceptual_2008,
crespoRelationalModelRobotic2017}. However, these architectures rarely include
multiple maps capturing the environment as seen by different sensor modalities
but rather abstract the environment to facilitate reasoning and human-robot
interaction.
Recently, motivated by the advancements in computer vision and machine learning,
interest has increased toward multi-layer mapping formalisms able to include
different sensor information. In \cite{zaenkerHypermapMappingFramework2019} a 2D
multi-layer mapping framework composed of a metric, semantic, and exploration
layers is proposed, and its application in the context of autonomous semantic
exploration is presented. In that work, the focus is left on 2D mapping only,
and no object shape completion is attempted. Rosinol \textit{et al.}{}
\cite{rosinol3DDynamicScene2020} propose a hierarchical graph comprised of
multiple layers, including one where object extents are estimated by fitting a
CAD model to the partial object. In this work, we propose a more general
approach to shape completion, by matching over a database of object models,
instead of a single one.
|
\section{Introduction}
Especially governmental organizations and agencies provide different types of open data sources in several countries today. While not new, open data has not always been available via convenient Application Programming Interfaces (APIs) as the data can also be provided only as documents \cite{Gonzalez2020model}. However, for example in Finland, this will be resolved because the recent Finnish legislation demands governmental organizations to provide APIs for their public data adhering to the European Union directive \cite{EUdirective}.
The same legislation, thus, increases the amount of available open data and open APIs as different data will be opened. Additionally, this enables more near real-time data when data will become available automatically through the APIs.
As the quantity of data and data sources grow massively, a need and opportunity emerge for data science services that can process huge amounts of data. Data can be seen as a service proposal per se, but with open data, the commercial product is rather the intelligence built on top of the data to solve the specific contextual problems of the customers. More and more often, this takes place in the form of artificial intelligence (AI) and machine learning (ML) systems.
However, provided APIs essentially define and control what operations can be performed on open data, by whom, and under what conditions.
For open data usage, there are challenges, such as legal or privacy issues and possible changes in governmental policies~\cite{Martin2013risk}.
Challenges are also associated with the creation and evolution of data science services, such as ensuring adequate efficiency in processing large data sets (data-intensive flows management)~\cite{Abello2017data}.
This experience report discusses the following problem: How to build data science services on top of a set of open data providers by combining open data sources and carrying out advanced analyses, such as machine learning so that the results are valuable for end-customers? Our firsthand experiences originate from the development of open data and APIs for marine traffic. This open data in association with other open data sources introduce new sustainable data science service opportunities for end-customers based on various AI technologies. Towards this end, we enumerate the challenges we have identified related to these opportunities and propose how API governance could alleviate the challenges.
\section{Case: Artificial intelligence for the Finnish–Swedish Winter Navigation System}
\subsection{Background: Marine traffic}
We base our firsthand experiences on marine traffic data in Finland and Sweden, especially during wintertime when the Finnish–Swedish Winter Navigation System (FSWNS) \cite{bergstrom2020simulation} is active. FSWNS maintains safe and efficient year-round navigation with agreements and information systems. The Finnish public authority (Fintraffic) provides near real-time monitoring of traffic through public APIs\footnote{https://www.digitraffic.fi/en/}.
This case was selected because Solita Ltd\footnote{http://www.solita.fi}, a consultancy company with over 1000 employees, developed the DigiTraffic API for Fintraffic and two authors of this paper work full time at Solita.
Therefore, we are familiar with the technologies, application domain, possible business opportunities, and challenges.
Marine traffic is quintessential for Finland: some 80-90\% of exported and imported goods are carried by sea \cite{tapaninen2020maritime}. In winter navigation, the changing ice conditions \cite{banda2015risk} cause relatively frequently accidents \cite{goerlandt2017analysis} which may trigger, for instance, oil spills and delays \cite{banda2016risk}.
The ecosystem related to marine traffic is large as the total number of companies working in the Finnish maritime cluster\footnote{https://shipowners.fi/en/maritime-cluster/} is almost 3000 \cite{tapaninen2020maritime}.
The marine environment of FSWNS is very special and challenging as it consists of shallow and narrow sea lanes, dense and rocky archipelago, and icy conditions as especially the Bay of Bothnia freezes during the wintertime \cite{lehtola2019finding}. These conditions do not only require piloting but also ice breaking in winter when ships may even be guided in a convoy or towed behind an ice breaker.
The icebreakers assist vessels free of charge to enable fluent foreign trade. However, unexpected delays of even tens of hours are typical\footnote{https://vayla.fi/en/transport-network/waterways/winter-navigation} due to changing ice conditions which then affect the inter-modal logistic chain \cite{tapaninen2020maritime}. Nowadays, the icebreaker captains try to interpret an ice forecast\footnote{http://baltice.org/weather/}, a wind forecast\footnote{https://www.windy.com} and satellite images to predict which vessels might get stuck into the ice and need assistance.
\subsection{Opportunities}
Novel data science service utilizing AI technologies could provide help for the decision making process for the icebreaker captains, but also several other parties in the ecosystem would gain benefits from better information: ports, shipping companies, cargo forwarding companies, and transport companies \cite{tapaninen2020maritime} -- and their customers.
Business-critical opportunities include how to predict and control the estimated time of arrival (ETA) or departure more accurately based on ice and weather conditions. Also, these opportunities may include identifying potential future traffic bottlenecks --
waiting does not only mean staying on hold and wasting time at the sea but also that the speed of the vessel could have been lower leading to savings if the waiting was anticipated. Respectively, sudden stop, acceleration, or un-optimized route of a large vessel is equally costly. All these accumulate CO$_2$ emissions.
Many manufacturing and assembly companies also depend on predictive just-in-time import and export in their business processes \cite{tapaninen2020maritime}. The logistics are not limited to marine but include storage, road, and rail.
Finally, there is an opportunity to enhance safety, such as predict and prevent a collision in a convoy \cite{jussila2018visualising}.
Future opportunities lie in technological advancements. Automation is increasingly important in the ports and heavily automated ports already exist outside Finland (e.g., Hamburg and Rotterdam). Also, the ships rely more and more on automation and autonomous cargo ships are being developed and tested.
The public authorities have already made several open data APIs available that are published following the aforementioned legislation. For example, the following APIs are relevant and available:
\begin{enumerate*}
\item Digitraffic (Ministry of Transport and Communications) traffic APIs including
marine\footnote{https://www.digitraffic.fi/en/marine-traffic/} and rail\footnote{https://www.digitraffic.fi/en/railway-traffic/};
\item Finpilot piloting status (government-owned piloting company)\footnote{ https://pilotonline.fi/traffic-info/api};
\item Finnish Meteorological Institute weather\footnote{https://en.ilmatieteenlaitos.fi/open-data-manual};
\item many others gathered to the Open data webpage\footnote{https://www.avoindata.fi/en}
\end{enumerate*}
The increasingly popular open data facilitates novel business opportunities to create data science services for end-customer.
That is, AI and especially ML can provide different stakeholders with advanced analytics and predictions based on the open data of these open APIs. Business-critical challenges include how to find a paying customer for open data and build a sustainable software ecosystem: The raw data cannot be the product as basically anyone can access it, so the value for a customer must come from, for example, user experience and good analytics in the right place and time. However, rather than focusing only on the challenges in algorithms and technical solutions, business models, or ecosystems, there are also more general software engineering challenges.
\section{Software Engineering Challenges}
In this section, we elaborate software engineering related challenges that we have encountered whilst considering different data science service opportunities based on intelligence built on top of open data and APIs for the maritime cluster.
\subsubsection{Relevant data}
REST APIs are today the dominant design in data APIs. It is customary that full data is provided through a REST API with no ability to customize what data is returned. This often leads to fetching a large amount of unnecessary data. For example, Maritime traffic API contains very large JSON messages but sometimes only one piece of data, such as the ETA of a vessel's portcall, is actually needed. That is, the data providers have just opened all data without much considering how data would be best usable. When multiple APIs are utilized and each provides much unnecessary data, this makes the development and running of the system inconvenient and unreliable. There are already be technologies that could potentially alleviate this. For example, GraphQL APIs can drastically reduce the sizes of transferred JSON files \cite{Brito2020rest}.
\subsubsection{Historical data}
Historical data, i.e. the data produced and collected over the years, is rarely made available. For example, the Maritime API shows only the current traffic although all data over years has been stored. The data providers might not have considered that someone could find value in the historical data on marine traffic. Alternatively, the historical data can be generalized and provided in a more-coarse grained manner in order not to cause too much load for API, such as only for limited time intervals or limited locations as appears to be for some weather data. However, historical data is required for learning in ML systems. Although a data user can start to store data in order to form a training dataset, storing is slow and inconvenient, and brings forward other challenges, such as licensing below.
\subsubsection{Licensing}
When multiple data sources are involved, different rights become an issue. Unfortunately, often open data and API licenses are even more difficult to manage than those of open source software. For example, common software licenses, such as GPL, BSD or MIT, or content licenses, such as Creative Commons (CC), are not necessarily used for data. Rather, open data providers create their own licenses or do not explicitly mark any license. The licenses can be even hard to find in the API specifications, such as for the pilot data above.
Moreover, when data is collected from different sources, it may be difficult to assess how different licenses are compatible and how the new combined data or solution inferred from data can be licensed and commercialized.
\subsubsection{Runtime quality}
A data science service based on multiple sources that need to be accessed near real-time, emphasizes different runtime quality characteristics, such as reliability and availability. With governmental open data, sudden changes to the open data policies are not perhaps as likely as with other organizations. However, there are no guarantees for dependability or service level agreements (SLAs) at least clearly stated in the data sources. The benefit of data science services often lies in near real-time inference and a discontinuity in source data APIs will immediately affect usefulness. In the worst case, the results can be incorrect rather than unavailable.
\subsubsection{API Evolution}
APIs evolve over time, and the changes often break the client developers' code (e.g. \cite{Brito2020}). Generally, the most common API breaking changes are due to refactorings \cite{Dig2006}. With open data APIs, the changes have the potential to affect several API users and end-customers -- likewise when multiple APIs are used any of them can change and break the solution. A specific problem in ML-based solutions is their lack of fault tolerance if something changes: an ML system can continue its operations and produce incorrect, drifted results if some of its data sources have changed.
Identifying the most likely changes in open APIs and preparing for them, e.g., by a means of fault tolerance, can help to mitigate potential API evolution problems. In addition, API evolution can also lead to changes in licensing and technical implementation. This in turn adds yet another layer of complexity in the development process.
\section{A need for governance models}
One unifying factor with all the challenges identified in the previous section is that they are all related, to a certain degree, to the governance of the provided APIs and data. \textit{API governance} is defined as \textit{"a task mainly applied inside an organization, typically aiming at achieving a certain harmonization of APIs in terms of their non-functional properties, best-practices-support, documentation quality or rule compliance in general"} \cite{Haupt2018api}. In \cite{De2017api}, API governance is seen encompassing a wide range of activities \textit{"starting with the API proposal all the way to its adoption, through requirements gathering, build and deploy, and operations during general availability"}.
Data governance, in turn, is associated with decisions regarding the data, i.e., \textit{"data governance refers to who holds the decision rights and is held accountable for an organization’s decision-making about its data assets"}~\cite{Khatri2010designing}.
API governance encompasses practices that need to be designed and executed to overcome the challenges of building intelligent data science services on top of open data APIs. As a summary, following aspects needs to be considered when designing API governance \cite{horkoff2019strategic}, \cite{krintz2013developing}, \cite{lourencco2019framework}:
\subsubsection{Change Control} When API changes are required, the effects of the change should be predictable and implemented in a uniform, consistent way. If changes need to be rolled back, the return to previous functionality should also be consistent, complete, and managed. This requires the development of efficient and automated change-impact analysis techniques that can determine the potential effects of a proposed change.
\subsubsection{Impact of Changes} As APIs are created in the context of business, the impact of changes in API should be carefully evaluated. Stakeholders of an API, such as consumers and business owners, should be informed of changes and the possible impact of those.
\subsubsection{Policy Specification and Analysis} Access control policies, their analysis, and application should be considered to only allow authorized clients to access resources. The accessibility to an API should also consider the business context.
\subsubsection{Consistent Policy Implementation} Policies that control the use of assets through API should be implemented independently of the technologies that are used to implement the assets. Decoupling API from asset implementation allows for API integrity to be kept changes to one do not influence the other.
\subsubsection{Life-cycle Alignment} The governance process should be involved in all the duration of the API life-cycle. A governance process should exist for the development, deployment, monitoring, and deprecation of an API.
\subsubsection{API Integrity} API should be able to interface on a newer version of the platform without conflicts and without effort. When planning new features, existing API should not require extensive refactoring, and backward compatibility should be ensured over a period of time.
\subsubsection{Monitoring and Auditing} API governance must incorporate a unified method of monitoring and auditing API activity.
\begin{table}[b]
\begin{center}
\caption{Relating API governance aspects and recognized challenges.} \label{tab:summary}
\begin{tabularx}{\columnwidth}{X | p{.25cm} | p{.25cm} | p{.25cm} | p{.25cm} | p{.25cm} | p{.25cm} | p{.25cm}|}
API governance aspect $\backslash$ challenge & \rotatebox{90}{Relevant data} &
\rotatebox{90}{Historical data} &
\rotatebox{90}{Licensing} &
\rotatebox{90}{Runtime quality} &
\rotatebox{90}{Evolution} \\
\hline
Change control & &X & & X& X \\
\hline
Impact of changes & X & & & X& X \\
\hline
Policy specification and analysis & & & X & & \\
\hline
Consistent policy implementation & & & X& & \\
\hline
Life-cycle alignment & X & & &X & X \\
\hline
API integrity & & & X &X & X \\
\hline
Monitoring and auditing && X& & X & X \\
\hline \end{tabularx}
\end{center}
\end{table}
Table \ref{tab:summary} summarizes the cross-mapping of API governance aspects and recognized challenges, which helps to understand how API governance could be designed in order to tackle the recognized challenges. To summarize, to guarantee that only \textit{relevant data} is published through API, there is a need to understand the development life-cycle, manage change and understand how changes impact data consumers. Moreover, creating an API that provides \textit{historical data} requires change control as the data structure behind the API can change, but API still needs to provide the same data, whereas \textit{licensing} requires policies to be designed and implemented to manage access to API. Also, API integrity needs to be considered with licensing to manage the compatibility and licenses of different API versions. \textit{Runtime quality} requires quality from both data and the API. API integrity, change control, and acknowledging the impact of changes are major factors to ensure runtime quality. Runtime quality can also be improved by API monitoring and auditing. As the need for runtime quality exists during the whole lifespan of an API, life-cycle alignment is required. Finally, successful \textit{evolution} of API requires managing changes from technical and business perspectives, making sure API integrity exists, aligning API life-cycle, and monitoring and auditing the API.
Moreover, technical API governance and data governance share some responsibilities when it comes to data quality management. As traditional APIs, such as REST and SOAP, are usually built on a separate layer that is not directly connected with data \cite{soni2019api}, most of the data quality dimensions, e.g. completeness, interpretability, accessibility, and representational consistency are shared responsibilities between data and API governance. In addition, as API governance also contains technical aspects, it needs to be considered as part of the IT governance as well.
API governance can be seen as part of the broader practise of API management, that is described as:
\textit{"An activity that enables organizations to design, publish and deploy their APIs for (external) developers to consume. API Management capabilities such as controlling API lifecycles, access and authentication to APIs, monitoring, throttling, and analyzing API usage, as well as providing security and documentation are often implemented through an integrated platform, which is supported by an API gateway."} \cite{mathijssen2020identification}
API governance can also be seen as a combination of data and IT governance, and as a part of broader API management. There is a need for a governance model that describes the structure of how API governance should be designed and executed. This structure should take into account the different aspects of API governance but enable organizational flexibility that exists because of the heterogeneous nature of business domains.
\section{Conclusions}
Open data and APIs are increasingly important and available in the future. Many different business opportunities can arise based on data science services, especially relying on machine learning built on top of open data and APIs. However, besides finding a viable business model and algorithms, there are several software engineering challenges when combining open data from several different open APIs for dependable data science services. We outlined our firsthand experiences about the challenges that we have encountered whilst working in the domain of maritime traffic and its open data and APIs. The challenges are not all domain-specific but pertain to data science services built on top of open data and APIs. As a solution to challenges, we discussed how better governance practices at open data and API providers could alleviate these challenges for those who design and operate data science services. We aim to carry out more research on data sciences services built on open data and APIs to gain experiences of fully exploiting the potential in a fully dependable manner. Especially, we are interested in different governance practices in the entire value network.
\bibliographystyle{IEEEtran}
|
\section{INTRODUCTION}\label{sec_introduction}
Like from an incomplete puzzle, we assemble reality from fragments of information incoming to us from the outside world.
This coarse-grained grasping of reality is mostly sufficient for successful and safe orientation in our
environment. Barring wrong interpretation of reality, the exception to this rule may occur in situations
when the partial information available to us carries no signatures of a global property, the knowledge of which
is crucial for our correct decision.
There is a parallel with quantum world here. Namely, the wave function contains all available information about
a state of a quantum system, but for many tasks we do not need to know it completely. However, unlike in the classical world,
the ``fragments'' of the wave function may not carry traces of the global property which is important for the particular
task, yet our knowledge gained from its parts can still be sufficient.
The states with that remarkable property share similarity with entangled states \cite{Schrodinger_35} as both exhibit a counterintuitive
relationship between the whole and its parts. It is not surprising then, that examples of states with a global property which can be inferred
from parts lacking the property were nonlocal correlations \cite{Wurflinger_12,Vertesi_14} and in particular multipartite entanglement
\cite{Toth_05,Toth_07,Toth_09,Chen_14,Miklin_16,Paraschiv_17}. Out of the many flavours of multipartite entanglement \cite{Dur_99}, the main attention naturally
fell on its strongest form, the genuine multipartite entanglement, which is behind the multipartite tests of quantum nonlocality \cite{Greenberger_89}, complex behaviour of strongly correlated
systems \cite{Amico_08}, certain models of quantum computing \cite{Raussendorf_01} and increased precision of quantum measurements \cite{Giovannetti_04}.
So far, only examples of qubit states carrying genuine multipartite entanglement which can be verified solely from separable two-qubit reduced states (marginals) were
found \cite{Chen_14,Miklin_16} and demonstrated experimentally \cite{Micuda_19}. In all these cases, the set of marginals used to certify the entanglement comprised
{\it all} two-qubit marginals. Interestingly, genuine multipartite entanglement can be detected even from a smaller set of separable marginals. Indeed, multiqubit
states can be found possessing all two-qubit marginals separable, whose genuine multipartite entanglement can be inferred only from the so called
minimal set of two-qubit marginals \cite{Paraschiv_17}. The minimal set covers any part of the entire system and it contains only marginals between
nearest-neighbours, which guarantees that knowledge of the set suffices to confirm global entanglement. In geometric terms, if we
represent parts of the global state as vertices of a graph \cite{Steinbach_99} and the bipartite marginals as its edges, then the minimal set
corresponds to a tree-like graph. States with genuine multipartite entanglement which can be confirmed using only the elements of the minimal set
where found for all configurations of up to six qubits using the iterative numerical search algorithm \cite{Paraschiv_17} combining the machinery of entanglement
witnesses \cite{Horodecki_96,Terhal_00} with the tools of semi-definite programming \cite{Vandenbergehe_96}. The best example obtained
was three-qubit \cite{Miklin_16} with the lowest witness mean being roughly
three times smaller than the witness mean for the scenario
in which all two-qubit marginals are known \cite{Miklin_16,Micuda_19}. Moreover, the difference is even more pronounced compared to
other theoretically predicted witness means \cite{Acin_01} of already successfully implemented multipartite entanglement witnesses experiments
\cite{Bourennane_04}. This indicates the complexity of the possible experimental demonstration of the studied effect.
In this paper we take a different approach to the problem by seeking states with the investigated property
in the realm of Gaussian states \cite{Braunstein_05a,Weedbrook_12}. More precisely, we look for Gaussian
states with all two-mode marginals separable and whose genuine multipartite entanglement can be proved only from the minimal set of the
marginals. For this purpose, we use the methods of Gaussian multipartite entanglement witnesses \cite{Hyllus_06} to assemble a
Gaussian analog of the search algorithm \cite{Paraschiv_17}. By running the
algorithm we then find examples of the studied
states for all configurations of up to four modes (see Fig.~\ref{fig1}). Our
simplest examples involve only three modes similarly to
the simplest known qubit example which consists of three qubits \cite{Miklin_16}. The three-mode example gives, for the Gaussian
analog of the genuine multipartite entanglement witness mean, a value which is roughly of the same size as the theoretically predicted
values \cite{Hyllus_06} for already realized similar Gaussian multimode entanglement experiments \cite{DiGuglielmo_11}. Further, the required squeezing is less than one third of a vacuum unit.
Given the promising role of Gaussian states in the current problem, we also propose a linear-optical circuit for preparation
of the three-mode state, which is based on interference of three correlatively displaced squeezed beams on three beam splitters.
Our results reveal that a minimal set of overlapping separable marginals may suffice to reveal genuine multipartite entanglement
also in Gaussian scenario. Besides, they indicate that Gaussian continuous variables represent a promising alternative
platform for experimental demonstration of the studied property of genuine multipartite entanglement.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.40\textwidth]{fig1.pdf}
\caption{Graphical representation of minimal sets of marginal CMs used to detect genuine multipartite entanglement for three and four modes. See text for details.}
\label{fig1}
\end{center}
\end{figure}
\section{GAUSSIAN STATES}\label{sec_Gaussian_states}
The scene of our considerations is the set of Gaussian states
of systems with infinite-dimensional Hilbert space, which we
shall call modes in what follows. A collection of $N$ modes $A_{j}$, $j=1,2,\ldots,N$,
can be characterized by a vector $\xi=(x_{A_1},p_{A_1},\ldots,x_{A_N},p_{A_N})^{T}$ of
position and momentum quadratures $x_{A_{j}}$ and $p_{A_{j}}$, respectively,
which obey the canonical commutation rules $[\xi_j,\xi_k]=i(\Omega_{N})_{jk}$ with
$\Omega_{N}=\oplus_{j=1}^{N}i\sigma_{y}$, where
$\sigma_{y}$ is the Pauli-$y$ matrix.
Gaussian states are defined as states with a Gaussian-shaped
phase-space Wigner function. An $N$-mode Gaussian state $\rho$ is thus fully described by a $2N\times 1$ vector
$\langle\xi\rangle=\mbox{Tr}[\xi\rho]$ of first moments and by a $2N\times 2N$ covariance matrix
(CM) $\gamma$ with entries $(\gamma)_{jk}=\langle\xi_{j}\xi_k+\xi_k\xi_j\rangle-2\langle\xi_{j}\rangle\langle\xi_k\rangle$.
The first moments can be nullified by local displacements and thus they are irrelevant as far as
the correlation properties investigated here are concerned. For this reason we set them to zero from now on.
Any CM $\gamma$ reflects the uncertainty principle by satisfying the inequality
\begin{eqnarray}\label{Heisenbergg}
\gamma+i\Omega_{N}\geq0,
\end{eqnarray}
which is not only a necessary but also a sufficient condition for a real symmetric $2N\times 2N$
matrix $\gamma$ to be a CM of a physical quantum state \cite{Simon_00}. Besides, a CM also carries complete
information about the separability properties of the corresponding Gaussian state.
Recall first, that a quantum state $\rho_{jk}$ of two subsystems $j$ and $k$ is separable
if it can be expressed as a convex mixture of product states $\rho_{j|k}^{\rm sep}\equiv\sum_{i}p_{i}\rho_{j}^{(i)}\otimes\rho_{k}^{(i)}$,
where $\rho_{j}^{(i)}$ and $\rho_{k}^{(i)}$ are local states of subsystems $j$ and $k$, respectively.
If the state cannot be written in this form it is called entangled. Separability of a two-mode Gaussian state $\rho_{jk}$
can be ascertained by the positive partial transposition (PPT) criterion \cite{Peres_96,Horodecki_96,Simon_00}. On the CM level
the partial transposition operation $T_{j}$ with respect to mode $j$ transforms the CM $\gamma_{jk}$ of
the state as $\gamma_{jk}^{(T_{j})}=(\sigma_{z}\oplus\openone)\gamma_{jk}(\sigma_{z}\oplus\openone)$,
where $\sigma_{z}$ is the Pauli-$z$ matrix and $\openone$ is the $2\times 2$ identity matrix. The PPT criterion then says \cite{Simon_00},
that the state $\rho_{jk}$ is separable if and only if (iff) the matrix $\gamma_{jk}^{(T_{j})}$ is a physical CM, i.e., iff
\begin{equation}\label{HeisenbergTj}
\gamma_{jk}^{(T_{j})}+i\Omega_{2}\geq0.
\end{equation}
The PPT criterion is a sufficient condition for separability only for two-mode \cite{Simon_00} and
$1\times M$-mode \cite{Werner_01} Gaussian states. For systems where each party holds more than one
mode, one has to use a more powerful criterion \cite{Werner_01} according to which an $N$-mode Gaussian state
with CM $\gamma$ consisting of an $l$-mode subsystem $A \equiv A_{1}A_{2}\ldots A_{l}$
and an $(N-l)$-mode subsystem $B \equiv A_{l+1}A_{l+2}\ldots A_{N}$, is separable iff there are
CMs $\gamma_{A}$ and $\gamma_{B}$ of the subsystems such, that
\begin{equation}\label{CMseparability}
\gamma-\gamma_{A}\oplus\gamma_{B}\geq0.
\end{equation}
The separability criterion (\ref{CMseparability}) is advantageous because it can be formulated as the following
semi-definite programme (SDP) \cite{Hyllus_06}:
\begin{equation}\label{sepprimal}
\begin{aligned}
& \underset{\gamma_{A},\gamma_{B},x_{e}}{\text{minimize}}
& & (-x_{e}) \\
& \text{subject to}
& &\gamma-\gamma_{A}\oplus\gamma_{B}\geq0,\\
& & &\gamma_{A}\oplus\gamma_{B}+(1+x_{e})i\Omega_{N}\geq0.
\end{aligned}
\end{equation}
If there is an optimal solution $x_{e}\geq0$, then CM $\gamma$ describes a separable state because there exist CMs
$\gamma_{A}$ and $\gamma_{B}$ such that the separability criterion (\ref{CMseparability}) is satisfied. If, on the other
hand, $x_{e}<0$, then the state with CM $\gamma$ is entangled.
\section{GAUSSIAN ENTANGLEMENT WITNESSES}\label{sec_Gaussian_entanglement_witnesses}
In practice, one needs most often to certify the presence of entanglement in a given state
rather than to show that it is separable. However, many entanglement criteria, including the PPT criterion
or criterion (\ref{CMseparability}) require knowledge of the entire quantum state and thus they are not
economical as far as the number of measurements is concerned. This also implies that the criteria
cannot be used in cases when we have access only to a part of the investigated state.
Nevertheless, it is still possible to detect entanglement provided that we have some a priori information
about the state. Namely, one can prove the presence of entanglement by measuring the so-called entanglement
witnesses \cite{Horodecki_96,Terhal_00}, which requires fewer measurements compared to the measurement of
the whole quantum state \cite{Guhne_02}.
\subsection{Bipartite entanglement witnesses}\label{subsec_Bipartite_witnesses}
For a bipartite state an entanglement witness is a Hermitian operator with a non-negative average for all separable states and a negative average on at least
one entangled state. However, the task of finding an entanglement witnesses for density matrices of continuous-variable modes is often hardly tractable
owing to their infinite dimension. A much more simple option, which is particularly suitable for Gaussian states,
is to seek entanglement witnesses for CMs \cite{Hyllus_06}. Such a witness is, for an $N$-mode state, represented by
a $2N\times2N$ real, symmetric and positive-semidefinite matrix $Z$, which satisfies the following conditions:
\begin{eqnarray}\label{Zconditions}
& \mbox{(i)} & \mbox{Tr}[\gamma Z]\geq 1,\quad \mbox{for all separable}\quad \gamma,\nonumber\\
& \mbox{(ii)} & \mbox{Tr}[\gamma Z]< 1,\quad \mbox{for some entangled}\quad \gamma.
\end{eqnarray}
Entanglement detection by means of matrix $Z$ possesses several advantages. First, the expression $\mbox{Tr}[\gamma Z]$ is a linear function of second moments and therefore it
can be measured by local homodyne detections followed by a suitable processing of the output photocurrents. More importantly, the expression also typically contains
only some elements of CM $\gamma$ and thus it requires fewer measurements than one needs to measure the entire CM. Another advantage of using the matrix
$Z$ is that for a given CM $\gamma$ it can be found numerically by solving the dual to program (\ref{sepprimal}) \cite{Hyllus_06}:
\begin{equation}\label{sepdual}
\begin{aligned}
& \underset{X_{1},X_{2}}{\text{minimize}}
& & \mbox{Tr}[\gamma X_{1}^{\rm re}]-1, \\
& \text{subject to}
& & X_{1}^{\rm bd, re}=X_{2}^{\rm bd, re},\quad X_{1}\geq0,\quad X_{2}\geq 0,\\
& & & \mbox{Tr}[i\Omega_{N} X_{2}]=-1.
\end{aligned}
\end{equation}
Here $X_{j}$, $j=1,2$, are $2N\times 2N$ Hermitian matrices, the symbol $X_{j}^{\rm re}$
stands for the real part of the matrix $X_{j}$, and $X_{j}^{\rm bd}=X_{j A}\oplus X_{j B}$, where $X_{j A}$ and $X_{j B}$
are diagonal blocks of the matrix $X_{j}$ corresponding to subsystems $A$ and $B$, respectively.
It can be shown \cite{Hyllus_06}, that for every feasible solution $X_{1}\oplus X_{2}$, the
matrix $X_{1}^{\rm re}$ satisfies
\begin{equation}\label{Wcondition1}
\mbox{Tr}[\gamma X_{1}^{\rm re}]\geq 1
\end{equation}
for every CM $\gamma$ of a separable state. Further, if $\gamma$ is a CM of an entangled state, then
\begin{equation}\label{Wcondition2}
\mbox{Tr}[\gamma X_{1}^{\rm re}]<1.
\end{equation}
This implies, that the real matrix $X_{1}^{\rm re}$ is an entanglement witness which is, in addition,
optimal in the sense that it yields the minimal value of $\mbox{Tr}[\gamma Z]$ out of all possible witnesses $Z$.
Needless to say, by adding more constraints into the SDP (\ref{sepdual}), one can seek witnesses with a special structure.
Below we will see, for instance, that one can seek witnesses which are `blind' to certain parts of CM $\gamma$.
\subsection{Genuine multipartite entanglement witnesses}\label{subsec_GME_witnesses}
Bipartite entanglement is just one particular kind of entanglement. In multipartite systems consisting of $N>2$ subsystems one
can investigate also multipartite entanglement, which occurs among more than two groups of subsystems. In general, it is possible
to split all subsystems into $k$ disjoint subsets and analyze entanglement with respect to the $k$-partite split.
We say that a state is $k$-separable if it is fully separable with respect to the $k$-partite split, i.e., if it
can be expressed as a convex mixture of product states with respect to the split. Otherwise, it is called entangled with
respect to the split. This allows us to classify multipartite states according to their separability properties with respect
to all possible $k$-partite splits for all possible $k$ \cite{Dur_99,Dur_00}. At the top of the hierarchy, there are
fully inseparable states which are not separable with respect to any $k$-partite split. Nevertheless,
even fully inseparable states in general do not carry the strongest form of multipartite entanglement. Namely,
some of them can be created by convex mixing of some $k$-separable states \cite{Guhne_09} and thus their preparation does not
require a collective operation on all subsystems as we would expect from truly multipartite entangled states.
For this reason, the concept of genuine $N$-partite entangled states was introduced as states that cannot be expressed as a
convex mixture of some $k$-separable states for any $k\geq2$ \cite{Acin_01}. Note, that any $k$-separable state with
$k>2$ is also $2$-separable. Consequently, a set of states that can be expressed as a convex mixture of some $k$-separable
states is a subset of the set of states that can be expressed as a convex mixture of some $2$-separable states, which are
fittingly called biseparable states. This reveals that for the presence of genuine multipartite entanglement in
a given quantum state it is sufficient to show that it is not biseparable.
The concept of biseparability carries over straightforwardly to CMs of $N$-mode Gaussian states.
For this purpose, let us collect modes $A_{j}$, $j=1,2,\ldots,N$, into the set $\mathcal{N}=\{A_{1},A_{2},\ldots,A_{N}\}$ and let $\mathcal{I}=\{1,2,\ldots,N\}$ be its index set.
Next, consider a nonempty proper index subset $\mathcal{J}_{k}=\{i_{1},i_{2},\ldots,i_{l}\}$ of $0<l<N$ elements of the index set $\mathcal{I}$ and let
$\bar{\mathcal{J}}_{k}=\mathcal{I}\backslash\mathcal{J}_{k}$ denotes its complement containing the remaining $N-l$ elements of $\mathcal{I}$. This allows
us to split the set $\mathcal{N}$ into $K\equiv 2^{N-1}-1$ different inequivalent $2$-partitions, called as bipartitions in what follows,
$\pi(k)\equiv\mathcal{M}_{\mathcal{J}_{k}}|\bar{\mathcal{M}}_{\mathcal{J}_{k}}$, $k=1,2,\ldots,K$, where
$\mathcal{M}_{\mathcal{J}_{k}}=\{A_{i_{1}},A_{i_{2}},\ldots,A_{i_{l}}\}$ and $\bar{\mathcal{M}}_{\mathcal{J}_{k}}=\mathcal{M}_{\bar{\mathcal{J}}_{k}}=\mathcal{N}\backslash\mathcal{M}_{\mathcal{J}_{k}}$.
Moving to the criterion of biseparability one can show \cite{Hyllus_06}, that an $N$-mode Gaussian state with CM $\gamma$ is
biseparable iff there exist bipartitions $\pi(k)$ and CMs $\gamma_{\pi(k)}$ which are block diagonal with respect to the bipartition $\pi(k)$, and probabilities $\lambda_{k}$ such that
\begin{equation}\label{biseparable}
\gamma-\sum_{k=1}^{K}\lambda_{k}\gamma_{\pi(k)}\geq0.
\end{equation}
Similarly as bipartite separability can be decided by solving the SDP (\ref{sepprimal}),
biseparability embodied by condition (\ref{biseparable}) can also be decided by solving an SDP \cite{Hyllus_06}.
Analogously, just like an optimal witness of bipartite entanglement can be obtained by solving the dual problem
(\ref{sepdual}) of the former SDP, the optimal witness of genuine $N$-partite entanglement can be found by solving the
dual problem of the corresponding SDP \cite{Hyllus_06}. Recall first, that the witness of genuine $N$-partite
entanglement is represented by a $2N\times 2N$ real, symmetric, and positive-semidefinite matrix $Z$ satisfying conditions \cite{Hyllus_06}
\begin{eqnarray}\label{ZGME}
& \mbox{(i)} & \mbox{Tr}[\gamma Z]\geq 1,\quad \mbox{for all biseparable}\quad \gamma,\nonumber\\
& \mbox{(ii)} & \mbox{Tr}[\gamma Z]< 1,\quad \mbox{for some entangled}\quad \gamma.
\end{eqnarray}
For a given CM $\gamma$ the witness can be found by solving the following dual problem \cite{Hyllus_06}:
\begin{widetext}
\begin{equation}\label{bisepdual}
\begin{aligned}
& \underset{X}{\text{minimize}}
& & \mbox{Tr}[\gamma X_{1}^{\mathrm{re}}]-1\\
& \text{subject to}
& &X_{1}^{{\rm re, bd}, \pi(k)}=X_{k+1}^{{\rm re, bd}, \pi(k)}\quad \mbox{for all}\quad k=1,\ldots,K,\\
& & &\mbox{Tr}[i\Omega_{N} X_{k+1}]+X_{K+2}-X_{K+3}+X_{K+3+k}=0,\quad \mbox{for all}\quad k=1,\ldots,K,\\
& & &X_{K+2}-X_{K+3}=1.
\end{aligned}
\end{equation}
\end{widetext}
Here, the minimization is preformed over Hermitian positive-semidefinite $[2N(K+1)+2+K]$-dimensional block-diagonal matrix
\begin{equation}\label{X}
X=\bigoplus_{j=1}^{2K+3}X_{j},
\end{equation}
with $X_{j}$, $j=1,2,\ldots,K+1$ being $2N\times 2N$ Hermitian matrices and $X_{j}$, $j=K+2,K+3,\ldots,2K+3$
being $1\times 1$ Hermitian matrices, i.e., real numbers. Further, the $k$-th equation $X_{1}^{{\rm re, bd}, \pi(k)}=X_{k+1}^{{\rm re, bd}, \pi(k)}$
imposes a constraint on diagonal blocks of the matrices $X_{1}$ and $X_{k+1}$ written in the block form with respect to the bipartition $\pi(k)$.
More precisely, let us express the matrix $X_{j}$ in the block-form with respect to the $N$-partite split $A_{1}|A_{2}|\ldots|A_{N}$,
\begin{eqnarray}\label{Xjblock}
X_{j}=\left(\begin{array}{cccc}
(X_{j})_{11} & (X_{j})_{12} & \ldots & (X_{j})_{1N} \\
(X_{j})_{12}^{\dag} & (X_{j})_{22} & \ldots & (X_{j})_{2N} \\
\vdots & \vdots & \ddots & \vdots\\
(X_{j})_{1N}^{\dag} & (X_{j})_{2N}^{\dag} & \ldots & (X_{j})_{NN}\\
\end{array}\right),
\end{eqnarray}
where $(X_{j})_{mn}$ is a $2\times 2$ block. Then, the matrix $X_{j}^{\mathrm{bd}, \pi(k)}$ is of the same block form with the $2\times 2$
blocks given by
\begin{equation}
\label{Xbd}
\left(X_{j}^{\mathrm{bd}, \pi(k)}\right)_{mn}=\left\{\begin{array}{lllll} (X_{j})_{mn}, & \!\textrm{if} & \hspace{-1.4cm} m,n\in\mathcal{J}_{k} & \textrm{or} & \bar{\mathcal{J}}_{k};\\
\mathbb{O}, & \textrm{otherwise}, & & &\\
\end{array}\right.
\end{equation}
where $\mathbb{O}$ is the $2\times 2$ zero matrix. For relevant cases $N=3$ and $N=4$ discussed in this paper an explicit form of the matrices
$X_{j}^{\mathrm{bd}, \pi(k)}$ can be found in Appendix~\ref{sec_app_I}.
According to the results of Ref.~\cite{Hyllus_06}, for every feasible solution $X$ of the dual program (\ref{bisepdual})
the matrix $X_{1}^{\mathrm{re}}$ is an optimal genuine multipartite entanglement witness.
\subsection{Blind genuine multipartite entanglement witnesses}\label{subsec_blind_GME_witnesses}
The witness obtained by solving the programme (\ref{bisepdual}) acts on the entire CM $\gamma$ and therefore enables us to certify genuine multipartite entanglement
provided that all elements of the CM are known. Viewed from a different perspective, it is equivalent to witnessing the entanglement from all two-mode marginal
CMs, because they completely determine the global CM. In this respect, the domain of Gaussian states differs from the qubit case, where the knowledge of all two-qubit marginals
is not generally equivalent to the knowledge of the whole density matrix. To make the task on inference of genuine multipartite entanglement from marginals
in Gaussian scenario meaningful, we thus have to work only with a proper subset of the set of all two-mode marginal CMs. In what follows, we utilize
the so-called minimal sets of bipartite marginals, which were introduced recently in Ref.~\cite{Paraschiv_17} to solve the task for qubits. Obviously, a necessary condition for the
set to allow detection of global entanglement is that it contains all modes and that one cannot divide it into a subset and its complement without having a common mode.
Among all such sets a particularly important role play further irreducible sets containing a minimum possible number of two-mode marginals.
A more convenient pictorial representation of such minimal sets was put forward in Ref.~\cite{Paraschiv_17} in the form on an unlabeled tree \cite{Steinbach_99}, which is a special form of an undirected
connected graph containing no cycles. Recall, that a graph is a pair $G=(V,E)$ of a set $V=\{1,2,\ldots,N\}$ of vertices and a set $E\subseteq K\equiv\{\{u,v\}|(u,v)\in V^2 \wedge u\ne v\}$
of edges \cite{West_01}. In our case a vertex $j$ of the graph represents mode $A_{j}$, whereas the edge connecting adjacent vertices $j$ and $k$ represents marginal CM $\gamma_{A_{j}A_{k}}$.
By definition, the minimal set contains two-mode marginal CMs corresponding to the edges in the respective tree denoted as $T=(V,E')$. A closed formula for the number
of non-isomorphic trees with $N$ vertices is not known, yet for small $N$ it can be found in Ref.~\cite{A000055}. In particular, all trees for the three-mode case ($N=3$) and
the four-mode case ($N=4$) are depicted in Fig.~\ref{fig1}, where we performed the following identification $A\equiv A_{1}, B\equiv A_{2}, C\equiv A_{3}$, and $D\equiv A_{4}$.
Ignorance of some sectors of CM $\gamma$ requires to impose some additional constraints onto the structure of the witness $X_{1}^{\mathrm{re}}$ being the solution of SDP (\ref{bisepdual}).
Specifically, as the respective tree is connected, the minimal set contains all single-mode CMs as well as $2\times2$ blocks of correlations between the
modes corresponding to the endpoints of the edges of the tree $T$. The part of the CM $\gamma$ which we do not know is therefore given by all $2\times2$ off-diagonal blocks of
correlations between pairs of modes carried by the marginal two-mode CMs contained in the complement of the minimal set. The elements of the complement correspond to the
edges in the complement graph $\bar{T}=(V,K\backslash E')$, i.e., to the edges which have to be added to the original tree $T$ to form the complete graph. Since for a given $N$ the complete graph
contains $\binom{N}{2}$ edges and the tree $T$ contains exactly $N-1$ edges \cite{West_01}, the number of unknown blocks of correlations is equal to $L\equiv(N-1)(N-2)/2$.
Further, as $\mbox{Tr}[\gamma X_{1}^{\rm re}]=\sum_{j,k}(\gamma)_{jk}(X_{1}^{\rm re})_{jk}$, in order for the witness
$X_{1}^{\mathrm{re}}$ not to act on the unknown blocks of CM $\gamma$, its blocks in places of the unknown blocks have to vanish. More precisely, if we
express the witness $X_{1}^{\mathrm{re}}$ in the block form with respect to $N$-partite split $A_{1}|A_{2}|\ldots|A_{N}$ similar to Eq.~(\ref{Xjblock}),
its $2\times 2$ off-diagonal blocks have to satisfy the following set of $L$ equations:
\begin{equation}\label{blind}
(X_{1}^{\rm re})_{mn}=\mathbb{O},\quad \mbox{if} \quad \{m,n\}\in K\backslash E',
\end{equation}
which have to be added to the SDP (\ref{bisepdual}) as additional constraints. For $N=3$ and the tree in Fig.~\ref{fig1}~a), the constraint reads explicitly as
\begin{eqnarray}\label{blind3}
(X_{1}^{\mathrm{re}})_{13}=\mathbb{O}.
\end{eqnarray}
Likewise, in the case $N=4$ and for the linear tree in Fig.~\ref{fig1}~b), the constraints are
\begin{eqnarray}\label{blind41}
(X_{1}^{\mathrm{re}})_{13}=(X_{1}^{\mathrm{re}})_{14}=(X_{1}^{\mathrm{re}})_{24}=\mathbb{O},
\end{eqnarray}
whereas for the `t'-shaped tree in Fig.~\ref{fig1}~c) one gets the constraints of the following form:
\begin{eqnarray}\label{blind42}
(X_{1}^{\mathrm{re}})_{13}=(X_{1}^{\mathrm{re}})_{14}=(X_{1}^{\mathrm{re}})_{34}=\mathbb{O}.
\end{eqnarray}
\section{SEARCH ALGORITHM}\label{sec_search_algorithm}
The goal of the present paper is to find an example of a Gaussian state with all two-mode marginals separable and whose genuine multipartite entanglement can be verified
solely from the minimal set of two-mode marginals. Recently, multiqubit examples of such states have been found \cite{Paraschiv_17} using a two-step
algorithm proposed in Ref.~\cite{Miklin_16}. Here, we employ the following Gaussian analog of the algorithm:
{\it Step 0:} Generate a random pure Gaussian state with CM $\gamma_{0}$ which has, for simplicity, no $x-p$ correlations.
{\it Step 1:} For CM $\gamma_{0}$, find a witness $X_{1}^{\mathrm{re}}$ by solving numerically the SDP (\ref{bisepdual}) supplemented with the constraints (\ref{blind}), which we
shall call as the SDP~1. Note, that the SDP~1 can be solved by modifying the freely available routine \cite{Lofberg} in Matlab by adding the constraints (\ref{blind}) into it.
{\it Step 2:} Find a CM $\gamma$ that gives the least value of $\mbox{Tr}[\gamma X_{1}^{\mathrm{re}}]$ for the witness $X_{1}^{\mathrm{re}}$ from the
first step under the constraint that the CM possesses all two-mode marginals separable. Again, the search can be accomplished by solving the following SDP:
\begin{equation}\label{step2}
\begin{aligned}
& \underset{\gamma}{\text{minimize}}
& & \mbox{Tr}[\gamma X_{1}^{\mathrm{re}}] \\
& \text{subject to}
& &\gamma+i\Omega_{N}\geq0,\\
& & & \gamma_{jk}^{(T_{j})}+i\Omega_{2}\geq0, \quad \mbox{for all}\quad j\ne k=1,\ldots,N,\\
& & & (\gamma)_{2j-1,2k}=(\gamma)_{2j,2k-1}=0, \quad j,k=1,\ldots,N,
\end{aligned}
\end{equation}
which is called as SDP~2 from now. Here, we carry out the minimization over all real symmetric $2N\times 2N$ matrices $\gamma$.
The first constraint guarantees that the matrix $\gamma$ is a CM of a physical quantum state, whereas the second constraint assures that
all its two-mode marginal CMs $\gamma_{jk}$ are separable. Finally, due to the third constraint we perform minimization
only over matrices $\gamma$ which do not contain any $x-p$ correlations.
By putting the obtained solution from Step~2 as an input to Step~1 we can iteratively seek the CM with
the desired properties. In the next section we give explicit examples of such CMs for all three-mode and four-mode minimal sets.
\section{Results}\label{sec_results}
\subsection{Three modes}
\label{sec_results_3modes}
First, we did a numerical search of a three-mode example of the investigated
effect. Running SDP~1 and SDP~2 successively for 10 iterations for $N=3$,
we found several examples of states with all two-mode marginals separable and
whose genuine three-mode entanglement can be verified solely from the nearest
neighbour marginal CMs $\gamma_{AB}$ and $\gamma_{BC}$ (see Fig.~\ref{fig1}~a)).
The CMs typically exhibited large diagonal entries and required high squeezing
for preparation. To get experimentally easier accessible CM, we therefore added another two
constraints to the SDP~2 (\ref{step2}). First, we limited the diagonal elements of
the CM to lie within the range $[1,10]$ and second, we also constrained the smallest
eigenvalue of the sought CM $\gamma$ to be above $0.2$. The best CM we got
in this way giving the least value of $\mbox{Tr}[\gamma Z]$ reads after the
rounding to two decimal places as
\begin{eqnarray*}\label{gamma3}
\gamma_{3} =\left(\begin{array}{cccccc}
1.34 & 0 & -0.35 & 0 & -0.82 & 0\\
0 & 10.00 & 0 & 8.45 & 0 & 1.87 \\
-0.35 & 0 & 7.8 0 & 0 & -8.05 & 0\\
0 & 8.45 & 0 & 7.92 & 0 & 2.09 \\
-0.82 & 0 & -8.05 & 0 & 10.00 & 0\\
0 & 1.87 & 0 & 2.09 & 0 & 1.62
\end{array}\right),\nonumber\\
\end{eqnarray*}
and by running the SDP~1 for the rounded CM $\gamma_{3}$ we got
$\mbox{Tr}[\gamma_{3} Z_{3}]-1\doteq-0.143$. The corresponding witness, which is
blind to the correlations between a pair of modes $(A,C)$, is after rounding to three
decimal places given by
\begin{eqnarray*}\label{Z}
Z_{3} \!\! = \!\! 10^{-2} \!\!
\left( \!\! \begin{array}{cccccc}
6.8 & 0 & -0.4 & 0 & 0 & 0\\
0 & 34.3 & 0 & -39.5 & 0 & 0\\
-0.4 & 0 & 25.1 & 0 & 20.9 & 0\\
0 & -39.5 & 0 & 46.1 & 0 & -2.0\\
0 & 0 & 20.9 & 0 & 17.5 & 0\\
0 & 0 & 0 & -2.0 & 0 & 6.6
\end{array} \!\! \right) \!\!.\nonumber\\
\end{eqnarray*}
The separability of all marginals is evidenced by Tab.~\ref{table1}
\begin{table}[ht]
\caption{Minimal eigenvalue $\varepsilon_{jk}\equiv\mathrm{min}\{\mathrm{eig}[\gamma_{3,jk}^{(T_{j})}+i\Omega_{2}]\}$.} \centering
\begin{tabular}{| c | c | c | c |}
\hline $jk$ & AB & AC & BC \\
\hline $\varepsilon_{jk}$ & 0.002 & 0.849 & 0.004 \\
\hline
\end{tabular}
\label{table1}
\end{table}
Inspection of Tab.~\ref{table1} reveals that all eigenvalues are strictly
positive and therefore all three two-mode marginal states are separable by PPT
criterion as required.
The present result can be compared with the results for qubits derived in
Ref.~\cite{Paraschiv_17}.
Note first that the value of $\mbox{Tr}[\gamma_{3} Z_{3}]-1\doteq-0.143$ found
here, for the simplest three-mode state, is slightly larger than the theoretical value of $-0.103$ for the same
quantity of the comparable effect of Gaussian bound entanglement \cite{Hyllus_06,Werner_01},
which was already observed experimentally \cite{DiGuglielmo_11}. On the other hand,
the best qubit mean of $\mbox{Tr}[\rho W]\doteq-6.58\cdot 10^{-3}$ obtained for the
three-qubit state \cite{Miklin_16} is approximately three times smaller
than the best theoretical witness mean of $-1.98\cdot 10^{-2}$ for the
case when all two-qubit marginals are known \cite{Miklin_16}, which was recently
demonstrated in \cite{Micuda_19}. Recall further, that in the qubit scenario
the noise tolerance is $5\%$ \cite{Miklin_16}. For comparison,
the produced state with CM $\gamma_{3}$ also tolerates the addition of a small amount of thermal noise,
i.e., the CM $\gamma_{p}=\gamma_{3}+p\openone$ exhibits the effect for up to $p\doteq0.1$,
yet the value is the same as one would get for the successfully demonstrated Gaussian bound
entanglement \cite{Hyllus_06}.
All these facts indicate the domain of Gaussian states to be a more promising platform for
the near-future experimental demonstration of the analyzed effect. Therefore, in the next section we present a
linear-optical scheme for preparation of a close approximation of the state with CM $\gamma_{3}$.
However, before doing so, we first construct also four-mode states carrying the investigated property.
\subsection{Four modes}
Next, we extended the search of example CMs to four modes. In this case there are two different minimal sets
of marginals corresponding to the linear tree and the `t'-shaped tree displayed in Figs.~\ref{fig1}~b) and c),
respectively. Through the same procedure as for the three-mode case, we found CMs with the desired
properties for both the minimal sets, which are given explicitly below.
\subsubsection{Linear tree}
First, we considered the minimal set of marginals given by the CMs
$\gamma_{AB}$, $\gamma_{BC}$ and $\gamma_{CD}$, corresponding to the edges in the linear tree in
Fig.~\ref{fig1}~b). This was reflected by inclusion of the constraints (\ref{blind41}) into our
search algorithm. By running the algorithm for 10 iterations, we produced several four-mode
CMs with the desired properties. The best such CM is
\begin{widetext}
\begin{eqnarray}\label{gammaLinear}
\gamma_{4}^{(1)}=\left(\begin{array}{cccccccc}
2.83 & 0 & -0.02 & 0 & -1.38 & 0 & 2.83 & 0\\
0 & 7.18 & 0 & 8.06 & 0 & 7.09 & 0 & -4.12\\
-0.02 & 0 & 3.91 & 0 & -2.46 & 0 & 4.73 & 0\\
0 & 8.06 & 0 & 9.79 & 0 & 8.47 & 0 & -4.81\\
-1.38 & 0 & -2.46 & 0 & 2.58 & 0 & -4.68 & 0\\
0 & 7.09 & 0 & 8.47 & 0 & 10.00 & 0 & -3.08\\
2.83 & 0 & 4.73 & 0 & -4.68 & 0 & 10.00 & 0\\
0 & -4.12 & 0 & -4.81 & 0 & -3.08 & 0 & 3.22
\end{array}\right).
\end{eqnarray}
\end{widetext}
The optimal witness $Z_{4}^{(1)}$, which is blind to correlations of modes $(A,C), (A,D)$ and $(B,D)$, gives the value of
$\mbox{Tr}[\gamma^{(1)}_4 Z_4^{(1)}]-1\doteq-0.069$ and it can be found in Appendix~\ref{sec_app_II}. The separability of all marginals can be confirmed
again by the PPT criterion (\ref{HeisenbergTj}) which is captured in Tab.~\ref{tableLinear}.
\begin{table}[ht!]
\caption{Minimal eigenvalue $\varepsilon_{jk}^{(1)}\equiv\mbox{min}\{\mbox{eig}[(\gamma_{4,jk}^{(1)})^{(T_j)} +
i \Omega_2]$.}
\centering
\begin{tabular}{| c | c | c | c | c | c | c |}
\hline $jk$ & AB & AC & AD & BC & BD & CD \\
\hline $\varepsilon_{jk}^{(1)}$ & 0.005 & 0.347 & 0.213 & 0.004 & 0.087 & 0.224 \\
\hline
\end{tabular}
\label{tableLinear}
\end{table}
As all entries in the second row of the Tab.~\ref{tableLinear} are strictly positive, all two-mode marginal CMs of CM $\gamma_{4}^{(1)}$
are separable as required. Note further, that the effect is roughly half that of the three-mode case,
which makes its experimental demonstration a bigger challenge.
\subsubsection{`t'-shaped tree}
Finally, we looked for states whose genuine four-mode entanglement can be witnessed
from its nearest-neighbour marginals as per the graph in Fig.~\ref{fig1} c).
This corresponds to the `t'-shaped tree for which the minimal set comprise
marginal CMs $\gamma_{AB}, \gamma_{BC}$ and $\gamma_{BD}$, and the
witness then fulfils the constraints (\ref{blind42}).
The best example CM found reads as
\begin{widetext}
\begin{eqnarray}\label{gammaTshape}
\gamma_{4}^{(2)}=\left(\begin{array}{cccccccc}
5.23 & 0 & 0.45 & 0 & -0.02 & 0 & -2.43 & 0\\
0 & 1.16 & 0 & 3.00 & 0 & 1.15 & 0 & 0.51 \\
0.45 & 0 & 3.35 & 0 & 0.91 & 0 & -5.20 & 0\\
0 & 3.00 & 0 & 10.00 & 0 & 3.52 & 0 & 2.06 \\
-0.02 & 0 & 0.91 & 0 & 4.09 & 0 & -2.97 & 0\\
0 & 1.15 & 0 & 3.52 & 0 & 1.62 & 0 & 0.62 \\
-2.43 & 0 & -5.20 & 0 & -2.97 & 0 & 10.00 & 0\\
0 & 0.51 & 0 & 2.06 & 0 & 0.62 & 0 & 1.49
\end{array}\right).
\end{eqnarray}
\end{widetext}
The corresponding optimal witness $Z^{(2)}_4$ is blind to intermodal correlations of pairs of modes $(A,C), (A,D)$ and $(C,D)$. It gives the value of
$\mbox{Tr}[\gamma^{(2)}_4 Z^{(2)}_4]-1\doteq-0.068$ and its explicit form can be found in Appendix~\ref{sec_app_II}. Once again, the separability of the
marginals can be verified via the PPT criterion. The results are summarized in Tab.~\ref{tableTshape}.
\begin{table}[ht]
\caption{Minimal eigenvalue $\varepsilon_{jk}^{(2)}\equiv\mbox{min}\{\mbox{eig}[(\gamma_{4,jk}^{(2)})^{(T_j)} + i \Omega_2]\}$.}
\centering
\begin{tabular}{| c | c | c | c | c | c | c |}
\hline $jk$ & AB & AC & AD & BC & BD & CD \\
\hline $\varepsilon_{jk}^{(2)}$
& 0.0481 & 0.0032 & 0.5256 & 0.1103 & 0.0001 & 0.5489 \\
\hline
\end{tabular}
\label{tableTshape}
\end{table}
The effect is about the same strength as for the linear tree.
A point to note is that, for qubits, a pure state example was found for
the `t'-shaped tree in Ref.~\cite{Paraschiv_17} while we found only a
mixed-state example in the Gaussian scenario.
\section{Experimental scheme}\label{sec_scheme}
In the previous section we have seen that the investigated effect is strongest in the three-mode case.
For this reason, we now derive a linear-optical scheme for preparing a Gaussian state with
the three-mode CM $\gamma_{3}$. The scheme is depicted in Fig.~\ref{fig2}.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.48\textwidth]{fig2.pdf}
\caption{Decomposition of symplectic transformation $S$ generating a Gaussian
state with CM $\gamma_{3}$ of three modes $A,B$ and $C$: $\nu_{j}$ -- thermal
states with mean number of thermal photons $(\nu_{j}-1)/2$, $j=A,B,C$ (red circles);
$U$ -- passive transformation consisting of beam splitters $\mathrm{BS}_{jk}^{(U)}$, $jk=AB, AC, BC$ (magenta block);
$V$ -- passive transformation consisting of beam splitters $\mathrm{BS}_{jk}^{(V)}$ (green block); $R$ -- squeezing transformation consisting of one squeezer in position quadrature, $R_{A}$,
and two squeezers in momentum quadrature, $R_{B}$ and $R_{C}$ (pink block). For rounded parameters as in Tabs.~\ref{tab:symp} and \ref{tab:bs} the circuit produces the
CM $\gamma'_{3}$, which closely approximates the CM $\gamma_{3}$, and retains its entanglement properties. See text for details.}
\label{fig2}
\end{center}
\end{figure}
The scheme follows from Williamson's symplectic diagonalisation of a CM
\cite{Williamson_36}, the Bloch-Messiah decomposition of a symplectic matrix \cite{Braunstein_05} and
the decomposition of an orthogonal symplectic matrix into an array of beam splitters and phase-shifters
\cite{Reck_94,vanLoock_02}. More precisely, according to Williamson's theorem \cite{Williamson_36}
for any CM $\gamma$ there is a symplectic transformation $\mathscr{S}$ which brings the CM to the normal
form,
\begin{equation}\label{Williamson}
\mathscr{S}\gamma\mathscr{S}^{T}=\bigoplus_{i=1}^{N}\nu_{i}\openone\equiv W,
\end{equation}
where $\nu_{1},\nu_{2},\ldots,\nu_{N}\geq1$ are the so called symplectic
eigenvalues of CM $\gamma$. In particular, $\nu_{1}=\nu_{2}=\ldots=\nu_{N}=1$ if the state is pure.
Consequently, making use of the symplectic transformation $S\equiv\mathscr{S}^{-1}$, one can write $\gamma = S W S^T$.
The symplectic eigenvalues are the magnitudes of the eigenvalues of the matrix $i \Omega \gamma$ \cite{Vidal_02} and
for CM $\gamma_3$ they are written in Tab. \ref{tab:symp}. The corresponding symplectic matrix $S$ can be found numerically
either using a method of Ref.~\cite{Serafini_05} or a method of Ref.~\cite{Pirandola_09}.
Making use of the Bloch-Messiah decomposition \cite{Braunstein_05} we further numerically decomposed the symplectic matrix $S$ into
passive transformations $U$ and $V$, and an active transformation $R$, as
\begin{equation}
\label{S}
S = V R\, U.
\end{equation}
Here, $U$ and $V$ are orthogonal and symplectic transformations and
$R = R_A(s_A) \oplus R_B(s_B^{-1}) \oplus R_C(s_C^{-1})$ is the squeezing transformation,
where $R_j(s_j) = \mbox{diag}(s_j,s_j^{-1})$, $j=A,B,C$, is the diagonal matrix and the squeezing
parameters $s_{j}<1$ may be found in Tab.~\ref{tab:symp}. These transformations are highlighted by the
coloured boxes in Fig.~\ref{fig2}.
\begin{table}[h]
\caption{Symplectic eigenvalues $\nu_j$ and the squeezing parameters $s_j$.}
\centering
\begin{tabular}{|c|c|c|c|}
\hline
$j$ & A & B & C \\
\hline
$\nu_{j}$ & 6.835 & 1.012 & 1.004 \\
\hline
$s_{j}$ & 0.396 & 0.851 & 0.478 \\
\hline
\end{tabular}
\label{tab:symp}
\end{table}
Next, following the method of Refs.~\cite{Reck_94,vanLoock_02} one can decompose the passive
transformations $U$ and $V$ into an array of three beam splitters as in Fig.~\ref{fig2},
\begin{eqnarray}
\label{UV}
U&=&B^{(U)}_{BC}(T_{BC}) B^{(U)}_{AC}(T_{AC}) B^{(U)}_{AB}(T_{AB}),\nonumber\\
V&=&B^{(V)}_{AB}(\tau_{AB})B^{(V)}_{AC}(\tau_{AC})B^{(V)}_{BC}(\tau_{BC}),
\end{eqnarray}
where the beam splitter matrices $B^{(U)}_{jk}(T_{jk})$ and $B^{(V)}_{jk}(\tau_{jk})$, $jk = AB,AC,BC$,
are given explicitly in Appendix~\ref{sec_app_beamsplitters}, and
the beam splitter transmissivities $T_{jk}$ and $\tau_{jk}$ can be found in Tab.~\ref{tab:bs}.
\begin{table}[h]
\caption{Amplitude transmissivities $T_{jk}$ and $\tau_{jk}$.}
\centering
\begin{tabular}{|c|c|c|c|}
\hline
$jk$ & AB & AC & BC \\
\hline
$T_{jk}$ & 0.555 & 0.947 & 0.492 \\
\hline
$\tau_{jk}$ & 0.716 & 0.904 & 0.657 \\
\hline
\end{tabular}
\label{tab:bs}
\end{table}
Needless to say, our decomposition is numerical and thus we rounded its
parameters to three decimal places.
Consequently, the output CM $\gamma_{3}'$ slightly deviates from the original CM $\gamma_{3}$,
yet it retains all relevant entanglement properties: it is genuinely multipartite entangled
with $\mbox{Tr}[\gamma_{3}' Z_{3}']-1 = -0.138$; and the marginals are all separable as per Tab.~\ref{tab:ppt_gamma'}.
\begin{table}[ht]
\caption{Minimal eigenvalue
${\varepsilon}_{jk}'\equiv\mathrm{min}\{\mathrm{eig}[\gamma_{3,jk}'^{(T_{j})}+i\Omega_{2}]\}$.} \centering
\begin{tabular}{| c | c | c | c |}
\hline $jk$ & AB & AC & BC \\
\hline
$\varepsilon_{jk}$ &0.005 & 0.852 & 0.010 \\
\hline
\end{tabular}
\label{tab:ppt_gamma'}
\end{table}
The CM $\gamma_{3}'$ and the corresponding witness $Z_{3}'$ can be found in Appendix \ref{app_circuit_experimental}.
In the next section, we present an equivalent, yet simpler, circuit whose output
CM still retains all required properties.
\subsection{Simplified circuit}\label{sec_scheme_simplified}
The scheme in Fig.~\ref{fig2} offers two simplifications which make its
experimental realization easier. First, inspecting Tab.~\ref{tab:symp}
one may see that the input states of modes $B$ and $C$ can be approximated by vacuua.
Second, the classically correlated state subject to the squeezing transformations can be
replaced by correlatively displaced squeezed vacuum states. This follows from the
fact that a thermal state at the input of mode $A$ can be prepared by the
displacements $x_{A}^{(0)}\rightarrow x_{A}^{(0)}+t$ and $p_{A}^{(0)}\rightarrow p_{A}^{(0)}+w$
of its position and momentum vacuum quadratures $x_{A}^{(0)}$ and $p_{A}^{(0)}$, respectively,
where $t$ and $w$ are uncorrelated classical zero mean Gaussian random variables with
second moments $\langle t^2 \rangle = \langle w^2 \rangle = (\nu_A - 1)/2$. As on the level of quadrature operators
the transformations $U$ and $R$ are linear, we can push the displacements through the transformations so that behind
the transformation $R$ they attain the following form:
\begin{equation}\label{displacements}
x_{j}\rightarrow x_{j}+\alpha_{j}t,\quad p_{j}\rightarrow p_{j}+\beta_{j}w,
\end{equation}
$j=A,B,C$, where the parameters $\alpha_{j}$ and $\beta_{j}$ after rounding read as in Tab.~\ref{tab:displacement}.
\begin{table}[h]
\caption{Parameters $\alpha_{j}$ and $\beta_{j}$ of displacements (\ref{displacements}).}
\centering
\begin{tabular}{|c|c|c|c|}
\hline
$j$ & A & B & C \\
\hline
$\alpha_j$
& 0.2
& - 0.7
& 1.3 \\
\hline
$\beta_j$
& 1.3
& - 0.5
& 0.3 \\
\hline
\end{tabular}
\label{tab:displacement}
\end{table}
Further, the first step of the obtained scheme consists of application of a passive transformation $U$ on three
vacuum states, which is nothing but a triple of vacuum states and thus the transformation $U$ can be
omitted completely. In this way, we arrive at the simplified scheme depicted in Fig.~\ref{fig3}.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=0.48\textwidth]{fig3.pdf}
\caption{Scheme for preparation of a Gaussian state with CM $\bar{\gamma}_{3}$ carrying genuine multipartite entanglement verifiable
from nearest-neighbour separable marginals. The input comprises of three vacuum states (red circles). The squeezing transformation $R$ (red box)
and the transformation $V$ (green box) are the same as in Fig.~\ref{fig2}. The block $D$ (gray box) contains correlated displacements
$D_{A}, D_{B}$ and $D_{C}$ (white boxes) given in Eq.~(\ref{displacements}), where the parameters $\alpha_{j}$ and $\beta_{j}$
are in Tab.~\ref{tab:displacement} and $\langle t^2 \rangle = \langle w^2 \rangle = (\nu_A-1)/2$. See text for details.}
\label{fig3}
\end{center}
\end{figure}
Using the squeezing parameters and transmissivities, found in the second row of
Tabs.~\ref{tab:symp} and \ref{tab:bs} respectively, as well as the displacements
in Tab.~\ref{tab:displacement}, the circuit in Fig.~\ref{fig3} produces a state
which is genuinely multipartite entangled and has all marginals separable.
Calling the CM of this state $\bar{\gamma}_3$, the optimal witness
for this CM gives $\mbox{Tr}[\bar{\gamma}_3 \bar{Z}_3] - 1 = -0.139$. The
numerical CM $\bar{\gamma}_3$ along with the corresponding entanglement witness
may be found in Appendix \ref{app_circuit_experimental}.
The simplified scheme in Fig.~\ref{fig3} makes experimental demonstration of the
investigated effect more viable. Primarily, preparation of squeezed states at
the input is easier than implementation of squeezing operations in between beam
splitter arrays $U$ and $V$ (compare positions of pink boxes $R$ in
Figs.~\ref{fig2} and \ref{fig3}). Further, the largest amount of squeezing,
$10 \mbox{Log}_{10}[(s_A)^2] \doteq -8\, \mbox{dB}$,
is well within the reach of the current technology \cite{Vahlbruch_16}, and what is more,
one may decrease the squeezing required at the cost of decreased effect strength.
Additionally, the effect is immune to rounding of CMs and some parameters of the circuit
components, which indicates, that perfect matching of the setup parameters with the theoretical
values is not critical for its demonstration. Finally, as we have already mentioned, the output state
tolerates the addition of a small amount of thermal noise, which is, however, of
the same size as for the comparably fragile, yet
already demonstrated similarly complex setup \cite{DiGuglielmo_11}. The extent to which the relatively low noise
tolerance and other imperfections are detrimental to observability of the
investigated phenomenon depends on the used experimental
platform and will be addressed elsewhere.
\section{Conclusions}\label{sec_discussion}
In this paper we extended the concept of genuine multipartite entanglement verifiable from separable marginals to the
domain of Gaussian states. We constructed many examples of Gaussian states possessing all two-mode marginals
separable and whose genuine multipartite entanglement can be certified solely from the set of nearest-neighbour
marginals. Each of the sets is characterized by a connected graph with no cycles, where the vertices represent the modes and the edges
the nearest-neighbour marginals. Our examples are numerical and result from an iterative search algorithm
relying on construction of a genuine multipartite witness in the space of covariance matrices. Moreover, the witness
is `blind' to correlations between modes corresponding to non-adjacent vertices in the respective graph.
Here, we gave examples for all configurations encompassing three and four modes thus complementing the study of the investigated
phenomenon in multi-qubit systems \cite{Paraschiv_17}. The three-mode state found by us
exhibits the strongest form of the property compared to the four-mode cases and therefore we also proposed a
scheme for preparation of the state, which consists of three quadrature squeezers sandwiched between two triples of
phase-free beam splitters. Further, we replaced the original scheme by a simpler scheme, which still produces the desired effect,
but requires only interference of three squeezed states subjected to correlated displacements
on three beam splitters. The squeezing used in the setup is well within the reach of the
current technology. Additionally, all relevant properties of the output state remain preserved after
contamination by a small amount of thermal noise which gives us a hope that the investigated property
of genuine multipartite entanglement could be observed. A successful realization of the proposed setup
would mean extension of the experimental analysis of the phenomenon of emergent genuine multipartite
entanglement \cite{Miklin_16} from qubits and the scenario when all bipartite marginals are known
\cite{Micuda_19}, to the realm of Gaussian states and more generic situation when only some bipartite
marginals are known.
The impact of the presented results is twofold. On one hand, they point at an alternative approach
towards experimental investigation of the remarkable concept of genuine multipartite entanglement
verifiable from incomplete sets of separable marginals. On the other hand, they also stimulate theoretical questions
concerning the existence of a Gaussian classical analog of the quantum marginal problem \cite{Leskovjanova_20} or the extendibility of the
entanglement marginal problem \cite{Navascues_20} to Gaussian case. On a more general level, our results
contribute to the development of methods of detection of global properties of multipartite quantum systems
from partial information.
\acknowledgments
O.L. and J.P. acknowledge support from IGA-PrF-2020-009 and IGA-PrF-2021-006.
J.P. also acknowledges project GA18-21285S of the Grant Agency of Czech
Republic (GAČR).
V.N. and N.K. have been supported by the Scottish Universities Physics
Alliance (SUPA) and by the Engineering and Physical Sciences Research Council
(EPSRC). N.K. was supported by the EU Flagship on Quantum Technologies, project
PhoG (820365).
\section*{\noindent Competing interests}
\vspace{-1.3em}
\noindent The authors declare that there are no competing interests.
\vspace{-2em}
\section*{Author contribution}
\vspace{-1.3em}
\noindent L.M. conceived the theory,
J.P., V.N. and O.L. contributed to writing of the code,
V.N., O.L., J.P. and L.M. carried out calculations,
V.N., L.M. and N.K. wrote the manuscript,
N.K. and L.M. supervised the project,
all authors discussed the paper.
\vspace{-2em}
\section*{Data availability}
\vspace{-1.3em}
\noindent The data generated or analysed during the current study are available
from the corresponding author upon reasonable request.
\vspace{-1.5em}
|
\section{Introduction}
Based on the seminal work of V. Hill \cite{Gesztelyi2012}, recent mathematical models
(validated by experiments) allow to predict the force response
to external stimulation.
They are presented and discussed in details in \cite{wilson2011} in the non-fatigue isometric case.
They were extended in particular by Ding et al. \cite{Ding2000,Ding2002,Ding2002b} to take
into account the muscular fatigue due to a long stimulation period
and later in \cite{Marion2013} to analyze the joint angular variable response, in the
non-isometric case aiming to produce movements.
Such models contain two basic nonlinearities to model the complexity of the dynamics.
Fist of all, the ionic conduction and the nonlinear effect of successive
pulses on the Ca$^{2+}$-concentration.
Second, the nonlinear dynamics relating the
muscular force response to such concentration,
modeled by the Michaelis-Menten-Hill functions \cite{menten1913}.
For each train of pulses, due to digital constraints, only a finite number of pulses can be applied
and from the optimal control point of view, the problem fits into the frame of optimal sampled-data control
problems, studied in particular in \cite{Bourdin2016} to derive Pontryagin necessary conditions.
They can be analyzed
to determine optimized train pulses and compared with direct optimizing schemes.
A previous series of articles described the optimal control problems related to track a reference force
or force strengthening, the control being either the interpulse $I_i=t_i-t_{i-1}$ between two
successive pulses or the amplitude of each pulse.
In particular, model predictive control (MPC) method is presented in \cite{Bakir2019} aiming the use
of online optimized closed loop control in the applications using force-fatigue model, as suggested in \cite{Doll2015}.
Direct methods vs indirect methods based on Pontryagin type necessary conditions are discussed and numerically
implemented in \cite{Bakir2020} for the isometric case or in \cite{bonnard2020} for the non-isometric case.
The conclusion of aforementioned articles is that the nonlinear dynamics is computationally expensive in the numerical
integration procedure and a challenging task is to reduce this time for real time computation in the
applications.
This article is motivated by the design of a smart electrostimulator, where the Ding et al. model
is used to adjust automatically the frequency and the amplitude of the stimulations and to compute the
sequence of stimulations and rest periods adapted to the task of the training program, e.g. endurance
program or force strengthening program.
The objective of this article being to bypass the computational difficulty by constructing
a finite dimensional approximation of the force response,
depending upon the parameters of each individual, which can be online estimated, aiming a
real time computation of the optimized amplitudes and times, for each training program.
Note that this approximation has been coded and the application scheme to the smart
electrostimulator is briefly presented in the final section.
The article is organized as follows.
In section \ref{sec2}, the mathematical model called the Ding et al. model \cite{Ding2000,Ding2002,Ding2002b}
is presented and the main properties of the dynamics are described, reflecting the features of the
muscular activity. Hence our analysis can be applied to different models discussed in \cite{wilson2011}.
The section \ref{sec3} presents the optimization problems, in relation with muscular dynamics and oriented towards
the design of a smart electrostimulator, where each training program must be translated into an
optimization problem.
Section \ref{sec4} is the technical contribution of this article,
that is the construction of the approximation for real time computation.
In section \ref{sec5}, we present some numerical simulations aiming to validate the approximation
and the optimizing scheme.
In the final section \ref{sec6}, we outline the application to the design of the smart electrostimulator.
It is based on a nonlinear output tracking \cite{hirschorn1987,hirschorn1988,Isidori1989}
as a general theoretical frame and is applied to produce a constant force in our situation, but it can
be extended to the non-isometric case to obtain any reference force.
The conclusion indicates directions to complete our analysis, related to online parameters estimation
of the problems \cite{Stein2013,wilson2011} and MPC-methods \cite{Richalet1993,Boyd2010} suitable for
practical applications.
\section{Mathematical model and main properties} \label{sec2}
We present the Ding et al. model force-fatigue model \cite{Ding2000,Ding2002,Ding2002b},
extension of the original Hill model \cite{Gesztelyi2012}.
\subsection{Ding et al. force-fatigue model \cite{Ding2000,Ding2002,Ding2002b}}
The FES input $u$ over a pulse train $[0,T]$ is given by
\begin{equation}
u(t) = \sum_{i=0}^n \eta_i \delta(t-t_i), \quad t\in [0,T],
\label{eq:dirac-peigne}
\end{equation}
where $0=t_0<t_1<\dots< t_n<T$ are the impulsion times with $n\in \mathbb{N}$ being fixed and
$\eta_i$ being the amplitudes of each pulses, which are convexified by taking $\eta_i\in [0,1]$,
$\delta(t-t_i)$ denoting the Dirac function at time $t_i$.
Such physical control will provide the FES-signal denoted by $E(t)$, which drives the force response
using electrical conduction and its dynamics is given by
\begin{equation}
\dot E(t) + \frac{E(t)}{\tau_c} = \frac{1}{\tau_c} \sum_{i=0}^n R_i \eta_i \delta(t-t_i), \ a.e.\ t\in [0,T],
\label{eq:Edot}
\end{equation}
with $E(0)=0$, depending upon the time response $\tau_c$ and the scaling function
$R_i$ defined by
\[R_i = \left\{
\begin{array}{ll}
1 & \text{ if } i=0
\\
1 + (\bar R-1)\, e^{-(t_i - t_{i-1})/\tau_c} & \text{ otherwise },
\end{array}
\right.\]
which codes the memory effect of successive muscle contractions and is associated to {\it tetanus} \cite{wilson2011}.
The first result is:
\begin{lem}
Integrating \eqref{eq:Edot}, one gets
\begin{equation*}
E(t) = \frac{1}{\tau_c} \sum_{i=0}^n R_i e^{-\frac{t-t_i}{\tau_c}} \eta_i H(t-t_i),
\label{eq:E}
\end{equation*}
where $H$ is the Heaviside function and the FES signal depends upon two parameters $(\tau_c,\bar R)$.
\end{lem}
\begin{defn}
Consider a control system of the form: $\frac{\mathrm{d} x}{\mathrm{d} t}=f(x,u)$ where $x\in \mathbb{R}^n$,
$u\in U\subset \mathbb{R}^m$.
It is said {\it permanent} if $u$ is a measurable bounded mapping valued in $U$.
It is called a sampled-data control system if the set of controls is restricted to the set of
piecewise constant mappings $[u_0,u_1,\dots,u_n]$, $u_i\in U$ over a set of times
$t_0=0<t_1<\dots <t_n<T$, where $n$ is a fixed integer.
\label{def:control}
\end{defn}
Our problem can be formulated in the sampled-data control frame.
One can write from \eqref{eq:E},
\begin{equation*}
E(t) = \frac{e^{-t/\tau_c}}{\tau_c}\ \sum_{i=0}^n R_i\eta_i e^{-t_i/\tau_c}\, H(t-t_i)
= \sum_{i=0}^n u_i(t),
\label{eq:cn}
\end{equation*}
where $u_i(t)$ is the effect of the pulse $\eta_i \delta(t-t_i)$ on the linear dynamics \eqref{eq:Edot}.
One introduces the following.
\begin{defn}
For each $i$ in $\{0,\dots,n\}$,
the restriction of $u_i$ to $[t_i,t_{i+1}]$ is called {\it the head} and the
restriction to $[t_{i+1},T]$ is called {\it the tail}.
\end{defn}
Clearly the FES-input is in the generalized frame of sampled data control system, provided
we take into account the time-dependence and the phenomenon of tetanus.
Observe also that each impulse have an effect on the whole train $[0,T]$.
The FES signal drives the evolution of the electrical conduction according to the linear
dynamics describing the evolution of $Ca^{2+}$-concentration $c_N$:
\begin{equation}
\dot c_N(t) + \frac{c_N(t)}{\tau_c} = E(t)
\label{eq:cNdot}
\end{equation}
and integrating the (resonant) system with $c_N(0)=0$ yields the following:
\begin{prop}
The concentration is
\begin{equation}
c_N(t) = \frac{1}{\tau_c} \sum_{i=0}^n R_i \eta_i (t-t_i)\,e^{-\frac{t-t_i}{\tau_c}}\, H(t-t_i),
\label{eq:cN}
\end{equation}
which are the superposition of lobes of the form
\begin{equation}
\ell_i(t) = \frac{1}{\tau_c} R_i \eta_i (t-t_i)\,e^{-\frac{t-t_i}{\tau_c}},
\label{eq:ci}
\end{equation}
whose restriction to $[t_i,t_{i+1}]$ forms the head of the corresponding lobe.
\label{prop:cN}
\end{prop}
\begin{pf}
Apply a time translation to the initial lobe with $R_0=1$.
\end{pf}
Introducing the functions
\begin{equation}
m_1(t) = \frac{c_N(t)}{K_m+c_N(t)},\quad m_2(t) = \frac{1}{\tau_1 + \tau_2\, m_1(t)},
\label{eq:m1m2}
\end{equation}
where $m_1$ is the {\it Michaelis-Menten-Hill function} \cite{menten1913}, the force response satisfies
the Hill dynamics
\begin{equation}
\dot F(t) = -m_2(t)\, F(t) + m_1(t) A,
\label{eq:dotF}
\end{equation}
and where $A,K_m,\tau_1,\tau_2$ being additional parameters and we denote by $\Lambda =
(\bar R, \tau_c, A,K_m,\tau_1,\tau_2)$ the whole set of parameters. \\
The model can be extended to take into account the fatigue.
Using sensitivity analysis from \cite{bonnard2020}, we shall restrict
our study to the case of the {\it force-fatigue Ding et al. model} with the single
equation:
\begin{equation}
\dot A(t) = -\frac{A(t) - A_{rest}}{\tau_{fat}} + \alpha_A \, F(t)
\label{eq:dotA}
\end{equation}
for all $t\in [0,t_f]$, where $t_f$ is the total time and $A(0)=A_{rest}$
corresponds to the fixed value of $A$ for the non fatigue model.
This leads to introduce additional parameters $\tau_{fat}, \alpha_A$.
Typical parameters values used in this article to validate numeric simulations
are reported in Table \ref{tb:params} .
\subsection{Mathematical rewriting}
For the previous force-fatigue model and for the sake of the analysis, the model is rewritten
as the control system:
\begin{equation*}
\dot x(t) = g(x(t)) + b(t)\, \sum_{i=0}^n G(t_{i-1},t_i)\eta_i\,H(t-t_i)\, \bm{e}
\label{eq:dotx}
\end{equation*}
with $x=(x_1,\dots,x_8)^\intercal =(c_N,F,A,\bar R, \tau_c, \tau_1,\tau_2,K_m)^\intercal$
which splits into state variables $(c_N,F,A)$ and fatigue parameters $\Lambda =
(\bar R, \tau_c,\tau_1,\tau_2,K_m)$ satisfying the dynamics
$
\dot \Lambda (t) = 0,
$
the system being integrated with the initial condition $x_0=(0,0,A_{rest}, \Lambda(0))^\intercal$
and
\begin{equation*}
\begin{array}{ll}
\bm{e}=(1,0\dots,0)^\intercal, & b(t)=\frac{1}{\tau_c}e^{-t/\tau_c},
\\
G(t_{i-1},t_i) = (\bar R-1) e^{t_{i-1}/\tau_c} + e^{t_i/\tau_c},
\end{array}
\end{equation*}
where $t_{-1}=-\infty, t_0=0$ and $t_{n+1}=T$.
This leads to a control system of the form
\[
\dot x(t) = g(x(t)) + b(t) \, \sum_{i=0}^n G(t_{i-1},t_i) \eta_i H(t-t_i)\ \bm{e}
\]
with $x(0)=x_0$.
The variable $\sigma=(t_1,\dots,t_n,\eta_0,\eta_1,\dots,\eta_n)$ denotes the finite dimensional input-space
with the constraints
\begin{equation*}
\begin{aligned}
&\eta_i \in [0,1],\ i=0,\dots , n
\\
&0<t_1<\dots< t_n<T, \quad t_i-t_{i-1}\ge I_{\min}, \quad i=1,\dots,n\ ,
\end{aligned}
\end{equation*}
where $I_{\min}$ is the smallest admissible interpulse.
Moreover the control is observed using the following observation function
\begin{equation}
y(t) = h(x(t)),
\label{eq:obs-function}
\end{equation}
and $h:x \mapsto (F,A)$ serves as a direct measure of the muscular force and the fatigue variable.
The following properties are straightforward but crucial in our analysis.
\begin{prop}
The input-output mapping $\sigma \mapsto y(t)$ is piecewise smooth over $[0,T]$ and smooth
if $t\neq t_i,\ i=0,\dots, n$ (impulse times).
\label{prop:ysmooth}
\end{prop}
\begin{prop}
For the non fatigue model, the force response can be integrated up to a time reparameterization as
\begin{equation}
F(s) = \int_0^s e^{u-s} m_3(u)\, \mathrm{d} u
\label{eq:Fs}
\end{equation}
with
\begin{equation}
m_3(s) = A \frac{m_1(s)}{m_2(s)}, \quad \mathrm{d} s = m_2(t)\,\mathrm{d} t.
\label{eq:m3s}
\end{equation}
\label{prop:Fs}
\end{prop}
\begin{pf}
Hill dynamics \eqref{eq:dotF} is rewritten as
\[
\frac{\mathrm{d} F}{\mathrm{d} s} = m_3(s) - F(s)
\]
and this linear dynamics can be integrated using Lagrange formula with $F(0)=0$.
This proves the assertion. $\hfill \qed$
\end{pf}
\section{Optimization problems related to the design of the electrostimulator} \label{sec3}
\subsection{Standard electrostimulators vs smart electrostimulators}
The standard commercial electrostimulators apply a sequence of pulses trains and rest periods,
where on each train $[0,T]$ the user only imposes the amplitude of the pulses trains and the constant frequency
is related to training program, typically low frequency for endurance program and high frequency
for force strengthening program.
Our aim is to introduce optimization problems related to the design of a smart electrostimulator,
which will be discussed in Section \ref{sec6}.
\subsection{Optimization problems}
\label{secOCP}
\subsubsection{The punch program}
In this case, our aim is to optimize the force at the end of the train over each train $[0,T]$.
This leads to:
{\small {\bf OCP1:} $\max_{\sigma} F(T)$.}\\
In this case, the amplitudes can be held at the constant maximal values $\eta_i=1,\ i=0,\dots,n$
and the optimization variables are the impulse times:
\[
0=t_0<t_1<\dots<t_n<T.
\]
Since one considers a single train, the force model is sufficient.
\subsubsection{The train endurance program}
We consider a single train $[0,T]$ on which the model is the force model and the
corresponding problem is
{\small {\bf OCP2:} $\min_{\sigma} \displaystyle \int_0^T |F(t)-F_{ref}|^2 \, \mathrm{d} t$.}\\
Here, the amplitudes are appended to the impulse times to form the optimization variables
and we use the convexified amplitudes constraints: $\eta_i \in [0,1],\ i=0,\dots,n$.
The force reference has to be adjusted in relation with the user and can be set to
$F_{\max}/k$, where $k$ is a suitable positive number greater than 1 and $F_{\max}$
is deduced from {\bf OCP1}.
\subsubsection{The endurance program}
We consider an interval $[0,t_f]$, where $t_f$ is the total training period formed
by sequences of stimulation and rest periods.
In this case, one must use a force-fatigue model and we take into account the constraint
$A\in [A_{rest},A_{rest}/k'']$, where $S=A_{rest}/k''$ corresponds to a {\it fatigue threshold} since,
as reported in \cite{Ding2000}, if the user is exhausted, the force signal is
{\it totally noisy}.
Moreover in the case of exhaustion, a large rest period is required.
This constraint can be penalized as follows
{\small {\bf OCP3:} $\min_{\sigma} \displaystyle \int_0^{t_f} |F(t)-F_{ref}|^2 \, \mathrm{d} t \, +\,
w_1\, \int_0^{t_f} |A(t)-A_S|^2 \, \mathrm{d} t$},\\
where $A_S$ is related to $S$, while $w_1$ is a weight parameter.
\section{Construction of an integrable model for real time application} \label{sec4}
\subsection{Mathematical analysis of $c_N$}
A pulses train is defined by a finite sequence of impulse times $\sigma = (t_i)_{0\le i\le n}$
such that $t_0<\dots<t_n$ and we extend it to the left by $t_{-1}=-\infty$ and to the right by
$t_{n+1}=T$.
The response $c_N$ can be decomposed as a sum of lobes defined as follows.
\begin{defn}
A lobe at $t_k$ is the representative curve of the function $\ell_k:\mathbb{R}\ni t\mapsto R_k\eta_k\, \frac{t-t_k}{\tau_c}\, e^{-(t-t_k)/\tau_c}\, H(t-t_k)$.
\label{def:lobe}
\end{defn}
\begin{property}
\begin{itemize}
\item A lobe at $t_k$ reaches its maximum
at $t=t_k+\tau_c$ and is equal to $R_k\eta_k/e$. It is strictly increasing on $[t_k, t_k+\tau_c]$ and
strictly decreasing $[t_k+\tau_c, t_{k+1}]$.
\item $\ddot \ell_k$ has a unique zero at $t_k + 2\tau_c$ and therefore,
$\ell_k$ is concave on $[t_k,t_k+2\tau_c]$ and convex on $[t_k+2\tau_c,t_{k+1}]$.
\item $\ell$ defines a density probability function
and more than $95\%$ of the values lie in $[t_k,t_k+5\tau_c]$.
For all $t\ge t_{k}+5\tau_c$, $|\ell_k(t)|\le R_k \eta_k\, 5e^{-5}$.
\label{prop:ell}
\end{itemize}
\end{property}
\begin{prop}
Denote for $k=0,\dots,n$, $c_N^k = c_{N_{\mid [t_k,t_{k+1}]}}$ and
$\displaystyle \bar c_N^k = \frac{1}{t_{k+1}-t_k}\ \int_{t_k}^{t_{k+1}} c_N(t)\, \mathrm{d} t$.
We have
\begin{equation}
\begin{aligned}
& c_N^k= \sum_{i=0}^k R_i \eta_i \frac{t-t_i}{\tau_c} e^{-(t-t_i)/\tau_c},
\\
& \bar c_N^k= \frac{1}{t_{k+1}-t_k} \, \sum_{i=0}^k R_i \eta_i
\left( \chi_i(t_k)-\chi_i(t_{k+1}) \right),
\end{aligned}
\label{eq:cnk}
\end{equation}
where $\chi_i(t)=
e^{-(t-t_i)/\tau_c}
\left(\tau_c + t-t_i\right)$.
\label{prop:cnk}
\end{prop}
\begin{defn}
The polynomial-exponential category for (piecewise) smooth functions
$[0,T] \mapsto \mathbb{R}$ is the category generated by sums, products of polynomials
$P(t)$ and exponential mappings to generate exponential-polynomials:
$\sum_n P_n(t) e^{\lambda_n t}$.
This category is stable with respect to derivation and integration.
\label{def:category}
\end{defn}
Using proposition \ref{prop:cN} one has:
\begin{lem}
For $t\neq t_i$, $c_N(t)$ is in the polynomial-exponential category.
Moreover, the coefficients are linear with respect to $\eta_i$ and polynomial-exponential with respect to $t_i$.
\label{lem:category}
\end{lem}
We introduce the notion of $p$-persistent pulses related to the case where $p$th successive lobes have an
influence on the $(p+1)$th lobe.
\begin{defn}
Let $t_1<\dots<t_{p}$ be $p$ successive impulses times of a pulses train satisfying
for any $i\in \llbracket 1,p \rrbracket$, $t_i\le t_{i-1} + 5$
and $t_i-t_{i-1}\ge I_{\min}$.
When such integer $p$ is maximal then the pulses train is said $p$-{\it persistent}.
\label{def:proper}
\end{defn}
\begin{rem}
Given an $1$-persistent pulses train $(t_i)_{1\le i\le n}$ ($n>1$),
there exists $j\in \llbracket 1,n\rrbracket$ such that
$t_j> t_{j-1}+5$. Then, by Property \ref{prop:ell}, $c_N^j$ is well approximated
by $t \mapsto (t-t_j)\, e^{-(t-t_j)}$ for $t\in [t_j,t_{j+1}]$ (the factor $R_j^\dagger$ has a negligible effect).
\end{rem}
We define now an approximation of $c_N$ denoted as $\tilde c_N$ that will be used
to construct an approximation of the force $F$ limiting the number of terms and the error
between $c_N$ and $\tilde c_N$ is analyzed in the following proposition.
\begin{prop}
Let $(t_i)_{1\le i\le n}$ be a $p$-persistent pulses train.
Denote \[c_N(t)\coloneqq\sum_{i=0}^k R_i^\dagger \eta_i (t-t_i)\, e^{-(t-t_i)}\]
and define its (lower) approximation by
\[\tilde c_N(t) \coloneqq \sum_{i=\max(0,k-p+1)}^k R_i^\dagger \eta_i(t-t_i)\, e^{-(t-t_i)}\]
for $t\in [t_k,t_{k+1}],\ k=0,\dots,n$.\\
Then, we have:
\[\sup_{t\in [t_k,t_{k+1}]} \ c_N(t)-\tilde c_N(t) \le \frac{\bar R}{e}\,
\kappa + 5e^{-5}\bar R (k-p-\kappa+1) ,\]
where $\lceil\cdot \rceil$ is the ceiling function and
$\kappa=\min\left(p,\left\lceil\frac{5\tau_c}{I_{\min}}\right\rceil\right)$
is independent of $k$.
\label{prop:err}
\end{prop}
\begin{pf}
For $t\in [t_k,t_{k+1}],\, k=0,\dots,n$, we have:
\[
c_N(t)
\ge \sum_{i=k-p+1}^k R_i^\dagger \eta_i (t-t_i) e^{-(t-t_i)} =\tilde c_N(t)
\]
and
\[
c_N(t) - \tilde c_N(t)
\le \bar R \sum_{i=0}^{k-p} (t-t_i) e^{-(t-t_i)}.
\]
The number of indices $i\in \{0,\dots,k-2\}$ for which $t-t_i\le 5\tau_c$ is at most
$\kappa\coloneqq\min\left(p,\left\lceil\frac{5\tau_c}{I_{\min}}\right\rceil\right)$ since
$(t_i)_i$ is $p$-persistent and $\left\lceil\frac{5\tau_c}{I_{\min}}\right\rceil$ stands for
the maximum number of impulse times satisfying the constraint $t_i-t_{i-1}\ge I_{\min}$
in an interval of length $5\tau_c$.
\end{pf}
\begin{prop}[Tail approximation of $c_N$]
Let $q\in \{0,\dots,n\}$.
Denote the tail of $c_N$ by $c_N^q = c_{N\rvert_{[t_q,T]}}$
and its average over $[t_q,T]$ by $\bar c_N^q$. We have:
\begin{equation}
\begin{aligned}
\bar c_N^q= \frac{1}{T-t_q} \, \sum_{i=0}^q R_i \eta_i
\left( \chi_i(t_q)-\chi_i(T) \right)
+
\frac{1}{T-t_q} \, \sum_{i=q+1}^n R_i \eta_i
\left( 1 - \chi_i(T) \right),
\end{aligned}
\label{eq:cnq}
\end{equation}
where $\chi_i(t)=
e^{-(t-t_i)}
\left(1+t-t_i\right)$.
\label{prop:cnq}
\end{prop}
\subsection{Approximations of $F$}
Integrating \eqref{eq:dotF}, the force with $F(0)=0$ can be written as
\begin{equation}
F(t) = A M(t)\, \int_{0}^t M^{-1}(s) m_1(s)\, \mathrm{d} s, \ t\in [0,T]
\label{eq:Fvrai}
\end{equation}
where $\displaystyle M(t) = \exp\left(-\int_{0}^t m_2(s)\, \mathrm{d} s\right)$.
The following properties show that it is natural to approximate $m_1$ and
$m_2$ by polynomial functions.
\begin{property}
Let $k\in \{0,\dots,n\}$.
\begin{itemize}
\item Denote $t^* = \underset{t_\in [t_k,t_{k+1}]}{\text{argmax }}\, c_N(t)$.
Then,
$m_1$ (resp. $m_2$) is strictly increasing (resp. decreasing) on $[t_k,t^*]$
and strictly decreasing (resp. increasing) of $[t^*,t_{k+1}]$.
\item
If $t_{k+1}<t_k+2\tau_c$ then ${m_1}_{\mid [t_k,t_{k+1}]}$ is concave and
${m_2}_{\mid [t_k,t_{k+1}]}$ is convex.
\end{itemize}
\end{property}
We consider a finer partition of $(t_i)_{1\le i\le n}$ denoted
as $(t_{i+j/p})_{0\le i\le n,\, 0\le j \le p-1}$, $p\in \mathbb{N}^*$,
such that it satisfies
$
t_i< t_{i+1/p} < \dots < t_{i+(p-1)/p} < t_{i+1}.
$
We approximate $m_1$ and $m_2$ on each interval $[t_{i+j/p},t_{i+(j+1)/p}]$ by a polynomial function
denoted respectively by $\tilde m_1$ and $\tilde m_2$.
\begin{exmp}[Triangular approximation of a lobe.]
\label{ex:affine-approx}
Since $\dot m_1 = K_m \dot c_N /(K_m+c_N)^2$ and $\dot m_2 = \tau_2\dot m_1/(\tau_1+ \tau_2m_1)^2$,
then $\dot m_1, \ \dot m_2$ are zero when
$c_N$ is maximal.
On $[t_{k+j/2},t_{k+(j+1)/2}]$, $j=0,1$, $m_i,\ i=1,2$ can be approximated by
\[\tilde m_i(t) = a_{ij,k}\,(t-t_{k+j/2})+b_{ij,k},\ k=0,\dots, n,\]
where $t_{k+1/2} = \underset{t\in [t_k,t_{k+1}]}{\text{argmax }} \ c_N(t)$.
Computing, we have
$t_{1/2} = \tau_c, \text{ and for } k=1,\dots,n:$
\[
t_{k+1/2} = \underset{t_\in [t_k,t_{k+1}]}{\text{argmax }}\, c_N(t) = \tau_c +
\frac{\sum_{i=1}^k R_i\eta_i\, t_i\, e^{t_{i}/\tau_c}}{\sum_{i=0}^k R_i\eta_i\, e^{t_{i}/\tau_c}}.
\]
Imposing $\tilde m_i(t_{k+j/2})=m_i(t_{k+j/2})$ and $\tilde m_i(t_{k+(j+1)/2})=m_i(t_{k+(j+1)/2})$, we get:
\[
a_{ij,k} = \frac{m_i(t_{k+(j+1)/2}) - m_i(t_{k+j/2})}{t_{k+(j+1)/2}-t_{k+j/2}},
\
b_{ij,k} = m_i(t_{k+j/2}).
\]
\end{exmp}
Take $k_s\in \{0,\dots,n\}$, $j_s\in \{0,\dots,p-1\}$ and $t\in [t_{k_s+j_s/p},t_{k_s+(j_s+1)/p}]$
and let $\Psi(u; i,j)$
be the primitive of $\tilde m_2$
on $[t_{i+j/p},t_{i+(j+1)/p}]$, zero at $t=t_{i+j/p}$.
We have for $t\in [t_{k_t+j_t/p},t_{k_t+(j_t+1)/p}]$:
\begin{equation}
\begin{aligned}
\tilde M(t) =
\exp\left(-
\sum_{k=0}^{k_s-1} \sum_{j=0}^{p-1}
\left[
\Psi(u;k,j)
\right]_{t_{k+j/p}}^{t_{k+(j+1)/p}}
-\sum_{j=0}^{j_s-1}
\left[
\Psi(u;k_s,j)
\right]_{t_{k_s+j/p}}^{t_{k_s+(j+1)/p}}
-\left[
\Psi(u;k_s,j_s)
\right]_{t_{k_s+j_s/p}}^{t}
\right),
\end{aligned}
\label{eq:Mkapp}
\end{equation}
and for $s\in [t_{k_s+j_s/p},t_{k_s+(j_s+1)/p}]$, $t\in [t_{k_t+j_t/p},t_{k_t+(j_t+1)/p}]$, we get:
\begin{equation}
\begin{aligned}
\tilde M(t) \tilde M^{-1}(s)
=\exp\Bigg( &-
\left[
\Psi(u;k_s,j_s)
\right]_s^{t_{k_s+j_s/p}}
+
\sum_{j=0}^{j_s-1} \left[
\Psi(u;k_s,j)
\right]_{t_{k_s+j/p}}^{t_{k_s+(j+1)/p}}
-
\sum_{j=0}^{j_t-1}\left[
\Psi(u;k_t,j)
\right]_{t_{k_t+j/p}}^{t_{k_t+(j+1)/p}}
\\
&-
\sum_{i=k_s}^{k_t-1}\sum_{j=0}^{p-1}
\left[
\Psi(u;i,j)
\right]_{t_{i+j/p}}^{t_{i+(j+1)/p}}
- \left[
\Psi(u;k_t,j_t)
\right]_{t_{k_t+j_t/p}}^t
\Bigg).
\end{aligned}
\label{eq:MtMsinv}
\end{equation}
To integrate the product $\tilde M(t) \tilde M^{-1}(s)\tilde m_1(s)$ with respect to $s$,
we gather the terms depending on $s$ in \eqref{eq:MtMsinv} together and we get,
for $t\in [t_{k_t+j_t/p},t_{k_t+(j_t+1)/p}]$:
\begin{equation}
\begin{aligned}
&\int_{t_{k_s+j_s/p}}^{t_{k_s+(j_s+1)/p}}
\tilde M(t)\tilde M^{-1}(s) \, \tilde m_1(s)\, \mathrm{d} s
=
\int_{t_{k_s+j_s/p}}^{t_{k_s+(j_s+1)/p}}\exp\left(
\Psi(s;k_s,j_s)
\right)\ \tilde m_1(s)\, \mathrm{d} s
\\
&\exp \Bigg(- \Psi(t;k_t,j_t)
+
\sum_{j=0}^{j_s-1}
\Psi(t_{k_s+(j+1)/p};k_s,j)
-
\sum_{j=0}^{j_t-1}
\Psi(t_{k_t+(j+1)/p};k_t,j)
-
\sum_{i=k_s}^{k_t-1}\sum_{j=0}^{p-1}
\Psi(t_{i+(j+1)/p};i,j)
\Bigg).
\end{aligned}
\label{eq:intMMm1}
\end{equation}
Consequently, we obtain
an approximation of $F$ on $[t_{k_t+j_t/p},t_{k_t+(j_t+1)/p}]$, $k_t=0,\dots,n$, $j_t=0,\dots,p-1$
writing:
\begin{equation}
\begin{aligned}
\tilde F(t)/A0 &=
\int_{0}^{t}
\tilde M(t)\tilde M^{-1}(s) \, \tilde m_1(s)\, \mathrm{d} s
\\
&=
\sum_{i=0}^{k_t-1}\sum_{j=0}^{p-1}
\int_{t_{i+j/p}}^{t_{i+(j+1)/p}}
\tilde M(t)\tilde M^{-1}(s) \, \tilde m_1(s)\, \mathrm{d} s
+
\sum_{j=0}^{j_t-1}
\int_{t_{k_t+j/p}}^{t_{k_t+(j+1)/p}}
\tilde M(t)\tilde M^{-1}(s) \, \tilde m_1(s)\, \mathrm{d} s
\\
&+
\int_{t_{k_t+j_t/p}}^{t}
\tilde M(t)\tilde M^{-1}(s) \, \tilde m_1(s)\, \mathrm{d} s.
\end{aligned}
\label{eq:Ftilde}
\end{equation}
\begin{prop}
Choosing $\tilde m_1$ as a piecewise polynomial function and $\tilde m_2$ as a piecewise
constant function on $[0,T]$, the function:
\begin{equation*}
\tilde F(t) =
A\,\int_{0}^{t}
\tilde M(t)\tilde M^{-1}(s) \, \tilde m_1(s)\, \mathrm{d} s,
\end{equation*}
where
$\displaystyle \tilde M(t) =
\exp\left(-\int_0^t \tilde m_2(s)\,\mathrm{d} s
\right),$
has a closed-form expression in the polynomial-exponential category.
\end{prop}
\begin{pf}
Decomposing the integral \eqref{eq:Ftilde} as a sum of integrals over
the partition $(t_{i+j/p})_{ij}$ gives the
expression:
\begin{align*}
\tilde F(t)/A &=
\sum_{i=0}^{k_t-1}\sum_{j=0}^{p-1}
\int_{t_{i+j/p}}^{t_{i+(j+1)/p}}
\tilde M(t)\tilde M^{-1}(s) \, \tilde m_1(s)\, \mathrm{d} s
\\
&+
\sum_{j=0}^{j_t-1}
\int_{t_{k_t+j/p}}^{t_{k_t+(j+1)/p}}
\tilde M(t)\tilde M^{-1}(s) \, \tilde m_1(s)\, \mathrm{d} s
\\
&+
\int_{t_{k_t+j_t/p}}^{t}
\tilde M(t)\tilde M^{-1}(s) \, \tilde m_1(s)\, \mathrm{d} s, \addtocounter{equation}{1}\tag{\theequation}
\label{eq:Ftilde2}
\end{align*}
for $t\in [t_{k_t+j_t/p},t_{k_t+(j_t+1)/p}]$, $k_t=0,\dots,n$, $j_t=0,\dots,p-1$,
and each of this integral term belongs to the polynomial-exponential category.
\end{pf}
\begin{rem}
To construct $\tilde F$, the functions $m_1$ and $m_2$ were considered independently in the sense that
the approximation
does not rely on the relation \eqref{eq:m1m2} between $m_1$ and $m_2$.
A direct consequence is that
an upper approximation of the force can be obtained from an upper approximation of $m_1$.
Outside the scope of this paper, this method may be applied for more general non-autonomous models.
\end{rem}
\begin{prop}
Adding a real parameter $\nu$ to the functions $\tilde m_1,\tilde m_2$ as follows
\begin{equation*}
\overset{\asymp}{m_1}(t;\nu) = \frac{c_N(t)}{\nu\, K_m + c_N(t)} \text{ and }
\overset{\asymp}{m_2}(t;\nu) = \frac{\nu}{\tau_1 + \tau_2\, m_1(t)}
\end{equation*}
allows to construct an upper (or lower) approximation $\tilde F$ of $F$ parameterized by $\nu$.
\label{prop:upper}
\end{prop}
\begin{rem}
A naive approach is to use classic integration schemes to define an explicit expression for $\tilde F$.
Namely, using an explicit Euler scheme for the force equation \eqref{eq:dotF}
adapted to the partition $(t_{i+j/p})_{ij}$
gives:
\[
\tilde F(t_{i+(j+1)/p}) =
\tilde F(t_{i+j/p})\, c_{i,j} + A\,d_{i,j},
\]
where $c_{ij}=(1 - h_{i,j}\,m_2(t_{i+j/p}))$,
$h_{i,j}=t_{i+(j+1)/p}-t_{i+j/p}$
and $d_{ij}=m_1(t_{i+j/p})$ for $i=0,\dots,n$
and $j=0,\dots,p-1$.
We deduce the following explicit expression for $\tilde F(t_{k_t+j_t/p}),\ k_t=0,\dots,n$, $j_t=0,\dots p-1$:
\begin{align*}
\tilde F(t_{k_t+j_t/p})/A &= \sum_{j=0}^{j_t-1}
h_{k_t,j}\, d_{k_t,j} \prod_{j'=j+1}^{j_t-1}\! c_{k_t,j'} \addtocounter{equation}{1}\tag{\theequation}
\\
&+\sum_{i=0}^{k_t-1}\sum_{j=0}^{p-1}
h_{i,j}\, d_{i,j}\left(
\prod_{j'=0}^{p-1}\prod_{i'=i+1}^{k_t-1}\! c_{i',j'}
\prod_{j'=j+1}^{p-1}\! c_{i,j'}
\prod_{j'=0}^{j_t-1}\! c_{k_t,j'}\right).
\label{eq:Ftilde}
\end{align*}
However, such method is not adapted for the design of our electrostimualtor (see Section \ref{sec6}).
Indeed, it does not exploit the structure of the Hill functions $m_1$ and
$m_2$ and yields worse results -- in terms of time complexity and approximation error --
compared to the approximation \eqref{eq:Ftilde}.
\end{rem}
{\bf Error estimate.}
We give a bound on the error between the approximation
$\tilde F$ and $F$ in the case where $\tilde {m_2}_{\mid [t_k,t_{k+1}]}$ is the average of $m_2$ on
$[t_k,t_{k+1}]$.
\begin{prop}
Consider the case where $0\le \tilde m_1(t)\le 1$ and $\tilde m_2$ is equal to the average
of $m_2$ on $[t_j,t_{j+1}]$, $j=0,\dots,n$.
Assume moreover
that each restriction on $[t_j,t_{j+1}]$, $j=0,\dots,n$ of $m_1$ (resp. $m_2$) is concave (resp. convex).
Then, the error between the force $F$ and its
approximation $\tilde F$ defined by \eqref{eq:Ftilde} satisfies for
$k=0,\dots,n$:\\
$\displaystyle
|F(t_k) - \tilde F(t_k)|/A \le
\int_0^{t_k} |m_1(s) - \tilde m_1(s)|\, \mathrm{d} s
+
t_k \int_0^{t_k} |m_2(s)-\tilde m_2(s)|\, \mathrm{d} s.
$ \label{prop:error}
\end{prop}
\begin{proof}
For $t_k>s$, we have:
\begin{equation*}
\begin{aligned}
|F(t_k)-\tilde F(t_k)|/A_0 &=
\left|\int_0^{t_k} M(t_k)M^{-1}(s) m_1(s) - \tilde M(t_k)\tilde M^{-1}(s) \tilde m_1(s)\, \mathrm{d} s\right|
\\
&\le
\int_0^{t_k} M(t_k)M^{-1}(s) |m_1(s) - \tilde m_1(s)|\, \mathrm{d} s
+ \left|\int_0^{t_k} \tilde m_1(s) M(t_k)M^{-1}(s) - \tilde M(t_k) \tilde M^{-1}(s)|\, \mathrm{d} s\right|
\\
&\le
\int_0^{t_k} |m_1(s) - \tilde m_1(s)|\, \mathrm{d} s +
\left|\int_0^{t_k} M(t_k)M^{-1}(s)- \tilde M(t_k) \tilde M^{-1}(s) \, \mathrm{d} s\right|.
\\
&= \int_0^{t_k} |m_1(s) - \tilde m_1(s)|\, \mathrm{d} s +
\left|\int_0^{t_k} \exp\left( -\int_{s}^{t_k} m_2(u)\, \mathrm{d} u \right)
- \exp\left( -\int_{s}^{t_k} \tilde m_2(u)\, \mathrm{d} u \right)\,\mathrm{d} s\right|
\\
&= \int_{0}^{t_{k}} |m_1(s) - \tilde m_1(s)|\, \mathrm{d} s +
\left|\sum_{i=0}^{k-1} \int_{t_i}^{t_{i+1}} \exp\left( -\int_{s}^{t_k} m_2(u)\, \mathrm{d} u \right)
- \exp\left( -\int_{s}^{t_k} \tilde m_2(u)\, \mathrm{d} u \right)\,\mathrm{d} s\right|.
\end{aligned}
\label{eq:proof-error}
\end{equation*}
Recall the function $m_2$ is decreasing on $[t_i,s_i]$ and increasing on $[s_i,t_{i+1}]$
where $s_i$ is the unique maximum of $c_N$ on $[t_i,t_{i+1}]$.
Define $\displaystyle \xi(s)\coloneqq \tilde m_2(s)- m_2(s)$.
Since $m_2$ is convex on $[t_i,t_{i+1}]$, we have three cases:
\begin{itemize}
\item[(i)] $\displaystyle \int_{t_i}^{s}\xi(u)\, \mathrm{d} u\le 0$ for $s\in [t_i,t_{i+1}]$.
We have, for $i=0,\dots,k-1$,
\begin{equation}
\begin{aligned}
\Bigg|\int_{t_i}^{t_{i+1}} &\exp\left( -\int_{s}^{t_k} m_2(u)\, \mathrm{d} u \right)
- \exp\left( -\int_{s}^{t_k} \tilde m_2(u)\, \mathrm{d} u \right)\,\mathrm{d} s\Bigg|
\\
&=
\left| \int_{t_i}^{t_{i+1}}
\exp\left( \int_{t_{k}}^s m_2(u)\, \mathrm{d} u \right) \left(1-
\exp\left( \int_{t_{k}}^s \xi(u)\, \mathrm{d} u \right)\right)\,
\mathrm{d} s \right|
\\
&\le
\int_{t_i}^{t_{i+1}}
\left|1-
\exp\left( \int_{t_{i}}^s \xi(u)\, \mathrm{d} u \right)\,
\right|\mathrm{d} s , \text{ (since } \int_{t_j}^{t_{j+1}} \xi(u)\, \mathrm{d} u =0)
\\
&\le
\int_{t_i}^{t_{i+1}}
\int_{t_i}^s -\xi(u)\, \mathrm{d} u\,
\mathrm{d} s
\\
&\le
(t_{i+1}-t_i)\
\int_{t_i}^{t_{i+1}} |\xi(u)|\, \mathrm{d} u.
\end{aligned}
\end{equation}
\item[(ii)] $\displaystyle \int_{t_i}^{s}\xi(u)\, \mathrm{d} u\ge 0$ for $s\in [t_i,t_{i+1}]$.
We obtain the same inequality as in the case (i) by replacing $\xi$ by $-\xi$.
\item[(iii)] There exists an unique $\theta_i \in [t_i,t_{i+1}]$ such that
$\displaystyle \int_{t_i}^{s}\xi(u)\, \mathrm{d} u\le 0$ for $s\in [t_i,\theta_{i}]$ and
$\ge 0$ for $s\in [\theta_{i},t_{i+1}]$.
Write
\begin{equation}
\begin{aligned}
\Bigg|\int_{t_i}^{t_{i+1}} &\exp\left( -\int_{s}^{t_k} m_2(u)\, \mathrm{d} u \right)
- \exp\left( -\int_{s}^{t_k} \tilde m_2(u)\, \mathrm{d} u \right)\,\mathrm{d} s\Bigg|
\\
&\le
\Bigg|\int_{t_i}^{\theta_i} \exp\left( -\int_{s}^{t_k} m_2(u)\, \mathrm{d} u \right)
- \exp\left( -\int_{s}^{t_k} \tilde m_2(u)\, \mathrm{d} u \right)\,\mathrm{d} s\Bigg|
\\
&\qquad +
\Bigg|\int_{\theta_i}^{t_{i+1}} \exp\left( -\int_{s}^{t_k} m_2(u)\, \mathrm{d} u \right)
- \exp\left( -\int_{s}^{t_k} \tilde m_2(u)\, \mathrm{d} u \right)\,\mathrm{d} s\Bigg|
\end{aligned}\label{eq:case3-proof}
\end{equation}
We have: $\int_{\theta_i}^{s}\xi(u)\, \mathrm{d} u\le 0$ for $s\in [t_i,\theta_i]$
and $\ge 0$ for $s\in [\theta_i,t_{i+1}]$.
As in the case (i), the first integral in the right hand side of \eqref{eq:case3-proof}
is bounded by
$(\theta_i-t_i)\ \int_{t_i}^{\theta_i} |\xi(u)|\, \mathrm{d} u$.
Likewise, the second integral in the right hand side of \eqref{eq:case3-proof} is bounded
by $(t_{i+1}-\theta_i)\ \int_{\theta_i}^{t_{i+1}} |\xi(u)|\, \mathrm{d} u$.
\end{itemize}
This concludes the proof.
\end{proof}
\section{Numerical solution to optimization problems} \label{sec5}
\subsection{Functional specification for the computation of a pulses train}
The aim is to compute a local minimum $\sigma^*=(\eta_0^*,\dots,\eta_n^*,t_1^*,\dots,t_n^*,T)\in \mathbb{R}_+^{2n+2}$
of a cost function denoted as $\Theta$. The free final time $T$ adjusts automatically the optimal
frequency of the pulses train.
The functional specification of the electrostimulator imposes efficient computation of this minimum
(real time computation) and
this prevents us (at least when $\Theta$ involves the force) from
using direct or indirect methods such as those presented in \cite{Bakir2020},
mainly because these methods are based on
a numerical scheme to approximate the variable $F$.
\subsection{Finite dimensional optimization methods}
\label{sec5p2}
We recall basic facts about finite dimensional optimization, see \cite{boyd2004} for details,
to emphasize that an optimal sampled-data control
problem can be viewed as an instance of such optimization problem.
The optimization problems, associated to the optimal sampled-data control problems
{\bf OCP1} and {\bf OCP2} presented in
Section \ref{secOCP}, can be written in the form:
\begin{equation} \label{pb:FDP}
\begin{aligned}
\begin{array}{ll}
\underset{\sigma}{\min} \quad& \Theta(\sigma)
\\
& \Im(\sigma) \le 0,
\end{array}
\end{aligned}
\end{equation}
where $\Im(\sigma) = (\Xi_1(\sigma),\dots,\Xi_{3n+5}(\sigma))$ is the vector of constraints defined by:
\begin{equation*}
\begin{array}{ll}
\Xi_i(\sigma^*)=t_{i-1}^*-t_{i}^*+I_{\min}, \ i=1,\dots n,
\\
\Xi_{n+1}(\sigma^*)=t_n^*-T, \\
\Xi_{n+2+i}(\sigma^*)=-\eta_{i}^*, \ i=0,\dots n+1,
\\
\Xi_{2n+4+i}(\sigma^*)=\eta_{i}^*-1, \ i=0,\dots n+1.
\end{array}
\end{equation*}
The cost
$
\Theta: \sigma \mapsto \Theta(\sigma)
$ related to the endurance or the force strengthening program
is smooth with respect to $\sigma$.\\
Consider the Lagrangian defined for all $(\sigma,\mu)\in \mathbb{R}^{2n+1}\times \mathbb{R}^{3n+5}_+$ by:
\[
\mathcal{L}(\sigma,\mu) \coloneqq \Theta(\sigma) + \mu \cdot \Im(\sigma).
\]
The problem \eqref{pb:FDP} is equivalent to the primal problem
\[\inf_{\sigma \in \mathbb{R}^{2n+1}} \sup_{\mu\in \mathbb{R_+}^{3n+5}}
\mathcal{L}(\sigma,\mu)\] and the first order necessary optimality conditions
for $\sigma^*$ to be a local minimizer,
assuming the vectors $\Xi_i'(\sigma^*),\ i\in \{i,\, \Xi_i(\sigma^*)=0\}$ to be linearly independent,
state that there exists a Lagrange multiplier $\lambda\in \mathbb{R}^{3n+5}$ such that
\begin{align*}
&\nabla_\sigma\Theta(\sigma^*) + \lambda \cdot \Im (\sigma^*) = 0,\, \lambda \cdot \Im(\sigma^*) = 0 \\
&\lambda_i\ge 0,\ \Xi_i(\sigma^*)\le 0, \ i=1,\dots,3n+5.
\end{align*}
We usually do not solve directly these optimality conditions to compute
an optimal pair $(\sigma^*,\lambda^*)$, but a relaxation of these conditions can lead to
efficient algorithm, namely the primal-dual interior point method \cite{boyd2004}.
\subsection{Force optimization}
We consider the problems the endurance and force strengthening optimization problems {\bf OCP1} and
{\bf OCP2}.
For each problem, we give the approximation
$\tilde \Theta$ of the cost functions $\Theta$
based on the approximation $\tilde F$ of the variable $F$ described in section \ref{sec4}.
We solve the associated problem \eqref{pb:FDP} -- where $\Theta$ is replaced by $\tilde \Theta$ --
using an interior point method
on a standard computer\footnote{4 Intel@Core$^{TM}$ i5 CPU @ 2.4Ghz}.
Note that $\tilde \Theta$ may consist of million of bytes, for that reason it is crucial to use
an approximation of the gradient of $\tilde \Theta$ with respect to $t_{i},\ i=1,\dots,n$, computed
via finite differences (vs formal computation).
We initialize the pulses train to a regular partition of $[0,1]$ and the initial amplitudes being equal to $1$.
We consider the force approximation $\tilde F$ defined by \eqref{eq:Ftilde} taking
the piecewise affine functions $\tilde m_1, \tilde m_2$ to be equal on
$[t_{k},t_{k+1}]$, $k=0,\dots,n$ to:
\begin{equation*}
\begin{aligned}
&\tilde m_1(t) =
\left\{
\begin{array}{ll}
m_1(t_{k+1/2}) &\text{ if } t\in [t_k,t_{k+1/2}] \\
a_{1j,k}\, (t-t_{k+1}) + b_{1j,k}, &\text{ if } t\in [t_{k+1/2},t_{k+1}]
\end{array}
\right.,
\\
&\tilde m_2(t) =
\left\{
\begin{array}{ll}
\frac{m_2(t_{k})+m_2(t_{k+1/2})}{2} &\text{ if } t\in [t_k,t_{k+1/2}] \\
\frac{m_2(t_{k+1/2})+m_2(t_{k+1})}{2} , &\text{ if } t\in [t_{k+1/2},t_{k+1}]
\end{array}
\right.,
\end{aligned}
\end{equation*}
where $t_{k+1/2} =\underset{u\in [t_k,t_{k+1}]}{\text{argmax}} c_N(u)$,
$a_{1j,k} =(m_1(t_{k+1}) - m_1(t_{k+1/2}))/(t_{k+1}-t_{k+1/2})$ and
$b_{1j,k} = m_1(t_{k+1})$.
\subsubsection{Problem OCP1: $\displaystyle {\Theta(\sigma) \coloneqq -F(T)}$.}
\paragraph*{Approximated cost.}
The objective function $\Theta(\sigma)=-F(T)$ is approximated by the function
$\tilde \Theta(\sigma)=-\tilde F(T)$.
The optimization variables consist in the impulse times while the amplitudes are fixed to $1$.
{\it Numerical result: }
The optimal solution $\sigma^*$, the force response $F$ and its approximation $\tilde F$ are depicted in
Fig.\ref{fig:maxFT20}.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.75\columnwidth]{maxFT7.pdf}
\caption{The dashed curve is the time evolution of $F$
associated to the optimal solution $\sigma^*=(t_1^*,\dots,t_n^*,T)$ of
$\max_{\sigma}\, \tilde F(T)$ ($T$ free)
(see \eqref{eq:Ftilde} for the definition of $\tilde F$)
under the constraints $\Xi_i\le 0,\ i=1,\dots, n+1$ (see \eqref{pb:FDP}).
The continuous curve is the response $t\mapsto F(t)$ to $\sigma^*$.
Values of the constants are $\tau_c=20$ms, $n=7$, $I_{\min}=20$ms. \label{fig:maxFT20}}
\end{figure}
\subsubsection{Problem OCP2: $\displaystyle {\Theta(\sigma) \coloneqq \int_0^T |F(t)-F_{ref}|^2\, \mathrm{d} t}$.}
The cost $\displaystyle \Theta(\sigma) = \int_0^T |F(s)-F_{ref}|^2\, \mathrm{d} s$ is approximated by:
\[\displaystyle \tilde \Theta(\sigma) = \sum_{k=0}^n \left( \tilde F(t_{k+1}) -F_{ref}\right)^2 (t_{k+1}-t_k),\]
where the functions $\tilde m_1$ and
$\tilde m_2$ are replaced by $\overset{\asymp}{m_1}(t;0.95)$ and $\overset{\asymp}{m_2}(t;0.95)$
respectively (see Proposition \ref{prop:upper} for their definition).
{\it Numerical result: }
The optimal solution $\sigma^*$, the force response $F$ and its approximation $\tilde F$ are depicted in
Fig.\ref{fig:minFref5_0p1}.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.75\columnwidth]{minFref_cref_0p1.pdf}
\caption{The dashed curve is associated to the optimal solution
$\sigma^*=(\eta_0^*,\dots,\eta_n^*,t_1^*,\dots,t_n^*,T)$ of
$\min_{\sigma}\, \sum_{k=0}^n \left( \tilde F(t_{k+1}) - F_{ref} \right)^2 (t_{k+1}-t_k)$ ($T=t_{n+1}$ is free),
where $\tilde F$ is the upper approximation of $F$ as described from Proposition \ref{prop:upper}
under the constraints $\Xi_i\le 0,\ i=1,\dots, 3n+5$ (see \eqref{pb:FDP}).
The continuous curve is the response $t\mapsto F(t)$ to $\sigma^*$.
Values of the constants are $\tau_c=20$ms, $n=5$, $I_{\min}=20$ms, $\nu=0.95$ and $F_{ref}=0.1$kN.
\label{fig:minFref5_0p1}}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.75\columnwidth]{minFref_cref_0p2.pdf}
\caption{The dashed curve is associated to the optimal solution
$\sigma^*=(\eta_0^*,\dots,\eta_n^*,t_1^*,\dots,t_n^*)$ of
$\min_{\sigma}\, \sum_{k=0}^n \left( \tilde F(t_{k+1}) - F_{ref} \right)^2 (t_{k+1}-t_k)$ ($T=t_{n+1}$ is free),
where $\tilde F$ is the approximated force given by \eqref{eq:Ftilde},
under the constraints $\Xi_i\le 0,\ i=1,\dots, 3n+5$ (see \eqref{pb:FDP}).
The continuous curve is the response $t\mapsto F(t)$ to $\sigma^*$.
Values of the constants are $\tau_c=20$ms, $n=7$, $I_m=20$ms and $F_{ref}=0.2$kN.
\label{fig:minFref7_0p2}}
\end{figure}
\subsection{$Ca^{2+}$ concentration optimization}
{\it Uniqueness of the optimal solution:}
Fix $\eta_i=1, i=0,\dots,n$.
The cost function $\displaystyle \Theta(\sigma) = \int_0^T (c_N(s)-c_{ref})^2\, \mathrm{d} s$ is smooth
with respect to $t_1$ and not convex on $[0,T]$.
In Fig.\ref{fig:cout}, we plot for $n=1$ the objective function $\Theta(t_1)$ and for $T$ fixed at some
specific values.
\begin{figure}
\centering
\includegraphics[width=0.75\columnwidth]{cost_n1.pdf}
\caption{Objective function $\displaystyle \Theta(\sigma) = \int_0^T (c_N(s)-c_{ref})^2\, \mathrm{d} s$,
where $T$ is assigned to specific values.
Constants for these simulations are
$\tau_c=20$ms, $n=1$, $I_m=20$ms.\label{fig:cout} }
\end{figure}
The well-posedness of this optimization problem for any number $n$ of impulsions times can be shown by
inductive reasoning.
\subsubsection{Cost: ${\Theta(\sigma) = -c_N(T)}$.}
\paragraph*{True cost.}
We have an explicit expression for $c_N$, this problem can be easily solved numerically.
We consider the finite dimensional optimization problem \ref{pb:FDP} where the cost is
\[
\Theta(\sigma) =
\sum_{i=0}^n R_i (T-t_i)\,e^{-\frac{T-t_i}{\tau_c}}.
\]
(Note that the amplitudes are fixed to $1$).
{\it Numerical result:}
Fig. \ref{fig:cnT20} represents the time evolution of $c_N$ associated to (locally) optimal
impulse times.
\begin{figure}
\centering
\includegraphics[width=0.75\columnwidth]{Cmax_Tfree_7.pdf}
\caption{Time evolution of $c_N$ associated to the optimal sampling times $\sigma^*$ for
$\max_{\sigma}\, c_N(T)$ ($T$ free) under the constraints $\Xi_i\le 0,\ i=1,\dots, n+1$
(see \eqref{pb:FDP}).
Values of the constants are $\tau_c=20$ms, $n=7$, $I_m=20$ms.\label{fig:cnT20} }
\end{figure}
\subsubsection{Cost: $\displaystyle {\Theta(\sigma) = \int_{0}^T |c_N(t)-c_{ref}|^2\, \mathrm{d} t}$, $T$ free.}
In this case, we approximate $\Theta$ by
\begin{equation}
\begin{aligned}
\tilde \Theta(\sigma) = \sum_{i=0}^n (\bar c_N^i - c_{N,ref})^2 (t_{i+1}-t_i),
\end{aligned}
\label{eq:approxThetaCnref}
\end{equation}
where the amplitudes and $T$ are free and $\bar c_N$ is given by Proposition \ref{prop:cnk}.
{\it Numerical result:}
The optimal solution $\sigma^*$ and its response $c_N$ are depicted in
Fig. \ref{fig:cnref1a}.
\begin{figure}
\centering
\includegraphics[width=0.75\columnwidth]{Cn_Tfree_7_Cref_0p5.pdf}
\caption{The dashed curve is the time evolution of $c_N$
associated to the optimal solution
$\sigma^*=(\eta_0^*,\dots,\eta_n^*,t_1^*,\dots,t_n^*)$ of
$\min_{\sigma}\, \sum_{i=0}^n (\bar c_N^i - c_{N,ref})^2 (t_{i+1}-t_i)$
($T$ free), where $\bar c_N$ is defined in Proposition \ref{prop:cnk},
under the constraints $\Xi_i\le 0,\ i=1,\dots, 3n+5$ (see \eqref{pb:FDP}).
The continuous curve is the response $t\mapsto c_N(t)$ to $\sigma^*$.
Constants for these simulations are
$\tau_c=20$ms, $n=7$, $I_m=20$ms and $c_{ref}=0.5$.\label{fig:cnref1a} }
\end{figure}
\section{Isometric case: design of a smart muscular electrostimulator} \label{sec6}
\label{sec6p1}
In this section we apply our study in the isometric case associated to the conception of a smart
electrostimulator. In this case, the task is to assign a constant reference force, but the general frame
is suboptimal motion planning, see \cite{hirschorn1987,hirschorn1988} for the theoretical foundations.
Advanced commercial muscular electrostimulator for training or reeducation purposes are based on the following.
First of all, the user defines a program training.
Basically endurance program with low frequency sequences of trains (with constant interpulse) or force
strengthening with high frequency trains. The program is a sequence of trains or rest periods.
Before starting, the muscle is scanned to determine the parameters.
A smart electrostimulator based on our study aims to design automatically such sequence,
each program being translated into an optimization problem.
Besides, in our framework, one can used VFT (Variable Frequency Trains) vs CFT (Constant Frequency Trains)
in the standard case to complete the tuning of the amplitude.
One needs the following proposition associated to the endurance program presented in Fig.\ref{fig:schema}
to illustrate the smart electrostimulator conception.
\begin{prop} Consider the endurance program, where a reference force $F_{ref}$ is given.
Plugging such $F_{ref}$ in $\dot{F}=0$ leads to solve the equation:
$A\tau_2m_1^2+A\tau_1m_1-F_{ref}=0$, which has an unique positive root $m_1^+$ giving
the reference concentration $c_{N,ref}$.
This root is stable and leads to design an optimized pulses train solving the $L^2$-optimization problem with cost:
$\displaystyle \int_0^T |c_N(t)-c_{N,ref}|^2\, \mathrm{d} t$.
\end{prop}
\begin{pf}
Note that the mapping $m_1$: $c_N \mapsto m_1(c_N)$ is one-to-one and $m_1$ can be taken as an accessory control
in place of $c_N$. Solving in $m_1$ the equation $\dot{F}=0$ leads to real roots denoted respectively $m_1^+>0$
and $m_1^-<0$. Taking $m_1^+$, stability is granted since $\lambda=-m_2(c_N^+)$ is negative,
where $c_N^+$ is given by $m_1(c_N^+)=m_1^+$. $\hfill \qed$
\end{pf}
\begin{rem}
The optimization problem $\min_{\sigma} \ \int_0^T |c_N(t)-c_{N,ref}|^2\, \mathrm{d} t$
can be efficiently solved using
the piecewise constant approximation of $c_N$ of Proposition \ref{prop:cnk}.
Indeed, we have:
\begin{equation*}
\int_0^T |c_N(t)-c_{N,ref}|^2\, \mathrm{d} t
\approx
\sum_{i=0}^n (\bar c_N^i - c_{N,ref})^2 (t_{i+1}-t_i).
\end{equation*}
\end{rem}
\section{Conclusion}
In this short article, we have mainly presented a finite dimensional approximation of the muscular force response
to FES-input exploiting the mathematical structure of the model.
The construction is based on the Ding et al. model but can be adapted to deal with the different models
discussed in \cite{wilson2011}.
We have presented one application of our study related to motion planning in the isometric case in view
to design a smart electrostimulator, which is an ongoing industrial project.
Our approximation can be used to parameters estimation \cite{wilson2011,Stein2013} and to design
MPC-optimized sampled-data control schemes, applying standard algorithms \cite{Richalet1993,Boyd2010}
to this situation.\\
Another application of our study is to track in the non-isometric case a path in the joint angle variable
and this will be developed in a forthcoming article.
\begin{table*}
\renewcommand{\arraystretch}{0.9}
\centering
\caption{List of variables and values of the constant parameters in the Ding et al.\ model}\label{tb:params}
\setlength{\tabcolsep}{2pt}
\begin{tabular}{llll}
Symbol & Unit & Value & Description \\
\hline
$C_{N}$&~---~ &~---~~&Normalized amount of $Ca^{2+}$-troponin complex\\
$F$& $\text{kN}$ &~---~&Force generated by muscle\\
$t_{i}$&$s$&~---~&Time of the $i^{th}$ pulse\\
$n$&---&~---~&Total number of the pulses before time $t$\\
$i$&---&~---~&Stimulation pulse index\\
$\tau_{c}$&$s$&$0.02$&Time constant that commands the rise and the decay of $C_{N}$\\
$\bar{R}$&---&$1.143$&Term of the enhancement in $C_{N}$ from successive stimuli\\
$A$&$\text{kN}\cdot s^{-1}$&~---~&Scaling factor for the force and the shortening velocity of muscle \\
$\tau_{1}$&$s$&$50.95$ ~&Force decline time constant when strongly bound cross-bridges absent\\
$\tau_{2}$&$s$&$0.1244$&Force decline time constant due to friction between actin and myosin\\
$K_{m}$&---&~---~&Sensitivity of strongly bound cross-bridges to $C_{N}$\\
$A_{\text{rest}}$&$\text{kN}\cdot s^{-1}$&$3.009$&Value of the parameter $A$ when muscle is not fatigued\\
$\alpha_{A}$&$s^{-2}$&$-4.0~10^{-1}$&Coefficient for the force-model parameter $A$ in the fatigue model\\
$\tau_{fat}$&$s$&$127$&Time constant controlling the recovery of $A$\\
\hline
\end{tabular}
\end{table*}
\begin{figure*}
\centering
\includegraphics[width=1.05\linewidth]{schemav5.pdf}
\caption{Endurance program for a smart electrostimulator.}
\label{fig:schema}
\end{figure*}
\FloatBarrier
{\footnotesize
|
\section*{References}}
\usepackage{amssymb}
\usepackage{amsmath,graphicx, xcolor,amssymb}
\usepackage[font=small,labelfont=bf,tableposition=top]{caption}
\usepackage{graphicx}
\usepackage{epstopdf}
\usepackage{mathabx}
\usepackage{mathrsfs}
\usepackage{ragged2e}
\usepackage[mathscr]{euscript}
\usepackage{multirow}
\usepackage{subfigure}
\usepackage{enumerate}
\usepackage{mathrsfs}
\usepackage{color,soul}
\usepackage{todonotes}
\usepackage{verbatim}
\usepackage{float}
\usepackage{framed}
\usepackage{algorithm, algorithmic}
\usepackage{lineno}
\begin{document}
\begin{frontmatter}
\title{Interface spaces based on physics for multiscale mixed methods applied to flows in fractured-like porous media}
\author[mymainaddress]{Franciane F. Rocha}
\ead{<EMAIL>}
\author[mymainaddress]{Fabricio S. Sousa\corref{mycorrespondingauthor}}
\cortext[mycorrespondingauthor]{Corresponding author}
\ead{<EMAIL>}
\author[mymainaddress]{Roberto F. Ausas}
\ead{<EMAIL>}
\author[mysecondaryaddress]{Felipe Pereira}
\ead{<EMAIL>}
\address[mymainaddress]{Instituto de Ci\^encias Matem\'aticas e de Computa\c c\~ao, Universidade de S\~ao Paulo,\\ Av. Trabalhador S\~ao-carlense, 400, 13566-590, S\~ao Carlos, SP, Brazil}
\address[mysecondaryaddress]{Department of Mathematical Sciences, The University of Texas at Dallas,\\ 800 W. Campbell Road, Richardson, TX 75080-3021, USA}
\author[mymainaddress]{Gustavo C. Buscaglia}
\ead{<EMAIL>}
\begin{abstract}
It is well known that domain-decomposition-based multiscale mixed methods rely on interface spaces, defined on the skeleton of the decomposition, to connect the solution among the non-overlapping subdomains. Usual spaces, such as polynomial-based ones, cannot properly represent high-contrast channelized features such as fractures (high permeability) and barriers (low permeability) for flows in heterogeneous porous media. We propose here new interface spaces, which are based on physics, to deal with permeability fields in the simultaneous presence of fractures and barriers, accommodated respectively, by the pressure and flux spaces.
Existing multiscale methods based on mixed formulations can take advantage of the proposed interface spaces, however, in order to present and test our results, we use the newly developed Multiscale Robin Coupled Method (MRCM) [Guiraldello, et al., \textit{J. Comput. Phys.}, 355 (2018) pp. 1-21], which generalizes most well-known multiscale mixed methods, and allows for the independent choice of the pressure and flux interface spaces.
An adaptive version of the MRCM [Rocha, et al., \textit{J. Comput. Phys.}, 409 (2020), 109316] is considered that automatically selects the physics-based pressure space for fractured structures and the physics-based flux space for regions with barriers, resulting in a procedure with unprecedented accuracy. The features of the proposed approach are investigated through several numerical simulations of single-phase and two-phase flows, in different heterogeneous porous media. The adaptive MRCM combined with the interface spaces based on physics provides promising results for challenging problems with the simultaneous presence of fractures and barriers.
\end{abstract}
\begin{keyword}
Multiscale Robin coupled method, multiscale mixed methods, interface spaces, two-phase flows, high-contrast porous media
\end{keyword}
\end{frontmatter}
\section{Introduction}
Simulations of petroleum reservoirs deal with highly heterogeneous permeability fields with multiple scales and high contrast. A representation of the solution that captures the heterogeneity requires several billion cells, making the numerical simulations extremely expensive \cite{chen2006computational}. The so-called {\em multiscale methods} have been introduced to exploit the multiscale structure of the problem and provide efficient approximations at a reduced computational cost \cite{kippe}.
The multiscale literature is vast and contemplates numerical implementations using techniques such as the finite volume method \cite{jenny2003multi,jenny2005adaptive}, the primal finite element method \cite{HouMulti, aarnes2002multiscale, efendiev2013generalized}, and several mixed finite element methods \cite{chen_hou, aarnes, arbogast, valentin, pereira, chung2015mixed}.
The multiscale approaches solve the problem on a coarse decomposition of the domain, incorporating the fine-grid information through local basis functions. The accuracy of the multiscale methods is related to the calculation of the basis functions. If the heterogeneities are not well represented by the basis, inaccurate solutions are obtained.
Spaces that are polynomial on the interfaces of the domain decomposition work well for smooth or Gaussian permeability fields, but their performance for high contrast, channelized ones is not satisfactory \cite{chung2014adaptive,chung2016enriched,cortinovis2014iterative,guiraldello2019interface}.
To remedy this, informed spaces obtained from sets of snapshots by algebraic dimensionality reduction were considered in \cite{efendiev2013generalized,chung2016enriched,guiraldello2019interface}. Another approach is to define the interface space through eigensolutions of local partial-differential problems \cite{madureira2017hybrid}. These approaches can also be coupled to a-posteriori error estimators \cite{chung2014adaptive, chung2017online, chung2018online}.
Other authors have looked for approaches more directly based on the geometry of the heterogeneities. In \cite{aarnes2006hierarchical, moyner2016multiscale, klemetsdal2019accelerating}, for example, the authors consider polynomial bases but adapt the grid to the geological properties, while in \cite{peszynska2002mortar,cortinovis2014iterative,cortinovis2017zonal} local enrichment functions were added on high-permeability regions. These geometrical strategies are well suited for permeability fields containing either channels or barriers, but not both. In this contribution, we propose a strategy to deal with the simultaneous presence of channels and barriers, since such situation is not unfrequent in reservoir simulation \cite{hosseinimehr2019dynamic}.
For this purpose, we take advantage of the Multiscale Robin Coupled Method (MRCM, \cite{guiraldello2018multiscale}), which allows for the independent choice of the pressure and flux interface spaces. In fact, the pressure space is designed so as to accommodate channels and the flux space to accommodate barriers, and the adaptive version of the MRCM \cite{bifasico} is used to automatically select the appropriate parameters at each location.
Our intention is to use high-definition volumetric grids that capture the large scale features of the fracture network, especially when they are relatively large as happens in fractured karst reservoirs \cite{baomin2009classification, popov2009multiphysics, huang2013efficient,lopes2019new}. Typically, fractures are handled with separate discrete models that represent the fractures as lower-dimensional objects so as to incorporate sub-grid resolution \cite{martin2005modeling, reichenberger2006mixed, formaggia2014reduced, schwenck2015dimensionally, berkowitz2002characterizing, flemisch2018benchmarks}. A popular approach is the Discrete Fracture Model (DFM) which uses unstructured grids to place fractures at the interface between matrix cells \cite{kim2000finite, karimi2003efficient, hoteit2008efficient}. Other approaches are the Embedded DFM \cite{ctene2016algebraic,chai2018efficient}, the multi-continuum model \cite{chung2017coupling, wang2020generalized} and the hierarchical fracture models \cite{hajibeygi2011hierarchical, li2008efficient, efendiev2015hierarchical}.
Since it is possible to combine discrete models with multiscale methods \cite{bosma2017multiscale, zhang2017multiscale, devloo2019multiscale, xia2018enriched}, the final goal of the improved MRCM proposed here is to allow a unified treatment of fractured karst reservoirs in which the modeling of the fractures is shared, depending on the fracture's size, between the volumetric grid and the discrete models. For this reason, we consider here permeability fields containing multiple narrow and relatively straight features (channels, barriers) that mimic the largest structures of a fractured porous medium, and refer to them as ``fractured-like'' fields. The accuracy and efficiency of the proposed method are investigated through several numerical simulations of fractured-like reservoirs both in single-phase and two-phase (oil-water) situations.
The paper is organized as follows. In Section \ref{Model problem} we introduce the model problem and numerical approximation. The MRCM is briefly recalled in Section \ref{The multiscale Robin coupled method} followed by Section \ref{physics-basead interface spaces}, where the strategies for building the physics-based interface spaces are presented along with some numerical experiments. In section \ref{aMRCM-PBS} we explain our proposed combination of the physics-based spaces with the adaptive MRCM.
Numerical experiments for single and two-phase flows are presented in Section \ref{results}. Finally, the paper is concluded in Section \ref{conclusions}.
\section{Model problem}\label{Model problem}
We consider a high-contrast heterogeneous oil reservoir through which an immiscible and incompressible single-phase or multiphase flow takes place. The heterogeneity of the medium is represented by a space-dependent permeability coefficient in the elliptic Darcy model. This coefficient exhibits variations of many orders of magnitude over short distances. For simplicity, the capillary pressure and gravity effects are not considered.
\subsection{Single-phase model}
The unknowns considered in the single-phase model are the Darcy velocity $\mathbf{u}(\mathbf x, t)$ and the fluid pressure $p(\mathbf x, t)$.
The governing equations are the Darcy's law with a statement of conservation of mass, given by
\begin{equation}\label{Darcy0}
\begin{array}{rll}
\mathbf{u}&=-K(\mathbf{x})\nabla p &\mbox{in}\ \Omega \\
\nabla \cdot \mathbf{u}&=q &\mbox{in}\ \Omega \\
p &= g &\mbox{on}\ \partial\Omega_{p}\\
\mathbf{u} \cdot \mathbf{n}&= z &\mbox{on}\ \partial\Omega_{u}
\end{array
\end{equation}
where $\Omega\subset\mathbb{R}^2 $ is the computational domain; $K(\mathbf{x})$ is the symmetric, uniformly positive definite absolute permeability tensor; $q = q(\mathbf x, t)$ is a source term; $g = g(\mathbf x, t)$ is the pressure condition specified at the boundary $\partial\Omega_{p}$ and $z = z(\mathbf x, t)$ is the normal velocity condition ($\mathbf{n}$ is the outward unit normal) specified at the boundary $\partial\Omega_{u}$.
\subsection{Two-phase model}
In the context of two-phase flows of water and oil (denoted by $w$ and $o$, respectively), we consider that the reservoir contains injection wells, from which water is injected to displace the trapped oil towards production wells. The oil and water saturations are the new unknowns. These saturations are related considering a fully saturated medium (the sum of both oil and water saturation is equal to one). Therefore the model considers only the water saturation $s(\mathbf x, t)$ in the transport problem. Here the elliptic problem (\ref{Darcy0}) is written as
\begin{equation}\label{Darcy}
\begin{array}{rll}
\mathbf{u}&=-\lambda(s)K(\mathbf{x})\nabla p &\mbox{in}\ \Omega \\
\nabla \cdot \mathbf{u}&=q &\mbox{in}\ \Omega \\
p &= g &\mbox{on}\ \partial\Omega_{p}\\
\mathbf{u} \cdot \mathbf{n}&= z &\mbox{on}\ \partial\Omega_{u}
\end{array
\end{equation}
and coupled with the hyperbolic conservation law for the transport of water saturation \cite{peaceman2000fundamentals,ewing1983mathematics}, which is given by
\begin{equation}\label{BL2D}
\begin{array}{rll}
\dfrac{\partial s}{\partial t} + \nabla \cdot \left(f(s)\mathbf{u}\right)& = 0 &\mbox{in}\ \Omega \\
s(\mathbf{x},t=0) &= s^0(\mathbf{x}) &\mbox{in}\ \Omega\\
s(\mathbf{x},t) &= \bar{s}(\mathbf{x},t) &\mbox{in}\ \partial\Omega^-
\end{array
\end{equation}
where $s^0$ is the initial condition for the saturation and $\bar{s}$ is the saturation at the inlet boundaries $\partial\Omega^-=\{\mathbf{x}\in\partial\Omega,\ \mathbf{u} \cdot \mathbf{n}<0\}$.
The total mobility $\lambda(s)=\lambda_w(s)+\lambda_o(s)$ and the fractional flow of water $f(s)$ are respectively given by
\begin{equation}
\lambda(s) =\dfrac{k_{rw}(s)}{\mu_w}+\dfrac{k_{ro}(s)}{\mu_o}\qquad \mbox{and} \qquad f(s) = \dfrac{\lambda_w(s)}{\lambda(s)},
\label{fluxos}
\end{equation}
where $k_{rj}(s)$ and $\mu_j$, $j \in \{w,o\} $, are respectively the relative permeability function and viscosity of phase $j$. For simplicity, the porosity is considered constant both in space and time and consequently, it can be easily scaled out by changing the time variable.
\subsection{Numerical approximation}
For computing single phase flows we solve directly the elliptic problem (\ref{Darcy0}) by the Multiscale Robin Coupled Method (MRCM). To simulate two-phase flows we solve the system (\ref{Darcy})-(\ref{BL2D}) using an operator splitting procedure, in which we solve (\ref{Darcy}) to compute $p(\mathbf x, t)$ and $\mathbf{u}(\mathbf x, t)$ and (\ref{BL2D}) for $s(\mathbf x, t)$, sequentially \cite{douglas1997numerical}.
We use the MRCM to solve the elliptic problem (\ref{Darcy}), and the upwind method \cite{leveque2002finite} for the transport of saturation (\ref{BL2D}), in which a CFL-type condition is enforced. The splitting procedure considers larger time steps for the elliptic problem compared to those used for the hyperbolic equation (\cite{douglas1997numerical}) improving the computational efficiency.
We consider the relation $\Delta t_p = C \Delta t_s$, where $\Delta t_s$ is the time step for the saturation equation (for simplicity assumed constant but in practice we allow for variable $\Delta t_s$), $\Delta t_p$ is the time step for the elliptic problem and $C$ is a positive integer.
The elliptic problem is solved at times $t_n = n\Delta t_p$, for $n=0,1,\dots$, where we compute $p^{n}(\mathbf x)$ and $\mathbf u^{n}(\mathbf x)$ using the water saturation $s^{n}(\mathbf x)$ at time $t=t_n$. The saturation equation is solved at intermediate times $t_{n,k} = t_n + k\Delta t_s$, for $k=1,2,\dots, C$, such that $t_n < t_{n,k} \leq t_{n+1}$.
For each saturation time step $t = t_{n,k}$ we set the velocity $\mathbf u^{n,k}$ to the following linear extrapolated value
\begin{equation}\label{extrapolation}
\mathbf u^{n,k}(\mathbf x, t) =
\left\{
\begin{array}{ll}
\mathbf u^0(\mathbf x), & \text{ if } 0 \leq t\leq t_1, \\
\dfrac{t-t_{n-1}}{\Delta t_p}\ \mathbf u^n(\mathbf x) -\dfrac{t-t_n}{\Delta t_p}\ \mathbf u^{n-1}(\mathbf x),
& \text{ if } t_n<t\leq t_{n+1},\\
\end{array} \right.
\end{equation}
where $\mathbf u^n (\mathbf x)$ is the velocity obtained from Equation (\ref{Darcy}) at time $t=t_n$, approximating $\mathbf u(\mathbf x, t_n)$ (see \cite{douglas1997numerical}).
\section{The multiscale Robin coupled method}\label{The multiscale Robin coupled method}
In this paper, we use the MRCM to approximate the Equation (\ref{Darcy0}) for single-phase flows and the Equation (\ref{Darcy}) for two-phase flows. For simplicity, we drop the time dependency and obtain the following elliptic problem
\begin{equation}\label{MRCM}
\begin{array}{rll}
\mathbf{u}&=-\kappa(\mathbf{x})\nabla p &\mbox{in}\ \Omega \\
\nabla \cdot \mathbf{u}&=q &\mbox{in}\ \Omega \\
p &= g &\mbox{on}\ \partial\Omega_{p}\\
\mathbf{u} \cdot \mathbf{n}&= z &\mbox{on}\ \partial\Omega_{u}
\end{array
\end{equation}
where $\kappa=K(\mathbf x)$ for single-phase and $\kappa = \lambda(s(\mathbf x)) K(\mathbf x)$ for two-phase flows.
The MRCM considers a decomposition of the domain $\Omega$ into non-overlapping subdomains $\Omega_i, \ i=1,2,\cdots,N$. Continuity of the normal fluxes and pressure are weakly imposed on a coarse scale $H$, which is significantly larger than the fine-scale of the discretization $h$. Here $H$ is the characteristic size of the subdomains.
The weak continuities are enforced to the multiscale solution $(\mathbf u_h,p_h)$ at the skeleton $\Gamma$ of the decomposition (union of all interfaces $\Gamma_{ij} = \Omega_i \cap \Omega_j$) by the following compatibility conditions
\begin{equation}
\int_\Gamma (\mathbf u_h^+ - \mathbf u_h^-)\cdot \check{\mathbf n} \ \psi \ d\Gamma=0 \quad \text{and}\quad
\int_\Gamma (p_h^+ - p_h^-)\ \phi \ d\Gamma=0
\label{eq:compatibility}
\end{equation}
for all $(\phi,\psi) \in \mathcal{U}_H \times \mathcal{P}_H$, which are the interface spaces defined over the edges $\mathcal{E}_h$ of the skeleton $\Gamma$.
Here $\check{\mathbf n}$ is a fixed normal vector to the skeleton $\Gamma$, pointing outwards from the sudomain with the smallest index. The solution on each side of the interface $\Gamma$ is represented by the $+$ and $-$ superscripts. The solution inside each subdomain
$(\mathbf u_h^i,p_h^i)$ is related to the normal flux and pressure unknowns at the interfaces $(U_H,P_H) \in \mathcal{U}_H \times \mathcal{P}_H$ by the Robin boundary conditions on the local problems
\begin{equation}
-\frac{\alpha(\mathbf x) H}{\kappa_i(\mathbf x)} \ \mathbf u_h^i \cdot \check{\mathbf n}^i + p_h^i =
-\frac{\alpha(\mathbf x) H}{\kappa_i(\mathbf x)}\ U_H \ \check{\mathbf n} \cdot \check{\mathbf n}^i + P_H, \quad \mathbf x \in \Gamma_{i,j},
\label{eq:robin}
\end{equation}
where the $\check{\mathbf n}^i$ is the normal vector to $\Gamma$ pointing outwards of $\Omega_i$.
The parameter for the Robin boundary condition in Equation (\ref{eq:robin}) is given by
\begin{equation}
\beta_i(\mathbf {x}) = \frac{\alpha(\mathbf{x}) H}{\kappa_i(\mathbf x)},
\label{eq:beta}
\end{equation}
where $\alpha(\mathbf {x})$ is a dimensionless algorithmic function that takes different values according to the permeability field (in the context of the adaptive MRCM).
We remark that by changing $\alpha$ we recover the Multiscale Mortar Mixed Finite Element Method (MMMFEM, \cite{arbogast}) when $\alpha \rightarrow 0$ and the Multiscale Hybrid-Mixed Finite Element Method (MHM, \cite{valentin}) when $\alpha \rightarrow +\infty$. The solution given by the Multiscale Mixed Method (MuMM, \cite{pereira}) can also be recovered under the right choice of parameters. The interested reader is referred to \cite{guiraldello2018multiscale} for more details.
The differential formulation of the MRCM consists in finding local solutions $(\mathbf u_h^i, p_h^i)$ into each subdomain $\Omega_i$, and global interface unknowns $(U_H,P_H)$ satisfying the local problems
\begin{equation}
\begin{array}{rcll}
\mathbf u_h^i &=& -\kappa(\mathbf{x)} \ \nabla p_h^i & \text{in }\Omega_i \\
\nabla \cdot \mathbf u_h^i &=& q & \text{in }\Omega_i \\
p_h^i & = & g & \text{on }\partial\Omega_i\cap \partial\Omega_p \\
\mathbf u_h^i \cdot \check{\mathbf n}^i &=& z & \text{on }\partial\Omega_i\cap \partial\Omega_u \\
-\beta_i \mathbf u_h^i \cdot \check{\mathbf n}^i + p_h^i &=& -\beta_i U_H \check{\mathbf n} \cdot \check{\mathbf n}^i + P_H & \text{on }\partial\Omega_i\cap \Gamma
\end{array}
\label{eq:mrc1}
\end{equation}
and the compatibility conditions
\begin{equation}
\begin{array}{rcl}
\displaystyle \sum_{i=1}^N \int_{\partial\Omega_i\cap \Gamma} (\mathbf u_h^i \cdot \check{\mathbf n}^i) \ \psi \ d\Gamma &=& 0\\
\displaystyle \sum_{i=1}^N \int_{\partial\Omega_i\cap \Gamma} \beta_i (\mathbf u_h^i \cdot \check{\mathbf n}^i - U_H\ \check{\mathbf n} \cdot \check{\mathbf n}^i)\ \phi \ (\check{\mathbf n} \cdot \check{\mathbf n}^i)\ d\Gamma &=& 0
\end{array}
\label{eq:mrc2}
\end{equation}
for all $(\phi,\psi)\in \mathcal{U}_H\times\mathcal{P}_H$.
The interface spaces $\mathcal{U}_H$ and $\mathcal{P}_H$ are defined over the skeleton $\Gamma$ as subspaces of
\begin{equation}
F_h(\mathcal{E}_h) = \left\{ f:\mathcal{E}_h\to \mathbb{R};~f|_e\,\in\,\mathbb{P}_0~,~\forall\,e\,\in\,\mathcal{E}_h \right\} ~,
\end{equation}
where $\mathcal{E}_h$ is the set of all edges of the skeleton $\Gamma$ \cite{guiraldello2018multiscale}.
These spaces are local and independently built on each interface $\Gamma_{i,j}$ and spanned by the multiscale basis functions $\{\phi_1, \phi_2, \cdots , \phi_{N_U}\}$ and $\{\psi_1, \psi_2, \cdots , \psi_{N_P}\}$, where $N_U$ and $N_P$ are the respective dimensions of the interface spaces (see \cite{guiraldello2018multiscale} for more details). The multiscale basis functions are obtained by solving the local problems in Equations (\ref{eq:mrc1}), that can be computed in parallel. These basis functions are further used to compute the global solution as a linear combination, whose coefficients are obtained by the solution of the global interface system generated by Equations (\ref{eq:mrc2}).
In the previous works, the authors consider the interface spaces $\mathcal{U}_H$ and $\mathcal{P}_H$ as low-dimensional polynomial spaces \cite{guiraldello2018multiscale} or informed spaces \cite{guiraldello2019interface} for the MRCM. In the next section, we propose alternative physics-based interface spaces to improve the accuracy of the multiscale solution of the elliptic problem. These interface spaces are built to capture variations of the permeability field, especially those characterized by fractures and barriers.
\section{Physics-based interface spaces} \label{physics-basead interface spaces}
Low-dimensional polynomial spaces are not robust to represent variations of high-contrast permeability fields containing channelized structures as fractures and barriers. Even informed spaces with a low number of degrees of freedom are not enough. To better represent the variations of fractured-like permeability fields we present two physics-based interface spaces. The idea is to build the multiscale basis functions based on the pressure and flux solutions at each channelized structure (fracture/barrier). We present one space for pressure and another for flux. The new pressure space is suited for high permeability channels, whereas the new flux space for barriers or low permeability barriers, that cross the interface between subdomains.
The support of the new basis functions are the interfaces that contain channelized structures. In the remaining interfaces the spaces can be freely chosen (any low-dimensional polynomial or informed spaces).
We denote $(\mathcal{U}_{H,k}, \mathcal{P}_{H,k})$ the choice of the interface spaces made up of the elementwise constant fine grid representation of polynomials over the interface elements, where $k$ is the degree of the polynomial. To introduce the concept we consider, for simplicity, the linear spaces $(\mathcal{U}_{H,1}, \mathcal{P}_{H,1})$ for both pressure and flux. This choice of interface spaces is used further in the numerical section. The basis of these spaces contain $N_P=2$ and $N_U=2$ functions on each interface between two subdomains.
If the permeability field contains fractures passing through the interface we propose to replace the space $\mathcal{P}_{H,1}$ by the physics-based pressure space. On the other hand, if the permeability filed contains barriers crossing the interface we replace the space $ \mathcal{U}_{H,1}$
by the physics-based flux space. In the following subsections, we describe how the physics-based interface spaces are built for capturing the high-contrast structures.
\subsection{A physics-based interface space for the pressure} \label{A physics-basead interface space for the pressure}
We define a simplified problem test to show the behavior of the solution to define the interface spaces.
In Figure \ref{fig:frac1} we consider a high-contrast permeability field containing a vertical fracture Figure \ref{fig:frac1}(a). We show the pressure Figure \ref{fig:frac1}(b) and flux Figure \ref{fig:frac1}(c) fine grid solutions. The flow is established by imposing a flux boundary condition from left to right and no-flow at top and bottom. This geometry induces a one-dimensional pressure solution which is plotted along a horizontal line in Figure \ref{fig:frac1}(d). Notice, the pressure is essentially constant over the fracture region.
Any domain decomposition with more than one subdomain in the $y$-direction contains horizontal interfaces through which the fracture passes. Let $\Gamma^{\text{frac}}$ be the set of all interfaces that contain at least one fine cell in which the absolute permeability is larger than a cutoff value $\zeta_{\max}$.
We intend to replace the pressure linear space $\mathcal{P}_{H,1}=\text{span}\{\psi_1, \psi_2\}$ at the interfaces $\Gamma_{i,j}\subset\Gamma^{\text{frac}}$ by a physics-based pressure space, denoted by $\mathcal{P}_{H}^*$.
\begin{figure}[htbp]
\center
\subfigure[ref1][Permeability]{\includegraphics[scale=0.178]{figs/1d/perm_high1.png}} \label{fig:frac1_a}
\subfigure[ref2][Pressure]{\includegraphics[scale=0.178]{figs/1d/press_high1.png}} \label{fig:frac1_b}
\subfigure[ref2][Flux]{\includegraphics[scale=0.178]{figs/1d/flux_high1.png}} \label{fig:frac1_c}
\subfigure[ref2][Pressure profile]{\includegraphics[scale=0.178]{figs/1d/press_1d_high1.pdf} } \label{fig:frac1_d}
\caption{Vertical fracture problem. (a) Permeability field (log-scaled) containing a vertical fracture. (b) Pressure field. (c) Flux. The colors in the flux plot refer to the log-scale flux magnitude. (d) Pressure along a horizontal line.
}
\label{fig:frac1}
\end{figure}
The new pressure space mimics the behavior of the pressure across the fractures. Let $\Gamma_{i,j}\subset\Gamma^{\text{frac}}$ be an interface with support in $[a,d]$ through which a fracture passes in $[b,c]\subset[a,d]$, as sketched in Figure \ref{fig:new_press_space}. The basis functions are defined as:
\begin{equation}\label{new_basis1}
\psi_1^*(x)=\left\{
\begin{array}{cl}
\dfrac{b-x}{b-a} &\mbox{if}\ x\in(a,b) \\
0\ &\mbox{otherwise}
\end{array} \right.
\end{equation}
\begin{equation}\label{new_basis2}
\psi_2^*(x)=\left\{
\begin{array}{cl}
\dfrac{x-a}{b-a} &\mbox{if}\ x\in(a,b) \\
1 &\mbox{if}\ x\in(b,c) \\
\dfrac{d-x}{d-c} &\mbox{if}\ x\in(c,d)
\end{array} \right.
\end{equation}
\begin{equation}\label{new_basis3}
\psi_3^*(x)=\left\{
\begin{array}{cl}
\dfrac{x-c}{d-c} &\mbox{if}\ x\in(c,d) \\
0\ &\mbox{otherwise}
\end{array} \right.
\end{equation}
The new interface pressure space is then defined as $\mathcal{P}_{H}^*=\text{span}\{\psi_1^*, \psi_2^*, \psi_3^*\}$ at $\Gamma_{i,j}\subset\Gamma^{\text{frac}}$. Since the definition of the basis depends only on the fine-grid discretization of the permeability at the interface, any fracture crossing the interface can be represented. Therefore, these basis functions are not restricted to fractures orthogonal to the interface. If the interface contains more than one fracture we need to define a new basis function with similar behavior to the $\psi_2^*$ for each fracture. The total number of basis functions per interface is thus $2+N_{\text{frac}}$, where $N_{\text{frac}}$ is the number of fractures at the interface.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.35]{figs/enrri_press.pdf}
\caption{Physics-based basis functions for pressure at the interfaces that contain fractures.
Note that the set of functions is able to capture the pressure profile across the fracture.
}
\label{fig:new_press_space}
\end{figure}
In Figure \ref{fig:frac2}, we show the pressure solution at the horizontal line $y=0.49$ delivered by the MMMFEM (by setting the $\alpha$ parameter of the MRCM to the value $10^{-6}$) in a domain decomposition with $2\times2$ coarse cells. We denote by MMMFEM-PBS the multiscale method combined with the physics-based spaces and by MMMFEM-POL the multiscale method combined with polynomial spaces (linear in this paper). We note that for both single (Figure \ref{fig:frac2}(a) and Figure \ref{fig:frac2}(c)) and multiple (Figure \ref{fig:frac2}(b) and Figure \ref{fig:frac2}(d)) fractures the correct pressure solution is only captured by the MMMFEM combined with physics-based interface spaces.
\begin{figure}[htbp]
\cente
\subfigure[ref1][Permeability]{\includegraphics[scale=0.172]{figs/1d/perm_high1_dd.png} } \label{fig:frac2_a}
\subfigure[ref2][Pressure]{\includegraphics[scale=0.172]{figs/1d/press_1d_high1_enrri.pdf} } \label{fig:frac2_b}
\subfigure[ref2][Permeability]{\includegraphics[scale=0.172]{figs/1d/perm_high2_dd.png} } \label{fig:frac2_c}
\subfigure[ref2][Pressure]{\includegraphics[scale=0.172]{figs/1d/press_1d_high2_enrri.pdf} } \label{fig:frac2_d}
\caption{High-contrast permeability field (log-scaled) with one (a) and two fractures (c). Pressure solutions for one (b) and two fractures (d) computed by the fine-grid solver and the multiscale ones MMMFEM-POL and MMMFEM-PBS. We note that for both one and two fractures the correct pressure solution is only captured by the MMMFEM-PBS.}
\label{fig:frac2}
\end{figure}
\subsection{A physics-based interface space for the flux} \label{A physics-basead interface space for the flux}
Now we focus on fields containing barriers. Once again we define a simplified problem to motivate the interface spaces.
In Figure \ref{fig:barrier1}(a) we consider a high-contrast permeability field containing a horizontal barrier. We show the fine grid solutions for pressure Figure \ref{fig:barrier1}(b) and for flux Figure \ref{fig:barrier1}(c). Here the flow is established by imposing a pressure gradient from left to right and no-flow at top and bottom. The $x$-component of the flux along a vertical line is illustrated in Figure \ref{fig:barrier1}(d), showing the discontinuities at the locations of transitions to barrier regions. Any domain decomposition with more than one subdomain in $x$-direction contains vertical interfaces through which the barrier passes. Let $\Gamma^{\text{barrier}}$ be the set of all the interfaces that contain at least one fine cell in which the absolute permeability is lower than a cutoff value $\zeta_{\min}$. We introduce the new flux spaces $\mathcal{U}_{H}^*$ according to the behavior of the flux solution by replacing the flux linear spaces $\mathcal{U}_{H,1}=\text{span}\{\phi_1, \phi_2\}$ at the interfaces $\Gamma_{i,j}\subset\Gamma^{\text{barrier}}$.
\begin{figure}[htbp]
\cente
\subfigure[ref1][Permeability]{\includegraphics[scale=0.173]{figs/1d/perm_low1.png}} \label{fig:barrier1_a}
\subfigure[ref2][Pressure]{\includegraphics[scale=0.173]{figs/1d/press_low1.png}} \label{fig:barrier1_b}
\subfigure[ref2][Flux]{\includegraphics[scale=0.173]{figs/1d/flux_low1.png}} \label{fig:barrier1_c}
\subfigure[ref2][Flux profile]{\includegraphics[scale=0.173]{figs/1d/flux_1d_low1.pdf} } \label{fig:barrier1_d}
\caption{Horizontal barrier problem. (a) Permeability field (log-scaled) containing a horizontal barrier. (b) Pressure field. (c) Flux. The colors in the flux plot refer to the log-scale flux magnitude.
Additionally, the $x$-component solution of the flux along a vertical line is illustrated in (d), where we note that the flux is discontinuous at the transitions to barrier regions.}
\label{fig:barrier1}
\end{figure}
Let $\Gamma_{i,j}\subset\Gamma^{\text{barrier}}$ be an interface with support in $[a,d]$ through which a barrier passes in $[b,c]\subset[a,d]$, as sketched in Figure \ref{fig:new_flux_space}. We define the following basis functions:
\begin{equation}\label{new_basis4}
\phi_1^*(x)=\left\{
\begin{array}{cl}
1 &\mbox{if}\ x\in(a,b) \\
0\ &\mbox{otherwise}
\end{array} \right.
\end{equation}
\begin{equation}\label{new_basis5}
\phi_2^*(x)=\left\{
\begin{array}{cl}
1 &\mbox{if}\ x\in(b,c) \\
0\ &\mbox{otherwise}
\end{array} \right.
\end{equation}
\begin{equation}\label{new_basis6}
\phi_3^*(x)=\left\{
\begin{array}{cl}
1 &\mbox{if}\ x\in(c,d) \\
0\ &\mbox{otherwise}
\end{array} \right.
\end{equation}
The new interface space is defined as $\mathcal{U}_{H}^*=\text{span}\{\phi_1^*, \phi_2^*, \phi_3^*\}$ at $\Gamma_{i,j}^{\text{barrier}}$. These basis functions are not restricted to barriers orthogonal to the interface, similar to the pressure basis. If the interface contains more than one barrier we need to define a new basis function with behavior similar to that of the $\phi_2^*$ for each barrier, plus a constant function for each region between two barriers. The total number of basis functions per interface is thus $1+2N_{\text{barrier}}$, where $N_{\text{barrier}}$ is the number of barriers.
\begin{figure}[htbp]
\centerin
\includegraphics[scale=0.35]{figs/enrri_flux.pdf}
\caption{Physics-based basis functions for flux at the interfaces that contain barriers. }
\label{fig:new_flux_space}
\end{figure}
In Figure \ref{fig:barrier2} we show the $x$-component of the flux along $x=0.5$ provided by the MHM (by setting the $\alpha$ parameter of the MRCM to the value $10^{6}$) in a domain decomposition with $2\times2$ coarse cells. We compare the approximations provided by the MHM-POL (MHM combined with the linear spaces) and MHM-PBS (MHM combined with the physics-based spaces) with the fine-grid solution.
We note that for both single (Figure \ref{fig:barrier2}(a) and Figure \ref{fig:barrier2}(c)) and multiple (Figure \ref{fig:barrier2}(b) and Figure \ref{fig:barrier2}(d)) barriers the correct $x$-component of the flux is only captured by MHM-PBS.
\begin{figure}[htbp]
\cente
\subfigure[ref1][Permeability]{\includegraphics[scale=0.166]{figs/1d/perm_low1_dd.png} } \label{fig:barrier2_a}
\subfigure[ref2][Flux]{\includegraphics[scale=0.166]{figs/1d/flux_1d_low1_enrri.pdf} } \label{fig:barrier2_b}
\subfigure[ref2][Permeability]{\includegraphics[scale=0.166]{figs/1d/perm_low2_dd.png} } \label{fig:barrier2_c}
\subfigure[ref2][Flux]{\includegraphics[scale=0.166]{figs/1d/flux_1d_low2_enrri.pdf} } \label{fig:barrier2_d}
\caption{High-contrast permeability field (log-scaled) with one (a) and two barriers (c). Fine-grid reference, MHM-POL and MHM-PBS solutions for the $x$-component of the flux considering one (c) and two barriers (d). Notice that for both one and two barriers the correct $x$-flux solution is only captured by the MHM-PBS.}
\label{fig:barrier2}
\end{figure}
\subsection{Experiments with the physics-based interface spaces}
The examples in Figure \ref{fig:frac2} and Figure \ref{fig:barrier2} illustrate that the usual linear interface spaces fail to approximate the solution in the presence of fractures and barriers even for simple problems. The proposed physics-based spaces have been able to represent the behavior of the solution in those problems.
In this subsection, we show that the physics-based spaces work well also in slightly more complex permeability fields.
Initially, we consider a permeability field with fractures and use the physics-based pressure space combined with the MMMFEM. Then we consider a permeability field with barriers and use the physics-based flux space combined with the MHM.
\subsubsection{The MMMFEM for permeability fields with fractures}
The first study considers a permeability field containing fractures. In this case, the physics-based pressure space is applied to the interfaces that contain cells in which the absolute permeability is larger than a cutoff value $\zeta_{\max}$. We combine the new interface space with the MMMFEM.
We consider the permeability in the fracture, ${K_{\max}}$ varying from $10$ to $10^8$ whereas the background is homogeneous with $K=1$, see Figure \ref{fig:press_for_fracs} (a). The cutoff value set to capture the fractures is $\zeta_{\max}=1$ in all cases. The domain considered is $\Omega=[0,1]\times[0,1]$ containing $160\times160$ cells. In Figure \ref{fig:press_for_fracs} (b) we show the relative $L^2(\Omega)$ errors for pressure as function of the permeability contrast. Three domain decompositions are considered: $4\times4$, $8\times8$ and $16\times16$ coarse cells, each one containing, respectively $40\times40$, $20\times20$ and $10\times10$ fine cells. The boundary conditions in the simulations of this subsection are no-flow at the top and bottom boundaries along with an imposed flux on the left and right boundaries. No source terms are considered.
We compare the multiscale solution considering the usual linear polynomial (MMMFEM-POL) and the physics-based (MMMFEM-PBS) pressure interface spaces. We note that for permeability contrasts larger than 100 the improvement with the physics-based spaces is significant for all the domain decompositions considered. Domain decompositions with more subdomains present smaller errors.
Figure \ref{fig:press_fracs} shows the pressure approximations for the decomposition of $8\times8$ coarse cells in the highest permeability contrast ${K_{\max}}/{K_{\min}}=10^8$. It is clear that the imprecisions of the MMMFEM solution with the linear interface spaces are corrected with the use of the physics-based ones.
\begin{figure}[htbp]
\cente
\subfigure[ref1][Permeability]{\includegraphics[scale=0.37]{figs_new/perm_high.png} } \label{fig:perm_fracs}
\subfigure[ref2][Pressure errors]{\includegraphics[scale=0.37]{figs_new/press_comp_contr_dd_high_mmmf.pdf} } \label{fig:error_fracs}
\caption{High-contrast permeability field with fractures (a). We consider the permeability in the fracture, ${K_{\max}}$ varying from $10$ to $10^8$ whereas the background is homogeneous with $K=1$.
Relative $L^2(\Omega)$ pressure errors as function of the contrast are shown for the MMMFEM-POL and MMMFEM-PBS (b). Three domain decompositions are considered: with $4\times4$, $8\times8$ and $16\times16$ coarse cells. We note a significant improvement for the MMMFEM-PBS in all the meshes, especially for high-contrast.}
\label{fig:press_for_fracs}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.25]{figs_new/press_8x8_high_fina.png}
\includegraphics[scale=0.25]{figs_new/press_8x8_high_mmmf.png}
\includegraphics[scale=0.25]{figs_new/press_8x8_high_mmmf_enrri.png}
\caption{Pressure approximations considering the contrast of ${K_{\max}}/{K_{\min}}=10^8$. Left to right: fine mesh, MMMFEM-POL, and MMMFEM-PBS solutions. The domain decomposition considered contains $8\times8$ coarse cells and is illustrated by the lines in the plot. We note that the MMMFEM-PBS solution is more accurate than the MMMFEM-POL.}
\label{fig:press_fracs}
\end{figure}
\subsubsection{The MHM for permeability fields with barriers}
In the next experiment, the same problem of the previous subsection is considered, except that the fractures are replaced by barriers of low permeability. The physics-based flux space is used at interfaces that contain cells in which the absolute permeability is lower than a cutoff value $\zeta_{\min}$. We combine the flux space with the MHM to approximate the solution. Similar to the previous example, we compare different permeability contrasts.
Here, we consider the permeability in the barrier, ${K_{\min}}$ varying from $10^{-8}$ to $10^{-1}$ whereas the background is homogeneous with $K=1$, see Figure \ref{fig:flux_for_barriers} (a). The cutoff value set to capture the barriers is $\zeta_{\min}=1$ in all cases.
The relative $L^2(\Omega)$ errors for flux as function of the contrast are displayed in Figure \ref{fig:flux_for_barriers} (b). Similar to the previous case, the solution is more accurate by using the physics-based space instead of the linear space. The flux errors provided by the linear spaces increase quickly with the contrast whereas the errors from the physics-based spaces are controlled. The results are consistent for the three domain decompositions, where the smaller errors are provided by the decompositions with more subdomains.
In Figure \ref{fig:flux_barriers} we show the flux approximations for the decomposition of $8\times8$ coarse cells in the highest permeability contrast ${K_{\max}}/{K_{\min}}=10^8$. We note that the MHM-POL solution is inaccurate and the MHM-PBS approximation captures the correct behavior of the reference solution.
\begin{figure}[htbp]
\cente
\subfigure[ref1][Permeability]{\includegraphics[scale=0.37]{figs_new/perm_low.png} } \label{fig:perm_barriers}
\subfigure[ref2][Flux errors]{\includegraphics[scale=0.37]{figs_new/flux_comp_contr_dd_low_mhm.pdf} } \label{fig:error_barriers}
\caption{High-contrast permeability field with barriers (a). We consider the permeability in the barrier, ${K_{\min}}$ varying from $10^{-8}$ to $10^{-1}$ whereas the background is homogeneous with $K=1$.
Relative $L^2(\Omega)$ flux errors as function of the contrast are shown for the MHM-POL and MHM-PBS (b). Three domain decompositions are considered: with $4\times4$, $8\times8$ and $16\times16$ coarse cells. We note that for high-contrast the only accurate approximations are produced by the MHM-PBS (in all the meshes).}
\label{fig:flux_for_barriers}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.25]{figs_new/flux_8x8_low_fina.png}
\includegraphics[scale=0.25]{figs_new/flux_8x8_low_mhm.png}
\includegraphics[scale=0.25]{figs_new/flux_8x8_low_mhm_enrri.png}
\caption{Flux approximations considering the contrast of ${K_{\max}}/{K_{\min}}=10^8$. Left to right: fine mesh, MHM-POL, and MHM-PBS solutions. The colors in the flux plot refer to the log-scale flux magnitude. The domain decomposition considered contains $8\times8$ coarse cells and is illustrated by the lines in the plot. We note that the MHM-PBS approximation is closely related to the reference solution and the MHM-POL is inaccurate.}
\label{fig:flux_barriers}
\end{figure}
\section{Adaptive MRCM with the physics-based interface spaces}\label{aMRCM-PBS}
We can conclude from the numerical studies reported above that the physics-based pressure (respectively, flux) space is fundamental to produce an accurate pressure (resp., flux) solution in presence of fractures (resp., barriers). Therefore if fractures and barriers appear in a single interface one can achieve a better approximation of pressure and flux using both pressure and flux physics-based interface spaces. To attain this goal one feature necessary for the multiscale method used is the ability to include the interface spaces independently. For this purpose, we combine the physics-based interface spaces with the MRCM, whose formulation enables to include the spaces separately. In the MRCM framework, we consider the adaptive version (denoting by $a$MRCM) that consists of setting the function $\alpha(\mathbf {x})$ depending on the variation of the permeability field at the interfaces.
In a previous work \cite{bifasico}, we found that the best choice is to set a small value $\alpha_{\text{small}}$ (pressure is favored) for regions where the permeability is larger than a cutoff value and a large value $\alpha_{\text{large}}$ (flux is favored) for the remaining areas. In this sense, the $a$MRCM controls the relative importance of each interface space at each location.
To solve the Equations (\ref{Darcy0}) or (\ref{Darcy}) by the MRCM we need to set the interface parameters (i.e., the interface spaces and the adaptive function $\alpha(\mathbf {x})$). In
Algorithm \ref{alg1} we detail the preprocessing operations to set the interface spaces and the adaptive coefficient $\alpha(\mathbf {x})$ from the permeability field. At each interface $\Gamma_{k},\ k=1,\cdots, M$ (where $M$ is the total number of interfaces, horizontal or vertical), and at each interface cell $e_l\in \Gamma_k,\ l=1,\cdots, m_k$ (where $m_k$ is the number of fine cells in $\Gamma_k$), it is defined whether $e_l$ corresponds to a fracture, a barrier, or to background, and then $\alpha({\bf x}_l)$ (at the cell's center) is set accordingly.
Then, we use the proposed physics-based interface spaces taking into account the observations presented in the previous section. We set the interface pressure (resp., flux) spaces $\mathcal{P}_{H,1}^*$ (respectively, $\mathcal{U}_{H,1}^*$) that consider the physics-based space for the interfaces containing fractures (resp., barriers) and linear spaces for the remaining interfaces.
Finally we are able to combine the $a$MRCM with the physics-based spaces $(\mathcal{U}_{H,1}^*, \mathcal{P}_{H,1}^*)$. We refer to this combination as $a$MRCM-PBS and consider for comparisons, the $a$MRCM-POL, which represents the $a$MRCM combined with linear polynomial spaces independently on the permeability field.
\begin{algorithm}
\caption{Setting interface parameters for solving Equations (\ref{Darcy0}) or (\ref{Darcy}) by the $a$MRCM-PBS}
\label{alg1}
\begin{algorithmic}[1]
\STATE Given $K(\mathbf x)$, $\zeta_{\min}$, $\zeta_{max}$, $\alpha_{\text{small}}$ and $\alpha_{\text{large}}$
\FOR{$k=1$ \TO $M$ }
\FOR{$l=1$ \TO $m_k$}
\STATE Evaluate the permeability in both sides ($K^-({\bf x}_l)$ and $K^+({\bf x}_l)$) of the interface cell $e_l\in\Gamma_k $
\IF {($\max\{K^-({\bf x}_l),K^+({\bf x}_l)\}>\zeta_{\max}$)}
\STATE $\alpha({\bf x}_l)=\alpha_{\text{small}}$
\STATE Add $\Gamma_{k}$ to $\Gamma^{\text{frac}}$
\ELSE
\STATE $\alpha({\bf x}_l)=\alpha_{\text{large}}$
\ENDIF
\IF {($\min\{K^-({\bf x}_l),K^+({\bf x}_l)\}<\zeta_{\min} $)}
\STATE Add $\Gamma_{k}$ to $\Gamma^{\text{barrier}}$
\ENDIF
\ENDFOR
\IF {($\Gamma_{k}\subset\Gamma^{\text{frac}}$)}
\STATE Compute the physics-based functions for pressure from Equations (\ref{new_basis1})-(\ref{new_basis3})
\STATE Set $\mathcal{P}_{H,1}^*=\mathcal{P}_{H}^*$ at $\Gamma_k$
\ELSE
\STATE Compute the linear polynomials functions for pressure
\STATE Set $\mathcal{P}_{H,1}^*=\mathcal{P}_{H,1}$ at $\Gamma_k$
\ENDIF
\IF {($\Gamma_{k}\subset\Gamma^{\text{barrier}}$)}
\STATE Compute the physics-based functions for flux from Equations (\ref{new_basis4})-(\ref{new_basis6})
\STATE Set $\mathcal{U}_{H,1}^*=\mathcal{U}_{H}^*$ at $\Gamma_k$
\ELSE
\STATE Compute the linear polynomials functions for flux
\STATE Set $\mathcal{U}_{H,1}^*=\mathcal{U}_{H,1}$ at $\Gamma_k$
\ENDIF
\ENDFOR
\STATE Given $(\mathcal{U}_{H,1}^*, \mathcal{P}_{H,1}^*)$ and $\alpha(\mathbf {x})$, solve Equations (\ref{Darcy0}) or (\ref{Darcy}) to obtain $p(\mathbf x)$ and $\mathbf u(\mathbf x)$ by the $a$MRCM-PBS
\end{algorithmic}
\end{algorithm}
In the next section, we explore the proposed $a$MRCM-PBS in challenging high-contrast fractured-like porous media through numerical experiments.
\section{Numerical experiments }\label{results}
We present representative numerical experiments to investigate the accuracy of the introduced physics-based interface spaces for the approximation of flows in high-contrast fractured-like porous media.
We start with the investigation of the physics-based interface spaces combined with the $a$MRCM for the elliptic problem. Then we study the influence of the physics-based interface spaces in the approximation of two-phase problems.
In all simulations the fine grid solution is used as the reference solution for evaluating the multiscale approximations. We consider the $a$MRCM (by setting $\alpha_{\text{small}}=10^{-2}$ and $\alpha_{\text{large}}=10^{2}$), MMMFEM (by setting $\alpha=10^{-6}$) and MHM (by setting $\alpha=10^{6}$).
To recover a continuous flux at the fine-scale we consider a velocity post-processing (downscaling)
\cite{guiraldello2019downscaling}. We choose the Stitch method, which has been indicated as the procedure that provides the best compromise between computational cost and precision.
The patch thicknesses of the oversampling regions for the Stitch method is fixed in two elements (that represents $10\%$ of the size of most subdomains considered in the examples).
\subsection{Single-phase flows in permeability fields with fractures and barriers}
We consider a permeability field containing both fractures and barriers, see Figure \ref{fig:perm_frac_and_barriers}. We fix the homogeneous background permeability with $K=1$ and consider the permeability in the barriers, ${K_{\min}}$ varying from $10^{-4}$ to $0.5$ whereas the permeability in the fractures, ${K_{\max}}$ varying from $10$ to $10^{4}$. The cutoff values considered are $\zeta_{\max}=\zeta_{\min}=1$.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.36]{figs_new/perm_high_and_low.png}
\caption{Permeability field containing both fractures (red) and barriers (blue). The contrast considered vary from $10$ to $10^8$.}
\label{fig:perm_frac_and_barriers}
\end{figure}
The relative $L^2(\Omega)$ error norms for pressure and flux as function of the contrast are shown in Figure \ref{fig:erro_frac_and_barriers} for three domain decompositions: with $4\times4$, $8\times8$ and $16\times16$ coarse cells. We observe that the pressure and flux errors related to the linear interface spaces increase quickly with the contrast. However, the errors provided by the physics-based spaces are moderate even for very large permeability contrasts. The results are consistent for all the domain decomposition tested, where the smaller errors are attained by the decompositions with more subdomains.
In Figure \ref{fig:comp_frac_and_barriers} we compare the solutions provided by the $a$MRCM, MMMFEM and MHM considering the domain decomposition of $8\times8$ subdomains having $20\times20$ fine cells into each one.
The first observation is that the MMMFEM and MHM approximations are not accurate for the permeability field with fractures and barriers even using the physics-based interface spaces. On the other hand, the $a$MRCM-PBS solutions are expressively more accurate than the ones obtained with the $a$MRCM-POL.
The pressure and flux approximations for the permeability contrast of $10^{8}$ are shown in Figure \ref{fig:press_frac_and_barriers} and Figure \ref{fig:flux_frac_and_barriers}, respectively. The plots confirm that the $a$MRCM combined with the physics-based interface spaces produces the most accurate solutions. We observe that all methods fail when using the linear interface spaces. We can conclude that the physics-based interface spaces are indispensable for a reasonable approximation of pressure, flux and further applications in two-phase flows.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.37]{figs_new/press_and_flux_comp_contr_dd.pdf}
\caption{Relative $L^2(\Omega)$ pressure (left) and flux (right) errors as function of the contrast are shown for the $a$MRCM-POL and $a$MRCM-PBS. Three domain decompositions are considered: with $4\times4$, $8\times8$ and $16\times16$ coarse cells. We note that the $a$MRCM-PBS is more accurate than the $a$MRCM-POL in all the contrast and meshes considered.}
\label{fig:erro_frac_and_barriers}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.38]{figs_new/press_comp_contr_methods.pdf}
\includegraphics[scale=0.38]{figs_new/flux_comp_contr_methods.pdf}
\caption{Relative $L^2(\Omega)$ pressure (left) and flux (right) errors as function of the contrast are shown for the $a$MRCM, MMMFEM and MHM considering both the linear and physics-based spaces. Three domain decompositions are considered: with $4\times4$, $8\times8$ and $16\times16$ coarse cells. We note that the $a$MRCM-PBS is more accurate than the $a$MRCM-POL in all the contrast and meshes considered.}
\label{fig:comp_frac_and_barriers}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.155]{figs_new/press_8x8_new.png}
\caption{Pressure approximations considering the contrast of ${K_{\max}}/{K_{\min}}=10^8$. First line, left to right: permeability field, fine mesh, $a$MRCM-POL, and $a$MRCM-PBS solutions. Second line, left to right: MMMFEM-POL, MMMFEM-PBS, MHM-POL, and MHM-PBS solutions. The domain decomposition considered contains $8\times8$ coarse cells and is illustrated by the lines in the plot. We note that the $a$MRCM-PBS approximation is the most closely related to the reference solution followed by the $a$MRCM-POL.}
\label{fig:press_frac_and_barriers}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.156]{figs_new/flux_8x8_new.png}
\caption{Flux approximations considering the contrast of ${K_{\max}}/{K_{\min}}=10^8$. First line, left to right: permeability field, fine mesh, $a$MRCM-POL, and $a$MRCM-PBS solutions. Second line, left to right: MMMFEM-POL, MMMFEM-PBS, MHM-POL, and MHM-PBS solutions. The colors refer to the log-scale flux magnitude. The domain decomposition considered contains $8\times8$ coarse cells and is illustrated by the lines in the plot. The only accurate procedure is the $a$MRCM-PBS.}
\label{fig:flux_frac_and_barriers}
\end{figure}
Finally, we present in Figure \ref{fig:varrido_frac_and_barriers}
the behavior of the errors for pressure and flux as a function of $\alpha$ (varying from $10^{-6}$ to $10^6$) for the permeability contrast of $10^8$ and maintaining the domain decomposition of $8\times8$ coarse cells. We compare the MRCM errors with linear and physics-based interface spaces. Two choices for the $a$MRCM are presented: setting $\alpha_{\text{small}}=10^{-2}$ and $\alpha_{\text{large}}=10^{2}$ or $\alpha_{\text{small}}=10^{-6}$ and $\alpha_{\text{large}}=10^{6}$. The errors for the adaptive version of the MRCM are illustrated as horizontal lines. We note a strong dependence on the parameter $\alpha$, where the minimum errors are attained at intermediate values (for choosing $\alpha$ constant), similarly to the previous works \cite{guiraldello2018multiscale, guiraldello2019interface, bifasico}. But the approximations by choosing any constant value of $\alpha$, even for the MRCM-PBS are inaccurate. We remark that the MMMFEM and MHM are also included in this observation. One can conclude that, besides the physics-based interface spaces, the $a$MRCM is necessary to obtain more accurate solutions. We note that the choice of $\alpha_{\text{small}}=10^{-2}$ and $\alpha_{\text{large}}=10^{2}$ or $\alpha_{\text{small}}=10^{-6}$ and $\alpha_{\text{large}}=10^{6}$ does not affect significantly the error.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.37]{figs_new/varrido_8x8.pdf}
\caption{Relative $L^2(\Omega)$ errors as a function of $\alpha$ for pressure (left) and flux (right) considering the permeability field plotted in Figure \ref{fig:perm_frac_and_barriers} with contrast of $10^8$ and the domain decomposition that contains $8\times8$ coarse cells. The physics-based and linear spaces are compared. We include the errors for two choices in the $a$MRCM: setting $\alpha_{\text{small}}=10^{-2}$ and $\alpha_{\text{large}}=10^{2}$ or $\alpha_{\text{small}}=10^{-6}$ and $\alpha_{\text{large}}=10^{6}$ (illustrated as horizontal lines). The improvement with the combination of the $a$MRCM and the physics-based spaces is expressive.}
\label{fig:varrido_frac_and_barriers}
\end{figure}
These results indicate that the linear interface spaces are not robust for capturing the effects of features as fractures and barriers. Even the physics-based spaces are not enough for complex fields if not combined with an appropriate multiscale method.
In order to study how these results are reflected in the corresponding saturation fields, in the next subsection we present numerical results for two-phase flows.
\subsection{Two-phase flow and transport problems}
Now we focus on a study of the MRCM performance for two-phase flows using the physics-based interface spaces for high-contrast fractured-like permeability fields. We introduce the saturation comparisons with a detailed analysis of fingering instabilities to show the impact of the design of the basis functions on the transport of the water saturation. Firstly we present numerical results for the permeability field with fractures and barriers of the previous section. Then we consider a high-contrast permeability field with channels and isolated inclusions that has frequently appeared in the literature \cite{jiang2017model}.
In all simulations the reservoir is initially filled with oil ($s^0=0$) and water is injected at a constant rate.
The relative permeabilities are given by $k_{ro}=(1-s)^2$ and $k_{rw}=s^2$, such that the fractional flow is written as
\begin{equation}
f(s) = \dfrac{M s^2}{M s^2 + (1-s)^2},
\label{frac_flow}
\end{equation}
where $ M = {\mu_o}/{\mu_w}$ is set as $M=10$.
In the operator splitting approach, we take the number of transport steps between the elliptic updates at most $C=20$. The time units employed are in pore volumes injected (PVI) that refers to the fraction of the total accessible pore volume injected into the domain \cite{chen2006computational}
\begin{equation}
T_{\text{PVI}}=-V_p^{-1} \int_0^t \int_{\partial\Omega_{\text{in}}}\mathbf{u}(\mathbf{x},\tau)\cdot \mathbf{n}\ dl\ d\tau,
\end{equation}
where $V_p$ is the total pore-volume of the reservoir, $t$ is the time taken for injection and $\partial\Omega_{\text{in}}$ the inflow well boundaries with the outward unit normal $\mathbf{n}$.
\subsubsection{High-contrast permeability field with fractures and barriers}
The objective of this study is to investigate the $a$MRCM, MMMFEM and MHM saturation solutions considering the physics-based interface spaces for the high-contrast permeability field with fractures and barriers. We consider the permeability field of the previous experiment, illustrated in Figure \ref{fig:perm_frac_and_barriers}. We fix the highest permeability contrast ${K_{\max}}/{K_{\min}}=10^8$ and the domain decomposition having $8\times8$ subdomains, each one discretized by $20\times20$ cells.
The first example considers the same slab geometry of the previous experiments, by imposing the flux on the left and right boundaries and no source terms. Figure \ref{fig:sat_maps_slab_N} shows the permeability field (log-scaled) along with the saturation profiles at $T_{\text{PVI}}=0.06$ (before breakthrough time) approximated by the multiscale methods with the linear and the physics-based interface spaces. The procedure that produces a saturation solution closer to the reference one is the $a$MRCM-PBS. The use of the physics-based spaces enables more accurate solutions for the $a$MRCM and the MMMFEM. However, the last one is still inaccurate. The MHM solutions are unacceptable either for linear or physics-based spaces. The corresponding relative $L^1(\Omega)$ errors throughout the simulation are presented in Figure \ref{fig:sat_errors_slab_N}. We note an expressive improvement provided by the physics-based interface spaces combined with the $a$MRCM. This combination enables the error to drop by one order of magnitude.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.52]{figs_new/sat_8x8_bc_slab_N.png}
\caption{Comparison of multiscale methods for the slab geometry with flux boundary conditions on the left and right. Saturation profiles at $T_{PVI}=0.06$ for the permeability field with fractures and barriers are shown. First line, left to right: high-contrast permeability field (log-scaled); reference fine grid solution; $a$MRCM-POL saturation profile; $a$MRCM-PBS saturation profile. Second line, left to right: MMMFEM-POL saturation profile; MMMFEM-PBS saturation profile; MHM-POL saturation profile; MHM-PBS saturation profile. The $a$MRCM-PBS provides the most accurate approximation.}
\label{fig:sat_maps_slab_N}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.55]{figs_new/sat_error_8x8_bc_slab_N.pdf}
\caption{Relative $L^{1}(\Omega)$ saturation errors as a function of time for the slab geometry with flux boundary conditions considering the field with fractures and barriers. We compare the $a$MRCM, MMMFEM and MHM with both physics-based and linear spaces. We note that the errors associated with the $a$MRCM-PBS are much smaller than all the others.}
\label{fig:sat_errors_slab_N}
\end{figure}
Now we maintain the slab geometry but using global boundary conditions of no-flow at top and bottom boundaries along with imposed pressure on the left and right boundaries.
In Figure \ref{fig:sat_maps_slab_D} we present the saturation profiles at $T_{\text{PVI}}=0.0001$ (before breakthrough time) approximated by the multiscale methods with the linear and the physics-based interface spaces. Similar results to the previous example are obtained. The corresponding relative errors throughout the simulation are presented in Figure \ref{fig:sat_errors_slab_D}, where it is clear the improved accuracy provided by combining the physics-based spaces with the $a$MRCM. Again, the MHM solutions are not acceptable with both interface spaces. The MMMFEM approximations improve significantly with the physics-based spaces, however, these solutions are comparable to the $a$MRCM-POL approximations.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.518]{figs_new/sat_8x8_bc_slab_D.png}
\caption{Comparison of multiscale methods for the slab geometry with pressure boundary conditions on the left and right. Saturation profiles at $T_{PVI}=0.0001$ for the permeability field with fractures and barriers are shown. First line, left to right: high-contrast permeability field (log-scaled); reference fine grid solution; $a$MRCM-POL saturation profile; $a$MRCM-PBS saturation profile. Second line, left to right: MMMFEM-POL saturation profile; MMMFEM-PBS saturation profile; MHM-POL saturation profile; MHM-PBS saturation profile. The $a$MRCM-PBS is clearly the most accurate procedure.}
\label{fig:sat_maps_slab_D}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.55]{figs_new/sat_error_8x8_bc_slab_D.pdf}
\caption{Relative $L^{1}(\Omega)$ saturation errors as a function of time for the slab geometry with pressure boundary conditions considering the field with fractures and barriers. We compare the $a$MRCM, MMMFEM and MHM with both physics-based and linear spaces. Similarly to the previous example, the errors associated with the $a$MRCM-PBS are much smaller than all the others.}
\label{fig:sat_errors_slab_D}
\end{figure}
Lastly, we test the multiscale methods in a quarter of a five-spot model, where we inject the water at the bottom-left cell and the sink is located at the top-right cell. The saturation profiles at $T_{\text{PVI}}=0.09$ (before breakthrough time) are shown in Figure \ref{fig:sat_maps_5spot}. The most accurate solutions are produced by the $a$MRCM considering both types of interface spaces.
The MMMFEM and MHM approximations present expressive fingering instabilities that are not present in the fine grid solution. In Figure \ref{fig:sat_errors_5spot}, the relative errors confirm that the $a$MRCM produces more accurate solutions. It is noticeable the poor quality solutions provided by the MMMFEM and MHM (even with the physics-based spaces), that are less accurate than the $a$MRCM-POL approximation. The solutions produced by the $a$MRCM-PBS are much more accurate than all the other approximations.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.52]{figs_new/sat_8x8_bc_5spot.png}
\caption{Comparison of multiscale methods for the quarter of a five-spot geometry. Saturation profiles at $T_{PVI}=0.09$ for the permeability field with fractures and barriers are shown. First line, left to right: high-contrast permeability field (log-scaled); reference fine grid solution; $a$MRCM-POL saturation profile; $a$MRCM-PBS saturation profile. Second line, left to right: MMMFEM-POL saturation profile; MMMFEM-PBS saturation profile; MHM-POL saturation profile; MHM-PBS saturation profile. The $a$MRCM is the only procedure that captures the details of the fingers.}
\label{fig:sat_maps_5spot}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.55]{figs_new/sat_error_8x8_bc_5spot.pdf}
\caption{Relative $L^{1}(\Omega)$ saturation errors as a function of time for the quarter of a five-spot geometry on the field with fractures and barriers. We compare the $a$MRCM, MMMFEM and MHM with both physics-based and linear spaces. The errors associated with the $a$MRCM-PBS are the smallest, followed by the $a$MRCM-POL.}
\label{fig:sat_errors_5spot}
\end{figure}
The high-contrast permeability fields are challenging for multiscale methods. We show that the methods fail in a field with both fractures and barriers. In all the models previously tested, the $a$MRCM-PBS is the only scheme that provides satisfactory approximations.
\subsubsection{High-contrast permeability field with channels and isolated inclusions}
In this subsection, we consider a high-contrast permeability field with channels and isolated inclusions that has frequently appeared in the literature \cite{efendiev2013generalized, chung2015mixed, chung2016enriched, chung2014adaptive, chung2017online, chung2018online, jiang2017model}. We consider a domain $\Omega$ containing $100\times100$ fine grid cells divided into $5\times5$ subdomains. The boundary conditions are no-flow at the top and bottom boundaries along with an imposed flux on the left and right boundaries. No source terms are considered. The permeability contrast considered is ${K_{\max}}/{K_{\min}} = 10^6$ and the cutoff values are set as $\zeta_{\max}=\zeta_{\min}=1$.
Figure \ref{fig:sat_high_efendiev} shows the permeability field (log-scaled) containing high-permeable channels and isolated inclusions and the saturation profiles at $T_{\text{PVI}}=0.07$ (before breakthrough). In this example, only the pressure physics-based spaces are used to handle the high-permeable structures, since there are no barriers. The more accurate solutions are provided by the $a$MRCM-PBS and the MMMFEM-PBS. For these two methods, the figure shows that the imprecisions that happen by using the linear spaces have completely disappeared with the physics-based spaces.
The MHM solutions are inaccurate. We remark that the MHM provides the same solution with the linear and the physics-based interface spaces, because this method considers only the flux space, which is always maintained as linear for this permeability field.
The relative errors are shown in Figure \ref{fig:sat_erro_high_efendiev} and reflect these observations throughout the whole simulation.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.52]{figs_new/sat_Efendiev_high.png}
\caption{Comparison of multiscale methods. Saturation profiles at $T_{PVI}=0.07$ for the permeability field with channels and isolated inclusions. First line, left to right: high-contrast permeability field (log-scaled); reference fine grid solution; $a$MRCM-POL saturation profile; $a$MRCM-PBS saturation profile. Second line, left to right: MMMFEM-POL saturation profile; MMMFEM-PBS saturation profile; MHM-POL saturation profile; MHM-PBS saturation profile. The more accurate approximations are attained by the $a$MRCM-PBS and MMMFEM-PBS.}
\label{fig:sat_high_efendiev}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.55]{figs_new/sat_error_Efendiev_high.pdf}
\caption{Relative $L^{1}(\Omega)$ saturation errors as a function of time on the field with channels and isolated inclusions. We compare the $a$MRCM, MMMFEM and MHM with both physics-based and linear spaces. The errors associated with the $a$MRCM-PBS and MMMFEM-PBS are the smallest.}
\label{fig:sat_erro_high_efendiev}
\end{figure}
As a final validation, we consider the performance of the methods in dealing with the same high-contrast permeability field replacing the type of channelized structures to barriers, as in \cite{jiang2017model}.
Figure \ref{fig:sat_low_efendiev} shows the permeability field (log-scaled) containing the low-permeable channels and the saturation profiles at $T_{\text{PVI}}=0.33$ (before breakthrough). In this case, only the flux physics-based spaces are used, since there are only low-permeable structures.
We note a considerable improvement in the $a$MRCM and MHM approximations replacing the linear interface spaces by the physics-based ones. The MMMFEM solutions present just some modest errors if compared to the $a$MRCM-POL and MHM-POL approximations. We remark that the MMMFEM solutions are the same with the linear and the physics-based interface spaces because, this method considers only the pressure space, which is always maintained as linear for this permeability field.
The relative errors throughout the whole simulation are shown in Figure \ref{fig:sat_low_erro_efendiev}. We confirm that the $a$MRCM-POL and MHM-POL are not accurate, the MMMFEM provides intermediate results and the $a$MRCM-PBS and MHM-PBS produce the most accurate approximations.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.52]{figs_new/sat_Efendiev_low.png}
\caption{Comparison of multiscale methods. Saturation profiles at $T_{PVI}=0.33$ for the permeability field with barriers are shown. First line, left to right: high-contrast permeability field (log-scaled); reference fine grid solution; $a$MRCM-POL saturation profile; $a$MRCM-PBS saturation profile. Second line, left to right: MMMFEM-POL saturation profile; MMMFEM-PBS saturation profile; MHM-POL saturation profile; MHM-PBS saturation profile. A considerable improvement is noticed in the $a$MRCM and MHM approximations replacing the linear spaces by the physics-based.}
\label{fig:sat_low_efendiev}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.55]{figs_new/sat_error_Efendiev_low.pdf}
\caption{Relative $L^{1}(\Omega)$ saturation errors as a function of time on the field with barriers. We compare the $a$MRCM, MMMFEM and MHM with both physics-based and linear spaces. The $a$MRCM-PBS and MMMFEM-PBS produce the smallest errors.}
\label{fig:sat_low_erro_efendiev}
\end{figure}
The numerical studies demonstrate the improvement obtained by using physics-based interface spaces. High-contrast fields were chosen that allowed us to assess the ability of the methods to handle problems in the presence of both high and low-permeable structures.
The results provide strong evidence that the adaptive MRCM combined with the physics-based spaces leads to improved transport approximations in high-contrast fractured-like porous media.
\section{Conclusions}\label{conclusions}
Two physics-based interface spaces (one for pressure and other for flux) have been proposed for better capturing the high-contrast effects of channelized structures. A careful investigation has been performed for the numerical solution of single and two-phase flows by combining the new spaces with multiscale mixed methods.
We show that the introduced physics-based pressure space can offer much better accuracy in comparison with the usual polynomial spaces in the presence of high-permeable structures. On the other hand, the proposed physics-based flux space can provide more accurate solutions in comparison with the polynomial spaces in the presence of low-permeable structures.
Two well known multiscale procedures have been applied to confirm these results: the Multiscale Mortar Mixed Finite Element Method (combined with the pressure space) and the Multiscale Hybrid-Mixed Finite Element Method (using the flux space). Aiming at using simultaneously both interface spaces we combine them with the MRCM, which allows for including the interface spaces independently. With this combination, we achieve the best accuracy in the approximation of challenging high-contrast flows in comparison with the other multiscale methods tested.
The MRCM formulation can take advantage of parallel computations with a computational cost comparable to existing procedures and providing a superior accuracy of the solutions in challenging high-contrast fields than the other multiscale methods that we have considered.
The development of interface spaces based on physics for 3D reservoir flow problems
is outside the scope of the work presented here. However, this is a topic that is
currently being considered by the authors.
Future work may include the use of the MRCM in more complex flow models, including the implicit solution of multiphase flow and transport problems.
\section*{Acknowledgments}
The authors gratefully acknowledge the financial support received from the Brazilian oil company Petrobras grant
2015/00400-4, and from the S\~ao Paulo Research Foundation FAPESP grant CEPID-CeMEAI 2013/07375-0; This study was
also financed in part by Brazilian government agencies CAPES (Finance Code 001) and CNPq grants 305599/2017-8 and 310990/2019-0; FP was
also funded in part by NSF-DMS 1514808, a Science Without Borders/CNPq-Brazil grant 400169/2014-2 and UT Dallas; FFR
acknowledges the hospitality provided by UT Dallas; Finally, the authors thank R.T. Guiraldello for providing the MRCM code and for
fruitful discussions about the method.
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
A wide range of tasks in computational biology start by determining, for a
given ``query gene'' $x$ in a species $A$, one or all genes $y$ in another
species $B$ that are most similar to $x$. Conceptually, these \emph{best
matches} $y$ of the query genes $x$ are meant to approximate the set of
genes in $B$ that are evolutionary most closely related to $x$. Best
matches can be identified by comparing evolutionary distances
\cite{Nei:06}, which in turn are usually obtained from sequence alignments
\cite{Chatzou:16}. In practice, fast approximation algorithms such as
\texttt{blast} and its more modern successors are often used for this
purpose \cite{MorenoHagelsieb:08,HernandezSalmeron:20}. Even if sequence
similarity is measured perfectly, deviations from a common molecular clock,
i.e., differences in the evolutionary rates of different genes, cause
discrepancies between \emph{best hits} (most similar sequences) and
\emph{best matches} (evolutionary most closely related sequences), see
\cite{Stadler:20a} for a detailed discussion.
The idea of best matches in the sense of closest evolutionary relatedness
pre-supposes an underlying tree $T$ that describes the phylogenetic
relationships among the genes, which correspond to the leaves of $T$, and a
map $\sigma$ assigning to each gene $x$ the species $\sigma(x)$ in which it
resides. Given such a leaf-colored tree $(T,\sigma)$, the best match graph
$\mathaccent"017E{G}(T,\sigma)$ has as its vertex set the leaves of $T$, i.e., the set of
genes, and as (directed) arcs the best matches. The latter are defined as
the pairs $(x,y)$ for which the last common ancestor of $x$ and $y$ is at
least as close to $x$ as the last common ancestor of $x$ and any other gene
$y'$ from the same spaces $\sigma(y')=\sigma(y)$. Best match graphs (BMGs),
i.e., graphs that are derived from a leaf-labeled tree $(T,\sigma)$ in this
manner, form a very restrictive class of colored digraphs
\cite{Geiss:19a,BMG-corrigendum}. Empirically determined best hit data
therefore will in general not satisfy the defining properties of BMGs. They
can be corrected in part, however, by considering the problem of editing a
given graph to the closest BMG. \citet{Schaller:20y} showed that the
arc modification problems for BMGs are NP-hard, but can be formulated as
integer linear programs (ILPs) allowing practical solutions for small
instances. However, in computational biology, applications to large gene
families would be of particular interest, creating the need for faster,
approximate solutions for BMG editing. Before embarking to develop software
for a BMG-based analysis of large sequence data sets, we need to understand
whether the editing problem for BMGs is tractable in practice with
sufficient accuracy and for interestingly large instances. The purpose of
this contributions is to establish that this is indeed the case.
Motivated by both theoretical and practical considerations, we are mainly
interested in heuristics that are consistent in the following sense: Let
$\mathbb{A}$ be an algorithm that takes an arbitrary vertex-colored digraph
$(\mathaccent"017E{G},\sigma)$ as input and outputs a BMG $\mathbb{A}(\mathaccent"017E{G},\sigma)$. Then
$\mathbb{A}$ is consistent if $\mathbb{A}(\mathaccent"017E{G},\sigma)=(\mathaccent"017E{G},\sigma)$ whenever
the input graph $(\mathaccent"017E{G},\sigma)$ is a BMG.
A vertex-colored digraph $(\mathaccent"017E{G}=(V,E),\sigma)$ is a BMG if and only if (a) a
certain set of informative triples $\mathscr{R}(\mathaccent"017E{G},\sigma)$, which can
easily be read off the input graph, is consistent and (b) the BMG
$\mathaccent"017E{G}(\hat T, \sigma)$ of the corresponding so-called Aho tree
$\hat T \coloneqq \Aho(\mathscr{R}(\mathaccent"017E{G},\sigma),V)$ coincides with $(\mathaccent"017E{G},\sigma)$
\cite{BMG-corrigendum}. In general, the Aho tree $\Aho(\mathscr{R}, V)$ of
a compatible set of triples $\mathscr{R}$ on a set $V$ is a least resolved
super-tree of all the triples in $\mathscr{R}$. For a BMG,
$\hat T \coloneqq \Aho(\mathscr{R}(\mathaccent"017E{G},\sigma), V)$ is the unique least
resolved tree (LRT) for $(\mathaccent"017E{G},\sigma)=\mathaccent"017E{G}(\hat T,\sigma)$. These close
connections between recognizing BMGs and constructing super-trees suggest
to adapt ideas from heuristic algorithms for triple consistency problems
and super-tree construction for BMG editing.
The simplest approach, therefore, is to extract a maximal consistent subset
$\mathscr{R}^*$ from $\mathscr{R}(\mathaccent"017E{G},\sigma)$ and to use
$\mathaccent"017E{G}(\Aho(\mathscr{R}^*,V),\sigma)$ as an approximation, see
Sec.~\ref{sect:simple}. A more detailed analysis of arcs in $(\mathaccent"017E{G},\sigma)$
that violate the property of being a BMG in Sec.~\ref{sect:UR}, however,
leads a notion of ``unsatisfiable relations'' (UR), which can be used to
count the arc modifications associated with a partition $\mathscr{V}$ of
the vertex set $V$ of $\mathaccent"017E{G}$. It also gives raise to a top-down algorithm in
which the vertex set of $\mathaccent"017E{G}$ is recursively edited and partitioned. A large
class of heuristics for BMG editing can be constructed depending on the
construction of the partition $\mathscr{V}$ in each recursion step. We
shall see that the arc edit sets in different steps of the recursion are
disjoint. A main result of this contribution,
Thm.~\ref{thm:algo-consistent}, links the partitions $\mathscr{V}$
appearing in BMG editing algorithms to the auxiliary graphs appearing in the
\texttt{BUILD} algorithm for supertree construction \cite{Aho:81}, see
Sec.~\ref{sect:notation}. This provides a guarantee that the BMG editing
algorithms are consistent provided the choice of $\mathscr{V}$ is such that it
does not enforce edits whenever an alternative partition with and empty UR
is available. For BMGs, this is in particular the case for the partitions
appearing in the \texttt{BUILD} algorithm. In Sec.~\ref{sect:UR-NP}, we
proceed to show by reduction from \PROBLEM{Set Splitting} that finding
a partition with a minimal number of unsatisfiable relations is NP-hard.
The theoretical results are complemented by computational experiments on
BMGs with randomly perturbed arc sets in
Sec.~\ref{sec:heur-computational-exp}. We focus on a comparison of
different algorithms to construct the partitions $\mathscr{V}$. Somewhat
surprisingly, we find that minimizing the cardinality of the UR alone is
not the best approach, since this tends to produce very unbalanced
partitions and thus requires a large number of steps in the recursions
whose costs add up. Instead, certain types of clustering or community
detection approaches that favor more balanced partitions tend to perform well.
\section{Notation and Preliminaries}
\label{sect:notation}
\paragraph{Partitions.} $\mathscr{V}=\{V_1,V_2,\dots,V_k\}$ is a partition
of a set $V$ if (i) $V_i\ne\emptyset$, (ii) $\bigcup_{i=1}^k V_i = V$ and
(iii) $V_i\cap V_j=\emptyset$ for $i\ne j$. A partition is non-trivial if
$|\mathscr{V}|\ge2$. Consider two partitions
$\mathscr{V}=\{V_1,\dots,V_k\}$ and $\mathscr{V}'=\{V'_1,\dots,V'_l\}$ of
$V$. If for every $1\le j'\le l$ there is a $j$ such that
$V'_{j'}\subseteq V_j$, i.e., if every set in $\mathscr{V}'$ is
completely contained in a set in $\mathscr{V}$, then $\mathscr{V}'$ is a
\emph{refinement} of $\mathscr{V}$, and $\mathscr{V}$ is a
\emph{coarse-graining} of $\mathscr{V}'$.
\paragraph{Graphs.}
Mostly, we consider simple directed graphs (digraphs) $\mathaccent"017E{G}=(V,E)$ with
vertex set $V$ and arc set
$E\subseteq V\times V \setminus \{(v,v)\mid v\in V\}$. We will frequently
write $V(\mathaccent"017E{G})$ and $E(\mathaccent"017E{G})$ to explicitly refer to the graph $\mathaccent"017E{G}$. For a
vertex $x\in V$, we say that $(y,x)$ is an \emph{in-arc} and $(x,z)$ is an
\emph{out-arc}. The subgraph induced by a subset $W\subseteq V$ is denoted
by $\mathaccent"017E{G}[W]$. Undirected graphs can be identified with symmetric digraphs,
i.e., the undirected graph $G$ underlying a digraph $\mathaccent"017E{G}$ is obtained by
dropping the direction of all arcs, or by symmetrizing the digraph, i.e.,
adding the arc $(y,x)$ to $E(\mathaccent"017E{G})$ for every arc $(x,y)\in E(\mathaccent"017E{G})$. When
referring to an undirected graph $G$, we write $xy$ for
$(x,y),(y,x)\in E(G)$ and call $xy$ an \emph{edge}. The (weakly) connected
components of $\mathaccent"017E{G}$ are the maximal connected subgraphs of the undirected
graph underlying $\mathaccent"017E{G}$ or, equivalently, the maximal strongly connected
components of the symmetrized digraph. Whenever the context is clear, we
will also refer to the partition of $V$ formed by the vertex sets of the
maximal connected subgraphs as the set of connected components.
A vertex coloring is a map $\sigma:V\to S$, where $S$ is a non-empty set of
colors. A vertex coloring of $\mathaccent"017E{G}$ is \emph{proper} if
$\sigma(x)\ne\sigma(y)$ whenever $(x,y)\in E(\mathaccent"017E{G})$. We write $(\mathaccent"017E{G},\sigma)$
for a vertex-colored digraph and denote by $V[r]$ the subset of vertices of
a graph $(\mathaccent"017E{G}=(V,E),\sigma)$ that have color $r$. Moreover, we define
$\sigma(W)\coloneqq \{\sigma(x) \mid x\in W\}$ for the subset of colors
present in a set $W\subseteq V$.
We write $N(x)$ for the set of out-neighbors of a vertex $x\in V(\mathaccent"017E{G})$ and
$N^-(x)$ for the set of in-neighbors of $x$. A graph $\mathaccent"017E{G}$ is called
\emph{sink-free} if $N(x)\ne\emptyset$ holds for all $x\in V(\mathaccent"017E{G})$.
We write
$A\symdiff B \coloneqq (A\setminus B)\cup(B\setminus A)$ for the symmetric
difference of two sets $A$ and $B$, and define, for a graph $\mathaccent"017E{G}=(V,E)$ and
arc set $F\subseteq (V\times V)\setminus\{(v,v)\mid v\in V\}$, the graph
$\mathaccent"017E{G}\symdiff F\coloneqq(V, E\symdiff F)$.
Analogously, we write $\mathaccent"017E{G}+ F\coloneqq(V, E\cup F)$ and $\mathaccent"017E{G}- F\coloneqq(V,
E\setminus F)$.
\paragraph{Phylogenetic trees.} Consider an undirected, rooted tree $T$
with leaf set $L(T)\subseteq V(T)$ and root $\rho_T\in V(T)$. Its inner
vertices are given by the set $V^0(T) = V(T) \setminus L(T)$. The
\emph{ancestor order} on $V(T)$ is defined such that
$u\preceq_T v$ if $v$ lies on the unique path from $u$ to the root
$\rho_T$, i.e., if $v$ is an ancestor of $v$. For brevity we set
$u \prec_T v$ if $u \preceq_{T} v$ and $u \neq v$. If $xy$ is an edge in
$T$ such that $y \prec_{T} x$, then $x$ is the \emph{parent} of $y$ and $y$
the \emph{child} of $x$. The set of children of a vertex $x\in V(T)$ is
denoted by $\mathsf{child}_T(x)$.
A tree is \emph{phylogenetic} if all its inner vertices have at least two
children. All trees considered in this contribution will be phylogenetic.
For a non-empty subset $A\subseteq V(T)$, we
define $\lca_T(A)$, the \emph{last common ancestor of $A$}, to be the
unique $\preceq_T$-minimal vertex of $T$ that is an ancestor of every
$u\in A$. Following e.g.\ \citet{Bryant:95}, we denote by $T_{L'}$ the
\emph{restriction} of $T$ to a subset $L'\subseteq L(T)$, i.e.\ $T_{L'}$ is
obtained by identifying the (unique) minimal subtree of $T$ that connects
all leaves in $L'$, and suppressing all vertices with degree two except
possibly the root $\rho_{T_{L'}}=\lca_T(L')$. We say that $T$
\emph{displays} or \emph{is a refinement of} a tree $T'$, in symbols
$T'\le T$, if $T'$ can be obtained from a restriction $T_{L'}$ of $T$ by a
series of inner edge contractions. $(T,\sigma)$ is a leaf-colored tree if
$\sigma: L(T)\to S$ is a map from the leaves of $T$ into a non-empty set of
colors. We say that $(T',\sigma')$ is displayed by $(T,\sigma)$ if
$T'\le T$ and $\sigma(v)=\sigma'(v)$ for all $v\in L(T')$.
\paragraph{Rooted triples.} A \emph{(rooted) triple} is a tree on three
leaves and with two inner vertices. We write $xy|z$ for the triple on the
leaves $x,y$ and $z$ if the path from $x$ to $y$ does not intersect the
path from $z$ to the root in $T$, i.e., if
$\lca_T(x,y)\prec_T \lca_T(x,z)=\lca_T(y,z)$. In this case we say that $T$
displays $xy|z$. We write
$\mathscr{R}_{|L'} \coloneqq\left\{ xy|z \in \mathscr{R} \,\colon x,y,z\in
L' \right\}$ for the restriction of a triple set $\mathscr{R}$ to a set
$L'$ of leaves. A set $\mathscr{R}$ of triples is \emph{consistent} if
there is a tree $T$ with leaf set $L:=\bigcup_{T'\in\mathscr{R}} L(T')$
that displays every triple in $\mathscr{R}$. The polynomial-time algorithm
\texttt{BUILD} decides for every triple set $\mathscr{R}$ whether it is
consistent, and if so, constructs a particular tree, the \emph{Aho tree}
$\Aho(\mathscr{R}, L)$, that displays every triple in $\mathscr{R}$
\cite{Aho:81}. The algorithm relies on the construction of an (undirected)
auxiliary graph, the \emph{Aho graph}, for a given triple set $\mathscr{R}$
on a set of leaves $L$. This graph, denoted by
$[\mathscr{R}, L]$, contains an edge $xy$ if and only if
$xy|z\in\mathscr{R}$ for some $z\in L$.
\section{Best match graphs}
\label{sect:BMG}
\begin{definition}
Let $(T,\sigma)$ be a leaf-colored tree. A leaf $y\in L(T)$ is a
\emph{best match} of the leaf $x\in L(T)$ if $\sigma(x)\neq\sigma(y)$ and
$\lca(x,y)\preceq_T \lca(x,y')$ holds for all leaves $y'$ of color
$\sigma(y')=\sigma(y)$.
\label{def:BMG}
\end{definition}
Given $(T,\sigma)$, the graph $\mathaccent"017E{G}(T,\sigma) = (V,E)$ with vertex set
$V=L(T)$, vertex-coloring $\sigma$, and with arcs $(x,y)\in E$ if and only
if $y$ is a best match of $x$ w.r.t.\ $(T,\sigma)$ is called the \emph{best
match graph} (BMG) of $(T,\sigma)$ \cite{Geiss:19a}.
\begin{definition}\label{def:BestMatchGraph}
An arbitrary vertex-colored graph $(\mathaccent"017E{G},\sigma)$ is a \emph{best match
graph (BMG)} if there exists a leaf-colored tree $(T,\sigma)$ such that
$(\mathaccent"017E{G},\sigma) = \mathaccent"017E{G}(T,\sigma)$. In this case, we say that $(T,\sigma)$
\emph{explains} $(\mathaccent"017E{G},\sigma)$.
\end{definition}
We say that $(\mathaccent"017E{G}=(V,E),\sigma)$ is an $\ell$-BMG if
$|\sigma(V)|=\ell$. By construction, there is at least one best match of
$x$ for every color $s\in\sigma(V)\setminus\{\sigma(x)\}$:
\begin{fact}
For every vertex $x$ and every color $s\neq \sigma(x)$ in a BMG
$(\mathaccent"017E{G},\sigma)$ there is some vertex $y\in N(x)$ with $\sigma(y)=s$.
Equivalently, the subgraph induced by every pair of colors is sink-free.
\label{fact:allcolors-out}
\end{fact}
In particular, therefore, BMGs are sink-free whenever they contain at least
two colors. We formalize this basic property of BMGs for colored graphs in
general:
\begin{definition}
Let $(G=(V,E),\sigma)$ be a colored graph. The coloring $\sigma$ is
\emph{sink-free} if it is proper and, for every vertex $x\in V$ and
every color $s\ne\sigma(x)$ in $\sigma(V)$, there is a vertex $y\in N(x)$
with $\sigma(y)=s$. A graph with a sink-free coloring is called
\emph{sf-colored}.
\end{definition}
Given a tree $T$ and an edge $e$, we denote by $T_e$ the tree obtained from
$T$ by contracting the edge $e$. We call an edge in $(T,\sigma)$
\emph{redundant (w.r.t.\ $(\mathaccent"017E{G},\sigma)$)} if both $(T,\sigma)$ and
$(T_e,\sigma)$ explain $(\mathaccent"017E{G},\sigma)$.
\begin{definition}
A tree $(T,\sigma)$ is \emph{least resolved} for a BMG $(\mathaccent"017E{G},\sigma)$ if
(i) it explains $(\mathaccent"017E{G},\sigma)$ and (ii) it does not contain any redundant
edges w.r.t.\ $(\mathaccent"017E{G},\sigma)$.
\label{def:LRT}
\end{definition}
\citet[Thm.~8]{Geiss:19a} showed that every BMG has a unique least resolved
tree (LRT).
\citet{Geiss:19a} gave a characterization of BMGs that makes use of
a set of informative triples, which can be defined compactly as
follows \cite{Schaller:20x}:
\begin{definition}\label{def:informative_triples}
Let $(\mathaccent"017E{G},\sigma)$ be a vertex-colored digraph. Then the set of
\emph{informative triples} is
\begin{align*}
\mathscr{R}(\mathaccent"017E{G},\sigma) &\coloneqq
\left\{ab|b' \colon
\sigma(a)\neq\sigma(b)=\sigma(b'),\,
(a,b)\in E(\mathaccent"017E{G}), \text{ and }
(a,b')\notin E(\mathaccent"017E{G}) \right\},
\intertext{and the set of \emph{forbidden triples} is}
\mathscr{F}(\mathaccent"017E{G},\sigma) &\coloneqq
\left\{ab|b' \colon
\sigma(a)\neq\sigma(b)=\sigma(b'),\,
b\ne b',\, \text{ and }
(a,b),(a,b')\in E(\mathaccent"017E{G}) \right\}.
\intertext{For the subclass of BMGs that can be explained by binary
trees, we will furthermore need}
\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G},\sigma) &\coloneqq
\mathscr{R}(\mathaccent"017E{G},\sigma) \cup \{bb'|a\colon ab|b'\in \mathscr{F}(\mathaccent"017E{G},\sigma)
\text{ and }\sigma(b)=\sigma(b')\}.
\end{align*}
\end{definition}
By definition, $a,b,b'$ must be pairwise distinct whenever
$ab|b'\in\mathscr{R}(\mathaccent"017E{G},\sigma)$, $ab|b'\in\mathscr{F}(\mathaccent"017E{G},\sigma)$, or
$ab|b'\in\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G},\sigma)$.
We extend the notion of consistency to pairs of triple sets in
\begin{definition}
Let $\mathscr{R}$ and $\mathscr{F}$ be sets of triples. The pair
$(\mathscr{R},\mathscr{F})$ is called \emph{consistent} if there is a
tree $T$ that displays all triples in $\mathscr{R}$ but none of the
triples in $\mathscr{F}$. In this case, we also say that $T$
\emph{agrees with} $(\mathscr{R},\mathscr{F})$.
\end{definition}
It can be decided in polynomial time whether such a pair
$(\mathscr{R},\mathscr{F})$ is consistent using the algorithm \texttt{MTT},
which was named after the corresponding \emph{mixed triplets problem
restricted to trees} and described by \citet{He:06}.
We continue with two simple observations concerning the restriction of
triple sets. Since informative and forbidden triples $xy|z$ are only
defined by the presence and absence of arcs in the subgraph of $\mathaccent"017E{G}$ induced
by $\{x,y,z\}$, this leads to the following
\begin{fact} \cite{Schaller:20p}
\label{obs:R-restriction}
Let $(\mathaccent"017E{G},\sigma)$ be a vertex-colored digraph and $V'\subseteq V(\mathaccent"017E{G})$.
Then $R(\mathaccent"017E{G},\sigma)_{|V'}=R(\mathaccent"017E{G}[V'],\sigma_{|V'})$ holds for every
$R\in\{\mathscr{R},\mathscr{F}, \mathop{\mathscr{R}^{\textrm{B}}} \}$.
\end{fact}
Moreover, any pair of triples $(\mathscr{R}',\mathscr{F}')$ such that
$\mathscr{R}'\subseteq\mathscr{R}$ and $\mathscr{F}'\subseteq\mathscr{F}$
for a consistent pair $(\mathscr{R},\mathscr{F})$ remains consistent since
any tree that agrees with $(\mathscr{R},\mathscr{F})$ clearly displays all
triples in $\mathscr{R}'$ and none of the triples in $\mathscr{F}'$.
Hence, we have
\begin{fact}
\label{obs:R-F-subsets}
Let $\mathscr{R}'\subseteq\mathscr{R}$ and
$\mathscr{F}'\subseteq\mathscr{F}$ for a consistent pair of triple sets
$(\mathscr{R},\mathscr{F})$. Then $(\mathscr{R}',\mathscr{F}')$ is
consistent.
\end{fact}
We summarize two characterizations of BMGs given in
\cite[Thm.~15]{BMG-corrigendum} and
\cite[Lemma~3.4 and Thm.~3.5]{Schaller:20y} in the following
\begin{proposition}\label{prop:BMG-charac}
Let $(\mathaccent"017E{G},\sigma)$ be a properly colored digraph with vertex set $L$.
Then the following three statements are equivalent:
\begin{enumerate}
\item $(\mathaccent"017E{G},\sigma)$ is a BMG.
\item $\mathscr{R}(\mathaccent"017E{G},\sigma)$ is consistent and
$\mathaccent"017E{G}(\Aho(\mathscr{R}(\mathaccent"017E{G},\sigma),L), \sigma) = (\mathaccent"017E{G},\sigma)$.
\item $(\mathaccent"017E{G},\sigma)$ is sf-colored and
$(\mathscr{R}(\mathaccent"017E{G},\sigma),\mathscr{F}(\mathaccent"017E{G},\sigma))$ is consistent.
\end{enumerate}
In this case, $(\Aho(\mathscr{R}(\mathaccent"017E{G},\sigma),L),\sigma)$ is the unique
LRT for $(\mathaccent"017E{G},\sigma)$, and a leaf-colored tree
$(T,\sigma)$ on $L$ explains $(\mathaccent"017E{G},\sigma)$ if and only if it agrees with
$(\mathscr{R}(\mathaccent"017E{G},\sigma), \mathscr{F}(\mathaccent"017E{G},\sigma))$.
\end{proposition}
\begin{figure}[htb]
\centering
\includegraphics[width=0.85\textwidth]{./bmg-aho-example.pdf}
\caption[]{Construction of the LRT for a BMG $(\mathaccent"017E{G},\sigma)$ using the
\texttt{BUILD} algorithm. Each recursion step (pink boxes) corresponds
to a vertex of the resulting tree (trivial steps on single vertices are
omitted in the drawing). The algorithm recurses on the connected
components (gray dashed boxes) of the Aho graphs and the corresponding
subsets of triples.}
\label{fig:bmg-aho-example}
\end{figure}
Prop.~\ref{prop:BMG-charac} states that the set of informative triples
$\mathscr{R}(\mathaccent"017E{G},\sigma)$ of a BMG $(\mathaccent"017E{G},\sigma)$ is consistent. Therefore,
it can be used to construct its LRT by means of the \texttt{BUILD}
algorithm, see Fig.~\ref{fig:bmg-aho-example} for an example.
\begin{figure}[htb]
\centering
\includegraphics[width=0.85\textwidth]{./introduce-inconsistency.pdf}
\caption[]{Both arc insertions and deletions into a BMG
$(\mathaccent"017E{G}=(V,E),\sigma)$ can introduce inconsistencies into the set of
informative triples. Top row: Leaf-colored tree $(T,\sigma)$ explaining
the BMG $(\mathaccent"017E{G},\sigma)$. Its set of informative triples is
$\mathscr{R}(\mathaccent"017E{G},\sigma)=\{ab_1|b_2,\, ab_1|b_3,\, c_1b_2|b_1,\,
c_1b_3|b_1,\, c_2b_2|b_1,\, c_2b_3|b_1\}$ giving the Aho graph
$H=[\mathscr{R}(\mathaccent"017E{G},\sigma), V]$. Bottom left: Insertion of the arc
$(a, b_2)$ creates a new informative triple $ab_2|b_3$ ($ab_1|b_2$ gets
lost) resulting in a connected Aho graph $H'$. Bottom right: Deletion
of the arc $(a, c_1)$ creates a new triple $ac_2|c_1$ resulting in a
connected Aho graph $H''$.}
\label{fig:introduce-inconsistency}
\end{figure}
It is important to note that both arc insertions and deletions
may lead to creation and loss of informative triples. In particular, when
starting from a BMG, both types of modifications have the potential to make
the triple set inconsistent as the example in
Fig.~\ref{fig:introduce-inconsistency} shows. This is indeed often the
case even for moderate disturbances of a BMG as we shall see in
Sec.~\ref{sec:heur-computational-exp}.
We expect that empirically estimated best match relations will typically
contain errors that correspond to both arc insertions and deletions w.r.t.\
the unknown underlying ``true'' best match graph. This motivates the
problem of editing a given vertex-colored digraph to a BMG:
\begin{problem}[\PROBLEM{$\ell$-BMG Editing}]\ \\
\label{prblm:ell-bmg-edit}
\begin{tabular}{ll}
\emph{Input:} & A properly $\ell$-colored digraph $(\mathaccent"017E{G} =(V,E),\sigma)$
and an integer $k$.\\
\emph{Question:} & Is there a subset $F\subseteq V\times V \setminus
\{(v,v)\mid v\in V\}$ such
that $|F|\leq k$ and\\ & $(\mathaccent"017E{G}\symdiff F,\sigma)$
is an $\ell$-BMG?
\end{tabular}
\end{problem}
Natural variants are \PROBLEM{$\ell$-BMG Completion} and
\PROBLEM{$\ell$-BMG Deletion} where $\mathaccent"017E{G}\symdiff F$ is replaced by
$\mathaccent"017E{G}+F$ and $\mathaccent"017E{G}-F$, respectively, i.e., only addition or
deletion of arcs is allowed. Both \PROBLEM{$\ell$-BMG Editing} and its
variants are NP-complete \cite{Schaller:20y}.
The heuristic algorithms considered in this contribution can be thought of
as maps $\mathbb{A}$ on the set of finite vertex-colored digraphs such that
$\mathbb{A}(\mathaccent"017E{G},\sigma)$ is a BMG for every vertex-colored input graph
$(\mathaccent"017E{G},\sigma)$. In particular, the following property of such algorithms is
desirable:
\begin{definition}
A (BMG-editing) algorithm is \emph{consistent} if
$\mathbb{A}(\mathaccent"017E{G},\sigma)=(\mathaccent"017E{G},\sigma)$ whenever $(\mathaccent"017E{G},\sigma)$ is a BMG.
\end{definition}
\section{A simple, triple-based heuristic}
\label{sect:simple}
The triple-based characterization summarized by Prop.~\ref{prop:BMG-charac}
suggests a simple heuristic for BMG editing that relies on replacing the
consistency checks for triple sets by the extraction of maximal sets of
consistent triples (see Alg.~\ref{alg:simple}). Both \PROBLEM{MaxRTC}, the
problem of extracting from a given set $\mathscr{R}$ of rooted triples a
maximum-size consistent subset, and \PROBLEM{MinRTI}, the problem of
finding a minimum-size subset $\mathscr{I}$ such that
$\mathscr{R}\setminus\mathscr{I}$ is consistent, are NP-hard
\cite{Jansson:01}. Furthermore, \PROBLEM{MaxRTC} is APX-hard and
\PROBLEM{MinRTI} is $\Omega(\ln n)$-inapproximable
\cite{Byrka:10}. However, because of their practical importance in
phylogenetics, a large number of practically useful heuristics have been
devised, see e.g.\ \cite{Gasieniec:99,Wu:04,Tazehkand:13}.
\begin{algorithm}
\caption{Simple Heuristic for BMG editing.}
\label{alg:simple}
\KwIn{Properly colored digraph $(\mathaccent"017E{G},\sigma)$}
\KwOut{BMG $(\mathaccent"017E{G}^*,\sigma)$}
$\mathscr{R}^*\leftarrow
\PROBLEM{MaxRTC}(\mathscr{R}(\mathaccent"017E{G},\sigma))$\\
\Return{$\mathaccent"017E{G}(\Aho(\mathscr{R}^*, V(\mathaccent"017E{G})),\sigma)$}
\end{algorithm}
As a consequence of Prop.~\ref{prop:BMG-charac}, Alg.~\ref{alg:simple}
is consistent, i.e., $(\mathaccent"017E{G}^*,\sigma)=(\mathaccent"017E{G},\sigma)$ if and only if the input
graph $(G,\sigma)$ is a BMG, if a consistent heuristic is
employed to solve \PROBLEM{MaxRTG}/\PROBLEM{MinRTI}, i.e., if consistent
triple sets remain unchanged by the method approximating
\PROBLEM{MaxRTG}/\PROBLEM{MinRTI}.
\begin{figure}[htb]
\centering
\includegraphics[width=0.65\textwidth]{./triple_heur_not_optimal.pdf}
\caption[]{Example for a graph (A) where Alg.~\ref{alg:simple} does not
lead to an optimal BMG editing. The set $\mathscr{R}(\mathaccent"017E{G},\sigma)$ is
empty and thus consistent. (B) The tree
$T=\Aho(\mathscr{R}(\mathaccent"017E{G},\sigma), V(\mathaccent"017E{G}))$ and (C) its corresponding BMG. The
two arcs $(b,a)$ and $(b,a')$ have been inserted. (D) A tree $(T',\sigma)$
and (E) its corresponding BMG $\mathaccent"017E{G}(T',\sigma)$ in which only the arc
$(b,a)$ has been inserted.}
\label{fig:heuristic_not_optimal}
\end{figure}
The heuristic Alg.~\ref{alg:simple} is not always optimal, even if
\PROBLEM{MaxRTC}/\PROBLEM{MinRTI} is solved optimally.
Fig.~\ref{fig:heuristic_not_optimal} shows an unconnected 2-colored graph
$(\mathaccent"017E{G},\sigma)$ on three vertices that is not a BMG and does not contain
informative triples. The BMG $(\mathaccent"017E{G}^*,\sigma)$ produced by
Alg.~\ref{alg:simple} introduces two arcs into $(\mathaccent"017E{G},\sigma)$. However,
$(\mathaccent"017E{G},\sigma)$ can also be edited to a BMG by inserting only one arc.
A simple improvement is to start by enforcing obvious arcs: If $v$ is the
only vertex with color $\sigma(v)$, then by definition there must be an arc
$(x,v)$ for every vertex $x\ne v$. The computation then starts from the
sets of informative triples of the modified graph. We shall see below that
these are the only arcs that can safely be added to $\mathaccent"017E{G}$ without other
additional knowledge or constraints (cf.\ Thm.~\ref{thm:trivial-unsat-rel}
below).
\section{Locally optimal splits}
\label{sect:UR}
Finding an optimal BMG editing of a digraph $(\mathaccent"017E{G}=(V,E),\sigma)$ is equivalent
to finding a tree $(T,\sigma)$ on $V$ that minimizes the cardinality of
\begin{equation}
\begin{split}
U(\mathaccent"017E{G},T) \coloneqq \{ (x,y)\in V\times V \mid\ &
(x,y)\in E \textrm{ and } (x,y)\notin E(\mathaccent"017E{G}(T,\sigma))
\text{, or} \\
& (x,y)\notin E \textrm{ and } (x,y)\in E(\mathaccent"017E{G}(T,\sigma)) \}.
\end{split}
\end{equation}
Clearly, $U(\mathaccent"017E{G},T) =\emptyset$ implies that $(\mathaccent"017E{G},\sigma) = \mathaccent"017E{G}(T,\sigma)$ is
a BMG. However, finding a tree $(T,\sigma)$ that minimizes $|U(\mathaccent"017E{G},T)|$ is
intractable (unless $P=NP$) since \PROBLEM{$\ell$-BMG Editing},
Problem~\ref{prblm:ell-bmg-edit} above, is NP-complete
\cite{Schaller:20y}.
We may ask, nevertheless, if trees $(T,\sigma)$ on $V$ contain information
about arcs and non-arcs in $(\mathaccent"017E{G},\sigma)$ that are ``unambiguously false''
in the sense that they are contained in every edit set that converts
$(\mathaccent"017E{G},\sigma)$ into a BMG. Denote by $\mathscr{T}_V$ the set of all
phylogenetic trees on $V$. The set of these ``unambiguously false''
\hbox{(non-)}arcs can then be expressed as
\begin{equation}
\label{eq:U-unambiguous}
U^*(\mathaccent"017E{G})\coloneqq \bigcap_{T\in \mathscr{T}_V} U(\mathaccent"017E{G},T).
\end{equation}
Since there are in general exponentially many trees on $V$ and thus, the
problem of determining $U^*(\mathaccent"017E{G})$ seems to be quite challenging on a first
glance. We shall see in Thm.~\ref{thm:trivial-unsat-rel}, however, that
$U^*(\mathaccent"017E{G})$ can be computed efficiently. We start with a conceptually
simpler construction.
\begin{definition}
\label{def:unsat-relations}
Let $(\mathaccent"017E{G}=(V,E),\sigma)$ be a properly vertex-colored digraph and
$\mathscr{V}$ a partition of $V$ with $|\mathscr{V}|\ge2$. Moreover, let
$\mathscr{T}(\mathscr{V})$ be the set of trees $T$ on $V$ that satisfy
$\mathscr{V} = \{L(T(v)) \mid v\in\mathsf{child}_{T}(\rho_T) \}$. The set of
\emph{unsatisfiable relations} (\textit{UR}\xspace), denoted by $U(\mathaccent"017E{G},\mathscr{V})$, is
defined as
\begin{equation}
\label{eq:U(G,V)}
U(\mathaccent"017E{G},\mathscr{V}) \coloneqq \bigcap_{T\in \mathscr{T}(\mathscr{V})}
U(\mathaccent"017E{G},T).
\end{equation}
The associated \textit{UR}\xspace-cost is $c(\mathaccent"017E{G},\mathscr{V})\coloneqq |U(\mathaccent"017E{G},\mathscr{V})|$.
\end{definition}
The set of (phylogenetic) trees $\mathscr{T}(\mathscr{V})$ is non-empty
since $|\mathscr{V}|\ge2$ in Def.~\ref{def:unsat-relations}. Moreover, by
construction, $(x,y) \in U(\mathaccent"017E{G},\mathscr{V})$ if and only if
\begin{align*}
& (x,y)\in E \textrm{ and } (x,y)\notin E(\mathaccent"017E{G}(T,\sigma)) \textrm{ for all }
T\in \mathscr{T}(\mathscr{V}) \textrm{, or}\\
& (x,y)\notin E \textrm{ and } (x,y)\in E(\mathaccent"017E{G}(T,\sigma)) \textrm{ for all }
T\in \mathscr{T}(\mathscr{V}).
\end{align*}
Intriguingly, the set $U(\mathaccent"017E{G},\mathscr{V})$, and thus the \textit{UR}\xspace-cost
$c(\mathaccent"017E{G},\mathscr{V})$, can be computed in polynomial time without any
explicit knowledge of the possible trees to determine the set
$U(\mathaccent"017E{G},\mathscr{V})$. To this end, we define the three sets
\begin{align*}
U_1(\mathaccent"017E{G},\mathscr{V}) &= \bigcup_{V_{i}\in\mathscr{V}} \{(x,y) \mid
(x,y)\in E,\; x\in V_{i},\;
y\in V\setminus V_{i},\;
\sigma(y)\in \sigma(V_{i})\},\\
U_2(\mathaccent"017E{G},\mathscr{V}) &= \bigcup_{V_{i}\in\mathscr{V}} \{(x,y) \mid
(x,y)\notin E,\; x\in V_{i},\;
y\in V\setminus V_{i},\;
\sigma(y)\notin \sigma(V_{i})\}, \textrm{ and}\\
U_3(\mathaccent"017E{G},\mathscr{V}) &= \bigcup_{V_{i}\in\mathscr{V}} \{(x,y) \mid
(x,y)\notin E,\; \textrm{distinct }x,y\in V_{i},\;
V_{i}[\sigma(y)]=\{y\}\}.
\end{align*}
\begin{lemma}
\label{lem:unsat-rel-charac}
Let $(\mathaccent"017E{G}=(V,E),\sigma)$ be a properly vertex-colored digraph and let
$\mathscr{V}=\{V_1,\dots,V_k\}$ be a partition of $V$ with
$|\mathscr{V}|=k\ge2$. Then
\begin{equation*}
U(\mathaccent"017E{G},\mathscr{V}) = U_1(\mathaccent"017E{G},\mathscr{V}) \;\charfusion[\mathbin]{\cup}{\cdot}\;
U_2(\mathaccent"017E{G},\mathscr{V}) \;\charfusion[\mathbin]{\cup}{\cdot}\; U_3(\mathaccent"017E{G},\mathscr{V})\,.
\end{equation*}
\end{lemma}
\begin{figure}[htb]
\centering
\includegraphics[width=0.75\textwidth]{./unsat-rel-example.pdf}
\caption[]{Example for unsatisfiable relations $U(\mathaccent"017E{G},\mathscr{V})$ of a
vertex-colored digraph $(\mathaccent"017E{G}=(V,E),\sigma)$ w.r.t.\ a partition
$\mathscr{V}=\{V_1,V_2\}$ (indicated by the gray boxes). In the
middle, the set of trees $\mathscr{T}(\mathscr{V})$ is illustrated,
i.e., the triangles represent all possible phylogenetic trees on the
respective subset of leaves. On the right, the arc modifications
implied by $\mathscr{V}$ (i.e., $U(\mathaccent"017E{G},\mathscr{V})$) are illustrated
where $U_1$, $U_2$, and $U_3$ indicate the type according to
Lemma~\ref{lem:unsat-rel-charac}.}
\label{fig:unsat-rel-example}
\end{figure}
The proof of Lemma~\ref{lem:unsat-rel-charac} relates the possible cases
between $\mathscr{V}$ and the tree set $\mathscr{T}(\mathscr{V})$ in a
straightforward manner. Since it is rather lengthy it is relegated to
Appendix. Fig.~\ref{fig:unsat-rel-example} gives examples for all three
types of unsatisfiable relations, i.e., for $U_1(\mathaccent"017E{G},\mathscr{V})$,
$U_2(\mathaccent"017E{G},\mathscr{V})$, and $U_3(\mathaccent"017E{G},\mathscr{V})$. In particular, we have
$(b', a)\in U_1(\mathaccent"017E{G},\mathscr{V})$ since it is an arc in $\mathaccent"017E{G}$ but $V_2$
contains another red vertex $a'$. Moreover, $(b,c)\in U_2(\mathaccent"017E{G},\mathscr{V})$
since it is not an arc in $\mathaccent"017E{G}$ but $V_1$ does not contain another green
vertex. Finally, we have $(a,b)\in U_3(\mathaccent"017E{G},\mathscr{V})$ since it is not an
arc in $\mathaccent"017E{G}$ but $b$ is the only blue vertex in $V_1$. In the example, the
graph $(\mathaccent"017E{G}\symdiff U(\mathaccent"017E{G},\mathscr{V}))$ is already a BMG which, however, is
not true in general.
\begin{corollary}
\label{cor:U-polytime}
The set $U(\mathaccent"017E{G},\mathscr{V})$ can be computed in quadratic time.
\end{corollary}
\begin{proof}
We first compute all numbers $n_{i,A}$ of vertices in $V_i$ with a given
color $A$. This can be done in $O(|V|)$ if we do not explicitly store the
zero-entries. Now, $\sigma(y)\in\sigma(V_i)$, i.e.\ $n_{i,\sigma(y)}>0$,
can be checked in constant time, and thus, it can also be decided in
constant time whether or not a pair $(x,y)$ is contained in
$U_1(\mathaccent"017E{G},\mathscr{V})$ or $U_2(\mathaccent"017E{G},\mathscr{V})$. Since, given $y\in V_i$,
the condition $V_i[\sigma(y)]=\{y\}$ is equivalent to
$n_{i,\sigma(y)}=1$, membership in $U_3(\mathaccent"017E{G},\mathscr{V})$ can also be
decided in constant time. Checking all ordered pairs $x,y\in V$ thus
requires a total effort of $O(|V|^2)$.
\qed
\end{proof}
\begin{algorithm}[t]
\caption{General BMG editing via locally optimal steps.\newline The parts
highlighted in color produce a tree $T$ explaining the edited graph
$(\vec G^*,\sigma)$. If the tree is not needed, these steps can be
omitted. The method for choosing the partition $\mathscr{V}$ (framed
box) determines different variants of the algorithm.}
\label{alg:general-local-optimal}
\newcommand{\treeline}[1]{\begingroup\color{blue}#1\endgroup}
\DontPrintSemicolon
\SetKwFunction{FRecurs}{void FnRecursive}%
\SetKwFunction{FRecurs}{Edit}
\SetKwProg{Fn}{Function}{}{}
\KwIn{Properly colored digraph $(\mathaccent"017E{G}=(V, E),\sigma)$}
\KwOut{BMG $(\mathaccent"017E{G}^*,\sigma)$.}
\BlankLine
initialize $(\mathaccent"017E{G}^*,\sigma) \leftarrow (\mathaccent"017E{G},\sigma)$ \label{line:init-G*}\;
\BlankLine
\Fn{\FRecurs{$V'$}}{
\uIf{$|V'|>1$}{
$\mathscr{V} \leftarrow$
\framebox{suitably chosen partition of $V'$ with $|\mathscr{V}|\ge2$}
\label{line:min-cost}\;
$\mathaccent"017E{G}^{*} \leftarrow \mathaccent"017E{G}^{*} \, \triangle \,
U(\mathaccent"017E{G}^{*}[V'],\mathscr{V})$ \label{line:apply-edits} \;
\treeline{create a tree $T'$ with root $\rho'$\;}
\ForEach{$V_i \in \mathscr{V}$}{
\treeline{\SetNlSty{bfseries}{\color{black}}{} attach the tree}
\FRecurs{$V_i$} \treeline{to $\rho'$}\;
}
\treeline{\Return $T'$\;}
}
\treeline{
\Else{
\Return a tree with the single element in $V'$ as root\;
}}
}
\BlankLine
\treeline{\SetNlSty{bfseries}{\color{black}}{} $T \leftarrow$}
\FRecurs{$V(\mathaccent"017E{G})$}\;
\Return $(\mathaccent"017E{G}^*,\sigma)$ \treeline{ and $T$} \label{line:return-edited}\;
\end{algorithm}
Our discussion so far suggests a recursive top-down approach, made precise
in Alg.~\ref{alg:general-local-optimal}. In each step, one determines a
``suitably chosen'' partition $\mathscr{V}$ and then recurses on the
subgraphs of the edited graph $\mathaccent"017E{G}^* \triangle
U(\mathaccent"017E{G}^*[V'],\mathscr{V})$. More details on such suitable partitions
$\mathscr{V}$ will be given in Thm.~\ref{thm:algo-consistent} below. The
parts in the algorithm highlighted in color can be omitted. They are
useful, however, if one is also interested in a tree $(T,\sigma)$ that
explains the editing result $(\mathaccent"017E{G}^*,\sigma)$ and to show that
$(\mathaccent"017E{G}^*,\sigma)$ is indeed a BMG (see below).
Alg.~\ref{alg:general-local-optimal} is designed to accumulate the
edit sets in each step, Line~\ref{line:apply-edits}. In particular, the
total edit cost and the scores $c(\mathaccent"017E{G}^*[V'],\mathscr{V})$ are closely tied
together, which follows from the following result:
\begin{lemma}
All edit sets $U(\mathaccent"017E{G}^*[V'],\mathscr{V})$ constructed in
Alg.~\ref{alg:general-local-optimal} are pairwise disjoint.
\label{lem:independent-U}
\end{lemma}
The proof of Lemma~\ref{lem:independent-U} and a technical result on which
it relies can be found in the Appendix. As an immediate consequence of
Lemma~\ref{lem:independent-U}, we have
\begin{corollary}
The edit cost of Alg.~\ref{alg:general-local-optimal} is the sum of
the \textit{UR}\xspace-costs $c(\mathaccent"017E{G}^*[V'],\mathscr{V})$ in each recursion step.
\label{cor:sum-c}
\end{corollary}
It is important to note that the edits $U(\mathaccent"017E{G}^*[V'],\mathscr{V})$ must be
applied immediately in each step (cf.\ Line~\ref{line:apply-edits} in
Alg.~\ref{alg:general-local-optimal}). In particular,
Lemma~\ref{lem:independent-U} and Cor.~\ref{cor:sum-c} pertain to the
partitioning of the edited graph $\mathaccent"017E{G}^*$, not to the original graph $\mathaccent"017E{G}$.
\begin{theorem}
\label{thm:algo-tree}
Every pair of edited graph $(\mathaccent"017E{G}^*,\sigma)$ and tree $T$ produced as
output by Alg.~\ref{alg:general-local-optimal} satisfies
$(\mathaccent"017E{G}^*,\sigma)=\mathaccent"017E{G}(T,\sigma)$. In particular, $(\mathaccent"017E{G}^*,\sigma)$ is a BMG.
\end{theorem}
\begin{proof}
By construction, the tree $T$ is phylogenetic and there is a
one-to-one correspondence between the vertices $u\in V(T)$ and the
recursion steps, which operate on the sets $V'=L(T(u))$. If
$|V'|\ge 2$ (or, equivalently, $u$ is an inner vertex of $T$), we
furthermore have $\mathscr{V}=\{L(T(v)) \mid v\in\mathsf{child}_{T}(u)\}$ for the
partition $\mathscr{V}$ of $V'$ chosen in that recursion step. In the
following, we denote by $(\mathaccent"017E{G}^*,\sigma)$ the graph during the editing
process, and by $(\mathaccent"017E{G},\sigma)$ the input graph, i.e., as in
Alg.~\ref{alg:general-local-optimal}. For brevity, we write
$E^*$ for the arc set of the final edited graph and
$E^T\coloneqq E(\mathaccent"017E{G}(T,\sigma))$.
Let us assume, for contradiction, that there exists (a)
$(x,y)\in E^*\setminus E^T\ne\emptyset$, or (b)
$(x,y)\in E^T\setminus E^*\ne\emptyset$. In either case, we set
$u\coloneqq \lca_T(x,y)$ and consider the recursion step on
$V'\coloneqq L(T(u))$ with the corresponding partition
$\mathscr{V}\coloneqq\{L(T(v)) \mid v\in\mathsf{child}_{T}(u)\}$ chosen for $V'$.
Note that $x\ne y$, and thus $u\in V^0(T)$. Moreover, let $v_x$ be the
child of $u$ such that $x\preceq_{T} v_x$, and
$V_x\coloneqq L(T(v_x))\in \mathscr{V}$.
Case~(a): $(x,y)\in E^*\setminus E^T\ne\emptyset$. \newline Since
$(x,y)\notin E^T$ and by the definition of best matches, there must be a
vertex $y'\in V_x$ of color $\sigma(y)$ such that
$\lca_T(x,y')\prec_T\lca_{T}(x,y)=u$, and thus
$\sigma(y)\in \sigma(V_x)$. Moreover, we have $V_x\in \mathscr{V}$,
$x\in V_x$ and $y\in V'\setminus V_x$. Two subcases need to be
considered, depending on whether or not $(x,y)$ is an arc in $\mathaccent"017E{G}^*$ at
the beginning of the recursion step. In the first case, the
arguments above imply that $(x,y)\in U_1(\mathaccent"017E{G}^*[V'], \mathscr{V})$,
and thus, $(x,y)\in U(\mathaccent"017E{G}^*[V'], \mathscr{V})$ by
Lemma~\ref{lem:unsat-rel-charac}. Hence, we delete the arc $(x,y)$ in
this step. In the second case, it is an easy task to verify that
none of the definitions of $U_1(\mathaccent"017E{G}^*[V'], \mathscr{V})$,
$U_2(\mathaccent"017E{G}^*[V'], \mathscr{V})$, and $U_3(\mathaccent"017E{G}^*[V'], \mathscr{V})$ matches
for $(x,y)$. Since this step is clearly the last one in the
recursion hierarchy that can affect the (non-)arc $(x,y)$, it follows
for both subcases that $(x,y)\notin E^*$; a contradiction.
Case~(b): $(x,y)\in E^T\setminus E^*\ne\emptyset$. \newline Since
$(x,y)\in E^T$ and by the definition of best matches, there cannot be a
vertex $y'\in V_x$ of color $\sigma(y)$ such that
$\lca_T(x,y')\prec_T\lca_{T}(x,y)=u$, and thus
$\sigma(y)\notin \sigma(V_x)$. Moreover, we have $V_x\in \mathscr{V}$,
$x\in V_x$ and $y\in V'\setminus V_x$. Again, two subcases need to
be distinguished depending on whether or not $(x,y)$ is an arc in
$\mathaccent"017E{G}^*$ at the beginning of the recursion step. In the first case,
the arguments above make it easy to verify that none of the
definitions of $U_1(\mathaccent"017E{G}^*[V'], \mathscr{V})$,
$U_2(\mathaccent"017E{G}^*[V'], \mathscr{V})$, and $U_3(\mathaccent"017E{G}^*[V'], \mathscr{V})$ matches
for $(x,y)$. In the second case, we obtain
$(x,y)\in U_2(\mathaccent"017E{G}^*[V'], \mathscr{V})$, and thus,
$(x,y)\in U(\mathaccent"017E{G}^*[V'], \mathscr{V})$ by
Lemma~\ref{lem:unsat-rel-charac}. Hence, we insert the arc $(x,y)$ in
this step. As before, the (non-)arc $(x,y)$ remains unaffected in any
deeper recursion step. Therefore, we have $(x,y)\in E^*$ in both
subcases; a contradiction.
Finally, $(\mathaccent"017E{G}^*,\sigma)=\mathaccent"017E{G}(T,\sigma)$ immediately implies that
$(\mathaccent"017E{G}^*,\sigma)$ is a BMG.
\qed
\end{proof}
\begin{figure}[htb]
\centering
\includegraphics[width=0.85\textwidth]{./global-vs-local-optimum.pdf}
\caption[]{(A) Example for a colored digraph $(\mathaccent"017E{G},\sigma)$ in which the
``locally'' optimal (first) split does not result in a global optimal
BMG editing. The minimal \textit{UR}\xspace-cost equals $3$ and is
attained only for the partition
$\mathscr{V}=\{ \{a_1,a_2,a_3,b_1,b_2,b_3\},
\{a_4,a_5,a_6,b_4,b_5,b_6\}\}$, which was verified by full enumeration
of all partitions and Lemma~\ref{lem:unsat-rel-charac}. For this
partition, $U(\mathaccent"017E{G},\mathscr{V})$ comprises the three purple arcs. (B)
The two (isomorphic) induced subgraphs obtained by applying the locally
optimal partition $\mathscr{V}$. Each of them has a (global) optimal
BMG editing cost of $4$. Therefore, the overall symmetric difference
of an edited graph (using the initial split $\mathscr{V}$ as specified)
comprises at least $c(\mathaccent"017E{G},\mathscr{V})+2\cdot 4=11$ arcs. (C) An
optimal editing removes the 8 green arcs and results in a digraph that
is explained by the tree in (D). The optimality of this solution was
verified using an implementation of the ILP formulation for BMG editing
given in \cite{Schaller:20y}.}
\label{fig:global-vs-local-optimum}
\end{figure}
Cor.~\ref{cor:sum-c} suggests a greedy-like ``local'' approach. In each
step, the partition $\mathscr{V}$ is chosen to minimize the score
$c(\mathaccent"017E{G}, \mathscr{V})$ in Line~\ref{line:min-cost}. The example in
Fig.~\ref{fig:global-vs-local-optimum} shows, however, that the greedy-like
choice of $\mathscr{V}$ does not necessarily yield a globally optimal
edit set.
In order to identify arcs that must be contained in every edit set, we first
clarify the relationship between the partitions $\mathfrak{P}_{\ge 2}$ on
$V$ and the partitions defined by the phylogenetic trees on $V$.
\begin{lemma}
\label{lem:tree-sets-equal}
Let $V$ be a set with $|V|\ge 2$. Let $\mathfrak{P}_{\ge 2}$ be the set
of all partitions $\mathscr{V}$ of $V$ with $|\mathscr{V}|\ge2$. Then
the set $\mathscr{T}_V$ of all phylogenetic trees with leaf set $V$
satisfies
$\mathscr{T}_V= \bigcup_{\mathscr{V}\in\mathfrak{P}_{\ge 2}}
\mathscr{T}(\mathscr{V})$.
\end{lemma}
\begin{proof}
For every $\mathscr{V}\in\mathfrak{P}_{\ge 2}$,
$\mathscr{T}(\mathscr{V})$ is a set of phylogenetic trees on $V$. Hence,
we conclude
$\bigcup_{\mathscr{V}\in\mathfrak{P}_{\ge 2}} \mathscr{T}(\mathscr{V})
\subseteq \mathscr{T}_V$. Conversely, assume that $T\in \mathscr{T}_V$.
Since $T$ (with root $\rho_T$) is a phylogenetic tree and has at least
two leaves, we have $|\mathsf{child}_{T}(\rho_T)|\ge 2$. Together with
$L(T(\rho_T))=L(T)=V$, this implies
$\mathscr{V}^*\coloneqq\{L(T(v)) \mid v\in \mathsf{child}_{T}(\rho_T)\} \in
\mathfrak{P}_{\ge 2}$. In particular, $T$ satisfies
$T\in\mathscr{T}(\mathscr{V}^*)$ for some
$\mathscr{V}^* \in \mathfrak{P}_{\ge 2}$, and is therefore contained in
$\bigcup_{\mathscr{V}\in\mathfrak{P}_{\ge 2}} \mathscr{T}(\mathscr{V})$.
\qed
\end{proof}
Using Lemma~\ref{lem:tree-sets-equal} and given that $|V|\ge 2$, we can
express the set of relations
that are unsatisfiable for every partition as follows
\begin{equation}
\begin{split}
\bigcap_{\mathscr{V}\in\mathfrak{P}_{\ge2}} U(\mathaccent"017E{G},\mathscr{V})
&= \bigcap_{\mathscr{V}\in\mathfrak{P}_{\ge2}}
\left( \bigcap_{T\in \mathscr{T}(\mathscr{V})} U(\mathaccent"017E{G},T) \right)
= \bigcap_{T\in\bigcup_{\mathscr{V}\in\mathfrak{P}_{\ge 2}}
\mathscr{T}(\mathscr{V})} U(\mathaccent"017E{G},T)\\
&= \bigcap_{T\in\mathscr{T}_V} U(\mathaccent"017E{G},T) = U^*(\mathaccent"017E{G})\;,
\end{split}
\label{eq:U*U-P}
\end{equation}
i.e., it coincides with the set of relations that are unsatisfiable for
every phylogenetic tree, and thus part of every edit set.
Note that $U^*(\mathaccent"017E{G})$ is trivially empty if $|V|<2$.
We next show that
$U^*(\mathaccent"017E{G})$ can be computed without considering the partitions of $V$
explicitly.
\begin{theorem}
\label{thm:trivial-unsat-rel}
Let $(\mathaccent"017E{G}=(V,E),\sigma)$ be a properly vertex-colored digraph with
$|V|\ge 2$ then
\begin{equation}
U^*(\mathaccent"017E{G}) = \left\{ (x,y) \mid
(x,y)\notin E,\; x\ne y,\; V[\sigma(y)]=\{y\} \right\}.
\label{eq:U-equalities}
\end{equation}
\end{theorem}
\begin{proof}
First note that $|V|\ge 2$ ensures that
$\mathfrak{P}_{\ge 2}\ne\emptyset$. Moreover, since $|\mathscr{V}|\ge2$
for any $\mathscr{V}\in\mathfrak{P}_{\ge 2}$, the sets
$\mathscr{T}(\mathscr{V})$ are all non-empty as well. With the
abbreviation $\hat U(\mathaccent"017E{G})$ for the right-hand side of
Eq.~(\ref{eq:U-equalities}), we show that
$\hat U(\mathaccent"017E{G})= \bigcap_{\mathscr{V}\in\mathfrak{P}_{\ge2}}
U(\mathaccent"017E{G},\mathscr{V})$ which by Eq.~(\ref{eq:U*U-P}) equals $U^*(\mathaccent"017E{G})$.
Suppose first that $(x,y)\in \hat U(\mathaccent"017E{G})$. Then $x\ne y$ and
$V[\sigma(y)]=\{y\}$ imply that $\sigma(x)\ne\sigma(y)$. This together
with the facts that (i) $y$ is the only vertex of its color in $V$, and
(ii) $L(T)=V$ for each $T\in \mathscr{T}(\mathscr{V})$ and any
$\mathscr{V}\in\mathfrak{P}_{\ge 2}$ implies that $y$ is a best match of
$x$ in every such tree $T$, i.e.\ $(x,y)\in E(\mathaccent"017E{G}(T,\sigma))$. Since in
addition $(x,y)\notin E$ by assumption, we conclude that
$(x,y)\in U^{*}(\mathaccent"017E{G})$.
Now suppose that $(x,y)\in U^{*}(\mathaccent"017E{G})$. Observe that
$\sigma(x)\ne\sigma(y)$ (and thus $x\ne y$) as a consequence of
Def.~\ref{def:unsat-relations} and the fact that $(\mathaccent"017E{G},\sigma)$ and all
BMGs are properly colored. If $V=\{x,y\}$ and thus $\{\{x\},\{y\}\}$ is
the only partition in $\mathfrak{P}_{\ge 2}$, the corresponding unique
tree $T$ consists of $x$ and $y$ connected to the root. In this case, we
clearly have $(x,y)\in E(\mathaccent"017E{G}(T,\sigma))$ since $\sigma(x)\ne\sigma(y)$.
On the other hand, if $\{x,y\}\subsetneq V$, then we can find a partition
$\mathscr{V}\in \mathfrak{P}_{\ge 2}$ such that $V_{i}=\{x,y\}$ for some
$V_{i}\in\mathscr{V}$. In this case, every tree
$T\in\mathscr{T}(\mathscr{V})$ has a vertex $v_{i}\in\mathsf{child}_{T}(\rho_T)$
with the leaves $x$ and $y$ as its single two children. Clearly,
$(x,y)\in E(\mathaccent"017E{G}(T,\sigma))$ holds for any such tree. In summary, there
always exists a partition $\mathscr{V}\in\mathfrak{P}_{\ge 2}$ such that
$(x,y)\in E(\mathaccent"017E{G}(T,\sigma))$ for some tree $T\in\mathscr{T}(\mathscr{V})$.
Therefore, by
$(x,y)\in\bigcap_{\mathscr{V}\in\mathfrak{P}_{\ge 2}} U(\mathaccent"017E{G},\mathscr{V})$
and Def.~\ref{def:unsat-relations}, we conclude that $(x,y)\notin E$. In
order to obtain $(x,y)\in \hat U(\mathaccent"017E{G})$, it remains to show that
$V[\sigma(y)]=\{y\}$. Since $(x,y)\notin E$ and
$(x,y)\in\bigcap_{\mathscr{V}\in\mathfrak{P}_{\ge 2}} U(\mathaccent"017E{G},\mathscr{V})$,
it must hold that $(x,y)\in E(\mathaccent"017E{G}(T,\sigma))$ for all
$T\in\mathscr{T}(\mathscr{V})$ and all
$\mathscr{V}\in\mathfrak{P}_{\ge 2}$. Now assume, for contradiction,
that there is a vertex $y'\ne y$ of color $\sigma(y')=\sigma(y)$. Since
$\sigma(x)\ne\sigma(y)$, the vertices $x,y,y'$ must be pairwise distinct.
Hence, we can find a partition $\mathscr{V}\in \mathfrak{P}_{\ge 2}$ such
that $V_{i}=\{x,y'\}$ for some $V_{i}\in\mathscr{V}$. In this case,
every tree $T\in\mathscr{T}(\mathscr{V})$ has a vertex
$v_{i}\in\mathsf{child}_{T}(\rho_T)$ with only the leaves $x$ and $y'$ as its
children. Clearly, $\lca_T(x,y')=v_{i}\prec_{T}\rho_{T}=\lca_{T}(x,y)$,
and thus $(x,y)\notin E(\mathaccent"017E{G}(T,\sigma))$; a contradiction. Therefore, we
conclude that $y$ is the only vertex of its color in $V$, and hence,
$(x,y)\in \hat U(\mathaccent"017E{G})$. In summary, therefore, we have
$U^{*}(\mathaccent"017E{G})=\hat U(\mathaccent"017E{G})$. \qed
\end{proof}
As a consequence of Thm.~\ref{thm:trivial-unsat-rel} and by similar
arguments as in the proof of Cor.~\ref{cor:U-polytime}, we observe
\begin{corollary}
The set $U^*(\mathaccent"017E{G})$ can be computed in quadratic time.
\end{corollary}
By Thm.~\ref{thm:trivial-unsat-rel}, $U^*(\mathaccent"017E{G})$ contains only non-arcs,
more precisely, missing arcs pointing towards a vertex that is the
only one of its color and thus, by definition, a best match of every
other vertex irrespective of the details of the gene tree. By definition,
furthermore, $U^*(\mathaccent"017E{G})$ is a subset of every edit set for
$(\mathaccent"017E{G},\sigma)$. We
therefore have the lower bound
\begin{equation}
|U^*(\mathaccent"017E{G})| \le c(\mathaccent"017E{G},\mathscr{V})
\end{equation}
for every $\mathscr{V}\in\mathfrak{P}_{\ge2}$.
The following result shows that if $(\mathaccent"017E{G},\sigma)$ is a BMG, then a suitable
partition $\mathscr{V}$ can be chosen such that
$c(\mathaccent"017E{G},\mathscr{V})=|U^*(\mathaccent"017E{G})|=0$.
\begin{lemma}
\label{lem:cost-0-BMG}
Let $(\mathaccent"017E{G}=(V,E),\sigma)$ be a BMG with $|V|\ge 2$ and $\mathscr{V}$ be the
connected components of the Aho graph $[\mathscr{R}(\mathaccent"017E{G},\sigma), V]$. Then
the partition $\mathscr{V}$ of $V$ satisfies $|\mathscr{V}|\ge 2$ and
$c(\mathaccent"017E{G},\mathscr{V})=0$.
\end{lemma}
\begin{proof}
Since $(\mathaccent"017E{G},\sigma)$ is a BMG, we can apply Prop.~\ref{prop:BMG-charac} to
conclude that $\mathscr{R}\coloneqq\mathscr{R}(\mathaccent"017E{G},\sigma)$ is consistent
and that $(T,\sigma)\coloneqq (\Aho(\mathscr{R}, V),\sigma)$ explains
$(\mathaccent"017E{G},\sigma)$, i.e., $\mathaccent"017E{G}(T,\sigma)=(\mathaccent"017E{G},\sigma)$. Hence,
$U(\mathaccent"017E{G},T)=\emptyset$. From $|V|\ge 2$ and consistency of $\mathscr{R}$,
it follows that $[\mathscr{R}, V]$ has at least two connected components
\cite{Aho:81}, and thus, by construction, $|\mathscr{V}|\ge 2$.
Moreover, we clearly have $T\in \mathscr{T}(\mathscr{V})$ by the
construction of $T$ via \texttt{BUILD}. Together with
$U(\mathaccent"017E{G},T)=\emptyset$, the latter implies $U(\mathaccent"017E{G},\mathscr{V})=\emptyset$,
and thus $c(\mathaccent"017E{G},\mathscr{V})=0$. \qed
\end{proof}
\begin{lemma}
\label{lem:cost-0-subgraphs-BMG}
Let $(\mathaccent"017E{G}=(V,E),\sigma)$ be a BMG, and $\mathscr{V}$ a partition of $V$
such that $c(\mathaccent"017E{G},\mathscr{V})=0$. Then the induced subgraph
$(\mathaccent"017E{G}[V'],\sigma_{|V'})$ is a BMG for every $V'\in\mathscr{V}$.
\end{lemma}
\begin{proof}
Set $\mathscr{R}\coloneqq\mathscr{R}(\mathaccent"017E{G},\sigma)$ and
$\mathscr{F}\coloneqq\mathscr{F}(\mathaccent"017E{G},\sigma)$ for the sets of informative
and forbidden triples of $(\mathaccent"017E{G},\sigma)$, respectively. Since
$(\mathaccent"017E{G},\sigma)$ is a BMG, we can apply Prop.~\ref{prop:BMG-charac} to
conclude that $(\mathscr{R},\mathscr{F})$ is consistent. Now we
choose an arbitrary set $V'\in\mathscr{V}$ and set
$(\mathaccent"017E{G}',\sigma')\coloneqq (\mathaccent"017E{G}[V'],\sigma_{|V'})$. By
Obs.~\ref{obs:R-restriction}, we obtain
$\mathscr{R}(\mathaccent"017E{G}',\sigma')=\mathscr{R}_{|V'}$ and
$\mathscr{F}(\mathaccent"017E{G}',\sigma')=\mathscr{F}_{|V'}$. This together with the fact
that $\mathscr{R}_{|V'}\subseteq\mathscr{R}$ and
$\mathscr{F}_{|V'}\subseteq\mathscr{F}$ and Obs.~\ref{obs:R-F-subsets}
implies that
$(\mathscr{R}_{|V'},\mathscr{F}_{|V'})=(\mathscr{R}(\mathaccent"017E{G}',\sigma'),
\mathscr{F}(\mathaccent"017E{G}',\sigma'))$ is consistent.
By Prop.~\ref{prop:BMG-charac}, it remains to show that $(\mathaccent"017E{G}',\sigma')$
is sf-colored to prove that it is a BMG. To this end, assume for
contradiction that there is a vertex $x\in V'$ and a color
$s\in\sigma(V')$ such that $x$ has no out-neighbor of color
$s\ne\sigma(x)$ in $V'$. However, since the color $s$ is contained in
$\sigma(V)$ and $(\mathaccent"017E{G},\sigma)$ is a BMG, and thus sf-colored, we conclude
that there must be a vertex $y\in V\setminus V'$ of color $s$ such that
$(x,y)\in E$. In summary, we obtain $(x,y)\in E$, $x\in V'$,
$y\in V\setminus V'$ and $\sigma(y)=s\in\sigma(V')$. Thus, we have
$(x,y)\in U_1(\mathaccent"017E{G},\mathscr{V})$. Hence, Lemma~\ref{lem:unsat-rel-charac}
implies that $(x,y)\in U(\mathaccent"017E{G},\mathscr{V})$ and, hence,
$c(\mathaccent"017E{G},\mathscr{V})>0$; a contradiction. Therefore, $(\mathaccent"017E{G}',\sigma')$ must
be sf-colored, which concludes the proof. \qed
\end{proof}
Lemma \ref{lem:cost-0-BMG} and \ref{lem:cost-0-subgraphs-BMG} allow us to
choose the partition $\mathscr{V}$ in each step of
Alg.~\ref{alg:general-local-optimal} in such a way that
Alg.~\ref{alg:general-local-optimal} is consistent, i.e., BMGs remain
unchanged.
\begin{theorem}
\label{thm:algo-consistent}
Alg.~\ref{alg:general-local-optimal} is consistent if, in each step
on $V'$ with $|V'|\ge 2$, the partition $\mathscr{V}$ in
Line~\ref{line:min-cost} is chosen according to one of the following
rules:
\begin{enumerate}
\item $\mathscr{V}$ has minimal \textit{UR}\xspace-cost among all possible partitions
$\mathscr{V}'$ of $V'$ with $|\mathscr{V}'|\ge 2$.
\item If the Aho graph $[\mathscr{R}(\mathaccent"017E{G}^*[V'],\sigma_{|V'}),V']$ is
disconnected with the set of connected components $\mathscr{V}_{\Aho}$,
and moreover $c(\mathaccent"017E{G}^*[V'],\mathscr{V}_{\Aho})=0$, then
$\mathscr{V}=\mathscr{V}_{\Aho}$.
\end{enumerate}
\end{theorem}
\begin{proof}
We have to show that the final edited graph $(\mathaccent"017E{G}^*,\sigma)$ returned in
Line~\ref{line:return-edited} equals the input graph $(\mathaccent"017E{G}=(V,E),\sigma)$
whenever $(\mathaccent"017E{G},\sigma)$ already is a BMG, i.e., nothing is edited. Thus
suppose that $(\mathaccent"017E{G},\sigma)$ is a BMG and first consider the top-level
recursion step on $V$ (where initially $\mathaccent"017E{G}^*=\mathaccent"017E{G}$ still holds at
Line~\ref{line:init-G*}). If $|V|=1$, neither $(\mathaccent"017E{G},\sigma)$ nor
$(\mathaccent"017E{G}^*,\sigma)$ contain any arcs, and thus, the edit cost is trivially
zero. Now suppose $|V|\ge 2$. Since $(\mathaccent"017E{G},\sigma)$ is a BMG,
Lemma~\ref{lem:cost-0-BMG} guarantees the existence of a partition
$\mathscr{V}$ satisfying $c(\mathaccent"017E{G},\mathscr{V})=0$, in particular, the
connected components $\mathscr{V}_{\Aho}$ of the Aho graph
$[\mathscr{R}(\mathaccent"017E{G},\sigma), V]$ form such a partition. Hence, for both
rules~(1) and~(2), we choose a partition $\mathscr{V}$ with (minimal)
\textit{UR}\xspace-cost $c(\mathaccent"017E{G},\mathscr{V})=0$. Now,
Lemma~\ref{lem:cost-0-subgraphs-BMG} implies that the induced subgraph
$(\mathaccent"017E{G}[V'],\sigma_{|V'})$ is a BMG for every $V'\in\mathscr{V}$. Since we
recurse on these subgraphs, we can repeat the arguments above along the
recursion hierarchy to conclude that the \textit{UR}\xspace-cost
$c(\mathaccent"017E{G}^*[V'],\mathscr{V}')$ vanishes in every recursion step. By
Cor.~\ref{cor:sum-c}, the total edit cost of
Alg.~\ref{alg:general-local-optimal} is the sum of the \textit{UR}\xspace-costs
$c(\mathaccent"017E{G}^*[V'],\mathscr{V}')$ in each recursion step, and thus, also zero.
Therefore, we conclude that we still have $(\mathaccent"017E{G}^*,\sigma)=(\mathaccent"017E{G},\sigma)$ in
Line~\ref{line:return-edited}. \qed
\end{proof}
By Thm.~\ref{thm:algo-consistent}, Alg.~\ref{alg:general-local-optimal} is
consistent whenever the choice of $\mathscr{V}$ minimizes the \textit{UR}\xspace-cost of
$\mathscr{V}$ in each step. We shall see in Sec.~\ref{sect:UR-NP} that
minimizing $c(\mathaccent"017E{G},\mathscr{V})$ is a difficult optimization problem in
general. Therefore, a good heuristic will be required for this step. This,
however, may not guarantee consistency of
Alg.~\ref{alg:general-local-optimal} in general. The second rule in
Thm.~\ref{thm:algo-consistent} provides a remedy: the Aho graph
$[\mathscr{R}(\mathaccent"017E{G}^*[V'],\sigma_{|V'}), V']$ can be computed
efficiently. Whenever $[\mathscr{R}(\mathaccent"017E{G}^*[V'],\sigma_{|V'}), V']$ is not
connected, the partition $\mathscr{V}_{\Aho}$ defined by the connected
components $[\mathscr{R}(\mathaccent"017E{G}^*[V'],\sigma_{|V'}), V']$ is chosen provided it
has \textit{UR}\xspace-cost zero. This procedure is effectively a generalization of the
algorithm \texttt{BUILD} using as input the set of informative triples
$\mathscr{R}(\mathaccent"017E{G},\sigma)$ of a properly vertex-colored graph
$(\mathaccent"017E{G},\sigma)$. If $(\mathaccent"017E{G},\sigma)$ is already a BMG, then the recursion in
Alg.~\ref{alg:general-local-optimal} is exactly the same as in
\texttt{BUILD}: it recurses on the connected components of the Aho graph
(cf.\ Prop.~\ref{prop:BMG-charac}). We can summarize this discussion as
\begin{corollary}
$(\mathaccent"017E{G},\sigma)$ is a BMG if and only if, in every step of the
\texttt{BUILD} algorithm operating on $\mathscr{R}(\mathaccent"017E{G},\sigma)_{|V'}$ and
$V'$, either $|V'|=1$, or $c(\mathaccent"017E{G}^*[V'], \mathscr{V}_{\Aho})=0$ for the
connected component partition $\mathscr{V}_{\Aho}$ of the disconnected
Aho graph $[\mathscr{R}(\mathaccent"017E{G}^*[V'],\sigma_{|V'}), V']$.
\end{corollary}
For recursion steps in which the Aho graph
$[\mathscr{R}(\mathaccent"017E{G}^*[V'],\sigma_{|V'}), V']$ is connected, and possibly also
in steps with non-zero \textit{UR}\xspace-cost, another (heuristic) rule has to be
employed. As a by-product, we obtain an approach for the case that
$\mathscr{R}(\mathaccent"017E{G},\sigma)$ is consistent: Following \texttt{BUILD} yields the
approximation $\mathaccent"017E{G}(\Aho(\mathscr{R}(\mathaccent"017E{G},\sigma),V(\mathaccent"017E{G})),\sigma)$ as a natural
choice.
\section{Binary-explainable BMGs}
\label{sect:beBMG}
Phylogenetic trees are often binary. Multifurcations are in many cases --
but not always -- the consequence of insufficient data
\cite{DeSalle:94,Sayyari:18,Schaller:20p}. It is therefore of practical
interest to consider BMGs that can be explained by a binary tree:
\begin{definition}
A properly colored digraph $(\mathaccent"017E{G},\sigma)$ is a \emph{binary-explainable best
match graph} (\emph{beBMG}) if there is a binary tree $T$ such that
$\mathaccent"017E{G}(T,\sigma)=(\mathaccent"017E{G},\sigma)$.
\end{definition}
Correspondingly, it is of interest to edit a properly colored digraph to a
beBMG, which translates to the following decision problem:
\begin{problem}[\PROBLEM{$\ell$-BMG Editing restricted to
Binary-Explainable\newline
Graphs (EBEG)}]\ \\
\label{prblm:ell-bmg-EBEG}
\begin{tabular}{ll}
\emph{Input:} & A properly $\ell$-colored digraph $(\mathaccent"017E{G} =(V,E),\sigma)$
and an integer $k$.\\
\emph{Question:} & Is there a subset $F\subseteq V\times V \setminus
\{(v,v)\mid v\in V\}$ such
that $|F|\leq k$ and\\ & $(\mathaccent"017E{G}\symdiff F,\sigma)$
is a binary-explainable $\ell$-BMG?
\end{tabular}
\end{problem}
We call the corresponding completion and deletion problem
\PROBLEM{$\ell$-BMG CBEG} and \PROBLEM{$\ell$-BMG DBEG}, respectively. As
their more general counterparts, all three variants are NP-complete as well,
cf.\ \cite[Cor.~6.2]{Schaller:20y} and \cite[Thm.~5]{Schaller:20p}.
Since the recursive partitioning in Alg.~\ref{alg:general-local-optimal}
defines a tree that explains the edited BMG, see Thm.~\ref{thm:algo-tree},
it is reasonable to restrict the optimization of $\mathscr{V}$ in
Line~\ref{line:min-cost} to bipartitions. The problem still remains hard,
however, since the corresponding decision problem (problem \PROBLEM{BPURC}
in Sec.~\ref{sect:UR-NP}) is NP-complete as shown in
Thm.~\ref{thm:BPURC-NPc} below. Similar to BMGs in general, beBMGs have a
characterization in terms of informative triples:
\begin{proposition}{\cite[Thm.~3.5]{Schaller:20p}}
\label{prop:Rbin-MAIN}
A properly vertex-colored graph $(\mathaccent"017E{G},\sigma)$ with vertex set $V$ is
binary-explainable if and only if (i) $(\mathaccent"017E{G},\sigma)$ is sf-colored, and
(ii) the triple set $\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G},\sigma)$ is consistent. In this case, the
BMG $(\mathaccent"017E{G},\sigma)$ is explained by every refinement of the \emph{binary
refinable tree} $(\Aho(\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G},\sigma), V), \sigma)$.
\end{proposition}
Using Prop.~\ref{prop:Rbin-MAIN}, we can apply analogous arguments as in
the proof of Lemma~\ref{lem:cost-0-BMG} for $\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G},\sigma)$ instead of
$\mathscr{R}(\mathaccent"017E{G},\sigma)$ to obtain
\begin{corollary}
\label{cor:cost-0-beBMG}
Let $(\mathaccent"017E{G}=(V,E),\sigma)$ be a beBMG with $|V|\ge 2$ and $\mathscr{V}$ be
the connected components of the Aho graph $[\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G},\sigma), V]$. Then
the partition $\mathscr{V}$ of $V$ satisfies $|\mathscr{V}|\ge 2$ and
$c(\mathaccent"017E{G},\mathscr{V})=0$.
\end{corollary}
Since a beBMG $(\mathaccent"017E{G},\sigma)$ is explained by every refinement of the Aho
tree constructed from $\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G},\sigma)$ (cf.\ Prop.~\ref{prop:Rbin-MAIN}),
we can obtain a slightly more general result.
\begin{lemma}
\label{lem:cost-0-beBMG-coarsements}
Let $(\mathaccent"017E{G}=(V,E),\sigma)$ be a beBMG with $|V|\ge 2$ and $\mathscr{V}$ be
the connected components of the Aho graph $[\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G},\sigma),V]$. Then,
every coarse-graining $\mathscr{V}'$ of $\mathscr{V}$ with
$|\mathscr{V}'|\ge 2$ satisfies $c(\mathaccent"017E{G},\mathscr{V}')=0$.
\end{lemma}
\begin{proof}
First note that $\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G},\sigma)$ is consistent by
Prop.~\ref{prop:Rbin-MAIN} since $(\mathaccent"017E{G},\sigma)$ is a beBMG. Therefore,
$|V|\ge 2$ implies $|\mathscr{V}|\ge 2$ \cite{Aho:81}. For the trivial
coarse-graining $\mathscr{V}'=\mathscr{V}$, Cor.~\ref{cor:cost-0-beBMG}
already implies the statement. Now assume $\mathscr{V}'\ne\mathscr{V}$.
Observe that the tree
$(T,\sigma) \coloneqq (\Aho(\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G},\sigma), V), \sigma)$ exists and
explains $(\mathaccent"017E{G},\sigma)$ by Prop.~\ref{prop:Rbin-MAIN}. Moreover, there
is, by construction, a one-to-one correspondence between the children $v_i$
of its root $\rho$ and the elements in $V_i\in\mathscr{V}$ given by
$L(T(v_i))=V_i$. We construct a refinement (tree) $T'$ of $T$ as follows:
Whenever we have multiple sets $V_i\in\mathscr{V}$ that are subsets of
the same set $V_j\in\mathscr{V}'$, we remove the edges $\rho v_i$ to the
corresponding vertices $v_i\in\mathsf{child}_{T}(\rho)$ in $T$, and collectively
connect these $v_i$ to a newly created vertex $w_j$. These vertices $w_j$ are
then reattached to the root $\rho$. Since $|\mathscr{V}'|\ge 2$ by
assumption, the
so-constructed tree $T'$ is still phylogenetic. Moreover, it satisfies
$\mathscr{V}'=\{L(T'(v)) \mid v\in\mathsf{child}_{T'}(\rho)\}$, and thus,
$T'\in \mathscr{T}(\mathscr{V}')$. It is a refinement of $T$ since
contraction of the edges $\rho w_j$ again yields $T$. Hence, we can
apply Prop.~\ref{prop:Rbin-MAIN} to conclude that $(T',\sigma)$ also
explains $(\mathaccent"017E{G},\sigma)$. It follows immediately that
$U(\mathaccent"017E{G},T')=\emptyset$. The latter together with
$T'\in \mathscr{T}(\mathscr{V}')$ implies $U(\mathaccent"017E{G},\mathscr{V}')=\emptyset$,
and thus $c(\mathaccent"017E{G},\mathscr{V}')=0$. \qed
\end{proof}
We are now in the position to formulate an analogue of
Thm.~\ref{thm:algo-consistent} for variants of
Alg.~\ref{alg:general-local-optimal} that aim to edit a properly-colored
digraph $(\mathaccent"017E{G},\sigma)$ to a beBMG.
\begin{theorem}
\label{thm:algo-consistent-binary}
Alg.~\ref{alg:general-local-optimal} is consistent for beBMGs
$(\mathaccent"017E{G},\sigma)$ if, in each step on $V'$ with $|V'|\ge 2$, a bipartition
$\mathscr{V}$ in Line~\ref{line:min-cost} is chosen according to one of
the following rules:
\begin{enumerate}
\item $\mathscr{V}$ has minimal \textit{UR}\xspace-cost among all possible bipartitions
$\mathscr{V}'$ of $V'$.
\item If the Aho graph $[\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G}^*[V'],\sigma_{|V'}),V']$ is
disconnected with the set of connected components $\mathscr{V}_{\Aho}$,
and moreover $c(\mathaccent"017E{G}^*[V'],\mathscr{V}_{\Aho})=0$, then $\mathscr{V}$ is
a coarse-graining of $\mathscr{V}_{\Aho}$.
\end{enumerate}
\end{theorem}
\begin{proof}
We have to show that the final edited graph $(\mathaccent"017E{G}^*,\sigma)$ returned in
Line~\ref{line:return-edited} equals the input graph $(\mathaccent"017E{G}=(V,E),\sigma)$
whenever $(\mathaccent"017E{G},\sigma)$ already is a beBMG, i.e., nothing is edited. Thus
suppose that $(\mathaccent"017E{G},\sigma)$ is a beBMG and first consider the top-level
recursion step on $V$ (where initially $\mathaccent"017E{G}^*=\mathaccent"017E{G}$ still holds at
Line~\ref{line:init-G*}). If $|V|=1$, neither $(\mathaccent"017E{G},\sigma)$ nor
$(\mathaccent"017E{G}^*,\sigma)$ contain any arcs, and thus, the edit cost is trivially
zero. Now suppose $|V|\ge 2$. Since $(\mathaccent"017E{G},\sigma)$ is a beBMG,
$\mathop{\mathscr{R}^{\textrm{B}}}\coloneqq\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G},\sigma)$ is consistent, and thus, the set of
connected components $\mathscr{V}_{\Aho}$ of the Aho graph $[\mathop{\mathscr{R}^{\textrm{B}}}, V]$
has a cardinality of at least two. If $|\mathscr{V}_{\Aho}|=2$,
$\mathscr{V}\coloneqq \mathscr{V}_{\Aho}$ is a bipartition satisfying
$c(\mathaccent"017E{G},\mathscr{V})=0$ by Cor.~\ref{cor:cost-0-beBMG}. If
$|\mathscr{V}_{\Aho}|>2$, we can find an arbitrary bipartition
$\mathscr{V}$ that is a coarsement of $\mathscr{V}_{\Aho}$. By
Lemma~\ref{lem:cost-0-beBMG-coarsements}, $\mathscr{V}$ also satisfies
$c(\mathaccent"017E{G},\mathscr{V})=0$ in this case. Hence, for both rules~(1) and~(2),
we choose a bipartition $\mathscr{V}$ with (minimal) \textit{UR}\xspace-cost
$c(\mathaccent"017E{G},\mathscr{V})=0$. Now, Lemma~\ref{lem:cost-0-subgraphs-BMG} implies
that the induced subgraph $(\mathaccent"017E{G}[V'],\sigma_{|V'})$ is a BMG for every
$V'\in\mathscr{V}$. To see that $(\mathaccent"017E{G}[V'],\sigma_{|V'})$ is also
binary-explainable, first note that
$\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G}[V'],\sigma_{|V'})=\mathop{\mathscr{R}^{\textrm{B}}}_{|V'}$ by Obs.~\ref{obs:R-restriction}.
This together with the fact that $\mathop{\mathscr{R}^{\textrm{B}}}_{|V'}\subseteq\mathop{\mathscr{R}^{\textrm{B}}}$ and
Obs.~\ref{obs:R-F-subsets} implies that $\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G}[V'],\sigma_{|V'})$ is
consistent. Moreover, Prop.~\ref{prop:BMG-charac} and
$(\mathaccent"017E{G}[V'],\sigma_{|V'})$ being a BMG together imply that
$(\mathaccent"017E{G}[V'],\sigma_{|V'})$ is sf-colored. Hence, we can apply
Prop.~\ref{prop:Rbin-MAIN} to conclude that $(\mathaccent"017E{G}[V'],\sigma_{|V'})$ is a
beBMG.
Since we recurse on the subgraphs $(\mathaccent"017E{G}[V'],\sigma_{|V'})$, which are
again beBMGs, we can repeat the arguments above along the recursion
hierarchy to conclude that the \textit{UR}\xspace-cost $c(\mathaccent"017E{G}^*[V'],\mathscr{V}')$
vanishes in every recursion step. By Cor.~\ref{cor:sum-c}, the total
edit cost of Alg.~\ref{alg:general-local-optimal} is the sum of the
\textit{UR}\xspace-costs $c(\mathaccent"017E{G}^*[V'],\mathscr{V}')$ in each recursion step, and thus,
also zero. Therefore, we conclude that we still have
$(\mathaccent"017E{G}^*,\sigma)=(\mathaccent"017E{G},\sigma)$ in Line~\ref{line:return-edited}. \qed
\end{proof}
\section{Minimizing the \textit{UR}\xspace-cost $c(\mathaccent"017E{G},\mathscr{V})$}
\label{sect:UR-NP}
The problem of minimizing $c(\mathaccent"017E{G},\mathscr{V})$ for a given properly
colored graph $(\mathaccent"017E{G},\sigma)$ corresponds to the following decision
problem.
\begin{problem}[\PROBLEM{(Bi)Partition with \textit{UR}\xspace-Cost ((B)PURC)}]\ \\
\begin{tabular}{ll}
\emph{Input:} & A properly $\ell$-colored digraph $(\mathaccent"017E{G} =(V,E),\sigma)$
and an integer $k\geq 0$.\\
\emph{Question:} & Is there a (bi)partition $\mathscr{V}$ of $V$
such that $c(\mathaccent"017E{G},\mathscr{V})\leq k$?
\end{tabular}
\end{problem}
In the Appendix, we show that \PROBLEM{(B)PURC} is NP-hard by reduction
from \PROBLEM{Set Splitting}, one of \citeauthor{Garey:79}'s
\citeyearpar{Garey:79} classical NP-complete problems.
\begin{theorem}
\label{thm:BPURC-NPc}
\PROBLEM{BPURC} is NP-complete.
\end{theorem}
Thm.~\ref{thm:algo-consistent} suggests to consider heuristics
for \PROBLEM{(B)PURC} that make use of the Aho graph in the following
manner:
\begin{enumerate}
\item Construct the Aho graph $H\coloneqq [\mathscr{R}(\mathaccent"017E{G},\sigma),V]$
based on the set of informative triples $\mathscr{R}(\mathaccent"017E{G},\sigma)$.
\item If $H$ has more than one connected component, we use the set of
connected components as the partition $\mathscr{V}$.
\item If $H$ is connected, a heuristic that operates on the Aho graph $H$
is used to find a partition $\mathscr{V}$ with small \textit{UR}\xspace-cost
$c(\mathaccent"017E{G},\mathscr{V})$.
\end{enumerate}
Plugging any algorithm of this type into Line~\ref{line:min-cost} of
Alg.~\ref{alg:general-local-optimal} reduces
the algorithm to \texttt{BUILD} if a BMG is used as
input and thus guarantees consistency (cf.\ Prop.~\ref{prop:BMG-charac}).
We note, however, that the connected components of a disconnected Aho graph
are not guaranteed to correspond to an optimal solution for
\PROBLEM{(B)PURC} in the general case.
\section{Computational experiments}
\label{sec:heur-computational-exp}
In this section, we compare different heuristics for the \PROBLEM{(B)PURC}
Problem and their performance in the context of BMG editing. Somewhat
unexpectedly, but in accordance with
Fig.~\ref{fig:global-vs-local-optimum}, our results suggest that a good (or
bad) performance of \PROBLEM{(B)PURC} is not directly linked to a good (or
bad) performance for BMG editing. Moreover, we find that, even for noisy
data, all analyzed methods are able to capture the tree structure of the
underlying ``true'' BMG at least to some extent. As we shall see, community
detection approaches in combination with the \textit{UR}\xspace-cost appear to be more
promising for BMG editing than optimal solutions of \PROBLEM{(B)PURC} alone.
\subsection{Heuristics for \PROBLEM{(B)PURC}}\label{subsec:exp_1}
\PROBLEM{(B)PURC} is a variation on graph partitioning problems. It seems
reasonable, therefore, to adapt graph partitioning algorithms for our
purposes.
\paragraph{MinCut.}
We solve the minimum edge cut problem for the connected undirected graph
$H$, i.e., we want to find a bipartition $\mathscr{V}=\{V_1,V_2\}$ such
that the number of edges between $V_1$ and $V_2$ is minimal in $H$. The
problem can be solved exactly in polynomial time using the Stoer-Wagner
algorithm \cite{Stoer:97}. Note, however, that the minimum edge cut in $H$
will in general not deliver an optimal solution of \PROBLEM{(B)PURC}.
\paragraph{Karger's algorithm} is a randomized algorithm that, in its
original form, also aims to find a minimum edge cut \cite{Karger:93}. In
brief, it merges vertices of the graph by randomly choosing and contracting
edges, until only two vertices remain, which induce a bipartition
$\mathscr{V}$ according to the vertices that were merged into them. By
repeating this process a sufficient number of times, a minimum edge cut can
be found with high probability. Here, we use the \textit{UR}\xspace-cost
$c(\mathaccent"017E{G},\mathscr{V})$ instead of the size of the edge cut as objective
function to select the best solution over multiple runs.
\paragraph{A simple greedy approach} starts with
$\mathscr{V}=\{V_1=\emptyset, V_2=V'\}$ and stepwise moves a vertex
$v\in V_2$ to $V_1$ such that $c(\mathaccent"017E{G},\{V_1\cup\{v\},V_2\setminus\{v\}\})$ is
optimized. Ties are broken at random. This produces $|V|-1$ ``locally
optimal'' bipartitions, from which the best one is selected.
\paragraph{Gradient walks.}
Here we interpret the space of all bipartitions $\mathscr{V}$ endowed with
the objective function $c(\mathaccent"017E{G},\mathscr{V})$ as a fitness landscape. We
start with a random but balanced bipartition $\mathscr{V}=\{V_1, V_2\}$.
As move set we allow moving one vertex from $V_1$ to $V_2$ or \textit{vice
versa}. In each step, we execute the move that best improves the
objective function, and stop when we reach a local optimum.
\paragraph{Louvain method.} This method for community detection in graphs
greedily optimizes the so-called modularity of a vertex partition
$\mathscr{V}$ \cite{Blondel:08}. Its objective function is
$q(\mathscr{V}) = \sum_{W\in\mathscr{V}}\sum_{u,v\in W}
(a_{uv}-d_ud_v/(2m))$, where $a_{uv}$ are the entries of the (possibly
weighted) adjacency matrix of a graph $H$, $d_u=\sum_v a_{uv}$ the vertex
degrees,
and $m$ is the sum of all edge weights in the graph. This favors so-called
\emph{communities} or \emph{modules} $W$ that are highly connected
internally but have only few edges between them. The Louvain method
operates in two phases starting from the discrete partition
$\mathscr{V}=\{\{u\}\,|\, u\in V\}$. In the first phase, it repeatedly
iterates over all vertices $x$ and moves $x$ into the community of one of
its neighbors that leads to the highest gain in modularity as long as a
move that increases $q(\mathscr{V})$ can be found. The second phase repeats
the first one on the weighted quotient graph $H/\mathscr{V}$ whose vertices
are the sets of $\mathscr{V}$ and whose edge weights are the sum of the
original weights between the communities. In addition to maximizing the
modularity, we also investigate a variant of the Louvain method that moves
vertices into the community of one of their neighbors if this results in a
lower \textit{UR}\xspace-cost $c(\mathaccent"017E{G},\mathscr{V})$, and otherwise proceeds analogously. We
exclude the merging of the last two vertices to ensure that a non-trivial
partition is returned. Since the Louvain method is sensitive to the order
in which the vertex set is traversed, we randomly permute the order of
vertices to allow multiple runs on the same input.
\smallskip With the exception of the Stoer-Wagner algorithm for solving the
minimum edge cut problem, all of these partitioning methods include random
decisions. One may therefore run them multiple times and use the partition
corresponding to the best objective value, i.e., the lowest \textit{UR}\xspace-cost
$c(\mathaccent"017E{G},\mathscr{V})$ or the highest modularity. If not stated otherwise, we
apply five runs for each of these methods in each recursion step (with a
connected Aho graph) in the following analyses.
\subsubsection*{Construction of test instances}
Since we are interested in the \PROBLEM{(B)PURC} problem in the context of
BMG editing, we test the heuristics on ensembles of perturbed BMGs that were
constructed as follows: We first generate leaf-colored trees $(T,\sigma)$
with a predefined number of vertices $N$ and colors $\ell$ and then compute
their BMGs $\mathaccent"017E{G}(T,\sigma)$. For each tree, we start from a single
vertex. We then repeatedly choose one of the existing vertices $v$
randomly, and, depending on whether $v$ is currently an inner vertex or a
leaf, attach either a single or two new leaves to it, respectively.
Hence, the number
of leaves increases by exactly one and the tree remains phylogenetic in
each step. We stop when the desired number $N$ of leaves is reached. In
the next step, colors are assigned randomly to the leaves under the
constraint that each of the $\ell$ colors appears at least once. We note
that trees created in this manner are usually not least resolved, and their
BMGs are in general not binary-explainable. Finally, we disturb these BMGs
by inserting and deleting arcs according to a specified insertion and
deletion probability, respectively. Since arcs between vertices of the
same color trivially cannot correspond to best matches, we do not insert
arcs between such vertices, i.e., the input graphs for the editing are all
properly vertex-colored digraphs.
For the purpose of benchmarking the heuristics for the \PROBLEM{(B)PURC}
problem, we only retain perturbed BMGs $(\mathaccent"017E{G},\sigma)$ with a connected Aho
graph $H\coloneqq[\mathscr{R}(\mathaccent"017E{G},\sigma), V(\mathaccent"017E{G})]$ because the heuristics
are not applied to instances with a disconnected Aho graph~$H$. Depending on
the insertion and deletion probabilities, we retained 93\% to
100\% of the initial sample, except in the case where arcs were only
inserted to obtain a disturbed graph. Here, the Aho graph $H$ was connected
in 60\% of the initial sample. Thus, even moderate perturbation of a BMG
introduces inconsistencies into the triple set $\mathscr{R}(\mathaccent"017E{G},\sigma)$ and
results in a connected Aho graph $H$ in the majority of cases. As shown in
Fig.~\ref{fig:introduce-inconsistency}, both arc insertions and deletions
can cause triple inconsistencies.
\begin{figure}[!t]
\centering
\includegraphics[width=0.85\textwidth]{./part-quality.pdf}
\caption{Performance of partitioning methods for minimizing
$c(\mathaccent"017E{G},\mathscr{V})$ on perturbed BMGs $(\mathaccent"017E{G},\sigma)$. The rows
correspond to different insertion and deletion probabilities (indicated
in the l.h.s.\ panels) used to disturb the original BMGs. The l.h.s.\
panels show the distribution of the no.\ of arc modifications in total,
arc insertions and arc deletions of the disturbed graphs w.r.t.\ the
original BMGs. The r.h.s.\ panels show the distribution of \textit{UR}\xspace-costs
$c(\mathaccent"017E{G},\mathscr{V})$ (red) obtained for each method, and of the no.\ of
arcs in $U_1(\mathaccent"017E{G},\mathscr{V})$, $U_2(\mathaccent"017E{G},\mathscr{V})$, and
$U_3(\mathaccent"017E{G},\mathscr{V})$ (i.e., the sets that contribute to the \textit{UR}\xspace-cost).
Example plot for $|V|=30$ vertices and $|\sigma(V)|=10$ colors in each
graph. Among the 200 generated graphs, only those with a connected Aho
graph $[\mathscr{R}(\mathaccent"017E{G},\sigma), V(\mathaccent"017E{G})]$ are included in each of the
five rows (93\%, 100\%, 100\%, 60\%, 95\%).}
\label{fig:part-quality}
\end{figure}
\subsubsection*{Benchmarking Results}
Fig.~\ref{fig:part-quality} suggests that the \emph{Simple Greedy} approach
is best suitable for the minimization of the \textit{UR}\xspace-cost $c(\mathaccent"017E{G},\mathscr{V})$
for any of the considered parameters for BMG disturbance. The Louvain
method based on graph modularity (\emph{Louvain (m)}) appears to have by
far the worst performance which, moreover, quickly produces higher
\textit{UR}\xspace-costs with an increasing intensity of the perturbations.
In order to better understand the behavior of the repeated application
of the partitioning heuristics of Alg.~\ref{alg:general-local-optimal}, it
is instructive to consider not only the score but also the structure of
partitions. We observe a strong tendency of some of the partitioning
methods to produce \emph{single-leaf splits}, i.e.,
(bi)partitions $\mathscr{V}$ in which at least one set $W\in\mathscr{V}$
is a singleton (i.e., $|W|=1$). Single-leaf splits in general seem to have
relatively low \textit{UR}\xspace-costs. Further details on the propensity of the
partitioning heuristics to produce single-leaf splits are given in
Appendix~\ref{sect:app-single}.
\subsection{Heuristics for BMG Editing}
In this section, we explore the performance of several variants of
Alg.~\ref{alg:simple} and~\ref{alg:general-local-optimal} for BMG
editing. The variants of Alg.~\ref{alg:general-local-optimal} correspond to
using the heuristics for \PROBLEM{(B)PURC} discussed above for processing
a connected Aho graph
$H\coloneqq [\mathscr{R}(\mathaccent"017E{G}^*[V'],\sigma_{|V'}),V']$ for the informative
triples $\mathscr{R}(\mathaccent"017E{G}^*[V'],\sigma_{|V'})$ in each step of the recursion.
We note that Alg.~\ref{alg:general-local-optimal} in combination with any
of the heuristics for \PROBLEM{(B)PURC} also serves as a heuristic for
\PROBLEM{MaxRTC} because the choice of the partition $\mathscr{V}$ in each
recursion step determines a set of
included triples $xy|z$, namely those
for which $x$ and $y$ are contained in one set of $\mathscr{V}$ while $z$
is contained in another. Another way of expressing that same fact is that
an approximation to \PROBLEM{MaxRTC} is given by the subset
$\mathscr{R}^*\subseteq \mathscr{R}(\mathaccent"017E{G},\sigma)$ of the informative triples
of the input graph $(\mathaccent"017E{G},\sigma)$ that are displayed by the tree $T$
constructed in Alg.~\ref{alg:general-local-optimal}. In particular,
Alg.~\ref{alg:general-local-optimal} together with the \emph{MinCut}
method has been described as a heuristic for \PROBLEM{MaxRTC} in earlier
work \cite{Gasieniec:99,Byrka:10}. For comparison, we will also consider
the following bottom-up approach as a component of Alg.~\ref{alg:simple}:
\paragraph{Best-Pair-Merge-First (BPMF)} was described by \citet{Wu:04},
and constructs a tree from a set of triples $\mathscr{R}$ in a bottom-up
fashion. We use here a modified version introduced by
\citet{Byrka:10}. BPMF operates similar to the well-known UPGMA clustering
algorithm \cite{Sokal:58}. Starting with each vertex $x\in V$ as its own
cluster, pairs of clusters are merged iteratively, thereby defining a
rooted binary tree with leaf set $V$. The choice of the two clusters to
merge depends on a similarity score with the property that any triple
$xy|z$ with $x$, $y$, and $z$ lying in distinct clusters $S_x$, $S_y$, and
$S_z$ contributes positively to $\textrm{score}(S_x,S_y)$ and negatively to
$\textrm{score}(S_x,S_z)$ and $\textrm{score}(S_y,S_z)$. Since BPMF
constructs the tree $T$ from the bottom, it does not imply a vertex
partitioning scheme that could be plugged into the top-down procedure of
Alg.~\ref{alg:general-local-optimal}. Importantly, BPMF is not a
consistent heuristic for \PROBLEM{MaxRTC}, i.e.\ it does not necessarily
recognize consistent triples sets. Hence, consistency in the application to
BMG editing is also not guaranteed, see Fig.~\ref{fig:BPMF-not-consistent}
in Appendix~\ref{sect:app-BPMF} for an example.
\smallskip
In summary, we have two distinct ways to obtain an edited BMG: We may
take either
\begin{enumerate}
\item $\mathaccent"017E{G}(T,\sigma)$, where $T$ is the output tree of
Alg.~\ref{alg:general-local-optimal} or BPMF, respectively, or
\item $\mathaccent"017E{G}(T^*, \sigma)$, where $T^*=\Aho(\mathscr{R}^*,V(\mathaccent"017E{G}))$ is
constructed from the consistent triple subset of triples
$\mathscr{R}^*$. This corresponds to Alg.~\ref{alg:simple}.
\end{enumerate}
Somewhat surprisingly, the results in Fig.~\ref{fig:edit_comparison1}
suggest that it is in general beneficial to extract the triple set
$\mathscr{R}^*$ and rerun the \texttt{BUILD} algorithm, i.e., to use
$\mathaccent"017E{G}(T^*,\sigma)$.
\subsubsection*{Benchmarking Results}
\begin{figure}[!t]
\centering
\includegraphics[width=0.85\textwidth]{./edit_comparison1.pdf}
\caption{Performance comparison of several BMG editing heuristics based
on the no.\ of arc differences. The rows correspond to different
insertion and deletion probabilities (indicated in the second column
panels) used to perturb the original BMGs. The l.h.s.\ panels show the
distribution of the no.\ of arcs in the original BMG and in the
perturbed graph. The second column panels show the distribution of the
no.\ of arc modifications in total, arc insertions and arc deletions of
the perturbed graphs w.r.t.\ the original BMGs. The red lines mark the
median values of the total no.\ of modifications. The r.h.s.\ panels
show the total no.\ of arc differences w.r.t.\ the original random BMGs
(blue) and the perturbed graphs (green). The light colors indicate the
``direct'' performance of each method, i.e., the graph $\mathaccent"017E{G}(T,\sigma)$
where $T$ is the tree that is directly constructed by each method. The
darker colors indicate the results if the methods are used as heuristic
for \PROBLEM{MaxRTC} in Alg.~\ref{alg:simple}. Example plot for
$|V|=30$ vertices and $|\sigma(V)|=10$ colors in each graph, 100 graphs
per row.}
\label{fig:edit_comparison1}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=0.85\textwidth]{./edit_comparison2.pdf}
\caption{Performance comparison of several BMG editing heuristics based
on recall, precision, specificity, and accuracy (rows 1 to 4). The
l.h.s.\ panels show the respective measure for the perturbed graph
w.r.t.\ the original random BMG. The red lines marks the median values
of the latter. The r.h.s.\ panels show the results for the edited
graphs w.r.t.\ the original BMGs (blue) and the perturbed graphs
(green). The light colors indicate the ``direct'' performance of each
method, i.e., the graph $\mathaccent"017E{G}(T,\sigma)$ where $T$ is the tree that is
directly constructed by each method. The darker colors indicate the
results if the methods are used as heuristic for \PROBLEM{MaxRTC} in
Alg.~\ref{alg:simple}. Example plot for $|V|=30$ vertices and
$|\sigma(V)|=10$ colors in each graph, insertion and deletion
probability 0.1, and 100 graphs.}
\label{fig:edit_comparison2}
\end{figure}
To assess the performance of the various heuristics, we consider the
differences between the editing result $(G^*,\sigma)$ from both the
original BMG $(G_{orig},\sigma)$ and the perturbed input graphs
$(G,\sigma)$. In Fig.~\ref{fig:edit_comparison1}, we summarize the absolute
values of the symmetric differences of the arc sets
$d_{orig} \coloneqq |E(G^*)\symdiff E(G_{orig})|$ and
$d \coloneqq |E(G^*)\symdiff E(G)|$, respectively. These results are
translated to usual normalized performance indicators (recall, precision,
specificity, and accuracy; all defined in terms of the arc sets) in
Fig.~\ref{fig:edit_comparison2}.
Comparing the distances $d_{orig}$ (blue boxplots) and $d$ (green
boxplots) of the editing result
$(G^*,\sigma)$ to original unperturbed BMG and the input graph, resp., we
find that, for the methods investigated here, on average $d_{orig}$ is
smaller than $d$. This indicates that all methods are able to capture the
underlying tree structure of the original BMG at least to some extent. The
discrepancy between $d_{orig}$ and $d$ tends to increase with the level of
perturbation, a trend that is most pronounced for \emph{Louvain (c)}. This
result is encouraging for practical applications of BMG modification to
correcting noisy best match data, where the eventual goal is to obtain a
good estimate of the underlying true BMG.
Intriguingly, the extraction of consistent informative triples
$\mathscr{R}^*$ from the reconstructed tree $T$ and rerunning
\texttt{BUILD}, i.e., using $\mathaccent"017E{G}(T^*,\sigma)$, in general improves the
estimation results for the majority of methods. In particular, this
increases the recall without a notable negative impact on precision and
specificity (cf.\ Fig.~\ref{fig:edit_comparison2}). A better recall,
corresponding to a higher proportion of correctly inferred arcs, is not
surprising in this context, since this additional step in essence reduces
the number of triples. We therefore expect the tree
$T^*=\Aho(\mathscr{R}^*,V(\mathaccent"017E{G}))$ to be on average less resolved than $T$.
The BMGs of less resolved trees tend to have more arcs than BMGs of highly
resolved tree (cf.\ \cite[Lemma~8]{Schaller:20x}). In good accordance with
this prediction, \emph{BPMF}, which shows a strong increase of recall,
always constructs a binary, i.e., fully-resolved, tree $T$ -- whereas the
corresponding tree $T^*$ in general is much less resolved.
Somewhat surprisingly, a good or bad performance for minimizing the
\textit{UR}\xspace-cost in individual steps apparently does not directly translate to the
performance in the overall editing procedure. In particular, the
modularity-based \emph{Louvain (m)} method seems to be a better choice than
the \emph{Simple Greedy} approach. The methods \emph{MinCut} and
\emph{Karger} do not seem to be suitable components for
Alg.~\ref{alg:general-local-optimal}, with the exception of the case where
perturbations are arc deletions only (Fig.~\ref{fig:edit_comparison1},
bottom row). Here, \emph{MinCut} produces reasonable estimates that compare
well with other methods. The bottom-up method for the \PROBLEM{MaxRTC}
problem \emph{BPMF} also produces relatively good results. It appears to be
robust at high levels of perturbation. For most of the parameter
combinations, we obtain the best results with the \textit{UR}\xspace-cost-based Louvain
method (\emph{Louvain~(c)}). Here, we often observe a symmetric difference
(w.r.t.\ the arcs sets) that is better than the difference between the
original and the perturbed graph. This trend is illustrated by the red
median lines in Fig.~\ref{fig:edit_comparison1}
and~\ref{fig:edit_comparison2}. Hence, we achieve two goals of BMG
editing: (i) the resulting graph $(\mathaccent"017E{G}^*,\sigma)$ is a BMG, i.e., it
satisfies Def.~\ref{def:BMG}, and (ii) it is closer to the original BMG
than the perturbed graph. We note that we observed similar trends across
all investigated combinations for the numbers of leaves $N$ (ranging from
$10$ to $40$) and of colors $\ell$ ($\ell<N$ ranging from $2$ to at most
$20$).
Our results show that minimization of the \textit{UR}\xspace-cost in each step is not the
best approach to BMG editing because this often produces very unbalanced
partitions. As a consequence, more recursion steps are needed in
Alg.~\ref{alg:general-local-optimal} resulting in higher accumulated number
of arc edits. Fig.~\ref{fig:louvain_better} shows that better solutions to
the BMG editing problem are not necessarily composed of vertex partitions
with minimal \textit{UR}\xspace-cost in each step. The perturbed graph $(\mathaccent"017E{G},\sigma)$ in
Fig.~\ref{fig:louvain_better} was obtained from the randomly simulated BMG
$(\mathaccent"017E{G}_\textrm{orig},\sigma)$ as described above using equal insertion and
deletion probabilities of $0.1$. As an example, the partitions
$\mathscr{V}_1$ and $\mathscr{V}_2$ as constructed by the \emph{MinCut} and
the \emph{Louvain~(c)} method in the first iteration step of
Alg.~\ref{alg:general-local-optimal} are shown as pink and green frames,
respectively. \emph{MinCut} produces a single-leaf split $\mathscr{V}_1$
with an isolated vertex $b_2$ and \textit{UR}\xspace-cost $c(\mathaccent"017E{G},\mathscr{V}_1)=1$
deriving from $U_1(\mathaccent"017E{G},\mathscr{V}_1)=\{(b_2,a_2)\}$. \emph{Louvain~(c)}
identifies the partition $\mathscr{V}_2$ with $c(\mathaccent"017E{G},\mathscr{V}_2)=3$
originating from
$U_2(\mathaccent"017E{G},\mathscr{V}_1)=\{(b_3,a_1),(c_2,a_1), (c_2,b_1)\}$, which
corresponds to the connected components of the Aho graph $H_\textrm{orig}$
of the unperturbed BMG and thus identifies the split in the original tree
$(T,\sigma)$. Here, the correct partition $\mathscr{V}_2$ has a strictly
larger \textit{UR}\xspace-cost than the misleading choice of $\mathscr{V}_1$. However,
\emph{MinCut} results in a higher total edit cost than \emph{Louvain~(c)}
for $(\mathaccent"017E{G},\sigma)$.
\begin{figure}[t]
\centering
\includegraphics[width=0.85\textwidth]{./louvain_better.pdf}
\caption{Example of an instance where the Louvain method performs
better due to more balanced partitions. The (least resolved) tree
$(T,\sigma)$ explains the BMG $(\mathaccent"017E{G}_\textrm{orig},\sigma)$ with vertex
set $V$. The graph
$H_\textrm{orig}=[\mathscr{R}(\mathaccent"017E{G}_\textrm{orig},\sigma), V]$ is the Aho
graph corresponding to the informative triple set
$\mathscr{R}(\mathaccent"017E{G}_\textrm{orig},\sigma)$. The perturbed graph
$(\mathaccent"017E{G},\sigma)$ is obtained from $(\mathaccent"017E{G}_\textrm{orig},\sigma)$ by inserting
the arcs $(b_3,a_1)$, $(c_2,a_1)$, and $(c_2, b_1)$ and deletion of
$(a_1,b_2)$. The corresponding Aho graph
$H=[\mathscr{R}(\mathaccent"017E{G},\sigma), V]$ is connected because the perturbation
introduced the additional informative triple $c_2b_1|b_2$. The green
and pink frames correspond to the partitions $\mathscr{V}_1$ and
$\mathscr{V}_2$ of $V$ constructed by the methods \emph{Louvain~(c)} and
\emph{MinCut}, respectively.}
\label{fig:louvain_better}
\end{figure}
In order to account for the issue of unbalanced partitions, we performed a
cursory analysis on maximizing a gain function rather than minimizing the
\textit{UR}\xspace-cost. In analogy to $c(\mathaccent"017E{G},\mathscr{V})$, we defined $g(\mathaccent"017E{G},\mathscr{V})$
as the number of arcs and non-arcs that are satisfied by the BMGs of
\emph{all} trees in $\mathscr{T}(\mathscr{V})$. Recapitulating the
arguments in the proof of Lemma~\ref{lem:unsat-rel-charac}, one can show
that these relations can also be determined as the union of three sets by
replacing ``$(x,y)\in E$'' with ``$(x,y)\notin E$'' and \textit{vice versa}
in the definitions of $U_1(\mathaccent"017E{G},\mathscr{V})$, $U_2(\mathaccent"017E{G},\mathscr{V})$, and
$U_3(\mathaccent"017E{G},\mathscr{V})$. The gain function $g(\mathaccent"017E{G},\mathscr{V})$ can be used
instead of the \textit{UR}\xspace-cost with \emph{Karger}, \emph{Simple Greedy},
\emph{Gradient Walk}, and in a gain-function-based \emph{Louvain}
method. For all these algorithms, however, maximizing $g(\mathaccent"017E{G},\mathscr{V})$
leads to partitions that appear to be \emph{too} balanced, and a
performance for BMG editing that is worse than the use of the \textit{UR}\xspace-cost. A
possible explanation for both unbalanced and too balanced partitions as
produced with a cost and gain function, resp., is the fact that
$U_1(\mathaccent"017E{G},\mathscr{V})$ and $U_2(\mathaccent"017E{G},\mathscr{V})$ (and their gain function
counterparts) contain pairs of vertices $(x,y)$ that lie in distinct sets
of $\mathscr{V}$. Hence, both single-leaf splits and perfectly balanced
partitions minimize (maximize, resp.) the number of pairs that could
potentially be contained in these arc sets.
\smallskip
\begin{figure}[t]
\centering
\includegraphics[width=0.85\textwidth]{./runtime.pdf}
\caption{Running times of the different methods for BMG editing. The
time only includes the construction of the tree $T$, i.e.,
Alg.~\ref{alg:general-local-optimal} or \emph{BPMF}, resp., but not the
extraction of the triple set $\mathscr{R}^*$ followed by rerunning
\texttt{BUILD}. For each number of leaves $N\in\{10,20,30,40\}$ and
each number of colors $\ell$ (taken from $\{2,5,10,20\}$ such that
$\ell<N$), 100 perturbed BMGs were generated using equal insertion and
deletion probabilities of $0.1$. In the right panel, the median values
are shown with logarithmic axes.}
\label{fig:runtime}
\end{figure}
All methods for BMG editing were implemented and compared using Python on
an off-the-shelf laptop. Fig.~\ref{fig:runtime} summarizes the running
times. The right panel shows that all methods appear to scale polynomially
in the size $|V|$ of the vertex set of the input graph. The methods that
explicitly rely on the \textit{UR}\xspace-cost are much slower than the other methods. We
suspect that this is largely due to the repeated $O(|V'|^2)$-computation of
$c(\mathaccent"017E{G},\mathscr{V})$ whenever a vertex is moved between the sets/communities
in $\mathscr{V}$. This could possibly be improved by an incremental
algorithm.
\subsection{Heuristics for binary-explainable BMG Editing}
\label{sect:beBMG-editing-results}
In order to test the heuristics for the slightly different task of
obtaining a binary-explainable BMG $(\mathaccent"017E{G}^*,\sigma)$, we constructed a
similar set of test instances. The only difference is that we ensured that
$T_\textrm{orig}$ is binary by modifying the attachment procedure (cf.\
Section \ref{subsec:exp_1}) such that in each growth step we only choose
among the vertices that are currently leaves for attaching two new
leaves. Thus, $(\mathaccent"017E{G}_\textrm{orig},\sigma)=\mathaccent"017E{G}(T_\textrm{orig},\sigma)$ is
binary-explainable. The editing heuristics are analogous, with two
straightforward modifications:
\begin{itemize}
\item In the Aho graphs, $\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G},\sigma)$ is used instead of
$\mathscr{R}(\mathaccent"017E{G},\sigma)$.
\item If we encounter a partition $\mathscr{V}$ of cardinality greater than
two in some recursion step, we use a coarse-graining $\mathscr{V}'$ of
$\mathscr{V}$ such that $|\mathscr{V}'|=2$ instead. This modification is
necessary whenever $[\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G},\sigma)[V'], V']$ itself has more than two
connected components, and for the partitions with $|\mathscr{V}|\ge 3$
returned by the Louvain method.
\end{itemize}
By Thm.~\ref{thm:algo-consistent-binary}, this procedure is consistent for
binary-explainable BMGs. Thm.~\ref{thm:algo-consistent-binary}, moreover,
guarantees some freedom in the choice of a coarse-graining
$\mathscr{V}'=\{V_1, V_2\}$ whenever $\mathscr{V}$ is not a bipartition.
We therefore aim to produce (locally) balanced trees in such situations,
i.e., we seek to minimize the difference of $|V_1|$ and $|V_2|$. Formally,
this corresponds to the well-known \PROBLEM{Number Partitioning} problem
with the multiset $\{|V_i| \,\mid\, V_i \in \mathscr{V}\}$ as input. We
use the efficient heuristic described by \citet{Karmarkar:83}, which in
general appears to yield very good solutions of the \PROBLEM{Number
Partitioning} problem \cite{Boettcher:08}.
To construct the second binary tree $T^*$ based on subset of triples
$\mathscr{R}^*\subseteq\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G},\sigma)$ that are displayed by $T$, we
employ an analogous coarse-graining in an otherwise unmodified
\texttt{BUILD} algorithm. We note, however, that one could incorporate
more sophisticated approaches which e.g.\ use some greedy coarse-graining
method based on the \textit{UR}\xspace-cost.
The results for beBMG editing in essence recapitulate the observations
for general BMG editing, see Appendix~\ref{sect:app-beBMG}:
Alg.~\ref{alg:general-local-optimal} in combination with \emph{Louvain~(c)}
appears to be the best choice for the majority of parameter combinations.
However, it is outperformed by the \emph{BPMF} heuristic at high levels of
perturbation (insertion and deletion probability $0.2$). As in the general
case, construction of $T^*$ and using
$(\mathaccent"017E{G}^*,\sigma)\coloneqq\mathaccent"017E{G}(T^*,\sigma)$ as editing result appears to be
advantageous. Moreover, the difference of the editing result
$(\mathaccent"017E{G}^*,\sigma)$ to the original beBMG $(\mathaccent"017E{G}_\textrm{orig},\sigma)$ is on
average smaller than the difference of $(\mathaccent"017E{G}^*,\sigma)$ to the perturbed
graph $(\mathaccent"017E{G},\sigma)$.
\section{Summary and discussion}
In this contribution, we have described a large class of heuristics for BMG
editing that operate in a recursive top-down fashion to (at least
implicitly) construct a tree $(T,\sigma)$ capturing the underlying
BMG-structure of an arbitrary input graph $(\mathaccent"017E{G},\sigma)$. We have shown
that this is closely related to a specific notion of locally good edits,
which we assess using the \textit{UR}\xspace-cost. The \textit{UR}\xspace-cost counts the minimum number
of arc insertions and deletions of the BMG-editing for $(\mathaccent"017E{G},\sigma)$ that
are linked to each inner node (and thus to their corresponding leaf
partitions) in $(T,\sigma)$ and cannot be reversed in subsequent recursion
steps. In particular, we showed that an optimal solution among all possible
partitions guarantees consistency of this class of heuristics (cf.\
Thm.~\ref{thm:algo-consistent} and~\ref{thm:algo-consistent-binary}).
Unfortunately, the corresponding problem \PROBLEM{BPURC} is itself
NP-complete.
We therefore suggested a number of approximation methods for finding
suitable partitions, and compared their performances in the context of
Alg.~\ref{alg:general-local-optimal}. We find that, even though good
solutions for \PROBLEM{(B)PURC} alone do not seem to be the most adequate
approach, the value of the \textit{UR}\xspace-costs appears most clearly in a combination
with a method for community detection, more precisely, a modification of
the Louvain method \cite{Blondel:08}.
For all of the methods investigated here, we found that the Aho graph
$H\coloneqq [\mathscr{R}(\mathaccent"017E{G},\sigma)[V'], V']$ serves as a useful starting
point for finding a suitable partition. This choice is based on the idea
that, due to the properties of BMGs and in particular the construction of
the tree $(T,\sigma)$ from informative triples of the BMG
$(\mathaccent"017E{G},\sigma)=\mathaccent"017E{G}(T,\sigma)$, arc insertions and deletions in
$(\mathaccent"017E{G},\sigma)$ should not add too many new edges between the connected
components of the originally disconnected Aho graph of
$\mathscr{R}(\mathaccent"017E{G},\sigma)$ (cf.\
Fig.~\ref{fig:introduce-inconsistency}). Therefore, we suggest that there
is a correlation between good partitions $\mathscr{V}$ of $V'$, i.e.\
partitions linked to few edits, and the minimization of the number of edges
in $H$ connecting vertices in distinct sets of $\mathscr{V}$.
For the general BMG editing problem, we did not make use of the information
contained in the set of forbidden triples $\mathscr{F}(\mathaccent"017E{G},\sigma)$ of the
input graph $(\mathaccent"017E{G},\sigma)$. It might be possible to adapt the algorithm
\texttt{MTT} \cite{He:06}, which identifies consistent pairs
$(\mathscr{R},\mathscr{F})$, instead of \texttt{BUILD}. \texttt{MTT}
constructs a coarse-graining $\mathscr{V}_\textrm{MTT}$ of the set of
connected components of the Aho graph (on $\mathscr{R}$) in order to
account for the forbidden triples in $\mathscr{F}$ in each recursion step.
Possibly, $\mathscr{V}_\textrm{MTT}$ (or some suitable graph
representation) yields a further improvement. However, in case of beBMG
editing, the extended triple set $\mathop{\mathscr{R}^{\textrm{B}}}(\mathaccent"017E{G},\sigma)$ and thus the
corresponding Aho graphs by construction already cover the information
contained in $\mathscr{F}(\mathaccent"017E{G},\sigma)$. Since no substantial improvement over
the general case was observed in this case (cf.\
Fig.~\ref{fig:edit-comparison-binary}), we opted against more detailed
benchmarking of $\mathscr{V}_\textrm{MTT}$ in comparison to partitions
based on the Aho graph.
The purpose of this contribution is to establish a sound theoretical
foundation for practical approaches to BMG editing and to demonstrate that
the problem can solved for interestingly large instances at reasonable
accuracy. In computational biology, however, much larger problems than
the ones considered here would also be of interest. Less emphasis has been
placed here on computational efficiency and scalability of different
variants. We leave this as topic for future research. Given the
performance advantage of community detection over minimization of the
\textit{UR}\xspace-cost in each step, it seems most promising to focus on community
detection methods that scale well for very large system. The Louvain
method seems to be a promising candidate, since it has been applied
successfully to large networks in the past \cite{Blondel:08}. This is largely
due to the fact that the change of modularity in response to moving a
vertex between modules can be computed efficiently. We suspect that a
comparably fast computation of the \textit{UR}\xspace-cost may also be possible; this
does not appear to be trivial, however. Moreover, the method could
probably be accelerated by moving vertices into the community of the first
neighbor such that this results in a (not necessarily optimal) improvement
of the \textit{UR}\xspace-cost. A similar randomization approach has already shown to
only slightly affect the clustering quality in terms of modularity
\cite{Traag:2015}.
Since the restriction of a (be)BMG to a subset of colors is again a
(be)BMG, it may also be possible to remove large parts of the noise by
editing induced subgraph on a moderate number of colors, possibly using
information of the phylogeny of the species to select species (= color)
sets. Presumably, color sets with sufficient overlaps will need to be
considered. A systematic analysis of this idea, however, depends on
scalable BMG editing for large instances and goes beyond the scope of
this contribution.
A potential shortcoming of the empirical analysis in
Sec.~\ref{sec:heur-computational-exp} is the simplistic error model, i.e.,
the independent perturbation of arcs (and non-arcs). Better models will
depend on the investigation of BMGs derived from real-life sequence data.
Such data is often burdened with systematic errors arising e.g.\ from the
fact that a common ancestry often cannot be detected for very large
evolutionary distances and from unequal mutation rates during the evolution
of gene families, see e.g.\ \cite{Rost:99,Lafond:18,Stadler:20a} for more
in-depth discussions of these issues. Benchmarking using real-life data,
however, is a difficult task because the ground truth is unknown and large,
well-curated data sets are not available. Our results so far suggest that a
good performance w.r.t.\ the input graph is also an indicator for a good
performance w.r.t.\ the true graph (cf.\ Fig.~\ref{fig:edit_comparison1}
and~\ref{fig:edit_comparison2}, green vs.\ blue boxplots). Moreover, they
at least suggest that realistic BMG data can be processed with
sufficient accuracy and efficiency to make BMGs an attractive alternative
to classical phylogenetic methods. The construction of bioinformatics
workflows to process best hit data, e.g.\ at the first processing stage of
\texttt{ProteinOrtho} \cite{Lechner:14a}, is a logical next step.
\subsection*{Acknowledgements}
This work was supported in part by the \emph{Deutsche
Forschungs\-gemeinschaft}, the Austrian Federal Ministries
BMK and BMDW, and the Province of Upper Austria in the frame of the COMET
Programme managed by FFG.
\subsection*{Conflict of interest}
The authors declare that they have no conflict of interest.
\begin{appendix}
\section{Proofs}
\subsection*{Proof of Lemma~\ref{lem:unsat-rel-charac}}
\begin{clemma}{\ref{lem:unsat-rel-charac}}
Let $(\mathaccent"017E{G}=(V,E),\sigma)$ be a properly vertex-colored digraph and let
$\mathscr{V}=\{V_1,\dots,V_k\}$ be a partition of $V$ with
$|\mathscr{V}|=k\ge2$. Then
\begin{equation*}
U(\mathaccent"017E{G},\mathscr{V}) = U_1(\mathaccent"017E{G},\mathscr{V}) \;\charfusion[\mathbin]{\cup}{\cdot}\;
U_2(\mathaccent"017E{G},\mathscr{V}) \;\charfusion[\mathbin]{\cup}{\cdot}\; U_3(\mathaccent"017E{G},\mathscr{V})\,.
\end{equation*}
\end{clemma}
\begin{proof}
We first note that $U_1\coloneqq U_1(\mathaccent"017E{G},\mathscr{V})$,
$U_2\coloneqq U_2(\mathaccent"017E{G},\mathscr{V})$ and $U_3\coloneqq U_3(\mathaccent"017E{G},\mathscr{V})$
are pairwise disjoint. Furthermore, we have $x\ne y$ and
$\sigma(x)\ne\sigma(y)$ for every
$(x,y)\in U_1 \,\charfusion[\mathbin]{\cup}{\cdot}\, U_2 \,\charfusion[\mathbin]{\cup}{\cdot}\, U_3 $ and every
$(x,y)\in U(\mathaccent"017E{G},\mathscr{V})$. Moreover, recall that
$\mathscr{T}(\mathscr{V})$ is the set of trees $T$ on $V$ that satisfy
$\mathscr{V} = \{L(T(v)) \mid v\in\mathsf{child}_{T}(\rho_T) \}$. Therefore,
there is a one-to-one correspondence between the $k\ge 2$ sets in
$\mathscr{V}$ and the children $\mathsf{child}_{T}(\rho_{T})$ of the root
$\rho_{T}$ for any $T\in \mathscr{T}(\mathscr{V})$. We denote by $v_{i}$
the child corresponding to $V_{i}\in\mathscr{V}$; thus
$V_{i}=L(T(v_{i}))$.
We first show that $(x,y)\in U_1 \,\charfusion[\mathbin]{\cup}{\cdot}\, U_2 \,\charfusion[\mathbin]{\cup}{\cdot}\, U_3$ implies
$(x,y)\in U(\mathaccent"017E{G},\mathscr{V})$. Let $T\in \mathscr{T}(\mathscr{V})$
be chosen arbitrarily, and let $\rho$ be its root. Suppose that
$(x,y)\in U_1$. Thus, we have $(x,y)\in E$,
$\sigma(y)\in \sigma(L(T(v_{i})))$, $x\preceq_{T} v_{i}$ and
$y\preceq_{T} v'$ for some $v'\in\mathsf{child}_{T}(\rho)\setminus\{v_{i}\}$.
Moreover, $\sigma(y)\in \sigma(L(T(v_{i})))$ implies that there is a
vertex $y'\preceq_{T}v_{i}$ with $\sigma(y')=\sigma(y)$. Taken together,
we obtain $\lca_{T}(x,y')\preceq_T v_{i} \prec_T \rho = \lca_{T}(x,y)$,
and thus $(x,y)\notin E(\mathaccent"017E{G}(T,\sigma))$. If $(x,y)\in U_2$, we have
$(x,y)\notin E$, $\sigma(y)\notin \sigma(L(T(v_{i})))$,
$x\preceq_{T}v_{i}$ and $y\preceq_{T} v'$ for some
$v'\in\mathsf{child}_{T}(\rho)\setminus\{v_{i}\}$. Moreover,
$\sigma(y)\notin \sigma(L(T(v_{i})))$ implies that there is no vertex
$y'\preceq_{T}v_{i}$ with $\sigma(y')=\sigma(y)$. Thus,
$\lca_{T}(x,y') = \lca_{T}(x,y)= \rho$ holds for all $y'$ of color
$\sigma(y')=\sigma(y)$, and thus $(x,y)\in E(\mathaccent"017E{G}(T,\sigma))$. Finally,
suppose $(x,y)\in U_3$. We have $(x,y)\notin E$ and $y$ is the only leaf
of its color in $L(T(v_{i}))$. Therefore, there is no vertex $y'$ with
$\sigma(y')=\sigma(y)$ and $\lca_{T}(x,y') \prec_{T} \lca_{T}(x,y)$, and
thus $(x,y)\in E(\mathaccent"017E{G}(T,\sigma))$. In summary, one of the conditions in
Def.~\ref{def:unsat-relations} is satisfied for $T$ in all three cases.
Since $T$ was chosen arbitrarily, we conclude
$(x,y)\in U(\mathaccent"017E{G},\mathscr{V})$ for any
$(x,y)\in U_1 \,\charfusion[\mathbin]{\cup}{\cdot}\, U_2 \,\charfusion[\mathbin]{\cup}{\cdot}\, U_3$.
In order to show that $(x,y)\in U(\mathaccent"017E{G},\mathscr{V})$ implies
$(x,y)\in U_1 \,\charfusion[\mathbin]{\cup}{\cdot}\, U_2 \,\charfusion[\mathbin]{\cup}{\cdot}\, U_3$, we distinguish
\textit{Case (a)}: $(x,y)\in E$ and $(x,y)\notin E(\mathaccent"017E{G}(T,\sigma))$ holds
for all $T\in\mathscr{T}(\mathscr{V})$, and \textit{Case (b)}:
$(x,y)\notin E$ and $(x,y)\in E(\mathaccent"017E{G}(T,\sigma))$ holds for all
$T\in\mathscr{T}(\mathscr{V})$.
\par\noindent\textit{Case (a).}
$(x,y)\in E$ implies $\sigma(x)\ne\sigma(y)$. Moreover,
there is a vertex $y'$ with $\sigma(y')=\sigma(y)$, and
$\lca_{T}(x,y')\prec_{T}\lca_{T}(x,y)$ for every
$T\in\mathscr{T}(\mathscr{V})$ because $(x,y)\notin E(\mathaccent"017E{G}(T,\sigma))$.
Since this is true for all trees in $\mathscr{T}(\mathscr{V})$, there
must be a set $V_{i}\in \mathscr{V}$ such that $x, y'\in V_{i}$, and in
particular $\sigma(y')=\sigma(y)\in\sigma(V_{i})$. Now suppose, for
contradiction, that $y\in V_{i}$ and thus $x,y,y'\in V_{i}$. In this
case, we can choose a tree $T\in\mathscr{T}(\mathscr{V})$ such that
$x,y\prec_{T}v_{i}$ for some child $v_{i}\in\mathsf{child}_{T}(\rho_T)$ and
$\lca_{T}(x,y)\preceq\lca_{T}(x,y')$ hold for all $y'$ of color
$\sigma(y')=\sigma(y)$. Hence, we obtain $(x,y)\in E(\mathaccent"017E{G}(T,\sigma))$ for
this tree; a contradiction. Therefore, we conclude that
$y\in V\setminus V_{i}$. In summary, all conditions for $U_1$ are
satisfied, and thus $(x,y)\in U_1$.
\par\noindent\textit{Case (b).} We have $(x,y)\notin E$ and
$(x,y)\in E(\mathaccent"017E{G}(T,\sigma))$ for all $T\in\mathscr{T}(\mathscr{V})$. Let
$V_{i}\in \mathscr{V}$ such that $x\in V_{i}$. We distinguish the two
cases (i) $y\notin V_{i}$, and (ii) $y\in V_{i}$. In Case~(i), suppose,
for contradiction, that $\sigma(y)\in\sigma(V_{i})$. Then, for every
tree $T\in\mathscr{T}(\mathscr{V})$, there must be a vertex $y'$ of color
$\sigma(y)$ such that
$\lca_{T}(x,y')\preceq_{T} v_{i}\prec_{T}\rho_T=\lca_{T}(x,y)$,
contradicting $(x,y)\in E(\mathaccent"017E{G}(T,\sigma))$. Therefore, we conclude
$\sigma(y)\notin\sigma(V_{i})$. It follows that $(x,y)\in U_2$. In
Case~(ii), assume, for contradiction, that there is a vertex
$y'\in V_{i}$ of color $\sigma(y)$ such that $y\ne y'$. This together
with $\sigma(y')=\sigma(y)\ne\sigma(x)$ implies that all three vertices
$x,y,y'$ are pairwise distinct. Since in addition $x,y,y'\in V_{i}$, we
can choose a tree $T\in\mathscr{T}(\mathscr{V})$ such that
$x,y,y'\prec_{T}v_{i}$ for some child $v_{i}\in\mathsf{child}_{T}(\rho_T)$ and
$\lca_{T}(x,y')\prec_{T}\lca_{T}(x,y)$; a contradiction to
$(x,y)\in E(\mathaccent"017E{G}(T,\sigma))$ for all $T\in\mathscr{T}(\mathscr{V})$.
Therefore, we conclude that $y$ is the only vertex of its color in
$V_{i}$. It follows that $(x,y)\in U_3$. \qed
\end{proof}
\subsection*{Proof of Lemma~\ref{lem:independent-U}}
In order to prove Lemma~\ref{lem:independent-U}, we first need the
following technical result which shows that the editing of an arc in
Alg.~\ref{alg:general-local-optimal} will not be reversed in the subsequent
recursion step.
\begin{lemma}
\label{lem:hierarchical-edits-independent}
Let $(\mathaccent"017E{G}=(V,E),\sigma)$ be a properly vertex-colored digraph,
$\mathscr{V}=\{V_1,\dots,V_k\}$ a partition of $V$ with
$|\mathscr{V}|=k\ge2$, and $\mathscr{V}_i=\{V_{i,1},\dots,V_{i,l}\}$,
$1\le i\le k$, a partition of $V_i$ with
$|\mathscr{V}_i|=l\ge2$. Moreover, let
$(\mathaccent"017E{G}'\coloneqq \mathaccent"017E{G} \triangle U(\mathaccent"017E{G},\mathscr{V}),\sigma)$ be the colored
digraph
that is obtained by applying the edits in $U(\mathaccent"017E{G},\mathscr{V})$ to
$(\mathaccent"017E{G},\sigma)$. Then
$U(\mathaccent"017E{G},\mathscr{V})\cap U(\mathaccent"017E{G}'[V_i],\mathscr{V}_i)=\emptyset$.
\end{lemma}
\begin{proof}
Let $\mathaccent"017E{G}'_i\coloneqq \mathaccent"017E{G}'[V_i]$. The sets of unsatisfiable relations
$U(\mathaccent"017E{G},\mathscr{V})$ and $U(\mathaccent"017E{G}'_i, \mathscr{V}_i)$ are given by the
(disjoint) unions
$U_1(\mathaccent"017E{G},\mathscr{V})\charfusion[\mathbin]{\cup}{\cdot} U_2(\mathaccent"017E{G},\mathscr{V})\charfusion[\mathbin]{\cup}{\cdot}
U_3(\mathaccent"017E{G},\mathscr{V})$ and
$U_1(\mathaccent"017E{G}'_i,\mathscr{V}_i)\charfusion[\mathbin]{\cup}{\cdot} U_2(\mathaccent"017E{G}'_i,\mathscr{V}_i)\charfusion[\mathbin]{\cup}{\cdot}
U_3(\mathaccent"017E{G}'_i,\mathscr{V}_i)$, respectively (cf.\
Lemma~\ref{lem:unsat-rel-charac}). First, let
$(x,y)\in U_1(\mathaccent"017E{G},\mathscr{V})$. Since, by definition of
$U_1(\mathaccent"017E{G},\mathscr{V})$, $x$ and $y$ are contained in different sets of the
partition $\mathscr{V}$, they cannot be both contained in $V_i$ and thus,
$U_1(\mathaccent"017E{G},\mathscr{V}) \cap U(\mathaccent"017E{G}'_i, \mathscr{V}_i)=\emptyset$. One
analogously argues that
$U_2(\mathaccent"017E{G},\mathscr{V}) \cap U(\mathaccent"017E{G}'_i, \mathscr{V}_i)=\emptyset$. Now, assume
for contradiction that
$(x,y)\in U_3(\mathaccent"017E{G},\mathscr{V})\cap U(\mathaccent"017E{G}'_i, \mathscr{V}_i)$. By
definition of $U(\mathaccent"017E{G}'_i, \mathscr{V}_i)$, this implies $x,y\in
V_i$. Moreover, by definition of $U_3(\mathaccent"017E{G},\mathscr{V})$, we have
$(x,y)\notin E$, which immediately implies $(x,y)\in E(\mathaccent"017E{G}'_i)$. By
Lemma~\ref{lem:unsat-rel-charac}, we therefore conclude
$(x,y)\in U_1(\mathaccent"017E{G}'_i,\mathscr{V}_i)$. Let $V_{i,j}$ be the set of the
partition $\mathscr{V}_i$ which contains $x$. Then, by definition of
$U_1(\mathaccent"017E{G}'_i,\mathscr{V}_i)$, the color of $y$ is contained in both
$V_{i,j}$ and $V_i\setminus V_{i,j}$, i.e., $V_i$ contains at least two
vertices of color $\sigma(y)$. However, $(x,y)\in U_3(\mathaccent"017E{G},\mathscr{V})$
and $y\in V_i$ together imply that $y$ is the only vertex of its color in
$V_i$; a contradiction. \qed
\end{proof}
We are now in the position to prove the more general
\begin{clemma}{\ref{lem:independent-U}}
All edit sets $U(\mathaccent"017E{G}^*[V'],\mathscr{V})$ constructed in
Alg.~\ref{alg:general-local-optimal} are pairwise disjoint.
\end{clemma}
\begin{proof}
First note that, by Lemma~\ref{lem:unsat-rel-charac}, we have
$\sigma(x)\ne\sigma(y)$ for all $(x,y)\in U(\mathaccent"017E{G}^*[V'],
\mathscr{V})$. Hence, the graph $(\mathaccent"017E{G}^*,\sigma)$ remains properly colored
during the whole recursion. Moreover, recursive calls on a set $V'$ with
$|V'|=1$ trivially contribute with a \textit{UR}\xspace-cost of zero.
By construction, the partitions in consecutive calls of \texttt{Edit()}
form a hierarchical refinement such that in each recursive call a single
element of $V_i\in\mathscr{V}$ is refined. Clearly edit sets encountered
in independent branches of the recursion tree are disjoint because they
pertain to disjoint vertex sets. For directly consecutive calls of
\texttt{Edit()}, Lemma~\ref{lem:hierarchical-edits-independent} states
that the edits sets are disjoint. Now consider two recursive call on
$V'$ and $V''$ with $V''\subset V'$ that are not directly consecutive.
Let $\mathscr{V}'$ and $\mathscr{V}''$, resp., be the partitions chosen
for the vertex sets $V'$ and $V''$ of $\mathaccent"017E{G}'$ and $\mathaccent"017E{G}''$ at the beginning
of the two recursion steps. We can apply the same arguments as in the
proof of Lemma~\ref{lem:hierarchical-edits-independent} to conclude that
$U_i(\mathaccent"017E{G}'[V'],\mathscr{V}') \cap U(\mathaccent"017E{G}''[V''],\mathscr{V}'')=\emptyset$,
$i\in \{1,2\}$. Finally, assume, for contradiction, that
$(x,y)\in U_3(\mathaccent"017E{G}'[V'],\mathscr{V}') \cap U(\mathaccent"017E{G}''[V''],\mathscr{V}'')$. By
definition of $U(\mathaccent"017E{G}''[V''], \mathscr{V}'')$, this implies $x,y\in
V''$. Moreover, by definition of $U_3(\mathaccent"017E{G}'[V'],\mathscr{V}')$, we have
$(x,y)\notin E(\mathaccent"017E{G}')$, which immediately implies
$(x,y)\in E(\mathaccent"017E{G}'\triangle U(\mathaccent"017E{G}'[V'],\mathscr{V}'))$, i.e., $(x,y)$ is an
arc after the editing in this step. Since both $x,y$ are contained in
$V''$, it follows from Lemma~\ref{lem:unsat-rel-charac} that all edit
steps on the way from $\mathaccent"017E{G}'[V'']$ to $\mathaccent"017E{G}''[V'']$ must be performed by the
set $U_3$, i.e., they exclusively correspond to arc
insertions. Therefore, $(x,y)$ is still an arc in $\mathaccent"017E{G}''[V'']$. By
Lemma~\ref{lem:unsat-rel-charac}, we therefore conclude that
$(x,y)\in U_1(\mathaccent"017E{G}''[V''],\mathscr{V}'')$. Let $V_{x}$ be the set of the
partition $\mathscr{V}''$ that contains $x$. Then, by definition of
$U_1(\mathaccent"017E{G}''[V''],\mathscr{V}'')$, the color of $y$ is contained in both
$V_{x}$ and $V''\setminus V_{x}$, i.e., $V''$ contains at least two
vertices of color $\sigma(y)$. However,
$(x,y)\in U_3(\mathaccent"017E{G}'[V'],\mathscr{V}')$ and $y\in V''\subset V_{x,y}$ for
some $V_{x,y}\in \mathscr{V}'$ together imply that $y$ is the only vertex
of its color in $V''$; a contradiction. \qed
\end{proof}
\subsection*{Proof of Theorem~\ref{thm:BPURC-NPc}}
In this section we show that \PROBLEM{(B)PURC} is NP-hard by reduction from
\PROBLEM{Set Splitting}.
\begin{problem}[\PROBLEM{Set Splitting}]\ \\
\begin{tabular}{ll}
\emph{Input:} & A collection $\mathfrak{C}$ of subsets of a finite set
$S$, denoted by $(\mathfrak{C},S)$.\\
\emph{Question:} & Is there a bipartition of $S$ into two subsets $S_1$
and
$S_2$ such that\\
&no subset in $\mathfrak{C}$ is entirely contained in either $S_1$ or
$S_2$?
\end{tabular}
\end{problem}
\begin{proposition}{\cite{Lovasz:73}}
\label{prop:set-splitting-NPc}
\PROBLEM{Set Splitting} is NP-complete.
\end{proposition}
\begin{ctheorem}{\ref{thm:BPURC-NPc}}
\PROBLEM{BPURC} is NP-complete.
\end{ctheorem}
\begin{proof}
Given a properly vertex-colored digraph $(\mathaccent"017E{G}=(V,E),\sigma)$ and a
bipartition $\mathscr{V}$ of $V$, the set $U(\mathaccent"017E{G},\mathscr{V})$ and thus
the \textit{UR}\xspace-cost $c(\mathaccent"017E{G},\mathscr{V})=|U(\mathaccent"017E{G},\mathscr{V})|$ can be computed in
polynomial time according to Cor.~\ref{cor:U-polytime}. Therefore,
\PROBLEM{BPURC} is contained in NP. To show NP-hardness, we use reduction
from \PROBLEM{Set Splitting}.
Let $(\mathfrak{C},S)$ be an instance of \PROBLEM{Set Splitting}. We may
assume w.l.o.g.\ that $|C|\ge 2$ holds for all $C\in \mathfrak{C}$, since
otherwise there is no solution at all for \PROBLEM{Set Splitting}. In
addition, we assume that $\bigcup_{C\in \mathfrak{C}}C = S$. To see that
this does not yield a loss of generality, suppose that
$\bigcup_{C\in \mathfrak{C}}C =S' \subsetneq S$. If $\{S'_1,S'_2\}$ is a
solution for $(\mathfrak{C},S')$ then no subset in $\mathfrak{C}$ is
entirely contained in either $S'_1$ or $S'_2$. Therefore, we can
construct a solution $(S_1,S_2)$ for $(\mathfrak{C},S)$ by arbitrarily
adding the elements in $S\setminus S'$ to either $S'_1$ or $S'_2$. In
contrast, $\{S_1\cap S',S_2\cap S'\}$ is a solution for
$(\mathfrak{C},S')$ provided that $\{S_1,S_2\}$ is a solution for
$(\mathfrak{C},S)$.
Now, let $(\mathfrak{C},S)$ be an instance of \PROBLEM{Set Splitting} and
define, for all $s\in S$, the set
$\mathfrak{C}(s) \coloneqq \{C\mid C\in \mathfrak{C}, s\in C\}$ as the
subset of $\mathfrak{C}$ that comprises all elements $C\in \mathfrak{C}$
that contain $s$. Note that $\mathfrak{C}(s) \neq \emptyset$ for all
$s\in S$, since we have assumed $\bigcup_{C\in \mathfrak{C}}C = S$, i.e.,
every $s\in S$ is contained in some element of $\mathfrak{C}$.
We construct a graph $(\mathaccent"017E{G}=(V,E),\sigma)$ that serves as input for
\PROBLEM{BPURC} as follows:
\begin{description}
\item[\emph{Step 1:}] For all $s\in S$, construct an $s$-gadget
$\mathscr{G}_s$ as follows:
\begin{enumerate}
\item[(i)] For all $C\in \mathfrak{C}(s)$, add four new vertices to
$\mathscr{G}_s$ of which two are colored with $(C,1)$ and the other
two with $(C,2)$.
\item[(ii)] Add arcs $(x,y),(y,x)$ between all $x,y\in V(\mathscr{G}_s)$
with $\sigma(x)\neq \sigma(y)$.
\end{enumerate} \smallskip
\item[\emph{Step 2:}] Set
$V \coloneqq \charfusion[\mathbin]{\bigcup}{\cdot}_{s\in S}\; V(\mathscr{G}_s)$,
$E \coloneqq \charfusion[\mathbin]{\bigcup}{\cdot}_{s\in S}\; E(\mathscr{G}_s)$ and preserve the
coloring of the vertices within the $s$-gadgets to obtain the graph
$(\mathaccent"017E{G},\sigma)$.
\end{description}
By construction, $|V(\mathscr{G}_s)| = 4|\mathfrak{C}(s)|$ and
$|V| = \sum_{s\in S}\; 4|\mathfrak{C}(s)|\le 4|\mathfrak{C}||S|$. Hence,
the construction of $(\mathaccent"017E{G},\sigma)$ can be achieved in polynomial time.
Moreover, by construction, $\{V(\mathscr{G}_s)\mid s\in S\}$ forms a
partition of $V$ and there are no arcs between vertices of distinct
$s$-gadgets. Furthermore, $\sigma(V) = \mathfrak{C}\times \{1,2\}$. An
illustrative example of such a constructed graph $(\mathaccent"017E{G},\sigma)$ is
provided in Fig.~\ref{fig:BPURC-reduction}.
\begin{figure}[htb]
\centering
\includegraphics[width=0.85\textwidth]{./BPURC-reduction.pdf}
\caption[]{Example for the reduction from an instance
$(\mathfrak{C},S)$ of \PROBLEM{Set Splitting} to an instance
$(\mathaccent"017E{G},\sigma)$ with $k=0$ of \PROBLEM{BPURC}, as specified in the proof
of Thm.~\ref{thm:BPURC-NPc}. In this example, we have $S=\{a,b,c,d\}$
and $\mathfrak{C}=\{C_1,C_2,C_3\}$. By construction, all arcs are
bidirectional and thus, arrow heads are omitted in the drawing of
$(\mathaccent"017E{G},\sigma)$. A solution for $(\mathfrak{C},S)$ is $S_1=\{a,d\}$ and
$S_2=\{b,c\}$. The latter is equivalent to a solution of
\PROBLEM{BPURC} by ``separating'' the $a$- and $d$-gadget from the
$b$- and $c$-gadget as indicated by the dashed line. The latter
yields a bipartition $\mathscr{V}=\{V_1,V_2\}$ of $V(\mathaccent"017E{G})$
that solves \PROBLEM{BPURC} with input $(\mathaccent"017E{G},\sigma,k=0)$.\\
Note, slight changes of the input $(\mathfrak{C},S)$ to
$S'=S\setminus\{d\}$ and
$\mathfrak{C}'=\{C_1,C_2,C_3\setminus \{d\}\}$ would yield an
instance of \PROBLEM{Set Splitting} that has no yes-answer. In this
case, the $d$-gadget would disappear from $(\mathaccent"017E{G},\sigma)$ resulting in
the graph $(\mathaccent"017E{G}',\sigma')$. It is easy to see that there is no
bipartition $\mathscr{V}=\{V_1,V_2\}$ of $V(\mathaccent"017E{G}')$ such that
$\sigma(V_1)=\sigma(V_2)=\sigma(V(\mathaccent"017E{G}'))$ and no gadget gets split up
between $V_1$ and $V_2$; two necessary properties to obtain a
solution for \PROBLEM{BPURC} with input $(\mathaccent"017E{G}',\sigma')$ and $k=0$
(cf.\ proof of Thm.~\ref{thm:BPURC-NPc}).}
\label{fig:BPURC-reduction}
\end{figure}
We continue by showing that an instance $(\mathfrak{C},S)$ of
\PROBLEM{Set Splitting} has a yes-answer if and only if \PROBLEM{BPURC}
has a yes-answer for the input graph $(\mathaccent"017E{G}=(V,E),\sigma)$ constructed
above and $k=0$. In particular, we will show that $\{S_1,S_2\}$ is a
solution of $(\mathfrak{C},S)$ if and only if $\mathscr{V} = \{V_1,V_2\}$
with $V_i=\cup_{s\in S_i} V(\mathscr{G}_s)$, $i\in \{1,2\}$ is a solution
for $(\mathaccent"017E{G},\sigma)$ where $c(\mathaccent"017E{G},\mathscr{V}) = 0$.
Recall that the set of unsatisfiable relations $U(\mathaccent"017E{G},\mathscr{V})$ of a
bipartition $\mathscr{V}$ of $V$ is given by the (disjoint) union
$U_1\charfusion[\mathbin]{\cup}{\cdot} U_2\charfusion[\mathbin]{\cup}{\cdot} U_3$ of the three sets
$U_1 \coloneqq U_1(\mathaccent"017E{G},\mathscr{V})$, $U_2 \coloneqq U_2(\mathaccent"017E{G},\mathscr{V})$
and $U_3 \coloneqq U_3(\mathaccent"017E{G},\mathscr{V})$ (cf.\
Lemma~\ref{lem:unsat-rel-charac}).
First suppose that \PROBLEM{Set Splitting} with input $(\mathfrak{C},S)$
has a yes-answer and let $\{S_1,S_2\}$ be one of its solutions. Hence,
no subset in $\mathfrak{C}$ is entirely contained in either $S_1$ or
$S_2$, and both sets must be non-empty. Consider the set
$\mathscr{V} = \{V_1,V_2\}$ with
$V_i=\cup_{s\in S_i}\, V(\mathscr{G}_s)$, $i\in \{1,2\}$. Since
$\{S_1,S_2\}$ is a bipartition of $S$ and
$\{V(\mathscr{G}_s)\mid s\in S\}$ is a partition of $V$, we conclude that
$\mathscr{V}$ is a bipartition of $V$ and that $V(\mathscr{G}_s)$ is
entirely contained in either $V_1$ or $V_2$ for all $s\in S$. Together
with the fact that there are no arcs in $\mathaccent"017E{G}$ between vertices of distinct
$s$-gadgets this implies that $U_1 =\emptyset$.
In order to verify that $U_2 = U_3 = \emptyset$, we first show that
$\sigma(V_1)=\sigma(V_2) = \sigma(V)$ and that $V_1$ and $V_2$ contain at
least two vertices of every color, respectively. Consider two arbitrary
pairs $(C,1), (C,2)\in \sigma(V)=\mathfrak{C}\times\{1,2\}$. Since
$\{S_1,S_2\}$ is a solution for \PROBLEM{Set Splitting} with input
$(\mathfrak{C},S)$, there are vertices $s\in C\cap S_1$ and
$s'\in C\cap S_2$ and thus, $V(\mathscr{G}_s)\subseteq V_1$ and
$V(\mathscr{G}_{s'})\subseteq V_2$. By construction, each of the sets
$V(\mathscr{G}_s)$ and $V(\mathscr{G}_{s'})$ contains two vertices of
color $(C,1)$ and two vertices of color $(C,2)$. Since
$V(\mathscr{G}_s)\subseteq V_1$ and $V(\mathscr{G}_{s'})\subseteq V_2$,
the sets $V_1$ and $V_2$ each contain two vertices of both colors $(C,1)$
and $(C,2)$. Since $(C,1), (C,2)\in \sigma(V)$ are arbitrary and
$\sigma(V)= \mathfrak{C}\times\{1,2\}$, we can conclude that
$\sigma(V_1)=\sigma(V_2) = \sigma(V)$, and that $V_1$ and $V_2$ contain
at least two vertices of every color. Now, $\sigma(V_1)=\sigma(V_2)$
implies that $U_2=\emptyset$. Moreover, since $V_1$ and $V_2$ contain at
least two vertices of every color, we also have that $U_3=\emptyset$. In
summary, we have
$U(\mathaccent"017E{G},\mathscr{V})=U_1\charfusion[\mathbin]{\cup}{\cdot} U_2\charfusion[\mathbin]{\cup}{\cdot} U_3 = \emptyset$, and thus,
$c(\mathaccent"017E{G},\mathscr{V})=0$. Therefore, \PROBLEM{BPURC} with input
$(\mathaccent"017E{G},\sigma,k=0)$ has a yes-answer.
Now suppose \PROBLEM{BPURC} with input $(\mathaccent"017E{G},\sigma,k=0)$ has a yes-answer
and thus, a solution $\mathscr{V}=\{V_1,V_2\}$. Consequently,
$U(\mathaccent"017E{G},\mathscr{V})=U_1\charfusion[\mathbin]{\cup}{\cdot} U_2\charfusion[\mathbin]{\cup}{\cdot} U_3 =\emptyset$. We first show
that both $V_1$ and $V_2$ must contain a vertex of every color in
$\sigma(V) = \mathfrak{C}\times\{1,2\}$. To this end, we assume for
contradiction that w.l.o.g.\ $V_1$ contains no vertex of color $(C,1)$
for some $C\in\mathfrak{C}$. Since $|C|\ge 2$, $C$ contains two distinct
elements $s, s'\in S$. Note that $C\in \mathfrak{C}(s)$ and
$C\in \mathfrak{C}(s')$. By construction in Step 1, there are vertices
$y\in V(\mathscr{G}_s)$ and $y'\in V(\mathscr{G}_{s}')$ of color
$\sigma(y)=\sigma(y')=(C,1)$. Since $(C,1)\notin\sigma(V_1)$, it must
hold that $y,y'\in V_2$. Now consider an arbitrary vertex $x\in V_1$.
Note that $(C,1)\notin\sigma(V_1)$ implies $\sigma(x)\ne(C,1)$. Since
$\mathscr{G}_s$ and $\mathscr{G}_{s'}$ are, by construction, vertex
disjoint, $x$ cannot belong two both gadgets $\mathscr G_s$ and
$\mathscr{G}_{s'}$. Therefore, we can choose $\tilde{y}\in\{y,y'\}$ such
that $x$ and $\tilde{y}$ belong to distinct gadgets, and we obtain
$(x,\tilde{y})\notin E$ by construction. This together with $x\in V_1$,
$\tilde{y}\in V_2=V\setminus V_1$ and
$\sigma(\tilde{y})=(C,1)\notin\sigma(V_1)$ implies
$(x,\tilde{y})\in U_2$. Hence, $U_2\neq \emptyset$; a contradiction.
Therefore, we conclude that both $V_1$ and $V_2$ contain vertices of all
colors in $\sigma(V)=\mathfrak{C}\times\{1,2\}$.
We continue by showing that $V(\mathscr{G}_s)$ is entirely contained in
either $V_1$ or $V_2$ for all $s\in S$. To this end, assume for
contradiction that there is a gadget $\mathscr{G}_s$ such that
$W_1\coloneqq V_1\cap V(\mathscr{G}_s)$ and
$W_2\coloneqq V_2\cap V(\mathscr{G}_s)$ are both non-empty. Since
$V(\mathscr{G}_s)$ forms a connected component in $(\mathaccent"017E{G},\sigma)$ and all
arcs are bidirectional by construction, we can find two vertices
$x\in W_1$ and $y\in W_2$ such that $(x,y)\in E$. This together with the
facts that $x$ and $y$ are in distinct sets $V_1$ and $V_2$ and that both
$V_1$ and $V_2$ contain all colors of $\sigma(V)$, implies that
$(x,y)\in U_1$. Hence, $U_1\neq \emptyset$; a contradiction. Therefore,
the vertex set of each $s$-gadget is entirely contained in either $V_1$
or $V_2$.
We can construct a well-defined partition $\{S_1,S_2\}$ of $S$ such that
$s\in S_i$ if and only if $V(\mathscr G_s)\subseteq V_{i}$,
$i\in\{1,2\}$. By construction, there are vertices of color $(C,1)$ and
$(C,2)$ in $\mathscr{G}_s$ if and only if $s\in C$. This together with
the fact that both $V_1$ and $V_2$ contain vertices of all colors
$\mathfrak{C}\times\{1,2\}$ implies that $S_1\cap C$ and $S_2\cap C$ are
both non-empty for every $C\in\mathfrak{C}$. Hence, $\{S_1, S_2\}$ is a
solution for \PROBLEM{Set Splitting} with input $(\mathfrak{C},S)$. \qed
\end{proof}
\clearpage
\section{BPMF is not a consistent heuristic for \PROBLEM{MaxRTC}}
\label{sect:app-BPMF}
The example in Fig.~\ref{fig:BPMF-not-consistent} shows that
Alg.~\ref{alg:simple} in combination with the Best-Pair-Merge-First (BPMF)
heuristic \cite{Wu:04,Byrka:10} is not a consistent heuristic for BMG editing.
In particular, BPMF is not consistent for \PROBLEM{MaxRTC}.
The input graph $(\mathaccent"017E{G}_\textrm{orig}=(V,E),\sigma)$ is a BMG and explained by
$(T_\textrm{orig},\sigma)$.
Therefore, its set of informative triples
\begin{equation*}
\mathscr{R}\coloneqq
\mathscr{R}(\mathaccent"017E{G}_\textrm{orig},\sigma) =
\{ ab_1|b_2,\, ac_1|c_2,\, ac_1|c_3,\,
b_1c_1|c_2,\, b_1c_1|c_3,\, b_2c_1|c_2,\, b_2c_1|c_3,\, c_1b_2|b_1\}
\end{equation*}
is consistent (cf.\ Prop.~\ref{prop:BMG-charac}). On the right-hand side
of Fig.~\ref{fig:BPMF-not-consistent}, the first three cluster merging
steps in BPMF with input $\mathscr{R}$ are shown where the numbers are the
scores $score(S_i, S_j)$ for each pair of clusters $S_i$ and $S_j$ as defined
in \cite{Byrka:10}. The pink arrows link inner vertices of the resulting
binary tree $(T,\sigma)$ and the corresponding cluster merging step based
on the maximal score. The tree $(T,\sigma)$ does not display the triple
$ab_1|b_2$. As a consequence, its BMG $\mathaccent"017E{G}(T,\sigma)$ contains the
additional arc $(a,b_2)$, and the triple set $\mathscr{R}^*$ extracted from
$T$ in Alg.~\ref{alg:simple} is a proper subset of $\mathscr{R}$. In
particular, the final editing result $\mathaccent"017E{G}(T^*,\sigma)$ with
$T^*=\Aho(\mathscr{R}^*,V)$ also contains the arc $(a,b_2)$ which was
not present in the original BMG.
\begin{figure}[htb]
\centering
\includegraphics[width=0.85\textwidth]{./BPMF-not-consistent.pdf}
\caption{Example showing that BPMF is not a consistent heuristic for
\PROBLEM{MaxRTC}, and that Alg.~\ref{alg:simple} with BPMF is not a
consistent heuristic for BMG editing. See the text for a detailed
description.}
\label{fig:BPMF-not-consistent}
\end{figure}
\clearpage
\section{Analysis of single-leaf splits}
\label{sect:app-single}
Fig.~\ref{fig:single-vertex-cut} quantifies the abundance of single-leaf
splits on the same instances as in Fig.~\ref{fig:part-quality}. We
distinguish between single-leaf splits that are correct w.r.t.\ the Aho
graph $H_\textrm{orig}$ of the orginal unperturbed graph, and single-leaf
splits that are not present in the unperturbed target. MinCut, Karger,
Simple Greedy and Gradient Walk frequently produce single-leaf splits that
are not present in $H_\textrm{orig}$. The modularity-based Louvain method,
in contrast, never returned a single-leaf split, even if it was present in
$H_\textrm{orig}$. The modified Louvain method is most often in good
agreement with $H_\textrm{orig}$ as far as single-leaf splits are
concerned, at least for pertubation levels of 10\% of insertions and
deletions.
\begin{figure}[htb]
\centering
\includegraphics[width=0.85\textwidth]{./single-vertex-cut.pdf}
\caption{Abundance of single-leaf splits for pairs of BMGs
$(\mathaccent"017E{G}_{\textrm{orig}},\sigma)$ and disturbed graphs $(\mathaccent"017E{G},\sigma)$ (both
with vertex set $V$). The partition $\mathscr{V}_{\textrm{orig}}$
corresponds to the connected components of the Aho graph
$H_\textrm{orig}\coloneqq [\mathscr{R}(\mathaccent"017E{G}_{\textrm{orig}},\sigma), V]$
and, hence, to the partition induced by the subtrees of the children of
the root of the LRT $(T,\sigma)$ of $(\mathaccent"017E{G}_{\textrm{orig}},\sigma)$ (cf.\
Prop.~\ref{prop:BMG-charac}). The partition $\mathscr{V}_\textrm{heur}$
corresponds to the partition of $V$ as determined by one of the
partitioning methods (based on $H\coloneqq[\mathscr{R}(\mathaccent"017E{G},\sigma), V]$).
The gray parts of the bars comprise those instances for which $H$ is
disconnected. The light and dark red bars indicate the amount of graphs
for which only $\mathscr{V}_{\textrm{orig}}$ or
$\mathscr{V}_\textrm{heur}$, resp., is a single-leaf split, while light
and dark green bars represent instances for which both and none of the
two partitions, resp., are single-leaf splits. Note that the partitions
were not compared explicitly, in particular, the identified singletons
in $\mathscr{V}_\textrm{heur}$ in the light green instances may deviate
from those in $\mathscr{V}_{\textrm{orig}}$ in some cases. Example plot
for $|V|=30$ vertices and $|\sigma(V)|=10$ colors in each graph. 200
generated graph pairs per combination of arc insertion (ins.) and
deletion (del.) probabilities.}
\label{fig:single-vertex-cut}
\end{figure}
\clearpage
\section{Performance analysis in beBMG editing}
\label{sect:app-beBMG}
\begin{figure}[htb]
\centering
\includegraphics[width=0.85\textwidth]{./edit_binary1.pdf}
\caption{Performance comparison of several beBMG editing heuristics based
on the no.\ of arc differences. See Fig.~\ref{fig:edit_comparison1} for
further description. Example plot for $|V|=30$ vertices and
$|\sigma(V)|=10$ colors in each graph, 100 graphs per row.}
\label{fig:edit-comparison-binary}
\end{figure}
\end{appendix}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.